Loading...
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 try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
37#include <linux/compaction.h>
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
40#include <linux/delay.h>
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43#include <linux/memcontrol.h>
44#include <linux/delayacct.h>
45#include <linux/sysctl.h>
46#include <linux/oom.h>
47#include <linux/pagevec.h>
48#include <linux/prefetch.h>
49#include <linux/printk.h>
50#include <linux/dax.h>
51#include <linux/psi.h>
52
53#include <asm/tlbflush.h>
54#include <asm/div64.h>
55
56#include <linux/swapops.h>
57#include <linux/balloon_compaction.h>
58
59#include "internal.h"
60
61#define CREATE_TRACE_POINTS
62#include <trace/events/vmscan.h>
63
64struct scan_control {
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
68 /*
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
70 * are scanned.
71 */
72 nodemask_t *nodemask;
73
74 /*
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
77 */
78 struct mem_cgroup *target_mem_cgroup;
79
80 /*
81 * Scan pressure balancing between anon and file LRUs
82 */
83 unsigned long anon_cost;
84 unsigned long file_cost;
85
86 /* Can active pages be deactivated as part of reclaim? */
87#define DEACTIVATE_ANON 1
88#define DEACTIVATE_FILE 2
89 unsigned int may_deactivate:2;
90 unsigned int force_deactivate:1;
91 unsigned int skipped_deactivate:1;
92
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage:1;
95
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap:1;
98
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap:1;
101
102 /*
103 * Cgroup memory below memory.low is protected as long as we
104 * don't threaten to OOM. If any cgroup is reclaimed at
105 * reduced force or passed over entirely due to its memory.low
106 * setting (memcg_low_skipped), and nothing is reclaimed as a
107 * result, then go back for one more cycle that reclaims the protected
108 * memory (memcg_low_reclaim) to avert OOM.
109 */
110 unsigned int memcg_low_reclaim:1;
111 unsigned int memcg_low_skipped:1;
112
113 unsigned int hibernation_mode:1;
114
115 /* One of the zones is ready for compaction */
116 unsigned int compaction_ready:1;
117
118 /* There is easily reclaimable cold cache in the current node */
119 unsigned int cache_trim_mode:1;
120
121 /* The file pages on the current node are dangerously low */
122 unsigned int file_is_tiny:1;
123
124 /* Allocation order */
125 s8 order;
126
127 /* Scan (total_size >> priority) pages at once */
128 s8 priority;
129
130 /* The highest zone to isolate pages for reclaim from */
131 s8 reclaim_idx;
132
133 /* This context's GFP mask */
134 gfp_t gfp_mask;
135
136 /* Incremented by the number of inactive pages that were scanned */
137 unsigned long nr_scanned;
138
139 /* Number of pages freed so far during a call to shrink_zones() */
140 unsigned long nr_reclaimed;
141
142 struct {
143 unsigned int dirty;
144 unsigned int unqueued_dirty;
145 unsigned int congested;
146 unsigned int writeback;
147 unsigned int immediate;
148 unsigned int file_taken;
149 unsigned int taken;
150 } nr;
151
152 /* for recording the reclaimed slab by now */
153 struct reclaim_state reclaim_state;
154};
155
156#ifdef ARCH_HAS_PREFETCHW
157#define prefetchw_prev_lru_page(_page, _base, _field) \
158 do { \
159 if ((_page)->lru.prev != _base) { \
160 struct page *prev; \
161 \
162 prev = lru_to_page(&(_page->lru)); \
163 prefetchw(&prev->_field); \
164 } \
165 } while (0)
166#else
167#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168#endif
169
170/*
171 * From 0 .. 200. Higher means more swappy.
172 */
173int vm_swappiness = 60;
174
175static void set_task_reclaim_state(struct task_struct *task,
176 struct reclaim_state *rs)
177{
178 /* Check for an overwrite */
179 WARN_ON_ONCE(rs && task->reclaim_state);
180
181 /* Check for the nulling of an already-nulled member */
182 WARN_ON_ONCE(!rs && !task->reclaim_state);
183
184 task->reclaim_state = rs;
185}
186
187static LIST_HEAD(shrinker_list);
188static DECLARE_RWSEM(shrinker_rwsem);
189
190#ifdef CONFIG_MEMCG
191static int shrinker_nr_max;
192
193/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
194static inline int shrinker_map_size(int nr_items)
195{
196 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
197}
198
199static inline int shrinker_defer_size(int nr_items)
200{
201 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
202}
203
204static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
205 int nid)
206{
207 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
208 lockdep_is_held(&shrinker_rwsem));
209}
210
211static int expand_one_shrinker_info(struct mem_cgroup *memcg,
212 int map_size, int defer_size,
213 int old_map_size, int old_defer_size)
214{
215 struct shrinker_info *new, *old;
216 struct mem_cgroup_per_node *pn;
217 int nid;
218 int size = map_size + defer_size;
219
220 for_each_node(nid) {
221 pn = memcg->nodeinfo[nid];
222 old = shrinker_info_protected(memcg, nid);
223 /* Not yet online memcg */
224 if (!old)
225 return 0;
226
227 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
228 if (!new)
229 return -ENOMEM;
230
231 new->nr_deferred = (atomic_long_t *)(new + 1);
232 new->map = (void *)new->nr_deferred + defer_size;
233
234 /* map: set all old bits, clear all new bits */
235 memset(new->map, (int)0xff, old_map_size);
236 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
237 /* nr_deferred: copy old values, clear all new values */
238 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
239 memset((void *)new->nr_deferred + old_defer_size, 0,
240 defer_size - old_defer_size);
241
242 rcu_assign_pointer(pn->shrinker_info, new);
243 kvfree_rcu(old, rcu);
244 }
245
246 return 0;
247}
248
249void free_shrinker_info(struct mem_cgroup *memcg)
250{
251 struct mem_cgroup_per_node *pn;
252 struct shrinker_info *info;
253 int nid;
254
255 for_each_node(nid) {
256 pn = memcg->nodeinfo[nid];
257 info = rcu_dereference_protected(pn->shrinker_info, true);
258 kvfree(info);
259 rcu_assign_pointer(pn->shrinker_info, NULL);
260 }
261}
262
263int alloc_shrinker_info(struct mem_cgroup *memcg)
264{
265 struct shrinker_info *info;
266 int nid, size, ret = 0;
267 int map_size, defer_size = 0;
268
269 down_write(&shrinker_rwsem);
270 map_size = shrinker_map_size(shrinker_nr_max);
271 defer_size = shrinker_defer_size(shrinker_nr_max);
272 size = map_size + defer_size;
273 for_each_node(nid) {
274 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
275 if (!info) {
276 free_shrinker_info(memcg);
277 ret = -ENOMEM;
278 break;
279 }
280 info->nr_deferred = (atomic_long_t *)(info + 1);
281 info->map = (void *)info->nr_deferred + defer_size;
282 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
283 }
284 up_write(&shrinker_rwsem);
285
286 return ret;
287}
288
289static inline bool need_expand(int nr_max)
290{
291 return round_up(nr_max, BITS_PER_LONG) >
292 round_up(shrinker_nr_max, BITS_PER_LONG);
293}
294
295static int expand_shrinker_info(int new_id)
296{
297 int ret = 0;
298 int new_nr_max = new_id + 1;
299 int map_size, defer_size = 0;
300 int old_map_size, old_defer_size = 0;
301 struct mem_cgroup *memcg;
302
303 if (!need_expand(new_nr_max))
304 goto out;
305
306 if (!root_mem_cgroup)
307 goto out;
308
309 lockdep_assert_held(&shrinker_rwsem);
310
311 map_size = shrinker_map_size(new_nr_max);
312 defer_size = shrinker_defer_size(new_nr_max);
313 old_map_size = shrinker_map_size(shrinker_nr_max);
314 old_defer_size = shrinker_defer_size(shrinker_nr_max);
315
316 memcg = mem_cgroup_iter(NULL, NULL, NULL);
317 do {
318 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
319 old_map_size, old_defer_size);
320 if (ret) {
321 mem_cgroup_iter_break(NULL, memcg);
322 goto out;
323 }
324 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
325out:
326 if (!ret)
327 shrinker_nr_max = new_nr_max;
328
329 return ret;
330}
331
332void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
333{
334 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
335 struct shrinker_info *info;
336
337 rcu_read_lock();
338 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
339 /* Pairs with smp mb in shrink_slab() */
340 smp_mb__before_atomic();
341 set_bit(shrinker_id, info->map);
342 rcu_read_unlock();
343 }
344}
345
346static DEFINE_IDR(shrinker_idr);
347
348static int prealloc_memcg_shrinker(struct shrinker *shrinker)
349{
350 int id, ret = -ENOMEM;
351
352 if (mem_cgroup_disabled())
353 return -ENOSYS;
354
355 down_write(&shrinker_rwsem);
356 /* This may call shrinker, so it must use down_read_trylock() */
357 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
358 if (id < 0)
359 goto unlock;
360
361 if (id >= shrinker_nr_max) {
362 if (expand_shrinker_info(id)) {
363 idr_remove(&shrinker_idr, id);
364 goto unlock;
365 }
366 }
367 shrinker->id = id;
368 ret = 0;
369unlock:
370 up_write(&shrinker_rwsem);
371 return ret;
372}
373
374static void unregister_memcg_shrinker(struct shrinker *shrinker)
375{
376 int id = shrinker->id;
377
378 BUG_ON(id < 0);
379
380 lockdep_assert_held(&shrinker_rwsem);
381
382 idr_remove(&shrinker_idr, id);
383}
384
385static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
386 struct mem_cgroup *memcg)
387{
388 struct shrinker_info *info;
389
390 info = shrinker_info_protected(memcg, nid);
391 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
392}
393
394static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
395 struct mem_cgroup *memcg)
396{
397 struct shrinker_info *info;
398
399 info = shrinker_info_protected(memcg, nid);
400 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
401}
402
403void reparent_shrinker_deferred(struct mem_cgroup *memcg)
404{
405 int i, nid;
406 long nr;
407 struct mem_cgroup *parent;
408 struct shrinker_info *child_info, *parent_info;
409
410 parent = parent_mem_cgroup(memcg);
411 if (!parent)
412 parent = root_mem_cgroup;
413
414 /* Prevent from concurrent shrinker_info expand */
415 down_read(&shrinker_rwsem);
416 for_each_node(nid) {
417 child_info = shrinker_info_protected(memcg, nid);
418 parent_info = shrinker_info_protected(parent, nid);
419 for (i = 0; i < shrinker_nr_max; i++) {
420 nr = atomic_long_read(&child_info->nr_deferred[i]);
421 atomic_long_add(nr, &parent_info->nr_deferred[i]);
422 }
423 }
424 up_read(&shrinker_rwsem);
425}
426
427static bool cgroup_reclaim(struct scan_control *sc)
428{
429 return sc->target_mem_cgroup;
430}
431
432/**
433 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
434 * @sc: scan_control in question
435 *
436 * The normal page dirty throttling mechanism in balance_dirty_pages() is
437 * completely broken with the legacy memcg and direct stalling in
438 * shrink_page_list() is used for throttling instead, which lacks all the
439 * niceties such as fairness, adaptive pausing, bandwidth proportional
440 * allocation and configurability.
441 *
442 * This function tests whether the vmscan currently in progress can assume
443 * that the normal dirty throttling mechanism is operational.
444 */
445static bool writeback_throttling_sane(struct scan_control *sc)
446{
447 if (!cgroup_reclaim(sc))
448 return true;
449#ifdef CONFIG_CGROUP_WRITEBACK
450 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
451 return true;
452#endif
453 return false;
454}
455#else
456static int prealloc_memcg_shrinker(struct shrinker *shrinker)
457{
458 return -ENOSYS;
459}
460
461static void unregister_memcg_shrinker(struct shrinker *shrinker)
462{
463}
464
465static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
466 struct mem_cgroup *memcg)
467{
468 return 0;
469}
470
471static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
472 struct mem_cgroup *memcg)
473{
474 return 0;
475}
476
477static bool cgroup_reclaim(struct scan_control *sc)
478{
479 return false;
480}
481
482static bool writeback_throttling_sane(struct scan_control *sc)
483{
484 return true;
485}
486#endif
487
488static long xchg_nr_deferred(struct shrinker *shrinker,
489 struct shrink_control *sc)
490{
491 int nid = sc->nid;
492
493 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
494 nid = 0;
495
496 if (sc->memcg &&
497 (shrinker->flags & SHRINKER_MEMCG_AWARE))
498 return xchg_nr_deferred_memcg(nid, shrinker,
499 sc->memcg);
500
501 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
502}
503
504
505static long add_nr_deferred(long nr, struct shrinker *shrinker,
506 struct shrink_control *sc)
507{
508 int nid = sc->nid;
509
510 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
511 nid = 0;
512
513 if (sc->memcg &&
514 (shrinker->flags & SHRINKER_MEMCG_AWARE))
515 return add_nr_deferred_memcg(nr, nid, shrinker,
516 sc->memcg);
517
518 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
519}
520
521/*
522 * This misses isolated pages which are not accounted for to save counters.
523 * As the data only determines if reclaim or compaction continues, it is
524 * not expected that isolated pages will be a dominating factor.
525 */
526unsigned long zone_reclaimable_pages(struct zone *zone)
527{
528 unsigned long nr;
529
530 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
531 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
532 if (get_nr_swap_pages() > 0)
533 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
534 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
535
536 return nr;
537}
538
539/**
540 * lruvec_lru_size - Returns the number of pages on the given LRU list.
541 * @lruvec: lru vector
542 * @lru: lru to use
543 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
544 */
545static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
546 int zone_idx)
547{
548 unsigned long size = 0;
549 int zid;
550
551 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
552 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
553
554 if (!managed_zone(zone))
555 continue;
556
557 if (!mem_cgroup_disabled())
558 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
559 else
560 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
561 }
562 return size;
563}
564
565/*
566 * Add a shrinker callback to be called from the vm.
567 */
568int prealloc_shrinker(struct shrinker *shrinker)
569{
570 unsigned int size;
571 int err;
572
573 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
574 err = prealloc_memcg_shrinker(shrinker);
575 if (err != -ENOSYS)
576 return err;
577
578 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
579 }
580
581 size = sizeof(*shrinker->nr_deferred);
582 if (shrinker->flags & SHRINKER_NUMA_AWARE)
583 size *= nr_node_ids;
584
585 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
586 if (!shrinker->nr_deferred)
587 return -ENOMEM;
588
589 return 0;
590}
591
592void free_prealloced_shrinker(struct shrinker *shrinker)
593{
594 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
595 down_write(&shrinker_rwsem);
596 unregister_memcg_shrinker(shrinker);
597 up_write(&shrinker_rwsem);
598 return;
599 }
600
601 kfree(shrinker->nr_deferred);
602 shrinker->nr_deferred = NULL;
603}
604
605void register_shrinker_prepared(struct shrinker *shrinker)
606{
607 down_write(&shrinker_rwsem);
608 list_add_tail(&shrinker->list, &shrinker_list);
609 shrinker->flags |= SHRINKER_REGISTERED;
610 up_write(&shrinker_rwsem);
611}
612
613int register_shrinker(struct shrinker *shrinker)
614{
615 int err = prealloc_shrinker(shrinker);
616
617 if (err)
618 return err;
619 register_shrinker_prepared(shrinker);
620 return 0;
621}
622EXPORT_SYMBOL(register_shrinker);
623
624/*
625 * Remove one
626 */
627void unregister_shrinker(struct shrinker *shrinker)
628{
629 if (!(shrinker->flags & SHRINKER_REGISTERED))
630 return;
631
632 down_write(&shrinker_rwsem);
633 list_del(&shrinker->list);
634 shrinker->flags &= ~SHRINKER_REGISTERED;
635 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
636 unregister_memcg_shrinker(shrinker);
637 up_write(&shrinker_rwsem);
638
639 kfree(shrinker->nr_deferred);
640 shrinker->nr_deferred = NULL;
641}
642EXPORT_SYMBOL(unregister_shrinker);
643
644#define SHRINK_BATCH 128
645
646static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
647 struct shrinker *shrinker, int priority)
648{
649 unsigned long freed = 0;
650 unsigned long long delta;
651 long total_scan;
652 long freeable;
653 long nr;
654 long new_nr;
655 long batch_size = shrinker->batch ? shrinker->batch
656 : SHRINK_BATCH;
657 long scanned = 0, next_deferred;
658
659 freeable = shrinker->count_objects(shrinker, shrinkctl);
660 if (freeable == 0 || freeable == SHRINK_EMPTY)
661 return freeable;
662
663 /*
664 * copy the current shrinker scan count into a local variable
665 * and zero it so that other concurrent shrinker invocations
666 * don't also do this scanning work.
667 */
668 nr = xchg_nr_deferred(shrinker, shrinkctl);
669
670 if (shrinker->seeks) {
671 delta = freeable >> priority;
672 delta *= 4;
673 do_div(delta, shrinker->seeks);
674 } else {
675 /*
676 * These objects don't require any IO to create. Trim
677 * them aggressively under memory pressure to keep
678 * them from causing refetches in the IO caches.
679 */
680 delta = freeable / 2;
681 }
682
683 total_scan = nr >> priority;
684 total_scan += delta;
685 total_scan = min(total_scan, (2 * freeable));
686
687 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
688 freeable, delta, total_scan, priority);
689
690 /*
691 * Normally, we should not scan less than batch_size objects in one
692 * pass to avoid too frequent shrinker calls, but if the slab has less
693 * than batch_size objects in total and we are really tight on memory,
694 * we will try to reclaim all available objects, otherwise we can end
695 * up failing allocations although there are plenty of reclaimable
696 * objects spread over several slabs with usage less than the
697 * batch_size.
698 *
699 * We detect the "tight on memory" situations by looking at the total
700 * number of objects we want to scan (total_scan). If it is greater
701 * than the total number of objects on slab (freeable), we must be
702 * scanning at high prio and therefore should try to reclaim as much as
703 * possible.
704 */
705 while (total_scan >= batch_size ||
706 total_scan >= freeable) {
707 unsigned long ret;
708 unsigned long nr_to_scan = min(batch_size, total_scan);
709
710 shrinkctl->nr_to_scan = nr_to_scan;
711 shrinkctl->nr_scanned = nr_to_scan;
712 ret = shrinker->scan_objects(shrinker, shrinkctl);
713 if (ret == SHRINK_STOP)
714 break;
715 freed += ret;
716
717 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
718 total_scan -= shrinkctl->nr_scanned;
719 scanned += shrinkctl->nr_scanned;
720
721 cond_resched();
722 }
723
724 /*
725 * The deferred work is increased by any new work (delta) that wasn't
726 * done, decreased by old deferred work that was done now.
727 *
728 * And it is capped to two times of the freeable items.
729 */
730 next_deferred = max_t(long, (nr + delta - scanned), 0);
731 next_deferred = min(next_deferred, (2 * freeable));
732
733 /*
734 * move the unused scan count back into the shrinker in a
735 * manner that handles concurrent updates.
736 */
737 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
738
739 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
740 return freed;
741}
742
743#ifdef CONFIG_MEMCG
744static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
745 struct mem_cgroup *memcg, int priority)
746{
747 struct shrinker_info *info;
748 unsigned long ret, freed = 0;
749 int i;
750
751 if (!mem_cgroup_online(memcg))
752 return 0;
753
754 if (!down_read_trylock(&shrinker_rwsem))
755 return 0;
756
757 info = shrinker_info_protected(memcg, nid);
758 if (unlikely(!info))
759 goto unlock;
760
761 for_each_set_bit(i, info->map, shrinker_nr_max) {
762 struct shrink_control sc = {
763 .gfp_mask = gfp_mask,
764 .nid = nid,
765 .memcg = memcg,
766 };
767 struct shrinker *shrinker;
768
769 shrinker = idr_find(&shrinker_idr, i);
770 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
771 if (!shrinker)
772 clear_bit(i, info->map);
773 continue;
774 }
775
776 /* Call non-slab shrinkers even though kmem is disabled */
777 if (!memcg_kmem_enabled() &&
778 !(shrinker->flags & SHRINKER_NONSLAB))
779 continue;
780
781 ret = do_shrink_slab(&sc, shrinker, priority);
782 if (ret == SHRINK_EMPTY) {
783 clear_bit(i, info->map);
784 /*
785 * After the shrinker reported that it had no objects to
786 * free, but before we cleared the corresponding bit in
787 * the memcg shrinker map, a new object might have been
788 * added. To make sure, we have the bit set in this
789 * case, we invoke the shrinker one more time and reset
790 * the bit if it reports that it is not empty anymore.
791 * The memory barrier here pairs with the barrier in
792 * set_shrinker_bit():
793 *
794 * list_lru_add() shrink_slab_memcg()
795 * list_add_tail() clear_bit()
796 * <MB> <MB>
797 * set_bit() do_shrink_slab()
798 */
799 smp_mb__after_atomic();
800 ret = do_shrink_slab(&sc, shrinker, priority);
801 if (ret == SHRINK_EMPTY)
802 ret = 0;
803 else
804 set_shrinker_bit(memcg, nid, i);
805 }
806 freed += ret;
807
808 if (rwsem_is_contended(&shrinker_rwsem)) {
809 freed = freed ? : 1;
810 break;
811 }
812 }
813unlock:
814 up_read(&shrinker_rwsem);
815 return freed;
816}
817#else /* CONFIG_MEMCG */
818static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
819 struct mem_cgroup *memcg, int priority)
820{
821 return 0;
822}
823#endif /* CONFIG_MEMCG */
824
825/**
826 * shrink_slab - shrink slab caches
827 * @gfp_mask: allocation context
828 * @nid: node whose slab caches to target
829 * @memcg: memory cgroup whose slab caches to target
830 * @priority: the reclaim priority
831 *
832 * Call the shrink functions to age shrinkable caches.
833 *
834 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
835 * unaware shrinkers will receive a node id of 0 instead.
836 *
837 * @memcg specifies the memory cgroup to target. Unaware shrinkers
838 * are called only if it is the root cgroup.
839 *
840 * @priority is sc->priority, we take the number of objects and >> by priority
841 * in order to get the scan target.
842 *
843 * Returns the number of reclaimed slab objects.
844 */
845static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
846 struct mem_cgroup *memcg,
847 int priority)
848{
849 unsigned long ret, freed = 0;
850 struct shrinker *shrinker;
851
852 /*
853 * The root memcg might be allocated even though memcg is disabled
854 * via "cgroup_disable=memory" boot parameter. This could make
855 * mem_cgroup_is_root() return false, then just run memcg slab
856 * shrink, but skip global shrink. This may result in premature
857 * oom.
858 */
859 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
860 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
861
862 if (!down_read_trylock(&shrinker_rwsem))
863 goto out;
864
865 list_for_each_entry(shrinker, &shrinker_list, list) {
866 struct shrink_control sc = {
867 .gfp_mask = gfp_mask,
868 .nid = nid,
869 .memcg = memcg,
870 };
871
872 ret = do_shrink_slab(&sc, shrinker, priority);
873 if (ret == SHRINK_EMPTY)
874 ret = 0;
875 freed += ret;
876 /*
877 * Bail out if someone want to register a new shrinker to
878 * prevent the registration from being stalled for long periods
879 * by parallel ongoing shrinking.
880 */
881 if (rwsem_is_contended(&shrinker_rwsem)) {
882 freed = freed ? : 1;
883 break;
884 }
885 }
886
887 up_read(&shrinker_rwsem);
888out:
889 cond_resched();
890 return freed;
891}
892
893void drop_slab_node(int nid)
894{
895 unsigned long freed;
896
897 do {
898 struct mem_cgroup *memcg = NULL;
899
900 if (fatal_signal_pending(current))
901 return;
902
903 freed = 0;
904 memcg = mem_cgroup_iter(NULL, NULL, NULL);
905 do {
906 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
907 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
908 } while (freed > 10);
909}
910
911void drop_slab(void)
912{
913 int nid;
914
915 for_each_online_node(nid)
916 drop_slab_node(nid);
917}
918
919static inline int is_page_cache_freeable(struct page *page)
920{
921 /*
922 * A freeable page cache page is referenced only by the caller
923 * that isolated the page, the page cache and optional buffer
924 * heads at page->private.
925 */
926 int page_cache_pins = thp_nr_pages(page);
927 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
928}
929
930static int may_write_to_inode(struct inode *inode)
931{
932 if (current->flags & PF_SWAPWRITE)
933 return 1;
934 if (!inode_write_congested(inode))
935 return 1;
936 if (inode_to_bdi(inode) == current->backing_dev_info)
937 return 1;
938 return 0;
939}
940
941/*
942 * We detected a synchronous write error writing a page out. Probably
943 * -ENOSPC. We need to propagate that into the address_space for a subsequent
944 * fsync(), msync() or close().
945 *
946 * The tricky part is that after writepage we cannot touch the mapping: nothing
947 * prevents it from being freed up. But we have a ref on the page and once
948 * that page is locked, the mapping is pinned.
949 *
950 * We're allowed to run sleeping lock_page() here because we know the caller has
951 * __GFP_FS.
952 */
953static void handle_write_error(struct address_space *mapping,
954 struct page *page, int error)
955{
956 lock_page(page);
957 if (page_mapping(page) == mapping)
958 mapping_set_error(mapping, error);
959 unlock_page(page);
960}
961
962/* possible outcome of pageout() */
963typedef enum {
964 /* failed to write page out, page is locked */
965 PAGE_KEEP,
966 /* move page to the active list, page is locked */
967 PAGE_ACTIVATE,
968 /* page has been sent to the disk successfully, page is unlocked */
969 PAGE_SUCCESS,
970 /* page is clean and locked */
971 PAGE_CLEAN,
972} pageout_t;
973
974/*
975 * pageout is called by shrink_page_list() for each dirty page.
976 * Calls ->writepage().
977 */
978static pageout_t pageout(struct page *page, struct address_space *mapping)
979{
980 /*
981 * If the page is dirty, only perform writeback if that write
982 * will be non-blocking. To prevent this allocation from being
983 * stalled by pagecache activity. But note that there may be
984 * stalls if we need to run get_block(). We could test
985 * PagePrivate for that.
986 *
987 * If this process is currently in __generic_file_write_iter() against
988 * this page's queue, we can perform writeback even if that
989 * will block.
990 *
991 * If the page is swapcache, write it back even if that would
992 * block, for some throttling. This happens by accident, because
993 * swap_backing_dev_info is bust: it doesn't reflect the
994 * congestion state of the swapdevs. Easy to fix, if needed.
995 */
996 if (!is_page_cache_freeable(page))
997 return PAGE_KEEP;
998 if (!mapping) {
999 /*
1000 * Some data journaling orphaned pages can have
1001 * page->mapping == NULL while being dirty with clean buffers.
1002 */
1003 if (page_has_private(page)) {
1004 if (try_to_free_buffers(page)) {
1005 ClearPageDirty(page);
1006 pr_info("%s: orphaned page\n", __func__);
1007 return PAGE_CLEAN;
1008 }
1009 }
1010 return PAGE_KEEP;
1011 }
1012 if (mapping->a_ops->writepage == NULL)
1013 return PAGE_ACTIVATE;
1014 if (!may_write_to_inode(mapping->host))
1015 return PAGE_KEEP;
1016
1017 if (clear_page_dirty_for_io(page)) {
1018 int res;
1019 struct writeback_control wbc = {
1020 .sync_mode = WB_SYNC_NONE,
1021 .nr_to_write = SWAP_CLUSTER_MAX,
1022 .range_start = 0,
1023 .range_end = LLONG_MAX,
1024 .for_reclaim = 1,
1025 };
1026
1027 SetPageReclaim(page);
1028 res = mapping->a_ops->writepage(page, &wbc);
1029 if (res < 0)
1030 handle_write_error(mapping, page, res);
1031 if (res == AOP_WRITEPAGE_ACTIVATE) {
1032 ClearPageReclaim(page);
1033 return PAGE_ACTIVATE;
1034 }
1035
1036 if (!PageWriteback(page)) {
1037 /* synchronous write or broken a_ops? */
1038 ClearPageReclaim(page);
1039 }
1040 trace_mm_vmscan_writepage(page);
1041 inc_node_page_state(page, NR_VMSCAN_WRITE);
1042 return PAGE_SUCCESS;
1043 }
1044
1045 return PAGE_CLEAN;
1046}
1047
1048/*
1049 * Same as remove_mapping, but if the page is removed from the mapping, it
1050 * gets returned with a refcount of 0.
1051 */
1052static int __remove_mapping(struct address_space *mapping, struct page *page,
1053 bool reclaimed, struct mem_cgroup *target_memcg)
1054{
1055 unsigned long flags;
1056 int refcount;
1057 void *shadow = NULL;
1058
1059 BUG_ON(!PageLocked(page));
1060 BUG_ON(mapping != page_mapping(page));
1061
1062 xa_lock_irqsave(&mapping->i_pages, flags);
1063 /*
1064 * The non racy check for a busy page.
1065 *
1066 * Must be careful with the order of the tests. When someone has
1067 * a ref to the page, it may be possible that they dirty it then
1068 * drop the reference. So if PageDirty is tested before page_count
1069 * here, then the following race may occur:
1070 *
1071 * get_user_pages(&page);
1072 * [user mapping goes away]
1073 * write_to(page);
1074 * !PageDirty(page) [good]
1075 * SetPageDirty(page);
1076 * put_page(page);
1077 * !page_count(page) [good, discard it]
1078 *
1079 * [oops, our write_to data is lost]
1080 *
1081 * Reversing the order of the tests ensures such a situation cannot
1082 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1083 * load is not satisfied before that of page->_refcount.
1084 *
1085 * Note that if SetPageDirty is always performed via set_page_dirty,
1086 * and thus under the i_pages lock, then this ordering is not required.
1087 */
1088 refcount = 1 + compound_nr(page);
1089 if (!page_ref_freeze(page, refcount))
1090 goto cannot_free;
1091 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1092 if (unlikely(PageDirty(page))) {
1093 page_ref_unfreeze(page, refcount);
1094 goto cannot_free;
1095 }
1096
1097 if (PageSwapCache(page)) {
1098 swp_entry_t swap = { .val = page_private(page) };
1099 mem_cgroup_swapout(page, swap);
1100 if (reclaimed && !mapping_exiting(mapping))
1101 shadow = workingset_eviction(page, target_memcg);
1102 __delete_from_swap_cache(page, swap, shadow);
1103 xa_unlock_irqrestore(&mapping->i_pages, flags);
1104 put_swap_page(page, swap);
1105 } else {
1106 void (*freepage)(struct page *);
1107
1108 freepage = mapping->a_ops->freepage;
1109 /*
1110 * Remember a shadow entry for reclaimed file cache in
1111 * order to detect refaults, thus thrashing, later on.
1112 *
1113 * But don't store shadows in an address space that is
1114 * already exiting. This is not just an optimization,
1115 * inode reclaim needs to empty out the radix tree or
1116 * the nodes are lost. Don't plant shadows behind its
1117 * back.
1118 *
1119 * We also don't store shadows for DAX mappings because the
1120 * only page cache pages found in these are zero pages
1121 * covering holes, and because we don't want to mix DAX
1122 * exceptional entries and shadow exceptional entries in the
1123 * same address_space.
1124 */
1125 if (reclaimed && page_is_file_lru(page) &&
1126 !mapping_exiting(mapping) && !dax_mapping(mapping))
1127 shadow = workingset_eviction(page, target_memcg);
1128 __delete_from_page_cache(page, shadow);
1129 xa_unlock_irqrestore(&mapping->i_pages, flags);
1130
1131 if (freepage != NULL)
1132 freepage(page);
1133 }
1134
1135 return 1;
1136
1137cannot_free:
1138 xa_unlock_irqrestore(&mapping->i_pages, flags);
1139 return 0;
1140}
1141
1142/*
1143 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1144 * someone else has a ref on the page, abort and return 0. If it was
1145 * successfully detached, return 1. Assumes the caller has a single ref on
1146 * this page.
1147 */
1148int remove_mapping(struct address_space *mapping, struct page *page)
1149{
1150 if (__remove_mapping(mapping, page, false, NULL)) {
1151 /*
1152 * Unfreezing the refcount with 1 rather than 2 effectively
1153 * drops the pagecache ref for us without requiring another
1154 * atomic operation.
1155 */
1156 page_ref_unfreeze(page, 1);
1157 return 1;
1158 }
1159 return 0;
1160}
1161
1162/**
1163 * putback_lru_page - put previously isolated page onto appropriate LRU list
1164 * @page: page to be put back to appropriate lru list
1165 *
1166 * Add previously isolated @page to appropriate LRU list.
1167 * Page may still be unevictable for other reasons.
1168 *
1169 * lru_lock must not be held, interrupts must be enabled.
1170 */
1171void putback_lru_page(struct page *page)
1172{
1173 lru_cache_add(page);
1174 put_page(page); /* drop ref from isolate */
1175}
1176
1177enum page_references {
1178 PAGEREF_RECLAIM,
1179 PAGEREF_RECLAIM_CLEAN,
1180 PAGEREF_KEEP,
1181 PAGEREF_ACTIVATE,
1182};
1183
1184static enum page_references page_check_references(struct page *page,
1185 struct scan_control *sc)
1186{
1187 int referenced_ptes, referenced_page;
1188 unsigned long vm_flags;
1189
1190 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1191 &vm_flags);
1192 referenced_page = TestClearPageReferenced(page);
1193
1194 /*
1195 * Mlock lost the isolation race with us. Let try_to_unmap()
1196 * move the page to the unevictable list.
1197 */
1198 if (vm_flags & VM_LOCKED)
1199 return PAGEREF_RECLAIM;
1200
1201 if (referenced_ptes) {
1202 /*
1203 * All mapped pages start out with page table
1204 * references from the instantiating fault, so we need
1205 * to look twice if a mapped file page is used more
1206 * than once.
1207 *
1208 * Mark it and spare it for another trip around the
1209 * inactive list. Another page table reference will
1210 * lead to its activation.
1211 *
1212 * Note: the mark is set for activated pages as well
1213 * so that recently deactivated but used pages are
1214 * quickly recovered.
1215 */
1216 SetPageReferenced(page);
1217
1218 if (referenced_page || referenced_ptes > 1)
1219 return PAGEREF_ACTIVATE;
1220
1221 /*
1222 * Activate file-backed executable pages after first usage.
1223 */
1224 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1225 return PAGEREF_ACTIVATE;
1226
1227 return PAGEREF_KEEP;
1228 }
1229
1230 /* Reclaim if clean, defer dirty pages to writeback */
1231 if (referenced_page && !PageSwapBacked(page))
1232 return PAGEREF_RECLAIM_CLEAN;
1233
1234 return PAGEREF_RECLAIM;
1235}
1236
1237/* Check if a page is dirty or under writeback */
1238static void page_check_dirty_writeback(struct page *page,
1239 bool *dirty, bool *writeback)
1240{
1241 struct address_space *mapping;
1242
1243 /*
1244 * Anonymous pages are not handled by flushers and must be written
1245 * from reclaim context. Do not stall reclaim based on them
1246 */
1247 if (!page_is_file_lru(page) ||
1248 (PageAnon(page) && !PageSwapBacked(page))) {
1249 *dirty = false;
1250 *writeback = false;
1251 return;
1252 }
1253
1254 /* By default assume that the page flags are accurate */
1255 *dirty = PageDirty(page);
1256 *writeback = PageWriteback(page);
1257
1258 /* Verify dirty/writeback state if the filesystem supports it */
1259 if (!page_has_private(page))
1260 return;
1261
1262 mapping = page_mapping(page);
1263 if (mapping && mapping->a_ops->is_dirty_writeback)
1264 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1265}
1266
1267/*
1268 * shrink_page_list() returns the number of reclaimed pages
1269 */
1270static unsigned int shrink_page_list(struct list_head *page_list,
1271 struct pglist_data *pgdat,
1272 struct scan_control *sc,
1273 struct reclaim_stat *stat,
1274 bool ignore_references)
1275{
1276 LIST_HEAD(ret_pages);
1277 LIST_HEAD(free_pages);
1278 unsigned int nr_reclaimed = 0;
1279 unsigned int pgactivate = 0;
1280
1281 memset(stat, 0, sizeof(*stat));
1282 cond_resched();
1283
1284 while (!list_empty(page_list)) {
1285 struct address_space *mapping;
1286 struct page *page;
1287 enum page_references references = PAGEREF_RECLAIM;
1288 bool dirty, writeback, may_enter_fs;
1289 unsigned int nr_pages;
1290
1291 cond_resched();
1292
1293 page = lru_to_page(page_list);
1294 list_del(&page->lru);
1295
1296 if (!trylock_page(page))
1297 goto keep;
1298
1299 VM_BUG_ON_PAGE(PageActive(page), page);
1300
1301 nr_pages = compound_nr(page);
1302
1303 /* Account the number of base pages even though THP */
1304 sc->nr_scanned += nr_pages;
1305
1306 if (unlikely(!page_evictable(page)))
1307 goto activate_locked;
1308
1309 if (!sc->may_unmap && page_mapped(page))
1310 goto keep_locked;
1311
1312 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1313 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1314
1315 /*
1316 * The number of dirty pages determines if a node is marked
1317 * reclaim_congested which affects wait_iff_congested. kswapd
1318 * will stall and start writing pages if the tail of the LRU
1319 * is all dirty unqueued pages.
1320 */
1321 page_check_dirty_writeback(page, &dirty, &writeback);
1322 if (dirty || writeback)
1323 stat->nr_dirty++;
1324
1325 if (dirty && !writeback)
1326 stat->nr_unqueued_dirty++;
1327
1328 /*
1329 * Treat this page as congested if the underlying BDI is or if
1330 * pages are cycling through the LRU so quickly that the
1331 * pages marked for immediate reclaim are making it to the
1332 * end of the LRU a second time.
1333 */
1334 mapping = page_mapping(page);
1335 if (((dirty || writeback) && mapping &&
1336 inode_write_congested(mapping->host)) ||
1337 (writeback && PageReclaim(page)))
1338 stat->nr_congested++;
1339
1340 /*
1341 * If a page at the tail of the LRU is under writeback, there
1342 * are three cases to consider.
1343 *
1344 * 1) If reclaim is encountering an excessive number of pages
1345 * under writeback and this page is both under writeback and
1346 * PageReclaim then it indicates that pages are being queued
1347 * for IO but are being recycled through the LRU before the
1348 * IO can complete. Waiting on the page itself risks an
1349 * indefinite stall if it is impossible to writeback the
1350 * page due to IO error or disconnected storage so instead
1351 * note that the LRU is being scanned too quickly and the
1352 * caller can stall after page list has been processed.
1353 *
1354 * 2) Global or new memcg reclaim encounters a page that is
1355 * not marked for immediate reclaim, or the caller does not
1356 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1357 * not to fs). In this case mark the page for immediate
1358 * reclaim and continue scanning.
1359 *
1360 * Require may_enter_fs because we would wait on fs, which
1361 * may not have submitted IO yet. And the loop driver might
1362 * enter reclaim, and deadlock if it waits on a page for
1363 * which it is needed to do the write (loop masks off
1364 * __GFP_IO|__GFP_FS for this reason); but more thought
1365 * would probably show more reasons.
1366 *
1367 * 3) Legacy memcg encounters a page that is already marked
1368 * PageReclaim. memcg does not have any dirty pages
1369 * throttling so we could easily OOM just because too many
1370 * pages are in writeback and there is nothing else to
1371 * reclaim. Wait for the writeback to complete.
1372 *
1373 * In cases 1) and 2) we activate the pages to get them out of
1374 * the way while we continue scanning for clean pages on the
1375 * inactive list and refilling from the active list. The
1376 * observation here is that waiting for disk writes is more
1377 * expensive than potentially causing reloads down the line.
1378 * Since they're marked for immediate reclaim, they won't put
1379 * memory pressure on the cache working set any longer than it
1380 * takes to write them to disk.
1381 */
1382 if (PageWriteback(page)) {
1383 /* Case 1 above */
1384 if (current_is_kswapd() &&
1385 PageReclaim(page) &&
1386 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1387 stat->nr_immediate++;
1388 goto activate_locked;
1389
1390 /* Case 2 above */
1391 } else if (writeback_throttling_sane(sc) ||
1392 !PageReclaim(page) || !may_enter_fs) {
1393 /*
1394 * This is slightly racy - end_page_writeback()
1395 * might have just cleared PageReclaim, then
1396 * setting PageReclaim here end up interpreted
1397 * as PageReadahead - but that does not matter
1398 * enough to care. What we do want is for this
1399 * page to have PageReclaim set next time memcg
1400 * reclaim reaches the tests above, so it will
1401 * then wait_on_page_writeback() to avoid OOM;
1402 * and it's also appropriate in global reclaim.
1403 */
1404 SetPageReclaim(page);
1405 stat->nr_writeback++;
1406 goto activate_locked;
1407
1408 /* Case 3 above */
1409 } else {
1410 unlock_page(page);
1411 wait_on_page_writeback(page);
1412 /* then go back and try same page again */
1413 list_add_tail(&page->lru, page_list);
1414 continue;
1415 }
1416 }
1417
1418 if (!ignore_references)
1419 references = page_check_references(page, sc);
1420
1421 switch (references) {
1422 case PAGEREF_ACTIVATE:
1423 goto activate_locked;
1424 case PAGEREF_KEEP:
1425 stat->nr_ref_keep += nr_pages;
1426 goto keep_locked;
1427 case PAGEREF_RECLAIM:
1428 case PAGEREF_RECLAIM_CLEAN:
1429 ; /* try to reclaim the page below */
1430 }
1431
1432 /*
1433 * Anonymous process memory has backing store?
1434 * Try to allocate it some swap space here.
1435 * Lazyfree page could be freed directly
1436 */
1437 if (PageAnon(page) && PageSwapBacked(page)) {
1438 if (!PageSwapCache(page)) {
1439 if (!(sc->gfp_mask & __GFP_IO))
1440 goto keep_locked;
1441 if (page_maybe_dma_pinned(page))
1442 goto keep_locked;
1443 if (PageTransHuge(page)) {
1444 /* cannot split THP, skip it */
1445 if (!can_split_huge_page(page, NULL))
1446 goto activate_locked;
1447 /*
1448 * Split pages without a PMD map right
1449 * away. Chances are some or all of the
1450 * tail pages can be freed without IO.
1451 */
1452 if (!compound_mapcount(page) &&
1453 split_huge_page_to_list(page,
1454 page_list))
1455 goto activate_locked;
1456 }
1457 if (!add_to_swap(page)) {
1458 if (!PageTransHuge(page))
1459 goto activate_locked_split;
1460 /* Fallback to swap normal pages */
1461 if (split_huge_page_to_list(page,
1462 page_list))
1463 goto activate_locked;
1464#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1465 count_vm_event(THP_SWPOUT_FALLBACK);
1466#endif
1467 if (!add_to_swap(page))
1468 goto activate_locked_split;
1469 }
1470
1471 may_enter_fs = true;
1472
1473 /* Adding to swap updated mapping */
1474 mapping = page_mapping(page);
1475 }
1476 } else if (unlikely(PageTransHuge(page))) {
1477 /* Split file THP */
1478 if (split_huge_page_to_list(page, page_list))
1479 goto keep_locked;
1480 }
1481
1482 /*
1483 * THP may get split above, need minus tail pages and update
1484 * nr_pages to avoid accounting tail pages twice.
1485 *
1486 * The tail pages that are added into swap cache successfully
1487 * reach here.
1488 */
1489 if ((nr_pages > 1) && !PageTransHuge(page)) {
1490 sc->nr_scanned -= (nr_pages - 1);
1491 nr_pages = 1;
1492 }
1493
1494 /*
1495 * The page is mapped into the page tables of one or more
1496 * processes. Try to unmap it here.
1497 */
1498 if (page_mapped(page)) {
1499 enum ttu_flags flags = TTU_BATCH_FLUSH;
1500 bool was_swapbacked = PageSwapBacked(page);
1501
1502 if (unlikely(PageTransHuge(page)))
1503 flags |= TTU_SPLIT_HUGE_PMD;
1504
1505 try_to_unmap(page, flags);
1506 if (page_mapped(page)) {
1507 stat->nr_unmap_fail += nr_pages;
1508 if (!was_swapbacked && PageSwapBacked(page))
1509 stat->nr_lazyfree_fail += nr_pages;
1510 goto activate_locked;
1511 }
1512 }
1513
1514 if (PageDirty(page)) {
1515 /*
1516 * Only kswapd can writeback filesystem pages
1517 * to avoid risk of stack overflow. But avoid
1518 * injecting inefficient single-page IO into
1519 * flusher writeback as much as possible: only
1520 * write pages when we've encountered many
1521 * dirty pages, and when we've already scanned
1522 * the rest of the LRU for clean pages and see
1523 * the same dirty pages again (PageReclaim).
1524 */
1525 if (page_is_file_lru(page) &&
1526 (!current_is_kswapd() || !PageReclaim(page) ||
1527 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1528 /*
1529 * Immediately reclaim when written back.
1530 * Similar in principal to deactivate_page()
1531 * except we already have the page isolated
1532 * and know it's dirty
1533 */
1534 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1535 SetPageReclaim(page);
1536
1537 goto activate_locked;
1538 }
1539
1540 if (references == PAGEREF_RECLAIM_CLEAN)
1541 goto keep_locked;
1542 if (!may_enter_fs)
1543 goto keep_locked;
1544 if (!sc->may_writepage)
1545 goto keep_locked;
1546
1547 /*
1548 * Page is dirty. Flush the TLB if a writable entry
1549 * potentially exists to avoid CPU writes after IO
1550 * starts and then write it out here.
1551 */
1552 try_to_unmap_flush_dirty();
1553 switch (pageout(page, mapping)) {
1554 case PAGE_KEEP:
1555 goto keep_locked;
1556 case PAGE_ACTIVATE:
1557 goto activate_locked;
1558 case PAGE_SUCCESS:
1559 stat->nr_pageout += thp_nr_pages(page);
1560
1561 if (PageWriteback(page))
1562 goto keep;
1563 if (PageDirty(page))
1564 goto keep;
1565
1566 /*
1567 * A synchronous write - probably a ramdisk. Go
1568 * ahead and try to reclaim the page.
1569 */
1570 if (!trylock_page(page))
1571 goto keep;
1572 if (PageDirty(page) || PageWriteback(page))
1573 goto keep_locked;
1574 mapping = page_mapping(page);
1575 fallthrough;
1576 case PAGE_CLEAN:
1577 ; /* try to free the page below */
1578 }
1579 }
1580
1581 /*
1582 * If the page has buffers, try to free the buffer mappings
1583 * associated with this page. If we succeed we try to free
1584 * the page as well.
1585 *
1586 * We do this even if the page is PageDirty().
1587 * try_to_release_page() does not perform I/O, but it is
1588 * possible for a page to have PageDirty set, but it is actually
1589 * clean (all its buffers are clean). This happens if the
1590 * buffers were written out directly, with submit_bh(). ext3
1591 * will do this, as well as the blockdev mapping.
1592 * try_to_release_page() will discover that cleanness and will
1593 * drop the buffers and mark the page clean - it can be freed.
1594 *
1595 * Rarely, pages can have buffers and no ->mapping. These are
1596 * the pages which were not successfully invalidated in
1597 * truncate_cleanup_page(). We try to drop those buffers here
1598 * and if that worked, and the page is no longer mapped into
1599 * process address space (page_count == 1) it can be freed.
1600 * Otherwise, leave the page on the LRU so it is swappable.
1601 */
1602 if (page_has_private(page)) {
1603 if (!try_to_release_page(page, sc->gfp_mask))
1604 goto activate_locked;
1605 if (!mapping && page_count(page) == 1) {
1606 unlock_page(page);
1607 if (put_page_testzero(page))
1608 goto free_it;
1609 else {
1610 /*
1611 * rare race with speculative reference.
1612 * the speculative reference will free
1613 * this page shortly, so we may
1614 * increment nr_reclaimed here (and
1615 * leave it off the LRU).
1616 */
1617 nr_reclaimed++;
1618 continue;
1619 }
1620 }
1621 }
1622
1623 if (PageAnon(page) && !PageSwapBacked(page)) {
1624 /* follow __remove_mapping for reference */
1625 if (!page_ref_freeze(page, 1))
1626 goto keep_locked;
1627 if (PageDirty(page)) {
1628 page_ref_unfreeze(page, 1);
1629 goto keep_locked;
1630 }
1631
1632 count_vm_event(PGLAZYFREED);
1633 count_memcg_page_event(page, PGLAZYFREED);
1634 } else if (!mapping || !__remove_mapping(mapping, page, true,
1635 sc->target_mem_cgroup))
1636 goto keep_locked;
1637
1638 unlock_page(page);
1639free_it:
1640 /*
1641 * THP may get swapped out in a whole, need account
1642 * all base pages.
1643 */
1644 nr_reclaimed += nr_pages;
1645
1646 /*
1647 * Is there need to periodically free_page_list? It would
1648 * appear not as the counts should be low
1649 */
1650 if (unlikely(PageTransHuge(page)))
1651 destroy_compound_page(page);
1652 else
1653 list_add(&page->lru, &free_pages);
1654 continue;
1655
1656activate_locked_split:
1657 /*
1658 * The tail pages that are failed to add into swap cache
1659 * reach here. Fixup nr_scanned and nr_pages.
1660 */
1661 if (nr_pages > 1) {
1662 sc->nr_scanned -= (nr_pages - 1);
1663 nr_pages = 1;
1664 }
1665activate_locked:
1666 /* Not a candidate for swapping, so reclaim swap space. */
1667 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1668 PageMlocked(page)))
1669 try_to_free_swap(page);
1670 VM_BUG_ON_PAGE(PageActive(page), page);
1671 if (!PageMlocked(page)) {
1672 int type = page_is_file_lru(page);
1673 SetPageActive(page);
1674 stat->nr_activate[type] += nr_pages;
1675 count_memcg_page_event(page, PGACTIVATE);
1676 }
1677keep_locked:
1678 unlock_page(page);
1679keep:
1680 list_add(&page->lru, &ret_pages);
1681 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1682 }
1683
1684 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1685
1686 mem_cgroup_uncharge_list(&free_pages);
1687 try_to_unmap_flush();
1688 free_unref_page_list(&free_pages);
1689
1690 list_splice(&ret_pages, page_list);
1691 count_vm_events(PGACTIVATE, pgactivate);
1692
1693 return nr_reclaimed;
1694}
1695
1696unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1697 struct list_head *page_list)
1698{
1699 struct scan_control sc = {
1700 .gfp_mask = GFP_KERNEL,
1701 .priority = DEF_PRIORITY,
1702 .may_unmap = 1,
1703 };
1704 struct reclaim_stat stat;
1705 unsigned int nr_reclaimed;
1706 struct page *page, *next;
1707 LIST_HEAD(clean_pages);
1708 unsigned int noreclaim_flag;
1709
1710 list_for_each_entry_safe(page, next, page_list, lru) {
1711 if (!PageHuge(page) && page_is_file_lru(page) &&
1712 !PageDirty(page) && !__PageMovable(page) &&
1713 !PageUnevictable(page)) {
1714 ClearPageActive(page);
1715 list_move(&page->lru, &clean_pages);
1716 }
1717 }
1718
1719 /*
1720 * We should be safe here since we are only dealing with file pages and
1721 * we are not kswapd and therefore cannot write dirty file pages. But
1722 * call memalloc_noreclaim_save() anyway, just in case these conditions
1723 * change in the future.
1724 */
1725 noreclaim_flag = memalloc_noreclaim_save();
1726 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1727 &stat, true);
1728 memalloc_noreclaim_restore(noreclaim_flag);
1729
1730 list_splice(&clean_pages, page_list);
1731 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1732 -(long)nr_reclaimed);
1733 /*
1734 * Since lazyfree pages are isolated from file LRU from the beginning,
1735 * they will rotate back to anonymous LRU in the end if it failed to
1736 * discard so isolated count will be mismatched.
1737 * Compensate the isolated count for both LRU lists.
1738 */
1739 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1740 stat.nr_lazyfree_fail);
1741 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1742 -(long)stat.nr_lazyfree_fail);
1743 return nr_reclaimed;
1744}
1745
1746/*
1747 * Attempt to remove the specified page from its LRU. Only take this page
1748 * if it is of the appropriate PageActive status. Pages which are being
1749 * freed elsewhere are also ignored.
1750 *
1751 * page: page to consider
1752 * mode: one of the LRU isolation modes defined above
1753 *
1754 * returns true on success, false on failure.
1755 */
1756bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1757{
1758 /* Only take pages on the LRU. */
1759 if (!PageLRU(page))
1760 return false;
1761
1762 /* Compaction should not handle unevictable pages but CMA can do so */
1763 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1764 return false;
1765
1766 /*
1767 * To minimise LRU disruption, the caller can indicate that it only
1768 * wants to isolate pages it will be able to operate on without
1769 * blocking - clean pages for the most part.
1770 *
1771 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1772 * that it is possible to migrate without blocking
1773 */
1774 if (mode & ISOLATE_ASYNC_MIGRATE) {
1775 /* All the caller can do on PageWriteback is block */
1776 if (PageWriteback(page))
1777 return false;
1778
1779 if (PageDirty(page)) {
1780 struct address_space *mapping;
1781 bool migrate_dirty;
1782
1783 /*
1784 * Only pages without mappings or that have a
1785 * ->migratepage callback are possible to migrate
1786 * without blocking. However, we can be racing with
1787 * truncation so it's necessary to lock the page
1788 * to stabilise the mapping as truncation holds
1789 * the page lock until after the page is removed
1790 * from the page cache.
1791 */
1792 if (!trylock_page(page))
1793 return false;
1794
1795 mapping = page_mapping(page);
1796 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1797 unlock_page(page);
1798 if (!migrate_dirty)
1799 return false;
1800 }
1801 }
1802
1803 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1804 return false;
1805
1806 return true;
1807}
1808
1809/*
1810 * Update LRU sizes after isolating pages. The LRU size updates must
1811 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1812 */
1813static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1814 enum lru_list lru, unsigned long *nr_zone_taken)
1815{
1816 int zid;
1817
1818 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1819 if (!nr_zone_taken[zid])
1820 continue;
1821
1822 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1823 }
1824
1825}
1826
1827/*
1828 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1829 *
1830 * lruvec->lru_lock is heavily contended. Some of the functions that
1831 * shrink the lists perform better by taking out a batch of pages
1832 * and working on them outside the LRU lock.
1833 *
1834 * For pagecache intensive workloads, this function is the hottest
1835 * spot in the kernel (apart from copy_*_user functions).
1836 *
1837 * Lru_lock must be held before calling this function.
1838 *
1839 * @nr_to_scan: The number of eligible pages to look through on the list.
1840 * @lruvec: The LRU vector to pull pages from.
1841 * @dst: The temp list to put pages on to.
1842 * @nr_scanned: The number of pages that were scanned.
1843 * @sc: The scan_control struct for this reclaim session
1844 * @lru: LRU list id for isolating
1845 *
1846 * returns how many pages were moved onto *@dst.
1847 */
1848static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1849 struct lruvec *lruvec, struct list_head *dst,
1850 unsigned long *nr_scanned, struct scan_control *sc,
1851 enum lru_list lru)
1852{
1853 struct list_head *src = &lruvec->lists[lru];
1854 unsigned long nr_taken = 0;
1855 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1856 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1857 unsigned long skipped = 0;
1858 unsigned long scan, total_scan, nr_pages;
1859 LIST_HEAD(pages_skipped);
1860 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1861
1862 total_scan = 0;
1863 scan = 0;
1864 while (scan < nr_to_scan && !list_empty(src)) {
1865 struct page *page;
1866
1867 page = lru_to_page(src);
1868 prefetchw_prev_lru_page(page, src, flags);
1869
1870 nr_pages = compound_nr(page);
1871 total_scan += nr_pages;
1872
1873 if (page_zonenum(page) > sc->reclaim_idx) {
1874 list_move(&page->lru, &pages_skipped);
1875 nr_skipped[page_zonenum(page)] += nr_pages;
1876 continue;
1877 }
1878
1879 /*
1880 * Do not count skipped pages because that makes the function
1881 * return with no isolated pages if the LRU mostly contains
1882 * ineligible pages. This causes the VM to not reclaim any
1883 * pages, triggering a premature OOM.
1884 *
1885 * Account all tail pages of THP. This would not cause
1886 * premature OOM since __isolate_lru_page() returns -EBUSY
1887 * only when the page is being freed somewhere else.
1888 */
1889 scan += nr_pages;
1890 if (!__isolate_lru_page_prepare(page, mode)) {
1891 /* It is being freed elsewhere */
1892 list_move(&page->lru, src);
1893 continue;
1894 }
1895 /*
1896 * Be careful not to clear PageLRU until after we're
1897 * sure the page is not being freed elsewhere -- the
1898 * page release code relies on it.
1899 */
1900 if (unlikely(!get_page_unless_zero(page))) {
1901 list_move(&page->lru, src);
1902 continue;
1903 }
1904
1905 if (!TestClearPageLRU(page)) {
1906 /* Another thread is already isolating this page */
1907 put_page(page);
1908 list_move(&page->lru, src);
1909 continue;
1910 }
1911
1912 nr_taken += nr_pages;
1913 nr_zone_taken[page_zonenum(page)] += nr_pages;
1914 list_move(&page->lru, dst);
1915 }
1916
1917 /*
1918 * Splice any skipped pages to the start of the LRU list. Note that
1919 * this disrupts the LRU order when reclaiming for lower zones but
1920 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1921 * scanning would soon rescan the same pages to skip and put the
1922 * system at risk of premature OOM.
1923 */
1924 if (!list_empty(&pages_skipped)) {
1925 int zid;
1926
1927 list_splice(&pages_skipped, src);
1928 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1929 if (!nr_skipped[zid])
1930 continue;
1931
1932 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1933 skipped += nr_skipped[zid];
1934 }
1935 }
1936 *nr_scanned = total_scan;
1937 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1938 total_scan, skipped, nr_taken, mode, lru);
1939 update_lru_sizes(lruvec, lru, nr_zone_taken);
1940 return nr_taken;
1941}
1942
1943/**
1944 * isolate_lru_page - tries to isolate a page from its LRU list
1945 * @page: page to isolate from its LRU list
1946 *
1947 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1948 * vmstat statistic corresponding to whatever LRU list the page was on.
1949 *
1950 * Returns 0 if the page was removed from an LRU list.
1951 * Returns -EBUSY if the page was not on an LRU list.
1952 *
1953 * The returned page will have PageLRU() cleared. If it was found on
1954 * the active list, it will have PageActive set. If it was found on
1955 * the unevictable list, it will have the PageUnevictable bit set. That flag
1956 * may need to be cleared by the caller before letting the page go.
1957 *
1958 * The vmstat statistic corresponding to the list on which the page was
1959 * found will be decremented.
1960 *
1961 * Restrictions:
1962 *
1963 * (1) Must be called with an elevated refcount on the page. This is a
1964 * fundamental difference from isolate_lru_pages (which is called
1965 * without a stable reference).
1966 * (2) the lru_lock must not be held.
1967 * (3) interrupts must be enabled.
1968 */
1969int isolate_lru_page(struct page *page)
1970{
1971 int ret = -EBUSY;
1972
1973 VM_BUG_ON_PAGE(!page_count(page), page);
1974 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1975
1976 if (TestClearPageLRU(page)) {
1977 struct lruvec *lruvec;
1978
1979 get_page(page);
1980 lruvec = lock_page_lruvec_irq(page);
1981 del_page_from_lru_list(page, lruvec);
1982 unlock_page_lruvec_irq(lruvec);
1983 ret = 0;
1984 }
1985
1986 return ret;
1987}
1988
1989/*
1990 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1991 * then get rescheduled. When there are massive number of tasks doing page
1992 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1993 * the LRU list will go small and be scanned faster than necessary, leading to
1994 * unnecessary swapping, thrashing and OOM.
1995 */
1996static int too_many_isolated(struct pglist_data *pgdat, int file,
1997 struct scan_control *sc)
1998{
1999 unsigned long inactive, isolated;
2000
2001 if (current_is_kswapd())
2002 return 0;
2003
2004 if (!writeback_throttling_sane(sc))
2005 return 0;
2006
2007 if (file) {
2008 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2009 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2010 } else {
2011 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2012 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2013 }
2014
2015 /*
2016 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2017 * won't get blocked by normal direct-reclaimers, forming a circular
2018 * deadlock.
2019 */
2020 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2021 inactive >>= 3;
2022
2023 return isolated > inactive;
2024}
2025
2026/*
2027 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2028 * On return, @list is reused as a list of pages to be freed by the caller.
2029 *
2030 * Returns the number of pages moved to the given lruvec.
2031 */
2032static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2033 struct list_head *list)
2034{
2035 int nr_pages, nr_moved = 0;
2036 LIST_HEAD(pages_to_free);
2037 struct page *page;
2038
2039 while (!list_empty(list)) {
2040 page = lru_to_page(list);
2041 VM_BUG_ON_PAGE(PageLRU(page), page);
2042 list_del(&page->lru);
2043 if (unlikely(!page_evictable(page))) {
2044 spin_unlock_irq(&lruvec->lru_lock);
2045 putback_lru_page(page);
2046 spin_lock_irq(&lruvec->lru_lock);
2047 continue;
2048 }
2049
2050 /*
2051 * The SetPageLRU needs to be kept here for list integrity.
2052 * Otherwise:
2053 * #0 move_pages_to_lru #1 release_pages
2054 * if !put_page_testzero
2055 * if (put_page_testzero())
2056 * !PageLRU //skip lru_lock
2057 * SetPageLRU()
2058 * list_add(&page->lru,)
2059 * list_add(&page->lru,)
2060 */
2061 SetPageLRU(page);
2062
2063 if (unlikely(put_page_testzero(page))) {
2064 __clear_page_lru_flags(page);
2065
2066 if (unlikely(PageCompound(page))) {
2067 spin_unlock_irq(&lruvec->lru_lock);
2068 destroy_compound_page(page);
2069 spin_lock_irq(&lruvec->lru_lock);
2070 } else
2071 list_add(&page->lru, &pages_to_free);
2072
2073 continue;
2074 }
2075
2076 /*
2077 * All pages were isolated from the same lruvec (and isolation
2078 * inhibits memcg migration).
2079 */
2080 VM_BUG_ON_PAGE(!page_matches_lruvec(page, lruvec), page);
2081 add_page_to_lru_list(page, lruvec);
2082 nr_pages = thp_nr_pages(page);
2083 nr_moved += nr_pages;
2084 if (PageActive(page))
2085 workingset_age_nonresident(lruvec, nr_pages);
2086 }
2087
2088 /*
2089 * To save our caller's stack, now use input list for pages to free.
2090 */
2091 list_splice(&pages_to_free, list);
2092
2093 return nr_moved;
2094}
2095
2096/*
2097 * If a kernel thread (such as nfsd for loop-back mounts) services
2098 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2099 * In that case we should only throttle if the backing device it is
2100 * writing to is congested. In other cases it is safe to throttle.
2101 */
2102static int current_may_throttle(void)
2103{
2104 return !(current->flags & PF_LOCAL_THROTTLE) ||
2105 current->backing_dev_info == NULL ||
2106 bdi_write_congested(current->backing_dev_info);
2107}
2108
2109/*
2110 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2111 * of reclaimed pages
2112 */
2113static unsigned long
2114shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2115 struct scan_control *sc, enum lru_list lru)
2116{
2117 LIST_HEAD(page_list);
2118 unsigned long nr_scanned;
2119 unsigned int nr_reclaimed = 0;
2120 unsigned long nr_taken;
2121 struct reclaim_stat stat;
2122 bool file = is_file_lru(lru);
2123 enum vm_event_item item;
2124 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2125 bool stalled = false;
2126
2127 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2128 if (stalled)
2129 return 0;
2130
2131 /* wait a bit for the reclaimer. */
2132 msleep(100);
2133 stalled = true;
2134
2135 /* We are about to die and free our memory. Return now. */
2136 if (fatal_signal_pending(current))
2137 return SWAP_CLUSTER_MAX;
2138 }
2139
2140 lru_add_drain();
2141
2142 spin_lock_irq(&lruvec->lru_lock);
2143
2144 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2145 &nr_scanned, sc, lru);
2146
2147 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2148 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2149 if (!cgroup_reclaim(sc))
2150 __count_vm_events(item, nr_scanned);
2151 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2152 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2153
2154 spin_unlock_irq(&lruvec->lru_lock);
2155
2156 if (nr_taken == 0)
2157 return 0;
2158
2159 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2160
2161 spin_lock_irq(&lruvec->lru_lock);
2162 move_pages_to_lru(lruvec, &page_list);
2163
2164 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2165 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2166 if (!cgroup_reclaim(sc))
2167 __count_vm_events(item, nr_reclaimed);
2168 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2169 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2170 spin_unlock_irq(&lruvec->lru_lock);
2171
2172 lru_note_cost(lruvec, file, stat.nr_pageout);
2173 mem_cgroup_uncharge_list(&page_list);
2174 free_unref_page_list(&page_list);
2175
2176 /*
2177 * If dirty pages are scanned that are not queued for IO, it
2178 * implies that flushers are not doing their job. This can
2179 * happen when memory pressure pushes dirty pages to the end of
2180 * the LRU before the dirty limits are breached and the dirty
2181 * data has expired. It can also happen when the proportion of
2182 * dirty pages grows not through writes but through memory
2183 * pressure reclaiming all the clean cache. And in some cases,
2184 * the flushers simply cannot keep up with the allocation
2185 * rate. Nudge the flusher threads in case they are asleep.
2186 */
2187 if (stat.nr_unqueued_dirty == nr_taken)
2188 wakeup_flusher_threads(WB_REASON_VMSCAN);
2189
2190 sc->nr.dirty += stat.nr_dirty;
2191 sc->nr.congested += stat.nr_congested;
2192 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2193 sc->nr.writeback += stat.nr_writeback;
2194 sc->nr.immediate += stat.nr_immediate;
2195 sc->nr.taken += nr_taken;
2196 if (file)
2197 sc->nr.file_taken += nr_taken;
2198
2199 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2200 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2201 return nr_reclaimed;
2202}
2203
2204/*
2205 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2206 *
2207 * We move them the other way if the page is referenced by one or more
2208 * processes.
2209 *
2210 * If the pages are mostly unmapped, the processing is fast and it is
2211 * appropriate to hold lru_lock across the whole operation. But if
2212 * the pages are mapped, the processing is slow (page_referenced()), so
2213 * we should drop lru_lock around each page. It's impossible to balance
2214 * this, so instead we remove the pages from the LRU while processing them.
2215 * It is safe to rely on PG_active against the non-LRU pages in here because
2216 * nobody will play with that bit on a non-LRU page.
2217 *
2218 * The downside is that we have to touch page->_refcount against each page.
2219 * But we had to alter page->flags anyway.
2220 */
2221static void shrink_active_list(unsigned long nr_to_scan,
2222 struct lruvec *lruvec,
2223 struct scan_control *sc,
2224 enum lru_list lru)
2225{
2226 unsigned long nr_taken;
2227 unsigned long nr_scanned;
2228 unsigned long vm_flags;
2229 LIST_HEAD(l_hold); /* The pages which were snipped off */
2230 LIST_HEAD(l_active);
2231 LIST_HEAD(l_inactive);
2232 struct page *page;
2233 unsigned nr_deactivate, nr_activate;
2234 unsigned nr_rotated = 0;
2235 int file = is_file_lru(lru);
2236 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2237
2238 lru_add_drain();
2239
2240 spin_lock_irq(&lruvec->lru_lock);
2241
2242 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2243 &nr_scanned, sc, lru);
2244
2245 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2246
2247 if (!cgroup_reclaim(sc))
2248 __count_vm_events(PGREFILL, nr_scanned);
2249 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2250
2251 spin_unlock_irq(&lruvec->lru_lock);
2252
2253 while (!list_empty(&l_hold)) {
2254 cond_resched();
2255 page = lru_to_page(&l_hold);
2256 list_del(&page->lru);
2257
2258 if (unlikely(!page_evictable(page))) {
2259 putback_lru_page(page);
2260 continue;
2261 }
2262
2263 if (unlikely(buffer_heads_over_limit)) {
2264 if (page_has_private(page) && trylock_page(page)) {
2265 if (page_has_private(page))
2266 try_to_release_page(page, 0);
2267 unlock_page(page);
2268 }
2269 }
2270
2271 if (page_referenced(page, 0, sc->target_mem_cgroup,
2272 &vm_flags)) {
2273 /*
2274 * Identify referenced, file-backed active pages and
2275 * give them one more trip around the active list. So
2276 * that executable code get better chances to stay in
2277 * memory under moderate memory pressure. Anon pages
2278 * are not likely to be evicted by use-once streaming
2279 * IO, plus JVM can create lots of anon VM_EXEC pages,
2280 * so we ignore them here.
2281 */
2282 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2283 nr_rotated += thp_nr_pages(page);
2284 list_add(&page->lru, &l_active);
2285 continue;
2286 }
2287 }
2288
2289 ClearPageActive(page); /* we are de-activating */
2290 SetPageWorkingset(page);
2291 list_add(&page->lru, &l_inactive);
2292 }
2293
2294 /*
2295 * Move pages back to the lru list.
2296 */
2297 spin_lock_irq(&lruvec->lru_lock);
2298
2299 nr_activate = move_pages_to_lru(lruvec, &l_active);
2300 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2301 /* Keep all free pages in l_active list */
2302 list_splice(&l_inactive, &l_active);
2303
2304 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2305 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2306
2307 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2308 spin_unlock_irq(&lruvec->lru_lock);
2309
2310 mem_cgroup_uncharge_list(&l_active);
2311 free_unref_page_list(&l_active);
2312 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2313 nr_deactivate, nr_rotated, sc->priority, file);
2314}
2315
2316unsigned long reclaim_pages(struct list_head *page_list)
2317{
2318 int nid = NUMA_NO_NODE;
2319 unsigned int nr_reclaimed = 0;
2320 LIST_HEAD(node_page_list);
2321 struct reclaim_stat dummy_stat;
2322 struct page *page;
2323 unsigned int noreclaim_flag;
2324 struct scan_control sc = {
2325 .gfp_mask = GFP_KERNEL,
2326 .priority = DEF_PRIORITY,
2327 .may_writepage = 1,
2328 .may_unmap = 1,
2329 .may_swap = 1,
2330 };
2331
2332 noreclaim_flag = memalloc_noreclaim_save();
2333
2334 while (!list_empty(page_list)) {
2335 page = lru_to_page(page_list);
2336 if (nid == NUMA_NO_NODE) {
2337 nid = page_to_nid(page);
2338 INIT_LIST_HEAD(&node_page_list);
2339 }
2340
2341 if (nid == page_to_nid(page)) {
2342 ClearPageActive(page);
2343 list_move(&page->lru, &node_page_list);
2344 continue;
2345 }
2346
2347 nr_reclaimed += shrink_page_list(&node_page_list,
2348 NODE_DATA(nid),
2349 &sc, &dummy_stat, false);
2350 while (!list_empty(&node_page_list)) {
2351 page = lru_to_page(&node_page_list);
2352 list_del(&page->lru);
2353 putback_lru_page(page);
2354 }
2355
2356 nid = NUMA_NO_NODE;
2357 }
2358
2359 if (!list_empty(&node_page_list)) {
2360 nr_reclaimed += shrink_page_list(&node_page_list,
2361 NODE_DATA(nid),
2362 &sc, &dummy_stat, false);
2363 while (!list_empty(&node_page_list)) {
2364 page = lru_to_page(&node_page_list);
2365 list_del(&page->lru);
2366 putback_lru_page(page);
2367 }
2368 }
2369
2370 memalloc_noreclaim_restore(noreclaim_flag);
2371
2372 return nr_reclaimed;
2373}
2374
2375static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2376 struct lruvec *lruvec, struct scan_control *sc)
2377{
2378 if (is_active_lru(lru)) {
2379 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2380 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2381 else
2382 sc->skipped_deactivate = 1;
2383 return 0;
2384 }
2385
2386 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2387}
2388
2389/*
2390 * The inactive anon list should be small enough that the VM never has
2391 * to do too much work.
2392 *
2393 * The inactive file list should be small enough to leave most memory
2394 * to the established workingset on the scan-resistant active list,
2395 * but large enough to avoid thrashing the aggregate readahead window.
2396 *
2397 * Both inactive lists should also be large enough that each inactive
2398 * page has a chance to be referenced again before it is reclaimed.
2399 *
2400 * If that fails and refaulting is observed, the inactive list grows.
2401 *
2402 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2403 * on this LRU, maintained by the pageout code. An inactive_ratio
2404 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2405 *
2406 * total target max
2407 * memory ratio inactive
2408 * -------------------------------------
2409 * 10MB 1 5MB
2410 * 100MB 1 50MB
2411 * 1GB 3 250MB
2412 * 10GB 10 0.9GB
2413 * 100GB 31 3GB
2414 * 1TB 101 10GB
2415 * 10TB 320 32GB
2416 */
2417static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2418{
2419 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2420 unsigned long inactive, active;
2421 unsigned long inactive_ratio;
2422 unsigned long gb;
2423
2424 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2425 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2426
2427 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2428 if (gb)
2429 inactive_ratio = int_sqrt(10 * gb);
2430 else
2431 inactive_ratio = 1;
2432
2433 return inactive * inactive_ratio < active;
2434}
2435
2436enum scan_balance {
2437 SCAN_EQUAL,
2438 SCAN_FRACT,
2439 SCAN_ANON,
2440 SCAN_FILE,
2441};
2442
2443/*
2444 * Determine how aggressively the anon and file LRU lists should be
2445 * scanned. The relative value of each set of LRU lists is determined
2446 * by looking at the fraction of the pages scanned we did rotate back
2447 * onto the active list instead of evict.
2448 *
2449 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2450 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2451 */
2452static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2453 unsigned long *nr)
2454{
2455 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2456 unsigned long anon_cost, file_cost, total_cost;
2457 int swappiness = mem_cgroup_swappiness(memcg);
2458 u64 fraction[ANON_AND_FILE];
2459 u64 denominator = 0; /* gcc */
2460 enum scan_balance scan_balance;
2461 unsigned long ap, fp;
2462 enum lru_list lru;
2463
2464 /* If we have no swap space, do not bother scanning anon pages. */
2465 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2466 scan_balance = SCAN_FILE;
2467 goto out;
2468 }
2469
2470 /*
2471 * Global reclaim will swap to prevent OOM even with no
2472 * swappiness, but memcg users want to use this knob to
2473 * disable swapping for individual groups completely when
2474 * using the memory controller's swap limit feature would be
2475 * too expensive.
2476 */
2477 if (cgroup_reclaim(sc) && !swappiness) {
2478 scan_balance = SCAN_FILE;
2479 goto out;
2480 }
2481
2482 /*
2483 * Do not apply any pressure balancing cleverness when the
2484 * system is close to OOM, scan both anon and file equally
2485 * (unless the swappiness setting disagrees with swapping).
2486 */
2487 if (!sc->priority && swappiness) {
2488 scan_balance = SCAN_EQUAL;
2489 goto out;
2490 }
2491
2492 /*
2493 * If the system is almost out of file pages, force-scan anon.
2494 */
2495 if (sc->file_is_tiny) {
2496 scan_balance = SCAN_ANON;
2497 goto out;
2498 }
2499
2500 /*
2501 * If there is enough inactive page cache, we do not reclaim
2502 * anything from the anonymous working right now.
2503 */
2504 if (sc->cache_trim_mode) {
2505 scan_balance = SCAN_FILE;
2506 goto out;
2507 }
2508
2509 scan_balance = SCAN_FRACT;
2510 /*
2511 * Calculate the pressure balance between anon and file pages.
2512 *
2513 * The amount of pressure we put on each LRU is inversely
2514 * proportional to the cost of reclaiming each list, as
2515 * determined by the share of pages that are refaulting, times
2516 * the relative IO cost of bringing back a swapped out
2517 * anonymous page vs reloading a filesystem page (swappiness).
2518 *
2519 * Although we limit that influence to ensure no list gets
2520 * left behind completely: at least a third of the pressure is
2521 * applied, before swappiness.
2522 *
2523 * With swappiness at 100, anon and file have equal IO cost.
2524 */
2525 total_cost = sc->anon_cost + sc->file_cost;
2526 anon_cost = total_cost + sc->anon_cost;
2527 file_cost = total_cost + sc->file_cost;
2528 total_cost = anon_cost + file_cost;
2529
2530 ap = swappiness * (total_cost + 1);
2531 ap /= anon_cost + 1;
2532
2533 fp = (200 - swappiness) * (total_cost + 1);
2534 fp /= file_cost + 1;
2535
2536 fraction[0] = ap;
2537 fraction[1] = fp;
2538 denominator = ap + fp;
2539out:
2540 for_each_evictable_lru(lru) {
2541 int file = is_file_lru(lru);
2542 unsigned long lruvec_size;
2543 unsigned long low, min;
2544 unsigned long scan;
2545
2546 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2547 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2548 &min, &low);
2549
2550 if (min || low) {
2551 /*
2552 * Scale a cgroup's reclaim pressure by proportioning
2553 * its current usage to its memory.low or memory.min
2554 * setting.
2555 *
2556 * This is important, as otherwise scanning aggression
2557 * becomes extremely binary -- from nothing as we
2558 * approach the memory protection threshold, to totally
2559 * nominal as we exceed it. This results in requiring
2560 * setting extremely liberal protection thresholds. It
2561 * also means we simply get no protection at all if we
2562 * set it too low, which is not ideal.
2563 *
2564 * If there is any protection in place, we reduce scan
2565 * pressure by how much of the total memory used is
2566 * within protection thresholds.
2567 *
2568 * There is one special case: in the first reclaim pass,
2569 * we skip over all groups that are within their low
2570 * protection. If that fails to reclaim enough pages to
2571 * satisfy the reclaim goal, we come back and override
2572 * the best-effort low protection. However, we still
2573 * ideally want to honor how well-behaved groups are in
2574 * that case instead of simply punishing them all
2575 * equally. As such, we reclaim them based on how much
2576 * memory they are using, reducing the scan pressure
2577 * again by how much of the total memory used is under
2578 * hard protection.
2579 */
2580 unsigned long cgroup_size = mem_cgroup_size(memcg);
2581 unsigned long protection;
2582
2583 /* memory.low scaling, make sure we retry before OOM */
2584 if (!sc->memcg_low_reclaim && low > min) {
2585 protection = low;
2586 sc->memcg_low_skipped = 1;
2587 } else {
2588 protection = min;
2589 }
2590
2591 /* Avoid TOCTOU with earlier protection check */
2592 cgroup_size = max(cgroup_size, protection);
2593
2594 scan = lruvec_size - lruvec_size * protection /
2595 (cgroup_size + 1);
2596
2597 /*
2598 * Minimally target SWAP_CLUSTER_MAX pages to keep
2599 * reclaim moving forwards, avoiding decrementing
2600 * sc->priority further than desirable.
2601 */
2602 scan = max(scan, SWAP_CLUSTER_MAX);
2603 } else {
2604 scan = lruvec_size;
2605 }
2606
2607 scan >>= sc->priority;
2608
2609 /*
2610 * If the cgroup's already been deleted, make sure to
2611 * scrape out the remaining cache.
2612 */
2613 if (!scan && !mem_cgroup_online(memcg))
2614 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2615
2616 switch (scan_balance) {
2617 case SCAN_EQUAL:
2618 /* Scan lists relative to size */
2619 break;
2620 case SCAN_FRACT:
2621 /*
2622 * Scan types proportional to swappiness and
2623 * their relative recent reclaim efficiency.
2624 * Make sure we don't miss the last page on
2625 * the offlined memory cgroups because of a
2626 * round-off error.
2627 */
2628 scan = mem_cgroup_online(memcg) ?
2629 div64_u64(scan * fraction[file], denominator) :
2630 DIV64_U64_ROUND_UP(scan * fraction[file],
2631 denominator);
2632 break;
2633 case SCAN_FILE:
2634 case SCAN_ANON:
2635 /* Scan one type exclusively */
2636 if ((scan_balance == SCAN_FILE) != file)
2637 scan = 0;
2638 break;
2639 default:
2640 /* Look ma, no brain */
2641 BUG();
2642 }
2643
2644 nr[lru] = scan;
2645 }
2646}
2647
2648static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2649{
2650 unsigned long nr[NR_LRU_LISTS];
2651 unsigned long targets[NR_LRU_LISTS];
2652 unsigned long nr_to_scan;
2653 enum lru_list lru;
2654 unsigned long nr_reclaimed = 0;
2655 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2656 struct blk_plug plug;
2657 bool scan_adjusted;
2658
2659 get_scan_count(lruvec, sc, nr);
2660
2661 /* Record the original scan target for proportional adjustments later */
2662 memcpy(targets, nr, sizeof(nr));
2663
2664 /*
2665 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2666 * event that can occur when there is little memory pressure e.g.
2667 * multiple streaming readers/writers. Hence, we do not abort scanning
2668 * when the requested number of pages are reclaimed when scanning at
2669 * DEF_PRIORITY on the assumption that the fact we are direct
2670 * reclaiming implies that kswapd is not keeping up and it is best to
2671 * do a batch of work at once. For memcg reclaim one check is made to
2672 * abort proportional reclaim if either the file or anon lru has already
2673 * dropped to zero at the first pass.
2674 */
2675 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2676 sc->priority == DEF_PRIORITY);
2677
2678 blk_start_plug(&plug);
2679 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2680 nr[LRU_INACTIVE_FILE]) {
2681 unsigned long nr_anon, nr_file, percentage;
2682 unsigned long nr_scanned;
2683
2684 for_each_evictable_lru(lru) {
2685 if (nr[lru]) {
2686 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2687 nr[lru] -= nr_to_scan;
2688
2689 nr_reclaimed += shrink_list(lru, nr_to_scan,
2690 lruvec, sc);
2691 }
2692 }
2693
2694 cond_resched();
2695
2696 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2697 continue;
2698
2699 /*
2700 * For kswapd and memcg, reclaim at least the number of pages
2701 * requested. Ensure that the anon and file LRUs are scanned
2702 * proportionally what was requested by get_scan_count(). We
2703 * stop reclaiming one LRU and reduce the amount scanning
2704 * proportional to the original scan target.
2705 */
2706 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2707 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2708
2709 /*
2710 * It's just vindictive to attack the larger once the smaller
2711 * has gone to zero. And given the way we stop scanning the
2712 * smaller below, this makes sure that we only make one nudge
2713 * towards proportionality once we've got nr_to_reclaim.
2714 */
2715 if (!nr_file || !nr_anon)
2716 break;
2717
2718 if (nr_file > nr_anon) {
2719 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2720 targets[LRU_ACTIVE_ANON] + 1;
2721 lru = LRU_BASE;
2722 percentage = nr_anon * 100 / scan_target;
2723 } else {
2724 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2725 targets[LRU_ACTIVE_FILE] + 1;
2726 lru = LRU_FILE;
2727 percentage = nr_file * 100 / scan_target;
2728 }
2729
2730 /* Stop scanning the smaller of the LRU */
2731 nr[lru] = 0;
2732 nr[lru + LRU_ACTIVE] = 0;
2733
2734 /*
2735 * Recalculate the other LRU scan count based on its original
2736 * scan target and the percentage scanning already complete
2737 */
2738 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2739 nr_scanned = targets[lru] - nr[lru];
2740 nr[lru] = targets[lru] * (100 - percentage) / 100;
2741 nr[lru] -= min(nr[lru], nr_scanned);
2742
2743 lru += LRU_ACTIVE;
2744 nr_scanned = targets[lru] - nr[lru];
2745 nr[lru] = targets[lru] * (100 - percentage) / 100;
2746 nr[lru] -= min(nr[lru], nr_scanned);
2747
2748 scan_adjusted = true;
2749 }
2750 blk_finish_plug(&plug);
2751 sc->nr_reclaimed += nr_reclaimed;
2752
2753 /*
2754 * Even if we did not try to evict anon pages at all, we want to
2755 * rebalance the anon lru active/inactive ratio.
2756 */
2757 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2758 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2759 sc, LRU_ACTIVE_ANON);
2760}
2761
2762/* Use reclaim/compaction for costly allocs or under memory pressure */
2763static bool in_reclaim_compaction(struct scan_control *sc)
2764{
2765 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2766 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2767 sc->priority < DEF_PRIORITY - 2))
2768 return true;
2769
2770 return false;
2771}
2772
2773/*
2774 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2775 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2776 * true if more pages should be reclaimed such that when the page allocator
2777 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2778 * It will give up earlier than that if there is difficulty reclaiming pages.
2779 */
2780static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2781 unsigned long nr_reclaimed,
2782 struct scan_control *sc)
2783{
2784 unsigned long pages_for_compaction;
2785 unsigned long inactive_lru_pages;
2786 int z;
2787
2788 /* If not in reclaim/compaction mode, stop */
2789 if (!in_reclaim_compaction(sc))
2790 return false;
2791
2792 /*
2793 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2794 * number of pages that were scanned. This will return to the caller
2795 * with the risk reclaim/compaction and the resulting allocation attempt
2796 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2797 * allocations through requiring that the full LRU list has been scanned
2798 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2799 * scan, but that approximation was wrong, and there were corner cases
2800 * where always a non-zero amount of pages were scanned.
2801 */
2802 if (!nr_reclaimed)
2803 return false;
2804
2805 /* If compaction would go ahead or the allocation would succeed, stop */
2806 for (z = 0; z <= sc->reclaim_idx; z++) {
2807 struct zone *zone = &pgdat->node_zones[z];
2808 if (!managed_zone(zone))
2809 continue;
2810
2811 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2812 case COMPACT_SUCCESS:
2813 case COMPACT_CONTINUE:
2814 return false;
2815 default:
2816 /* check next zone */
2817 ;
2818 }
2819 }
2820
2821 /*
2822 * If we have not reclaimed enough pages for compaction and the
2823 * inactive lists are large enough, continue reclaiming
2824 */
2825 pages_for_compaction = compact_gap(sc->order);
2826 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2827 if (get_nr_swap_pages() > 0)
2828 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2829
2830 return inactive_lru_pages > pages_for_compaction;
2831}
2832
2833static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2834{
2835 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2836 struct mem_cgroup *memcg;
2837
2838 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2839 do {
2840 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2841 unsigned long reclaimed;
2842 unsigned long scanned;
2843
2844 /*
2845 * This loop can become CPU-bound when target memcgs
2846 * aren't eligible for reclaim - either because they
2847 * don't have any reclaimable pages, or because their
2848 * memory is explicitly protected. Avoid soft lockups.
2849 */
2850 cond_resched();
2851
2852 mem_cgroup_calculate_protection(target_memcg, memcg);
2853
2854 if (mem_cgroup_below_min(memcg)) {
2855 /*
2856 * Hard protection.
2857 * If there is no reclaimable memory, OOM.
2858 */
2859 continue;
2860 } else if (mem_cgroup_below_low(memcg)) {
2861 /*
2862 * Soft protection.
2863 * Respect the protection only as long as
2864 * there is an unprotected supply
2865 * of reclaimable memory from other cgroups.
2866 */
2867 if (!sc->memcg_low_reclaim) {
2868 sc->memcg_low_skipped = 1;
2869 continue;
2870 }
2871 memcg_memory_event(memcg, MEMCG_LOW);
2872 }
2873
2874 reclaimed = sc->nr_reclaimed;
2875 scanned = sc->nr_scanned;
2876
2877 shrink_lruvec(lruvec, sc);
2878
2879 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2880 sc->priority);
2881
2882 /* Record the group's reclaim efficiency */
2883 vmpressure(sc->gfp_mask, memcg, false,
2884 sc->nr_scanned - scanned,
2885 sc->nr_reclaimed - reclaimed);
2886
2887 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2888}
2889
2890static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2891{
2892 struct reclaim_state *reclaim_state = current->reclaim_state;
2893 unsigned long nr_reclaimed, nr_scanned;
2894 struct lruvec *target_lruvec;
2895 bool reclaimable = false;
2896 unsigned long file;
2897
2898 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2899
2900again:
2901 memset(&sc->nr, 0, sizeof(sc->nr));
2902
2903 nr_reclaimed = sc->nr_reclaimed;
2904 nr_scanned = sc->nr_scanned;
2905
2906 /*
2907 * Determine the scan balance between anon and file LRUs.
2908 */
2909 spin_lock_irq(&target_lruvec->lru_lock);
2910 sc->anon_cost = target_lruvec->anon_cost;
2911 sc->file_cost = target_lruvec->file_cost;
2912 spin_unlock_irq(&target_lruvec->lru_lock);
2913
2914 /*
2915 * Target desirable inactive:active list ratios for the anon
2916 * and file LRU lists.
2917 */
2918 if (!sc->force_deactivate) {
2919 unsigned long refaults;
2920
2921 refaults = lruvec_page_state(target_lruvec,
2922 WORKINGSET_ACTIVATE_ANON);
2923 if (refaults != target_lruvec->refaults[0] ||
2924 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2925 sc->may_deactivate |= DEACTIVATE_ANON;
2926 else
2927 sc->may_deactivate &= ~DEACTIVATE_ANON;
2928
2929 /*
2930 * When refaults are being observed, it means a new
2931 * workingset is being established. Deactivate to get
2932 * rid of any stale active pages quickly.
2933 */
2934 refaults = lruvec_page_state(target_lruvec,
2935 WORKINGSET_ACTIVATE_FILE);
2936 if (refaults != target_lruvec->refaults[1] ||
2937 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2938 sc->may_deactivate |= DEACTIVATE_FILE;
2939 else
2940 sc->may_deactivate &= ~DEACTIVATE_FILE;
2941 } else
2942 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2943
2944 /*
2945 * If we have plenty of inactive file pages that aren't
2946 * thrashing, try to reclaim those first before touching
2947 * anonymous pages.
2948 */
2949 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2950 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2951 sc->cache_trim_mode = 1;
2952 else
2953 sc->cache_trim_mode = 0;
2954
2955 /*
2956 * Prevent the reclaimer from falling into the cache trap: as
2957 * cache pages start out inactive, every cache fault will tip
2958 * the scan balance towards the file LRU. And as the file LRU
2959 * shrinks, so does the window for rotation from references.
2960 * This means we have a runaway feedback loop where a tiny
2961 * thrashing file LRU becomes infinitely more attractive than
2962 * anon pages. Try to detect this based on file LRU size.
2963 */
2964 if (!cgroup_reclaim(sc)) {
2965 unsigned long total_high_wmark = 0;
2966 unsigned long free, anon;
2967 int z;
2968
2969 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2970 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2971 node_page_state(pgdat, NR_INACTIVE_FILE);
2972
2973 for (z = 0; z < MAX_NR_ZONES; z++) {
2974 struct zone *zone = &pgdat->node_zones[z];
2975 if (!managed_zone(zone))
2976 continue;
2977
2978 total_high_wmark += high_wmark_pages(zone);
2979 }
2980
2981 /*
2982 * Consider anon: if that's low too, this isn't a
2983 * runaway file reclaim problem, but rather just
2984 * extreme pressure. Reclaim as per usual then.
2985 */
2986 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2987
2988 sc->file_is_tiny =
2989 file + free <= total_high_wmark &&
2990 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2991 anon >> sc->priority;
2992 }
2993
2994 shrink_node_memcgs(pgdat, sc);
2995
2996 if (reclaim_state) {
2997 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2998 reclaim_state->reclaimed_slab = 0;
2999 }
3000
3001 /* Record the subtree's reclaim efficiency */
3002 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3003 sc->nr_scanned - nr_scanned,
3004 sc->nr_reclaimed - nr_reclaimed);
3005
3006 if (sc->nr_reclaimed - nr_reclaimed)
3007 reclaimable = true;
3008
3009 if (current_is_kswapd()) {
3010 /*
3011 * If reclaim is isolating dirty pages under writeback,
3012 * it implies that the long-lived page allocation rate
3013 * is exceeding the page laundering rate. Either the
3014 * global limits are not being effective at throttling
3015 * processes due to the page distribution throughout
3016 * zones or there is heavy usage of a slow backing
3017 * device. The only option is to throttle from reclaim
3018 * context which is not ideal as there is no guarantee
3019 * the dirtying process is throttled in the same way
3020 * balance_dirty_pages() manages.
3021 *
3022 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3023 * count the number of pages under pages flagged for
3024 * immediate reclaim and stall if any are encountered
3025 * in the nr_immediate check below.
3026 */
3027 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3028 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3029
3030 /* Allow kswapd to start writing pages during reclaim.*/
3031 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3032 set_bit(PGDAT_DIRTY, &pgdat->flags);
3033
3034 /*
3035 * If kswapd scans pages marked for immediate
3036 * reclaim and under writeback (nr_immediate), it
3037 * implies that pages are cycling through the LRU
3038 * faster than they are written so also forcibly stall.
3039 */
3040 if (sc->nr.immediate)
3041 congestion_wait(BLK_RW_ASYNC, HZ/10);
3042 }
3043
3044 /*
3045 * Tag a node/memcg as congested if all the dirty pages
3046 * scanned were backed by a congested BDI and
3047 * wait_iff_congested will stall.
3048 *
3049 * Legacy memcg will stall in page writeback so avoid forcibly
3050 * stalling in wait_iff_congested().
3051 */
3052 if ((current_is_kswapd() ||
3053 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3054 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3055 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3056
3057 /*
3058 * Stall direct reclaim for IO completions if underlying BDIs
3059 * and node is congested. Allow kswapd to continue until it
3060 * starts encountering unqueued dirty pages or cycling through
3061 * the LRU too quickly.
3062 */
3063 if (!current_is_kswapd() && current_may_throttle() &&
3064 !sc->hibernation_mode &&
3065 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3066 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
3067
3068 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3069 sc))
3070 goto again;
3071
3072 /*
3073 * Kswapd gives up on balancing particular nodes after too
3074 * many failures to reclaim anything from them and goes to
3075 * sleep. On reclaim progress, reset the failure counter. A
3076 * successful direct reclaim run will revive a dormant kswapd.
3077 */
3078 if (reclaimable)
3079 pgdat->kswapd_failures = 0;
3080}
3081
3082/*
3083 * Returns true if compaction should go ahead for a costly-order request, or
3084 * the allocation would already succeed without compaction. Return false if we
3085 * should reclaim first.
3086 */
3087static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3088{
3089 unsigned long watermark;
3090 enum compact_result suitable;
3091
3092 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3093 if (suitable == COMPACT_SUCCESS)
3094 /* Allocation should succeed already. Don't reclaim. */
3095 return true;
3096 if (suitable == COMPACT_SKIPPED)
3097 /* Compaction cannot yet proceed. Do reclaim. */
3098 return false;
3099
3100 /*
3101 * Compaction is already possible, but it takes time to run and there
3102 * are potentially other callers using the pages just freed. So proceed
3103 * with reclaim to make a buffer of free pages available to give
3104 * compaction a reasonable chance of completing and allocating the page.
3105 * Note that we won't actually reclaim the whole buffer in one attempt
3106 * as the target watermark in should_continue_reclaim() is lower. But if
3107 * we are already above the high+gap watermark, don't reclaim at all.
3108 */
3109 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3110
3111 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3112}
3113
3114/*
3115 * This is the direct reclaim path, for page-allocating processes. We only
3116 * try to reclaim pages from zones which will satisfy the caller's allocation
3117 * request.
3118 *
3119 * If a zone is deemed to be full of pinned pages then just give it a light
3120 * scan then give up on it.
3121 */
3122static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3123{
3124 struct zoneref *z;
3125 struct zone *zone;
3126 unsigned long nr_soft_reclaimed;
3127 unsigned long nr_soft_scanned;
3128 gfp_t orig_mask;
3129 pg_data_t *last_pgdat = NULL;
3130
3131 /*
3132 * If the number of buffer_heads in the machine exceeds the maximum
3133 * allowed level, force direct reclaim to scan the highmem zone as
3134 * highmem pages could be pinning lowmem pages storing buffer_heads
3135 */
3136 orig_mask = sc->gfp_mask;
3137 if (buffer_heads_over_limit) {
3138 sc->gfp_mask |= __GFP_HIGHMEM;
3139 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3140 }
3141
3142 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3143 sc->reclaim_idx, sc->nodemask) {
3144 /*
3145 * Take care memory controller reclaiming has small influence
3146 * to global LRU.
3147 */
3148 if (!cgroup_reclaim(sc)) {
3149 if (!cpuset_zone_allowed(zone,
3150 GFP_KERNEL | __GFP_HARDWALL))
3151 continue;
3152
3153 /*
3154 * If we already have plenty of memory free for
3155 * compaction in this zone, don't free any more.
3156 * Even though compaction is invoked for any
3157 * non-zero order, only frequent costly order
3158 * reclamation is disruptive enough to become a
3159 * noticeable problem, like transparent huge
3160 * page allocations.
3161 */
3162 if (IS_ENABLED(CONFIG_COMPACTION) &&
3163 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3164 compaction_ready(zone, sc)) {
3165 sc->compaction_ready = true;
3166 continue;
3167 }
3168
3169 /*
3170 * Shrink each node in the zonelist once. If the
3171 * zonelist is ordered by zone (not the default) then a
3172 * node may be shrunk multiple times but in that case
3173 * the user prefers lower zones being preserved.
3174 */
3175 if (zone->zone_pgdat == last_pgdat)
3176 continue;
3177
3178 /*
3179 * This steals pages from memory cgroups over softlimit
3180 * and returns the number of reclaimed pages and
3181 * scanned pages. This works for global memory pressure
3182 * and balancing, not for a memcg's limit.
3183 */
3184 nr_soft_scanned = 0;
3185 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3186 sc->order, sc->gfp_mask,
3187 &nr_soft_scanned);
3188 sc->nr_reclaimed += nr_soft_reclaimed;
3189 sc->nr_scanned += nr_soft_scanned;
3190 /* need some check for avoid more shrink_zone() */
3191 }
3192
3193 /* See comment about same check for global reclaim above */
3194 if (zone->zone_pgdat == last_pgdat)
3195 continue;
3196 last_pgdat = zone->zone_pgdat;
3197 shrink_node(zone->zone_pgdat, sc);
3198 }
3199
3200 /*
3201 * Restore to original mask to avoid the impact on the caller if we
3202 * promoted it to __GFP_HIGHMEM.
3203 */
3204 sc->gfp_mask = orig_mask;
3205}
3206
3207static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3208{
3209 struct lruvec *target_lruvec;
3210 unsigned long refaults;
3211
3212 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3213 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3214 target_lruvec->refaults[0] = refaults;
3215 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3216 target_lruvec->refaults[1] = refaults;
3217}
3218
3219/*
3220 * This is the main entry point to direct page reclaim.
3221 *
3222 * If a full scan of the inactive list fails to free enough memory then we
3223 * are "out of memory" and something needs to be killed.
3224 *
3225 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3226 * high - the zone may be full of dirty or under-writeback pages, which this
3227 * caller can't do much about. We kick the writeback threads and take explicit
3228 * naps in the hope that some of these pages can be written. But if the
3229 * allocating task holds filesystem locks which prevent writeout this might not
3230 * work, and the allocation attempt will fail.
3231 *
3232 * returns: 0, if no pages reclaimed
3233 * else, the number of pages reclaimed
3234 */
3235static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3236 struct scan_control *sc)
3237{
3238 int initial_priority = sc->priority;
3239 pg_data_t *last_pgdat;
3240 struct zoneref *z;
3241 struct zone *zone;
3242retry:
3243 delayacct_freepages_start();
3244
3245 if (!cgroup_reclaim(sc))
3246 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3247
3248 do {
3249 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3250 sc->priority);
3251 sc->nr_scanned = 0;
3252 shrink_zones(zonelist, sc);
3253
3254 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3255 break;
3256
3257 if (sc->compaction_ready)
3258 break;
3259
3260 /*
3261 * If we're getting trouble reclaiming, start doing
3262 * writepage even in laptop mode.
3263 */
3264 if (sc->priority < DEF_PRIORITY - 2)
3265 sc->may_writepage = 1;
3266 } while (--sc->priority >= 0);
3267
3268 last_pgdat = NULL;
3269 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3270 sc->nodemask) {
3271 if (zone->zone_pgdat == last_pgdat)
3272 continue;
3273 last_pgdat = zone->zone_pgdat;
3274
3275 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3276
3277 if (cgroup_reclaim(sc)) {
3278 struct lruvec *lruvec;
3279
3280 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3281 zone->zone_pgdat);
3282 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3283 }
3284 }
3285
3286 delayacct_freepages_end();
3287
3288 if (sc->nr_reclaimed)
3289 return sc->nr_reclaimed;
3290
3291 /* Aborted reclaim to try compaction? don't OOM, then */
3292 if (sc->compaction_ready)
3293 return 1;
3294
3295 /*
3296 * We make inactive:active ratio decisions based on the node's
3297 * composition of memory, but a restrictive reclaim_idx or a
3298 * memory.low cgroup setting can exempt large amounts of
3299 * memory from reclaim. Neither of which are very common, so
3300 * instead of doing costly eligibility calculations of the
3301 * entire cgroup subtree up front, we assume the estimates are
3302 * good, and retry with forcible deactivation if that fails.
3303 */
3304 if (sc->skipped_deactivate) {
3305 sc->priority = initial_priority;
3306 sc->force_deactivate = 1;
3307 sc->skipped_deactivate = 0;
3308 goto retry;
3309 }
3310
3311 /* Untapped cgroup reserves? Don't OOM, retry. */
3312 if (sc->memcg_low_skipped) {
3313 sc->priority = initial_priority;
3314 sc->force_deactivate = 0;
3315 sc->memcg_low_reclaim = 1;
3316 sc->memcg_low_skipped = 0;
3317 goto retry;
3318 }
3319
3320 return 0;
3321}
3322
3323static bool allow_direct_reclaim(pg_data_t *pgdat)
3324{
3325 struct zone *zone;
3326 unsigned long pfmemalloc_reserve = 0;
3327 unsigned long free_pages = 0;
3328 int i;
3329 bool wmark_ok;
3330
3331 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3332 return true;
3333
3334 for (i = 0; i <= ZONE_NORMAL; i++) {
3335 zone = &pgdat->node_zones[i];
3336 if (!managed_zone(zone))
3337 continue;
3338
3339 if (!zone_reclaimable_pages(zone))
3340 continue;
3341
3342 pfmemalloc_reserve += min_wmark_pages(zone);
3343 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3344 }
3345
3346 /* If there are no reserves (unexpected config) then do not throttle */
3347 if (!pfmemalloc_reserve)
3348 return true;
3349
3350 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3351
3352 /* kswapd must be awake if processes are being throttled */
3353 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3354 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3355 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3356
3357 wake_up_interruptible(&pgdat->kswapd_wait);
3358 }
3359
3360 return wmark_ok;
3361}
3362
3363/*
3364 * Throttle direct reclaimers if backing storage is backed by the network
3365 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3366 * depleted. kswapd will continue to make progress and wake the processes
3367 * when the low watermark is reached.
3368 *
3369 * Returns true if a fatal signal was delivered during throttling. If this
3370 * happens, the page allocator should not consider triggering the OOM killer.
3371 */
3372static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3373 nodemask_t *nodemask)
3374{
3375 struct zoneref *z;
3376 struct zone *zone;
3377 pg_data_t *pgdat = NULL;
3378
3379 /*
3380 * Kernel threads should not be throttled as they may be indirectly
3381 * responsible for cleaning pages necessary for reclaim to make forward
3382 * progress. kjournald for example may enter direct reclaim while
3383 * committing a transaction where throttling it could forcing other
3384 * processes to block on log_wait_commit().
3385 */
3386 if (current->flags & PF_KTHREAD)
3387 goto out;
3388
3389 /*
3390 * If a fatal signal is pending, this process should not throttle.
3391 * It should return quickly so it can exit and free its memory
3392 */
3393 if (fatal_signal_pending(current))
3394 goto out;
3395
3396 /*
3397 * Check if the pfmemalloc reserves are ok by finding the first node
3398 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3399 * GFP_KERNEL will be required for allocating network buffers when
3400 * swapping over the network so ZONE_HIGHMEM is unusable.
3401 *
3402 * Throttling is based on the first usable node and throttled processes
3403 * wait on a queue until kswapd makes progress and wakes them. There
3404 * is an affinity then between processes waking up and where reclaim
3405 * progress has been made assuming the process wakes on the same node.
3406 * More importantly, processes running on remote nodes will not compete
3407 * for remote pfmemalloc reserves and processes on different nodes
3408 * should make reasonable progress.
3409 */
3410 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3411 gfp_zone(gfp_mask), nodemask) {
3412 if (zone_idx(zone) > ZONE_NORMAL)
3413 continue;
3414
3415 /* Throttle based on the first usable node */
3416 pgdat = zone->zone_pgdat;
3417 if (allow_direct_reclaim(pgdat))
3418 goto out;
3419 break;
3420 }
3421
3422 /* If no zone was usable by the allocation flags then do not throttle */
3423 if (!pgdat)
3424 goto out;
3425
3426 /* Account for the throttling */
3427 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3428
3429 /*
3430 * If the caller cannot enter the filesystem, it's possible that it
3431 * is due to the caller holding an FS lock or performing a journal
3432 * transaction in the case of a filesystem like ext[3|4]. In this case,
3433 * it is not safe to block on pfmemalloc_wait as kswapd could be
3434 * blocked waiting on the same lock. Instead, throttle for up to a
3435 * second before continuing.
3436 */
3437 if (!(gfp_mask & __GFP_FS)) {
3438 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3439 allow_direct_reclaim(pgdat), HZ);
3440
3441 goto check_pending;
3442 }
3443
3444 /* Throttle until kswapd wakes the process */
3445 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3446 allow_direct_reclaim(pgdat));
3447
3448check_pending:
3449 if (fatal_signal_pending(current))
3450 return true;
3451
3452out:
3453 return false;
3454}
3455
3456unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3457 gfp_t gfp_mask, nodemask_t *nodemask)
3458{
3459 unsigned long nr_reclaimed;
3460 struct scan_control sc = {
3461 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3462 .gfp_mask = current_gfp_context(gfp_mask),
3463 .reclaim_idx = gfp_zone(gfp_mask),
3464 .order = order,
3465 .nodemask = nodemask,
3466 .priority = DEF_PRIORITY,
3467 .may_writepage = !laptop_mode,
3468 .may_unmap = 1,
3469 .may_swap = 1,
3470 };
3471
3472 /*
3473 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3474 * Confirm they are large enough for max values.
3475 */
3476 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3477 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3478 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3479
3480 /*
3481 * Do not enter reclaim if fatal signal was delivered while throttled.
3482 * 1 is returned so that the page allocator does not OOM kill at this
3483 * point.
3484 */
3485 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3486 return 1;
3487
3488 set_task_reclaim_state(current, &sc.reclaim_state);
3489 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3490
3491 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3492
3493 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3494 set_task_reclaim_state(current, NULL);
3495
3496 return nr_reclaimed;
3497}
3498
3499#ifdef CONFIG_MEMCG
3500
3501/* Only used by soft limit reclaim. Do not reuse for anything else. */
3502unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3503 gfp_t gfp_mask, bool noswap,
3504 pg_data_t *pgdat,
3505 unsigned long *nr_scanned)
3506{
3507 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3508 struct scan_control sc = {
3509 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3510 .target_mem_cgroup = memcg,
3511 .may_writepage = !laptop_mode,
3512 .may_unmap = 1,
3513 .reclaim_idx = MAX_NR_ZONES - 1,
3514 .may_swap = !noswap,
3515 };
3516
3517 WARN_ON_ONCE(!current->reclaim_state);
3518
3519 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3520 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3521
3522 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3523 sc.gfp_mask);
3524
3525 /*
3526 * NOTE: Although we can get the priority field, using it
3527 * here is not a good idea, since it limits the pages we can scan.
3528 * if we don't reclaim here, the shrink_node from balance_pgdat
3529 * will pick up pages from other mem cgroup's as well. We hack
3530 * the priority and make it zero.
3531 */
3532 shrink_lruvec(lruvec, &sc);
3533
3534 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3535
3536 *nr_scanned = sc.nr_scanned;
3537
3538 return sc.nr_reclaimed;
3539}
3540
3541unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3542 unsigned long nr_pages,
3543 gfp_t gfp_mask,
3544 bool may_swap)
3545{
3546 unsigned long nr_reclaimed;
3547 unsigned int noreclaim_flag;
3548 struct scan_control sc = {
3549 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3550 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3551 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3552 .reclaim_idx = MAX_NR_ZONES - 1,
3553 .target_mem_cgroup = memcg,
3554 .priority = DEF_PRIORITY,
3555 .may_writepage = !laptop_mode,
3556 .may_unmap = 1,
3557 .may_swap = may_swap,
3558 };
3559 /*
3560 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3561 * equal pressure on all the nodes. This is based on the assumption that
3562 * the reclaim does not bail out early.
3563 */
3564 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3565
3566 set_task_reclaim_state(current, &sc.reclaim_state);
3567 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3568 noreclaim_flag = memalloc_noreclaim_save();
3569
3570 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3571
3572 memalloc_noreclaim_restore(noreclaim_flag);
3573 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3574 set_task_reclaim_state(current, NULL);
3575
3576 return nr_reclaimed;
3577}
3578#endif
3579
3580static void age_active_anon(struct pglist_data *pgdat,
3581 struct scan_control *sc)
3582{
3583 struct mem_cgroup *memcg;
3584 struct lruvec *lruvec;
3585
3586 if (!total_swap_pages)
3587 return;
3588
3589 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3590 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3591 return;
3592
3593 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3594 do {
3595 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3596 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3597 sc, LRU_ACTIVE_ANON);
3598 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3599 } while (memcg);
3600}
3601
3602static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3603{
3604 int i;
3605 struct zone *zone;
3606
3607 /*
3608 * Check for watermark boosts top-down as the higher zones
3609 * are more likely to be boosted. Both watermarks and boosts
3610 * should not be checked at the same time as reclaim would
3611 * start prematurely when there is no boosting and a lower
3612 * zone is balanced.
3613 */
3614 for (i = highest_zoneidx; i >= 0; i--) {
3615 zone = pgdat->node_zones + i;
3616 if (!managed_zone(zone))
3617 continue;
3618
3619 if (zone->watermark_boost)
3620 return true;
3621 }
3622
3623 return false;
3624}
3625
3626/*
3627 * Returns true if there is an eligible zone balanced for the request order
3628 * and highest_zoneidx
3629 */
3630static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3631{
3632 int i;
3633 unsigned long mark = -1;
3634 struct zone *zone;
3635
3636 /*
3637 * Check watermarks bottom-up as lower zones are more likely to
3638 * meet watermarks.
3639 */
3640 for (i = 0; i <= highest_zoneidx; i++) {
3641 zone = pgdat->node_zones + i;
3642
3643 if (!managed_zone(zone))
3644 continue;
3645
3646 mark = high_wmark_pages(zone);
3647 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3648 return true;
3649 }
3650
3651 /*
3652 * If a node has no populated zone within highest_zoneidx, it does not
3653 * need balancing by definition. This can happen if a zone-restricted
3654 * allocation tries to wake a remote kswapd.
3655 */
3656 if (mark == -1)
3657 return true;
3658
3659 return false;
3660}
3661
3662/* Clear pgdat state for congested, dirty or under writeback. */
3663static void clear_pgdat_congested(pg_data_t *pgdat)
3664{
3665 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3666
3667 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3668 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3669 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3670}
3671
3672/*
3673 * Prepare kswapd for sleeping. This verifies that there are no processes
3674 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3675 *
3676 * Returns true if kswapd is ready to sleep
3677 */
3678static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3679 int highest_zoneidx)
3680{
3681 /*
3682 * The throttled processes are normally woken up in balance_pgdat() as
3683 * soon as allow_direct_reclaim() is true. But there is a potential
3684 * race between when kswapd checks the watermarks and a process gets
3685 * throttled. There is also a potential race if processes get
3686 * throttled, kswapd wakes, a large process exits thereby balancing the
3687 * zones, which causes kswapd to exit balance_pgdat() before reaching
3688 * the wake up checks. If kswapd is going to sleep, no process should
3689 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3690 * the wake up is premature, processes will wake kswapd and get
3691 * throttled again. The difference from wake ups in balance_pgdat() is
3692 * that here we are under prepare_to_wait().
3693 */
3694 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3695 wake_up_all(&pgdat->pfmemalloc_wait);
3696
3697 /* Hopeless node, leave it to direct reclaim */
3698 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3699 return true;
3700
3701 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3702 clear_pgdat_congested(pgdat);
3703 return true;
3704 }
3705
3706 return false;
3707}
3708
3709/*
3710 * kswapd shrinks a node of pages that are at or below the highest usable
3711 * zone that is currently unbalanced.
3712 *
3713 * Returns true if kswapd scanned at least the requested number of pages to
3714 * reclaim or if the lack of progress was due to pages under writeback.
3715 * This is used to determine if the scanning priority needs to be raised.
3716 */
3717static bool kswapd_shrink_node(pg_data_t *pgdat,
3718 struct scan_control *sc)
3719{
3720 struct zone *zone;
3721 int z;
3722
3723 /* Reclaim a number of pages proportional to the number of zones */
3724 sc->nr_to_reclaim = 0;
3725 for (z = 0; z <= sc->reclaim_idx; z++) {
3726 zone = pgdat->node_zones + z;
3727 if (!managed_zone(zone))
3728 continue;
3729
3730 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3731 }
3732
3733 /*
3734 * Historically care was taken to put equal pressure on all zones but
3735 * now pressure is applied based on node LRU order.
3736 */
3737 shrink_node(pgdat, sc);
3738
3739 /*
3740 * Fragmentation may mean that the system cannot be rebalanced for
3741 * high-order allocations. If twice the allocation size has been
3742 * reclaimed then recheck watermarks only at order-0 to prevent
3743 * excessive reclaim. Assume that a process requested a high-order
3744 * can direct reclaim/compact.
3745 */
3746 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3747 sc->order = 0;
3748
3749 return sc->nr_scanned >= sc->nr_to_reclaim;
3750}
3751
3752/* Page allocator PCP high watermark is lowered if reclaim is active. */
3753static inline void
3754update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
3755{
3756 int i;
3757 struct zone *zone;
3758
3759 for (i = 0; i <= highest_zoneidx; i++) {
3760 zone = pgdat->node_zones + i;
3761
3762 if (!managed_zone(zone))
3763 continue;
3764
3765 if (active)
3766 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3767 else
3768 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3769 }
3770}
3771
3772static inline void
3773set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3774{
3775 update_reclaim_active(pgdat, highest_zoneidx, true);
3776}
3777
3778static inline void
3779clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3780{
3781 update_reclaim_active(pgdat, highest_zoneidx, false);
3782}
3783
3784/*
3785 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3786 * that are eligible for use by the caller until at least one zone is
3787 * balanced.
3788 *
3789 * Returns the order kswapd finished reclaiming at.
3790 *
3791 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3792 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3793 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3794 * or lower is eligible for reclaim until at least one usable zone is
3795 * balanced.
3796 */
3797static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3798{
3799 int i;
3800 unsigned long nr_soft_reclaimed;
3801 unsigned long nr_soft_scanned;
3802 unsigned long pflags;
3803 unsigned long nr_boost_reclaim;
3804 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3805 bool boosted;
3806 struct zone *zone;
3807 struct scan_control sc = {
3808 .gfp_mask = GFP_KERNEL,
3809 .order = order,
3810 .may_unmap = 1,
3811 };
3812
3813 set_task_reclaim_state(current, &sc.reclaim_state);
3814 psi_memstall_enter(&pflags);
3815 __fs_reclaim_acquire();
3816
3817 count_vm_event(PAGEOUTRUN);
3818
3819 /*
3820 * Account for the reclaim boost. Note that the zone boost is left in
3821 * place so that parallel allocations that are near the watermark will
3822 * stall or direct reclaim until kswapd is finished.
3823 */
3824 nr_boost_reclaim = 0;
3825 for (i = 0; i <= highest_zoneidx; i++) {
3826 zone = pgdat->node_zones + i;
3827 if (!managed_zone(zone))
3828 continue;
3829
3830 nr_boost_reclaim += zone->watermark_boost;
3831 zone_boosts[i] = zone->watermark_boost;
3832 }
3833 boosted = nr_boost_reclaim;
3834
3835restart:
3836 set_reclaim_active(pgdat, highest_zoneidx);
3837 sc.priority = DEF_PRIORITY;
3838 do {
3839 unsigned long nr_reclaimed = sc.nr_reclaimed;
3840 bool raise_priority = true;
3841 bool balanced;
3842 bool ret;
3843
3844 sc.reclaim_idx = highest_zoneidx;
3845
3846 /*
3847 * If the number of buffer_heads exceeds the maximum allowed
3848 * then consider reclaiming from all zones. This has a dual
3849 * purpose -- on 64-bit systems it is expected that
3850 * buffer_heads are stripped during active rotation. On 32-bit
3851 * systems, highmem pages can pin lowmem memory and shrinking
3852 * buffers can relieve lowmem pressure. Reclaim may still not
3853 * go ahead if all eligible zones for the original allocation
3854 * request are balanced to avoid excessive reclaim from kswapd.
3855 */
3856 if (buffer_heads_over_limit) {
3857 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3858 zone = pgdat->node_zones + i;
3859 if (!managed_zone(zone))
3860 continue;
3861
3862 sc.reclaim_idx = i;
3863 break;
3864 }
3865 }
3866
3867 /*
3868 * If the pgdat is imbalanced then ignore boosting and preserve
3869 * the watermarks for a later time and restart. Note that the
3870 * zone watermarks will be still reset at the end of balancing
3871 * on the grounds that the normal reclaim should be enough to
3872 * re-evaluate if boosting is required when kswapd next wakes.
3873 */
3874 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3875 if (!balanced && nr_boost_reclaim) {
3876 nr_boost_reclaim = 0;
3877 goto restart;
3878 }
3879
3880 /*
3881 * If boosting is not active then only reclaim if there are no
3882 * eligible zones. Note that sc.reclaim_idx is not used as
3883 * buffer_heads_over_limit may have adjusted it.
3884 */
3885 if (!nr_boost_reclaim && balanced)
3886 goto out;
3887
3888 /* Limit the priority of boosting to avoid reclaim writeback */
3889 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3890 raise_priority = false;
3891
3892 /*
3893 * Do not writeback or swap pages for boosted reclaim. The
3894 * intent is to relieve pressure not issue sub-optimal IO
3895 * from reclaim context. If no pages are reclaimed, the
3896 * reclaim will be aborted.
3897 */
3898 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3899 sc.may_swap = !nr_boost_reclaim;
3900
3901 /*
3902 * Do some background aging of the anon list, to give
3903 * pages a chance to be referenced before reclaiming. All
3904 * pages are rotated regardless of classzone as this is
3905 * about consistent aging.
3906 */
3907 age_active_anon(pgdat, &sc);
3908
3909 /*
3910 * If we're getting trouble reclaiming, start doing writepage
3911 * even in laptop mode.
3912 */
3913 if (sc.priority < DEF_PRIORITY - 2)
3914 sc.may_writepage = 1;
3915
3916 /* Call soft limit reclaim before calling shrink_node. */
3917 sc.nr_scanned = 0;
3918 nr_soft_scanned = 0;
3919 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3920 sc.gfp_mask, &nr_soft_scanned);
3921 sc.nr_reclaimed += nr_soft_reclaimed;
3922
3923 /*
3924 * There should be no need to raise the scanning priority if
3925 * enough pages are already being scanned that that high
3926 * watermark would be met at 100% efficiency.
3927 */
3928 if (kswapd_shrink_node(pgdat, &sc))
3929 raise_priority = false;
3930
3931 /*
3932 * If the low watermark is met there is no need for processes
3933 * to be throttled on pfmemalloc_wait as they should not be
3934 * able to safely make forward progress. Wake them
3935 */
3936 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3937 allow_direct_reclaim(pgdat))
3938 wake_up_all(&pgdat->pfmemalloc_wait);
3939
3940 /* Check if kswapd should be suspending */
3941 __fs_reclaim_release();
3942 ret = try_to_freeze();
3943 __fs_reclaim_acquire();
3944 if (ret || kthread_should_stop())
3945 break;
3946
3947 /*
3948 * Raise priority if scanning rate is too low or there was no
3949 * progress in reclaiming pages
3950 */
3951 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3952 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3953
3954 /*
3955 * If reclaim made no progress for a boost, stop reclaim as
3956 * IO cannot be queued and it could be an infinite loop in
3957 * extreme circumstances.
3958 */
3959 if (nr_boost_reclaim && !nr_reclaimed)
3960 break;
3961
3962 if (raise_priority || !nr_reclaimed)
3963 sc.priority--;
3964 } while (sc.priority >= 1);
3965
3966 if (!sc.nr_reclaimed)
3967 pgdat->kswapd_failures++;
3968
3969out:
3970 clear_reclaim_active(pgdat, highest_zoneidx);
3971
3972 /* If reclaim was boosted, account for the reclaim done in this pass */
3973 if (boosted) {
3974 unsigned long flags;
3975
3976 for (i = 0; i <= highest_zoneidx; i++) {
3977 if (!zone_boosts[i])
3978 continue;
3979
3980 /* Increments are under the zone lock */
3981 zone = pgdat->node_zones + i;
3982 spin_lock_irqsave(&zone->lock, flags);
3983 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3984 spin_unlock_irqrestore(&zone->lock, flags);
3985 }
3986
3987 /*
3988 * As there is now likely space, wakeup kcompact to defragment
3989 * pageblocks.
3990 */
3991 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3992 }
3993
3994 snapshot_refaults(NULL, pgdat);
3995 __fs_reclaim_release();
3996 psi_memstall_leave(&pflags);
3997 set_task_reclaim_state(current, NULL);
3998
3999 /*
4000 * Return the order kswapd stopped reclaiming at as
4001 * prepare_kswapd_sleep() takes it into account. If another caller
4002 * entered the allocator slow path while kswapd was awake, order will
4003 * remain at the higher level.
4004 */
4005 return sc.order;
4006}
4007
4008/*
4009 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4010 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4011 * not a valid index then either kswapd runs for first time or kswapd couldn't
4012 * sleep after previous reclaim attempt (node is still unbalanced). In that
4013 * case return the zone index of the previous kswapd reclaim cycle.
4014 */
4015static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4016 enum zone_type prev_highest_zoneidx)
4017{
4018 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4019
4020 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4021}
4022
4023static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4024 unsigned int highest_zoneidx)
4025{
4026 long remaining = 0;
4027 DEFINE_WAIT(wait);
4028
4029 if (freezing(current) || kthread_should_stop())
4030 return;
4031
4032 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4033
4034 /*
4035 * Try to sleep for a short interval. Note that kcompactd will only be
4036 * woken if it is possible to sleep for a short interval. This is
4037 * deliberate on the assumption that if reclaim cannot keep an
4038 * eligible zone balanced that it's also unlikely that compaction will
4039 * succeed.
4040 */
4041 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4042 /*
4043 * Compaction records what page blocks it recently failed to
4044 * isolate pages from and skips them in the future scanning.
4045 * When kswapd is going to sleep, it is reasonable to assume
4046 * that pages and compaction may succeed so reset the cache.
4047 */
4048 reset_isolation_suitable(pgdat);
4049
4050 /*
4051 * We have freed the memory, now we should compact it to make
4052 * allocation of the requested order possible.
4053 */
4054 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4055
4056 remaining = schedule_timeout(HZ/10);
4057
4058 /*
4059 * If woken prematurely then reset kswapd_highest_zoneidx and
4060 * order. The values will either be from a wakeup request or
4061 * the previous request that slept prematurely.
4062 */
4063 if (remaining) {
4064 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4065 kswapd_highest_zoneidx(pgdat,
4066 highest_zoneidx));
4067
4068 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4069 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4070 }
4071
4072 finish_wait(&pgdat->kswapd_wait, &wait);
4073 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4074 }
4075
4076 /*
4077 * After a short sleep, check if it was a premature sleep. If not, then
4078 * go fully to sleep until explicitly woken up.
4079 */
4080 if (!remaining &&
4081 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4082 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4083
4084 /*
4085 * vmstat counters are not perfectly accurate and the estimated
4086 * value for counters such as NR_FREE_PAGES can deviate from the
4087 * true value by nr_online_cpus * threshold. To avoid the zone
4088 * watermarks being breached while under pressure, we reduce the
4089 * per-cpu vmstat threshold while kswapd is awake and restore
4090 * them before going back to sleep.
4091 */
4092 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4093
4094 if (!kthread_should_stop())
4095 schedule();
4096
4097 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4098 } else {
4099 if (remaining)
4100 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4101 else
4102 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4103 }
4104 finish_wait(&pgdat->kswapd_wait, &wait);
4105}
4106
4107/*
4108 * The background pageout daemon, started as a kernel thread
4109 * from the init process.
4110 *
4111 * This basically trickles out pages so that we have _some_
4112 * free memory available even if there is no other activity
4113 * that frees anything up. This is needed for things like routing
4114 * etc, where we otherwise might have all activity going on in
4115 * asynchronous contexts that cannot page things out.
4116 *
4117 * If there are applications that are active memory-allocators
4118 * (most normal use), this basically shouldn't matter.
4119 */
4120static int kswapd(void *p)
4121{
4122 unsigned int alloc_order, reclaim_order;
4123 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4124 pg_data_t *pgdat = (pg_data_t *)p;
4125 struct task_struct *tsk = current;
4126 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4127
4128 if (!cpumask_empty(cpumask))
4129 set_cpus_allowed_ptr(tsk, cpumask);
4130
4131 /*
4132 * Tell the memory management that we're a "memory allocator",
4133 * and that if we need more memory we should get access to it
4134 * regardless (see "__alloc_pages()"). "kswapd" should
4135 * never get caught in the normal page freeing logic.
4136 *
4137 * (Kswapd normally doesn't need memory anyway, but sometimes
4138 * you need a small amount of memory in order to be able to
4139 * page out something else, and this flag essentially protects
4140 * us from recursively trying to free more memory as we're
4141 * trying to free the first piece of memory in the first place).
4142 */
4143 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4144 set_freezable();
4145
4146 WRITE_ONCE(pgdat->kswapd_order, 0);
4147 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4148 for ( ; ; ) {
4149 bool ret;
4150
4151 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4152 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4153 highest_zoneidx);
4154
4155kswapd_try_sleep:
4156 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4157 highest_zoneidx);
4158
4159 /* Read the new order and highest_zoneidx */
4160 alloc_order = READ_ONCE(pgdat->kswapd_order);
4161 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4162 highest_zoneidx);
4163 WRITE_ONCE(pgdat->kswapd_order, 0);
4164 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4165
4166 ret = try_to_freeze();
4167 if (kthread_should_stop())
4168 break;
4169
4170 /*
4171 * We can speed up thawing tasks if we don't call balance_pgdat
4172 * after returning from the refrigerator
4173 */
4174 if (ret)
4175 continue;
4176
4177 /*
4178 * Reclaim begins at the requested order but if a high-order
4179 * reclaim fails then kswapd falls back to reclaiming for
4180 * order-0. If that happens, kswapd will consider sleeping
4181 * for the order it finished reclaiming at (reclaim_order)
4182 * but kcompactd is woken to compact for the original
4183 * request (alloc_order).
4184 */
4185 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4186 alloc_order);
4187 reclaim_order = balance_pgdat(pgdat, alloc_order,
4188 highest_zoneidx);
4189 if (reclaim_order < alloc_order)
4190 goto kswapd_try_sleep;
4191 }
4192
4193 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4194
4195 return 0;
4196}
4197
4198/*
4199 * A zone is low on free memory or too fragmented for high-order memory. If
4200 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4201 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4202 * has failed or is not needed, still wake up kcompactd if only compaction is
4203 * needed.
4204 */
4205void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4206 enum zone_type highest_zoneidx)
4207{
4208 pg_data_t *pgdat;
4209 enum zone_type curr_idx;
4210
4211 if (!managed_zone(zone))
4212 return;
4213
4214 if (!cpuset_zone_allowed(zone, gfp_flags))
4215 return;
4216
4217 pgdat = zone->zone_pgdat;
4218 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4219
4220 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4221 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4222
4223 if (READ_ONCE(pgdat->kswapd_order) < order)
4224 WRITE_ONCE(pgdat->kswapd_order, order);
4225
4226 if (!waitqueue_active(&pgdat->kswapd_wait))
4227 return;
4228
4229 /* Hopeless node, leave it to direct reclaim if possible */
4230 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4231 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4232 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4233 /*
4234 * There may be plenty of free memory available, but it's too
4235 * fragmented for high-order allocations. Wake up kcompactd
4236 * and rely on compaction_suitable() to determine if it's
4237 * needed. If it fails, it will defer subsequent attempts to
4238 * ratelimit its work.
4239 */
4240 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4241 wakeup_kcompactd(pgdat, order, highest_zoneidx);
4242 return;
4243 }
4244
4245 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4246 gfp_flags);
4247 wake_up_interruptible(&pgdat->kswapd_wait);
4248}
4249
4250#ifdef CONFIG_HIBERNATION
4251/*
4252 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4253 * freed pages.
4254 *
4255 * Rather than trying to age LRUs the aim is to preserve the overall
4256 * LRU order by reclaiming preferentially
4257 * inactive > active > active referenced > active mapped
4258 */
4259unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4260{
4261 struct scan_control sc = {
4262 .nr_to_reclaim = nr_to_reclaim,
4263 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4264 .reclaim_idx = MAX_NR_ZONES - 1,
4265 .priority = DEF_PRIORITY,
4266 .may_writepage = 1,
4267 .may_unmap = 1,
4268 .may_swap = 1,
4269 .hibernation_mode = 1,
4270 };
4271 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4272 unsigned long nr_reclaimed;
4273 unsigned int noreclaim_flag;
4274
4275 fs_reclaim_acquire(sc.gfp_mask);
4276 noreclaim_flag = memalloc_noreclaim_save();
4277 set_task_reclaim_state(current, &sc.reclaim_state);
4278
4279 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4280
4281 set_task_reclaim_state(current, NULL);
4282 memalloc_noreclaim_restore(noreclaim_flag);
4283 fs_reclaim_release(sc.gfp_mask);
4284
4285 return nr_reclaimed;
4286}
4287#endif /* CONFIG_HIBERNATION */
4288
4289/*
4290 * This kswapd start function will be called by init and node-hot-add.
4291 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4292 */
4293int kswapd_run(int nid)
4294{
4295 pg_data_t *pgdat = NODE_DATA(nid);
4296 int ret = 0;
4297
4298 if (pgdat->kswapd)
4299 return 0;
4300
4301 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4302 if (IS_ERR(pgdat->kswapd)) {
4303 /* failure at boot is fatal */
4304 BUG_ON(system_state < SYSTEM_RUNNING);
4305 pr_err("Failed to start kswapd on node %d\n", nid);
4306 ret = PTR_ERR(pgdat->kswapd);
4307 pgdat->kswapd = NULL;
4308 }
4309 return ret;
4310}
4311
4312/*
4313 * Called by memory hotplug when all memory in a node is offlined. Caller must
4314 * hold mem_hotplug_begin/end().
4315 */
4316void kswapd_stop(int nid)
4317{
4318 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4319
4320 if (kswapd) {
4321 kthread_stop(kswapd);
4322 NODE_DATA(nid)->kswapd = NULL;
4323 }
4324}
4325
4326static int __init kswapd_init(void)
4327{
4328 int nid;
4329
4330 swap_setup();
4331 for_each_node_state(nid, N_MEMORY)
4332 kswapd_run(nid);
4333 return 0;
4334}
4335
4336module_init(kswapd_init)
4337
4338#ifdef CONFIG_NUMA
4339/*
4340 * Node reclaim mode
4341 *
4342 * If non-zero call node_reclaim when the number of free pages falls below
4343 * the watermarks.
4344 */
4345int node_reclaim_mode __read_mostly;
4346
4347/*
4348 * Priority for NODE_RECLAIM. This determines the fraction of pages
4349 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4350 * a zone.
4351 */
4352#define NODE_RECLAIM_PRIORITY 4
4353
4354/*
4355 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4356 * occur.
4357 */
4358int sysctl_min_unmapped_ratio = 1;
4359
4360/*
4361 * If the number of slab pages in a zone grows beyond this percentage then
4362 * slab reclaim needs to occur.
4363 */
4364int sysctl_min_slab_ratio = 5;
4365
4366static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4367{
4368 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4369 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4370 node_page_state(pgdat, NR_ACTIVE_FILE);
4371
4372 /*
4373 * It's possible for there to be more file mapped pages than
4374 * accounted for by the pages on the file LRU lists because
4375 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4376 */
4377 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4378}
4379
4380/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4381static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4382{
4383 unsigned long nr_pagecache_reclaimable;
4384 unsigned long delta = 0;
4385
4386 /*
4387 * If RECLAIM_UNMAP is set, then all file pages are considered
4388 * potentially reclaimable. Otherwise, we have to worry about
4389 * pages like swapcache and node_unmapped_file_pages() provides
4390 * a better estimate
4391 */
4392 if (node_reclaim_mode & RECLAIM_UNMAP)
4393 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4394 else
4395 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4396
4397 /* If we can't clean pages, remove dirty pages from consideration */
4398 if (!(node_reclaim_mode & RECLAIM_WRITE))
4399 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4400
4401 /* Watch for any possible underflows due to delta */
4402 if (unlikely(delta > nr_pagecache_reclaimable))
4403 delta = nr_pagecache_reclaimable;
4404
4405 return nr_pagecache_reclaimable - delta;
4406}
4407
4408/*
4409 * Try to free up some pages from this node through reclaim.
4410 */
4411static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4412{
4413 /* Minimum pages needed in order to stay on node */
4414 const unsigned long nr_pages = 1 << order;
4415 struct task_struct *p = current;
4416 unsigned int noreclaim_flag;
4417 struct scan_control sc = {
4418 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4419 .gfp_mask = current_gfp_context(gfp_mask),
4420 .order = order,
4421 .priority = NODE_RECLAIM_PRIORITY,
4422 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4423 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4424 .may_swap = 1,
4425 .reclaim_idx = gfp_zone(gfp_mask),
4426 };
4427 unsigned long pflags;
4428
4429 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4430 sc.gfp_mask);
4431
4432 cond_resched();
4433 psi_memstall_enter(&pflags);
4434 fs_reclaim_acquire(sc.gfp_mask);
4435 /*
4436 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4437 * and we also need to be able to write out pages for RECLAIM_WRITE
4438 * and RECLAIM_UNMAP.
4439 */
4440 noreclaim_flag = memalloc_noreclaim_save();
4441 p->flags |= PF_SWAPWRITE;
4442 set_task_reclaim_state(p, &sc.reclaim_state);
4443
4444 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4445 /*
4446 * Free memory by calling shrink node with increasing
4447 * priorities until we have enough memory freed.
4448 */
4449 do {
4450 shrink_node(pgdat, &sc);
4451 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4452 }
4453
4454 set_task_reclaim_state(p, NULL);
4455 current->flags &= ~PF_SWAPWRITE;
4456 memalloc_noreclaim_restore(noreclaim_flag);
4457 fs_reclaim_release(sc.gfp_mask);
4458 psi_memstall_leave(&pflags);
4459
4460 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4461
4462 return sc.nr_reclaimed >= nr_pages;
4463}
4464
4465int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4466{
4467 int ret;
4468
4469 /*
4470 * Node reclaim reclaims unmapped file backed pages and
4471 * slab pages if we are over the defined limits.
4472 *
4473 * A small portion of unmapped file backed pages is needed for
4474 * file I/O otherwise pages read by file I/O will be immediately
4475 * thrown out if the node is overallocated. So we do not reclaim
4476 * if less than a specified percentage of the node is used by
4477 * unmapped file backed pages.
4478 */
4479 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4480 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4481 pgdat->min_slab_pages)
4482 return NODE_RECLAIM_FULL;
4483
4484 /*
4485 * Do not scan if the allocation should not be delayed.
4486 */
4487 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4488 return NODE_RECLAIM_NOSCAN;
4489
4490 /*
4491 * Only run node reclaim on the local node or on nodes that do not
4492 * have associated processors. This will favor the local processor
4493 * over remote processors and spread off node memory allocations
4494 * as wide as possible.
4495 */
4496 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4497 return NODE_RECLAIM_NOSCAN;
4498
4499 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4500 return NODE_RECLAIM_NOSCAN;
4501
4502 ret = __node_reclaim(pgdat, gfp_mask, order);
4503 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4504
4505 if (!ret)
4506 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4507
4508 return ret;
4509}
4510#endif
4511
4512/**
4513 * check_move_unevictable_pages - check pages for evictability and move to
4514 * appropriate zone lru list
4515 * @pvec: pagevec with lru pages to check
4516 *
4517 * Checks pages for evictability, if an evictable page is in the unevictable
4518 * lru list, moves it to the appropriate evictable lru list. This function
4519 * should be only used for lru pages.
4520 */
4521void check_move_unevictable_pages(struct pagevec *pvec)
4522{
4523 struct lruvec *lruvec = NULL;
4524 int pgscanned = 0;
4525 int pgrescued = 0;
4526 int i;
4527
4528 for (i = 0; i < pvec->nr; i++) {
4529 struct page *page = pvec->pages[i];
4530 int nr_pages;
4531
4532 if (PageTransTail(page))
4533 continue;
4534
4535 nr_pages = thp_nr_pages(page);
4536 pgscanned += nr_pages;
4537
4538 /* block memcg migration during page moving between lru */
4539 if (!TestClearPageLRU(page))
4540 continue;
4541
4542 lruvec = relock_page_lruvec_irq(page, lruvec);
4543 if (page_evictable(page) && PageUnevictable(page)) {
4544 del_page_from_lru_list(page, lruvec);
4545 ClearPageUnevictable(page);
4546 add_page_to_lru_list(page, lruvec);
4547 pgrescued += nr_pages;
4548 }
4549 SetPageLRU(page);
4550 }
4551
4552 if (lruvec) {
4553 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4554 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4555 unlock_page_lruvec_irq(lruvec);
4556 } else if (pgscanned) {
4557 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4558 }
4559}
4560EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
1/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16#include <linux/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 try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
37#include <linux/compaction.h>
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
40#include <linux/delay.h>
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43#include <linux/memcontrol.h>
44#include <linux/delayacct.h>
45#include <linux/sysctl.h>
46#include <linux/oom.h>
47#include <linux/prefetch.h>
48#include <linux/printk.h>
49#include <linux/dax.h>
50
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
55#include <linux/balloon_compaction.h>
56
57#include "internal.h"
58
59#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
62struct scan_control {
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
66 /* This context's GFP mask */
67 gfp_t gfp_mask;
68
69 /* Allocation order */
70 int order;
71
72 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
77
78 /*
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
83
84 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
89
90 unsigned int may_writepage:1;
91
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
94
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
97
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
100
101 unsigned int hibernation_mode:1;
102
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
105
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
108
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
111};
112
113#ifdef ARCH_HAS_PREFETCH
114#define prefetch_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127#ifdef ARCH_HAS_PREFETCHW
128#define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
132 \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
135 } \
136 } while (0)
137#else
138#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139#endif
140
141/*
142 * From 0 .. 100. Higher means more swappy.
143 */
144int vm_swappiness = 60;
145/*
146 * The total number of pages which are beyond the high watermark within all
147 * zones.
148 */
149unsigned long vm_total_pages;
150
151static LIST_HEAD(shrinker_list);
152static DECLARE_RWSEM(shrinker_rwsem);
153
154#ifdef CONFIG_MEMCG
155static bool global_reclaim(struct scan_control *sc)
156{
157 return !sc->target_mem_cgroup;
158}
159
160/**
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
163 *
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
169 *
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
172 */
173static bool sane_reclaim(struct scan_control *sc)
174{
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
176
177 if (!memcg)
178 return true;
179#ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
181 return true;
182#endif
183 return false;
184}
185#else
186static bool global_reclaim(struct scan_control *sc)
187{
188 return true;
189}
190
191static bool sane_reclaim(struct scan_control *sc)
192{
193 return true;
194}
195#endif
196
197/*
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
201 */
202unsigned long zone_reclaimable_pages(struct zone *zone)
203{
204 unsigned long nr;
205
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
211
212 return nr;
213}
214
215unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
216{
217 unsigned long nr;
218
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
222
223 if (get_nr_swap_pages() > 0)
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
227
228 return nr;
229}
230
231bool pgdat_reclaimable(struct pglist_data *pgdat)
232{
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
235}
236
237/**
238 * lruvec_lru_size - Returns the number of pages on the given LRU list.
239 * @lruvec: lru vector
240 * @lru: lru to use
241 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
242 */
243unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
244{
245 unsigned long lru_size;
246 int zid;
247
248 if (!mem_cgroup_disabled())
249 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
250 else
251 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
252
253 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
254 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
255 unsigned long size;
256
257 if (!managed_zone(zone))
258 continue;
259
260 if (!mem_cgroup_disabled())
261 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
262 else
263 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
264 NR_ZONE_LRU_BASE + lru);
265 lru_size -= min(size, lru_size);
266 }
267
268 return lru_size;
269
270}
271
272/*
273 * Add a shrinker callback to be called from the vm.
274 */
275int register_shrinker(struct shrinker *shrinker)
276{
277 size_t size = sizeof(*shrinker->nr_deferred);
278
279 if (shrinker->flags & SHRINKER_NUMA_AWARE)
280 size *= nr_node_ids;
281
282 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
283 if (!shrinker->nr_deferred)
284 return -ENOMEM;
285
286 down_write(&shrinker_rwsem);
287 list_add_tail(&shrinker->list, &shrinker_list);
288 up_write(&shrinker_rwsem);
289 return 0;
290}
291EXPORT_SYMBOL(register_shrinker);
292
293/*
294 * Remove one
295 */
296void unregister_shrinker(struct shrinker *shrinker)
297{
298 down_write(&shrinker_rwsem);
299 list_del(&shrinker->list);
300 up_write(&shrinker_rwsem);
301 kfree(shrinker->nr_deferred);
302}
303EXPORT_SYMBOL(unregister_shrinker);
304
305#define SHRINK_BATCH 128
306
307static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
308 struct shrinker *shrinker,
309 unsigned long nr_scanned,
310 unsigned long nr_eligible)
311{
312 unsigned long freed = 0;
313 unsigned long long delta;
314 long total_scan;
315 long freeable;
316 long nr;
317 long new_nr;
318 int nid = shrinkctl->nid;
319 long batch_size = shrinker->batch ? shrinker->batch
320 : SHRINK_BATCH;
321 long scanned = 0, next_deferred;
322
323 freeable = shrinker->count_objects(shrinker, shrinkctl);
324 if (freeable == 0)
325 return 0;
326
327 /*
328 * copy the current shrinker scan count into a local variable
329 * and zero it so that other concurrent shrinker invocations
330 * don't also do this scanning work.
331 */
332 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
333
334 total_scan = nr;
335 delta = (4 * nr_scanned) / shrinker->seeks;
336 delta *= freeable;
337 do_div(delta, nr_eligible + 1);
338 total_scan += delta;
339 if (total_scan < 0) {
340 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
341 shrinker->scan_objects, total_scan);
342 total_scan = freeable;
343 next_deferred = nr;
344 } else
345 next_deferred = total_scan;
346
347 /*
348 * We need to avoid excessive windup on filesystem shrinkers
349 * due to large numbers of GFP_NOFS allocations causing the
350 * shrinkers to return -1 all the time. This results in a large
351 * nr being built up so when a shrink that can do some work
352 * comes along it empties the entire cache due to nr >>>
353 * freeable. This is bad for sustaining a working set in
354 * memory.
355 *
356 * Hence only allow the shrinker to scan the entire cache when
357 * a large delta change is calculated directly.
358 */
359 if (delta < freeable / 4)
360 total_scan = min(total_scan, freeable / 2);
361
362 /*
363 * Avoid risking looping forever due to too large nr value:
364 * never try to free more than twice the estimate number of
365 * freeable entries.
366 */
367 if (total_scan > freeable * 2)
368 total_scan = freeable * 2;
369
370 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
371 nr_scanned, nr_eligible,
372 freeable, delta, total_scan);
373
374 /*
375 * Normally, we should not scan less than batch_size objects in one
376 * pass to avoid too frequent shrinker calls, but if the slab has less
377 * than batch_size objects in total and we are really tight on memory,
378 * we will try to reclaim all available objects, otherwise we can end
379 * up failing allocations although there are plenty of reclaimable
380 * objects spread over several slabs with usage less than the
381 * batch_size.
382 *
383 * We detect the "tight on memory" situations by looking at the total
384 * number of objects we want to scan (total_scan). If it is greater
385 * than the total number of objects on slab (freeable), we must be
386 * scanning at high prio and therefore should try to reclaim as much as
387 * possible.
388 */
389 while (total_scan >= batch_size ||
390 total_scan >= freeable) {
391 unsigned long ret;
392 unsigned long nr_to_scan = min(batch_size, total_scan);
393
394 shrinkctl->nr_to_scan = nr_to_scan;
395 ret = shrinker->scan_objects(shrinker, shrinkctl);
396 if (ret == SHRINK_STOP)
397 break;
398 freed += ret;
399
400 count_vm_events(SLABS_SCANNED, nr_to_scan);
401 total_scan -= nr_to_scan;
402 scanned += nr_to_scan;
403
404 cond_resched();
405 }
406
407 if (next_deferred >= scanned)
408 next_deferred -= scanned;
409 else
410 next_deferred = 0;
411 /*
412 * move the unused scan count back into the shrinker in a
413 * manner that handles concurrent updates. If we exhausted the
414 * scan, there is no need to do an update.
415 */
416 if (next_deferred > 0)
417 new_nr = atomic_long_add_return(next_deferred,
418 &shrinker->nr_deferred[nid]);
419 else
420 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
421
422 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
423 return freed;
424}
425
426/**
427 * shrink_slab - shrink slab caches
428 * @gfp_mask: allocation context
429 * @nid: node whose slab caches to target
430 * @memcg: memory cgroup whose slab caches to target
431 * @nr_scanned: pressure numerator
432 * @nr_eligible: pressure denominator
433 *
434 * Call the shrink functions to age shrinkable caches.
435 *
436 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
437 * unaware shrinkers will receive a node id of 0 instead.
438 *
439 * @memcg specifies the memory cgroup to target. If it is not NULL,
440 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
441 * objects from the memory cgroup specified. Otherwise, only unaware
442 * shrinkers are called.
443 *
444 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
445 * the available objects should be scanned. Page reclaim for example
446 * passes the number of pages scanned and the number of pages on the
447 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
448 * when it encountered mapped pages. The ratio is further biased by
449 * the ->seeks setting of the shrink function, which indicates the
450 * cost to recreate an object relative to that of an LRU page.
451 *
452 * Returns the number of reclaimed slab objects.
453 */
454static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
455 struct mem_cgroup *memcg,
456 unsigned long nr_scanned,
457 unsigned long nr_eligible)
458{
459 struct shrinker *shrinker;
460 unsigned long freed = 0;
461
462 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
463 return 0;
464
465 if (nr_scanned == 0)
466 nr_scanned = SWAP_CLUSTER_MAX;
467
468 if (!down_read_trylock(&shrinker_rwsem)) {
469 /*
470 * If we would return 0, our callers would understand that we
471 * have nothing else to shrink and give up trying. By returning
472 * 1 we keep it going and assume we'll be able to shrink next
473 * time.
474 */
475 freed = 1;
476 goto out;
477 }
478
479 list_for_each_entry(shrinker, &shrinker_list, list) {
480 struct shrink_control sc = {
481 .gfp_mask = gfp_mask,
482 .nid = nid,
483 .memcg = memcg,
484 };
485
486 /*
487 * If kernel memory accounting is disabled, we ignore
488 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
489 * passing NULL for memcg.
490 */
491 if (memcg_kmem_enabled() &&
492 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
493 continue;
494
495 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
496 sc.nid = 0;
497
498 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
499 }
500
501 up_read(&shrinker_rwsem);
502out:
503 cond_resched();
504 return freed;
505}
506
507void drop_slab_node(int nid)
508{
509 unsigned long freed;
510
511 do {
512 struct mem_cgroup *memcg = NULL;
513
514 freed = 0;
515 do {
516 freed += shrink_slab(GFP_KERNEL, nid, memcg,
517 1000, 1000);
518 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
519 } while (freed > 10);
520}
521
522void drop_slab(void)
523{
524 int nid;
525
526 for_each_online_node(nid)
527 drop_slab_node(nid);
528}
529
530static inline int is_page_cache_freeable(struct page *page)
531{
532 /*
533 * A freeable page cache page is referenced only by the caller
534 * that isolated the page, the page cache radix tree and
535 * optional buffer heads at page->private.
536 */
537 return page_count(page) - page_has_private(page) == 2;
538}
539
540static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
541{
542 if (current->flags & PF_SWAPWRITE)
543 return 1;
544 if (!inode_write_congested(inode))
545 return 1;
546 if (inode_to_bdi(inode) == current->backing_dev_info)
547 return 1;
548 return 0;
549}
550
551/*
552 * We detected a synchronous write error writing a page out. Probably
553 * -ENOSPC. We need to propagate that into the address_space for a subsequent
554 * fsync(), msync() or close().
555 *
556 * The tricky part is that after writepage we cannot touch the mapping: nothing
557 * prevents it from being freed up. But we have a ref on the page and once
558 * that page is locked, the mapping is pinned.
559 *
560 * We're allowed to run sleeping lock_page() here because we know the caller has
561 * __GFP_FS.
562 */
563static void handle_write_error(struct address_space *mapping,
564 struct page *page, int error)
565{
566 lock_page(page);
567 if (page_mapping(page) == mapping)
568 mapping_set_error(mapping, error);
569 unlock_page(page);
570}
571
572/* possible outcome of pageout() */
573typedef enum {
574 /* failed to write page out, page is locked */
575 PAGE_KEEP,
576 /* move page to the active list, page is locked */
577 PAGE_ACTIVATE,
578 /* page has been sent to the disk successfully, page is unlocked */
579 PAGE_SUCCESS,
580 /* page is clean and locked */
581 PAGE_CLEAN,
582} pageout_t;
583
584/*
585 * pageout is called by shrink_page_list() for each dirty page.
586 * Calls ->writepage().
587 */
588static pageout_t pageout(struct page *page, struct address_space *mapping,
589 struct scan_control *sc)
590{
591 /*
592 * If the page is dirty, only perform writeback if that write
593 * will be non-blocking. To prevent this allocation from being
594 * stalled by pagecache activity. But note that there may be
595 * stalls if we need to run get_block(). We could test
596 * PagePrivate for that.
597 *
598 * If this process is currently in __generic_file_write_iter() against
599 * this page's queue, we can perform writeback even if that
600 * will block.
601 *
602 * If the page is swapcache, write it back even if that would
603 * block, for some throttling. This happens by accident, because
604 * swap_backing_dev_info is bust: it doesn't reflect the
605 * congestion state of the swapdevs. Easy to fix, if needed.
606 */
607 if (!is_page_cache_freeable(page))
608 return PAGE_KEEP;
609 if (!mapping) {
610 /*
611 * Some data journaling orphaned pages can have
612 * page->mapping == NULL while being dirty with clean buffers.
613 */
614 if (page_has_private(page)) {
615 if (try_to_free_buffers(page)) {
616 ClearPageDirty(page);
617 pr_info("%s: orphaned page\n", __func__);
618 return PAGE_CLEAN;
619 }
620 }
621 return PAGE_KEEP;
622 }
623 if (mapping->a_ops->writepage == NULL)
624 return PAGE_ACTIVATE;
625 if (!may_write_to_inode(mapping->host, sc))
626 return PAGE_KEEP;
627
628 if (clear_page_dirty_for_io(page)) {
629 int res;
630 struct writeback_control wbc = {
631 .sync_mode = WB_SYNC_NONE,
632 .nr_to_write = SWAP_CLUSTER_MAX,
633 .range_start = 0,
634 .range_end = LLONG_MAX,
635 .for_reclaim = 1,
636 };
637
638 SetPageReclaim(page);
639 res = mapping->a_ops->writepage(page, &wbc);
640 if (res < 0)
641 handle_write_error(mapping, page, res);
642 if (res == AOP_WRITEPAGE_ACTIVATE) {
643 ClearPageReclaim(page);
644 return PAGE_ACTIVATE;
645 }
646
647 if (!PageWriteback(page)) {
648 /* synchronous write or broken a_ops? */
649 ClearPageReclaim(page);
650 }
651 trace_mm_vmscan_writepage(page);
652 inc_node_page_state(page, NR_VMSCAN_WRITE);
653 return PAGE_SUCCESS;
654 }
655
656 return PAGE_CLEAN;
657}
658
659/*
660 * Same as remove_mapping, but if the page is removed from the mapping, it
661 * gets returned with a refcount of 0.
662 */
663static int __remove_mapping(struct address_space *mapping, struct page *page,
664 bool reclaimed)
665{
666 unsigned long flags;
667
668 BUG_ON(!PageLocked(page));
669 BUG_ON(mapping != page_mapping(page));
670
671 spin_lock_irqsave(&mapping->tree_lock, flags);
672 /*
673 * The non racy check for a busy page.
674 *
675 * Must be careful with the order of the tests. When someone has
676 * a ref to the page, it may be possible that they dirty it then
677 * drop the reference. So if PageDirty is tested before page_count
678 * here, then the following race may occur:
679 *
680 * get_user_pages(&page);
681 * [user mapping goes away]
682 * write_to(page);
683 * !PageDirty(page) [good]
684 * SetPageDirty(page);
685 * put_page(page);
686 * !page_count(page) [good, discard it]
687 *
688 * [oops, our write_to data is lost]
689 *
690 * Reversing the order of the tests ensures such a situation cannot
691 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
692 * load is not satisfied before that of page->_refcount.
693 *
694 * Note that if SetPageDirty is always performed via set_page_dirty,
695 * and thus under tree_lock, then this ordering is not required.
696 */
697 if (!page_ref_freeze(page, 2))
698 goto cannot_free;
699 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
700 if (unlikely(PageDirty(page))) {
701 page_ref_unfreeze(page, 2);
702 goto cannot_free;
703 }
704
705 if (PageSwapCache(page)) {
706 swp_entry_t swap = { .val = page_private(page) };
707 mem_cgroup_swapout(page, swap);
708 __delete_from_swap_cache(page);
709 spin_unlock_irqrestore(&mapping->tree_lock, flags);
710 swapcache_free(swap);
711 } else {
712 void (*freepage)(struct page *);
713 void *shadow = NULL;
714
715 freepage = mapping->a_ops->freepage;
716 /*
717 * Remember a shadow entry for reclaimed file cache in
718 * order to detect refaults, thus thrashing, later on.
719 *
720 * But don't store shadows in an address space that is
721 * already exiting. This is not just an optizimation,
722 * inode reclaim needs to empty out the radix tree or
723 * the nodes are lost. Don't plant shadows behind its
724 * back.
725 *
726 * We also don't store shadows for DAX mappings because the
727 * only page cache pages found in these are zero pages
728 * covering holes, and because we don't want to mix DAX
729 * exceptional entries and shadow exceptional entries in the
730 * same page_tree.
731 */
732 if (reclaimed && page_is_file_cache(page) &&
733 !mapping_exiting(mapping) && !dax_mapping(mapping))
734 shadow = workingset_eviction(mapping, page);
735 __delete_from_page_cache(page, shadow);
736 spin_unlock_irqrestore(&mapping->tree_lock, flags);
737
738 if (freepage != NULL)
739 freepage(page);
740 }
741
742 return 1;
743
744cannot_free:
745 spin_unlock_irqrestore(&mapping->tree_lock, flags);
746 return 0;
747}
748
749/*
750 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
751 * someone else has a ref on the page, abort and return 0. If it was
752 * successfully detached, return 1. Assumes the caller has a single ref on
753 * this page.
754 */
755int remove_mapping(struct address_space *mapping, struct page *page)
756{
757 if (__remove_mapping(mapping, page, false)) {
758 /*
759 * Unfreezing the refcount with 1 rather than 2 effectively
760 * drops the pagecache ref for us without requiring another
761 * atomic operation.
762 */
763 page_ref_unfreeze(page, 1);
764 return 1;
765 }
766 return 0;
767}
768
769/**
770 * putback_lru_page - put previously isolated page onto appropriate LRU list
771 * @page: page to be put back to appropriate lru list
772 *
773 * Add previously isolated @page to appropriate LRU list.
774 * Page may still be unevictable for other reasons.
775 *
776 * lru_lock must not be held, interrupts must be enabled.
777 */
778void putback_lru_page(struct page *page)
779{
780 bool is_unevictable;
781 int was_unevictable = PageUnevictable(page);
782
783 VM_BUG_ON_PAGE(PageLRU(page), page);
784
785redo:
786 ClearPageUnevictable(page);
787
788 if (page_evictable(page)) {
789 /*
790 * For evictable pages, we can use the cache.
791 * In event of a race, worst case is we end up with an
792 * unevictable page on [in]active list.
793 * We know how to handle that.
794 */
795 is_unevictable = false;
796 lru_cache_add(page);
797 } else {
798 /*
799 * Put unevictable pages directly on zone's unevictable
800 * list.
801 */
802 is_unevictable = true;
803 add_page_to_unevictable_list(page);
804 /*
805 * When racing with an mlock or AS_UNEVICTABLE clearing
806 * (page is unlocked) make sure that if the other thread
807 * does not observe our setting of PG_lru and fails
808 * isolation/check_move_unevictable_pages,
809 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
810 * the page back to the evictable list.
811 *
812 * The other side is TestClearPageMlocked() or shmem_lock().
813 */
814 smp_mb();
815 }
816
817 /*
818 * page's status can change while we move it among lru. If an evictable
819 * page is on unevictable list, it never be freed. To avoid that,
820 * check after we added it to the list, again.
821 */
822 if (is_unevictable && page_evictable(page)) {
823 if (!isolate_lru_page(page)) {
824 put_page(page);
825 goto redo;
826 }
827 /* This means someone else dropped this page from LRU
828 * So, it will be freed or putback to LRU again. There is
829 * nothing to do here.
830 */
831 }
832
833 if (was_unevictable && !is_unevictable)
834 count_vm_event(UNEVICTABLE_PGRESCUED);
835 else if (!was_unevictable && is_unevictable)
836 count_vm_event(UNEVICTABLE_PGCULLED);
837
838 put_page(page); /* drop ref from isolate */
839}
840
841enum page_references {
842 PAGEREF_RECLAIM,
843 PAGEREF_RECLAIM_CLEAN,
844 PAGEREF_KEEP,
845 PAGEREF_ACTIVATE,
846};
847
848static enum page_references page_check_references(struct page *page,
849 struct scan_control *sc)
850{
851 int referenced_ptes, referenced_page;
852 unsigned long vm_flags;
853
854 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
855 &vm_flags);
856 referenced_page = TestClearPageReferenced(page);
857
858 /*
859 * Mlock lost the isolation race with us. Let try_to_unmap()
860 * move the page to the unevictable list.
861 */
862 if (vm_flags & VM_LOCKED)
863 return PAGEREF_RECLAIM;
864
865 if (referenced_ptes) {
866 if (PageSwapBacked(page))
867 return PAGEREF_ACTIVATE;
868 /*
869 * All mapped pages start out with page table
870 * references from the instantiating fault, so we need
871 * to look twice if a mapped file page is used more
872 * than once.
873 *
874 * Mark it and spare it for another trip around the
875 * inactive list. Another page table reference will
876 * lead to its activation.
877 *
878 * Note: the mark is set for activated pages as well
879 * so that recently deactivated but used pages are
880 * quickly recovered.
881 */
882 SetPageReferenced(page);
883
884 if (referenced_page || referenced_ptes > 1)
885 return PAGEREF_ACTIVATE;
886
887 /*
888 * Activate file-backed executable pages after first usage.
889 */
890 if (vm_flags & VM_EXEC)
891 return PAGEREF_ACTIVATE;
892
893 return PAGEREF_KEEP;
894 }
895
896 /* Reclaim if clean, defer dirty pages to writeback */
897 if (referenced_page && !PageSwapBacked(page))
898 return PAGEREF_RECLAIM_CLEAN;
899
900 return PAGEREF_RECLAIM;
901}
902
903/* Check if a page is dirty or under writeback */
904static void page_check_dirty_writeback(struct page *page,
905 bool *dirty, bool *writeback)
906{
907 struct address_space *mapping;
908
909 /*
910 * Anonymous pages are not handled by flushers and must be written
911 * from reclaim context. Do not stall reclaim based on them
912 */
913 if (!page_is_file_cache(page)) {
914 *dirty = false;
915 *writeback = false;
916 return;
917 }
918
919 /* By default assume that the page flags are accurate */
920 *dirty = PageDirty(page);
921 *writeback = PageWriteback(page);
922
923 /* Verify dirty/writeback state if the filesystem supports it */
924 if (!page_has_private(page))
925 return;
926
927 mapping = page_mapping(page);
928 if (mapping && mapping->a_ops->is_dirty_writeback)
929 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
930}
931
932/*
933 * shrink_page_list() returns the number of reclaimed pages
934 */
935static unsigned long shrink_page_list(struct list_head *page_list,
936 struct pglist_data *pgdat,
937 struct scan_control *sc,
938 enum ttu_flags ttu_flags,
939 unsigned long *ret_nr_dirty,
940 unsigned long *ret_nr_unqueued_dirty,
941 unsigned long *ret_nr_congested,
942 unsigned long *ret_nr_writeback,
943 unsigned long *ret_nr_immediate,
944 bool force_reclaim)
945{
946 LIST_HEAD(ret_pages);
947 LIST_HEAD(free_pages);
948 int pgactivate = 0;
949 unsigned long nr_unqueued_dirty = 0;
950 unsigned long nr_dirty = 0;
951 unsigned long nr_congested = 0;
952 unsigned long nr_reclaimed = 0;
953 unsigned long nr_writeback = 0;
954 unsigned long nr_immediate = 0;
955
956 cond_resched();
957
958 while (!list_empty(page_list)) {
959 struct address_space *mapping;
960 struct page *page;
961 int may_enter_fs;
962 enum page_references references = PAGEREF_RECLAIM_CLEAN;
963 bool dirty, writeback;
964 bool lazyfree = false;
965 int ret = SWAP_SUCCESS;
966
967 cond_resched();
968
969 page = lru_to_page(page_list);
970 list_del(&page->lru);
971
972 if (!trylock_page(page))
973 goto keep;
974
975 VM_BUG_ON_PAGE(PageActive(page), page);
976
977 sc->nr_scanned++;
978
979 if (unlikely(!page_evictable(page)))
980 goto cull_mlocked;
981
982 if (!sc->may_unmap && page_mapped(page))
983 goto keep_locked;
984
985 /* Double the slab pressure for mapped and swapcache pages */
986 if (page_mapped(page) || PageSwapCache(page))
987 sc->nr_scanned++;
988
989 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
990 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
991
992 /*
993 * The number of dirty pages determines if a zone is marked
994 * reclaim_congested which affects wait_iff_congested. kswapd
995 * will stall and start writing pages if the tail of the LRU
996 * is all dirty unqueued pages.
997 */
998 page_check_dirty_writeback(page, &dirty, &writeback);
999 if (dirty || writeback)
1000 nr_dirty++;
1001
1002 if (dirty && !writeback)
1003 nr_unqueued_dirty++;
1004
1005 /*
1006 * Treat this page as congested if the underlying BDI is or if
1007 * pages are cycling through the LRU so quickly that the
1008 * pages marked for immediate reclaim are making it to the
1009 * end of the LRU a second time.
1010 */
1011 mapping = page_mapping(page);
1012 if (((dirty || writeback) && mapping &&
1013 inode_write_congested(mapping->host)) ||
1014 (writeback && PageReclaim(page)))
1015 nr_congested++;
1016
1017 /*
1018 * If a page at the tail of the LRU is under writeback, there
1019 * are three cases to consider.
1020 *
1021 * 1) If reclaim is encountering an excessive number of pages
1022 * under writeback and this page is both under writeback and
1023 * PageReclaim then it indicates that pages are being queued
1024 * for IO but are being recycled through the LRU before the
1025 * IO can complete. Waiting on the page itself risks an
1026 * indefinite stall if it is impossible to writeback the
1027 * page due to IO error or disconnected storage so instead
1028 * note that the LRU is being scanned too quickly and the
1029 * caller can stall after page list has been processed.
1030 *
1031 * 2) Global or new memcg reclaim encounters a page that is
1032 * not marked for immediate reclaim, or the caller does not
1033 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1034 * not to fs). In this case mark the page for immediate
1035 * reclaim and continue scanning.
1036 *
1037 * Require may_enter_fs because we would wait on fs, which
1038 * may not have submitted IO yet. And the loop driver might
1039 * enter reclaim, and deadlock if it waits on a page for
1040 * which it is needed to do the write (loop masks off
1041 * __GFP_IO|__GFP_FS for this reason); but more thought
1042 * would probably show more reasons.
1043 *
1044 * 3) Legacy memcg encounters a page that is already marked
1045 * PageReclaim. memcg does not have any dirty pages
1046 * throttling so we could easily OOM just because too many
1047 * pages are in writeback and there is nothing else to
1048 * reclaim. Wait for the writeback to complete.
1049 */
1050 if (PageWriteback(page)) {
1051 /* Case 1 above */
1052 if (current_is_kswapd() &&
1053 PageReclaim(page) &&
1054 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1055 nr_immediate++;
1056 goto keep_locked;
1057
1058 /* Case 2 above */
1059 } else if (sane_reclaim(sc) ||
1060 !PageReclaim(page) || !may_enter_fs) {
1061 /*
1062 * This is slightly racy - end_page_writeback()
1063 * might have just cleared PageReclaim, then
1064 * setting PageReclaim here end up interpreted
1065 * as PageReadahead - but that does not matter
1066 * enough to care. What we do want is for this
1067 * page to have PageReclaim set next time memcg
1068 * reclaim reaches the tests above, so it will
1069 * then wait_on_page_writeback() to avoid OOM;
1070 * and it's also appropriate in global reclaim.
1071 */
1072 SetPageReclaim(page);
1073 nr_writeback++;
1074 goto keep_locked;
1075
1076 /* Case 3 above */
1077 } else {
1078 unlock_page(page);
1079 wait_on_page_writeback(page);
1080 /* then go back and try same page again */
1081 list_add_tail(&page->lru, page_list);
1082 continue;
1083 }
1084 }
1085
1086 if (!force_reclaim)
1087 references = page_check_references(page, sc);
1088
1089 switch (references) {
1090 case PAGEREF_ACTIVATE:
1091 goto activate_locked;
1092 case PAGEREF_KEEP:
1093 goto keep_locked;
1094 case PAGEREF_RECLAIM:
1095 case PAGEREF_RECLAIM_CLEAN:
1096 ; /* try to reclaim the page below */
1097 }
1098
1099 /*
1100 * Anonymous process memory has backing store?
1101 * Try to allocate it some swap space here.
1102 */
1103 if (PageAnon(page) && !PageSwapCache(page)) {
1104 if (!(sc->gfp_mask & __GFP_IO))
1105 goto keep_locked;
1106 if (!add_to_swap(page, page_list))
1107 goto activate_locked;
1108 lazyfree = true;
1109 may_enter_fs = 1;
1110
1111 /* Adding to swap updated mapping */
1112 mapping = page_mapping(page);
1113 } else if (unlikely(PageTransHuge(page))) {
1114 /* Split file THP */
1115 if (split_huge_page_to_list(page, page_list))
1116 goto keep_locked;
1117 }
1118
1119 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1120
1121 /*
1122 * The page is mapped into the page tables of one or more
1123 * processes. Try to unmap it here.
1124 */
1125 if (page_mapped(page) && mapping) {
1126 switch (ret = try_to_unmap(page, lazyfree ?
1127 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1128 (ttu_flags | TTU_BATCH_FLUSH))) {
1129 case SWAP_FAIL:
1130 goto activate_locked;
1131 case SWAP_AGAIN:
1132 goto keep_locked;
1133 case SWAP_MLOCK:
1134 goto cull_mlocked;
1135 case SWAP_LZFREE:
1136 goto lazyfree;
1137 case SWAP_SUCCESS:
1138 ; /* try to free the page below */
1139 }
1140 }
1141
1142 if (PageDirty(page)) {
1143 /*
1144 * Only kswapd can writeback filesystem pages to
1145 * avoid risk of stack overflow but only writeback
1146 * if many dirty pages have been encountered.
1147 */
1148 if (page_is_file_cache(page) &&
1149 (!current_is_kswapd() ||
1150 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1151 /*
1152 * Immediately reclaim when written back.
1153 * Similar in principal to deactivate_page()
1154 * except we already have the page isolated
1155 * and know it's dirty
1156 */
1157 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1158 SetPageReclaim(page);
1159
1160 goto keep_locked;
1161 }
1162
1163 if (references == PAGEREF_RECLAIM_CLEAN)
1164 goto keep_locked;
1165 if (!may_enter_fs)
1166 goto keep_locked;
1167 if (!sc->may_writepage)
1168 goto keep_locked;
1169
1170 /*
1171 * Page is dirty. Flush the TLB if a writable entry
1172 * potentially exists to avoid CPU writes after IO
1173 * starts and then write it out here.
1174 */
1175 try_to_unmap_flush_dirty();
1176 switch (pageout(page, mapping, sc)) {
1177 case PAGE_KEEP:
1178 goto keep_locked;
1179 case PAGE_ACTIVATE:
1180 goto activate_locked;
1181 case PAGE_SUCCESS:
1182 if (PageWriteback(page))
1183 goto keep;
1184 if (PageDirty(page))
1185 goto keep;
1186
1187 /*
1188 * A synchronous write - probably a ramdisk. Go
1189 * ahead and try to reclaim the page.
1190 */
1191 if (!trylock_page(page))
1192 goto keep;
1193 if (PageDirty(page) || PageWriteback(page))
1194 goto keep_locked;
1195 mapping = page_mapping(page);
1196 case PAGE_CLEAN:
1197 ; /* try to free the page below */
1198 }
1199 }
1200
1201 /*
1202 * If the page has buffers, try to free the buffer mappings
1203 * associated with this page. If we succeed we try to free
1204 * the page as well.
1205 *
1206 * We do this even if the page is PageDirty().
1207 * try_to_release_page() does not perform I/O, but it is
1208 * possible for a page to have PageDirty set, but it is actually
1209 * clean (all its buffers are clean). This happens if the
1210 * buffers were written out directly, with submit_bh(). ext3
1211 * will do this, as well as the blockdev mapping.
1212 * try_to_release_page() will discover that cleanness and will
1213 * drop the buffers and mark the page clean - it can be freed.
1214 *
1215 * Rarely, pages can have buffers and no ->mapping. These are
1216 * the pages which were not successfully invalidated in
1217 * truncate_complete_page(). We try to drop those buffers here
1218 * and if that worked, and the page is no longer mapped into
1219 * process address space (page_count == 1) it can be freed.
1220 * Otherwise, leave the page on the LRU so it is swappable.
1221 */
1222 if (page_has_private(page)) {
1223 if (!try_to_release_page(page, sc->gfp_mask))
1224 goto activate_locked;
1225 if (!mapping && page_count(page) == 1) {
1226 unlock_page(page);
1227 if (put_page_testzero(page))
1228 goto free_it;
1229 else {
1230 /*
1231 * rare race with speculative reference.
1232 * the speculative reference will free
1233 * this page shortly, so we may
1234 * increment nr_reclaimed here (and
1235 * leave it off the LRU).
1236 */
1237 nr_reclaimed++;
1238 continue;
1239 }
1240 }
1241 }
1242
1243lazyfree:
1244 if (!mapping || !__remove_mapping(mapping, page, true))
1245 goto keep_locked;
1246
1247 /*
1248 * At this point, we have no other references and there is
1249 * no way to pick any more up (removed from LRU, removed
1250 * from pagecache). Can use non-atomic bitops now (and
1251 * we obviously don't have to worry about waking up a process
1252 * waiting on the page lock, because there are no references.
1253 */
1254 __ClearPageLocked(page);
1255free_it:
1256 if (ret == SWAP_LZFREE)
1257 count_vm_event(PGLAZYFREED);
1258
1259 nr_reclaimed++;
1260
1261 /*
1262 * Is there need to periodically free_page_list? It would
1263 * appear not as the counts should be low
1264 */
1265 list_add(&page->lru, &free_pages);
1266 continue;
1267
1268cull_mlocked:
1269 if (PageSwapCache(page))
1270 try_to_free_swap(page);
1271 unlock_page(page);
1272 list_add(&page->lru, &ret_pages);
1273 continue;
1274
1275activate_locked:
1276 /* Not a candidate for swapping, so reclaim swap space. */
1277 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1278 try_to_free_swap(page);
1279 VM_BUG_ON_PAGE(PageActive(page), page);
1280 SetPageActive(page);
1281 pgactivate++;
1282keep_locked:
1283 unlock_page(page);
1284keep:
1285 list_add(&page->lru, &ret_pages);
1286 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1287 }
1288
1289 mem_cgroup_uncharge_list(&free_pages);
1290 try_to_unmap_flush();
1291 free_hot_cold_page_list(&free_pages, true);
1292
1293 list_splice(&ret_pages, page_list);
1294 count_vm_events(PGACTIVATE, pgactivate);
1295
1296 *ret_nr_dirty += nr_dirty;
1297 *ret_nr_congested += nr_congested;
1298 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1299 *ret_nr_writeback += nr_writeback;
1300 *ret_nr_immediate += nr_immediate;
1301 return nr_reclaimed;
1302}
1303
1304unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1305 struct list_head *page_list)
1306{
1307 struct scan_control sc = {
1308 .gfp_mask = GFP_KERNEL,
1309 .priority = DEF_PRIORITY,
1310 .may_unmap = 1,
1311 };
1312 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1313 struct page *page, *next;
1314 LIST_HEAD(clean_pages);
1315
1316 list_for_each_entry_safe(page, next, page_list, lru) {
1317 if (page_is_file_cache(page) && !PageDirty(page) &&
1318 !__PageMovable(page)) {
1319 ClearPageActive(page);
1320 list_move(&page->lru, &clean_pages);
1321 }
1322 }
1323
1324 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1325 TTU_UNMAP|TTU_IGNORE_ACCESS,
1326 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1327 list_splice(&clean_pages, page_list);
1328 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1329 return ret;
1330}
1331
1332/*
1333 * Attempt to remove the specified page from its LRU. Only take this page
1334 * if it is of the appropriate PageActive status. Pages which are being
1335 * freed elsewhere are also ignored.
1336 *
1337 * page: page to consider
1338 * mode: one of the LRU isolation modes defined above
1339 *
1340 * returns 0 on success, -ve errno on failure.
1341 */
1342int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1343{
1344 int ret = -EINVAL;
1345
1346 /* Only take pages on the LRU. */
1347 if (!PageLRU(page))
1348 return ret;
1349
1350 /* Compaction should not handle unevictable pages but CMA can do so */
1351 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1352 return ret;
1353
1354 ret = -EBUSY;
1355
1356 /*
1357 * To minimise LRU disruption, the caller can indicate that it only
1358 * wants to isolate pages it will be able to operate on without
1359 * blocking - clean pages for the most part.
1360 *
1361 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1362 * is used by reclaim when it is cannot write to backing storage
1363 *
1364 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1365 * that it is possible to migrate without blocking
1366 */
1367 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1368 /* All the caller can do on PageWriteback is block */
1369 if (PageWriteback(page))
1370 return ret;
1371
1372 if (PageDirty(page)) {
1373 struct address_space *mapping;
1374
1375 /* ISOLATE_CLEAN means only clean pages */
1376 if (mode & ISOLATE_CLEAN)
1377 return ret;
1378
1379 /*
1380 * Only pages without mappings or that have a
1381 * ->migratepage callback are possible to migrate
1382 * without blocking
1383 */
1384 mapping = page_mapping(page);
1385 if (mapping && !mapping->a_ops->migratepage)
1386 return ret;
1387 }
1388 }
1389
1390 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1391 return ret;
1392
1393 if (likely(get_page_unless_zero(page))) {
1394 /*
1395 * Be careful not to clear PageLRU until after we're
1396 * sure the page is not being freed elsewhere -- the
1397 * page release code relies on it.
1398 */
1399 ClearPageLRU(page);
1400 ret = 0;
1401 }
1402
1403 return ret;
1404}
1405
1406
1407/*
1408 * Update LRU sizes after isolating pages. The LRU size updates must
1409 * be complete before mem_cgroup_update_lru_size due to a santity check.
1410 */
1411static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1412 enum lru_list lru, unsigned long *nr_zone_taken)
1413{
1414 int zid;
1415
1416 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1417 if (!nr_zone_taken[zid])
1418 continue;
1419
1420 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1421#ifdef CONFIG_MEMCG
1422 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1423#endif
1424 }
1425
1426}
1427
1428/*
1429 * zone_lru_lock is heavily contended. Some of the functions that
1430 * shrink the lists perform better by taking out a batch of pages
1431 * and working on them outside the LRU lock.
1432 *
1433 * For pagecache intensive workloads, this function is the hottest
1434 * spot in the kernel (apart from copy_*_user functions).
1435 *
1436 * Appropriate locks must be held before calling this function.
1437 *
1438 * @nr_to_scan: The number of pages to look through on the list.
1439 * @lruvec: The LRU vector to pull pages from.
1440 * @dst: The temp list to put pages on to.
1441 * @nr_scanned: The number of pages that were scanned.
1442 * @sc: The scan_control struct for this reclaim session
1443 * @mode: One of the LRU isolation modes
1444 * @lru: LRU list id for isolating
1445 *
1446 * returns how many pages were moved onto *@dst.
1447 */
1448static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1449 struct lruvec *lruvec, struct list_head *dst,
1450 unsigned long *nr_scanned, struct scan_control *sc,
1451 isolate_mode_t mode, enum lru_list lru)
1452{
1453 struct list_head *src = &lruvec->lists[lru];
1454 unsigned long nr_taken = 0;
1455 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1456 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1457 unsigned long scan, nr_pages;
1458 LIST_HEAD(pages_skipped);
1459
1460 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1461 !list_empty(src);) {
1462 struct page *page;
1463
1464 page = lru_to_page(src);
1465 prefetchw_prev_lru_page(page, src, flags);
1466
1467 VM_BUG_ON_PAGE(!PageLRU(page), page);
1468
1469 if (page_zonenum(page) > sc->reclaim_idx) {
1470 list_move(&page->lru, &pages_skipped);
1471 nr_skipped[page_zonenum(page)]++;
1472 continue;
1473 }
1474
1475 /*
1476 * Account for scanned and skipped separetly to avoid the pgdat
1477 * being prematurely marked unreclaimable by pgdat_reclaimable.
1478 */
1479 scan++;
1480
1481 switch (__isolate_lru_page(page, mode)) {
1482 case 0:
1483 nr_pages = hpage_nr_pages(page);
1484 nr_taken += nr_pages;
1485 nr_zone_taken[page_zonenum(page)] += nr_pages;
1486 list_move(&page->lru, dst);
1487 break;
1488
1489 case -EBUSY:
1490 /* else it is being freed elsewhere */
1491 list_move(&page->lru, src);
1492 continue;
1493
1494 default:
1495 BUG();
1496 }
1497 }
1498
1499 /*
1500 * Splice any skipped pages to the start of the LRU list. Note that
1501 * this disrupts the LRU order when reclaiming for lower zones but
1502 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1503 * scanning would soon rescan the same pages to skip and put the
1504 * system at risk of premature OOM.
1505 */
1506 if (!list_empty(&pages_skipped)) {
1507 int zid;
1508 unsigned long total_skipped = 0;
1509
1510 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1511 if (!nr_skipped[zid])
1512 continue;
1513
1514 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1515 total_skipped += nr_skipped[zid];
1516 }
1517
1518 /*
1519 * Account skipped pages as a partial scan as the pgdat may be
1520 * close to unreclaimable. If the LRU list is empty, account
1521 * skipped pages as a full scan.
1522 */
1523 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1524
1525 list_splice(&pages_skipped, src);
1526 }
1527 *nr_scanned = scan;
1528 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
1529 nr_taken, mode, is_file_lru(lru));
1530 update_lru_sizes(lruvec, lru, nr_zone_taken);
1531 return nr_taken;
1532}
1533
1534/**
1535 * isolate_lru_page - tries to isolate a page from its LRU list
1536 * @page: page to isolate from its LRU list
1537 *
1538 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1539 * vmstat statistic corresponding to whatever LRU list the page was on.
1540 *
1541 * Returns 0 if the page was removed from an LRU list.
1542 * Returns -EBUSY if the page was not on an LRU list.
1543 *
1544 * The returned page will have PageLRU() cleared. If it was found on
1545 * the active list, it will have PageActive set. If it was found on
1546 * the unevictable list, it will have the PageUnevictable bit set. That flag
1547 * may need to be cleared by the caller before letting the page go.
1548 *
1549 * The vmstat statistic corresponding to the list on which the page was
1550 * found will be decremented.
1551 *
1552 * Restrictions:
1553 * (1) Must be called with an elevated refcount on the page. This is a
1554 * fundamentnal difference from isolate_lru_pages (which is called
1555 * without a stable reference).
1556 * (2) the lru_lock must not be held.
1557 * (3) interrupts must be enabled.
1558 */
1559int isolate_lru_page(struct page *page)
1560{
1561 int ret = -EBUSY;
1562
1563 VM_BUG_ON_PAGE(!page_count(page), page);
1564 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1565
1566 if (PageLRU(page)) {
1567 struct zone *zone = page_zone(page);
1568 struct lruvec *lruvec;
1569
1570 spin_lock_irq(zone_lru_lock(zone));
1571 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1572 if (PageLRU(page)) {
1573 int lru = page_lru(page);
1574 get_page(page);
1575 ClearPageLRU(page);
1576 del_page_from_lru_list(page, lruvec, lru);
1577 ret = 0;
1578 }
1579 spin_unlock_irq(zone_lru_lock(zone));
1580 }
1581 return ret;
1582}
1583
1584/*
1585 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1586 * then get resheduled. When there are massive number of tasks doing page
1587 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1588 * the LRU list will go small and be scanned faster than necessary, leading to
1589 * unnecessary swapping, thrashing and OOM.
1590 */
1591static int too_many_isolated(struct pglist_data *pgdat, int file,
1592 struct scan_control *sc)
1593{
1594 unsigned long inactive, isolated;
1595
1596 if (current_is_kswapd())
1597 return 0;
1598
1599 if (!sane_reclaim(sc))
1600 return 0;
1601
1602 if (file) {
1603 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1604 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1605 } else {
1606 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1607 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1608 }
1609
1610 /*
1611 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1612 * won't get blocked by normal direct-reclaimers, forming a circular
1613 * deadlock.
1614 */
1615 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1616 inactive >>= 3;
1617
1618 return isolated > inactive;
1619}
1620
1621static noinline_for_stack void
1622putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1623{
1624 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1625 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1626 LIST_HEAD(pages_to_free);
1627
1628 /*
1629 * Put back any unfreeable pages.
1630 */
1631 while (!list_empty(page_list)) {
1632 struct page *page = lru_to_page(page_list);
1633 int lru;
1634
1635 VM_BUG_ON_PAGE(PageLRU(page), page);
1636 list_del(&page->lru);
1637 if (unlikely(!page_evictable(page))) {
1638 spin_unlock_irq(&pgdat->lru_lock);
1639 putback_lru_page(page);
1640 spin_lock_irq(&pgdat->lru_lock);
1641 continue;
1642 }
1643
1644 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1645
1646 SetPageLRU(page);
1647 lru = page_lru(page);
1648 add_page_to_lru_list(page, lruvec, lru);
1649
1650 if (is_active_lru(lru)) {
1651 int file = is_file_lru(lru);
1652 int numpages = hpage_nr_pages(page);
1653 reclaim_stat->recent_rotated[file] += numpages;
1654 }
1655 if (put_page_testzero(page)) {
1656 __ClearPageLRU(page);
1657 __ClearPageActive(page);
1658 del_page_from_lru_list(page, lruvec, lru);
1659
1660 if (unlikely(PageCompound(page))) {
1661 spin_unlock_irq(&pgdat->lru_lock);
1662 mem_cgroup_uncharge(page);
1663 (*get_compound_page_dtor(page))(page);
1664 spin_lock_irq(&pgdat->lru_lock);
1665 } else
1666 list_add(&page->lru, &pages_to_free);
1667 }
1668 }
1669
1670 /*
1671 * To save our caller's stack, now use input list for pages to free.
1672 */
1673 list_splice(&pages_to_free, page_list);
1674}
1675
1676/*
1677 * If a kernel thread (such as nfsd for loop-back mounts) services
1678 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1679 * In that case we should only throttle if the backing device it is
1680 * writing to is congested. In other cases it is safe to throttle.
1681 */
1682static int current_may_throttle(void)
1683{
1684 return !(current->flags & PF_LESS_THROTTLE) ||
1685 current->backing_dev_info == NULL ||
1686 bdi_write_congested(current->backing_dev_info);
1687}
1688
1689static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1690 struct scan_control *sc, enum lru_list lru)
1691{
1692 int zid;
1693 struct zone *zone;
1694 int file = is_file_lru(lru);
1695 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1696
1697 if (!global_reclaim(sc))
1698 return true;
1699
1700 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1701 zone = &pgdat->node_zones[zid];
1702 if (!managed_zone(zone))
1703 continue;
1704
1705 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1706 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1707 return true;
1708 }
1709
1710 return false;
1711}
1712
1713/*
1714 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1715 * of reclaimed pages
1716 */
1717static noinline_for_stack unsigned long
1718shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1719 struct scan_control *sc, enum lru_list lru)
1720{
1721 LIST_HEAD(page_list);
1722 unsigned long nr_scanned;
1723 unsigned long nr_reclaimed = 0;
1724 unsigned long nr_taken;
1725 unsigned long nr_dirty = 0;
1726 unsigned long nr_congested = 0;
1727 unsigned long nr_unqueued_dirty = 0;
1728 unsigned long nr_writeback = 0;
1729 unsigned long nr_immediate = 0;
1730 isolate_mode_t isolate_mode = 0;
1731 int file = is_file_lru(lru);
1732 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1733 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1734
1735 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1736 return 0;
1737
1738 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1739 congestion_wait(BLK_RW_ASYNC, HZ/10);
1740
1741 /* We are about to die and free our memory. Return now. */
1742 if (fatal_signal_pending(current))
1743 return SWAP_CLUSTER_MAX;
1744 }
1745
1746 lru_add_drain();
1747
1748 if (!sc->may_unmap)
1749 isolate_mode |= ISOLATE_UNMAPPED;
1750 if (!sc->may_writepage)
1751 isolate_mode |= ISOLATE_CLEAN;
1752
1753 spin_lock_irq(&pgdat->lru_lock);
1754
1755 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1756 &nr_scanned, sc, isolate_mode, lru);
1757
1758 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1759 reclaim_stat->recent_scanned[file] += nr_taken;
1760
1761 if (global_reclaim(sc)) {
1762 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1763 if (current_is_kswapd())
1764 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1765 else
1766 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1767 }
1768 spin_unlock_irq(&pgdat->lru_lock);
1769
1770 if (nr_taken == 0)
1771 return 0;
1772
1773 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1774 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1775 &nr_writeback, &nr_immediate,
1776 false);
1777
1778 spin_lock_irq(&pgdat->lru_lock);
1779
1780 if (global_reclaim(sc)) {
1781 if (current_is_kswapd())
1782 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1783 else
1784 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1785 }
1786
1787 putback_inactive_pages(lruvec, &page_list);
1788
1789 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1790
1791 spin_unlock_irq(&pgdat->lru_lock);
1792
1793 mem_cgroup_uncharge_list(&page_list);
1794 free_hot_cold_page_list(&page_list, true);
1795
1796 /*
1797 * If reclaim is isolating dirty pages under writeback, it implies
1798 * that the long-lived page allocation rate is exceeding the page
1799 * laundering rate. Either the global limits are not being effective
1800 * at throttling processes due to the page distribution throughout
1801 * zones or there is heavy usage of a slow backing device. The
1802 * only option is to throttle from reclaim context which is not ideal
1803 * as there is no guarantee the dirtying process is throttled in the
1804 * same way balance_dirty_pages() manages.
1805 *
1806 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1807 * of pages under pages flagged for immediate reclaim and stall if any
1808 * are encountered in the nr_immediate check below.
1809 */
1810 if (nr_writeback && nr_writeback == nr_taken)
1811 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1812
1813 /*
1814 * Legacy memcg will stall in page writeback so avoid forcibly
1815 * stalling here.
1816 */
1817 if (sane_reclaim(sc)) {
1818 /*
1819 * Tag a zone as congested if all the dirty pages scanned were
1820 * backed by a congested BDI and wait_iff_congested will stall.
1821 */
1822 if (nr_dirty && nr_dirty == nr_congested)
1823 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1824
1825 /*
1826 * If dirty pages are scanned that are not queued for IO, it
1827 * implies that flushers are not keeping up. In this case, flag
1828 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1829 * reclaim context.
1830 */
1831 if (nr_unqueued_dirty == nr_taken)
1832 set_bit(PGDAT_DIRTY, &pgdat->flags);
1833
1834 /*
1835 * If kswapd scans pages marked marked for immediate
1836 * reclaim and under writeback (nr_immediate), it implies
1837 * that pages are cycling through the LRU faster than
1838 * they are written so also forcibly stall.
1839 */
1840 if (nr_immediate && current_may_throttle())
1841 congestion_wait(BLK_RW_ASYNC, HZ/10);
1842 }
1843
1844 /*
1845 * Stall direct reclaim for IO completions if underlying BDIs or zone
1846 * is congested. Allow kswapd to continue until it starts encountering
1847 * unqueued dirty pages or cycling through the LRU too quickly.
1848 */
1849 if (!sc->hibernation_mode && !current_is_kswapd() &&
1850 current_may_throttle())
1851 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1852
1853 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1854 nr_scanned, nr_reclaimed,
1855 sc->priority, file);
1856 return nr_reclaimed;
1857}
1858
1859/*
1860 * This moves pages from the active list to the inactive list.
1861 *
1862 * We move them the other way if the page is referenced by one or more
1863 * processes, from rmap.
1864 *
1865 * If the pages are mostly unmapped, the processing is fast and it is
1866 * appropriate to hold zone_lru_lock across the whole operation. But if
1867 * the pages are mapped, the processing is slow (page_referenced()) so we
1868 * should drop zone_lru_lock around each page. It's impossible to balance
1869 * this, so instead we remove the pages from the LRU while processing them.
1870 * It is safe to rely on PG_active against the non-LRU pages in here because
1871 * nobody will play with that bit on a non-LRU page.
1872 *
1873 * The downside is that we have to touch page->_refcount against each page.
1874 * But we had to alter page->flags anyway.
1875 */
1876
1877static void move_active_pages_to_lru(struct lruvec *lruvec,
1878 struct list_head *list,
1879 struct list_head *pages_to_free,
1880 enum lru_list lru)
1881{
1882 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1883 unsigned long pgmoved = 0;
1884 struct page *page;
1885 int nr_pages;
1886
1887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
1889 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1890
1891 VM_BUG_ON_PAGE(PageLRU(page), page);
1892 SetPageLRU(page);
1893
1894 nr_pages = hpage_nr_pages(page);
1895 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1896 list_move(&page->lru, &lruvec->lists[lru]);
1897 pgmoved += nr_pages;
1898
1899 if (put_page_testzero(page)) {
1900 __ClearPageLRU(page);
1901 __ClearPageActive(page);
1902 del_page_from_lru_list(page, lruvec, lru);
1903
1904 if (unlikely(PageCompound(page))) {
1905 spin_unlock_irq(&pgdat->lru_lock);
1906 mem_cgroup_uncharge(page);
1907 (*get_compound_page_dtor(page))(page);
1908 spin_lock_irq(&pgdat->lru_lock);
1909 } else
1910 list_add(&page->lru, pages_to_free);
1911 }
1912 }
1913
1914 if (!is_active_lru(lru))
1915 __count_vm_events(PGDEACTIVATE, pgmoved);
1916}
1917
1918static void shrink_active_list(unsigned long nr_to_scan,
1919 struct lruvec *lruvec,
1920 struct scan_control *sc,
1921 enum lru_list lru)
1922{
1923 unsigned long nr_taken;
1924 unsigned long nr_scanned;
1925 unsigned long vm_flags;
1926 LIST_HEAD(l_hold); /* The pages which were snipped off */
1927 LIST_HEAD(l_active);
1928 LIST_HEAD(l_inactive);
1929 struct page *page;
1930 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1931 unsigned long nr_rotated = 0;
1932 isolate_mode_t isolate_mode = 0;
1933 int file = is_file_lru(lru);
1934 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1935
1936 lru_add_drain();
1937
1938 if (!sc->may_unmap)
1939 isolate_mode |= ISOLATE_UNMAPPED;
1940 if (!sc->may_writepage)
1941 isolate_mode |= ISOLATE_CLEAN;
1942
1943 spin_lock_irq(&pgdat->lru_lock);
1944
1945 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1946 &nr_scanned, sc, isolate_mode, lru);
1947
1948 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1949 reclaim_stat->recent_scanned[file] += nr_taken;
1950
1951 if (global_reclaim(sc))
1952 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1953 __count_vm_events(PGREFILL, nr_scanned);
1954
1955 spin_unlock_irq(&pgdat->lru_lock);
1956
1957 while (!list_empty(&l_hold)) {
1958 cond_resched();
1959 page = lru_to_page(&l_hold);
1960 list_del(&page->lru);
1961
1962 if (unlikely(!page_evictable(page))) {
1963 putback_lru_page(page);
1964 continue;
1965 }
1966
1967 if (unlikely(buffer_heads_over_limit)) {
1968 if (page_has_private(page) && trylock_page(page)) {
1969 if (page_has_private(page))
1970 try_to_release_page(page, 0);
1971 unlock_page(page);
1972 }
1973 }
1974
1975 if (page_referenced(page, 0, sc->target_mem_cgroup,
1976 &vm_flags)) {
1977 nr_rotated += hpage_nr_pages(page);
1978 /*
1979 * Identify referenced, file-backed active pages and
1980 * give them one more trip around the active list. So
1981 * that executable code get better chances to stay in
1982 * memory under moderate memory pressure. Anon pages
1983 * are not likely to be evicted by use-once streaming
1984 * IO, plus JVM can create lots of anon VM_EXEC pages,
1985 * so we ignore them here.
1986 */
1987 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1988 list_add(&page->lru, &l_active);
1989 continue;
1990 }
1991 }
1992
1993 ClearPageActive(page); /* we are de-activating */
1994 list_add(&page->lru, &l_inactive);
1995 }
1996
1997 /*
1998 * Move pages back to the lru list.
1999 */
2000 spin_lock_irq(&pgdat->lru_lock);
2001 /*
2002 * Count referenced pages from currently used mappings as rotated,
2003 * even though only some of them are actually re-activated. This
2004 * helps balance scan pressure between file and anonymous pages in
2005 * get_scan_count.
2006 */
2007 reclaim_stat->recent_rotated[file] += nr_rotated;
2008
2009 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2010 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2011 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2012 spin_unlock_irq(&pgdat->lru_lock);
2013
2014 mem_cgroup_uncharge_list(&l_hold);
2015 free_hot_cold_page_list(&l_hold, true);
2016}
2017
2018/*
2019 * The inactive anon list should be small enough that the VM never has
2020 * to do too much work.
2021 *
2022 * The inactive file list should be small enough to leave most memory
2023 * to the established workingset on the scan-resistant active list,
2024 * but large enough to avoid thrashing the aggregate readahead window.
2025 *
2026 * Both inactive lists should also be large enough that each inactive
2027 * page has a chance to be referenced again before it is reclaimed.
2028 *
2029 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2030 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2031 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2032 *
2033 * total target max
2034 * memory ratio inactive
2035 * -------------------------------------
2036 * 10MB 1 5MB
2037 * 100MB 1 50MB
2038 * 1GB 3 250MB
2039 * 10GB 10 0.9GB
2040 * 100GB 31 3GB
2041 * 1TB 101 10GB
2042 * 10TB 320 32GB
2043 */
2044static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2045 struct scan_control *sc)
2046{
2047 unsigned long inactive_ratio;
2048 unsigned long inactive, active;
2049 enum lru_list inactive_lru = file * LRU_FILE;
2050 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2051 unsigned long gb;
2052
2053 /*
2054 * If we don't have swap space, anonymous page deactivation
2055 * is pointless.
2056 */
2057 if (!file && !total_swap_pages)
2058 return false;
2059
2060 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2061 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2062
2063 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2064 if (gb)
2065 inactive_ratio = int_sqrt(10 * gb);
2066 else
2067 inactive_ratio = 1;
2068
2069 return inactive * inactive_ratio < active;
2070}
2071
2072static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2073 struct lruvec *lruvec, struct scan_control *sc)
2074{
2075 if (is_active_lru(lru)) {
2076 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
2077 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2078 return 0;
2079 }
2080
2081 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2082}
2083
2084enum scan_balance {
2085 SCAN_EQUAL,
2086 SCAN_FRACT,
2087 SCAN_ANON,
2088 SCAN_FILE,
2089};
2090
2091/*
2092 * Determine how aggressively the anon and file LRU lists should be
2093 * scanned. The relative value of each set of LRU lists is determined
2094 * by looking at the fraction of the pages scanned we did rotate back
2095 * onto the active list instead of evict.
2096 *
2097 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2098 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2099 */
2100static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2101 struct scan_control *sc, unsigned long *nr,
2102 unsigned long *lru_pages)
2103{
2104 int swappiness = mem_cgroup_swappiness(memcg);
2105 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2106 u64 fraction[2];
2107 u64 denominator = 0; /* gcc */
2108 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2109 unsigned long anon_prio, file_prio;
2110 enum scan_balance scan_balance;
2111 unsigned long anon, file;
2112 bool force_scan = false;
2113 unsigned long ap, fp;
2114 enum lru_list lru;
2115 bool some_scanned;
2116 int pass;
2117
2118 /*
2119 * If the zone or memcg is small, nr[l] can be 0. This
2120 * results in no scanning on this priority and a potential
2121 * priority drop. Global direct reclaim can go to the next
2122 * zone and tends to have no problems. Global kswapd is for
2123 * zone balancing and it needs to scan a minimum amount. When
2124 * reclaiming for a memcg, a priority drop can cause high
2125 * latencies, so it's better to scan a minimum amount there as
2126 * well.
2127 */
2128 if (current_is_kswapd()) {
2129 if (!pgdat_reclaimable(pgdat))
2130 force_scan = true;
2131 if (!mem_cgroup_online(memcg))
2132 force_scan = true;
2133 }
2134 if (!global_reclaim(sc))
2135 force_scan = true;
2136
2137 /* If we have no swap space, do not bother scanning anon pages. */
2138 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2139 scan_balance = SCAN_FILE;
2140 goto out;
2141 }
2142
2143 /*
2144 * Global reclaim will swap to prevent OOM even with no
2145 * swappiness, but memcg users want to use this knob to
2146 * disable swapping for individual groups completely when
2147 * using the memory controller's swap limit feature would be
2148 * too expensive.
2149 */
2150 if (!global_reclaim(sc) && !swappiness) {
2151 scan_balance = SCAN_FILE;
2152 goto out;
2153 }
2154
2155 /*
2156 * Do not apply any pressure balancing cleverness when the
2157 * system is close to OOM, scan both anon and file equally
2158 * (unless the swappiness setting disagrees with swapping).
2159 */
2160 if (!sc->priority && swappiness) {
2161 scan_balance = SCAN_EQUAL;
2162 goto out;
2163 }
2164
2165 /*
2166 * Prevent the reclaimer from falling into the cache trap: as
2167 * cache pages start out inactive, every cache fault will tip
2168 * the scan balance towards the file LRU. And as the file LRU
2169 * shrinks, so does the window for rotation from references.
2170 * This means we have a runaway feedback loop where a tiny
2171 * thrashing file LRU becomes infinitely more attractive than
2172 * anon pages. Try to detect this based on file LRU size.
2173 */
2174 if (global_reclaim(sc)) {
2175 unsigned long pgdatfile;
2176 unsigned long pgdatfree;
2177 int z;
2178 unsigned long total_high_wmark = 0;
2179
2180 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2181 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2182 node_page_state(pgdat, NR_INACTIVE_FILE);
2183
2184 for (z = 0; z < MAX_NR_ZONES; z++) {
2185 struct zone *zone = &pgdat->node_zones[z];
2186 if (!managed_zone(zone))
2187 continue;
2188
2189 total_high_wmark += high_wmark_pages(zone);
2190 }
2191
2192 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2193 scan_balance = SCAN_ANON;
2194 goto out;
2195 }
2196 }
2197
2198 /*
2199 * If there is enough inactive page cache, i.e. if the size of the
2200 * inactive list is greater than that of the active list *and* the
2201 * inactive list actually has some pages to scan on this priority, we
2202 * do not reclaim anything from the anonymous working set right now.
2203 * Without the second condition we could end up never scanning an
2204 * lruvec even if it has plenty of old anonymous pages unless the
2205 * system is under heavy pressure.
2206 */
2207 if (!inactive_list_is_low(lruvec, true, sc) &&
2208 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2209 scan_balance = SCAN_FILE;
2210 goto out;
2211 }
2212
2213 scan_balance = SCAN_FRACT;
2214
2215 /*
2216 * With swappiness at 100, anonymous and file have the same priority.
2217 * This scanning priority is essentially the inverse of IO cost.
2218 */
2219 anon_prio = swappiness;
2220 file_prio = 200 - anon_prio;
2221
2222 /*
2223 * OK, so we have swap space and a fair amount of page cache
2224 * pages. We use the recently rotated / recently scanned
2225 * ratios to determine how valuable each cache is.
2226 *
2227 * Because workloads change over time (and to avoid overflow)
2228 * we keep these statistics as a floating average, which ends
2229 * up weighing recent references more than old ones.
2230 *
2231 * anon in [0], file in [1]
2232 */
2233
2234 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2235 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2236 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2237 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2238
2239 spin_lock_irq(&pgdat->lru_lock);
2240 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2241 reclaim_stat->recent_scanned[0] /= 2;
2242 reclaim_stat->recent_rotated[0] /= 2;
2243 }
2244
2245 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2246 reclaim_stat->recent_scanned[1] /= 2;
2247 reclaim_stat->recent_rotated[1] /= 2;
2248 }
2249
2250 /*
2251 * The amount of pressure on anon vs file pages is inversely
2252 * proportional to the fraction of recently scanned pages on
2253 * each list that were recently referenced and in active use.
2254 */
2255 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2256 ap /= reclaim_stat->recent_rotated[0] + 1;
2257
2258 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2259 fp /= reclaim_stat->recent_rotated[1] + 1;
2260 spin_unlock_irq(&pgdat->lru_lock);
2261
2262 fraction[0] = ap;
2263 fraction[1] = fp;
2264 denominator = ap + fp + 1;
2265out:
2266 some_scanned = false;
2267 /* Only use force_scan on second pass. */
2268 for (pass = 0; !some_scanned && pass < 2; pass++) {
2269 *lru_pages = 0;
2270 for_each_evictable_lru(lru) {
2271 int file = is_file_lru(lru);
2272 unsigned long size;
2273 unsigned long scan;
2274
2275 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2276 scan = size >> sc->priority;
2277
2278 if (!scan && pass && force_scan)
2279 scan = min(size, SWAP_CLUSTER_MAX);
2280
2281 switch (scan_balance) {
2282 case SCAN_EQUAL:
2283 /* Scan lists relative to size */
2284 break;
2285 case SCAN_FRACT:
2286 /*
2287 * Scan types proportional to swappiness and
2288 * their relative recent reclaim efficiency.
2289 */
2290 scan = div64_u64(scan * fraction[file],
2291 denominator);
2292 break;
2293 case SCAN_FILE:
2294 case SCAN_ANON:
2295 /* Scan one type exclusively */
2296 if ((scan_balance == SCAN_FILE) != file) {
2297 size = 0;
2298 scan = 0;
2299 }
2300 break;
2301 default:
2302 /* Look ma, no brain */
2303 BUG();
2304 }
2305
2306 *lru_pages += size;
2307 nr[lru] = scan;
2308
2309 /*
2310 * Skip the second pass and don't force_scan,
2311 * if we found something to scan.
2312 */
2313 some_scanned |= !!scan;
2314 }
2315 }
2316}
2317
2318/*
2319 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2320 */
2321static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2322 struct scan_control *sc, unsigned long *lru_pages)
2323{
2324 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2325 unsigned long nr[NR_LRU_LISTS];
2326 unsigned long targets[NR_LRU_LISTS];
2327 unsigned long nr_to_scan;
2328 enum lru_list lru;
2329 unsigned long nr_reclaimed = 0;
2330 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2331 struct blk_plug plug;
2332 bool scan_adjusted;
2333
2334 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2335
2336 /* Record the original scan target for proportional adjustments later */
2337 memcpy(targets, nr, sizeof(nr));
2338
2339 /*
2340 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2341 * event that can occur when there is little memory pressure e.g.
2342 * multiple streaming readers/writers. Hence, we do not abort scanning
2343 * when the requested number of pages are reclaimed when scanning at
2344 * DEF_PRIORITY on the assumption that the fact we are direct
2345 * reclaiming implies that kswapd is not keeping up and it is best to
2346 * do a batch of work at once. For memcg reclaim one check is made to
2347 * abort proportional reclaim if either the file or anon lru has already
2348 * dropped to zero at the first pass.
2349 */
2350 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2351 sc->priority == DEF_PRIORITY);
2352
2353 blk_start_plug(&plug);
2354 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2355 nr[LRU_INACTIVE_FILE]) {
2356 unsigned long nr_anon, nr_file, percentage;
2357 unsigned long nr_scanned;
2358
2359 for_each_evictable_lru(lru) {
2360 if (nr[lru]) {
2361 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2362 nr[lru] -= nr_to_scan;
2363
2364 nr_reclaimed += shrink_list(lru, nr_to_scan,
2365 lruvec, sc);
2366 }
2367 }
2368
2369 cond_resched();
2370
2371 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2372 continue;
2373
2374 /*
2375 * For kswapd and memcg, reclaim at least the number of pages
2376 * requested. Ensure that the anon and file LRUs are scanned
2377 * proportionally what was requested by get_scan_count(). We
2378 * stop reclaiming one LRU and reduce the amount scanning
2379 * proportional to the original scan target.
2380 */
2381 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2382 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2383
2384 /*
2385 * It's just vindictive to attack the larger once the smaller
2386 * has gone to zero. And given the way we stop scanning the
2387 * smaller below, this makes sure that we only make one nudge
2388 * towards proportionality once we've got nr_to_reclaim.
2389 */
2390 if (!nr_file || !nr_anon)
2391 break;
2392
2393 if (nr_file > nr_anon) {
2394 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2395 targets[LRU_ACTIVE_ANON] + 1;
2396 lru = LRU_BASE;
2397 percentage = nr_anon * 100 / scan_target;
2398 } else {
2399 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2400 targets[LRU_ACTIVE_FILE] + 1;
2401 lru = LRU_FILE;
2402 percentage = nr_file * 100 / scan_target;
2403 }
2404
2405 /* Stop scanning the smaller of the LRU */
2406 nr[lru] = 0;
2407 nr[lru + LRU_ACTIVE] = 0;
2408
2409 /*
2410 * Recalculate the other LRU scan count based on its original
2411 * scan target and the percentage scanning already complete
2412 */
2413 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2414 nr_scanned = targets[lru] - nr[lru];
2415 nr[lru] = targets[lru] * (100 - percentage) / 100;
2416 nr[lru] -= min(nr[lru], nr_scanned);
2417
2418 lru += LRU_ACTIVE;
2419 nr_scanned = targets[lru] - nr[lru];
2420 nr[lru] = targets[lru] * (100 - percentage) / 100;
2421 nr[lru] -= min(nr[lru], nr_scanned);
2422
2423 scan_adjusted = true;
2424 }
2425 blk_finish_plug(&plug);
2426 sc->nr_reclaimed += nr_reclaimed;
2427
2428 /*
2429 * Even if we did not try to evict anon pages at all, we want to
2430 * rebalance the anon lru active/inactive ratio.
2431 */
2432 if (inactive_list_is_low(lruvec, false, sc))
2433 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2434 sc, LRU_ACTIVE_ANON);
2435}
2436
2437/* Use reclaim/compaction for costly allocs or under memory pressure */
2438static bool in_reclaim_compaction(struct scan_control *sc)
2439{
2440 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2441 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2442 sc->priority < DEF_PRIORITY - 2))
2443 return true;
2444
2445 return false;
2446}
2447
2448/*
2449 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2450 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2451 * true if more pages should be reclaimed such that when the page allocator
2452 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2453 * It will give up earlier than that if there is difficulty reclaiming pages.
2454 */
2455static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2456 unsigned long nr_reclaimed,
2457 unsigned long nr_scanned,
2458 struct scan_control *sc)
2459{
2460 unsigned long pages_for_compaction;
2461 unsigned long inactive_lru_pages;
2462 int z;
2463
2464 /* If not in reclaim/compaction mode, stop */
2465 if (!in_reclaim_compaction(sc))
2466 return false;
2467
2468 /* Consider stopping depending on scan and reclaim activity */
2469 if (sc->gfp_mask & __GFP_REPEAT) {
2470 /*
2471 * For __GFP_REPEAT allocations, stop reclaiming if the
2472 * full LRU list has been scanned and we are still failing
2473 * to reclaim pages. This full LRU scan is potentially
2474 * expensive but a __GFP_REPEAT caller really wants to succeed
2475 */
2476 if (!nr_reclaimed && !nr_scanned)
2477 return false;
2478 } else {
2479 /*
2480 * For non-__GFP_REPEAT allocations which can presumably
2481 * fail without consequence, stop if we failed to reclaim
2482 * any pages from the last SWAP_CLUSTER_MAX number of
2483 * pages that were scanned. This will return to the
2484 * caller faster at the risk reclaim/compaction and
2485 * the resulting allocation attempt fails
2486 */
2487 if (!nr_reclaimed)
2488 return false;
2489 }
2490
2491 /*
2492 * If we have not reclaimed enough pages for compaction and the
2493 * inactive lists are large enough, continue reclaiming
2494 */
2495 pages_for_compaction = compact_gap(sc->order);
2496 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2497 if (get_nr_swap_pages() > 0)
2498 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2499 if (sc->nr_reclaimed < pages_for_compaction &&
2500 inactive_lru_pages > pages_for_compaction)
2501 return true;
2502
2503 /* If compaction would go ahead or the allocation would succeed, stop */
2504 for (z = 0; z <= sc->reclaim_idx; z++) {
2505 struct zone *zone = &pgdat->node_zones[z];
2506 if (!managed_zone(zone))
2507 continue;
2508
2509 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2510 case COMPACT_SUCCESS:
2511 case COMPACT_CONTINUE:
2512 return false;
2513 default:
2514 /* check next zone */
2515 ;
2516 }
2517 }
2518 return true;
2519}
2520
2521static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2522{
2523 struct reclaim_state *reclaim_state = current->reclaim_state;
2524 unsigned long nr_reclaimed, nr_scanned;
2525 bool reclaimable = false;
2526
2527 do {
2528 struct mem_cgroup *root = sc->target_mem_cgroup;
2529 struct mem_cgroup_reclaim_cookie reclaim = {
2530 .pgdat = pgdat,
2531 .priority = sc->priority,
2532 };
2533 unsigned long node_lru_pages = 0;
2534 struct mem_cgroup *memcg;
2535
2536 nr_reclaimed = sc->nr_reclaimed;
2537 nr_scanned = sc->nr_scanned;
2538
2539 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2540 do {
2541 unsigned long lru_pages;
2542 unsigned long reclaimed;
2543 unsigned long scanned;
2544
2545 if (mem_cgroup_low(root, memcg)) {
2546 if (!sc->may_thrash)
2547 continue;
2548 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2549 }
2550
2551 reclaimed = sc->nr_reclaimed;
2552 scanned = sc->nr_scanned;
2553
2554 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2555 node_lru_pages += lru_pages;
2556
2557 if (memcg)
2558 shrink_slab(sc->gfp_mask, pgdat->node_id,
2559 memcg, sc->nr_scanned - scanned,
2560 lru_pages);
2561
2562 /* Record the group's reclaim efficiency */
2563 vmpressure(sc->gfp_mask, memcg, false,
2564 sc->nr_scanned - scanned,
2565 sc->nr_reclaimed - reclaimed);
2566
2567 /*
2568 * Direct reclaim and kswapd have to scan all memory
2569 * cgroups to fulfill the overall scan target for the
2570 * node.
2571 *
2572 * Limit reclaim, on the other hand, only cares about
2573 * nr_to_reclaim pages to be reclaimed and it will
2574 * retry with decreasing priority if one round over the
2575 * whole hierarchy is not sufficient.
2576 */
2577 if (!global_reclaim(sc) &&
2578 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2579 mem_cgroup_iter_break(root, memcg);
2580 break;
2581 }
2582 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2583
2584 /*
2585 * Shrink the slab caches in the same proportion that
2586 * the eligible LRU pages were scanned.
2587 */
2588 if (global_reclaim(sc))
2589 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2590 sc->nr_scanned - nr_scanned,
2591 node_lru_pages);
2592
2593 if (reclaim_state) {
2594 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2595 reclaim_state->reclaimed_slab = 0;
2596 }
2597
2598 /* Record the subtree's reclaim efficiency */
2599 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2600 sc->nr_scanned - nr_scanned,
2601 sc->nr_reclaimed - nr_reclaimed);
2602
2603 if (sc->nr_reclaimed - nr_reclaimed)
2604 reclaimable = true;
2605
2606 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2607 sc->nr_scanned - nr_scanned, sc));
2608
2609 return reclaimable;
2610}
2611
2612/*
2613 * Returns true if compaction should go ahead for a costly-order request, or
2614 * the allocation would already succeed without compaction. Return false if we
2615 * should reclaim first.
2616 */
2617static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2618{
2619 unsigned long watermark;
2620 enum compact_result suitable;
2621
2622 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2623 if (suitable == COMPACT_SUCCESS)
2624 /* Allocation should succeed already. Don't reclaim. */
2625 return true;
2626 if (suitable == COMPACT_SKIPPED)
2627 /* Compaction cannot yet proceed. Do reclaim. */
2628 return false;
2629
2630 /*
2631 * Compaction is already possible, but it takes time to run and there
2632 * are potentially other callers using the pages just freed. So proceed
2633 * with reclaim to make a buffer of free pages available to give
2634 * compaction a reasonable chance of completing and allocating the page.
2635 * Note that we won't actually reclaim the whole buffer in one attempt
2636 * as the target watermark in should_continue_reclaim() is lower. But if
2637 * we are already above the high+gap watermark, don't reclaim at all.
2638 */
2639 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2640
2641 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2642}
2643
2644/*
2645 * This is the direct reclaim path, for page-allocating processes. We only
2646 * try to reclaim pages from zones which will satisfy the caller's allocation
2647 * request.
2648 *
2649 * If a zone is deemed to be full of pinned pages then just give it a light
2650 * scan then give up on it.
2651 */
2652static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2653{
2654 struct zoneref *z;
2655 struct zone *zone;
2656 unsigned long nr_soft_reclaimed;
2657 unsigned long nr_soft_scanned;
2658 gfp_t orig_mask;
2659 pg_data_t *last_pgdat = NULL;
2660
2661 /*
2662 * If the number of buffer_heads in the machine exceeds the maximum
2663 * allowed level, force direct reclaim to scan the highmem zone as
2664 * highmem pages could be pinning lowmem pages storing buffer_heads
2665 */
2666 orig_mask = sc->gfp_mask;
2667 if (buffer_heads_over_limit) {
2668 sc->gfp_mask |= __GFP_HIGHMEM;
2669 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2670 }
2671
2672 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2673 sc->reclaim_idx, sc->nodemask) {
2674 /*
2675 * Take care memory controller reclaiming has small influence
2676 * to global LRU.
2677 */
2678 if (global_reclaim(sc)) {
2679 if (!cpuset_zone_allowed(zone,
2680 GFP_KERNEL | __GFP_HARDWALL))
2681 continue;
2682
2683 if (sc->priority != DEF_PRIORITY &&
2684 !pgdat_reclaimable(zone->zone_pgdat))
2685 continue; /* Let kswapd poll it */
2686
2687 /*
2688 * If we already have plenty of memory free for
2689 * compaction in this zone, don't free any more.
2690 * Even though compaction is invoked for any
2691 * non-zero order, only frequent costly order
2692 * reclamation is disruptive enough to become a
2693 * noticeable problem, like transparent huge
2694 * page allocations.
2695 */
2696 if (IS_ENABLED(CONFIG_COMPACTION) &&
2697 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2698 compaction_ready(zone, sc)) {
2699 sc->compaction_ready = true;
2700 continue;
2701 }
2702
2703 /*
2704 * Shrink each node in the zonelist once. If the
2705 * zonelist is ordered by zone (not the default) then a
2706 * node may be shrunk multiple times but in that case
2707 * the user prefers lower zones being preserved.
2708 */
2709 if (zone->zone_pgdat == last_pgdat)
2710 continue;
2711
2712 /*
2713 * This steals pages from memory cgroups over softlimit
2714 * and returns the number of reclaimed pages and
2715 * scanned pages. This works for global memory pressure
2716 * and balancing, not for a memcg's limit.
2717 */
2718 nr_soft_scanned = 0;
2719 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2720 sc->order, sc->gfp_mask,
2721 &nr_soft_scanned);
2722 sc->nr_reclaimed += nr_soft_reclaimed;
2723 sc->nr_scanned += nr_soft_scanned;
2724 /* need some check for avoid more shrink_zone() */
2725 }
2726
2727 /* See comment about same check for global reclaim above */
2728 if (zone->zone_pgdat == last_pgdat)
2729 continue;
2730 last_pgdat = zone->zone_pgdat;
2731 shrink_node(zone->zone_pgdat, sc);
2732 }
2733
2734 /*
2735 * Restore to original mask to avoid the impact on the caller if we
2736 * promoted it to __GFP_HIGHMEM.
2737 */
2738 sc->gfp_mask = orig_mask;
2739}
2740
2741/*
2742 * This is the main entry point to direct page reclaim.
2743 *
2744 * If a full scan of the inactive list fails to free enough memory then we
2745 * are "out of memory" and something needs to be killed.
2746 *
2747 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2748 * high - the zone may be full of dirty or under-writeback pages, which this
2749 * caller can't do much about. We kick the writeback threads and take explicit
2750 * naps in the hope that some of these pages can be written. But if the
2751 * allocating task holds filesystem locks which prevent writeout this might not
2752 * work, and the allocation attempt will fail.
2753 *
2754 * returns: 0, if no pages reclaimed
2755 * else, the number of pages reclaimed
2756 */
2757static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2758 struct scan_control *sc)
2759{
2760 int initial_priority = sc->priority;
2761 unsigned long total_scanned = 0;
2762 unsigned long writeback_threshold;
2763retry:
2764 delayacct_freepages_start();
2765
2766 if (global_reclaim(sc))
2767 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2768
2769 do {
2770 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2771 sc->priority);
2772 sc->nr_scanned = 0;
2773 shrink_zones(zonelist, sc);
2774
2775 total_scanned += sc->nr_scanned;
2776 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2777 break;
2778
2779 if (sc->compaction_ready)
2780 break;
2781
2782 /*
2783 * If we're getting trouble reclaiming, start doing
2784 * writepage even in laptop mode.
2785 */
2786 if (sc->priority < DEF_PRIORITY - 2)
2787 sc->may_writepage = 1;
2788
2789 /*
2790 * Try to write back as many pages as we just scanned. This
2791 * tends to cause slow streaming writers to write data to the
2792 * disk smoothly, at the dirtying rate, which is nice. But
2793 * that's undesirable in laptop mode, where we *want* lumpy
2794 * writeout. So in laptop mode, write out the whole world.
2795 */
2796 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2797 if (total_scanned > writeback_threshold) {
2798 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2799 WB_REASON_TRY_TO_FREE_PAGES);
2800 sc->may_writepage = 1;
2801 }
2802 } while (--sc->priority >= 0);
2803
2804 delayacct_freepages_end();
2805
2806 if (sc->nr_reclaimed)
2807 return sc->nr_reclaimed;
2808
2809 /* Aborted reclaim to try compaction? don't OOM, then */
2810 if (sc->compaction_ready)
2811 return 1;
2812
2813 /* Untapped cgroup reserves? Don't OOM, retry. */
2814 if (!sc->may_thrash) {
2815 sc->priority = initial_priority;
2816 sc->may_thrash = 1;
2817 goto retry;
2818 }
2819
2820 return 0;
2821}
2822
2823static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2824{
2825 struct zone *zone;
2826 unsigned long pfmemalloc_reserve = 0;
2827 unsigned long free_pages = 0;
2828 int i;
2829 bool wmark_ok;
2830
2831 for (i = 0; i <= ZONE_NORMAL; i++) {
2832 zone = &pgdat->node_zones[i];
2833 if (!managed_zone(zone) ||
2834 pgdat_reclaimable_pages(pgdat) == 0)
2835 continue;
2836
2837 pfmemalloc_reserve += min_wmark_pages(zone);
2838 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2839 }
2840
2841 /* If there are no reserves (unexpected config) then do not throttle */
2842 if (!pfmemalloc_reserve)
2843 return true;
2844
2845 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2846
2847 /* kswapd must be awake if processes are being throttled */
2848 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2849 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2850 (enum zone_type)ZONE_NORMAL);
2851 wake_up_interruptible(&pgdat->kswapd_wait);
2852 }
2853
2854 return wmark_ok;
2855}
2856
2857/*
2858 * Throttle direct reclaimers if backing storage is backed by the network
2859 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2860 * depleted. kswapd will continue to make progress and wake the processes
2861 * when the low watermark is reached.
2862 *
2863 * Returns true if a fatal signal was delivered during throttling. If this
2864 * happens, the page allocator should not consider triggering the OOM killer.
2865 */
2866static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2867 nodemask_t *nodemask)
2868{
2869 struct zoneref *z;
2870 struct zone *zone;
2871 pg_data_t *pgdat = NULL;
2872
2873 /*
2874 * Kernel threads should not be throttled as they may be indirectly
2875 * responsible for cleaning pages necessary for reclaim to make forward
2876 * progress. kjournald for example may enter direct reclaim while
2877 * committing a transaction where throttling it could forcing other
2878 * processes to block on log_wait_commit().
2879 */
2880 if (current->flags & PF_KTHREAD)
2881 goto out;
2882
2883 /*
2884 * If a fatal signal is pending, this process should not throttle.
2885 * It should return quickly so it can exit and free its memory
2886 */
2887 if (fatal_signal_pending(current))
2888 goto out;
2889
2890 /*
2891 * Check if the pfmemalloc reserves are ok by finding the first node
2892 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2893 * GFP_KERNEL will be required for allocating network buffers when
2894 * swapping over the network so ZONE_HIGHMEM is unusable.
2895 *
2896 * Throttling is based on the first usable node and throttled processes
2897 * wait on a queue until kswapd makes progress and wakes them. There
2898 * is an affinity then between processes waking up and where reclaim
2899 * progress has been made assuming the process wakes on the same node.
2900 * More importantly, processes running on remote nodes will not compete
2901 * for remote pfmemalloc reserves and processes on different nodes
2902 * should make reasonable progress.
2903 */
2904 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2905 gfp_zone(gfp_mask), nodemask) {
2906 if (zone_idx(zone) > ZONE_NORMAL)
2907 continue;
2908
2909 /* Throttle based on the first usable node */
2910 pgdat = zone->zone_pgdat;
2911 if (pfmemalloc_watermark_ok(pgdat))
2912 goto out;
2913 break;
2914 }
2915
2916 /* If no zone was usable by the allocation flags then do not throttle */
2917 if (!pgdat)
2918 goto out;
2919
2920 /* Account for the throttling */
2921 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2922
2923 /*
2924 * If the caller cannot enter the filesystem, it's possible that it
2925 * is due to the caller holding an FS lock or performing a journal
2926 * transaction in the case of a filesystem like ext[3|4]. In this case,
2927 * it is not safe to block on pfmemalloc_wait as kswapd could be
2928 * blocked waiting on the same lock. Instead, throttle for up to a
2929 * second before continuing.
2930 */
2931 if (!(gfp_mask & __GFP_FS)) {
2932 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2933 pfmemalloc_watermark_ok(pgdat), HZ);
2934
2935 goto check_pending;
2936 }
2937
2938 /* Throttle until kswapd wakes the process */
2939 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2940 pfmemalloc_watermark_ok(pgdat));
2941
2942check_pending:
2943 if (fatal_signal_pending(current))
2944 return true;
2945
2946out:
2947 return false;
2948}
2949
2950unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2951 gfp_t gfp_mask, nodemask_t *nodemask)
2952{
2953 unsigned long nr_reclaimed;
2954 struct scan_control sc = {
2955 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2956 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2957 .reclaim_idx = gfp_zone(gfp_mask),
2958 .order = order,
2959 .nodemask = nodemask,
2960 .priority = DEF_PRIORITY,
2961 .may_writepage = !laptop_mode,
2962 .may_unmap = 1,
2963 .may_swap = 1,
2964 };
2965
2966 /*
2967 * Do not enter reclaim if fatal signal was delivered while throttled.
2968 * 1 is returned so that the page allocator does not OOM kill at this
2969 * point.
2970 */
2971 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2972 return 1;
2973
2974 trace_mm_vmscan_direct_reclaim_begin(order,
2975 sc.may_writepage,
2976 gfp_mask,
2977 sc.reclaim_idx);
2978
2979 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2980
2981 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2982
2983 return nr_reclaimed;
2984}
2985
2986#ifdef CONFIG_MEMCG
2987
2988unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2989 gfp_t gfp_mask, bool noswap,
2990 pg_data_t *pgdat,
2991 unsigned long *nr_scanned)
2992{
2993 struct scan_control sc = {
2994 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2995 .target_mem_cgroup = memcg,
2996 .may_writepage = !laptop_mode,
2997 .may_unmap = 1,
2998 .reclaim_idx = MAX_NR_ZONES - 1,
2999 .may_swap = !noswap,
3000 };
3001 unsigned long lru_pages;
3002
3003 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3004 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3005
3006 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3007 sc.may_writepage,
3008 sc.gfp_mask,
3009 sc.reclaim_idx);
3010
3011 /*
3012 * NOTE: Although we can get the priority field, using it
3013 * here is not a good idea, since it limits the pages we can scan.
3014 * if we don't reclaim here, the shrink_node from balance_pgdat
3015 * will pick up pages from other mem cgroup's as well. We hack
3016 * the priority and make it zero.
3017 */
3018 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3019
3020 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3021
3022 *nr_scanned = sc.nr_scanned;
3023 return sc.nr_reclaimed;
3024}
3025
3026unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3027 unsigned long nr_pages,
3028 gfp_t gfp_mask,
3029 bool may_swap)
3030{
3031 struct zonelist *zonelist;
3032 unsigned long nr_reclaimed;
3033 int nid;
3034 struct scan_control sc = {
3035 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3036 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3037 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3038 .reclaim_idx = MAX_NR_ZONES - 1,
3039 .target_mem_cgroup = memcg,
3040 .priority = DEF_PRIORITY,
3041 .may_writepage = !laptop_mode,
3042 .may_unmap = 1,
3043 .may_swap = may_swap,
3044 };
3045
3046 /*
3047 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3048 * take care of from where we get pages. So the node where we start the
3049 * scan does not need to be the current node.
3050 */
3051 nid = mem_cgroup_select_victim_node(memcg);
3052
3053 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3054
3055 trace_mm_vmscan_memcg_reclaim_begin(0,
3056 sc.may_writepage,
3057 sc.gfp_mask,
3058 sc.reclaim_idx);
3059
3060 current->flags |= PF_MEMALLOC;
3061 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3062 current->flags &= ~PF_MEMALLOC;
3063
3064 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3065
3066 return nr_reclaimed;
3067}
3068#endif
3069
3070static void age_active_anon(struct pglist_data *pgdat,
3071 struct scan_control *sc)
3072{
3073 struct mem_cgroup *memcg;
3074
3075 if (!total_swap_pages)
3076 return;
3077
3078 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3079 do {
3080 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3081
3082 if (inactive_list_is_low(lruvec, false, sc))
3083 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3084 sc, LRU_ACTIVE_ANON);
3085
3086 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3087 } while (memcg);
3088}
3089
3090static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3091{
3092 unsigned long mark = high_wmark_pages(zone);
3093
3094 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3095 return false;
3096
3097 /*
3098 * If any eligible zone is balanced then the node is not considered
3099 * to be congested or dirty
3100 */
3101 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3102 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3103
3104 return true;
3105}
3106
3107/*
3108 * Prepare kswapd for sleeping. This verifies that there are no processes
3109 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3110 *
3111 * Returns true if kswapd is ready to sleep
3112 */
3113static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3114{
3115 int i;
3116
3117 /*
3118 * The throttled processes are normally woken up in balance_pgdat() as
3119 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3120 * race between when kswapd checks the watermarks and a process gets
3121 * throttled. There is also a potential race if processes get
3122 * throttled, kswapd wakes, a large process exits thereby balancing the
3123 * zones, which causes kswapd to exit balance_pgdat() before reaching
3124 * the wake up checks. If kswapd is going to sleep, no process should
3125 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3126 * the wake up is premature, processes will wake kswapd and get
3127 * throttled again. The difference from wake ups in balance_pgdat() is
3128 * that here we are under prepare_to_wait().
3129 */
3130 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3131 wake_up_all(&pgdat->pfmemalloc_wait);
3132
3133 for (i = 0; i <= classzone_idx; i++) {
3134 struct zone *zone = pgdat->node_zones + i;
3135
3136 if (!managed_zone(zone))
3137 continue;
3138
3139 if (!zone_balanced(zone, order, classzone_idx))
3140 return false;
3141 }
3142
3143 return true;
3144}
3145
3146/*
3147 * kswapd shrinks a node of pages that are at or below the highest usable
3148 * zone that is currently unbalanced.
3149 *
3150 * Returns true if kswapd scanned at least the requested number of pages to
3151 * reclaim or if the lack of progress was due to pages under writeback.
3152 * This is used to determine if the scanning priority needs to be raised.
3153 */
3154static bool kswapd_shrink_node(pg_data_t *pgdat,
3155 struct scan_control *sc)
3156{
3157 struct zone *zone;
3158 int z;
3159
3160 /* Reclaim a number of pages proportional to the number of zones */
3161 sc->nr_to_reclaim = 0;
3162 for (z = 0; z <= sc->reclaim_idx; z++) {
3163 zone = pgdat->node_zones + z;
3164 if (!managed_zone(zone))
3165 continue;
3166
3167 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3168 }
3169
3170 /*
3171 * Historically care was taken to put equal pressure on all zones but
3172 * now pressure is applied based on node LRU order.
3173 */
3174 shrink_node(pgdat, sc);
3175
3176 /*
3177 * Fragmentation may mean that the system cannot be rebalanced for
3178 * high-order allocations. If twice the allocation size has been
3179 * reclaimed then recheck watermarks only at order-0 to prevent
3180 * excessive reclaim. Assume that a process requested a high-order
3181 * can direct reclaim/compact.
3182 */
3183 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3184 sc->order = 0;
3185
3186 return sc->nr_scanned >= sc->nr_to_reclaim;
3187}
3188
3189/*
3190 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3191 * that are eligible for use by the caller until at least one zone is
3192 * balanced.
3193 *
3194 * Returns the order kswapd finished reclaiming at.
3195 *
3196 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3197 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3198 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3199 * or lower is eligible for reclaim until at least one usable zone is
3200 * balanced.
3201 */
3202static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3203{
3204 int i;
3205 unsigned long nr_soft_reclaimed;
3206 unsigned long nr_soft_scanned;
3207 struct zone *zone;
3208 struct scan_control sc = {
3209 .gfp_mask = GFP_KERNEL,
3210 .order = order,
3211 .priority = DEF_PRIORITY,
3212 .may_writepage = !laptop_mode,
3213 .may_unmap = 1,
3214 .may_swap = 1,
3215 };
3216 count_vm_event(PAGEOUTRUN);
3217
3218 do {
3219 bool raise_priority = true;
3220
3221 sc.nr_reclaimed = 0;
3222 sc.reclaim_idx = classzone_idx;
3223
3224 /*
3225 * If the number of buffer_heads exceeds the maximum allowed
3226 * then consider reclaiming from all zones. This has a dual
3227 * purpose -- on 64-bit systems it is expected that
3228 * buffer_heads are stripped during active rotation. On 32-bit
3229 * systems, highmem pages can pin lowmem memory and shrinking
3230 * buffers can relieve lowmem pressure. Reclaim may still not
3231 * go ahead if all eligible zones for the original allocation
3232 * request are balanced to avoid excessive reclaim from kswapd.
3233 */
3234 if (buffer_heads_over_limit) {
3235 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3236 zone = pgdat->node_zones + i;
3237 if (!managed_zone(zone))
3238 continue;
3239
3240 sc.reclaim_idx = i;
3241 break;
3242 }
3243 }
3244
3245 /*
3246 * Only reclaim if there are no eligible zones. Check from
3247 * high to low zone as allocations prefer higher zones.
3248 * Scanning from low to high zone would allow congestion to be
3249 * cleared during a very small window when a small low
3250 * zone was balanced even under extreme pressure when the
3251 * overall node may be congested. Note that sc.reclaim_idx
3252 * is not used as buffer_heads_over_limit may have adjusted
3253 * it.
3254 */
3255 for (i = classzone_idx; i >= 0; i--) {
3256 zone = pgdat->node_zones + i;
3257 if (!managed_zone(zone))
3258 continue;
3259
3260 if (zone_balanced(zone, sc.order, classzone_idx))
3261 goto out;
3262 }
3263
3264 /*
3265 * Do some background aging of the anon list, to give
3266 * pages a chance to be referenced before reclaiming. All
3267 * pages are rotated regardless of classzone as this is
3268 * about consistent aging.
3269 */
3270 age_active_anon(pgdat, &sc);
3271
3272 /*
3273 * If we're getting trouble reclaiming, start doing writepage
3274 * even in laptop mode.
3275 */
3276 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3277 sc.may_writepage = 1;
3278
3279 /* Call soft limit reclaim before calling shrink_node. */
3280 sc.nr_scanned = 0;
3281 nr_soft_scanned = 0;
3282 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3283 sc.gfp_mask, &nr_soft_scanned);
3284 sc.nr_reclaimed += nr_soft_reclaimed;
3285
3286 /*
3287 * There should be no need to raise the scanning priority if
3288 * enough pages are already being scanned that that high
3289 * watermark would be met at 100% efficiency.
3290 */
3291 if (kswapd_shrink_node(pgdat, &sc))
3292 raise_priority = false;
3293
3294 /*
3295 * If the low watermark is met there is no need for processes
3296 * to be throttled on pfmemalloc_wait as they should not be
3297 * able to safely make forward progress. Wake them
3298 */
3299 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3300 pfmemalloc_watermark_ok(pgdat))
3301 wake_up_all(&pgdat->pfmemalloc_wait);
3302
3303 /* Check if kswapd should be suspending */
3304 if (try_to_freeze() || kthread_should_stop())
3305 break;
3306
3307 /*
3308 * Raise priority if scanning rate is too low or there was no
3309 * progress in reclaiming pages
3310 */
3311 if (raise_priority || !sc.nr_reclaimed)
3312 sc.priority--;
3313 } while (sc.priority >= 1);
3314
3315out:
3316 /*
3317 * Return the order kswapd stopped reclaiming at as
3318 * prepare_kswapd_sleep() takes it into account. If another caller
3319 * entered the allocator slow path while kswapd was awake, order will
3320 * remain at the higher level.
3321 */
3322 return sc.order;
3323}
3324
3325static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3326 unsigned int classzone_idx)
3327{
3328 long remaining = 0;
3329 DEFINE_WAIT(wait);
3330
3331 if (freezing(current) || kthread_should_stop())
3332 return;
3333
3334 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3335
3336 /* Try to sleep for a short interval */
3337 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3338 /*
3339 * Compaction records what page blocks it recently failed to
3340 * isolate pages from and skips them in the future scanning.
3341 * When kswapd is going to sleep, it is reasonable to assume
3342 * that pages and compaction may succeed so reset the cache.
3343 */
3344 reset_isolation_suitable(pgdat);
3345
3346 /*
3347 * We have freed the memory, now we should compact it to make
3348 * allocation of the requested order possible.
3349 */
3350 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3351
3352 remaining = schedule_timeout(HZ/10);
3353
3354 /*
3355 * If woken prematurely then reset kswapd_classzone_idx and
3356 * order. The values will either be from a wakeup request or
3357 * the previous request that slept prematurely.
3358 */
3359 if (remaining) {
3360 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3361 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3362 }
3363
3364 finish_wait(&pgdat->kswapd_wait, &wait);
3365 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3366 }
3367
3368 /*
3369 * After a short sleep, check if it was a premature sleep. If not, then
3370 * go fully to sleep until explicitly woken up.
3371 */
3372 if (!remaining &&
3373 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3374 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3375
3376 /*
3377 * vmstat counters are not perfectly accurate and the estimated
3378 * value for counters such as NR_FREE_PAGES can deviate from the
3379 * true value by nr_online_cpus * threshold. To avoid the zone
3380 * watermarks being breached while under pressure, we reduce the
3381 * per-cpu vmstat threshold while kswapd is awake and restore
3382 * them before going back to sleep.
3383 */
3384 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3385
3386 if (!kthread_should_stop())
3387 schedule();
3388
3389 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3390 } else {
3391 if (remaining)
3392 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3393 else
3394 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3395 }
3396 finish_wait(&pgdat->kswapd_wait, &wait);
3397}
3398
3399/*
3400 * The background pageout daemon, started as a kernel thread
3401 * from the init process.
3402 *
3403 * This basically trickles out pages so that we have _some_
3404 * free memory available even if there is no other activity
3405 * that frees anything up. This is needed for things like routing
3406 * etc, where we otherwise might have all activity going on in
3407 * asynchronous contexts that cannot page things out.
3408 *
3409 * If there are applications that are active memory-allocators
3410 * (most normal use), this basically shouldn't matter.
3411 */
3412static int kswapd(void *p)
3413{
3414 unsigned int alloc_order, reclaim_order, classzone_idx;
3415 pg_data_t *pgdat = (pg_data_t*)p;
3416 struct task_struct *tsk = current;
3417
3418 struct reclaim_state reclaim_state = {
3419 .reclaimed_slab = 0,
3420 };
3421 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3422
3423 lockdep_set_current_reclaim_state(GFP_KERNEL);
3424
3425 if (!cpumask_empty(cpumask))
3426 set_cpus_allowed_ptr(tsk, cpumask);
3427 current->reclaim_state = &reclaim_state;
3428
3429 /*
3430 * Tell the memory management that we're a "memory allocator",
3431 * and that if we need more memory we should get access to it
3432 * regardless (see "__alloc_pages()"). "kswapd" should
3433 * never get caught in the normal page freeing logic.
3434 *
3435 * (Kswapd normally doesn't need memory anyway, but sometimes
3436 * you need a small amount of memory in order to be able to
3437 * page out something else, and this flag essentially protects
3438 * us from recursively trying to free more memory as we're
3439 * trying to free the first piece of memory in the first place).
3440 */
3441 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3442 set_freezable();
3443
3444 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3445 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3446 for ( ; ; ) {
3447 bool ret;
3448
3449kswapd_try_sleep:
3450 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3451 classzone_idx);
3452
3453 /* Read the new order and classzone_idx */
3454 alloc_order = reclaim_order = pgdat->kswapd_order;
3455 classzone_idx = pgdat->kswapd_classzone_idx;
3456 pgdat->kswapd_order = 0;
3457 pgdat->kswapd_classzone_idx = 0;
3458
3459 ret = try_to_freeze();
3460 if (kthread_should_stop())
3461 break;
3462
3463 /*
3464 * We can speed up thawing tasks if we don't call balance_pgdat
3465 * after returning from the refrigerator
3466 */
3467 if (ret)
3468 continue;
3469
3470 /*
3471 * Reclaim begins at the requested order but if a high-order
3472 * reclaim fails then kswapd falls back to reclaiming for
3473 * order-0. If that happens, kswapd will consider sleeping
3474 * for the order it finished reclaiming at (reclaim_order)
3475 * but kcompactd is woken to compact for the original
3476 * request (alloc_order).
3477 */
3478 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3479 alloc_order);
3480 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3481 if (reclaim_order < alloc_order)
3482 goto kswapd_try_sleep;
3483
3484 alloc_order = reclaim_order = pgdat->kswapd_order;
3485 classzone_idx = pgdat->kswapd_classzone_idx;
3486 }
3487
3488 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3489 current->reclaim_state = NULL;
3490 lockdep_clear_current_reclaim_state();
3491
3492 return 0;
3493}
3494
3495/*
3496 * A zone is low on free memory, so wake its kswapd task to service it.
3497 */
3498void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3499{
3500 pg_data_t *pgdat;
3501 int z;
3502
3503 if (!managed_zone(zone))
3504 return;
3505
3506 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3507 return;
3508 pgdat = zone->zone_pgdat;
3509 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3510 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3511 if (!waitqueue_active(&pgdat->kswapd_wait))
3512 return;
3513
3514 /* Only wake kswapd if all zones are unbalanced */
3515 for (z = 0; z <= classzone_idx; z++) {
3516 zone = pgdat->node_zones + z;
3517 if (!managed_zone(zone))
3518 continue;
3519
3520 if (zone_balanced(zone, order, classzone_idx))
3521 return;
3522 }
3523
3524 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3525 wake_up_interruptible(&pgdat->kswapd_wait);
3526}
3527
3528#ifdef CONFIG_HIBERNATION
3529/*
3530 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3531 * freed pages.
3532 *
3533 * Rather than trying to age LRUs the aim is to preserve the overall
3534 * LRU order by reclaiming preferentially
3535 * inactive > active > active referenced > active mapped
3536 */
3537unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3538{
3539 struct reclaim_state reclaim_state;
3540 struct scan_control sc = {
3541 .nr_to_reclaim = nr_to_reclaim,
3542 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3543 .reclaim_idx = MAX_NR_ZONES - 1,
3544 .priority = DEF_PRIORITY,
3545 .may_writepage = 1,
3546 .may_unmap = 1,
3547 .may_swap = 1,
3548 .hibernation_mode = 1,
3549 };
3550 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3551 struct task_struct *p = current;
3552 unsigned long nr_reclaimed;
3553
3554 p->flags |= PF_MEMALLOC;
3555 lockdep_set_current_reclaim_state(sc.gfp_mask);
3556 reclaim_state.reclaimed_slab = 0;
3557 p->reclaim_state = &reclaim_state;
3558
3559 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3560
3561 p->reclaim_state = NULL;
3562 lockdep_clear_current_reclaim_state();
3563 p->flags &= ~PF_MEMALLOC;
3564
3565 return nr_reclaimed;
3566}
3567#endif /* CONFIG_HIBERNATION */
3568
3569/* It's optimal to keep kswapds on the same CPUs as their memory, but
3570 not required for correctness. So if the last cpu in a node goes
3571 away, we get changed to run anywhere: as the first one comes back,
3572 restore their cpu bindings. */
3573static int kswapd_cpu_online(unsigned int cpu)
3574{
3575 int nid;
3576
3577 for_each_node_state(nid, N_MEMORY) {
3578 pg_data_t *pgdat = NODE_DATA(nid);
3579 const struct cpumask *mask;
3580
3581 mask = cpumask_of_node(pgdat->node_id);
3582
3583 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3584 /* One of our CPUs online: restore mask */
3585 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3586 }
3587 return 0;
3588}
3589
3590/*
3591 * This kswapd start function will be called by init and node-hot-add.
3592 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3593 */
3594int kswapd_run(int nid)
3595{
3596 pg_data_t *pgdat = NODE_DATA(nid);
3597 int ret = 0;
3598
3599 if (pgdat->kswapd)
3600 return 0;
3601
3602 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3603 if (IS_ERR(pgdat->kswapd)) {
3604 /* failure at boot is fatal */
3605 BUG_ON(system_state == SYSTEM_BOOTING);
3606 pr_err("Failed to start kswapd on node %d\n", nid);
3607 ret = PTR_ERR(pgdat->kswapd);
3608 pgdat->kswapd = NULL;
3609 }
3610 return ret;
3611}
3612
3613/*
3614 * Called by memory hotplug when all memory in a node is offlined. Caller must
3615 * hold mem_hotplug_begin/end().
3616 */
3617void kswapd_stop(int nid)
3618{
3619 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3620
3621 if (kswapd) {
3622 kthread_stop(kswapd);
3623 NODE_DATA(nid)->kswapd = NULL;
3624 }
3625}
3626
3627static int __init kswapd_init(void)
3628{
3629 int nid, ret;
3630
3631 swap_setup();
3632 for_each_node_state(nid, N_MEMORY)
3633 kswapd_run(nid);
3634 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3635 "mm/vmscan:online", kswapd_cpu_online,
3636 NULL);
3637 WARN_ON(ret < 0);
3638 return 0;
3639}
3640
3641module_init(kswapd_init)
3642
3643#ifdef CONFIG_NUMA
3644/*
3645 * Node reclaim mode
3646 *
3647 * If non-zero call node_reclaim when the number of free pages falls below
3648 * the watermarks.
3649 */
3650int node_reclaim_mode __read_mostly;
3651
3652#define RECLAIM_OFF 0
3653#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3654#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3655#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3656
3657/*
3658 * Priority for NODE_RECLAIM. This determines the fraction of pages
3659 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3660 * a zone.
3661 */
3662#define NODE_RECLAIM_PRIORITY 4
3663
3664/*
3665 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3666 * occur.
3667 */
3668int sysctl_min_unmapped_ratio = 1;
3669
3670/*
3671 * If the number of slab pages in a zone grows beyond this percentage then
3672 * slab reclaim needs to occur.
3673 */
3674int sysctl_min_slab_ratio = 5;
3675
3676static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3677{
3678 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3679 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3680 node_page_state(pgdat, NR_ACTIVE_FILE);
3681
3682 /*
3683 * It's possible for there to be more file mapped pages than
3684 * accounted for by the pages on the file LRU lists because
3685 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3686 */
3687 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3688}
3689
3690/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3691static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3692{
3693 unsigned long nr_pagecache_reclaimable;
3694 unsigned long delta = 0;
3695
3696 /*
3697 * If RECLAIM_UNMAP is set, then all file pages are considered
3698 * potentially reclaimable. Otherwise, we have to worry about
3699 * pages like swapcache and node_unmapped_file_pages() provides
3700 * a better estimate
3701 */
3702 if (node_reclaim_mode & RECLAIM_UNMAP)
3703 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3704 else
3705 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3706
3707 /* If we can't clean pages, remove dirty pages from consideration */
3708 if (!(node_reclaim_mode & RECLAIM_WRITE))
3709 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3710
3711 /* Watch for any possible underflows due to delta */
3712 if (unlikely(delta > nr_pagecache_reclaimable))
3713 delta = nr_pagecache_reclaimable;
3714
3715 return nr_pagecache_reclaimable - delta;
3716}
3717
3718/*
3719 * Try to free up some pages from this node through reclaim.
3720 */
3721static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3722{
3723 /* Minimum pages needed in order to stay on node */
3724 const unsigned long nr_pages = 1 << order;
3725 struct task_struct *p = current;
3726 struct reclaim_state reclaim_state;
3727 int classzone_idx = gfp_zone(gfp_mask);
3728 struct scan_control sc = {
3729 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3730 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3731 .order = order,
3732 .priority = NODE_RECLAIM_PRIORITY,
3733 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3734 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3735 .may_swap = 1,
3736 .reclaim_idx = classzone_idx,
3737 };
3738
3739 cond_resched();
3740 /*
3741 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3742 * and we also need to be able to write out pages for RECLAIM_WRITE
3743 * and RECLAIM_UNMAP.
3744 */
3745 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3746 lockdep_set_current_reclaim_state(gfp_mask);
3747 reclaim_state.reclaimed_slab = 0;
3748 p->reclaim_state = &reclaim_state;
3749
3750 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3751 /*
3752 * Free memory by calling shrink zone with increasing
3753 * priorities until we have enough memory freed.
3754 */
3755 do {
3756 shrink_node(pgdat, &sc);
3757 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3758 }
3759
3760 p->reclaim_state = NULL;
3761 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3762 lockdep_clear_current_reclaim_state();
3763 return sc.nr_reclaimed >= nr_pages;
3764}
3765
3766int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3767{
3768 int ret;
3769
3770 /*
3771 * Node reclaim reclaims unmapped file backed pages and
3772 * slab pages if we are over the defined limits.
3773 *
3774 * A small portion of unmapped file backed pages is needed for
3775 * file I/O otherwise pages read by file I/O will be immediately
3776 * thrown out if the node is overallocated. So we do not reclaim
3777 * if less than a specified percentage of the node is used by
3778 * unmapped file backed pages.
3779 */
3780 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3781 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3782 return NODE_RECLAIM_FULL;
3783
3784 if (!pgdat_reclaimable(pgdat))
3785 return NODE_RECLAIM_FULL;
3786
3787 /*
3788 * Do not scan if the allocation should not be delayed.
3789 */
3790 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3791 return NODE_RECLAIM_NOSCAN;
3792
3793 /*
3794 * Only run node reclaim on the local node or on nodes that do not
3795 * have associated processors. This will favor the local processor
3796 * over remote processors and spread off node memory allocations
3797 * as wide as possible.
3798 */
3799 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3800 return NODE_RECLAIM_NOSCAN;
3801
3802 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3803 return NODE_RECLAIM_NOSCAN;
3804
3805 ret = __node_reclaim(pgdat, gfp_mask, order);
3806 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3807
3808 if (!ret)
3809 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3810
3811 return ret;
3812}
3813#endif
3814
3815/*
3816 * page_evictable - test whether a page is evictable
3817 * @page: the page to test
3818 *
3819 * Test whether page is evictable--i.e., should be placed on active/inactive
3820 * lists vs unevictable list.
3821 *
3822 * Reasons page might not be evictable:
3823 * (1) page's mapping marked unevictable
3824 * (2) page is part of an mlocked VMA
3825 *
3826 */
3827int page_evictable(struct page *page)
3828{
3829 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3830}
3831
3832#ifdef CONFIG_SHMEM
3833/**
3834 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3835 * @pages: array of pages to check
3836 * @nr_pages: number of pages to check
3837 *
3838 * Checks pages for evictability and moves them to the appropriate lru list.
3839 *
3840 * This function is only used for SysV IPC SHM_UNLOCK.
3841 */
3842void check_move_unevictable_pages(struct page **pages, int nr_pages)
3843{
3844 struct lruvec *lruvec;
3845 struct pglist_data *pgdat = NULL;
3846 int pgscanned = 0;
3847 int pgrescued = 0;
3848 int i;
3849
3850 for (i = 0; i < nr_pages; i++) {
3851 struct page *page = pages[i];
3852 struct pglist_data *pagepgdat = page_pgdat(page);
3853
3854 pgscanned++;
3855 if (pagepgdat != pgdat) {
3856 if (pgdat)
3857 spin_unlock_irq(&pgdat->lru_lock);
3858 pgdat = pagepgdat;
3859 spin_lock_irq(&pgdat->lru_lock);
3860 }
3861 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3862
3863 if (!PageLRU(page) || !PageUnevictable(page))
3864 continue;
3865
3866 if (page_evictable(page)) {
3867 enum lru_list lru = page_lru_base_type(page);
3868
3869 VM_BUG_ON_PAGE(PageActive(page), page);
3870 ClearPageUnevictable(page);
3871 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3872 add_page_to_lru_list(page, lruvec, lru);
3873 pgrescued++;
3874 }
3875 }
3876
3877 if (pgdat) {
3878 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3879 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3880 spin_unlock_irq(&pgdat->lru_lock);
3881 }
3882}
3883#endif /* CONFIG_SHMEM */