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 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
98 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
108};
109
110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
151#ifdef CONFIG_MEMCG
152static bool global_reclaim(struct scan_control *sc)
153{
154 return !sc->target_mem_cgroup;
155}
156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
178 return true;
179#endif
180 return false;
181}
182#else
183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
192#endif
193
194static unsigned long zone_reclaimable_pages(struct zone *zone)
195{
196 unsigned long nr;
197
198 nr = zone_page_state_snapshot(zone, NR_ACTIVE_FILE) +
199 zone_page_state_snapshot(zone, NR_INACTIVE_FILE) +
200 zone_page_state_snapshot(zone, NR_ISOLATED_FILE);
201
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state_snapshot(zone, NR_ACTIVE_ANON) +
204 zone_page_state_snapshot(zone, NR_INACTIVE_ANON) +
205 zone_page_state_snapshot(zone, NR_ISOLATED_ANON);
206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
212 return zone_page_state_snapshot(zone, NR_PAGES_SCANNED) <
213 zone_reclaimable_pages(zone) * 6;
214}
215
216unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
217{
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec, lru);
220
221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
222}
223
224/*
225 * Add a shrinker callback to be called from the vm.
226 */
227int register_shrinker(struct shrinker *shrinker)
228{
229 size_t size = sizeof(*shrinker->nr_deferred);
230
231 if (shrinker->flags & SHRINKER_NUMA_AWARE)
232 size *= nr_node_ids;
233
234 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
235 if (!shrinker->nr_deferred)
236 return -ENOMEM;
237
238 down_write(&shrinker_rwsem);
239 list_add_tail(&shrinker->list, &shrinker_list);
240 up_write(&shrinker_rwsem);
241 return 0;
242}
243EXPORT_SYMBOL(register_shrinker);
244
245/*
246 * Remove one
247 */
248void unregister_shrinker(struct shrinker *shrinker)
249{
250 down_write(&shrinker_rwsem);
251 list_del(&shrinker->list);
252 up_write(&shrinker_rwsem);
253 kfree(shrinker->nr_deferred);
254}
255EXPORT_SYMBOL(unregister_shrinker);
256
257#define SHRINK_BATCH 128
258
259static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
260 struct shrinker *shrinker,
261 unsigned long nr_scanned,
262 unsigned long nr_eligible)
263{
264 unsigned long freed = 0;
265 unsigned long long delta;
266 long total_scan;
267 long freeable;
268 long nr;
269 long new_nr;
270 int nid = shrinkctl->nid;
271 long batch_size = shrinker->batch ? shrinker->batch
272 : SHRINK_BATCH;
273
274 freeable = shrinker->count_objects(shrinker, shrinkctl);
275 if (freeable == 0)
276 return 0;
277
278 /*
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
282 */
283 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
284
285 total_scan = nr;
286 delta = (4 * nr_scanned) / shrinker->seeks;
287 delta *= freeable;
288 do_div(delta, nr_eligible + 1);
289 total_scan += delta;
290 if (total_scan < 0) {
291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
292 shrinker->scan_objects, total_scan);
293 total_scan = freeable;
294 }
295
296 /*
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
302 * freeable. This is bad for sustaining a working set in
303 * memory.
304 *
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
307 */
308 if (delta < freeable / 4)
309 total_scan = min(total_scan, freeable / 2);
310
311 /*
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
314 * freeable entries.
315 */
316 if (total_scan > freeable * 2)
317 total_scan = freeable * 2;
318
319 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
320 nr_scanned, nr_eligible,
321 freeable, delta, total_scan);
322
323 /*
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
330 * batch_size.
331 *
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
334 * than the total number of objects on slab (freeable), we must be
335 * scanning at high prio and therefore should try to reclaim as much as
336 * possible.
337 */
338 while (total_scan >= batch_size ||
339 total_scan >= freeable) {
340 unsigned long ret;
341 unsigned long nr_to_scan = min(batch_size, total_scan);
342
343 shrinkctl->nr_to_scan = nr_to_scan;
344 ret = shrinker->scan_objects(shrinker, shrinkctl);
345 if (ret == SHRINK_STOP)
346 break;
347 freed += ret;
348
349 count_vm_events(SLABS_SCANNED, nr_to_scan);
350 total_scan -= nr_to_scan;
351
352 cond_resched();
353 }
354
355 /*
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
359 */
360 if (total_scan > 0)
361 new_nr = atomic_long_add_return(total_scan,
362 &shrinker->nr_deferred[nid]);
363 else
364 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
365
366 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
367 return freed;
368}
369
370/**
371 * shrink_slab - shrink slab caches
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
374 * @memcg: memory cgroup whose slab caches to target
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
377 *
378 * Call the shrink functions to age shrinkable caches.
379 *
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
382 *
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
387 *
388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
395 *
396 * Returns the number of reclaimed slab objects.
397 */
398static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
399 struct mem_cgroup *memcg,
400 unsigned long nr_scanned,
401 unsigned long nr_eligible)
402{
403 struct shrinker *shrinker;
404 unsigned long freed = 0;
405
406 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
407 return 0;
408
409 if (nr_scanned == 0)
410 nr_scanned = SWAP_CLUSTER_MAX;
411
412 if (!down_read_trylock(&shrinker_rwsem)) {
413 /*
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
417 * time.
418 */
419 freed = 1;
420 goto out;
421 }
422
423 list_for_each_entry(shrinker, &shrinker_list, list) {
424 struct shrink_control sc = {
425 .gfp_mask = gfp_mask,
426 .nid = nid,
427 .memcg = memcg,
428 };
429
430 /*
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
434 */
435 if (memcg_kmem_enabled() &&
436 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
437 continue;
438
439 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
440 sc.nid = 0;
441
442 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
443 }
444
445 up_read(&shrinker_rwsem);
446out:
447 cond_resched();
448 return freed;
449}
450
451void drop_slab_node(int nid)
452{
453 unsigned long freed;
454
455 do {
456 struct mem_cgroup *memcg = NULL;
457
458 freed = 0;
459 do {
460 freed += shrink_slab(GFP_KERNEL, nid, memcg,
461 1000, 1000);
462 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
463 } while (freed > 10);
464}
465
466void drop_slab(void)
467{
468 int nid;
469
470 for_each_online_node(nid)
471 drop_slab_node(nid);
472}
473
474static inline int is_page_cache_freeable(struct page *page)
475{
476 /*
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
480 */
481 return page_count(page) - page_has_private(page) == 2;
482}
483
484static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
485{
486 if (current->flags & PF_SWAPWRITE)
487 return 1;
488 if (!inode_write_congested(inode))
489 return 1;
490 if (inode_to_bdi(inode) == current->backing_dev_info)
491 return 1;
492 return 0;
493}
494
495/*
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
499 *
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
503 *
504 * We're allowed to run sleeping lock_page() here because we know the caller has
505 * __GFP_FS.
506 */
507static void handle_write_error(struct address_space *mapping,
508 struct page *page, int error)
509{
510 lock_page(page);
511 if (page_mapping(page) == mapping)
512 mapping_set_error(mapping, error);
513 unlock_page(page);
514}
515
516/* possible outcome of pageout() */
517typedef enum {
518 /* failed to write page out, page is locked */
519 PAGE_KEEP,
520 /* move page to the active list, page is locked */
521 PAGE_ACTIVATE,
522 /* page has been sent to the disk successfully, page is unlocked */
523 PAGE_SUCCESS,
524 /* page is clean and locked */
525 PAGE_CLEAN,
526} pageout_t;
527
528/*
529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
531 */
532static pageout_t pageout(struct page *page, struct address_space *mapping,
533 struct scan_control *sc)
534{
535 /*
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
541 *
542 * If this process is currently in __generic_file_write_iter() against
543 * this page's queue, we can perform writeback even if that
544 * will block.
545 *
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
550 */
551 if (!is_page_cache_freeable(page))
552 return PAGE_KEEP;
553 if (!mapping) {
554 /*
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
557 */
558 if (page_has_private(page)) {
559 if (try_to_free_buffers(page)) {
560 ClearPageDirty(page);
561 pr_info("%s: orphaned page\n", __func__);
562 return PAGE_CLEAN;
563 }
564 }
565 return PAGE_KEEP;
566 }
567 if (mapping->a_ops->writepage == NULL)
568 return PAGE_ACTIVATE;
569 if (!may_write_to_inode(mapping->host, sc))
570 return PAGE_KEEP;
571
572 if (clear_page_dirty_for_io(page)) {
573 int res;
574 struct writeback_control wbc = {
575 .sync_mode = WB_SYNC_NONE,
576 .nr_to_write = SWAP_CLUSTER_MAX,
577 .range_start = 0,
578 .range_end = LLONG_MAX,
579 .for_reclaim = 1,
580 };
581
582 SetPageReclaim(page);
583 res = mapping->a_ops->writepage(page, &wbc);
584 if (res < 0)
585 handle_write_error(mapping, page, res);
586 if (res == AOP_WRITEPAGE_ACTIVATE) {
587 ClearPageReclaim(page);
588 return PAGE_ACTIVATE;
589 }
590
591 if (!PageWriteback(page)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page);
594 }
595 trace_mm_vmscan_writepage(page);
596 inc_zone_page_state(page, NR_VMSCAN_WRITE);
597 return PAGE_SUCCESS;
598 }
599
600 return PAGE_CLEAN;
601}
602
603/*
604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
606 */
607static int __remove_mapping(struct address_space *mapping, struct page *page,
608 bool reclaimed)
609{
610 unsigned long flags;
611
612 BUG_ON(!PageLocked(page));
613 BUG_ON(mapping != page_mapping(page));
614
615 spin_lock_irqsave(&mapping->tree_lock, flags);
616 /*
617 * The non racy check for a busy page.
618 *
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
623 *
624 * get_user_pages(&page);
625 * [user mapping goes away]
626 * write_to(page);
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
629 * put_page(page);
630 * !page_count(page) [good, discard it]
631 *
632 * [oops, our write_to data is lost]
633 *
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
636 * load is not satisfied before that of page->_count.
637 *
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
640 */
641 if (!page_ref_freeze(page, 2))
642 goto cannot_free;
643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page))) {
645 page_ref_unfreeze(page, 2);
646 goto cannot_free;
647 }
648
649 if (PageSwapCache(page)) {
650 swp_entry_t swap = { .val = page_private(page) };
651 mem_cgroup_swapout(page, swap);
652 __delete_from_swap_cache(page);
653 spin_unlock_irqrestore(&mapping->tree_lock, flags);
654 swapcache_free(swap);
655 } else {
656 void (*freepage)(struct page *);
657 void *shadow = NULL;
658
659 freepage = mapping->a_ops->freepage;
660 /*
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
663 *
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
668 * back.
669 *
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
674 * same page_tree.
675 */
676 if (reclaimed && page_is_file_cache(page) &&
677 !mapping_exiting(mapping) && !dax_mapping(mapping))
678 shadow = workingset_eviction(mapping, page);
679 __delete_from_page_cache(page, shadow);
680 spin_unlock_irqrestore(&mapping->tree_lock, flags);
681
682 if (freepage != NULL)
683 freepage(page);
684 }
685
686 return 1;
687
688cannot_free:
689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 return 0;
691}
692
693/*
694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
697 * this page.
698 */
699int remove_mapping(struct address_space *mapping, struct page *page)
700{
701 if (__remove_mapping(mapping, page, false)) {
702 /*
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
705 * atomic operation.
706 */
707 page_ref_unfreeze(page, 1);
708 return 1;
709 }
710 return 0;
711}
712
713/**
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
716 *
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
719 *
720 * lru_lock must not be held, interrupts must be enabled.
721 */
722void putback_lru_page(struct page *page)
723{
724 bool is_unevictable;
725 int was_unevictable = PageUnevictable(page);
726
727 VM_BUG_ON_PAGE(PageLRU(page), page);
728
729redo:
730 ClearPageUnevictable(page);
731
732 if (page_evictable(page)) {
733 /*
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
738 */
739 is_unevictable = false;
740 lru_cache_add(page);
741 } else {
742 /*
743 * Put unevictable pages directly on zone's unevictable
744 * list.
745 */
746 is_unevictable = true;
747 add_page_to_unevictable_list(page);
748 /*
749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
752 * isolation/check_move_unevictable_pages,
753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
754 * the page back to the evictable list.
755 *
756 * The other side is TestClearPageMlocked() or shmem_lock().
757 */
758 smp_mb();
759 }
760
761 /*
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
765 */
766 if (is_unevictable && page_evictable(page)) {
767 if (!isolate_lru_page(page)) {
768 put_page(page);
769 goto redo;
770 }
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
774 */
775 }
776
777 if (was_unevictable && !is_unevictable)
778 count_vm_event(UNEVICTABLE_PGRESCUED);
779 else if (!was_unevictable && is_unevictable)
780 count_vm_event(UNEVICTABLE_PGCULLED);
781
782 put_page(page); /* drop ref from isolate */
783}
784
785enum page_references {
786 PAGEREF_RECLAIM,
787 PAGEREF_RECLAIM_CLEAN,
788 PAGEREF_KEEP,
789 PAGEREF_ACTIVATE,
790};
791
792static enum page_references page_check_references(struct page *page,
793 struct scan_control *sc)
794{
795 int referenced_ptes, referenced_page;
796 unsigned long vm_flags;
797
798 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
799 &vm_flags);
800 referenced_page = TestClearPageReferenced(page);
801
802 /*
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
805 */
806 if (vm_flags & VM_LOCKED)
807 return PAGEREF_RECLAIM;
808
809 if (referenced_ptes) {
810 if (PageSwapBacked(page))
811 return PAGEREF_ACTIVATE;
812 /*
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
816 * than once.
817 *
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
821 *
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
824 * quickly recovered.
825 */
826 SetPageReferenced(page);
827
828 if (referenced_page || referenced_ptes > 1)
829 return PAGEREF_ACTIVATE;
830
831 /*
832 * Activate file-backed executable pages after first usage.
833 */
834 if (vm_flags & VM_EXEC)
835 return PAGEREF_ACTIVATE;
836
837 return PAGEREF_KEEP;
838 }
839
840 /* Reclaim if clean, defer dirty pages to writeback */
841 if (referenced_page && !PageSwapBacked(page))
842 return PAGEREF_RECLAIM_CLEAN;
843
844 return PAGEREF_RECLAIM;
845}
846
847/* Check if a page is dirty or under writeback */
848static void page_check_dirty_writeback(struct page *page,
849 bool *dirty, bool *writeback)
850{
851 struct address_space *mapping;
852
853 /*
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
856 */
857 if (!page_is_file_cache(page)) {
858 *dirty = false;
859 *writeback = false;
860 return;
861 }
862
863 /* By default assume that the page flags are accurate */
864 *dirty = PageDirty(page);
865 *writeback = PageWriteback(page);
866
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page))
869 return;
870
871 mapping = page_mapping(page);
872 if (mapping && mapping->a_ops->is_dirty_writeback)
873 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
874}
875
876/*
877 * shrink_page_list() returns the number of reclaimed pages
878 */
879static unsigned long shrink_page_list(struct list_head *page_list,
880 struct zone *zone,
881 struct scan_control *sc,
882 enum ttu_flags ttu_flags,
883 unsigned long *ret_nr_dirty,
884 unsigned long *ret_nr_unqueued_dirty,
885 unsigned long *ret_nr_congested,
886 unsigned long *ret_nr_writeback,
887 unsigned long *ret_nr_immediate,
888 bool force_reclaim)
889{
890 LIST_HEAD(ret_pages);
891 LIST_HEAD(free_pages);
892 int pgactivate = 0;
893 unsigned long nr_unqueued_dirty = 0;
894 unsigned long nr_dirty = 0;
895 unsigned long nr_congested = 0;
896 unsigned long nr_reclaimed = 0;
897 unsigned long nr_writeback = 0;
898 unsigned long nr_immediate = 0;
899
900 cond_resched();
901
902 while (!list_empty(page_list)) {
903 struct address_space *mapping;
904 struct page *page;
905 int may_enter_fs;
906 enum page_references references = PAGEREF_RECLAIM_CLEAN;
907 bool dirty, writeback;
908 bool lazyfree = false;
909 int ret = SWAP_SUCCESS;
910
911 cond_resched();
912
913 page = lru_to_page(page_list);
914 list_del(&page->lru);
915
916 if (!trylock_page(page))
917 goto keep;
918
919 VM_BUG_ON_PAGE(PageActive(page), page);
920 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
921
922 sc->nr_scanned++;
923
924 if (unlikely(!page_evictable(page)))
925 goto cull_mlocked;
926
927 if (!sc->may_unmap && page_mapped(page))
928 goto keep_locked;
929
930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page) || PageSwapCache(page))
932 sc->nr_scanned++;
933
934 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
935 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
936
937 /*
938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
942 */
943 page_check_dirty_writeback(page, &dirty, &writeback);
944 if (dirty || writeback)
945 nr_dirty++;
946
947 if (dirty && !writeback)
948 nr_unqueued_dirty++;
949
950 /*
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
955 */
956 mapping = page_mapping(page);
957 if (((dirty || writeback) && mapping &&
958 inode_write_congested(mapping->host)) ||
959 (writeback && PageReclaim(page)))
960 nr_congested++;
961
962 /*
963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
965 *
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
975 *
976 * 2) Global or new memcg reclaim encounters a page that is
977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
980 * reclaim and continue scanning.
981 *
982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
988 *
989 * 3) Legacy memcg encounters a page that is already marked
990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
994 */
995 if (PageWriteback(page)) {
996 /* Case 1 above */
997 if (current_is_kswapd() &&
998 PageReclaim(page) &&
999 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1000 nr_immediate++;
1001 goto keep_locked;
1002
1003 /* Case 2 above */
1004 } else if (sane_reclaim(sc) ||
1005 !PageReclaim(page) || !may_enter_fs) {
1006 /*
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1016 */
1017 SetPageReclaim(page);
1018 nr_writeback++;
1019 goto keep_locked;
1020
1021 /* Case 3 above */
1022 } else {
1023 unlock_page(page);
1024 wait_on_page_writeback(page);
1025 /* then go back and try same page again */
1026 list_add_tail(&page->lru, page_list);
1027 continue;
1028 }
1029 }
1030
1031 if (!force_reclaim)
1032 references = page_check_references(page, sc);
1033
1034 switch (references) {
1035 case PAGEREF_ACTIVATE:
1036 goto activate_locked;
1037 case PAGEREF_KEEP:
1038 goto keep_locked;
1039 case PAGEREF_RECLAIM:
1040 case PAGEREF_RECLAIM_CLEAN:
1041 ; /* try to reclaim the page below */
1042 }
1043
1044 /*
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1047 */
1048 if (PageAnon(page) && !PageSwapCache(page)) {
1049 if (!(sc->gfp_mask & __GFP_IO))
1050 goto keep_locked;
1051 if (!add_to_swap(page, page_list))
1052 goto activate_locked;
1053 lazyfree = true;
1054 may_enter_fs = 1;
1055
1056 /* Adding to swap updated mapping */
1057 mapping = page_mapping(page);
1058 }
1059
1060 /*
1061 * The page is mapped into the page tables of one or more
1062 * processes. Try to unmap it here.
1063 */
1064 if (page_mapped(page) && mapping) {
1065 switch (ret = try_to_unmap(page, lazyfree ?
1066 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1067 (ttu_flags | TTU_BATCH_FLUSH))) {
1068 case SWAP_FAIL:
1069 goto activate_locked;
1070 case SWAP_AGAIN:
1071 goto keep_locked;
1072 case SWAP_MLOCK:
1073 goto cull_mlocked;
1074 case SWAP_LZFREE:
1075 goto lazyfree;
1076 case SWAP_SUCCESS:
1077 ; /* try to free the page below */
1078 }
1079 }
1080
1081 if (PageDirty(page)) {
1082 /*
1083 * Only kswapd can writeback filesystem pages to
1084 * avoid risk of stack overflow but only writeback
1085 * if many dirty pages have been encountered.
1086 */
1087 if (page_is_file_cache(page) &&
1088 (!current_is_kswapd() ||
1089 !test_bit(ZONE_DIRTY, &zone->flags))) {
1090 /*
1091 * Immediately reclaim when written back.
1092 * Similar in principal to deactivate_page()
1093 * except we already have the page isolated
1094 * and know it's dirty
1095 */
1096 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1097 SetPageReclaim(page);
1098
1099 goto keep_locked;
1100 }
1101
1102 if (references == PAGEREF_RECLAIM_CLEAN)
1103 goto keep_locked;
1104 if (!may_enter_fs)
1105 goto keep_locked;
1106 if (!sc->may_writepage)
1107 goto keep_locked;
1108
1109 /*
1110 * Page is dirty. Flush the TLB if a writable entry
1111 * potentially exists to avoid CPU writes after IO
1112 * starts and then write it out here.
1113 */
1114 try_to_unmap_flush_dirty();
1115 switch (pageout(page, mapping, sc)) {
1116 case PAGE_KEEP:
1117 goto keep_locked;
1118 case PAGE_ACTIVATE:
1119 goto activate_locked;
1120 case PAGE_SUCCESS:
1121 if (PageWriteback(page))
1122 goto keep;
1123 if (PageDirty(page))
1124 goto keep;
1125
1126 /*
1127 * A synchronous write - probably a ramdisk. Go
1128 * ahead and try to reclaim the page.
1129 */
1130 if (!trylock_page(page))
1131 goto keep;
1132 if (PageDirty(page) || PageWriteback(page))
1133 goto keep_locked;
1134 mapping = page_mapping(page);
1135 case PAGE_CLEAN:
1136 ; /* try to free the page below */
1137 }
1138 }
1139
1140 /*
1141 * If the page has buffers, try to free the buffer mappings
1142 * associated with this page. If we succeed we try to free
1143 * the page as well.
1144 *
1145 * We do this even if the page is PageDirty().
1146 * try_to_release_page() does not perform I/O, but it is
1147 * possible for a page to have PageDirty set, but it is actually
1148 * clean (all its buffers are clean). This happens if the
1149 * buffers were written out directly, with submit_bh(). ext3
1150 * will do this, as well as the blockdev mapping.
1151 * try_to_release_page() will discover that cleanness and will
1152 * drop the buffers and mark the page clean - it can be freed.
1153 *
1154 * Rarely, pages can have buffers and no ->mapping. These are
1155 * the pages which were not successfully invalidated in
1156 * truncate_complete_page(). We try to drop those buffers here
1157 * and if that worked, and the page is no longer mapped into
1158 * process address space (page_count == 1) it can be freed.
1159 * Otherwise, leave the page on the LRU so it is swappable.
1160 */
1161 if (page_has_private(page)) {
1162 if (!try_to_release_page(page, sc->gfp_mask))
1163 goto activate_locked;
1164 if (!mapping && page_count(page) == 1) {
1165 unlock_page(page);
1166 if (put_page_testzero(page))
1167 goto free_it;
1168 else {
1169 /*
1170 * rare race with speculative reference.
1171 * the speculative reference will free
1172 * this page shortly, so we may
1173 * increment nr_reclaimed here (and
1174 * leave it off the LRU).
1175 */
1176 nr_reclaimed++;
1177 continue;
1178 }
1179 }
1180 }
1181
1182lazyfree:
1183 if (!mapping || !__remove_mapping(mapping, page, true))
1184 goto keep_locked;
1185
1186 /*
1187 * At this point, we have no other references and there is
1188 * no way to pick any more up (removed from LRU, removed
1189 * from pagecache). Can use non-atomic bitops now (and
1190 * we obviously don't have to worry about waking up a process
1191 * waiting on the page lock, because there are no references.
1192 */
1193 __ClearPageLocked(page);
1194free_it:
1195 if (ret == SWAP_LZFREE)
1196 count_vm_event(PGLAZYFREED);
1197
1198 nr_reclaimed++;
1199
1200 /*
1201 * Is there need to periodically free_page_list? It would
1202 * appear not as the counts should be low
1203 */
1204 list_add(&page->lru, &free_pages);
1205 continue;
1206
1207cull_mlocked:
1208 if (PageSwapCache(page))
1209 try_to_free_swap(page);
1210 unlock_page(page);
1211 list_add(&page->lru, &ret_pages);
1212 continue;
1213
1214activate_locked:
1215 /* Not a candidate for swapping, so reclaim swap space. */
1216 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1217 try_to_free_swap(page);
1218 VM_BUG_ON_PAGE(PageActive(page), page);
1219 SetPageActive(page);
1220 pgactivate++;
1221keep_locked:
1222 unlock_page(page);
1223keep:
1224 list_add(&page->lru, &ret_pages);
1225 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1226 }
1227
1228 mem_cgroup_uncharge_list(&free_pages);
1229 try_to_unmap_flush();
1230 free_hot_cold_page_list(&free_pages, true);
1231
1232 list_splice(&ret_pages, page_list);
1233 count_vm_events(PGACTIVATE, pgactivate);
1234
1235 *ret_nr_dirty += nr_dirty;
1236 *ret_nr_congested += nr_congested;
1237 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1238 *ret_nr_writeback += nr_writeback;
1239 *ret_nr_immediate += nr_immediate;
1240 return nr_reclaimed;
1241}
1242
1243unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1244 struct list_head *page_list)
1245{
1246 struct scan_control sc = {
1247 .gfp_mask = GFP_KERNEL,
1248 .priority = DEF_PRIORITY,
1249 .may_unmap = 1,
1250 };
1251 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1252 struct page *page, *next;
1253 LIST_HEAD(clean_pages);
1254
1255 list_for_each_entry_safe(page, next, page_list, lru) {
1256 if (page_is_file_cache(page) && !PageDirty(page) &&
1257 !isolated_balloon_page(page)) {
1258 ClearPageActive(page);
1259 list_move(&page->lru, &clean_pages);
1260 }
1261 }
1262
1263 ret = shrink_page_list(&clean_pages, zone, &sc,
1264 TTU_UNMAP|TTU_IGNORE_ACCESS,
1265 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1266 list_splice(&clean_pages, page_list);
1267 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1268 return ret;
1269}
1270
1271/*
1272 * Attempt to remove the specified page from its LRU. Only take this page
1273 * if it is of the appropriate PageActive status. Pages which are being
1274 * freed elsewhere are also ignored.
1275 *
1276 * page: page to consider
1277 * mode: one of the LRU isolation modes defined above
1278 *
1279 * returns 0 on success, -ve errno on failure.
1280 */
1281int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1282{
1283 int ret = -EINVAL;
1284
1285 /* Only take pages on the LRU. */
1286 if (!PageLRU(page))
1287 return ret;
1288
1289 /* Compaction should not handle unevictable pages but CMA can do so */
1290 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1291 return ret;
1292
1293 ret = -EBUSY;
1294
1295 /*
1296 * To minimise LRU disruption, the caller can indicate that it only
1297 * wants to isolate pages it will be able to operate on without
1298 * blocking - clean pages for the most part.
1299 *
1300 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1301 * is used by reclaim when it is cannot write to backing storage
1302 *
1303 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1304 * that it is possible to migrate without blocking
1305 */
1306 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1307 /* All the caller can do on PageWriteback is block */
1308 if (PageWriteback(page))
1309 return ret;
1310
1311 if (PageDirty(page)) {
1312 struct address_space *mapping;
1313
1314 /* ISOLATE_CLEAN means only clean pages */
1315 if (mode & ISOLATE_CLEAN)
1316 return ret;
1317
1318 /*
1319 * Only pages without mappings or that have a
1320 * ->migratepage callback are possible to migrate
1321 * without blocking
1322 */
1323 mapping = page_mapping(page);
1324 if (mapping && !mapping->a_ops->migratepage)
1325 return ret;
1326 }
1327 }
1328
1329 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 return ret;
1331
1332 if (likely(get_page_unless_zero(page))) {
1333 /*
1334 * Be careful not to clear PageLRU until after we're
1335 * sure the page is not being freed elsewhere -- the
1336 * page release code relies on it.
1337 */
1338 ClearPageLRU(page);
1339 ret = 0;
1340 }
1341
1342 return ret;
1343}
1344
1345/*
1346 * zone->lru_lock is heavily contended. Some of the functions that
1347 * shrink the lists perform better by taking out a batch of pages
1348 * and working on them outside the LRU lock.
1349 *
1350 * For pagecache intensive workloads, this function is the hottest
1351 * spot in the kernel (apart from copy_*_user functions).
1352 *
1353 * Appropriate locks must be held before calling this function.
1354 *
1355 * @nr_to_scan: The number of pages to look through on the list.
1356 * @lruvec: The LRU vector to pull pages from.
1357 * @dst: The temp list to put pages on to.
1358 * @nr_scanned: The number of pages that were scanned.
1359 * @sc: The scan_control struct for this reclaim session
1360 * @mode: One of the LRU isolation modes
1361 * @lru: LRU list id for isolating
1362 *
1363 * returns how many pages were moved onto *@dst.
1364 */
1365static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1366 struct lruvec *lruvec, struct list_head *dst,
1367 unsigned long *nr_scanned, struct scan_control *sc,
1368 isolate_mode_t mode, enum lru_list lru)
1369{
1370 struct list_head *src = &lruvec->lists[lru];
1371 unsigned long nr_taken = 0;
1372 unsigned long scan;
1373
1374 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1375 !list_empty(src); scan++) {
1376 struct page *page;
1377 int nr_pages;
1378
1379 page = lru_to_page(src);
1380 prefetchw_prev_lru_page(page, src, flags);
1381
1382 VM_BUG_ON_PAGE(!PageLRU(page), page);
1383
1384 switch (__isolate_lru_page(page, mode)) {
1385 case 0:
1386 nr_pages = hpage_nr_pages(page);
1387 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1388 list_move(&page->lru, dst);
1389 nr_taken += nr_pages;
1390 break;
1391
1392 case -EBUSY:
1393 /* else it is being freed elsewhere */
1394 list_move(&page->lru, src);
1395 continue;
1396
1397 default:
1398 BUG();
1399 }
1400 }
1401
1402 *nr_scanned = scan;
1403 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1404 nr_taken, mode, is_file_lru(lru));
1405 return nr_taken;
1406}
1407
1408/**
1409 * isolate_lru_page - tries to isolate a page from its LRU list
1410 * @page: page to isolate from its LRU list
1411 *
1412 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1413 * vmstat statistic corresponding to whatever LRU list the page was on.
1414 *
1415 * Returns 0 if the page was removed from an LRU list.
1416 * Returns -EBUSY if the page was not on an LRU list.
1417 *
1418 * The returned page will have PageLRU() cleared. If it was found on
1419 * the active list, it will have PageActive set. If it was found on
1420 * the unevictable list, it will have the PageUnevictable bit set. That flag
1421 * may need to be cleared by the caller before letting the page go.
1422 *
1423 * The vmstat statistic corresponding to the list on which the page was
1424 * found will be decremented.
1425 *
1426 * Restrictions:
1427 * (1) Must be called with an elevated refcount on the page. This is a
1428 * fundamentnal difference from isolate_lru_pages (which is called
1429 * without a stable reference).
1430 * (2) the lru_lock must not be held.
1431 * (3) interrupts must be enabled.
1432 */
1433int isolate_lru_page(struct page *page)
1434{
1435 int ret = -EBUSY;
1436
1437 VM_BUG_ON_PAGE(!page_count(page), page);
1438 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1439
1440 if (PageLRU(page)) {
1441 struct zone *zone = page_zone(page);
1442 struct lruvec *lruvec;
1443
1444 spin_lock_irq(&zone->lru_lock);
1445 lruvec = mem_cgroup_page_lruvec(page, zone);
1446 if (PageLRU(page)) {
1447 int lru = page_lru(page);
1448 get_page(page);
1449 ClearPageLRU(page);
1450 del_page_from_lru_list(page, lruvec, lru);
1451 ret = 0;
1452 }
1453 spin_unlock_irq(&zone->lru_lock);
1454 }
1455 return ret;
1456}
1457
1458/*
1459 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1460 * then get resheduled. When there are massive number of tasks doing page
1461 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1462 * the LRU list will go small and be scanned faster than necessary, leading to
1463 * unnecessary swapping, thrashing and OOM.
1464 */
1465static int too_many_isolated(struct zone *zone, int file,
1466 struct scan_control *sc)
1467{
1468 unsigned long inactive, isolated;
1469
1470 if (current_is_kswapd())
1471 return 0;
1472
1473 if (!sane_reclaim(sc))
1474 return 0;
1475
1476 if (file) {
1477 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1478 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1479 } else {
1480 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1481 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1482 }
1483
1484 /*
1485 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1486 * won't get blocked by normal direct-reclaimers, forming a circular
1487 * deadlock.
1488 */
1489 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1490 inactive >>= 3;
1491
1492 return isolated > inactive;
1493}
1494
1495static noinline_for_stack void
1496putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1497{
1498 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1499 struct zone *zone = lruvec_zone(lruvec);
1500 LIST_HEAD(pages_to_free);
1501
1502 /*
1503 * Put back any unfreeable pages.
1504 */
1505 while (!list_empty(page_list)) {
1506 struct page *page = lru_to_page(page_list);
1507 int lru;
1508
1509 VM_BUG_ON_PAGE(PageLRU(page), page);
1510 list_del(&page->lru);
1511 if (unlikely(!page_evictable(page))) {
1512 spin_unlock_irq(&zone->lru_lock);
1513 putback_lru_page(page);
1514 spin_lock_irq(&zone->lru_lock);
1515 continue;
1516 }
1517
1518 lruvec = mem_cgroup_page_lruvec(page, zone);
1519
1520 SetPageLRU(page);
1521 lru = page_lru(page);
1522 add_page_to_lru_list(page, lruvec, lru);
1523
1524 if (is_active_lru(lru)) {
1525 int file = is_file_lru(lru);
1526 int numpages = hpage_nr_pages(page);
1527 reclaim_stat->recent_rotated[file] += numpages;
1528 }
1529 if (put_page_testzero(page)) {
1530 __ClearPageLRU(page);
1531 __ClearPageActive(page);
1532 del_page_from_lru_list(page, lruvec, lru);
1533
1534 if (unlikely(PageCompound(page))) {
1535 spin_unlock_irq(&zone->lru_lock);
1536 mem_cgroup_uncharge(page);
1537 (*get_compound_page_dtor(page))(page);
1538 spin_lock_irq(&zone->lru_lock);
1539 } else
1540 list_add(&page->lru, &pages_to_free);
1541 }
1542 }
1543
1544 /*
1545 * To save our caller's stack, now use input list for pages to free.
1546 */
1547 list_splice(&pages_to_free, page_list);
1548}
1549
1550/*
1551 * If a kernel thread (such as nfsd for loop-back mounts) services
1552 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1553 * In that case we should only throttle if the backing device it is
1554 * writing to is congested. In other cases it is safe to throttle.
1555 */
1556static int current_may_throttle(void)
1557{
1558 return !(current->flags & PF_LESS_THROTTLE) ||
1559 current->backing_dev_info == NULL ||
1560 bdi_write_congested(current->backing_dev_info);
1561}
1562
1563/*
1564 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1565 * of reclaimed pages
1566 */
1567static noinline_for_stack unsigned long
1568shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1569 struct scan_control *sc, enum lru_list lru)
1570{
1571 LIST_HEAD(page_list);
1572 unsigned long nr_scanned;
1573 unsigned long nr_reclaimed = 0;
1574 unsigned long nr_taken;
1575 unsigned long nr_dirty = 0;
1576 unsigned long nr_congested = 0;
1577 unsigned long nr_unqueued_dirty = 0;
1578 unsigned long nr_writeback = 0;
1579 unsigned long nr_immediate = 0;
1580 isolate_mode_t isolate_mode = 0;
1581 int file = is_file_lru(lru);
1582 struct zone *zone = lruvec_zone(lruvec);
1583 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1584
1585 while (unlikely(too_many_isolated(zone, file, sc))) {
1586 congestion_wait(BLK_RW_ASYNC, HZ/10);
1587
1588 /* We are about to die and free our memory. Return now. */
1589 if (fatal_signal_pending(current))
1590 return SWAP_CLUSTER_MAX;
1591 }
1592
1593 lru_add_drain();
1594
1595 if (!sc->may_unmap)
1596 isolate_mode |= ISOLATE_UNMAPPED;
1597 if (!sc->may_writepage)
1598 isolate_mode |= ISOLATE_CLEAN;
1599
1600 spin_lock_irq(&zone->lru_lock);
1601
1602 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1603 &nr_scanned, sc, isolate_mode, lru);
1604
1605 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1606 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1607
1608 if (global_reclaim(sc)) {
1609 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1610 if (current_is_kswapd())
1611 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1612 else
1613 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1614 }
1615 spin_unlock_irq(&zone->lru_lock);
1616
1617 if (nr_taken == 0)
1618 return 0;
1619
1620 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1621 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1622 &nr_writeback, &nr_immediate,
1623 false);
1624
1625 spin_lock_irq(&zone->lru_lock);
1626
1627 reclaim_stat->recent_scanned[file] += nr_taken;
1628
1629 if (global_reclaim(sc)) {
1630 if (current_is_kswapd())
1631 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1632 nr_reclaimed);
1633 else
1634 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1635 nr_reclaimed);
1636 }
1637
1638 putback_inactive_pages(lruvec, &page_list);
1639
1640 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1641
1642 spin_unlock_irq(&zone->lru_lock);
1643
1644 mem_cgroup_uncharge_list(&page_list);
1645 free_hot_cold_page_list(&page_list, true);
1646
1647 /*
1648 * If reclaim is isolating dirty pages under writeback, it implies
1649 * that the long-lived page allocation rate is exceeding the page
1650 * laundering rate. Either the global limits are not being effective
1651 * at throttling processes due to the page distribution throughout
1652 * zones or there is heavy usage of a slow backing device. The
1653 * only option is to throttle from reclaim context which is not ideal
1654 * as there is no guarantee the dirtying process is throttled in the
1655 * same way balance_dirty_pages() manages.
1656 *
1657 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1658 * of pages under pages flagged for immediate reclaim and stall if any
1659 * are encountered in the nr_immediate check below.
1660 */
1661 if (nr_writeback && nr_writeback == nr_taken)
1662 set_bit(ZONE_WRITEBACK, &zone->flags);
1663
1664 /*
1665 * Legacy memcg will stall in page writeback so avoid forcibly
1666 * stalling here.
1667 */
1668 if (sane_reclaim(sc)) {
1669 /*
1670 * Tag a zone as congested if all the dirty pages scanned were
1671 * backed by a congested BDI and wait_iff_congested will stall.
1672 */
1673 if (nr_dirty && nr_dirty == nr_congested)
1674 set_bit(ZONE_CONGESTED, &zone->flags);
1675
1676 /*
1677 * If dirty pages are scanned that are not queued for IO, it
1678 * implies that flushers are not keeping up. In this case, flag
1679 * the zone ZONE_DIRTY and kswapd will start writing pages from
1680 * reclaim context.
1681 */
1682 if (nr_unqueued_dirty == nr_taken)
1683 set_bit(ZONE_DIRTY, &zone->flags);
1684
1685 /*
1686 * If kswapd scans pages marked marked for immediate
1687 * reclaim and under writeback (nr_immediate), it implies
1688 * that pages are cycling through the LRU faster than
1689 * they are written so also forcibly stall.
1690 */
1691 if (nr_immediate && current_may_throttle())
1692 congestion_wait(BLK_RW_ASYNC, HZ/10);
1693 }
1694
1695 /*
1696 * Stall direct reclaim for IO completions if underlying BDIs or zone
1697 * is congested. Allow kswapd to continue until it starts encountering
1698 * unqueued dirty pages or cycling through the LRU too quickly.
1699 */
1700 if (!sc->hibernation_mode && !current_is_kswapd() &&
1701 current_may_throttle())
1702 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1703
1704 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1705 sc->priority, file);
1706 return nr_reclaimed;
1707}
1708
1709/*
1710 * This moves pages from the active list to the inactive list.
1711 *
1712 * We move them the other way if the page is referenced by one or more
1713 * processes, from rmap.
1714 *
1715 * If the pages are mostly unmapped, the processing is fast and it is
1716 * appropriate to hold zone->lru_lock across the whole operation. But if
1717 * the pages are mapped, the processing is slow (page_referenced()) so we
1718 * should drop zone->lru_lock around each page. It's impossible to balance
1719 * this, so instead we remove the pages from the LRU while processing them.
1720 * It is safe to rely on PG_active against the non-LRU pages in here because
1721 * nobody will play with that bit on a non-LRU page.
1722 *
1723 * The downside is that we have to touch page->_count against each page.
1724 * But we had to alter page->flags anyway.
1725 */
1726
1727static void move_active_pages_to_lru(struct lruvec *lruvec,
1728 struct list_head *list,
1729 struct list_head *pages_to_free,
1730 enum lru_list lru)
1731{
1732 struct zone *zone = lruvec_zone(lruvec);
1733 unsigned long pgmoved = 0;
1734 struct page *page;
1735 int nr_pages;
1736
1737 while (!list_empty(list)) {
1738 page = lru_to_page(list);
1739 lruvec = mem_cgroup_page_lruvec(page, zone);
1740
1741 VM_BUG_ON_PAGE(PageLRU(page), page);
1742 SetPageLRU(page);
1743
1744 nr_pages = hpage_nr_pages(page);
1745 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1746 list_move(&page->lru, &lruvec->lists[lru]);
1747 pgmoved += nr_pages;
1748
1749 if (put_page_testzero(page)) {
1750 __ClearPageLRU(page);
1751 __ClearPageActive(page);
1752 del_page_from_lru_list(page, lruvec, lru);
1753
1754 if (unlikely(PageCompound(page))) {
1755 spin_unlock_irq(&zone->lru_lock);
1756 mem_cgroup_uncharge(page);
1757 (*get_compound_page_dtor(page))(page);
1758 spin_lock_irq(&zone->lru_lock);
1759 } else
1760 list_add(&page->lru, pages_to_free);
1761 }
1762 }
1763 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1764 if (!is_active_lru(lru))
1765 __count_vm_events(PGDEACTIVATE, pgmoved);
1766}
1767
1768static void shrink_active_list(unsigned long nr_to_scan,
1769 struct lruvec *lruvec,
1770 struct scan_control *sc,
1771 enum lru_list lru)
1772{
1773 unsigned long nr_taken;
1774 unsigned long nr_scanned;
1775 unsigned long vm_flags;
1776 LIST_HEAD(l_hold); /* The pages which were snipped off */
1777 LIST_HEAD(l_active);
1778 LIST_HEAD(l_inactive);
1779 struct page *page;
1780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1781 unsigned long nr_rotated = 0;
1782 isolate_mode_t isolate_mode = 0;
1783 int file = is_file_lru(lru);
1784 struct zone *zone = lruvec_zone(lruvec);
1785
1786 lru_add_drain();
1787
1788 if (!sc->may_unmap)
1789 isolate_mode |= ISOLATE_UNMAPPED;
1790 if (!sc->may_writepage)
1791 isolate_mode |= ISOLATE_CLEAN;
1792
1793 spin_lock_irq(&zone->lru_lock);
1794
1795 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1796 &nr_scanned, sc, isolate_mode, lru);
1797 if (global_reclaim(sc))
1798 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1799
1800 reclaim_stat->recent_scanned[file] += nr_taken;
1801
1802 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1803 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1804 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1805 spin_unlock_irq(&zone->lru_lock);
1806
1807 while (!list_empty(&l_hold)) {
1808 cond_resched();
1809 page = lru_to_page(&l_hold);
1810 list_del(&page->lru);
1811
1812 if (unlikely(!page_evictable(page))) {
1813 putback_lru_page(page);
1814 continue;
1815 }
1816
1817 if (unlikely(buffer_heads_over_limit)) {
1818 if (page_has_private(page) && trylock_page(page)) {
1819 if (page_has_private(page))
1820 try_to_release_page(page, 0);
1821 unlock_page(page);
1822 }
1823 }
1824
1825 if (page_referenced(page, 0, sc->target_mem_cgroup,
1826 &vm_flags)) {
1827 nr_rotated += hpage_nr_pages(page);
1828 /*
1829 * Identify referenced, file-backed active pages and
1830 * give them one more trip around the active list. So
1831 * that executable code get better chances to stay in
1832 * memory under moderate memory pressure. Anon pages
1833 * are not likely to be evicted by use-once streaming
1834 * IO, plus JVM can create lots of anon VM_EXEC pages,
1835 * so we ignore them here.
1836 */
1837 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1838 list_add(&page->lru, &l_active);
1839 continue;
1840 }
1841 }
1842
1843 ClearPageActive(page); /* we are de-activating */
1844 list_add(&page->lru, &l_inactive);
1845 }
1846
1847 /*
1848 * Move pages back to the lru list.
1849 */
1850 spin_lock_irq(&zone->lru_lock);
1851 /*
1852 * Count referenced pages from currently used mappings as rotated,
1853 * even though only some of them are actually re-activated. This
1854 * helps balance scan pressure between file and anonymous pages in
1855 * get_scan_count.
1856 */
1857 reclaim_stat->recent_rotated[file] += nr_rotated;
1858
1859 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1860 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1861 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1862 spin_unlock_irq(&zone->lru_lock);
1863
1864 mem_cgroup_uncharge_list(&l_hold);
1865 free_hot_cold_page_list(&l_hold, true);
1866}
1867
1868#ifdef CONFIG_SWAP
1869static bool inactive_anon_is_low_global(struct zone *zone)
1870{
1871 unsigned long active, inactive;
1872
1873 active = zone_page_state(zone, NR_ACTIVE_ANON);
1874 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1875
1876 return inactive * zone->inactive_ratio < active;
1877}
1878
1879/**
1880 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1881 * @lruvec: LRU vector to check
1882 *
1883 * Returns true if the zone does not have enough inactive anon pages,
1884 * meaning some active anon pages need to be deactivated.
1885 */
1886static bool inactive_anon_is_low(struct lruvec *lruvec)
1887{
1888 /*
1889 * If we don't have swap space, anonymous page deactivation
1890 * is pointless.
1891 */
1892 if (!total_swap_pages)
1893 return false;
1894
1895 if (!mem_cgroup_disabled())
1896 return mem_cgroup_inactive_anon_is_low(lruvec);
1897
1898 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1899}
1900#else
1901static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1902{
1903 return false;
1904}
1905#endif
1906
1907/**
1908 * inactive_file_is_low - check if file pages need to be deactivated
1909 * @lruvec: LRU vector to check
1910 *
1911 * When the system is doing streaming IO, memory pressure here
1912 * ensures that active file pages get deactivated, until more
1913 * than half of the file pages are on the inactive list.
1914 *
1915 * Once we get to that situation, protect the system's working
1916 * set from being evicted by disabling active file page aging.
1917 *
1918 * This uses a different ratio than the anonymous pages, because
1919 * the page cache uses a use-once replacement algorithm.
1920 */
1921static bool inactive_file_is_low(struct lruvec *lruvec)
1922{
1923 unsigned long inactive;
1924 unsigned long active;
1925
1926 inactive = lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
1927 active = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
1928
1929 return active > inactive;
1930}
1931
1932static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1933{
1934 if (is_file_lru(lru))
1935 return inactive_file_is_low(lruvec);
1936 else
1937 return inactive_anon_is_low(lruvec);
1938}
1939
1940static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1941 struct lruvec *lruvec, struct scan_control *sc)
1942{
1943 if (is_active_lru(lru)) {
1944 if (inactive_list_is_low(lruvec, lru))
1945 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1946 return 0;
1947 }
1948
1949 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1950}
1951
1952enum scan_balance {
1953 SCAN_EQUAL,
1954 SCAN_FRACT,
1955 SCAN_ANON,
1956 SCAN_FILE,
1957};
1958
1959/*
1960 * Determine how aggressively the anon and file LRU lists should be
1961 * scanned. The relative value of each set of LRU lists is determined
1962 * by looking at the fraction of the pages scanned we did rotate back
1963 * onto the active list instead of evict.
1964 *
1965 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1966 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1967 */
1968static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
1969 struct scan_control *sc, unsigned long *nr,
1970 unsigned long *lru_pages)
1971{
1972 int swappiness = mem_cgroup_swappiness(memcg);
1973 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1974 u64 fraction[2];
1975 u64 denominator = 0; /* gcc */
1976 struct zone *zone = lruvec_zone(lruvec);
1977 unsigned long anon_prio, file_prio;
1978 enum scan_balance scan_balance;
1979 unsigned long anon, file;
1980 bool force_scan = false;
1981 unsigned long ap, fp;
1982 enum lru_list lru;
1983 bool some_scanned;
1984 int pass;
1985
1986 /*
1987 * If the zone or memcg is small, nr[l] can be 0. This
1988 * results in no scanning on this priority and a potential
1989 * priority drop. Global direct reclaim can go to the next
1990 * zone and tends to have no problems. Global kswapd is for
1991 * zone balancing and it needs to scan a minimum amount. When
1992 * reclaiming for a memcg, a priority drop can cause high
1993 * latencies, so it's better to scan a minimum amount there as
1994 * well.
1995 */
1996 if (current_is_kswapd()) {
1997 if (!zone_reclaimable(zone))
1998 force_scan = true;
1999 if (!mem_cgroup_online(memcg))
2000 force_scan = true;
2001 }
2002 if (!global_reclaim(sc))
2003 force_scan = true;
2004
2005 /* If we have no swap space, do not bother scanning anon pages. */
2006 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2007 scan_balance = SCAN_FILE;
2008 goto out;
2009 }
2010
2011 /*
2012 * Global reclaim will swap to prevent OOM even with no
2013 * swappiness, but memcg users want to use this knob to
2014 * disable swapping for individual groups completely when
2015 * using the memory controller's swap limit feature would be
2016 * too expensive.
2017 */
2018 if (!global_reclaim(sc) && !swappiness) {
2019 scan_balance = SCAN_FILE;
2020 goto out;
2021 }
2022
2023 /*
2024 * Do not apply any pressure balancing cleverness when the
2025 * system is close to OOM, scan both anon and file equally
2026 * (unless the swappiness setting disagrees with swapping).
2027 */
2028 if (!sc->priority && swappiness) {
2029 scan_balance = SCAN_EQUAL;
2030 goto out;
2031 }
2032
2033 /*
2034 * Prevent the reclaimer from falling into the cache trap: as
2035 * cache pages start out inactive, every cache fault will tip
2036 * the scan balance towards the file LRU. And as the file LRU
2037 * shrinks, so does the window for rotation from references.
2038 * This means we have a runaway feedback loop where a tiny
2039 * thrashing file LRU becomes infinitely more attractive than
2040 * anon pages. Try to detect this based on file LRU size.
2041 */
2042 if (global_reclaim(sc)) {
2043 unsigned long zonefile;
2044 unsigned long zonefree;
2045
2046 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2047 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2048 zone_page_state(zone, NR_INACTIVE_FILE);
2049
2050 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2051 scan_balance = SCAN_ANON;
2052 goto out;
2053 }
2054 }
2055
2056 /*
2057 * If there is enough inactive page cache, i.e. if the size of the
2058 * inactive list is greater than that of the active list *and* the
2059 * inactive list actually has some pages to scan on this priority, we
2060 * do not reclaim anything from the anonymous working set right now.
2061 * Without the second condition we could end up never scanning an
2062 * lruvec even if it has plenty of old anonymous pages unless the
2063 * system is under heavy pressure.
2064 */
2065 if (!inactive_file_is_low(lruvec) &&
2066 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2067 scan_balance = SCAN_FILE;
2068 goto out;
2069 }
2070
2071 scan_balance = SCAN_FRACT;
2072
2073 /*
2074 * With swappiness at 100, anonymous and file have the same priority.
2075 * This scanning priority is essentially the inverse of IO cost.
2076 */
2077 anon_prio = swappiness;
2078 file_prio = 200 - anon_prio;
2079
2080 /*
2081 * OK, so we have swap space and a fair amount of page cache
2082 * pages. We use the recently rotated / recently scanned
2083 * ratios to determine how valuable each cache is.
2084 *
2085 * Because workloads change over time (and to avoid overflow)
2086 * we keep these statistics as a floating average, which ends
2087 * up weighing recent references more than old ones.
2088 *
2089 * anon in [0], file in [1]
2090 */
2091
2092 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2093 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2094 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2095 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2096
2097 spin_lock_irq(&zone->lru_lock);
2098 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2099 reclaim_stat->recent_scanned[0] /= 2;
2100 reclaim_stat->recent_rotated[0] /= 2;
2101 }
2102
2103 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2104 reclaim_stat->recent_scanned[1] /= 2;
2105 reclaim_stat->recent_rotated[1] /= 2;
2106 }
2107
2108 /*
2109 * The amount of pressure on anon vs file pages is inversely
2110 * proportional to the fraction of recently scanned pages on
2111 * each list that were recently referenced and in active use.
2112 */
2113 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2114 ap /= reclaim_stat->recent_rotated[0] + 1;
2115
2116 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2117 fp /= reclaim_stat->recent_rotated[1] + 1;
2118 spin_unlock_irq(&zone->lru_lock);
2119
2120 fraction[0] = ap;
2121 fraction[1] = fp;
2122 denominator = ap + fp + 1;
2123out:
2124 some_scanned = false;
2125 /* Only use force_scan on second pass. */
2126 for (pass = 0; !some_scanned && pass < 2; pass++) {
2127 *lru_pages = 0;
2128 for_each_evictable_lru(lru) {
2129 int file = is_file_lru(lru);
2130 unsigned long size;
2131 unsigned long scan;
2132
2133 size = lruvec_lru_size(lruvec, lru);
2134 scan = size >> sc->priority;
2135
2136 if (!scan && pass && force_scan)
2137 scan = min(size, SWAP_CLUSTER_MAX);
2138
2139 switch (scan_balance) {
2140 case SCAN_EQUAL:
2141 /* Scan lists relative to size */
2142 break;
2143 case SCAN_FRACT:
2144 /*
2145 * Scan types proportional to swappiness and
2146 * their relative recent reclaim efficiency.
2147 */
2148 scan = div64_u64(scan * fraction[file],
2149 denominator);
2150 break;
2151 case SCAN_FILE:
2152 case SCAN_ANON:
2153 /* Scan one type exclusively */
2154 if ((scan_balance == SCAN_FILE) != file) {
2155 size = 0;
2156 scan = 0;
2157 }
2158 break;
2159 default:
2160 /* Look ma, no brain */
2161 BUG();
2162 }
2163
2164 *lru_pages += size;
2165 nr[lru] = scan;
2166
2167 /*
2168 * Skip the second pass and don't force_scan,
2169 * if we found something to scan.
2170 */
2171 some_scanned |= !!scan;
2172 }
2173 }
2174}
2175
2176#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2177static void init_tlb_ubc(void)
2178{
2179 /*
2180 * This deliberately does not clear the cpumask as it's expensive
2181 * and unnecessary. If there happens to be data in there then the
2182 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2183 * then will be cleared.
2184 */
2185 current->tlb_ubc.flush_required = false;
2186}
2187#else
2188static inline void init_tlb_ubc(void)
2189{
2190}
2191#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2192
2193/*
2194 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2195 */
2196static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *lru_pages)
2198{
2199 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2200 unsigned long nr[NR_LRU_LISTS];
2201 unsigned long targets[NR_LRU_LISTS];
2202 unsigned long nr_to_scan;
2203 enum lru_list lru;
2204 unsigned long nr_reclaimed = 0;
2205 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2206 struct blk_plug plug;
2207 bool scan_adjusted;
2208
2209 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2210
2211 /* Record the original scan target for proportional adjustments later */
2212 memcpy(targets, nr, sizeof(nr));
2213
2214 /*
2215 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2216 * event that can occur when there is little memory pressure e.g.
2217 * multiple streaming readers/writers. Hence, we do not abort scanning
2218 * when the requested number of pages are reclaimed when scanning at
2219 * DEF_PRIORITY on the assumption that the fact we are direct
2220 * reclaiming implies that kswapd is not keeping up and it is best to
2221 * do a batch of work at once. For memcg reclaim one check is made to
2222 * abort proportional reclaim if either the file or anon lru has already
2223 * dropped to zero at the first pass.
2224 */
2225 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2226 sc->priority == DEF_PRIORITY);
2227
2228 init_tlb_ubc();
2229
2230 blk_start_plug(&plug);
2231 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2232 nr[LRU_INACTIVE_FILE]) {
2233 unsigned long nr_anon, nr_file, percentage;
2234 unsigned long nr_scanned;
2235
2236 for_each_evictable_lru(lru) {
2237 if (nr[lru]) {
2238 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2239 nr[lru] -= nr_to_scan;
2240
2241 nr_reclaimed += shrink_list(lru, nr_to_scan,
2242 lruvec, sc);
2243 }
2244 }
2245
2246 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2247 continue;
2248
2249 /*
2250 * For kswapd and memcg, reclaim at least the number of pages
2251 * requested. Ensure that the anon and file LRUs are scanned
2252 * proportionally what was requested by get_scan_count(). We
2253 * stop reclaiming one LRU and reduce the amount scanning
2254 * proportional to the original scan target.
2255 */
2256 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2257 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2258
2259 /*
2260 * It's just vindictive to attack the larger once the smaller
2261 * has gone to zero. And given the way we stop scanning the
2262 * smaller below, this makes sure that we only make one nudge
2263 * towards proportionality once we've got nr_to_reclaim.
2264 */
2265 if (!nr_file || !nr_anon)
2266 break;
2267
2268 if (nr_file > nr_anon) {
2269 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2270 targets[LRU_ACTIVE_ANON] + 1;
2271 lru = LRU_BASE;
2272 percentage = nr_anon * 100 / scan_target;
2273 } else {
2274 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2275 targets[LRU_ACTIVE_FILE] + 1;
2276 lru = LRU_FILE;
2277 percentage = nr_file * 100 / scan_target;
2278 }
2279
2280 /* Stop scanning the smaller of the LRU */
2281 nr[lru] = 0;
2282 nr[lru + LRU_ACTIVE] = 0;
2283
2284 /*
2285 * Recalculate the other LRU scan count based on its original
2286 * scan target and the percentage scanning already complete
2287 */
2288 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2289 nr_scanned = targets[lru] - nr[lru];
2290 nr[lru] = targets[lru] * (100 - percentage) / 100;
2291 nr[lru] -= min(nr[lru], nr_scanned);
2292
2293 lru += LRU_ACTIVE;
2294 nr_scanned = targets[lru] - nr[lru];
2295 nr[lru] = targets[lru] * (100 - percentage) / 100;
2296 nr[lru] -= min(nr[lru], nr_scanned);
2297
2298 scan_adjusted = true;
2299 }
2300 blk_finish_plug(&plug);
2301 sc->nr_reclaimed += nr_reclaimed;
2302
2303 /*
2304 * Even if we did not try to evict anon pages at all, we want to
2305 * rebalance the anon lru active/inactive ratio.
2306 */
2307 if (inactive_anon_is_low(lruvec))
2308 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2309 sc, LRU_ACTIVE_ANON);
2310
2311 throttle_vm_writeout(sc->gfp_mask);
2312}
2313
2314/* Use reclaim/compaction for costly allocs or under memory pressure */
2315static bool in_reclaim_compaction(struct scan_control *sc)
2316{
2317 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2318 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2319 sc->priority < DEF_PRIORITY - 2))
2320 return true;
2321
2322 return false;
2323}
2324
2325/*
2326 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2327 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2328 * true if more pages should be reclaimed such that when the page allocator
2329 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2330 * It will give up earlier than that if there is difficulty reclaiming pages.
2331 */
2332static inline bool should_continue_reclaim(struct zone *zone,
2333 unsigned long nr_reclaimed,
2334 unsigned long nr_scanned,
2335 struct scan_control *sc)
2336{
2337 unsigned long pages_for_compaction;
2338 unsigned long inactive_lru_pages;
2339
2340 /* If not in reclaim/compaction mode, stop */
2341 if (!in_reclaim_compaction(sc))
2342 return false;
2343
2344 /* Consider stopping depending on scan and reclaim activity */
2345 if (sc->gfp_mask & __GFP_REPEAT) {
2346 /*
2347 * For __GFP_REPEAT allocations, stop reclaiming if the
2348 * full LRU list has been scanned and we are still failing
2349 * to reclaim pages. This full LRU scan is potentially
2350 * expensive but a __GFP_REPEAT caller really wants to succeed
2351 */
2352 if (!nr_reclaimed && !nr_scanned)
2353 return false;
2354 } else {
2355 /*
2356 * For non-__GFP_REPEAT allocations which can presumably
2357 * fail without consequence, stop if we failed to reclaim
2358 * any pages from the last SWAP_CLUSTER_MAX number of
2359 * pages that were scanned. This will return to the
2360 * caller faster at the risk reclaim/compaction and
2361 * the resulting allocation attempt fails
2362 */
2363 if (!nr_reclaimed)
2364 return false;
2365 }
2366
2367 /*
2368 * If we have not reclaimed enough pages for compaction and the
2369 * inactive lists are large enough, continue reclaiming
2370 */
2371 pages_for_compaction = (2UL << sc->order);
2372 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2373 if (get_nr_swap_pages() > 0)
2374 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2375 if (sc->nr_reclaimed < pages_for_compaction &&
2376 inactive_lru_pages > pages_for_compaction)
2377 return true;
2378
2379 /* If compaction would go ahead or the allocation would succeed, stop */
2380 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2381 case COMPACT_PARTIAL:
2382 case COMPACT_CONTINUE:
2383 return false;
2384 default:
2385 return true;
2386 }
2387}
2388
2389static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2390 bool is_classzone)
2391{
2392 struct reclaim_state *reclaim_state = current->reclaim_state;
2393 unsigned long nr_reclaimed, nr_scanned;
2394 bool reclaimable = false;
2395
2396 do {
2397 struct mem_cgroup *root = sc->target_mem_cgroup;
2398 struct mem_cgroup_reclaim_cookie reclaim = {
2399 .zone = zone,
2400 .priority = sc->priority,
2401 };
2402 unsigned long zone_lru_pages = 0;
2403 struct mem_cgroup *memcg;
2404
2405 nr_reclaimed = sc->nr_reclaimed;
2406 nr_scanned = sc->nr_scanned;
2407
2408 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2409 do {
2410 unsigned long lru_pages;
2411 unsigned long reclaimed;
2412 unsigned long scanned;
2413
2414 if (mem_cgroup_low(root, memcg)) {
2415 if (!sc->may_thrash)
2416 continue;
2417 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2418 }
2419
2420 reclaimed = sc->nr_reclaimed;
2421 scanned = sc->nr_scanned;
2422
2423 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
2424 zone_lru_pages += lru_pages;
2425
2426 if (memcg && is_classzone)
2427 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2428 memcg, sc->nr_scanned - scanned,
2429 lru_pages);
2430
2431 /* Record the group's reclaim efficiency */
2432 vmpressure(sc->gfp_mask, memcg, false,
2433 sc->nr_scanned - scanned,
2434 sc->nr_reclaimed - reclaimed);
2435
2436 /*
2437 * Direct reclaim and kswapd have to scan all memory
2438 * cgroups to fulfill the overall scan target for the
2439 * zone.
2440 *
2441 * Limit reclaim, on the other hand, only cares about
2442 * nr_to_reclaim pages to be reclaimed and it will
2443 * retry with decreasing priority if one round over the
2444 * whole hierarchy is not sufficient.
2445 */
2446 if (!global_reclaim(sc) &&
2447 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2448 mem_cgroup_iter_break(root, memcg);
2449 break;
2450 }
2451 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2452
2453 /*
2454 * Shrink the slab caches in the same proportion that
2455 * the eligible LRU pages were scanned.
2456 */
2457 if (global_reclaim(sc) && is_classzone)
2458 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2459 sc->nr_scanned - nr_scanned,
2460 zone_lru_pages);
2461
2462 if (reclaim_state) {
2463 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2464 reclaim_state->reclaimed_slab = 0;
2465 }
2466
2467 /* Record the subtree's reclaim efficiency */
2468 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2469 sc->nr_scanned - nr_scanned,
2470 sc->nr_reclaimed - nr_reclaimed);
2471
2472 if (sc->nr_reclaimed - nr_reclaimed)
2473 reclaimable = true;
2474
2475 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2476 sc->nr_scanned - nr_scanned, sc));
2477
2478 return reclaimable;
2479}
2480
2481/*
2482 * Returns true if compaction should go ahead for a high-order request, or
2483 * the high-order allocation would succeed without compaction.
2484 */
2485static inline bool compaction_ready(struct zone *zone, int order)
2486{
2487 unsigned long balance_gap, watermark;
2488 bool watermark_ok;
2489
2490 /*
2491 * Compaction takes time to run and there are potentially other
2492 * callers using the pages just freed. Continue reclaiming until
2493 * there is a buffer of free pages available to give compaction
2494 * a reasonable chance of completing and allocating the page
2495 */
2496 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2497 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2498 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2499 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2500
2501 /*
2502 * If compaction is deferred, reclaim up to a point where
2503 * compaction will have a chance of success when re-enabled
2504 */
2505 if (compaction_deferred(zone, order))
2506 return watermark_ok;
2507
2508 /*
2509 * If compaction is not ready to start and allocation is not likely
2510 * to succeed without it, then keep reclaiming.
2511 */
2512 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2513 return false;
2514
2515 return watermark_ok;
2516}
2517
2518/*
2519 * This is the direct reclaim path, for page-allocating processes. We only
2520 * try to reclaim pages from zones which will satisfy the caller's allocation
2521 * request.
2522 *
2523 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2524 * Because:
2525 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2526 * allocation or
2527 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2528 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2529 * zone defense algorithm.
2530 *
2531 * If a zone is deemed to be full of pinned pages then just give it a light
2532 * scan then give up on it.
2533 *
2534 * Returns true if a zone was reclaimable.
2535 */
2536static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2537{
2538 struct zoneref *z;
2539 struct zone *zone;
2540 unsigned long nr_soft_reclaimed;
2541 unsigned long nr_soft_scanned;
2542 gfp_t orig_mask;
2543 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2544 bool reclaimable = false;
2545
2546 /*
2547 * If the number of buffer_heads in the machine exceeds the maximum
2548 * allowed level, force direct reclaim to scan the highmem zone as
2549 * highmem pages could be pinning lowmem pages storing buffer_heads
2550 */
2551 orig_mask = sc->gfp_mask;
2552 if (buffer_heads_over_limit)
2553 sc->gfp_mask |= __GFP_HIGHMEM;
2554
2555 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2556 gfp_zone(sc->gfp_mask), sc->nodemask) {
2557 enum zone_type classzone_idx;
2558
2559 if (!populated_zone(zone))
2560 continue;
2561
2562 classzone_idx = requested_highidx;
2563 while (!populated_zone(zone->zone_pgdat->node_zones +
2564 classzone_idx))
2565 classzone_idx--;
2566
2567 /*
2568 * Take care memory controller reclaiming has small influence
2569 * to global LRU.
2570 */
2571 if (global_reclaim(sc)) {
2572 if (!cpuset_zone_allowed(zone,
2573 GFP_KERNEL | __GFP_HARDWALL))
2574 continue;
2575
2576 if (sc->priority != DEF_PRIORITY &&
2577 !zone_reclaimable(zone))
2578 continue; /* Let kswapd poll it */
2579
2580 /*
2581 * If we already have plenty of memory free for
2582 * compaction in this zone, don't free any more.
2583 * Even though compaction is invoked for any
2584 * non-zero order, only frequent costly order
2585 * reclamation is disruptive enough to become a
2586 * noticeable problem, like transparent huge
2587 * page allocations.
2588 */
2589 if (IS_ENABLED(CONFIG_COMPACTION) &&
2590 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2591 zonelist_zone_idx(z) <= requested_highidx &&
2592 compaction_ready(zone, sc->order)) {
2593 sc->compaction_ready = true;
2594 continue;
2595 }
2596
2597 /*
2598 * This steals pages from memory cgroups over softlimit
2599 * and returns the number of reclaimed pages and
2600 * scanned pages. This works for global memory pressure
2601 * and balancing, not for a memcg's limit.
2602 */
2603 nr_soft_scanned = 0;
2604 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2605 sc->order, sc->gfp_mask,
2606 &nr_soft_scanned);
2607 sc->nr_reclaimed += nr_soft_reclaimed;
2608 sc->nr_scanned += nr_soft_scanned;
2609 if (nr_soft_reclaimed)
2610 reclaimable = true;
2611 /* need some check for avoid more shrink_zone() */
2612 }
2613
2614 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2615 reclaimable = true;
2616
2617 if (global_reclaim(sc) &&
2618 !reclaimable && zone_reclaimable(zone))
2619 reclaimable = true;
2620 }
2621
2622 /*
2623 * Restore to original mask to avoid the impact on the caller if we
2624 * promoted it to __GFP_HIGHMEM.
2625 */
2626 sc->gfp_mask = orig_mask;
2627
2628 return reclaimable;
2629}
2630
2631/*
2632 * This is the main entry point to direct page reclaim.
2633 *
2634 * If a full scan of the inactive list fails to free enough memory then we
2635 * are "out of memory" and something needs to be killed.
2636 *
2637 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2638 * high - the zone may be full of dirty or under-writeback pages, which this
2639 * caller can't do much about. We kick the writeback threads and take explicit
2640 * naps in the hope that some of these pages can be written. But if the
2641 * allocating task holds filesystem locks which prevent writeout this might not
2642 * work, and the allocation attempt will fail.
2643 *
2644 * returns: 0, if no pages reclaimed
2645 * else, the number of pages reclaimed
2646 */
2647static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2648 struct scan_control *sc)
2649{
2650 int initial_priority = sc->priority;
2651 unsigned long total_scanned = 0;
2652 unsigned long writeback_threshold;
2653 bool zones_reclaimable;
2654retry:
2655 delayacct_freepages_start();
2656
2657 if (global_reclaim(sc))
2658 count_vm_event(ALLOCSTALL);
2659
2660 do {
2661 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2662 sc->priority);
2663 sc->nr_scanned = 0;
2664 zones_reclaimable = shrink_zones(zonelist, sc);
2665
2666 total_scanned += sc->nr_scanned;
2667 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2668 break;
2669
2670 if (sc->compaction_ready)
2671 break;
2672
2673 /*
2674 * If we're getting trouble reclaiming, start doing
2675 * writepage even in laptop mode.
2676 */
2677 if (sc->priority < DEF_PRIORITY - 2)
2678 sc->may_writepage = 1;
2679
2680 /*
2681 * Try to write back as many pages as we just scanned. This
2682 * tends to cause slow streaming writers to write data to the
2683 * disk smoothly, at the dirtying rate, which is nice. But
2684 * that's undesirable in laptop mode, where we *want* lumpy
2685 * writeout. So in laptop mode, write out the whole world.
2686 */
2687 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2688 if (total_scanned > writeback_threshold) {
2689 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2690 WB_REASON_TRY_TO_FREE_PAGES);
2691 sc->may_writepage = 1;
2692 }
2693 } while (--sc->priority >= 0);
2694
2695 delayacct_freepages_end();
2696
2697 if (sc->nr_reclaimed)
2698 return sc->nr_reclaimed;
2699
2700 /* Aborted reclaim to try compaction? don't OOM, then */
2701 if (sc->compaction_ready)
2702 return 1;
2703
2704 /* Untapped cgroup reserves? Don't OOM, retry. */
2705 if (!sc->may_thrash) {
2706 sc->priority = initial_priority;
2707 sc->may_thrash = 1;
2708 goto retry;
2709 }
2710
2711 /* Any of the zones still reclaimable? Don't OOM. */
2712 if (zones_reclaimable)
2713 return 1;
2714
2715 return 0;
2716}
2717
2718static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2719{
2720 struct zone *zone;
2721 unsigned long pfmemalloc_reserve = 0;
2722 unsigned long free_pages = 0;
2723 int i;
2724 bool wmark_ok;
2725
2726 for (i = 0; i <= ZONE_NORMAL; i++) {
2727 zone = &pgdat->node_zones[i];
2728 if (!populated_zone(zone) ||
2729 zone_reclaimable_pages(zone) == 0)
2730 continue;
2731
2732 pfmemalloc_reserve += min_wmark_pages(zone);
2733 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2734 }
2735
2736 /* If there are no reserves (unexpected config) then do not throttle */
2737 if (!pfmemalloc_reserve)
2738 return true;
2739
2740 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2741
2742 /* kswapd must be awake if processes are being throttled */
2743 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2744 pgdat->classzone_idx = min(pgdat->classzone_idx,
2745 (enum zone_type)ZONE_NORMAL);
2746 wake_up_interruptible(&pgdat->kswapd_wait);
2747 }
2748
2749 return wmark_ok;
2750}
2751
2752/*
2753 * Throttle direct reclaimers if backing storage is backed by the network
2754 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2755 * depleted. kswapd will continue to make progress and wake the processes
2756 * when the low watermark is reached.
2757 *
2758 * Returns true if a fatal signal was delivered during throttling. If this
2759 * happens, the page allocator should not consider triggering the OOM killer.
2760 */
2761static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2762 nodemask_t *nodemask)
2763{
2764 struct zoneref *z;
2765 struct zone *zone;
2766 pg_data_t *pgdat = NULL;
2767
2768 /*
2769 * Kernel threads should not be throttled as they may be indirectly
2770 * responsible for cleaning pages necessary for reclaim to make forward
2771 * progress. kjournald for example may enter direct reclaim while
2772 * committing a transaction where throttling it could forcing other
2773 * processes to block on log_wait_commit().
2774 */
2775 if (current->flags & PF_KTHREAD)
2776 goto out;
2777
2778 /*
2779 * If a fatal signal is pending, this process should not throttle.
2780 * It should return quickly so it can exit and free its memory
2781 */
2782 if (fatal_signal_pending(current))
2783 goto out;
2784
2785 /*
2786 * Check if the pfmemalloc reserves are ok by finding the first node
2787 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2788 * GFP_KERNEL will be required for allocating network buffers when
2789 * swapping over the network so ZONE_HIGHMEM is unusable.
2790 *
2791 * Throttling is based on the first usable node and throttled processes
2792 * wait on a queue until kswapd makes progress and wakes them. There
2793 * is an affinity then between processes waking up and where reclaim
2794 * progress has been made assuming the process wakes on the same node.
2795 * More importantly, processes running on remote nodes will not compete
2796 * for remote pfmemalloc reserves and processes on different nodes
2797 * should make reasonable progress.
2798 */
2799 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2800 gfp_zone(gfp_mask), nodemask) {
2801 if (zone_idx(zone) > ZONE_NORMAL)
2802 continue;
2803
2804 /* Throttle based on the first usable node */
2805 pgdat = zone->zone_pgdat;
2806 if (pfmemalloc_watermark_ok(pgdat))
2807 goto out;
2808 break;
2809 }
2810
2811 /* If no zone was usable by the allocation flags then do not throttle */
2812 if (!pgdat)
2813 goto out;
2814
2815 /* Account for the throttling */
2816 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2817
2818 /*
2819 * If the caller cannot enter the filesystem, it's possible that it
2820 * is due to the caller holding an FS lock or performing a journal
2821 * transaction in the case of a filesystem like ext[3|4]. In this case,
2822 * it is not safe to block on pfmemalloc_wait as kswapd could be
2823 * blocked waiting on the same lock. Instead, throttle for up to a
2824 * second before continuing.
2825 */
2826 if (!(gfp_mask & __GFP_FS)) {
2827 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2828 pfmemalloc_watermark_ok(pgdat), HZ);
2829
2830 goto check_pending;
2831 }
2832
2833 /* Throttle until kswapd wakes the process */
2834 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2835 pfmemalloc_watermark_ok(pgdat));
2836
2837check_pending:
2838 if (fatal_signal_pending(current))
2839 return true;
2840
2841out:
2842 return false;
2843}
2844
2845unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2846 gfp_t gfp_mask, nodemask_t *nodemask)
2847{
2848 unsigned long nr_reclaimed;
2849 struct scan_control sc = {
2850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2851 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2852 .order = order,
2853 .nodemask = nodemask,
2854 .priority = DEF_PRIORITY,
2855 .may_writepage = !laptop_mode,
2856 .may_unmap = 1,
2857 .may_swap = 1,
2858 };
2859
2860 /*
2861 * Do not enter reclaim if fatal signal was delivered while throttled.
2862 * 1 is returned so that the page allocator does not OOM kill at this
2863 * point.
2864 */
2865 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2866 return 1;
2867
2868 trace_mm_vmscan_direct_reclaim_begin(order,
2869 sc.may_writepage,
2870 gfp_mask);
2871
2872 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2873
2874 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2875
2876 return nr_reclaimed;
2877}
2878
2879#ifdef CONFIG_MEMCG
2880
2881unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2882 gfp_t gfp_mask, bool noswap,
2883 struct zone *zone,
2884 unsigned long *nr_scanned)
2885{
2886 struct scan_control sc = {
2887 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2888 .target_mem_cgroup = memcg,
2889 .may_writepage = !laptop_mode,
2890 .may_unmap = 1,
2891 .may_swap = !noswap,
2892 };
2893 unsigned long lru_pages;
2894
2895 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2896 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2897
2898 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2899 sc.may_writepage,
2900 sc.gfp_mask);
2901
2902 /*
2903 * NOTE: Although we can get the priority field, using it
2904 * here is not a good idea, since it limits the pages we can scan.
2905 * if we don't reclaim here, the shrink_zone from balance_pgdat
2906 * will pick up pages from other mem cgroup's as well. We hack
2907 * the priority and make it zero.
2908 */
2909 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
2910
2911 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2912
2913 *nr_scanned = sc.nr_scanned;
2914 return sc.nr_reclaimed;
2915}
2916
2917unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2918 unsigned long nr_pages,
2919 gfp_t gfp_mask,
2920 bool may_swap)
2921{
2922 struct zonelist *zonelist;
2923 unsigned long nr_reclaimed;
2924 int nid;
2925 struct scan_control sc = {
2926 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2927 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2928 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2929 .target_mem_cgroup = memcg,
2930 .priority = DEF_PRIORITY,
2931 .may_writepage = !laptop_mode,
2932 .may_unmap = 1,
2933 .may_swap = may_swap,
2934 };
2935
2936 /*
2937 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2938 * take care of from where we get pages. So the node where we start the
2939 * scan does not need to be the current node.
2940 */
2941 nid = mem_cgroup_select_victim_node(memcg);
2942
2943 zonelist = NODE_DATA(nid)->node_zonelists;
2944
2945 trace_mm_vmscan_memcg_reclaim_begin(0,
2946 sc.may_writepage,
2947 sc.gfp_mask);
2948
2949 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2950
2951 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2952
2953 return nr_reclaimed;
2954}
2955#endif
2956
2957static void age_active_anon(struct zone *zone, struct scan_control *sc)
2958{
2959 struct mem_cgroup *memcg;
2960
2961 if (!total_swap_pages)
2962 return;
2963
2964 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2965 do {
2966 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2967
2968 if (inactive_anon_is_low(lruvec))
2969 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2970 sc, LRU_ACTIVE_ANON);
2971
2972 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2973 } while (memcg);
2974}
2975
2976static bool zone_balanced(struct zone *zone, int order, bool highorder,
2977 unsigned long balance_gap, int classzone_idx)
2978{
2979 unsigned long mark = high_wmark_pages(zone) + balance_gap;
2980
2981 /*
2982 * When checking from pgdat_balanced(), kswapd should stop and sleep
2983 * when it reaches the high order-0 watermark and let kcompactd take
2984 * over. Other callers such as wakeup_kswapd() want to determine the
2985 * true high-order watermark.
2986 */
2987 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
2988 mark += (1UL << order);
2989 order = 0;
2990 }
2991
2992 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
2993}
2994
2995/*
2996 * pgdat_balanced() is used when checking if a node is balanced.
2997 *
2998 * For order-0, all zones must be balanced!
2999 *
3000 * For high-order allocations only zones that meet watermarks and are in a
3001 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3002 * total of balanced pages must be at least 25% of the zones allowed by
3003 * classzone_idx for the node to be considered balanced. Forcing all zones to
3004 * be balanced for high orders can cause excessive reclaim when there are
3005 * imbalanced zones.
3006 * The choice of 25% is due to
3007 * o a 16M DMA zone that is balanced will not balance a zone on any
3008 * reasonable sized machine
3009 * o On all other machines, the top zone must be at least a reasonable
3010 * percentage of the middle zones. For example, on 32-bit x86, highmem
3011 * would need to be at least 256M for it to be balance a whole node.
3012 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3013 * to balance a node on its own. These seemed like reasonable ratios.
3014 */
3015static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3016{
3017 unsigned long managed_pages = 0;
3018 unsigned long balanced_pages = 0;
3019 int i;
3020
3021 /* Check the watermark levels */
3022 for (i = 0; i <= classzone_idx; i++) {
3023 struct zone *zone = pgdat->node_zones + i;
3024
3025 if (!populated_zone(zone))
3026 continue;
3027
3028 managed_pages += zone->managed_pages;
3029
3030 /*
3031 * A special case here:
3032 *
3033 * balance_pgdat() skips over all_unreclaimable after
3034 * DEF_PRIORITY. Effectively, it considers them balanced so
3035 * they must be considered balanced here as well!
3036 */
3037 if (!zone_reclaimable(zone)) {
3038 balanced_pages += zone->managed_pages;
3039 continue;
3040 }
3041
3042 if (zone_balanced(zone, order, false, 0, i))
3043 balanced_pages += zone->managed_pages;
3044 else if (!order)
3045 return false;
3046 }
3047
3048 if (order)
3049 return balanced_pages >= (managed_pages >> 2);
3050 else
3051 return true;
3052}
3053
3054/*
3055 * Prepare kswapd for sleeping. This verifies that there are no processes
3056 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3057 *
3058 * Returns true if kswapd is ready to sleep
3059 */
3060static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3061 int classzone_idx)
3062{
3063 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3064 if (remaining)
3065 return false;
3066
3067 /*
3068 * The throttled processes are normally woken up in balance_pgdat() as
3069 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3070 * race between when kswapd checks the watermarks and a process gets
3071 * throttled. There is also a potential race if processes get
3072 * throttled, kswapd wakes, a large process exits thereby balancing the
3073 * zones, which causes kswapd to exit balance_pgdat() before reaching
3074 * the wake up checks. If kswapd is going to sleep, no process should
3075 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3076 * the wake up is premature, processes will wake kswapd and get
3077 * throttled again. The difference from wake ups in balance_pgdat() is
3078 * that here we are under prepare_to_wait().
3079 */
3080 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3081 wake_up_all(&pgdat->pfmemalloc_wait);
3082
3083 return pgdat_balanced(pgdat, order, classzone_idx);
3084}
3085
3086/*
3087 * kswapd shrinks the zone by the number of pages required to reach
3088 * the high watermark.
3089 *
3090 * Returns true if kswapd scanned at least the requested number of pages to
3091 * reclaim or if the lack of progress was due to pages under writeback.
3092 * This is used to determine if the scanning priority needs to be raised.
3093 */
3094static bool kswapd_shrink_zone(struct zone *zone,
3095 int classzone_idx,
3096 struct scan_control *sc)
3097{
3098 unsigned long balance_gap;
3099 bool lowmem_pressure;
3100
3101 /* Reclaim above the high watermark. */
3102 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3103
3104 /*
3105 * We put equal pressure on every zone, unless one zone has way too
3106 * many pages free already. The "too many pages" is defined as the
3107 * high wmark plus a "gap" where the gap is either the low
3108 * watermark or 1% of the zone, whichever is smaller.
3109 */
3110 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3111 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3112
3113 /*
3114 * If there is no low memory pressure or the zone is balanced then no
3115 * reclaim is necessary
3116 */
3117 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3118 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
3119 balance_gap, classzone_idx))
3120 return true;
3121
3122 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3123
3124 clear_bit(ZONE_WRITEBACK, &zone->flags);
3125
3126 /*
3127 * If a zone reaches its high watermark, consider it to be no longer
3128 * congested. It's possible there are dirty pages backed by congested
3129 * BDIs but as pressure is relieved, speculatively avoid congestion
3130 * waits.
3131 */
3132 if (zone_reclaimable(zone) &&
3133 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
3134 clear_bit(ZONE_CONGESTED, &zone->flags);
3135 clear_bit(ZONE_DIRTY, &zone->flags);
3136 }
3137
3138 return sc->nr_scanned >= sc->nr_to_reclaim;
3139}
3140
3141/*
3142 * For kswapd, balance_pgdat() will work across all this node's zones until
3143 * they are all at high_wmark_pages(zone).
3144 *
3145 * Returns the highest zone idx kswapd was reclaiming at
3146 *
3147 * There is special handling here for zones which are full of pinned pages.
3148 * This can happen if the pages are all mlocked, or if they are all used by
3149 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3150 * What we do is to detect the case where all pages in the zone have been
3151 * scanned twice and there has been zero successful reclaim. Mark the zone as
3152 * dead and from now on, only perform a short scan. Basically we're polling
3153 * the zone for when the problem goes away.
3154 *
3155 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3156 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3157 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3158 * lower zones regardless of the number of free pages in the lower zones. This
3159 * interoperates with the page allocator fallback scheme to ensure that aging
3160 * of pages is balanced across the zones.
3161 */
3162static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3163{
3164 int i;
3165 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3166 unsigned long nr_soft_reclaimed;
3167 unsigned long nr_soft_scanned;
3168 struct scan_control sc = {
3169 .gfp_mask = GFP_KERNEL,
3170 .order = order,
3171 .priority = DEF_PRIORITY,
3172 .may_writepage = !laptop_mode,
3173 .may_unmap = 1,
3174 .may_swap = 1,
3175 };
3176 count_vm_event(PAGEOUTRUN);
3177
3178 do {
3179 bool raise_priority = true;
3180
3181 sc.nr_reclaimed = 0;
3182
3183 /*
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3186 */
3187 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188 struct zone *zone = pgdat->node_zones + i;
3189
3190 if (!populated_zone(zone))
3191 continue;
3192
3193 if (sc.priority != DEF_PRIORITY &&
3194 !zone_reclaimable(zone))
3195 continue;
3196
3197 /*
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3200 */
3201 age_active_anon(zone, &sc);
3202
3203 /*
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3208 */
3209 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210 end_zone = i;
3211 break;
3212 }
3213
3214 if (!zone_balanced(zone, order, false, 0, 0)) {
3215 end_zone = i;
3216 break;
3217 } else {
3218 /*
3219 * If balanced, clear the dirty and congested
3220 * flags
3221 */
3222 clear_bit(ZONE_CONGESTED, &zone->flags);
3223 clear_bit(ZONE_DIRTY, &zone->flags);
3224 }
3225 }
3226
3227 if (i < 0)
3228 goto out;
3229
3230 /*
3231 * If we're getting trouble reclaiming, start doing writepage
3232 * even in laptop mode.
3233 */
3234 if (sc.priority < DEF_PRIORITY - 2)
3235 sc.may_writepage = 1;
3236
3237 /*
3238 * Now scan the zone in the dma->highmem direction, stopping
3239 * at the last zone which needs scanning.
3240 *
3241 * We do this because the page allocator works in the opposite
3242 * direction. This prevents the page allocator from allocating
3243 * pages behind kswapd's direction of progress, which would
3244 * cause too much scanning of the lower zones.
3245 */
3246 for (i = 0; i <= end_zone; i++) {
3247 struct zone *zone = pgdat->node_zones + i;
3248
3249 if (!populated_zone(zone))
3250 continue;
3251
3252 if (sc.priority != DEF_PRIORITY &&
3253 !zone_reclaimable(zone))
3254 continue;
3255
3256 sc.nr_scanned = 0;
3257
3258 nr_soft_scanned = 0;
3259 /*
3260 * Call soft limit reclaim before calling shrink_zone.
3261 */
3262 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3263 order, sc.gfp_mask,
3264 &nr_soft_scanned);
3265 sc.nr_reclaimed += nr_soft_reclaimed;
3266
3267 /*
3268 * There should be no need to raise the scanning
3269 * priority if enough pages are already being scanned
3270 * that that high watermark would be met at 100%
3271 * efficiency.
3272 */
3273 if (kswapd_shrink_zone(zone, end_zone, &sc))
3274 raise_priority = false;
3275 }
3276
3277 /*
3278 * If the low watermark is met there is no need for processes
3279 * to be throttled on pfmemalloc_wait as they should not be
3280 * able to safely make forward progress. Wake them
3281 */
3282 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3283 pfmemalloc_watermark_ok(pgdat))
3284 wake_up_all(&pgdat->pfmemalloc_wait);
3285
3286 /* Check if kswapd should be suspending */
3287 if (try_to_freeze() || kthread_should_stop())
3288 break;
3289
3290 /*
3291 * Raise priority if scanning rate is too low or there was no
3292 * progress in reclaiming pages
3293 */
3294 if (raise_priority || !sc.nr_reclaimed)
3295 sc.priority--;
3296 } while (sc.priority >= 1 &&
3297 !pgdat_balanced(pgdat, order, classzone_idx));
3298
3299out:
3300 /*
3301 * Return the highest zone idx we were reclaiming at so
3302 * prepare_kswapd_sleep() makes the same decisions as here.
3303 */
3304 return end_zone;
3305}
3306
3307static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3308 int classzone_idx, int balanced_classzone_idx)
3309{
3310 long remaining = 0;
3311 DEFINE_WAIT(wait);
3312
3313 if (freezing(current) || kthread_should_stop())
3314 return;
3315
3316 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3317
3318 /* Try to sleep for a short interval */
3319 if (prepare_kswapd_sleep(pgdat, order, remaining,
3320 balanced_classzone_idx)) {
3321 /*
3322 * Compaction records what page blocks it recently failed to
3323 * isolate pages from and skips them in the future scanning.
3324 * When kswapd is going to sleep, it is reasonable to assume
3325 * that pages and compaction may succeed so reset the cache.
3326 */
3327 reset_isolation_suitable(pgdat);
3328
3329 /*
3330 * We have freed the memory, now we should compact it to make
3331 * allocation of the requested order possible.
3332 */
3333 wakeup_kcompactd(pgdat, order, classzone_idx);
3334
3335 remaining = schedule_timeout(HZ/10);
3336 finish_wait(&pgdat->kswapd_wait, &wait);
3337 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3338 }
3339
3340 /*
3341 * After a short sleep, check if it was a premature sleep. If not, then
3342 * go fully to sleep until explicitly woken up.
3343 */
3344 if (prepare_kswapd_sleep(pgdat, order, remaining,
3345 balanced_classzone_idx)) {
3346 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3347
3348 /*
3349 * vmstat counters are not perfectly accurate and the estimated
3350 * value for counters such as NR_FREE_PAGES can deviate from the
3351 * true value by nr_online_cpus * threshold. To avoid the zone
3352 * watermarks being breached while under pressure, we reduce the
3353 * per-cpu vmstat threshold while kswapd is awake and restore
3354 * them before going back to sleep.
3355 */
3356 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3357
3358 if (!kthread_should_stop())
3359 schedule();
3360
3361 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3362 } else {
3363 if (remaining)
3364 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3365 else
3366 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3367 }
3368 finish_wait(&pgdat->kswapd_wait, &wait);
3369}
3370
3371/*
3372 * The background pageout daemon, started as a kernel thread
3373 * from the init process.
3374 *
3375 * This basically trickles out pages so that we have _some_
3376 * free memory available even if there is no other activity
3377 * that frees anything up. This is needed for things like routing
3378 * etc, where we otherwise might have all activity going on in
3379 * asynchronous contexts that cannot page things out.
3380 *
3381 * If there are applications that are active memory-allocators
3382 * (most normal use), this basically shouldn't matter.
3383 */
3384static int kswapd(void *p)
3385{
3386 unsigned long order, new_order;
3387 int classzone_idx, new_classzone_idx;
3388 int balanced_classzone_idx;
3389 pg_data_t *pgdat = (pg_data_t*)p;
3390 struct task_struct *tsk = current;
3391
3392 struct reclaim_state reclaim_state = {
3393 .reclaimed_slab = 0,
3394 };
3395 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3396
3397 lockdep_set_current_reclaim_state(GFP_KERNEL);
3398
3399 if (!cpumask_empty(cpumask))
3400 set_cpus_allowed_ptr(tsk, cpumask);
3401 current->reclaim_state = &reclaim_state;
3402
3403 /*
3404 * Tell the memory management that we're a "memory allocator",
3405 * and that if we need more memory we should get access to it
3406 * regardless (see "__alloc_pages()"). "kswapd" should
3407 * never get caught in the normal page freeing logic.
3408 *
3409 * (Kswapd normally doesn't need memory anyway, but sometimes
3410 * you need a small amount of memory in order to be able to
3411 * page out something else, and this flag essentially protects
3412 * us from recursively trying to free more memory as we're
3413 * trying to free the first piece of memory in the first place).
3414 */
3415 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3416 set_freezable();
3417
3418 order = new_order = 0;
3419 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3420 balanced_classzone_idx = classzone_idx;
3421 for ( ; ; ) {
3422 bool ret;
3423
3424 /*
3425 * While we were reclaiming, there might have been another
3426 * wakeup, so check the values.
3427 */
3428 new_order = pgdat->kswapd_max_order;
3429 new_classzone_idx = pgdat->classzone_idx;
3430 pgdat->kswapd_max_order = 0;
3431 pgdat->classzone_idx = pgdat->nr_zones - 1;
3432
3433 if (order < new_order || classzone_idx > new_classzone_idx) {
3434 /*
3435 * Don't sleep if someone wants a larger 'order'
3436 * allocation or has tigher zone constraints
3437 */
3438 order = new_order;
3439 classzone_idx = new_classzone_idx;
3440 } else {
3441 kswapd_try_to_sleep(pgdat, order, classzone_idx,
3442 balanced_classzone_idx);
3443 order = pgdat->kswapd_max_order;
3444 classzone_idx = pgdat->classzone_idx;
3445 new_order = order;
3446 new_classzone_idx = classzone_idx;
3447 pgdat->kswapd_max_order = 0;
3448 pgdat->classzone_idx = pgdat->nr_zones - 1;
3449 }
3450
3451 ret = try_to_freeze();
3452 if (kthread_should_stop())
3453 break;
3454
3455 /*
3456 * We can speed up thawing tasks if we don't call balance_pgdat
3457 * after returning from the refrigerator
3458 */
3459 if (!ret) {
3460 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3461 balanced_classzone_idx = balance_pgdat(pgdat, order,
3462 classzone_idx);
3463 }
3464 }
3465
3466 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3467 current->reclaim_state = NULL;
3468 lockdep_clear_current_reclaim_state();
3469
3470 return 0;
3471}
3472
3473/*
3474 * A zone is low on free memory, so wake its kswapd task to service it.
3475 */
3476void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3477{
3478 pg_data_t *pgdat;
3479
3480 if (!populated_zone(zone))
3481 return;
3482
3483 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3484 return;
3485 pgdat = zone->zone_pgdat;
3486 if (pgdat->kswapd_max_order < order) {
3487 pgdat->kswapd_max_order = order;
3488 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3489 }
3490 if (!waitqueue_active(&pgdat->kswapd_wait))
3491 return;
3492 if (zone_balanced(zone, order, true, 0, 0))
3493 return;
3494
3495 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3496 wake_up_interruptible(&pgdat->kswapd_wait);
3497}
3498
3499#ifdef CONFIG_HIBERNATION
3500/*
3501 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3502 * freed pages.
3503 *
3504 * Rather than trying to age LRUs the aim is to preserve the overall
3505 * LRU order by reclaiming preferentially
3506 * inactive > active > active referenced > active mapped
3507 */
3508unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3509{
3510 struct reclaim_state reclaim_state;
3511 struct scan_control sc = {
3512 .nr_to_reclaim = nr_to_reclaim,
3513 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3514 .priority = DEF_PRIORITY,
3515 .may_writepage = 1,
3516 .may_unmap = 1,
3517 .may_swap = 1,
3518 .hibernation_mode = 1,
3519 };
3520 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3521 struct task_struct *p = current;
3522 unsigned long nr_reclaimed;
3523
3524 p->flags |= PF_MEMALLOC;
3525 lockdep_set_current_reclaim_state(sc.gfp_mask);
3526 reclaim_state.reclaimed_slab = 0;
3527 p->reclaim_state = &reclaim_state;
3528
3529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3530
3531 p->reclaim_state = NULL;
3532 lockdep_clear_current_reclaim_state();
3533 p->flags &= ~PF_MEMALLOC;
3534
3535 return nr_reclaimed;
3536}
3537#endif /* CONFIG_HIBERNATION */
3538
3539/* It's optimal to keep kswapds on the same CPUs as their memory, but
3540 not required for correctness. So if the last cpu in a node goes
3541 away, we get changed to run anywhere: as the first one comes back,
3542 restore their cpu bindings. */
3543static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3544 void *hcpu)
3545{
3546 int nid;
3547
3548 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3549 for_each_node_state(nid, N_MEMORY) {
3550 pg_data_t *pgdat = NODE_DATA(nid);
3551 const struct cpumask *mask;
3552
3553 mask = cpumask_of_node(pgdat->node_id);
3554
3555 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3556 /* One of our CPUs online: restore mask */
3557 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3558 }
3559 }
3560 return NOTIFY_OK;
3561}
3562
3563/*
3564 * This kswapd start function will be called by init and node-hot-add.
3565 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3566 */
3567int kswapd_run(int nid)
3568{
3569 pg_data_t *pgdat = NODE_DATA(nid);
3570 int ret = 0;
3571
3572 if (pgdat->kswapd)
3573 return 0;
3574
3575 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3576 if (IS_ERR(pgdat->kswapd)) {
3577 /* failure at boot is fatal */
3578 BUG_ON(system_state == SYSTEM_BOOTING);
3579 pr_err("Failed to start kswapd on node %d\n", nid);
3580 ret = PTR_ERR(pgdat->kswapd);
3581 pgdat->kswapd = NULL;
3582 }
3583 return ret;
3584}
3585
3586/*
3587 * Called by memory hotplug when all memory in a node is offlined. Caller must
3588 * hold mem_hotplug_begin/end().
3589 */
3590void kswapd_stop(int nid)
3591{
3592 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3593
3594 if (kswapd) {
3595 kthread_stop(kswapd);
3596 NODE_DATA(nid)->kswapd = NULL;
3597 }
3598}
3599
3600static int __init kswapd_init(void)
3601{
3602 int nid;
3603
3604 swap_setup();
3605 for_each_node_state(nid, N_MEMORY)
3606 kswapd_run(nid);
3607 hotcpu_notifier(cpu_callback, 0);
3608 return 0;
3609}
3610
3611module_init(kswapd_init)
3612
3613#ifdef CONFIG_NUMA
3614/*
3615 * Zone reclaim mode
3616 *
3617 * If non-zero call zone_reclaim when the number of free pages falls below
3618 * the watermarks.
3619 */
3620int zone_reclaim_mode __read_mostly;
3621
3622#define RECLAIM_OFF 0
3623#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3624#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3625#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3626
3627/*
3628 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3629 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3630 * a zone.
3631 */
3632#define ZONE_RECLAIM_PRIORITY 4
3633
3634/*
3635 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3636 * occur.
3637 */
3638int sysctl_min_unmapped_ratio = 1;
3639
3640/*
3641 * If the number of slab pages in a zone grows beyond this percentage then
3642 * slab reclaim needs to occur.
3643 */
3644int sysctl_min_slab_ratio = 5;
3645
3646static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3647{
3648 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3649 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3650 zone_page_state(zone, NR_ACTIVE_FILE);
3651
3652 /*
3653 * It's possible for there to be more file mapped pages than
3654 * accounted for by the pages on the file LRU lists because
3655 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3656 */
3657 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3658}
3659
3660/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3661static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3662{
3663 unsigned long nr_pagecache_reclaimable;
3664 unsigned long delta = 0;
3665
3666 /*
3667 * If RECLAIM_UNMAP is set, then all file pages are considered
3668 * potentially reclaimable. Otherwise, we have to worry about
3669 * pages like swapcache and zone_unmapped_file_pages() provides
3670 * a better estimate
3671 */
3672 if (zone_reclaim_mode & RECLAIM_UNMAP)
3673 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3674 else
3675 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3676
3677 /* If we can't clean pages, remove dirty pages from consideration */
3678 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3679 delta += zone_page_state(zone, NR_FILE_DIRTY);
3680
3681 /* Watch for any possible underflows due to delta */
3682 if (unlikely(delta > nr_pagecache_reclaimable))
3683 delta = nr_pagecache_reclaimable;
3684
3685 return nr_pagecache_reclaimable - delta;
3686}
3687
3688/*
3689 * Try to free up some pages from this zone through reclaim.
3690 */
3691static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3692{
3693 /* Minimum pages needed in order to stay on node */
3694 const unsigned long nr_pages = 1 << order;
3695 struct task_struct *p = current;
3696 struct reclaim_state reclaim_state;
3697 struct scan_control sc = {
3698 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3699 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3700 .order = order,
3701 .priority = ZONE_RECLAIM_PRIORITY,
3702 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3703 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3704 .may_swap = 1,
3705 };
3706
3707 cond_resched();
3708 /*
3709 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3710 * and we also need to be able to write out pages for RECLAIM_WRITE
3711 * and RECLAIM_UNMAP.
3712 */
3713 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3714 lockdep_set_current_reclaim_state(gfp_mask);
3715 reclaim_state.reclaimed_slab = 0;
3716 p->reclaim_state = &reclaim_state;
3717
3718 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3719 /*
3720 * Free memory by calling shrink zone with increasing
3721 * priorities until we have enough memory freed.
3722 */
3723 do {
3724 shrink_zone(zone, &sc, true);
3725 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3726 }
3727
3728 p->reclaim_state = NULL;
3729 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3730 lockdep_clear_current_reclaim_state();
3731 return sc.nr_reclaimed >= nr_pages;
3732}
3733
3734int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3735{
3736 int node_id;
3737 int ret;
3738
3739 /*
3740 * Zone reclaim reclaims unmapped file backed pages and
3741 * slab pages if we are over the defined limits.
3742 *
3743 * A small portion of unmapped file backed pages is needed for
3744 * file I/O otherwise pages read by file I/O will be immediately
3745 * thrown out if the zone is overallocated. So we do not reclaim
3746 * if less than a specified percentage of the zone is used by
3747 * unmapped file backed pages.
3748 */
3749 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3750 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3751 return ZONE_RECLAIM_FULL;
3752
3753 if (!zone_reclaimable(zone))
3754 return ZONE_RECLAIM_FULL;
3755
3756 /*
3757 * Do not scan if the allocation should not be delayed.
3758 */
3759 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3760 return ZONE_RECLAIM_NOSCAN;
3761
3762 /*
3763 * Only run zone reclaim on the local zone or on zones that do not
3764 * have associated processors. This will favor the local processor
3765 * over remote processors and spread off node memory allocations
3766 * as wide as possible.
3767 */
3768 node_id = zone_to_nid(zone);
3769 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3770 return ZONE_RECLAIM_NOSCAN;
3771
3772 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3773 return ZONE_RECLAIM_NOSCAN;
3774
3775 ret = __zone_reclaim(zone, gfp_mask, order);
3776 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3777
3778 if (!ret)
3779 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3780
3781 return ret;
3782}
3783#endif
3784
3785/*
3786 * page_evictable - test whether a page is evictable
3787 * @page: the page to test
3788 *
3789 * Test whether page is evictable--i.e., should be placed on active/inactive
3790 * lists vs unevictable list.
3791 *
3792 * Reasons page might not be evictable:
3793 * (1) page's mapping marked unevictable
3794 * (2) page is part of an mlocked VMA
3795 *
3796 */
3797int page_evictable(struct page *page)
3798{
3799 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3800}
3801
3802#ifdef CONFIG_SHMEM
3803/**
3804 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3805 * @pages: array of pages to check
3806 * @nr_pages: number of pages to check
3807 *
3808 * Checks pages for evictability and moves them to the appropriate lru list.
3809 *
3810 * This function is only used for SysV IPC SHM_UNLOCK.
3811 */
3812void check_move_unevictable_pages(struct page **pages, int nr_pages)
3813{
3814 struct lruvec *lruvec;
3815 struct zone *zone = NULL;
3816 int pgscanned = 0;
3817 int pgrescued = 0;
3818 int i;
3819
3820 for (i = 0; i < nr_pages; i++) {
3821 struct page *page = pages[i];
3822 struct zone *pagezone;
3823
3824 pgscanned++;
3825 pagezone = page_zone(page);
3826 if (pagezone != zone) {
3827 if (zone)
3828 spin_unlock_irq(&zone->lru_lock);
3829 zone = pagezone;
3830 spin_lock_irq(&zone->lru_lock);
3831 }
3832 lruvec = mem_cgroup_page_lruvec(page, zone);
3833
3834 if (!PageLRU(page) || !PageUnevictable(page))
3835 continue;
3836
3837 if (page_evictable(page)) {
3838 enum lru_list lru = page_lru_base_type(page);
3839
3840 VM_BUG_ON_PAGE(PageActive(page), page);
3841 ClearPageUnevictable(page);
3842 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3843 add_page_to_lru_list(page, lruvec, lru);
3844 pgrescued++;
3845 }
3846 }
3847
3848 if (zone) {
3849 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3850 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3851 spin_unlock_irq(&zone->lru_lock);
3852 }
3853}
3854#endif /* CONFIG_SHMEM */