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#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/gfp.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
22#include <linux/vmstat.h>
23#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/backing-dev.h>
30#include <linux/rmap.h>
31#include <linux/topology.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
34#include <linux/compaction.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
37#include <linux/delay.h>
38#include <linux/kthread.h>
39#include <linux/freezer.h>
40#include <linux/memcontrol.h>
41#include <linux/delayacct.h>
42#include <linux/sysctl.h>
43#include <linux/oom.h>
44#include <linux/prefetch.h>
45
46#include <asm/tlbflush.h>
47#include <asm/div64.h>
48
49#include <linux/swapops.h>
50
51#include "internal.h"
52
53#define CREATE_TRACE_POINTS
54#include <trace/events/vmscan.h>
55
56struct scan_control {
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned;
59
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed;
62
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
66 unsigned long hibernation_mode;
67
68 /* This context's GFP mask */
69 gfp_t gfp_mask;
70
71 int may_writepage;
72
73 /* Can mapped pages be reclaimed? */
74 int may_unmap;
75
76 /* Can pages be swapped as part of reclaim? */
77 int may_swap;
78
79 int order;
80
81 /* Scan (total_size >> priority) pages at once */
82 int priority;
83
84 /*
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
87 */
88 struct mem_cgroup *target_mem_cgroup;
89
90 /*
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
92 * are scanned.
93 */
94 nodemask_t *nodemask;
95};
96
97#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99#ifdef ARCH_HAS_PREFETCH
100#define prefetch_prev_lru_page(_page, _base, _field) \
101 do { \
102 if ((_page)->lru.prev != _base) { \
103 struct page *prev; \
104 \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
107 } \
108 } while (0)
109#else
110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111#endif
112
113#ifdef ARCH_HAS_PREFETCHW
114#define prefetchw_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127/*
128 * From 0 .. 100. Higher means more swappy.
129 */
130int vm_swappiness = 60;
131long vm_total_pages; /* The total number of pages which the VM controls */
132
133static LIST_HEAD(shrinker_list);
134static DECLARE_RWSEM(shrinker_rwsem);
135
136#ifdef CONFIG_CGROUP_MEM_RES_CTLR
137static bool global_reclaim(struct scan_control *sc)
138{
139 return !sc->target_mem_cgroup;
140}
141#else
142static bool global_reclaim(struct scan_control *sc)
143{
144 return true;
145}
146#endif
147
148static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
149{
150 if (!mem_cgroup_disabled())
151 return mem_cgroup_get_lru_size(lruvec, lru);
152
153 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
154}
155
156/*
157 * Add a shrinker callback to be called from the vm
158 */
159void register_shrinker(struct shrinker *shrinker)
160{
161 atomic_long_set(&shrinker->nr_in_batch, 0);
162 down_write(&shrinker_rwsem);
163 list_add_tail(&shrinker->list, &shrinker_list);
164 up_write(&shrinker_rwsem);
165}
166EXPORT_SYMBOL(register_shrinker);
167
168/*
169 * Remove one
170 */
171void unregister_shrinker(struct shrinker *shrinker)
172{
173 down_write(&shrinker_rwsem);
174 list_del(&shrinker->list);
175 up_write(&shrinker_rwsem);
176}
177EXPORT_SYMBOL(unregister_shrinker);
178
179static inline int do_shrinker_shrink(struct shrinker *shrinker,
180 struct shrink_control *sc,
181 unsigned long nr_to_scan)
182{
183 sc->nr_to_scan = nr_to_scan;
184 return (*shrinker->shrink)(shrinker, sc);
185}
186
187#define SHRINK_BATCH 128
188/*
189 * Call the shrink functions to age shrinkable caches
190 *
191 * Here we assume it costs one seek to replace a lru page and that it also
192 * takes a seek to recreate a cache object. With this in mind we age equal
193 * percentages of the lru and ageable caches. This should balance the seeks
194 * generated by these structures.
195 *
196 * If the vm encountered mapped pages on the LRU it increase the pressure on
197 * slab to avoid swapping.
198 *
199 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
200 *
201 * `lru_pages' represents the number of on-LRU pages in all the zones which
202 * are eligible for the caller's allocation attempt. It is used for balancing
203 * slab reclaim versus page reclaim.
204 *
205 * Returns the number of slab objects which we shrunk.
206 */
207unsigned long shrink_slab(struct shrink_control *shrink,
208 unsigned long nr_pages_scanned,
209 unsigned long lru_pages)
210{
211 struct shrinker *shrinker;
212 unsigned long ret = 0;
213
214 if (nr_pages_scanned == 0)
215 nr_pages_scanned = SWAP_CLUSTER_MAX;
216
217 if (!down_read_trylock(&shrinker_rwsem)) {
218 /* Assume we'll be able to shrink next time */
219 ret = 1;
220 goto out;
221 }
222
223 list_for_each_entry(shrinker, &shrinker_list, list) {
224 unsigned long long delta;
225 long total_scan;
226 long max_pass;
227 int shrink_ret = 0;
228 long nr;
229 long new_nr;
230 long batch_size = shrinker->batch ? shrinker->batch
231 : SHRINK_BATCH;
232
233 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
234 if (max_pass <= 0)
235 continue;
236
237 /*
238 * copy the current shrinker scan count into a local variable
239 * and zero it so that other concurrent shrinker invocations
240 * don't also do this scanning work.
241 */
242 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
243
244 total_scan = nr;
245 delta = (4 * nr_pages_scanned) / shrinker->seeks;
246 delta *= max_pass;
247 do_div(delta, lru_pages + 1);
248 total_scan += delta;
249 if (total_scan < 0) {
250 printk(KERN_ERR "shrink_slab: %pF negative objects to "
251 "delete nr=%ld\n",
252 shrinker->shrink, total_scan);
253 total_scan = max_pass;
254 }
255
256 /*
257 * We need to avoid excessive windup on filesystem shrinkers
258 * due to large numbers of GFP_NOFS allocations causing the
259 * shrinkers to return -1 all the time. This results in a large
260 * nr being built up so when a shrink that can do some work
261 * comes along it empties the entire cache due to nr >>>
262 * max_pass. This is bad for sustaining a working set in
263 * memory.
264 *
265 * Hence only allow the shrinker to scan the entire cache when
266 * a large delta change is calculated directly.
267 */
268 if (delta < max_pass / 4)
269 total_scan = min(total_scan, max_pass / 2);
270
271 /*
272 * Avoid risking looping forever due to too large nr value:
273 * never try to free more than twice the estimate number of
274 * freeable entries.
275 */
276 if (total_scan > max_pass * 2)
277 total_scan = max_pass * 2;
278
279 trace_mm_shrink_slab_start(shrinker, shrink, nr,
280 nr_pages_scanned, lru_pages,
281 max_pass, delta, total_scan);
282
283 while (total_scan >= batch_size) {
284 int nr_before;
285
286 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287 shrink_ret = do_shrinker_shrink(shrinker, shrink,
288 batch_size);
289 if (shrink_ret == -1)
290 break;
291 if (shrink_ret < nr_before)
292 ret += nr_before - shrink_ret;
293 count_vm_events(SLABS_SCANNED, batch_size);
294 total_scan -= batch_size;
295
296 cond_resched();
297 }
298
299 /*
300 * move the unused scan count back into the shrinker in a
301 * manner that handles concurrent updates. If we exhausted the
302 * scan, there is no need to do an update.
303 */
304 if (total_scan > 0)
305 new_nr = atomic_long_add_return(total_scan,
306 &shrinker->nr_in_batch);
307 else
308 new_nr = atomic_long_read(&shrinker->nr_in_batch);
309
310 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
311 }
312 up_read(&shrinker_rwsem);
313out:
314 cond_resched();
315 return ret;
316}
317
318static inline int is_page_cache_freeable(struct page *page)
319{
320 /*
321 * A freeable page cache page is referenced only by the caller
322 * that isolated the page, the page cache radix tree and
323 * optional buffer heads at page->private.
324 */
325 return page_count(page) - page_has_private(page) == 2;
326}
327
328static int may_write_to_queue(struct backing_dev_info *bdi,
329 struct scan_control *sc)
330{
331 if (current->flags & PF_SWAPWRITE)
332 return 1;
333 if (!bdi_write_congested(bdi))
334 return 1;
335 if (bdi == current->backing_dev_info)
336 return 1;
337 return 0;
338}
339
340/*
341 * We detected a synchronous write error writing a page out. Probably
342 * -ENOSPC. We need to propagate that into the address_space for a subsequent
343 * fsync(), msync() or close().
344 *
345 * The tricky part is that after writepage we cannot touch the mapping: nothing
346 * prevents it from being freed up. But we have a ref on the page and once
347 * that page is locked, the mapping is pinned.
348 *
349 * We're allowed to run sleeping lock_page() here because we know the caller has
350 * __GFP_FS.
351 */
352static void handle_write_error(struct address_space *mapping,
353 struct page *page, int error)
354{
355 lock_page(page);
356 if (page_mapping(page) == mapping)
357 mapping_set_error(mapping, error);
358 unlock_page(page);
359}
360
361/* possible outcome of pageout() */
362typedef enum {
363 /* failed to write page out, page is locked */
364 PAGE_KEEP,
365 /* move page to the active list, page is locked */
366 PAGE_ACTIVATE,
367 /* page has been sent to the disk successfully, page is unlocked */
368 PAGE_SUCCESS,
369 /* page is clean and locked */
370 PAGE_CLEAN,
371} pageout_t;
372
373/*
374 * pageout is called by shrink_page_list() for each dirty page.
375 * Calls ->writepage().
376 */
377static pageout_t pageout(struct page *page, struct address_space *mapping,
378 struct scan_control *sc)
379{
380 /*
381 * If the page is dirty, only perform writeback if that write
382 * will be non-blocking. To prevent this allocation from being
383 * stalled by pagecache activity. But note that there may be
384 * stalls if we need to run get_block(). We could test
385 * PagePrivate for that.
386 *
387 * If this process is currently in __generic_file_aio_write() against
388 * this page's queue, we can perform writeback even if that
389 * will block.
390 *
391 * If the page is swapcache, write it back even if that would
392 * block, for some throttling. This happens by accident, because
393 * swap_backing_dev_info is bust: it doesn't reflect the
394 * congestion state of the swapdevs. Easy to fix, if needed.
395 */
396 if (!is_page_cache_freeable(page))
397 return PAGE_KEEP;
398 if (!mapping) {
399 /*
400 * Some data journaling orphaned pages can have
401 * page->mapping == NULL while being dirty with clean buffers.
402 */
403 if (page_has_private(page)) {
404 if (try_to_free_buffers(page)) {
405 ClearPageDirty(page);
406 printk("%s: orphaned page\n", __func__);
407 return PAGE_CLEAN;
408 }
409 }
410 return PAGE_KEEP;
411 }
412 if (mapping->a_ops->writepage == NULL)
413 return PAGE_ACTIVATE;
414 if (!may_write_to_queue(mapping->backing_dev_info, sc))
415 return PAGE_KEEP;
416
417 if (clear_page_dirty_for_io(page)) {
418 int res;
419 struct writeback_control wbc = {
420 .sync_mode = WB_SYNC_NONE,
421 .nr_to_write = SWAP_CLUSTER_MAX,
422 .range_start = 0,
423 .range_end = LLONG_MAX,
424 .for_reclaim = 1,
425 };
426
427 SetPageReclaim(page);
428 res = mapping->a_ops->writepage(page, &wbc);
429 if (res < 0)
430 handle_write_error(mapping, page, res);
431 if (res == AOP_WRITEPAGE_ACTIVATE) {
432 ClearPageReclaim(page);
433 return PAGE_ACTIVATE;
434 }
435
436 if (!PageWriteback(page)) {
437 /* synchronous write or broken a_ops? */
438 ClearPageReclaim(page);
439 }
440 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
441 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442 return PAGE_SUCCESS;
443 }
444
445 return PAGE_CLEAN;
446}
447
448/*
449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
451 */
452static int __remove_mapping(struct address_space *mapping, struct page *page)
453{
454 BUG_ON(!PageLocked(page));
455 BUG_ON(mapping != page_mapping(page));
456
457 spin_lock_irq(&mapping->tree_lock);
458 /*
459 * The non racy check for a busy page.
460 *
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
465 *
466 * get_user_pages(&page);
467 * [user mapping goes away]
468 * write_to(page);
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
471 * put_page(page);
472 * !page_count(page) [good, discard it]
473 *
474 * [oops, our write_to data is lost]
475 *
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
479 *
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
482 */
483 if (!page_freeze_refs(page, 2))
484 goto cannot_free;
485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page))) {
487 page_unfreeze_refs(page, 2);
488 goto cannot_free;
489 }
490
491 if (PageSwapCache(page)) {
492 swp_entry_t swap = { .val = page_private(page) };
493 __delete_from_swap_cache(page);
494 spin_unlock_irq(&mapping->tree_lock);
495 swapcache_free(swap, page);
496 } else {
497 void (*freepage)(struct page *);
498
499 freepage = mapping->a_ops->freepage;
500
501 __delete_from_page_cache(page);
502 spin_unlock_irq(&mapping->tree_lock);
503 mem_cgroup_uncharge_cache_page(page);
504
505 if (freepage != NULL)
506 freepage(page);
507 }
508
509 return 1;
510
511cannot_free:
512 spin_unlock_irq(&mapping->tree_lock);
513 return 0;
514}
515
516/*
517 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
518 * someone else has a ref on the page, abort and return 0. If it was
519 * successfully detached, return 1. Assumes the caller has a single ref on
520 * this page.
521 */
522int remove_mapping(struct address_space *mapping, struct page *page)
523{
524 if (__remove_mapping(mapping, page)) {
525 /*
526 * Unfreezing the refcount with 1 rather than 2 effectively
527 * drops the pagecache ref for us without requiring another
528 * atomic operation.
529 */
530 page_unfreeze_refs(page, 1);
531 return 1;
532 }
533 return 0;
534}
535
536/**
537 * putback_lru_page - put previously isolated page onto appropriate LRU list
538 * @page: page to be put back to appropriate lru list
539 *
540 * Add previously isolated @page to appropriate LRU list.
541 * Page may still be unevictable for other reasons.
542 *
543 * lru_lock must not be held, interrupts must be enabled.
544 */
545void putback_lru_page(struct page *page)
546{
547 int lru;
548 int active = !!TestClearPageActive(page);
549 int was_unevictable = PageUnevictable(page);
550
551 VM_BUG_ON(PageLRU(page));
552
553redo:
554 ClearPageUnevictable(page);
555
556 if (page_evictable(page, NULL)) {
557 /*
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
562 */
563 lru = active + page_lru_base_type(page);
564 lru_cache_add_lru(page, lru);
565 } else {
566 /*
567 * Put unevictable pages directly on zone's unevictable
568 * list.
569 */
570 lru = LRU_UNEVICTABLE;
571 add_page_to_unevictable_list(page);
572 /*
573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
576 * isolation/check_move_unevictable_pages,
577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578 * the page back to the evictable list.
579 *
580 * The other side is TestClearPageMlocked() or shmem_lock().
581 */
582 smp_mb();
583 }
584
585 /*
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
589 */
590 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
591 if (!isolate_lru_page(page)) {
592 put_page(page);
593 goto redo;
594 }
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
598 */
599 }
600
601 if (was_unevictable && lru != LRU_UNEVICTABLE)
602 count_vm_event(UNEVICTABLE_PGRESCUED);
603 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604 count_vm_event(UNEVICTABLE_PGCULLED);
605
606 put_page(page); /* drop ref from isolate */
607}
608
609enum page_references {
610 PAGEREF_RECLAIM,
611 PAGEREF_RECLAIM_CLEAN,
612 PAGEREF_KEEP,
613 PAGEREF_ACTIVATE,
614};
615
616static enum page_references page_check_references(struct page *page,
617 struct scan_control *sc)
618{
619 int referenced_ptes, referenced_page;
620 unsigned long vm_flags;
621
622 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623 &vm_flags);
624 referenced_page = TestClearPageReferenced(page);
625
626 /*
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
629 */
630 if (vm_flags & VM_LOCKED)
631 return PAGEREF_RECLAIM;
632
633 if (referenced_ptes) {
634 if (PageSwapBacked(page))
635 return PAGEREF_ACTIVATE;
636 /*
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
640 * than once.
641 *
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
645 *
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
648 * quickly recovered.
649 */
650 SetPageReferenced(page);
651
652 if (referenced_page || referenced_ptes > 1)
653 return PAGEREF_ACTIVATE;
654
655 /*
656 * Activate file-backed executable pages after first usage.
657 */
658 if (vm_flags & VM_EXEC)
659 return PAGEREF_ACTIVATE;
660
661 return PAGEREF_KEEP;
662 }
663
664 /* Reclaim if clean, defer dirty pages to writeback */
665 if (referenced_page && !PageSwapBacked(page))
666 return PAGEREF_RECLAIM_CLEAN;
667
668 return PAGEREF_RECLAIM;
669}
670
671/*
672 * shrink_page_list() returns the number of reclaimed pages
673 */
674static unsigned long shrink_page_list(struct list_head *page_list,
675 struct zone *zone,
676 struct scan_control *sc,
677 unsigned long *ret_nr_dirty,
678 unsigned long *ret_nr_writeback)
679{
680 LIST_HEAD(ret_pages);
681 LIST_HEAD(free_pages);
682 int pgactivate = 0;
683 unsigned long nr_dirty = 0;
684 unsigned long nr_congested = 0;
685 unsigned long nr_reclaimed = 0;
686 unsigned long nr_writeback = 0;
687
688 cond_resched();
689
690 while (!list_empty(page_list)) {
691 enum page_references references;
692 struct address_space *mapping;
693 struct page *page;
694 int may_enter_fs;
695
696 cond_resched();
697
698 page = lru_to_page(page_list);
699 list_del(&page->lru);
700
701 if (!trylock_page(page))
702 goto keep;
703
704 VM_BUG_ON(PageActive(page));
705 VM_BUG_ON(page_zone(page) != zone);
706
707 sc->nr_scanned++;
708
709 if (unlikely(!page_evictable(page, NULL)))
710 goto cull_mlocked;
711
712 if (!sc->may_unmap && page_mapped(page))
713 goto keep_locked;
714
715 /* Double the slab pressure for mapped and swapcache pages */
716 if (page_mapped(page) || PageSwapCache(page))
717 sc->nr_scanned++;
718
719 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
720 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
721
722 if (PageWriteback(page)) {
723 /*
724 * memcg doesn't have any dirty pages throttling so we
725 * could easily OOM just because too many pages are in
726 * writeback and there is nothing else to reclaim.
727 *
728 * Check __GFP_IO, certainly because a loop driver
729 * thread might enter reclaim, and deadlock if it waits
730 * on a page for which it is needed to do the write
731 * (loop masks off __GFP_IO|__GFP_FS for this reason);
732 * but more thought would probably show more reasons.
733 *
734 * Don't require __GFP_FS, since we're not going into
735 * the FS, just waiting on its writeback completion.
736 * Worryingly, ext4 gfs2 and xfs allocate pages with
737 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
738 * testing may_enter_fs here is liable to OOM on them.
739 */
740 if (global_reclaim(sc) ||
741 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
742 /*
743 * This is slightly racy - end_page_writeback()
744 * might have just cleared PageReclaim, then
745 * setting PageReclaim here end up interpreted
746 * as PageReadahead - but that does not matter
747 * enough to care. What we do want is for this
748 * page to have PageReclaim set next time memcg
749 * reclaim reaches the tests above, so it will
750 * then wait_on_page_writeback() to avoid OOM;
751 * and it's also appropriate in global reclaim.
752 */
753 SetPageReclaim(page);
754 nr_writeback++;
755 goto keep_locked;
756 }
757 wait_on_page_writeback(page);
758 }
759
760 references = page_check_references(page, sc);
761 switch (references) {
762 case PAGEREF_ACTIVATE:
763 goto activate_locked;
764 case PAGEREF_KEEP:
765 goto keep_locked;
766 case PAGEREF_RECLAIM:
767 case PAGEREF_RECLAIM_CLEAN:
768 ; /* try to reclaim the page below */
769 }
770
771 /*
772 * Anonymous process memory has backing store?
773 * Try to allocate it some swap space here.
774 */
775 if (PageAnon(page) && !PageSwapCache(page)) {
776 if (!(sc->gfp_mask & __GFP_IO))
777 goto keep_locked;
778 if (!add_to_swap(page))
779 goto activate_locked;
780 may_enter_fs = 1;
781 }
782
783 mapping = page_mapping(page);
784
785 /*
786 * The page is mapped into the page tables of one or more
787 * processes. Try to unmap it here.
788 */
789 if (page_mapped(page) && mapping) {
790 switch (try_to_unmap(page, TTU_UNMAP)) {
791 case SWAP_FAIL:
792 goto activate_locked;
793 case SWAP_AGAIN:
794 goto keep_locked;
795 case SWAP_MLOCK:
796 goto cull_mlocked;
797 case SWAP_SUCCESS:
798 ; /* try to free the page below */
799 }
800 }
801
802 if (PageDirty(page)) {
803 nr_dirty++;
804
805 /*
806 * Only kswapd can writeback filesystem pages to
807 * avoid risk of stack overflow but do not writeback
808 * unless under significant pressure.
809 */
810 if (page_is_file_cache(page) &&
811 (!current_is_kswapd() ||
812 sc->priority >= DEF_PRIORITY - 2)) {
813 /*
814 * Immediately reclaim when written back.
815 * Similar in principal to deactivate_page()
816 * except we already have the page isolated
817 * and know it's dirty
818 */
819 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
820 SetPageReclaim(page);
821
822 goto keep_locked;
823 }
824
825 if (references == PAGEREF_RECLAIM_CLEAN)
826 goto keep_locked;
827 if (!may_enter_fs)
828 goto keep_locked;
829 if (!sc->may_writepage)
830 goto keep_locked;
831
832 /* Page is dirty, try to write it out here */
833 switch (pageout(page, mapping, sc)) {
834 case PAGE_KEEP:
835 nr_congested++;
836 goto keep_locked;
837 case PAGE_ACTIVATE:
838 goto activate_locked;
839 case PAGE_SUCCESS:
840 if (PageWriteback(page))
841 goto keep;
842 if (PageDirty(page))
843 goto keep;
844
845 /*
846 * A synchronous write - probably a ramdisk. Go
847 * ahead and try to reclaim the page.
848 */
849 if (!trylock_page(page))
850 goto keep;
851 if (PageDirty(page) || PageWriteback(page))
852 goto keep_locked;
853 mapping = page_mapping(page);
854 case PAGE_CLEAN:
855 ; /* try to free the page below */
856 }
857 }
858
859 /*
860 * If the page has buffers, try to free the buffer mappings
861 * associated with this page. If we succeed we try to free
862 * the page as well.
863 *
864 * We do this even if the page is PageDirty().
865 * try_to_release_page() does not perform I/O, but it is
866 * possible for a page to have PageDirty set, but it is actually
867 * clean (all its buffers are clean). This happens if the
868 * buffers were written out directly, with submit_bh(). ext3
869 * will do this, as well as the blockdev mapping.
870 * try_to_release_page() will discover that cleanness and will
871 * drop the buffers and mark the page clean - it can be freed.
872 *
873 * Rarely, pages can have buffers and no ->mapping. These are
874 * the pages which were not successfully invalidated in
875 * truncate_complete_page(). We try to drop those buffers here
876 * and if that worked, and the page is no longer mapped into
877 * process address space (page_count == 1) it can be freed.
878 * Otherwise, leave the page on the LRU so it is swappable.
879 */
880 if (page_has_private(page)) {
881 if (!try_to_release_page(page, sc->gfp_mask))
882 goto activate_locked;
883 if (!mapping && page_count(page) == 1) {
884 unlock_page(page);
885 if (put_page_testzero(page))
886 goto free_it;
887 else {
888 /*
889 * rare race with speculative reference.
890 * the speculative reference will free
891 * this page shortly, so we may
892 * increment nr_reclaimed here (and
893 * leave it off the LRU).
894 */
895 nr_reclaimed++;
896 continue;
897 }
898 }
899 }
900
901 if (!mapping || !__remove_mapping(mapping, page))
902 goto keep_locked;
903
904 /*
905 * At this point, we have no other references and there is
906 * no way to pick any more up (removed from LRU, removed
907 * from pagecache). Can use non-atomic bitops now (and
908 * we obviously don't have to worry about waking up a process
909 * waiting on the page lock, because there are no references.
910 */
911 __clear_page_locked(page);
912free_it:
913 nr_reclaimed++;
914
915 /*
916 * Is there need to periodically free_page_list? It would
917 * appear not as the counts should be low
918 */
919 list_add(&page->lru, &free_pages);
920 continue;
921
922cull_mlocked:
923 if (PageSwapCache(page))
924 try_to_free_swap(page);
925 unlock_page(page);
926 putback_lru_page(page);
927 continue;
928
929activate_locked:
930 /* Not a candidate for swapping, so reclaim swap space. */
931 if (PageSwapCache(page) && vm_swap_full())
932 try_to_free_swap(page);
933 VM_BUG_ON(PageActive(page));
934 SetPageActive(page);
935 pgactivate++;
936keep_locked:
937 unlock_page(page);
938keep:
939 list_add(&page->lru, &ret_pages);
940 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
941 }
942
943 /*
944 * Tag a zone as congested if all the dirty pages encountered were
945 * backed by a congested BDI. In this case, reclaimers should just
946 * back off and wait for congestion to clear because further reclaim
947 * will encounter the same problem
948 */
949 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
950 zone_set_flag(zone, ZONE_CONGESTED);
951
952 free_hot_cold_page_list(&free_pages, 1);
953
954 list_splice(&ret_pages, page_list);
955 count_vm_events(PGACTIVATE, pgactivate);
956 *ret_nr_dirty += nr_dirty;
957 *ret_nr_writeback += nr_writeback;
958 return nr_reclaimed;
959}
960
961/*
962 * Attempt to remove the specified page from its LRU. Only take this page
963 * if it is of the appropriate PageActive status. Pages which are being
964 * freed elsewhere are also ignored.
965 *
966 * page: page to consider
967 * mode: one of the LRU isolation modes defined above
968 *
969 * returns 0 on success, -ve errno on failure.
970 */
971int __isolate_lru_page(struct page *page, isolate_mode_t mode)
972{
973 int ret = -EINVAL;
974
975 /* Only take pages on the LRU. */
976 if (!PageLRU(page))
977 return ret;
978
979 /* Do not give back unevictable pages for compaction */
980 if (PageUnevictable(page))
981 return ret;
982
983 ret = -EBUSY;
984
985 /*
986 * To minimise LRU disruption, the caller can indicate that it only
987 * wants to isolate pages it will be able to operate on without
988 * blocking - clean pages for the most part.
989 *
990 * ISOLATE_CLEAN means that only clean pages should be isolated. This
991 * is used by reclaim when it is cannot write to backing storage
992 *
993 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
994 * that it is possible to migrate without blocking
995 */
996 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
997 /* All the caller can do on PageWriteback is block */
998 if (PageWriteback(page))
999 return ret;
1000
1001 if (PageDirty(page)) {
1002 struct address_space *mapping;
1003
1004 /* ISOLATE_CLEAN means only clean pages */
1005 if (mode & ISOLATE_CLEAN)
1006 return ret;
1007
1008 /*
1009 * Only pages without mappings or that have a
1010 * ->migratepage callback are possible to migrate
1011 * without blocking
1012 */
1013 mapping = page_mapping(page);
1014 if (mapping && !mapping->a_ops->migratepage)
1015 return ret;
1016 }
1017 }
1018
1019 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1020 return ret;
1021
1022 if (likely(get_page_unless_zero(page))) {
1023 /*
1024 * Be careful not to clear PageLRU until after we're
1025 * sure the page is not being freed elsewhere -- the
1026 * page release code relies on it.
1027 */
1028 ClearPageLRU(page);
1029 ret = 0;
1030 }
1031
1032 return ret;
1033}
1034
1035/*
1036 * zone->lru_lock is heavily contended. Some of the functions that
1037 * shrink the lists perform better by taking out a batch of pages
1038 * and working on them outside the LRU lock.
1039 *
1040 * For pagecache intensive workloads, this function is the hottest
1041 * spot in the kernel (apart from copy_*_user functions).
1042 *
1043 * Appropriate locks must be held before calling this function.
1044 *
1045 * @nr_to_scan: The number of pages to look through on the list.
1046 * @lruvec: The LRU vector to pull pages from.
1047 * @dst: The temp list to put pages on to.
1048 * @nr_scanned: The number of pages that were scanned.
1049 * @sc: The scan_control struct for this reclaim session
1050 * @mode: One of the LRU isolation modes
1051 * @lru: LRU list id for isolating
1052 *
1053 * returns how many pages were moved onto *@dst.
1054 */
1055static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1056 struct lruvec *lruvec, struct list_head *dst,
1057 unsigned long *nr_scanned, struct scan_control *sc,
1058 isolate_mode_t mode, enum lru_list lru)
1059{
1060 struct list_head *src = &lruvec->lists[lru];
1061 unsigned long nr_taken = 0;
1062 unsigned long scan;
1063
1064 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1065 struct page *page;
1066 int nr_pages;
1067
1068 page = lru_to_page(src);
1069 prefetchw_prev_lru_page(page, src, flags);
1070
1071 VM_BUG_ON(!PageLRU(page));
1072
1073 switch (__isolate_lru_page(page, mode)) {
1074 case 0:
1075 nr_pages = hpage_nr_pages(page);
1076 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1077 list_move(&page->lru, dst);
1078 nr_taken += nr_pages;
1079 break;
1080
1081 case -EBUSY:
1082 /* else it is being freed elsewhere */
1083 list_move(&page->lru, src);
1084 continue;
1085
1086 default:
1087 BUG();
1088 }
1089 }
1090
1091 *nr_scanned = scan;
1092 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1093 nr_taken, mode, is_file_lru(lru));
1094 return nr_taken;
1095}
1096
1097/**
1098 * isolate_lru_page - tries to isolate a page from its LRU list
1099 * @page: page to isolate from its LRU list
1100 *
1101 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1102 * vmstat statistic corresponding to whatever LRU list the page was on.
1103 *
1104 * Returns 0 if the page was removed from an LRU list.
1105 * Returns -EBUSY if the page was not on an LRU list.
1106 *
1107 * The returned page will have PageLRU() cleared. If it was found on
1108 * the active list, it will have PageActive set. If it was found on
1109 * the unevictable list, it will have the PageUnevictable bit set. That flag
1110 * may need to be cleared by the caller before letting the page go.
1111 *
1112 * The vmstat statistic corresponding to the list on which the page was
1113 * found will be decremented.
1114 *
1115 * Restrictions:
1116 * (1) Must be called with an elevated refcount on the page. This is a
1117 * fundamentnal difference from isolate_lru_pages (which is called
1118 * without a stable reference).
1119 * (2) the lru_lock must not be held.
1120 * (3) interrupts must be enabled.
1121 */
1122int isolate_lru_page(struct page *page)
1123{
1124 int ret = -EBUSY;
1125
1126 VM_BUG_ON(!page_count(page));
1127
1128 if (PageLRU(page)) {
1129 struct zone *zone = page_zone(page);
1130 struct lruvec *lruvec;
1131
1132 spin_lock_irq(&zone->lru_lock);
1133 lruvec = mem_cgroup_page_lruvec(page, zone);
1134 if (PageLRU(page)) {
1135 int lru = page_lru(page);
1136 get_page(page);
1137 ClearPageLRU(page);
1138 del_page_from_lru_list(page, lruvec, lru);
1139 ret = 0;
1140 }
1141 spin_unlock_irq(&zone->lru_lock);
1142 }
1143 return ret;
1144}
1145
1146/*
1147 * Are there way too many processes in the direct reclaim path already?
1148 */
1149static int too_many_isolated(struct zone *zone, int file,
1150 struct scan_control *sc)
1151{
1152 unsigned long inactive, isolated;
1153
1154 if (current_is_kswapd())
1155 return 0;
1156
1157 if (!global_reclaim(sc))
1158 return 0;
1159
1160 if (file) {
1161 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1162 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1163 } else {
1164 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1165 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1166 }
1167
1168 return isolated > inactive;
1169}
1170
1171static noinline_for_stack void
1172putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1173{
1174 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1175 struct zone *zone = lruvec_zone(lruvec);
1176 LIST_HEAD(pages_to_free);
1177
1178 /*
1179 * Put back any unfreeable pages.
1180 */
1181 while (!list_empty(page_list)) {
1182 struct page *page = lru_to_page(page_list);
1183 int lru;
1184
1185 VM_BUG_ON(PageLRU(page));
1186 list_del(&page->lru);
1187 if (unlikely(!page_evictable(page, NULL))) {
1188 spin_unlock_irq(&zone->lru_lock);
1189 putback_lru_page(page);
1190 spin_lock_irq(&zone->lru_lock);
1191 continue;
1192 }
1193
1194 lruvec = mem_cgroup_page_lruvec(page, zone);
1195
1196 SetPageLRU(page);
1197 lru = page_lru(page);
1198 add_page_to_lru_list(page, lruvec, lru);
1199
1200 if (is_active_lru(lru)) {
1201 int file = is_file_lru(lru);
1202 int numpages = hpage_nr_pages(page);
1203 reclaim_stat->recent_rotated[file] += numpages;
1204 }
1205 if (put_page_testzero(page)) {
1206 __ClearPageLRU(page);
1207 __ClearPageActive(page);
1208 del_page_from_lru_list(page, lruvec, lru);
1209
1210 if (unlikely(PageCompound(page))) {
1211 spin_unlock_irq(&zone->lru_lock);
1212 (*get_compound_page_dtor(page))(page);
1213 spin_lock_irq(&zone->lru_lock);
1214 } else
1215 list_add(&page->lru, &pages_to_free);
1216 }
1217 }
1218
1219 /*
1220 * To save our caller's stack, now use input list for pages to free.
1221 */
1222 list_splice(&pages_to_free, page_list);
1223}
1224
1225/*
1226 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1227 * of reclaimed pages
1228 */
1229static noinline_for_stack unsigned long
1230shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1231 struct scan_control *sc, enum lru_list lru)
1232{
1233 LIST_HEAD(page_list);
1234 unsigned long nr_scanned;
1235 unsigned long nr_reclaimed = 0;
1236 unsigned long nr_taken;
1237 unsigned long nr_dirty = 0;
1238 unsigned long nr_writeback = 0;
1239 isolate_mode_t isolate_mode = 0;
1240 int file = is_file_lru(lru);
1241 struct zone *zone = lruvec_zone(lruvec);
1242 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1243
1244 while (unlikely(too_many_isolated(zone, file, sc))) {
1245 congestion_wait(BLK_RW_ASYNC, HZ/10);
1246
1247 /* We are about to die and free our memory. Return now. */
1248 if (fatal_signal_pending(current))
1249 return SWAP_CLUSTER_MAX;
1250 }
1251
1252 lru_add_drain();
1253
1254 if (!sc->may_unmap)
1255 isolate_mode |= ISOLATE_UNMAPPED;
1256 if (!sc->may_writepage)
1257 isolate_mode |= ISOLATE_CLEAN;
1258
1259 spin_lock_irq(&zone->lru_lock);
1260
1261 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1262 &nr_scanned, sc, isolate_mode, lru);
1263
1264 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1265 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1266
1267 if (global_reclaim(sc)) {
1268 zone->pages_scanned += nr_scanned;
1269 if (current_is_kswapd())
1270 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1271 else
1272 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1273 }
1274 spin_unlock_irq(&zone->lru_lock);
1275
1276 if (nr_taken == 0)
1277 return 0;
1278
1279 nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1280 &nr_dirty, &nr_writeback);
1281
1282 spin_lock_irq(&zone->lru_lock);
1283
1284 reclaim_stat->recent_scanned[file] += nr_taken;
1285
1286 if (global_reclaim(sc)) {
1287 if (current_is_kswapd())
1288 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1289 nr_reclaimed);
1290 else
1291 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1292 nr_reclaimed);
1293 }
1294
1295 putback_inactive_pages(lruvec, &page_list);
1296
1297 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1298
1299 spin_unlock_irq(&zone->lru_lock);
1300
1301 free_hot_cold_page_list(&page_list, 1);
1302
1303 /*
1304 * If reclaim is isolating dirty pages under writeback, it implies
1305 * that the long-lived page allocation rate is exceeding the page
1306 * laundering rate. Either the global limits are not being effective
1307 * at throttling processes due to the page distribution throughout
1308 * zones or there is heavy usage of a slow backing device. The
1309 * only option is to throttle from reclaim context which is not ideal
1310 * as there is no guarantee the dirtying process is throttled in the
1311 * same way balance_dirty_pages() manages.
1312 *
1313 * This scales the number of dirty pages that must be under writeback
1314 * before throttling depending on priority. It is a simple backoff
1315 * function that has the most effect in the range DEF_PRIORITY to
1316 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1317 * in trouble and reclaim is considered to be in trouble.
1318 *
1319 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1320 * DEF_PRIORITY-1 50% must be PageWriteback
1321 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1322 * ...
1323 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1324 * isolated page is PageWriteback
1325 */
1326 if (nr_writeback && nr_writeback >=
1327 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1328 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1329
1330 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1331 zone_idx(zone),
1332 nr_scanned, nr_reclaimed,
1333 sc->priority,
1334 trace_shrink_flags(file));
1335 return nr_reclaimed;
1336}
1337
1338/*
1339 * This moves pages from the active list to the inactive list.
1340 *
1341 * We move them the other way if the page is referenced by one or more
1342 * processes, from rmap.
1343 *
1344 * If the pages are mostly unmapped, the processing is fast and it is
1345 * appropriate to hold zone->lru_lock across the whole operation. But if
1346 * the pages are mapped, the processing is slow (page_referenced()) so we
1347 * should drop zone->lru_lock around each page. It's impossible to balance
1348 * this, so instead we remove the pages from the LRU while processing them.
1349 * It is safe to rely on PG_active against the non-LRU pages in here because
1350 * nobody will play with that bit on a non-LRU page.
1351 *
1352 * The downside is that we have to touch page->_count against each page.
1353 * But we had to alter page->flags anyway.
1354 */
1355
1356static void move_active_pages_to_lru(struct lruvec *lruvec,
1357 struct list_head *list,
1358 struct list_head *pages_to_free,
1359 enum lru_list lru)
1360{
1361 struct zone *zone = lruvec_zone(lruvec);
1362 unsigned long pgmoved = 0;
1363 struct page *page;
1364 int nr_pages;
1365
1366 while (!list_empty(list)) {
1367 page = lru_to_page(list);
1368 lruvec = mem_cgroup_page_lruvec(page, zone);
1369
1370 VM_BUG_ON(PageLRU(page));
1371 SetPageLRU(page);
1372
1373 nr_pages = hpage_nr_pages(page);
1374 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1375 list_move(&page->lru, &lruvec->lists[lru]);
1376 pgmoved += nr_pages;
1377
1378 if (put_page_testzero(page)) {
1379 __ClearPageLRU(page);
1380 __ClearPageActive(page);
1381 del_page_from_lru_list(page, lruvec, lru);
1382
1383 if (unlikely(PageCompound(page))) {
1384 spin_unlock_irq(&zone->lru_lock);
1385 (*get_compound_page_dtor(page))(page);
1386 spin_lock_irq(&zone->lru_lock);
1387 } else
1388 list_add(&page->lru, pages_to_free);
1389 }
1390 }
1391 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1392 if (!is_active_lru(lru))
1393 __count_vm_events(PGDEACTIVATE, pgmoved);
1394}
1395
1396static void shrink_active_list(unsigned long nr_to_scan,
1397 struct lruvec *lruvec,
1398 struct scan_control *sc,
1399 enum lru_list lru)
1400{
1401 unsigned long nr_taken;
1402 unsigned long nr_scanned;
1403 unsigned long vm_flags;
1404 LIST_HEAD(l_hold); /* The pages which were snipped off */
1405 LIST_HEAD(l_active);
1406 LIST_HEAD(l_inactive);
1407 struct page *page;
1408 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1409 unsigned long nr_rotated = 0;
1410 isolate_mode_t isolate_mode = 0;
1411 int file = is_file_lru(lru);
1412 struct zone *zone = lruvec_zone(lruvec);
1413
1414 lru_add_drain();
1415
1416 if (!sc->may_unmap)
1417 isolate_mode |= ISOLATE_UNMAPPED;
1418 if (!sc->may_writepage)
1419 isolate_mode |= ISOLATE_CLEAN;
1420
1421 spin_lock_irq(&zone->lru_lock);
1422
1423 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1424 &nr_scanned, sc, isolate_mode, lru);
1425 if (global_reclaim(sc))
1426 zone->pages_scanned += nr_scanned;
1427
1428 reclaim_stat->recent_scanned[file] += nr_taken;
1429
1430 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1431 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1432 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1433 spin_unlock_irq(&zone->lru_lock);
1434
1435 while (!list_empty(&l_hold)) {
1436 cond_resched();
1437 page = lru_to_page(&l_hold);
1438 list_del(&page->lru);
1439
1440 if (unlikely(!page_evictable(page, NULL))) {
1441 putback_lru_page(page);
1442 continue;
1443 }
1444
1445 if (unlikely(buffer_heads_over_limit)) {
1446 if (page_has_private(page) && trylock_page(page)) {
1447 if (page_has_private(page))
1448 try_to_release_page(page, 0);
1449 unlock_page(page);
1450 }
1451 }
1452
1453 if (page_referenced(page, 0, sc->target_mem_cgroup,
1454 &vm_flags)) {
1455 nr_rotated += hpage_nr_pages(page);
1456 /*
1457 * Identify referenced, file-backed active pages and
1458 * give them one more trip around the active list. So
1459 * that executable code get better chances to stay in
1460 * memory under moderate memory pressure. Anon pages
1461 * are not likely to be evicted by use-once streaming
1462 * IO, plus JVM can create lots of anon VM_EXEC pages,
1463 * so we ignore them here.
1464 */
1465 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1466 list_add(&page->lru, &l_active);
1467 continue;
1468 }
1469 }
1470
1471 ClearPageActive(page); /* we are de-activating */
1472 list_add(&page->lru, &l_inactive);
1473 }
1474
1475 /*
1476 * Move pages back to the lru list.
1477 */
1478 spin_lock_irq(&zone->lru_lock);
1479 /*
1480 * Count referenced pages from currently used mappings as rotated,
1481 * even though only some of them are actually re-activated. This
1482 * helps balance scan pressure between file and anonymous pages in
1483 * get_scan_ratio.
1484 */
1485 reclaim_stat->recent_rotated[file] += nr_rotated;
1486
1487 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1488 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1489 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1490 spin_unlock_irq(&zone->lru_lock);
1491
1492 free_hot_cold_page_list(&l_hold, 1);
1493}
1494
1495#ifdef CONFIG_SWAP
1496static int inactive_anon_is_low_global(struct zone *zone)
1497{
1498 unsigned long active, inactive;
1499
1500 active = zone_page_state(zone, NR_ACTIVE_ANON);
1501 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1502
1503 if (inactive * zone->inactive_ratio < active)
1504 return 1;
1505
1506 return 0;
1507}
1508
1509/**
1510 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1511 * @lruvec: LRU vector to check
1512 *
1513 * Returns true if the zone does not have enough inactive anon pages,
1514 * meaning some active anon pages need to be deactivated.
1515 */
1516static int inactive_anon_is_low(struct lruvec *lruvec)
1517{
1518 /*
1519 * If we don't have swap space, anonymous page deactivation
1520 * is pointless.
1521 */
1522 if (!total_swap_pages)
1523 return 0;
1524
1525 if (!mem_cgroup_disabled())
1526 return mem_cgroup_inactive_anon_is_low(lruvec);
1527
1528 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1529}
1530#else
1531static inline int inactive_anon_is_low(struct lruvec *lruvec)
1532{
1533 return 0;
1534}
1535#endif
1536
1537static int inactive_file_is_low_global(struct zone *zone)
1538{
1539 unsigned long active, inactive;
1540
1541 active = zone_page_state(zone, NR_ACTIVE_FILE);
1542 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1543
1544 return (active > inactive);
1545}
1546
1547/**
1548 * inactive_file_is_low - check if file pages need to be deactivated
1549 * @lruvec: LRU vector to check
1550 *
1551 * When the system is doing streaming IO, memory pressure here
1552 * ensures that active file pages get deactivated, until more
1553 * than half of the file pages are on the inactive list.
1554 *
1555 * Once we get to that situation, protect the system's working
1556 * set from being evicted by disabling active file page aging.
1557 *
1558 * This uses a different ratio than the anonymous pages, because
1559 * the page cache uses a use-once replacement algorithm.
1560 */
1561static int inactive_file_is_low(struct lruvec *lruvec)
1562{
1563 if (!mem_cgroup_disabled())
1564 return mem_cgroup_inactive_file_is_low(lruvec);
1565
1566 return inactive_file_is_low_global(lruvec_zone(lruvec));
1567}
1568
1569static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1570{
1571 if (is_file_lru(lru))
1572 return inactive_file_is_low(lruvec);
1573 else
1574 return inactive_anon_is_low(lruvec);
1575}
1576
1577static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1578 struct lruvec *lruvec, struct scan_control *sc)
1579{
1580 if (is_active_lru(lru)) {
1581 if (inactive_list_is_low(lruvec, lru))
1582 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1583 return 0;
1584 }
1585
1586 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1587}
1588
1589static int vmscan_swappiness(struct scan_control *sc)
1590{
1591 if (global_reclaim(sc))
1592 return vm_swappiness;
1593 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1594}
1595
1596/*
1597 * Determine how aggressively the anon and file LRU lists should be
1598 * scanned. The relative value of each set of LRU lists is determined
1599 * by looking at the fraction of the pages scanned we did rotate back
1600 * onto the active list instead of evict.
1601 *
1602 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1603 */
1604static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1605 unsigned long *nr)
1606{
1607 unsigned long anon, file, free;
1608 unsigned long anon_prio, file_prio;
1609 unsigned long ap, fp;
1610 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1611 u64 fraction[2], denominator;
1612 enum lru_list lru;
1613 int noswap = 0;
1614 bool force_scan = false;
1615 struct zone *zone = lruvec_zone(lruvec);
1616
1617 /*
1618 * If the zone or memcg is small, nr[l] can be 0. This
1619 * results in no scanning on this priority and a potential
1620 * priority drop. Global direct reclaim can go to the next
1621 * zone and tends to have no problems. Global kswapd is for
1622 * zone balancing and it needs to scan a minimum amount. When
1623 * reclaiming for a memcg, a priority drop can cause high
1624 * latencies, so it's better to scan a minimum amount there as
1625 * well.
1626 */
1627 if (current_is_kswapd() && zone->all_unreclaimable)
1628 force_scan = true;
1629 if (!global_reclaim(sc))
1630 force_scan = true;
1631
1632 /* If we have no swap space, do not bother scanning anon pages. */
1633 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1634 noswap = 1;
1635 fraction[0] = 0;
1636 fraction[1] = 1;
1637 denominator = 1;
1638 goto out;
1639 }
1640
1641 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1642 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1643 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1644 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1645
1646 if (global_reclaim(sc)) {
1647 free = zone_page_state(zone, NR_FREE_PAGES);
1648 /* If we have very few page cache pages,
1649 force-scan anon pages. */
1650 if (unlikely(file + free <= high_wmark_pages(zone))) {
1651 fraction[0] = 1;
1652 fraction[1] = 0;
1653 denominator = 1;
1654 goto out;
1655 }
1656 }
1657
1658 /*
1659 * With swappiness at 100, anonymous and file have the same priority.
1660 * This scanning priority is essentially the inverse of IO cost.
1661 */
1662 anon_prio = vmscan_swappiness(sc);
1663 file_prio = 200 - anon_prio;
1664
1665 /*
1666 * OK, so we have swap space and a fair amount of page cache
1667 * pages. We use the recently rotated / recently scanned
1668 * ratios to determine how valuable each cache is.
1669 *
1670 * Because workloads change over time (and to avoid overflow)
1671 * we keep these statistics as a floating average, which ends
1672 * up weighing recent references more than old ones.
1673 *
1674 * anon in [0], file in [1]
1675 */
1676 spin_lock_irq(&zone->lru_lock);
1677 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1678 reclaim_stat->recent_scanned[0] /= 2;
1679 reclaim_stat->recent_rotated[0] /= 2;
1680 }
1681
1682 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1683 reclaim_stat->recent_scanned[1] /= 2;
1684 reclaim_stat->recent_rotated[1] /= 2;
1685 }
1686
1687 /*
1688 * The amount of pressure on anon vs file pages is inversely
1689 * proportional to the fraction of recently scanned pages on
1690 * each list that were recently referenced and in active use.
1691 */
1692 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1693 ap /= reclaim_stat->recent_rotated[0] + 1;
1694
1695 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1696 fp /= reclaim_stat->recent_rotated[1] + 1;
1697 spin_unlock_irq(&zone->lru_lock);
1698
1699 fraction[0] = ap;
1700 fraction[1] = fp;
1701 denominator = ap + fp + 1;
1702out:
1703 for_each_evictable_lru(lru) {
1704 int file = is_file_lru(lru);
1705 unsigned long scan;
1706
1707 scan = get_lru_size(lruvec, lru);
1708 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1709 scan >>= sc->priority;
1710 if (!scan && force_scan)
1711 scan = SWAP_CLUSTER_MAX;
1712 scan = div64_u64(scan * fraction[file], denominator);
1713 }
1714 nr[lru] = scan;
1715 }
1716}
1717
1718/* Use reclaim/compaction for costly allocs or under memory pressure */
1719static bool in_reclaim_compaction(struct scan_control *sc)
1720{
1721 if (COMPACTION_BUILD && sc->order &&
1722 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1723 sc->priority < DEF_PRIORITY - 2))
1724 return true;
1725
1726 return false;
1727}
1728
1729/*
1730 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1731 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1732 * true if more pages should be reclaimed such that when the page allocator
1733 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1734 * It will give up earlier than that if there is difficulty reclaiming pages.
1735 */
1736static inline bool should_continue_reclaim(struct lruvec *lruvec,
1737 unsigned long nr_reclaimed,
1738 unsigned long nr_scanned,
1739 struct scan_control *sc)
1740{
1741 unsigned long pages_for_compaction;
1742 unsigned long inactive_lru_pages;
1743
1744 /* If not in reclaim/compaction mode, stop */
1745 if (!in_reclaim_compaction(sc))
1746 return false;
1747
1748 /* Consider stopping depending on scan and reclaim activity */
1749 if (sc->gfp_mask & __GFP_REPEAT) {
1750 /*
1751 * For __GFP_REPEAT allocations, stop reclaiming if the
1752 * full LRU list has been scanned and we are still failing
1753 * to reclaim pages. This full LRU scan is potentially
1754 * expensive but a __GFP_REPEAT caller really wants to succeed
1755 */
1756 if (!nr_reclaimed && !nr_scanned)
1757 return false;
1758 } else {
1759 /*
1760 * For non-__GFP_REPEAT allocations which can presumably
1761 * fail without consequence, stop if we failed to reclaim
1762 * any pages from the last SWAP_CLUSTER_MAX number of
1763 * pages that were scanned. This will return to the
1764 * caller faster at the risk reclaim/compaction and
1765 * the resulting allocation attempt fails
1766 */
1767 if (!nr_reclaimed)
1768 return false;
1769 }
1770
1771 /*
1772 * If we have not reclaimed enough pages for compaction and the
1773 * inactive lists are large enough, continue reclaiming
1774 */
1775 pages_for_compaction = (2UL << sc->order);
1776 inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1777 if (nr_swap_pages > 0)
1778 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
1779 if (sc->nr_reclaimed < pages_for_compaction &&
1780 inactive_lru_pages > pages_for_compaction)
1781 return true;
1782
1783 /* If compaction would go ahead or the allocation would succeed, stop */
1784 switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1785 case COMPACT_PARTIAL:
1786 case COMPACT_CONTINUE:
1787 return false;
1788 default:
1789 return true;
1790 }
1791}
1792
1793/*
1794 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1795 */
1796static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1797{
1798 unsigned long nr[NR_LRU_LISTS];
1799 unsigned long nr_to_scan;
1800 enum lru_list lru;
1801 unsigned long nr_reclaimed, nr_scanned;
1802 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1803 struct blk_plug plug;
1804
1805restart:
1806 nr_reclaimed = 0;
1807 nr_scanned = sc->nr_scanned;
1808 get_scan_count(lruvec, sc, nr);
1809
1810 blk_start_plug(&plug);
1811 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1812 nr[LRU_INACTIVE_FILE]) {
1813 for_each_evictable_lru(lru) {
1814 if (nr[lru]) {
1815 nr_to_scan = min_t(unsigned long,
1816 nr[lru], SWAP_CLUSTER_MAX);
1817 nr[lru] -= nr_to_scan;
1818
1819 nr_reclaimed += shrink_list(lru, nr_to_scan,
1820 lruvec, sc);
1821 }
1822 }
1823 /*
1824 * On large memory systems, scan >> priority can become
1825 * really large. This is fine for the starting priority;
1826 * we want to put equal scanning pressure on each zone.
1827 * However, if the VM has a harder time of freeing pages,
1828 * with multiple processes reclaiming pages, the total
1829 * freeing target can get unreasonably large.
1830 */
1831 if (nr_reclaimed >= nr_to_reclaim &&
1832 sc->priority < DEF_PRIORITY)
1833 break;
1834 }
1835 blk_finish_plug(&plug);
1836 sc->nr_reclaimed += nr_reclaimed;
1837
1838 /*
1839 * Even if we did not try to evict anon pages at all, we want to
1840 * rebalance the anon lru active/inactive ratio.
1841 */
1842 if (inactive_anon_is_low(lruvec))
1843 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1844 sc, LRU_ACTIVE_ANON);
1845
1846 /* reclaim/compaction might need reclaim to continue */
1847 if (should_continue_reclaim(lruvec, nr_reclaimed,
1848 sc->nr_scanned - nr_scanned, sc))
1849 goto restart;
1850
1851 throttle_vm_writeout(sc->gfp_mask);
1852}
1853
1854static void shrink_zone(struct zone *zone, struct scan_control *sc)
1855{
1856 struct mem_cgroup *root = sc->target_mem_cgroup;
1857 struct mem_cgroup_reclaim_cookie reclaim = {
1858 .zone = zone,
1859 .priority = sc->priority,
1860 };
1861 struct mem_cgroup *memcg;
1862
1863 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1864 do {
1865 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1866
1867 shrink_lruvec(lruvec, sc);
1868
1869 /*
1870 * Limit reclaim has historically picked one memcg and
1871 * scanned it with decreasing priority levels until
1872 * nr_to_reclaim had been reclaimed. This priority
1873 * cycle is thus over after a single memcg.
1874 *
1875 * Direct reclaim and kswapd, on the other hand, have
1876 * to scan all memory cgroups to fulfill the overall
1877 * scan target for the zone.
1878 */
1879 if (!global_reclaim(sc)) {
1880 mem_cgroup_iter_break(root, memcg);
1881 break;
1882 }
1883 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1884 } while (memcg);
1885}
1886
1887/* Returns true if compaction should go ahead for a high-order request */
1888static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1889{
1890 unsigned long balance_gap, watermark;
1891 bool watermark_ok;
1892
1893 /* Do not consider compaction for orders reclaim is meant to satisfy */
1894 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1895 return false;
1896
1897 /*
1898 * Compaction takes time to run and there are potentially other
1899 * callers using the pages just freed. Continue reclaiming until
1900 * there is a buffer of free pages available to give compaction
1901 * a reasonable chance of completing and allocating the page
1902 */
1903 balance_gap = min(low_wmark_pages(zone),
1904 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1905 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1906 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1907 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1908
1909 /*
1910 * If compaction is deferred, reclaim up to a point where
1911 * compaction will have a chance of success when re-enabled
1912 */
1913 if (compaction_deferred(zone, sc->order))
1914 return watermark_ok;
1915
1916 /* If compaction is not ready to start, keep reclaiming */
1917 if (!compaction_suitable(zone, sc->order))
1918 return false;
1919
1920 return watermark_ok;
1921}
1922
1923/*
1924 * This is the direct reclaim path, for page-allocating processes. We only
1925 * try to reclaim pages from zones which will satisfy the caller's allocation
1926 * request.
1927 *
1928 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1929 * Because:
1930 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1931 * allocation or
1932 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1933 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1934 * zone defense algorithm.
1935 *
1936 * If a zone is deemed to be full of pinned pages then just give it a light
1937 * scan then give up on it.
1938 *
1939 * This function returns true if a zone is being reclaimed for a costly
1940 * high-order allocation and compaction is ready to begin. This indicates to
1941 * the caller that it should consider retrying the allocation instead of
1942 * further reclaim.
1943 */
1944static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1945{
1946 struct zoneref *z;
1947 struct zone *zone;
1948 unsigned long nr_soft_reclaimed;
1949 unsigned long nr_soft_scanned;
1950 bool aborted_reclaim = false;
1951
1952 /*
1953 * If the number of buffer_heads in the machine exceeds the maximum
1954 * allowed level, force direct reclaim to scan the highmem zone as
1955 * highmem pages could be pinning lowmem pages storing buffer_heads
1956 */
1957 if (buffer_heads_over_limit)
1958 sc->gfp_mask |= __GFP_HIGHMEM;
1959
1960 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1961 gfp_zone(sc->gfp_mask), sc->nodemask) {
1962 if (!populated_zone(zone))
1963 continue;
1964 /*
1965 * Take care memory controller reclaiming has small influence
1966 * to global LRU.
1967 */
1968 if (global_reclaim(sc)) {
1969 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1970 continue;
1971 if (zone->all_unreclaimable &&
1972 sc->priority != DEF_PRIORITY)
1973 continue; /* Let kswapd poll it */
1974 if (COMPACTION_BUILD) {
1975 /*
1976 * If we already have plenty of memory free for
1977 * compaction in this zone, don't free any more.
1978 * Even though compaction is invoked for any
1979 * non-zero order, only frequent costly order
1980 * reclamation is disruptive enough to become a
1981 * noticeable problem, like transparent huge
1982 * page allocations.
1983 */
1984 if (compaction_ready(zone, sc)) {
1985 aborted_reclaim = true;
1986 continue;
1987 }
1988 }
1989 /*
1990 * This steals pages from memory cgroups over softlimit
1991 * and returns the number of reclaimed pages and
1992 * scanned pages. This works for global memory pressure
1993 * and balancing, not for a memcg's limit.
1994 */
1995 nr_soft_scanned = 0;
1996 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
1997 sc->order, sc->gfp_mask,
1998 &nr_soft_scanned);
1999 sc->nr_reclaimed += nr_soft_reclaimed;
2000 sc->nr_scanned += nr_soft_scanned;
2001 /* need some check for avoid more shrink_zone() */
2002 }
2003
2004 shrink_zone(zone, sc);
2005 }
2006
2007 return aborted_reclaim;
2008}
2009
2010static bool zone_reclaimable(struct zone *zone)
2011{
2012 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2013}
2014
2015/* All zones in zonelist are unreclaimable? */
2016static bool all_unreclaimable(struct zonelist *zonelist,
2017 struct scan_control *sc)
2018{
2019 struct zoneref *z;
2020 struct zone *zone;
2021
2022 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2023 gfp_zone(sc->gfp_mask), sc->nodemask) {
2024 if (!populated_zone(zone))
2025 continue;
2026 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2027 continue;
2028 if (!zone->all_unreclaimable)
2029 return false;
2030 }
2031
2032 return true;
2033}
2034
2035/*
2036 * This is the main entry point to direct page reclaim.
2037 *
2038 * If a full scan of the inactive list fails to free enough memory then we
2039 * are "out of memory" and something needs to be killed.
2040 *
2041 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2042 * high - the zone may be full of dirty or under-writeback pages, which this
2043 * caller can't do much about. We kick the writeback threads and take explicit
2044 * naps in the hope that some of these pages can be written. But if the
2045 * allocating task holds filesystem locks which prevent writeout this might not
2046 * work, and the allocation attempt will fail.
2047 *
2048 * returns: 0, if no pages reclaimed
2049 * else, the number of pages reclaimed
2050 */
2051static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2052 struct scan_control *sc,
2053 struct shrink_control *shrink)
2054{
2055 unsigned long total_scanned = 0;
2056 struct reclaim_state *reclaim_state = current->reclaim_state;
2057 struct zoneref *z;
2058 struct zone *zone;
2059 unsigned long writeback_threshold;
2060 bool aborted_reclaim;
2061
2062 delayacct_freepages_start();
2063
2064 if (global_reclaim(sc))
2065 count_vm_event(ALLOCSTALL);
2066
2067 do {
2068 sc->nr_scanned = 0;
2069 aborted_reclaim = shrink_zones(zonelist, sc);
2070
2071 /*
2072 * Don't shrink slabs when reclaiming memory from
2073 * over limit cgroups
2074 */
2075 if (global_reclaim(sc)) {
2076 unsigned long lru_pages = 0;
2077 for_each_zone_zonelist(zone, z, zonelist,
2078 gfp_zone(sc->gfp_mask)) {
2079 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2080 continue;
2081
2082 lru_pages += zone_reclaimable_pages(zone);
2083 }
2084
2085 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2086 if (reclaim_state) {
2087 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2088 reclaim_state->reclaimed_slab = 0;
2089 }
2090 }
2091 total_scanned += sc->nr_scanned;
2092 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2093 goto out;
2094
2095 /*
2096 * Try to write back as many pages as we just scanned. This
2097 * tends to cause slow streaming writers to write data to the
2098 * disk smoothly, at the dirtying rate, which is nice. But
2099 * that's undesirable in laptop mode, where we *want* lumpy
2100 * writeout. So in laptop mode, write out the whole world.
2101 */
2102 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2103 if (total_scanned > writeback_threshold) {
2104 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2105 WB_REASON_TRY_TO_FREE_PAGES);
2106 sc->may_writepage = 1;
2107 }
2108
2109 /* Take a nap, wait for some writeback to complete */
2110 if (!sc->hibernation_mode && sc->nr_scanned &&
2111 sc->priority < DEF_PRIORITY - 2) {
2112 struct zone *preferred_zone;
2113
2114 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2115 &cpuset_current_mems_allowed,
2116 &preferred_zone);
2117 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2118 }
2119 } while (--sc->priority >= 0);
2120
2121out:
2122 delayacct_freepages_end();
2123
2124 if (sc->nr_reclaimed)
2125 return sc->nr_reclaimed;
2126
2127 /*
2128 * As hibernation is going on, kswapd is freezed so that it can't mark
2129 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2130 * check.
2131 */
2132 if (oom_killer_disabled)
2133 return 0;
2134
2135 /* Aborted reclaim to try compaction? don't OOM, then */
2136 if (aborted_reclaim)
2137 return 1;
2138
2139 /* top priority shrink_zones still had more to do? don't OOM, then */
2140 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2141 return 1;
2142
2143 return 0;
2144}
2145
2146unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2147 gfp_t gfp_mask, nodemask_t *nodemask)
2148{
2149 unsigned long nr_reclaimed;
2150 struct scan_control sc = {
2151 .gfp_mask = gfp_mask,
2152 .may_writepage = !laptop_mode,
2153 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2154 .may_unmap = 1,
2155 .may_swap = 1,
2156 .order = order,
2157 .priority = DEF_PRIORITY,
2158 .target_mem_cgroup = NULL,
2159 .nodemask = nodemask,
2160 };
2161 struct shrink_control shrink = {
2162 .gfp_mask = sc.gfp_mask,
2163 };
2164
2165 trace_mm_vmscan_direct_reclaim_begin(order,
2166 sc.may_writepage,
2167 gfp_mask);
2168
2169 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2170
2171 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2172
2173 return nr_reclaimed;
2174}
2175
2176#ifdef CONFIG_CGROUP_MEM_RES_CTLR
2177
2178unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2179 gfp_t gfp_mask, bool noswap,
2180 struct zone *zone,
2181 unsigned long *nr_scanned)
2182{
2183 struct scan_control sc = {
2184 .nr_scanned = 0,
2185 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2186 .may_writepage = !laptop_mode,
2187 .may_unmap = 1,
2188 .may_swap = !noswap,
2189 .order = 0,
2190 .priority = 0,
2191 .target_mem_cgroup = memcg,
2192 };
2193 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2194
2195 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2196 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2197
2198 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2199 sc.may_writepage,
2200 sc.gfp_mask);
2201
2202 /*
2203 * NOTE: Although we can get the priority field, using it
2204 * here is not a good idea, since it limits the pages we can scan.
2205 * if we don't reclaim here, the shrink_zone from balance_pgdat
2206 * will pick up pages from other mem cgroup's as well. We hack
2207 * the priority and make it zero.
2208 */
2209 shrink_lruvec(lruvec, &sc);
2210
2211 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2212
2213 *nr_scanned = sc.nr_scanned;
2214 return sc.nr_reclaimed;
2215}
2216
2217unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2218 gfp_t gfp_mask,
2219 bool noswap)
2220{
2221 struct zonelist *zonelist;
2222 unsigned long nr_reclaimed;
2223 int nid;
2224 struct scan_control sc = {
2225 .may_writepage = !laptop_mode,
2226 .may_unmap = 1,
2227 .may_swap = !noswap,
2228 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2229 .order = 0,
2230 .priority = DEF_PRIORITY,
2231 .target_mem_cgroup = memcg,
2232 .nodemask = NULL, /* we don't care the placement */
2233 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2234 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2235 };
2236 struct shrink_control shrink = {
2237 .gfp_mask = sc.gfp_mask,
2238 };
2239
2240 /*
2241 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2242 * take care of from where we get pages. So the node where we start the
2243 * scan does not need to be the current node.
2244 */
2245 nid = mem_cgroup_select_victim_node(memcg);
2246
2247 zonelist = NODE_DATA(nid)->node_zonelists;
2248
2249 trace_mm_vmscan_memcg_reclaim_begin(0,
2250 sc.may_writepage,
2251 sc.gfp_mask);
2252
2253 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2254
2255 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2256
2257 return nr_reclaimed;
2258}
2259#endif
2260
2261static void age_active_anon(struct zone *zone, struct scan_control *sc)
2262{
2263 struct mem_cgroup *memcg;
2264
2265 if (!total_swap_pages)
2266 return;
2267
2268 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2269 do {
2270 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2271
2272 if (inactive_anon_is_low(lruvec))
2273 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2274 sc, LRU_ACTIVE_ANON);
2275
2276 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2277 } while (memcg);
2278}
2279
2280/*
2281 * pgdat_balanced is used when checking if a node is balanced for high-order
2282 * allocations. Only zones that meet watermarks and are in a zone allowed
2283 * by the callers classzone_idx are added to balanced_pages. The total of
2284 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2285 * for the node to be considered balanced. Forcing all zones to be balanced
2286 * for high orders can cause excessive reclaim when there are imbalanced zones.
2287 * The choice of 25% is due to
2288 * o a 16M DMA zone that is balanced will not balance a zone on any
2289 * reasonable sized machine
2290 * o On all other machines, the top zone must be at least a reasonable
2291 * percentage of the middle zones. For example, on 32-bit x86, highmem
2292 * would need to be at least 256M for it to be balance a whole node.
2293 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2294 * to balance a node on its own. These seemed like reasonable ratios.
2295 */
2296static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2297 int classzone_idx)
2298{
2299 unsigned long present_pages = 0;
2300 int i;
2301
2302 for (i = 0; i <= classzone_idx; i++)
2303 present_pages += pgdat->node_zones[i].present_pages;
2304
2305 /* A special case here: if zone has no page, we think it's balanced */
2306 return balanced_pages >= (present_pages >> 2);
2307}
2308
2309/* is kswapd sleeping prematurely? */
2310static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2311 int classzone_idx)
2312{
2313 int i;
2314 unsigned long balanced = 0;
2315 bool all_zones_ok = true;
2316
2317 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2318 if (remaining)
2319 return true;
2320
2321 /* Check the watermark levels */
2322 for (i = 0; i <= classzone_idx; i++) {
2323 struct zone *zone = pgdat->node_zones + i;
2324
2325 if (!populated_zone(zone))
2326 continue;
2327
2328 /*
2329 * balance_pgdat() skips over all_unreclaimable after
2330 * DEF_PRIORITY. Effectively, it considers them balanced so
2331 * they must be considered balanced here as well if kswapd
2332 * is to sleep
2333 */
2334 if (zone->all_unreclaimable) {
2335 balanced += zone->present_pages;
2336 continue;
2337 }
2338
2339 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2340 i, 0))
2341 all_zones_ok = false;
2342 else
2343 balanced += zone->present_pages;
2344 }
2345
2346 /*
2347 * For high-order requests, the balanced zones must contain at least
2348 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2349 * must be balanced
2350 */
2351 if (order)
2352 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2353 else
2354 return !all_zones_ok;
2355}
2356
2357/*
2358 * For kswapd, balance_pgdat() will work across all this node's zones until
2359 * they are all at high_wmark_pages(zone).
2360 *
2361 * Returns the final order kswapd was reclaiming at
2362 *
2363 * There is special handling here for zones which are full of pinned pages.
2364 * This can happen if the pages are all mlocked, or if they are all used by
2365 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2366 * What we do is to detect the case where all pages in the zone have been
2367 * scanned twice and there has been zero successful reclaim. Mark the zone as
2368 * dead and from now on, only perform a short scan. Basically we're polling
2369 * the zone for when the problem goes away.
2370 *
2371 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2372 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2373 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2374 * lower zones regardless of the number of free pages in the lower zones. This
2375 * interoperates with the page allocator fallback scheme to ensure that aging
2376 * of pages is balanced across the zones.
2377 */
2378static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2379 int *classzone_idx)
2380{
2381 int all_zones_ok;
2382 unsigned long balanced;
2383 int i;
2384 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2385 unsigned long total_scanned;
2386 struct reclaim_state *reclaim_state = current->reclaim_state;
2387 unsigned long nr_soft_reclaimed;
2388 unsigned long nr_soft_scanned;
2389 struct scan_control sc = {
2390 .gfp_mask = GFP_KERNEL,
2391 .may_unmap = 1,
2392 .may_swap = 1,
2393 /*
2394 * kswapd doesn't want to be bailed out while reclaim. because
2395 * we want to put equal scanning pressure on each zone.
2396 */
2397 .nr_to_reclaim = ULONG_MAX,
2398 .order = order,
2399 .target_mem_cgroup = NULL,
2400 };
2401 struct shrink_control shrink = {
2402 .gfp_mask = sc.gfp_mask,
2403 };
2404loop_again:
2405 total_scanned = 0;
2406 sc.priority = DEF_PRIORITY;
2407 sc.nr_reclaimed = 0;
2408 sc.may_writepage = !laptop_mode;
2409 count_vm_event(PAGEOUTRUN);
2410
2411 do {
2412 unsigned long lru_pages = 0;
2413 int has_under_min_watermark_zone = 0;
2414
2415 all_zones_ok = 1;
2416 balanced = 0;
2417
2418 /*
2419 * Scan in the highmem->dma direction for the highest
2420 * zone which needs scanning
2421 */
2422 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2423 struct zone *zone = pgdat->node_zones + i;
2424
2425 if (!populated_zone(zone))
2426 continue;
2427
2428 if (zone->all_unreclaimable &&
2429 sc.priority != DEF_PRIORITY)
2430 continue;
2431
2432 /*
2433 * Do some background aging of the anon list, to give
2434 * pages a chance to be referenced before reclaiming.
2435 */
2436 age_active_anon(zone, &sc);
2437
2438 /*
2439 * If the number of buffer_heads in the machine
2440 * exceeds the maximum allowed level and this node
2441 * has a highmem zone, force kswapd to reclaim from
2442 * it to relieve lowmem pressure.
2443 */
2444 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2445 end_zone = i;
2446 break;
2447 }
2448
2449 if (!zone_watermark_ok_safe(zone, order,
2450 high_wmark_pages(zone), 0, 0)) {
2451 end_zone = i;
2452 break;
2453 } else {
2454 /* If balanced, clear the congested flag */
2455 zone_clear_flag(zone, ZONE_CONGESTED);
2456 }
2457 }
2458 if (i < 0)
2459 goto out;
2460
2461 for (i = 0; i <= end_zone; i++) {
2462 struct zone *zone = pgdat->node_zones + i;
2463
2464 lru_pages += zone_reclaimable_pages(zone);
2465 }
2466
2467 /*
2468 * Now scan the zone in the dma->highmem direction, stopping
2469 * at the last zone which needs scanning.
2470 *
2471 * We do this because the page allocator works in the opposite
2472 * direction. This prevents the page allocator from allocating
2473 * pages behind kswapd's direction of progress, which would
2474 * cause too much scanning of the lower zones.
2475 */
2476 for (i = 0; i <= end_zone; i++) {
2477 struct zone *zone = pgdat->node_zones + i;
2478 int nr_slab, testorder;
2479 unsigned long balance_gap;
2480
2481 if (!populated_zone(zone))
2482 continue;
2483
2484 if (zone->all_unreclaimable &&
2485 sc.priority != DEF_PRIORITY)
2486 continue;
2487
2488 sc.nr_scanned = 0;
2489
2490 nr_soft_scanned = 0;
2491 /*
2492 * Call soft limit reclaim before calling shrink_zone.
2493 */
2494 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2495 order, sc.gfp_mask,
2496 &nr_soft_scanned);
2497 sc.nr_reclaimed += nr_soft_reclaimed;
2498 total_scanned += nr_soft_scanned;
2499
2500 /*
2501 * We put equal pressure on every zone, unless
2502 * one zone has way too many pages free
2503 * already. The "too many pages" is defined
2504 * as the high wmark plus a "gap" where the
2505 * gap is either the low watermark or 1%
2506 * of the zone, whichever is smaller.
2507 */
2508 balance_gap = min(low_wmark_pages(zone),
2509 (zone->present_pages +
2510 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2511 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2512 /*
2513 * Kswapd reclaims only single pages with compaction
2514 * enabled. Trying too hard to reclaim until contiguous
2515 * free pages have become available can hurt performance
2516 * by evicting too much useful data from memory.
2517 * Do not reclaim more than needed for compaction.
2518 */
2519 testorder = order;
2520 if (COMPACTION_BUILD && order &&
2521 compaction_suitable(zone, order) !=
2522 COMPACT_SKIPPED)
2523 testorder = 0;
2524
2525 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2526 !zone_watermark_ok_safe(zone, testorder,
2527 high_wmark_pages(zone) + balance_gap,
2528 end_zone, 0)) {
2529 shrink_zone(zone, &sc);
2530
2531 reclaim_state->reclaimed_slab = 0;
2532 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2533 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2534 total_scanned += sc.nr_scanned;
2535
2536 if (nr_slab == 0 && !zone_reclaimable(zone))
2537 zone->all_unreclaimable = 1;
2538 }
2539
2540 /*
2541 * If we've done a decent amount of scanning and
2542 * the reclaim ratio is low, start doing writepage
2543 * even in laptop mode
2544 */
2545 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2546 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2547 sc.may_writepage = 1;
2548
2549 if (zone->all_unreclaimable) {
2550 if (end_zone && end_zone == i)
2551 end_zone--;
2552 continue;
2553 }
2554
2555 if (!zone_watermark_ok_safe(zone, testorder,
2556 high_wmark_pages(zone), end_zone, 0)) {
2557 all_zones_ok = 0;
2558 /*
2559 * We are still under min water mark. This
2560 * means that we have a GFP_ATOMIC allocation
2561 * failure risk. Hurry up!
2562 */
2563 if (!zone_watermark_ok_safe(zone, order,
2564 min_wmark_pages(zone), end_zone, 0))
2565 has_under_min_watermark_zone = 1;
2566 } else {
2567 /*
2568 * If a zone reaches its high watermark,
2569 * consider it to be no longer congested. It's
2570 * possible there are dirty pages backed by
2571 * congested BDIs but as pressure is relieved,
2572 * spectulatively avoid congestion waits
2573 */
2574 zone_clear_flag(zone, ZONE_CONGESTED);
2575 if (i <= *classzone_idx)
2576 balanced += zone->present_pages;
2577 }
2578
2579 }
2580 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2581 break; /* kswapd: all done */
2582 /*
2583 * OK, kswapd is getting into trouble. Take a nap, then take
2584 * another pass across the zones.
2585 */
2586 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2587 if (has_under_min_watermark_zone)
2588 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2589 else
2590 congestion_wait(BLK_RW_ASYNC, HZ/10);
2591 }
2592
2593 /*
2594 * We do this so kswapd doesn't build up large priorities for
2595 * example when it is freeing in parallel with allocators. It
2596 * matches the direct reclaim path behaviour in terms of impact
2597 * on zone->*_priority.
2598 */
2599 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2600 break;
2601 } while (--sc.priority >= 0);
2602out:
2603
2604 /*
2605 * order-0: All zones must meet high watermark for a balanced node
2606 * high-order: Balanced zones must make up at least 25% of the node
2607 * for the node to be balanced
2608 */
2609 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2610 cond_resched();
2611
2612 try_to_freeze();
2613
2614 /*
2615 * Fragmentation may mean that the system cannot be
2616 * rebalanced for high-order allocations in all zones.
2617 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2618 * it means the zones have been fully scanned and are still
2619 * not balanced. For high-order allocations, there is
2620 * little point trying all over again as kswapd may
2621 * infinite loop.
2622 *
2623 * Instead, recheck all watermarks at order-0 as they
2624 * are the most important. If watermarks are ok, kswapd will go
2625 * back to sleep. High-order users can still perform direct
2626 * reclaim if they wish.
2627 */
2628 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2629 order = sc.order = 0;
2630
2631 goto loop_again;
2632 }
2633
2634 /*
2635 * If kswapd was reclaiming at a higher order, it has the option of
2636 * sleeping without all zones being balanced. Before it does, it must
2637 * ensure that the watermarks for order-0 on *all* zones are met and
2638 * that the congestion flags are cleared. The congestion flag must
2639 * be cleared as kswapd is the only mechanism that clears the flag
2640 * and it is potentially going to sleep here.
2641 */
2642 if (order) {
2643 int zones_need_compaction = 1;
2644
2645 for (i = 0; i <= end_zone; i++) {
2646 struct zone *zone = pgdat->node_zones + i;
2647
2648 if (!populated_zone(zone))
2649 continue;
2650
2651 if (zone->all_unreclaimable &&
2652 sc.priority != DEF_PRIORITY)
2653 continue;
2654
2655 /* Would compaction fail due to lack of free memory? */
2656 if (COMPACTION_BUILD &&
2657 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2658 goto loop_again;
2659
2660 /* Confirm the zone is balanced for order-0 */
2661 if (!zone_watermark_ok(zone, 0,
2662 high_wmark_pages(zone), 0, 0)) {
2663 order = sc.order = 0;
2664 goto loop_again;
2665 }
2666
2667 /* Check if the memory needs to be defragmented. */
2668 if (zone_watermark_ok(zone, order,
2669 low_wmark_pages(zone), *classzone_idx, 0))
2670 zones_need_compaction = 0;
2671
2672 /* If balanced, clear the congested flag */
2673 zone_clear_flag(zone, ZONE_CONGESTED);
2674 }
2675
2676 if (zones_need_compaction)
2677 compact_pgdat(pgdat, order);
2678 }
2679
2680 /*
2681 * Return the order we were reclaiming at so sleeping_prematurely()
2682 * makes a decision on the order we were last reclaiming at. However,
2683 * if another caller entered the allocator slow path while kswapd
2684 * was awake, order will remain at the higher level
2685 */
2686 *classzone_idx = end_zone;
2687 return order;
2688}
2689
2690static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2691{
2692 long remaining = 0;
2693 DEFINE_WAIT(wait);
2694
2695 if (freezing(current) || kthread_should_stop())
2696 return;
2697
2698 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2699
2700 /* Try to sleep for a short interval */
2701 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2702 remaining = schedule_timeout(HZ/10);
2703 finish_wait(&pgdat->kswapd_wait, &wait);
2704 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2705 }
2706
2707 /*
2708 * After a short sleep, check if it was a premature sleep. If not, then
2709 * go fully to sleep until explicitly woken up.
2710 */
2711 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2712 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2713
2714 /*
2715 * vmstat counters are not perfectly accurate and the estimated
2716 * value for counters such as NR_FREE_PAGES can deviate from the
2717 * true value by nr_online_cpus * threshold. To avoid the zone
2718 * watermarks being breached while under pressure, we reduce the
2719 * per-cpu vmstat threshold while kswapd is awake and restore
2720 * them before going back to sleep.
2721 */
2722 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2723
2724 if (!kthread_should_stop())
2725 schedule();
2726
2727 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2728 } else {
2729 if (remaining)
2730 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2731 else
2732 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2733 }
2734 finish_wait(&pgdat->kswapd_wait, &wait);
2735}
2736
2737/*
2738 * The background pageout daemon, started as a kernel thread
2739 * from the init process.
2740 *
2741 * This basically trickles out pages so that we have _some_
2742 * free memory available even if there is no other activity
2743 * that frees anything up. This is needed for things like routing
2744 * etc, where we otherwise might have all activity going on in
2745 * asynchronous contexts that cannot page things out.
2746 *
2747 * If there are applications that are active memory-allocators
2748 * (most normal use), this basically shouldn't matter.
2749 */
2750static int kswapd(void *p)
2751{
2752 unsigned long order, new_order;
2753 unsigned balanced_order;
2754 int classzone_idx, new_classzone_idx;
2755 int balanced_classzone_idx;
2756 pg_data_t *pgdat = (pg_data_t*)p;
2757 struct task_struct *tsk = current;
2758
2759 struct reclaim_state reclaim_state = {
2760 .reclaimed_slab = 0,
2761 };
2762 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2763
2764 lockdep_set_current_reclaim_state(GFP_KERNEL);
2765
2766 if (!cpumask_empty(cpumask))
2767 set_cpus_allowed_ptr(tsk, cpumask);
2768 current->reclaim_state = &reclaim_state;
2769
2770 /*
2771 * Tell the memory management that we're a "memory allocator",
2772 * and that if we need more memory we should get access to it
2773 * regardless (see "__alloc_pages()"). "kswapd" should
2774 * never get caught in the normal page freeing logic.
2775 *
2776 * (Kswapd normally doesn't need memory anyway, but sometimes
2777 * you need a small amount of memory in order to be able to
2778 * page out something else, and this flag essentially protects
2779 * us from recursively trying to free more memory as we're
2780 * trying to free the first piece of memory in the first place).
2781 */
2782 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2783 set_freezable();
2784
2785 order = new_order = 0;
2786 balanced_order = 0;
2787 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2788 balanced_classzone_idx = classzone_idx;
2789 for ( ; ; ) {
2790 int ret;
2791
2792 /*
2793 * If the last balance_pgdat was unsuccessful it's unlikely a
2794 * new request of a similar or harder type will succeed soon
2795 * so consider going to sleep on the basis we reclaimed at
2796 */
2797 if (balanced_classzone_idx >= new_classzone_idx &&
2798 balanced_order == new_order) {
2799 new_order = pgdat->kswapd_max_order;
2800 new_classzone_idx = pgdat->classzone_idx;
2801 pgdat->kswapd_max_order = 0;
2802 pgdat->classzone_idx = pgdat->nr_zones - 1;
2803 }
2804
2805 if (order < new_order || classzone_idx > new_classzone_idx) {
2806 /*
2807 * Don't sleep if someone wants a larger 'order'
2808 * allocation or has tigher zone constraints
2809 */
2810 order = new_order;
2811 classzone_idx = new_classzone_idx;
2812 } else {
2813 kswapd_try_to_sleep(pgdat, balanced_order,
2814 balanced_classzone_idx);
2815 order = pgdat->kswapd_max_order;
2816 classzone_idx = pgdat->classzone_idx;
2817 new_order = order;
2818 new_classzone_idx = classzone_idx;
2819 pgdat->kswapd_max_order = 0;
2820 pgdat->classzone_idx = pgdat->nr_zones - 1;
2821 }
2822
2823 ret = try_to_freeze();
2824 if (kthread_should_stop())
2825 break;
2826
2827 /*
2828 * We can speed up thawing tasks if we don't call balance_pgdat
2829 * after returning from the refrigerator
2830 */
2831 if (!ret) {
2832 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2833 balanced_classzone_idx = classzone_idx;
2834 balanced_order = balance_pgdat(pgdat, order,
2835 &balanced_classzone_idx);
2836 }
2837 }
2838 return 0;
2839}
2840
2841/*
2842 * A zone is low on free memory, so wake its kswapd task to service it.
2843 */
2844void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2845{
2846 pg_data_t *pgdat;
2847
2848 if (!populated_zone(zone))
2849 return;
2850
2851 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2852 return;
2853 pgdat = zone->zone_pgdat;
2854 if (pgdat->kswapd_max_order < order) {
2855 pgdat->kswapd_max_order = order;
2856 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2857 }
2858 if (!waitqueue_active(&pgdat->kswapd_wait))
2859 return;
2860 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2861 return;
2862
2863 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2864 wake_up_interruptible(&pgdat->kswapd_wait);
2865}
2866
2867/*
2868 * The reclaimable count would be mostly accurate.
2869 * The less reclaimable pages may be
2870 * - mlocked pages, which will be moved to unevictable list when encountered
2871 * - mapped pages, which may require several travels to be reclaimed
2872 * - dirty pages, which is not "instantly" reclaimable
2873 */
2874unsigned long global_reclaimable_pages(void)
2875{
2876 int nr;
2877
2878 nr = global_page_state(NR_ACTIVE_FILE) +
2879 global_page_state(NR_INACTIVE_FILE);
2880
2881 if (nr_swap_pages > 0)
2882 nr += global_page_state(NR_ACTIVE_ANON) +
2883 global_page_state(NR_INACTIVE_ANON);
2884
2885 return nr;
2886}
2887
2888unsigned long zone_reclaimable_pages(struct zone *zone)
2889{
2890 int nr;
2891
2892 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2893 zone_page_state(zone, NR_INACTIVE_FILE);
2894
2895 if (nr_swap_pages > 0)
2896 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2897 zone_page_state(zone, NR_INACTIVE_ANON);
2898
2899 return nr;
2900}
2901
2902#ifdef CONFIG_HIBERNATION
2903/*
2904 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2905 * freed pages.
2906 *
2907 * Rather than trying to age LRUs the aim is to preserve the overall
2908 * LRU order by reclaiming preferentially
2909 * inactive > active > active referenced > active mapped
2910 */
2911unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2912{
2913 struct reclaim_state reclaim_state;
2914 struct scan_control sc = {
2915 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2916 .may_swap = 1,
2917 .may_unmap = 1,
2918 .may_writepage = 1,
2919 .nr_to_reclaim = nr_to_reclaim,
2920 .hibernation_mode = 1,
2921 .order = 0,
2922 .priority = DEF_PRIORITY,
2923 };
2924 struct shrink_control shrink = {
2925 .gfp_mask = sc.gfp_mask,
2926 };
2927 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2928 struct task_struct *p = current;
2929 unsigned long nr_reclaimed;
2930
2931 p->flags |= PF_MEMALLOC;
2932 lockdep_set_current_reclaim_state(sc.gfp_mask);
2933 reclaim_state.reclaimed_slab = 0;
2934 p->reclaim_state = &reclaim_state;
2935
2936 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2937
2938 p->reclaim_state = NULL;
2939 lockdep_clear_current_reclaim_state();
2940 p->flags &= ~PF_MEMALLOC;
2941
2942 return nr_reclaimed;
2943}
2944#endif /* CONFIG_HIBERNATION */
2945
2946/* It's optimal to keep kswapds on the same CPUs as their memory, but
2947 not required for correctness. So if the last cpu in a node goes
2948 away, we get changed to run anywhere: as the first one comes back,
2949 restore their cpu bindings. */
2950static int __devinit cpu_callback(struct notifier_block *nfb,
2951 unsigned long action, void *hcpu)
2952{
2953 int nid;
2954
2955 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2956 for_each_node_state(nid, N_HIGH_MEMORY) {
2957 pg_data_t *pgdat = NODE_DATA(nid);
2958 const struct cpumask *mask;
2959
2960 mask = cpumask_of_node(pgdat->node_id);
2961
2962 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2963 /* One of our CPUs online: restore mask */
2964 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2965 }
2966 }
2967 return NOTIFY_OK;
2968}
2969
2970/*
2971 * This kswapd start function will be called by init and node-hot-add.
2972 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2973 */
2974int kswapd_run(int nid)
2975{
2976 pg_data_t *pgdat = NODE_DATA(nid);
2977 int ret = 0;
2978
2979 if (pgdat->kswapd)
2980 return 0;
2981
2982 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2983 if (IS_ERR(pgdat->kswapd)) {
2984 /* failure at boot is fatal */
2985 BUG_ON(system_state == SYSTEM_BOOTING);
2986 printk("Failed to start kswapd on node %d\n",nid);
2987 ret = -1;
2988 }
2989 return ret;
2990}
2991
2992/*
2993 * Called by memory hotplug when all memory in a node is offlined. Caller must
2994 * hold lock_memory_hotplug().
2995 */
2996void kswapd_stop(int nid)
2997{
2998 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2999
3000 if (kswapd) {
3001 kthread_stop(kswapd);
3002 NODE_DATA(nid)->kswapd = NULL;
3003 }
3004}
3005
3006static int __init kswapd_init(void)
3007{
3008 int nid;
3009
3010 swap_setup();
3011 for_each_node_state(nid, N_HIGH_MEMORY)
3012 kswapd_run(nid);
3013 hotcpu_notifier(cpu_callback, 0);
3014 return 0;
3015}
3016
3017module_init(kswapd_init)
3018
3019#ifdef CONFIG_NUMA
3020/*
3021 * Zone reclaim mode
3022 *
3023 * If non-zero call zone_reclaim when the number of free pages falls below
3024 * the watermarks.
3025 */
3026int zone_reclaim_mode __read_mostly;
3027
3028#define RECLAIM_OFF 0
3029#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3030#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3031#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3032
3033/*
3034 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3035 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3036 * a zone.
3037 */
3038#define ZONE_RECLAIM_PRIORITY 4
3039
3040/*
3041 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3042 * occur.
3043 */
3044int sysctl_min_unmapped_ratio = 1;
3045
3046/*
3047 * If the number of slab pages in a zone grows beyond this percentage then
3048 * slab reclaim needs to occur.
3049 */
3050int sysctl_min_slab_ratio = 5;
3051
3052static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3053{
3054 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3055 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3056 zone_page_state(zone, NR_ACTIVE_FILE);
3057
3058 /*
3059 * It's possible for there to be more file mapped pages than
3060 * accounted for by the pages on the file LRU lists because
3061 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3062 */
3063 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3064}
3065
3066/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3067static long zone_pagecache_reclaimable(struct zone *zone)
3068{
3069 long nr_pagecache_reclaimable;
3070 long delta = 0;
3071
3072 /*
3073 * If RECLAIM_SWAP is set, then all file pages are considered
3074 * potentially reclaimable. Otherwise, we have to worry about
3075 * pages like swapcache and zone_unmapped_file_pages() provides
3076 * a better estimate
3077 */
3078 if (zone_reclaim_mode & RECLAIM_SWAP)
3079 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3080 else
3081 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3082
3083 /* If we can't clean pages, remove dirty pages from consideration */
3084 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3085 delta += zone_page_state(zone, NR_FILE_DIRTY);
3086
3087 /* Watch for any possible underflows due to delta */
3088 if (unlikely(delta > nr_pagecache_reclaimable))
3089 delta = nr_pagecache_reclaimable;
3090
3091 return nr_pagecache_reclaimable - delta;
3092}
3093
3094/*
3095 * Try to free up some pages from this zone through reclaim.
3096 */
3097static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3098{
3099 /* Minimum pages needed in order to stay on node */
3100 const unsigned long nr_pages = 1 << order;
3101 struct task_struct *p = current;
3102 struct reclaim_state reclaim_state;
3103 struct scan_control sc = {
3104 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3105 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3106 .may_swap = 1,
3107 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3108 SWAP_CLUSTER_MAX),
3109 .gfp_mask = gfp_mask,
3110 .order = order,
3111 .priority = ZONE_RECLAIM_PRIORITY,
3112 };
3113 struct shrink_control shrink = {
3114 .gfp_mask = sc.gfp_mask,
3115 };
3116 unsigned long nr_slab_pages0, nr_slab_pages1;
3117
3118 cond_resched();
3119 /*
3120 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3121 * and we also need to be able to write out pages for RECLAIM_WRITE
3122 * and RECLAIM_SWAP.
3123 */
3124 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3125 lockdep_set_current_reclaim_state(gfp_mask);
3126 reclaim_state.reclaimed_slab = 0;
3127 p->reclaim_state = &reclaim_state;
3128
3129 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3130 /*
3131 * Free memory by calling shrink zone with increasing
3132 * priorities until we have enough memory freed.
3133 */
3134 do {
3135 shrink_zone(zone, &sc);
3136 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3137 }
3138
3139 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3140 if (nr_slab_pages0 > zone->min_slab_pages) {
3141 /*
3142 * shrink_slab() does not currently allow us to determine how
3143 * many pages were freed in this zone. So we take the current
3144 * number of slab pages and shake the slab until it is reduced
3145 * by the same nr_pages that we used for reclaiming unmapped
3146 * pages.
3147 *
3148 * Note that shrink_slab will free memory on all zones and may
3149 * take a long time.
3150 */
3151 for (;;) {
3152 unsigned long lru_pages = zone_reclaimable_pages(zone);
3153
3154 /* No reclaimable slab or very low memory pressure */
3155 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3156 break;
3157
3158 /* Freed enough memory */
3159 nr_slab_pages1 = zone_page_state(zone,
3160 NR_SLAB_RECLAIMABLE);
3161 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3162 break;
3163 }
3164
3165 /*
3166 * Update nr_reclaimed by the number of slab pages we
3167 * reclaimed from this zone.
3168 */
3169 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3170 if (nr_slab_pages1 < nr_slab_pages0)
3171 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3172 }
3173
3174 p->reclaim_state = NULL;
3175 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3176 lockdep_clear_current_reclaim_state();
3177 return sc.nr_reclaimed >= nr_pages;
3178}
3179
3180int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3181{
3182 int node_id;
3183 int ret;
3184
3185 /*
3186 * Zone reclaim reclaims unmapped file backed pages and
3187 * slab pages if we are over the defined limits.
3188 *
3189 * A small portion of unmapped file backed pages is needed for
3190 * file I/O otherwise pages read by file I/O will be immediately
3191 * thrown out if the zone is overallocated. So we do not reclaim
3192 * if less than a specified percentage of the zone is used by
3193 * unmapped file backed pages.
3194 */
3195 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3196 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3197 return ZONE_RECLAIM_FULL;
3198
3199 if (zone->all_unreclaimable)
3200 return ZONE_RECLAIM_FULL;
3201
3202 /*
3203 * Do not scan if the allocation should not be delayed.
3204 */
3205 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3206 return ZONE_RECLAIM_NOSCAN;
3207
3208 /*
3209 * Only run zone reclaim on the local zone or on zones that do not
3210 * have associated processors. This will favor the local processor
3211 * over remote processors and spread off node memory allocations
3212 * as wide as possible.
3213 */
3214 node_id = zone_to_nid(zone);
3215 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3216 return ZONE_RECLAIM_NOSCAN;
3217
3218 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3219 return ZONE_RECLAIM_NOSCAN;
3220
3221 ret = __zone_reclaim(zone, gfp_mask, order);
3222 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3223
3224 if (!ret)
3225 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3226
3227 return ret;
3228}
3229#endif
3230
3231/*
3232 * page_evictable - test whether a page is evictable
3233 * @page: the page to test
3234 * @vma: the VMA in which the page is or will be mapped, may be NULL
3235 *
3236 * Test whether page is evictable--i.e., should be placed on active/inactive
3237 * lists vs unevictable list. The vma argument is !NULL when called from the
3238 * fault path to determine how to instantate a new page.
3239 *
3240 * Reasons page might not be evictable:
3241 * (1) page's mapping marked unevictable
3242 * (2) page is part of an mlocked VMA
3243 *
3244 */
3245int page_evictable(struct page *page, struct vm_area_struct *vma)
3246{
3247
3248 if (mapping_unevictable(page_mapping(page)))
3249 return 0;
3250
3251 if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3252 return 0;
3253
3254 return 1;
3255}
3256
3257#ifdef CONFIG_SHMEM
3258/**
3259 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3260 * @pages: array of pages to check
3261 * @nr_pages: number of pages to check
3262 *
3263 * Checks pages for evictability and moves them to the appropriate lru list.
3264 *
3265 * This function is only used for SysV IPC SHM_UNLOCK.
3266 */
3267void check_move_unevictable_pages(struct page **pages, int nr_pages)
3268{
3269 struct lruvec *lruvec;
3270 struct zone *zone = NULL;
3271 int pgscanned = 0;
3272 int pgrescued = 0;
3273 int i;
3274
3275 for (i = 0; i < nr_pages; i++) {
3276 struct page *page = pages[i];
3277 struct zone *pagezone;
3278
3279 pgscanned++;
3280 pagezone = page_zone(page);
3281 if (pagezone != zone) {
3282 if (zone)
3283 spin_unlock_irq(&zone->lru_lock);
3284 zone = pagezone;
3285 spin_lock_irq(&zone->lru_lock);
3286 }
3287 lruvec = mem_cgroup_page_lruvec(page, zone);
3288
3289 if (!PageLRU(page) || !PageUnevictable(page))
3290 continue;
3291
3292 if (page_evictable(page, NULL)) {
3293 enum lru_list lru = page_lru_base_type(page);
3294
3295 VM_BUG_ON(PageActive(page));
3296 ClearPageUnevictable(page);
3297 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3298 add_page_to_lru_list(page, lruvec, lru);
3299 pgrescued++;
3300 }
3301 }
3302
3303 if (zone) {
3304 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3305 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3306 spin_unlock_irq(&zone->lru_lock);
3307 }
3308}
3309#endif /* CONFIG_SHMEM */
3310
3311static void warn_scan_unevictable_pages(void)
3312{
3313 printk_once(KERN_WARNING
3314 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3315 "disabled for lack of a legitimate use case. If you have "
3316 "one, please send an email to linux-mm@kvack.org.\n",
3317 current->comm);
3318}
3319
3320/*
3321 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3322 * all nodes' unevictable lists for evictable pages
3323 */
3324unsigned long scan_unevictable_pages;
3325
3326int scan_unevictable_handler(struct ctl_table *table, int write,
3327 void __user *buffer,
3328 size_t *length, loff_t *ppos)
3329{
3330 warn_scan_unevictable_pages();
3331 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3332 scan_unevictable_pages = 0;
3333 return 0;
3334}
3335
3336#ifdef CONFIG_NUMA
3337/*
3338 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3339 * a specified node's per zone unevictable lists for evictable pages.
3340 */
3341
3342static ssize_t read_scan_unevictable_node(struct device *dev,
3343 struct device_attribute *attr,
3344 char *buf)
3345{
3346 warn_scan_unevictable_pages();
3347 return sprintf(buf, "0\n"); /* always zero; should fit... */
3348}
3349
3350static ssize_t write_scan_unevictable_node(struct device *dev,
3351 struct device_attribute *attr,
3352 const char *buf, size_t count)
3353{
3354 warn_scan_unevictable_pages();
3355 return 1;
3356}
3357
3358
3359static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3360 read_scan_unevictable_node,
3361 write_scan_unevictable_node);
3362
3363int scan_unevictable_register_node(struct node *node)
3364{
3365 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3366}
3367
3368void scan_unevictable_unregister_node(struct node *node)
3369{
3370 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3371}
3372#endif