Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
17#include <linux/mm.h>
18#include <linux/sched/mm.h>
19#include <linux/module.h>
20#include <linux/gfp.h>
21#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
26#include <linux/vmpressure.h>
27#include <linux/vmstat.h>
28#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
34#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
39#include <linux/compaction.h>
40#include <linux/notifier.h>
41#include <linux/rwsem.h>
42#include <linux/delay.h>
43#include <linux/kthread.h>
44#include <linux/freezer.h>
45#include <linux/memcontrol.h>
46#include <linux/delayacct.h>
47#include <linux/sysctl.h>
48#include <linux/oom.h>
49#include <linux/pagevec.h>
50#include <linux/prefetch.h>
51#include <linux/printk.h>
52#include <linux/dax.h>
53#include <linux/psi.h>
54
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
59#include <linux/balloon_compaction.h>
60
61#include "internal.h"
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
66struct scan_control {
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
70 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
75
76 /*
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
81
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
84
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
87
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
90
91 /*
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
95 */
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
98
99 unsigned int hibernation_mode:1;
100
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
103
104 /* Allocation order */
105 s8 order;
106
107 /* Scan (total_size >> priority) pages at once */
108 s8 priority;
109
110 /* The highest zone to isolate pages for reclaim from */
111 s8 reclaim_idx;
112
113 /* This context's GFP mask */
114 gfp_t gfp_mask;
115
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned;
118
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed;
121
122 struct {
123 unsigned int dirty;
124 unsigned int unqueued_dirty;
125 unsigned int congested;
126 unsigned int writeback;
127 unsigned int immediate;
128 unsigned int file_taken;
129 unsigned int taken;
130 } nr;
131
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state;
134};
135
136#ifdef ARCH_HAS_PREFETCH
137#define prefetch_prev_lru_page(_page, _base, _field) \
138 do { \
139 if ((_page)->lru.prev != _base) { \
140 struct page *prev; \
141 \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
144 } \
145 } while (0)
146#else
147#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148#endif
149
150#ifdef ARCH_HAS_PREFETCHW
151#define prefetchw_prev_lru_page(_page, _base, _field) \
152 do { \
153 if ((_page)->lru.prev != _base) { \
154 struct page *prev; \
155 \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
158 } \
159 } while (0)
160#else
161#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162#endif
163
164/*
165 * From 0 .. 100. Higher means more swappy.
166 */
167int vm_swappiness = 60;
168/*
169 * The total number of pages which are beyond the high watermark within all
170 * zones.
171 */
172unsigned long vm_total_pages;
173
174static void set_task_reclaim_state(struct task_struct *task,
175 struct reclaim_state *rs)
176{
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs && task->reclaim_state);
179
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs && !task->reclaim_state);
182
183 task->reclaim_state = rs;
184}
185
186static LIST_HEAD(shrinker_list);
187static DECLARE_RWSEM(shrinker_rwsem);
188
189#ifdef CONFIG_MEMCG
190/*
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
197 *
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
200 */
201#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202
203static DEFINE_IDR(shrinker_idr);
204static int shrinker_nr_max;
205
206static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207{
208 int id, ret = -ENOMEM;
209
210 down_write(&shrinker_rwsem);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
213 if (id < 0)
214 goto unlock;
215
216 if (id >= shrinker_nr_max) {
217 if (memcg_expand_shrinker_maps(id)) {
218 idr_remove(&shrinker_idr, id);
219 goto unlock;
220 }
221
222 shrinker_nr_max = id + 1;
223 }
224 shrinker->id = id;
225 ret = 0;
226unlock:
227 up_write(&shrinker_rwsem);
228 return ret;
229}
230
231static void unregister_memcg_shrinker(struct shrinker *shrinker)
232{
233 int id = shrinker->id;
234
235 BUG_ON(id < 0);
236
237 down_write(&shrinker_rwsem);
238 idr_remove(&shrinker_idr, id);
239 up_write(&shrinker_rwsem);
240}
241
242static bool global_reclaim(struct scan_control *sc)
243{
244 return !sc->target_mem_cgroup;
245}
246
247/**
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
250 *
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
256 *
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
259 */
260static bool sane_reclaim(struct scan_control *sc)
261{
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
263
264 if (!memcg)
265 return true;
266#ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
268 return true;
269#endif
270 return false;
271}
272
273static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
275 bool congested)
276{
277 struct mem_cgroup_per_node *mn;
278
279 if (!memcg)
280 return;
281
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
284}
285
286static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
288{
289 struct mem_cgroup_per_node *mn;
290
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
293
294}
295#else
296static int prealloc_memcg_shrinker(struct shrinker *shrinker)
297{
298 return 0;
299}
300
301static void unregister_memcg_shrinker(struct shrinker *shrinker)
302{
303}
304
305static bool global_reclaim(struct scan_control *sc)
306{
307 return true;
308}
309
310static bool sane_reclaim(struct scan_control *sc)
311{
312 return true;
313}
314
315static inline void set_memcg_congestion(struct pglist_data *pgdat,
316 struct mem_cgroup *memcg, bool congested)
317{
318}
319
320static inline bool memcg_congested(struct pglist_data *pgdat,
321 struct mem_cgroup *memcg)
322{
323 return false;
324
325}
326#endif
327
328/*
329 * This misses isolated pages which are not accounted for to save counters.
330 * As the data only determines if reclaim or compaction continues, it is
331 * not expected that isolated pages will be a dominating factor.
332 */
333unsigned long zone_reclaimable_pages(struct zone *zone)
334{
335 unsigned long nr;
336
337 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
338 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
339 if (get_nr_swap_pages() > 0)
340 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
341 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
342
343 return nr;
344}
345
346/**
347 * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 * @lruvec: lru vector
349 * @lru: lru to use
350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
351 */
352unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
353{
354 unsigned long lru_size = 0;
355 int zid;
356
357 if (!mem_cgroup_disabled()) {
358 for (zid = 0; zid < MAX_NR_ZONES; zid++)
359 lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
360 } else
361 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
362
363 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
364 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
365 unsigned long size;
366
367 if (!managed_zone(zone))
368 continue;
369
370 if (!mem_cgroup_disabled())
371 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
372 else
373 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
374 NR_ZONE_LRU_BASE + lru);
375 lru_size -= min(size, lru_size);
376 }
377
378 return lru_size;
379
380}
381
382/*
383 * Add a shrinker callback to be called from the vm.
384 */
385int prealloc_shrinker(struct shrinker *shrinker)
386{
387 unsigned int size = sizeof(*shrinker->nr_deferred);
388
389 if (shrinker->flags & SHRINKER_NUMA_AWARE)
390 size *= nr_node_ids;
391
392 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
393 if (!shrinker->nr_deferred)
394 return -ENOMEM;
395
396 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
397 if (prealloc_memcg_shrinker(shrinker))
398 goto free_deferred;
399 }
400
401 return 0;
402
403free_deferred:
404 kfree(shrinker->nr_deferred);
405 shrinker->nr_deferred = NULL;
406 return -ENOMEM;
407}
408
409void free_prealloced_shrinker(struct shrinker *shrinker)
410{
411 if (!shrinker->nr_deferred)
412 return;
413
414 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
415 unregister_memcg_shrinker(shrinker);
416
417 kfree(shrinker->nr_deferred);
418 shrinker->nr_deferred = NULL;
419}
420
421void register_shrinker_prepared(struct shrinker *shrinker)
422{
423 down_write(&shrinker_rwsem);
424 list_add_tail(&shrinker->list, &shrinker_list);
425#ifdef CONFIG_MEMCG_KMEM
426 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
427 idr_replace(&shrinker_idr, shrinker, shrinker->id);
428#endif
429 up_write(&shrinker_rwsem);
430}
431
432int register_shrinker(struct shrinker *shrinker)
433{
434 int err = prealloc_shrinker(shrinker);
435
436 if (err)
437 return err;
438 register_shrinker_prepared(shrinker);
439 return 0;
440}
441EXPORT_SYMBOL(register_shrinker);
442
443/*
444 * Remove one
445 */
446void unregister_shrinker(struct shrinker *shrinker)
447{
448 if (!shrinker->nr_deferred)
449 return;
450 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
451 unregister_memcg_shrinker(shrinker);
452 down_write(&shrinker_rwsem);
453 list_del(&shrinker->list);
454 up_write(&shrinker_rwsem);
455 kfree(shrinker->nr_deferred);
456 shrinker->nr_deferred = NULL;
457}
458EXPORT_SYMBOL(unregister_shrinker);
459
460#define SHRINK_BATCH 128
461
462static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
463 struct shrinker *shrinker, int priority)
464{
465 unsigned long freed = 0;
466 unsigned long long delta;
467 long total_scan;
468 long freeable;
469 long nr;
470 long new_nr;
471 int nid = shrinkctl->nid;
472 long batch_size = shrinker->batch ? shrinker->batch
473 : SHRINK_BATCH;
474 long scanned = 0, next_deferred;
475
476 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
477 nid = 0;
478
479 freeable = shrinker->count_objects(shrinker, shrinkctl);
480 if (freeable == 0 || freeable == SHRINK_EMPTY)
481 return freeable;
482
483 /*
484 * copy the current shrinker scan count into a local variable
485 * and zero it so that other concurrent shrinker invocations
486 * don't also do this scanning work.
487 */
488 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
489
490 total_scan = nr;
491 if (shrinker->seeks) {
492 delta = freeable >> priority;
493 delta *= 4;
494 do_div(delta, shrinker->seeks);
495 } else {
496 /*
497 * These objects don't require any IO to create. Trim
498 * them aggressively under memory pressure to keep
499 * them from causing refetches in the IO caches.
500 */
501 delta = freeable / 2;
502 }
503
504 total_scan += delta;
505 if (total_scan < 0) {
506 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 shrinker->scan_objects, total_scan);
508 total_scan = freeable;
509 next_deferred = nr;
510 } else
511 next_deferred = total_scan;
512
513 /*
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
520 * memory.
521 *
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
524 */
525 if (delta < freeable / 4)
526 total_scan = min(total_scan, freeable / 2);
527
528 /*
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
531 * freeable entries.
532 */
533 if (total_scan > freeable * 2)
534 total_scan = freeable * 2;
535
536 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 freeable, delta, total_scan, priority);
538
539 /*
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
546 * batch_size.
547 *
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
552 * possible.
553 */
554 while (total_scan >= batch_size ||
555 total_scan >= freeable) {
556 unsigned long ret;
557 unsigned long nr_to_scan = min(batch_size, total_scan);
558
559 shrinkctl->nr_to_scan = nr_to_scan;
560 shrinkctl->nr_scanned = nr_to_scan;
561 ret = shrinker->scan_objects(shrinker, shrinkctl);
562 if (ret == SHRINK_STOP)
563 break;
564 freed += ret;
565
566 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 total_scan -= shrinkctl->nr_scanned;
568 scanned += shrinkctl->nr_scanned;
569
570 cond_resched();
571 }
572
573 if (next_deferred >= scanned)
574 next_deferred -= scanned;
575 else
576 next_deferred = 0;
577 /*
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
581 */
582 if (next_deferred > 0)
583 new_nr = atomic_long_add_return(next_deferred,
584 &shrinker->nr_deferred[nid]);
585 else
586 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
587
588 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
589 return freed;
590}
591
592#ifdef CONFIG_MEMCG
593static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 struct mem_cgroup *memcg, int priority)
595{
596 struct memcg_shrinker_map *map;
597 unsigned long ret, freed = 0;
598 int i;
599
600 if (!mem_cgroup_online(memcg))
601 return 0;
602
603 if (!down_read_trylock(&shrinker_rwsem))
604 return 0;
605
606 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
607 true);
608 if (unlikely(!map))
609 goto unlock;
610
611 for_each_set_bit(i, map->map, shrinker_nr_max) {
612 struct shrink_control sc = {
613 .gfp_mask = gfp_mask,
614 .nid = nid,
615 .memcg = memcg,
616 };
617 struct shrinker *shrinker;
618
619 shrinker = idr_find(&shrinker_idr, i);
620 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
621 if (!shrinker)
622 clear_bit(i, map->map);
623 continue;
624 }
625
626 /* Call non-slab shrinkers even though kmem is disabled */
627 if (!memcg_kmem_enabled() &&
628 !(shrinker->flags & SHRINKER_NONSLAB))
629 continue;
630
631 ret = do_shrink_slab(&sc, shrinker, priority);
632 if (ret == SHRINK_EMPTY) {
633 clear_bit(i, map->map);
634 /*
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
643 *
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
646 * <MB> <MB>
647 * set_bit() do_shrink_slab()
648 */
649 smp_mb__after_atomic();
650 ret = do_shrink_slab(&sc, shrinker, priority);
651 if (ret == SHRINK_EMPTY)
652 ret = 0;
653 else
654 memcg_set_shrinker_bit(memcg, nid, i);
655 }
656 freed += ret;
657
658 if (rwsem_is_contended(&shrinker_rwsem)) {
659 freed = freed ? : 1;
660 break;
661 }
662 }
663unlock:
664 up_read(&shrinker_rwsem);
665 return freed;
666}
667#else /* CONFIG_MEMCG */
668static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 struct mem_cgroup *memcg, int priority)
670{
671 return 0;
672}
673#endif /* CONFIG_MEMCG */
674
675/**
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
681 *
682 * Call the shrink functions to age shrinkable caches.
683 *
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
686 *
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
689 *
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
692 *
693 * Returns the number of reclaimed slab objects.
694 */
695static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 struct mem_cgroup *memcg,
697 int priority)
698{
699 unsigned long ret, freed = 0;
700 struct shrinker *shrinker;
701
702 /*
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
707 * oom.
708 */
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
711
712 if (!down_read_trylock(&shrinker_rwsem))
713 goto out;
714
715 list_for_each_entry(shrinker, &shrinker_list, list) {
716 struct shrink_control sc = {
717 .gfp_mask = gfp_mask,
718 .nid = nid,
719 .memcg = memcg,
720 };
721
722 ret = do_shrink_slab(&sc, shrinker, priority);
723 if (ret == SHRINK_EMPTY)
724 ret = 0;
725 freed += ret;
726 /*
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
730 */
731 if (rwsem_is_contended(&shrinker_rwsem)) {
732 freed = freed ? : 1;
733 break;
734 }
735 }
736
737 up_read(&shrinker_rwsem);
738out:
739 cond_resched();
740 return freed;
741}
742
743void drop_slab_node(int nid)
744{
745 unsigned long freed;
746
747 do {
748 struct mem_cgroup *memcg = NULL;
749
750 freed = 0;
751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
752 do {
753 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 } while (freed > 10);
756}
757
758void drop_slab(void)
759{
760 int nid;
761
762 for_each_online_node(nid)
763 drop_slab_node(nid);
764}
765
766static inline int is_page_cache_freeable(struct page *page)
767{
768 /*
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
772 */
773 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
774 HPAGE_PMD_NR : 1;
775 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
776}
777
778static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
779{
780 if (current->flags & PF_SWAPWRITE)
781 return 1;
782 if (!inode_write_congested(inode))
783 return 1;
784 if (inode_to_bdi(inode) == current->backing_dev_info)
785 return 1;
786 return 0;
787}
788
789/*
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
793 *
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
797 *
798 * We're allowed to run sleeping lock_page() here because we know the caller has
799 * __GFP_FS.
800 */
801static void handle_write_error(struct address_space *mapping,
802 struct page *page, int error)
803{
804 lock_page(page);
805 if (page_mapping(page) == mapping)
806 mapping_set_error(mapping, error);
807 unlock_page(page);
808}
809
810/* possible outcome of pageout() */
811typedef enum {
812 /* failed to write page out, page is locked */
813 PAGE_KEEP,
814 /* move page to the active list, page is locked */
815 PAGE_ACTIVATE,
816 /* page has been sent to the disk successfully, page is unlocked */
817 PAGE_SUCCESS,
818 /* page is clean and locked */
819 PAGE_CLEAN,
820} pageout_t;
821
822/*
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
825 */
826static pageout_t pageout(struct page *page, struct address_space *mapping,
827 struct scan_control *sc)
828{
829 /*
830 * If the page is dirty, only perform writeback if that write
831 * will be non-blocking. To prevent this allocation from being
832 * stalled by pagecache activity. But note that there may be
833 * stalls if we need to run get_block(). We could test
834 * PagePrivate for that.
835 *
836 * If this process is currently in __generic_file_write_iter() against
837 * this page's queue, we can perform writeback even if that
838 * will block.
839 *
840 * If the page is swapcache, write it back even if that would
841 * block, for some throttling. This happens by accident, because
842 * swap_backing_dev_info is bust: it doesn't reflect the
843 * congestion state of the swapdevs. Easy to fix, if needed.
844 */
845 if (!is_page_cache_freeable(page))
846 return PAGE_KEEP;
847 if (!mapping) {
848 /*
849 * Some data journaling orphaned pages can have
850 * page->mapping == NULL while being dirty with clean buffers.
851 */
852 if (page_has_private(page)) {
853 if (try_to_free_buffers(page)) {
854 ClearPageDirty(page);
855 pr_info("%s: orphaned page\n", __func__);
856 return PAGE_CLEAN;
857 }
858 }
859 return PAGE_KEEP;
860 }
861 if (mapping->a_ops->writepage == NULL)
862 return PAGE_ACTIVATE;
863 if (!may_write_to_inode(mapping->host, sc))
864 return PAGE_KEEP;
865
866 if (clear_page_dirty_for_io(page)) {
867 int res;
868 struct writeback_control wbc = {
869 .sync_mode = WB_SYNC_NONE,
870 .nr_to_write = SWAP_CLUSTER_MAX,
871 .range_start = 0,
872 .range_end = LLONG_MAX,
873 .for_reclaim = 1,
874 };
875
876 SetPageReclaim(page);
877 res = mapping->a_ops->writepage(page, &wbc);
878 if (res < 0)
879 handle_write_error(mapping, page, res);
880 if (res == AOP_WRITEPAGE_ACTIVATE) {
881 ClearPageReclaim(page);
882 return PAGE_ACTIVATE;
883 }
884
885 if (!PageWriteback(page)) {
886 /* synchronous write or broken a_ops? */
887 ClearPageReclaim(page);
888 }
889 trace_mm_vmscan_writepage(page);
890 inc_node_page_state(page, NR_VMSCAN_WRITE);
891 return PAGE_SUCCESS;
892 }
893
894 return PAGE_CLEAN;
895}
896
897/*
898 * Same as remove_mapping, but if the page is removed from the mapping, it
899 * gets returned with a refcount of 0.
900 */
901static int __remove_mapping(struct address_space *mapping, struct page *page,
902 bool reclaimed)
903{
904 unsigned long flags;
905 int refcount;
906
907 BUG_ON(!PageLocked(page));
908 BUG_ON(mapping != page_mapping(page));
909
910 xa_lock_irqsave(&mapping->i_pages, flags);
911 /*
912 * The non racy check for a busy page.
913 *
914 * Must be careful with the order of the tests. When someone has
915 * a ref to the page, it may be possible that they dirty it then
916 * drop the reference. So if PageDirty is tested before page_count
917 * here, then the following race may occur:
918 *
919 * get_user_pages(&page);
920 * [user mapping goes away]
921 * write_to(page);
922 * !PageDirty(page) [good]
923 * SetPageDirty(page);
924 * put_page(page);
925 * !page_count(page) [good, discard it]
926 *
927 * [oops, our write_to data is lost]
928 *
929 * Reversing the order of the tests ensures such a situation cannot
930 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 * load is not satisfied before that of page->_refcount.
932 *
933 * Note that if SetPageDirty is always performed via set_page_dirty,
934 * and thus under the i_pages lock, then this ordering is not required.
935 */
936 refcount = 1 + compound_nr(page);
937 if (!page_ref_freeze(page, refcount))
938 goto cannot_free;
939 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 if (unlikely(PageDirty(page))) {
941 page_ref_unfreeze(page, refcount);
942 goto cannot_free;
943 }
944
945 if (PageSwapCache(page)) {
946 swp_entry_t swap = { .val = page_private(page) };
947 mem_cgroup_swapout(page, swap);
948 __delete_from_swap_cache(page, swap);
949 xa_unlock_irqrestore(&mapping->i_pages, flags);
950 put_swap_page(page, swap);
951 } else {
952 void (*freepage)(struct page *);
953 void *shadow = NULL;
954
955 freepage = mapping->a_ops->freepage;
956 /*
957 * Remember a shadow entry for reclaimed file cache in
958 * order to detect refaults, thus thrashing, later on.
959 *
960 * But don't store shadows in an address space that is
961 * already exiting. This is not just an optizimation,
962 * inode reclaim needs to empty out the radix tree or
963 * the nodes are lost. Don't plant shadows behind its
964 * back.
965 *
966 * We also don't store shadows for DAX mappings because the
967 * only page cache pages found in these are zero pages
968 * covering holes, and because we don't want to mix DAX
969 * exceptional entries and shadow exceptional entries in the
970 * same address_space.
971 */
972 if (reclaimed && page_is_file_cache(page) &&
973 !mapping_exiting(mapping) && !dax_mapping(mapping))
974 shadow = workingset_eviction(page);
975 __delete_from_page_cache(page, shadow);
976 xa_unlock_irqrestore(&mapping->i_pages, flags);
977
978 if (freepage != NULL)
979 freepage(page);
980 }
981
982 return 1;
983
984cannot_free:
985 xa_unlock_irqrestore(&mapping->i_pages, flags);
986 return 0;
987}
988
989/*
990 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 * someone else has a ref on the page, abort and return 0. If it was
992 * successfully detached, return 1. Assumes the caller has a single ref on
993 * this page.
994 */
995int remove_mapping(struct address_space *mapping, struct page *page)
996{
997 if (__remove_mapping(mapping, page, false)) {
998 /*
999 * Unfreezing the refcount with 1 rather than 2 effectively
1000 * drops the pagecache ref for us without requiring another
1001 * atomic operation.
1002 */
1003 page_ref_unfreeze(page, 1);
1004 return 1;
1005 }
1006 return 0;
1007}
1008
1009/**
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1012 *
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1015 *
1016 * lru_lock must not be held, interrupts must be enabled.
1017 */
1018void putback_lru_page(struct page *page)
1019{
1020 lru_cache_add(page);
1021 put_page(page); /* drop ref from isolate */
1022}
1023
1024enum page_references {
1025 PAGEREF_RECLAIM,
1026 PAGEREF_RECLAIM_CLEAN,
1027 PAGEREF_KEEP,
1028 PAGEREF_ACTIVATE,
1029};
1030
1031static enum page_references page_check_references(struct page *page,
1032 struct scan_control *sc)
1033{
1034 int referenced_ptes, referenced_page;
1035 unsigned long vm_flags;
1036
1037 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038 &vm_flags);
1039 referenced_page = TestClearPageReferenced(page);
1040
1041 /*
1042 * Mlock lost the isolation race with us. Let try_to_unmap()
1043 * move the page to the unevictable list.
1044 */
1045 if (vm_flags & VM_LOCKED)
1046 return PAGEREF_RECLAIM;
1047
1048 if (referenced_ptes) {
1049 if (PageSwapBacked(page))
1050 return PAGEREF_ACTIVATE;
1051 /*
1052 * All mapped pages start out with page table
1053 * references from the instantiating fault, so we need
1054 * to look twice if a mapped file page is used more
1055 * than once.
1056 *
1057 * Mark it and spare it for another trip around the
1058 * inactive list. Another page table reference will
1059 * lead to its activation.
1060 *
1061 * Note: the mark is set for activated pages as well
1062 * so that recently deactivated but used pages are
1063 * quickly recovered.
1064 */
1065 SetPageReferenced(page);
1066
1067 if (referenced_page || referenced_ptes > 1)
1068 return PAGEREF_ACTIVATE;
1069
1070 /*
1071 * Activate file-backed executable pages after first usage.
1072 */
1073 if (vm_flags & VM_EXEC)
1074 return PAGEREF_ACTIVATE;
1075
1076 return PAGEREF_KEEP;
1077 }
1078
1079 /* Reclaim if clean, defer dirty pages to writeback */
1080 if (referenced_page && !PageSwapBacked(page))
1081 return PAGEREF_RECLAIM_CLEAN;
1082
1083 return PAGEREF_RECLAIM;
1084}
1085
1086/* Check if a page is dirty or under writeback */
1087static void page_check_dirty_writeback(struct page *page,
1088 bool *dirty, bool *writeback)
1089{
1090 struct address_space *mapping;
1091
1092 /*
1093 * Anonymous pages are not handled by flushers and must be written
1094 * from reclaim context. Do not stall reclaim based on them
1095 */
1096 if (!page_is_file_cache(page) ||
1097 (PageAnon(page) && !PageSwapBacked(page))) {
1098 *dirty = false;
1099 *writeback = false;
1100 return;
1101 }
1102
1103 /* By default assume that the page flags are accurate */
1104 *dirty = PageDirty(page);
1105 *writeback = PageWriteback(page);
1106
1107 /* Verify dirty/writeback state if the filesystem supports it */
1108 if (!page_has_private(page))
1109 return;
1110
1111 mapping = page_mapping(page);
1112 if (mapping && mapping->a_ops->is_dirty_writeback)
1113 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1114}
1115
1116/*
1117 * shrink_page_list() returns the number of reclaimed pages
1118 */
1119static unsigned long shrink_page_list(struct list_head *page_list,
1120 struct pglist_data *pgdat,
1121 struct scan_control *sc,
1122 enum ttu_flags ttu_flags,
1123 struct reclaim_stat *stat,
1124 bool ignore_references)
1125{
1126 LIST_HEAD(ret_pages);
1127 LIST_HEAD(free_pages);
1128 unsigned nr_reclaimed = 0;
1129 unsigned pgactivate = 0;
1130
1131 memset(stat, 0, sizeof(*stat));
1132 cond_resched();
1133
1134 while (!list_empty(page_list)) {
1135 struct address_space *mapping;
1136 struct page *page;
1137 int may_enter_fs;
1138 enum page_references references = PAGEREF_RECLAIM;
1139 bool dirty, writeback;
1140 unsigned int nr_pages;
1141
1142 cond_resched();
1143
1144 page = lru_to_page(page_list);
1145 list_del(&page->lru);
1146
1147 if (!trylock_page(page))
1148 goto keep;
1149
1150 VM_BUG_ON_PAGE(PageActive(page), page);
1151
1152 nr_pages = compound_nr(page);
1153
1154 /* Account the number of base pages even though THP */
1155 sc->nr_scanned += nr_pages;
1156
1157 if (unlikely(!page_evictable(page)))
1158 goto activate_locked;
1159
1160 if (!sc->may_unmap && page_mapped(page))
1161 goto keep_locked;
1162
1163 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165
1166 /*
1167 * The number of dirty pages determines if a node is marked
1168 * reclaim_congested which affects wait_iff_congested. kswapd
1169 * will stall and start writing pages if the tail of the LRU
1170 * is all dirty unqueued pages.
1171 */
1172 page_check_dirty_writeback(page, &dirty, &writeback);
1173 if (dirty || writeback)
1174 stat->nr_dirty++;
1175
1176 if (dirty && !writeback)
1177 stat->nr_unqueued_dirty++;
1178
1179 /*
1180 * Treat this page as congested if the underlying BDI is or if
1181 * pages are cycling through the LRU so quickly that the
1182 * pages marked for immediate reclaim are making it to the
1183 * end of the LRU a second time.
1184 */
1185 mapping = page_mapping(page);
1186 if (((dirty || writeback) && mapping &&
1187 inode_write_congested(mapping->host)) ||
1188 (writeback && PageReclaim(page)))
1189 stat->nr_congested++;
1190
1191 /*
1192 * If a page at the tail of the LRU is under writeback, there
1193 * are three cases to consider.
1194 *
1195 * 1) If reclaim is encountering an excessive number of pages
1196 * under writeback and this page is both under writeback and
1197 * PageReclaim then it indicates that pages are being queued
1198 * for IO but are being recycled through the LRU before the
1199 * IO can complete. Waiting on the page itself risks an
1200 * indefinite stall if it is impossible to writeback the
1201 * page due to IO error or disconnected storage so instead
1202 * note that the LRU is being scanned too quickly and the
1203 * caller can stall after page list has been processed.
1204 *
1205 * 2) Global or new memcg reclaim encounters a page that is
1206 * not marked for immediate reclaim, or the caller does not
1207 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 * not to fs). In this case mark the page for immediate
1209 * reclaim and continue scanning.
1210 *
1211 * Require may_enter_fs because we would wait on fs, which
1212 * may not have submitted IO yet. And the loop driver might
1213 * enter reclaim, and deadlock if it waits on a page for
1214 * which it is needed to do the write (loop masks off
1215 * __GFP_IO|__GFP_FS for this reason); but more thought
1216 * would probably show more reasons.
1217 *
1218 * 3) Legacy memcg encounters a page that is already marked
1219 * PageReclaim. memcg does not have any dirty pages
1220 * throttling so we could easily OOM just because too many
1221 * pages are in writeback and there is nothing else to
1222 * reclaim. Wait for the writeback to complete.
1223 *
1224 * In cases 1) and 2) we activate the pages to get them out of
1225 * the way while we continue scanning for clean pages on the
1226 * inactive list and refilling from the active list. The
1227 * observation here is that waiting for disk writes is more
1228 * expensive than potentially causing reloads down the line.
1229 * Since they're marked for immediate reclaim, they won't put
1230 * memory pressure on the cache working set any longer than it
1231 * takes to write them to disk.
1232 */
1233 if (PageWriteback(page)) {
1234 /* Case 1 above */
1235 if (current_is_kswapd() &&
1236 PageReclaim(page) &&
1237 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238 stat->nr_immediate++;
1239 goto activate_locked;
1240
1241 /* Case 2 above */
1242 } else if (sane_reclaim(sc) ||
1243 !PageReclaim(page) || !may_enter_fs) {
1244 /*
1245 * This is slightly racy - end_page_writeback()
1246 * might have just cleared PageReclaim, then
1247 * setting PageReclaim here end up interpreted
1248 * as PageReadahead - but that does not matter
1249 * enough to care. What we do want is for this
1250 * page to have PageReclaim set next time memcg
1251 * reclaim reaches the tests above, so it will
1252 * then wait_on_page_writeback() to avoid OOM;
1253 * and it's also appropriate in global reclaim.
1254 */
1255 SetPageReclaim(page);
1256 stat->nr_writeback++;
1257 goto activate_locked;
1258
1259 /* Case 3 above */
1260 } else {
1261 unlock_page(page);
1262 wait_on_page_writeback(page);
1263 /* then go back and try same page again */
1264 list_add_tail(&page->lru, page_list);
1265 continue;
1266 }
1267 }
1268
1269 if (!ignore_references)
1270 references = page_check_references(page, sc);
1271
1272 switch (references) {
1273 case PAGEREF_ACTIVATE:
1274 goto activate_locked;
1275 case PAGEREF_KEEP:
1276 stat->nr_ref_keep += nr_pages;
1277 goto keep_locked;
1278 case PAGEREF_RECLAIM:
1279 case PAGEREF_RECLAIM_CLEAN:
1280 ; /* try to reclaim the page below */
1281 }
1282
1283 /*
1284 * Anonymous process memory has backing store?
1285 * Try to allocate it some swap space here.
1286 * Lazyfree page could be freed directly
1287 */
1288 if (PageAnon(page) && PageSwapBacked(page)) {
1289 if (!PageSwapCache(page)) {
1290 if (!(sc->gfp_mask & __GFP_IO))
1291 goto keep_locked;
1292 if (PageTransHuge(page)) {
1293 /* cannot split THP, skip it */
1294 if (!can_split_huge_page(page, NULL))
1295 goto activate_locked;
1296 /*
1297 * Split pages without a PMD map right
1298 * away. Chances are some or all of the
1299 * tail pages can be freed without IO.
1300 */
1301 if (!compound_mapcount(page) &&
1302 split_huge_page_to_list(page,
1303 page_list))
1304 goto activate_locked;
1305 }
1306 if (!add_to_swap(page)) {
1307 if (!PageTransHuge(page))
1308 goto activate_locked_split;
1309 /* Fallback to swap normal pages */
1310 if (split_huge_page_to_list(page,
1311 page_list))
1312 goto activate_locked;
1313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 count_vm_event(THP_SWPOUT_FALLBACK);
1315#endif
1316 if (!add_to_swap(page))
1317 goto activate_locked_split;
1318 }
1319
1320 may_enter_fs = 1;
1321
1322 /* Adding to swap updated mapping */
1323 mapping = page_mapping(page);
1324 }
1325 } else if (unlikely(PageTransHuge(page))) {
1326 /* Split file THP */
1327 if (split_huge_page_to_list(page, page_list))
1328 goto keep_locked;
1329 }
1330
1331 /*
1332 * THP may get split above, need minus tail pages and update
1333 * nr_pages to avoid accounting tail pages twice.
1334 *
1335 * The tail pages that are added into swap cache successfully
1336 * reach here.
1337 */
1338 if ((nr_pages > 1) && !PageTransHuge(page)) {
1339 sc->nr_scanned -= (nr_pages - 1);
1340 nr_pages = 1;
1341 }
1342
1343 /*
1344 * The page is mapped into the page tables of one or more
1345 * processes. Try to unmap it here.
1346 */
1347 if (page_mapped(page)) {
1348 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349
1350 if (unlikely(PageTransHuge(page)))
1351 flags |= TTU_SPLIT_HUGE_PMD;
1352 if (!try_to_unmap(page, flags)) {
1353 stat->nr_unmap_fail += nr_pages;
1354 goto activate_locked;
1355 }
1356 }
1357
1358 if (PageDirty(page)) {
1359 /*
1360 * Only kswapd can writeback filesystem pages
1361 * to avoid risk of stack overflow. But avoid
1362 * injecting inefficient single-page IO into
1363 * flusher writeback as much as possible: only
1364 * write pages when we've encountered many
1365 * dirty pages, and when we've already scanned
1366 * the rest of the LRU for clean pages and see
1367 * the same dirty pages again (PageReclaim).
1368 */
1369 if (page_is_file_cache(page) &&
1370 (!current_is_kswapd() || !PageReclaim(page) ||
1371 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372 /*
1373 * Immediately reclaim when written back.
1374 * Similar in principal to deactivate_page()
1375 * except we already have the page isolated
1376 * and know it's dirty
1377 */
1378 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379 SetPageReclaim(page);
1380
1381 goto activate_locked;
1382 }
1383
1384 if (references == PAGEREF_RECLAIM_CLEAN)
1385 goto keep_locked;
1386 if (!may_enter_fs)
1387 goto keep_locked;
1388 if (!sc->may_writepage)
1389 goto keep_locked;
1390
1391 /*
1392 * Page is dirty. Flush the TLB if a writable entry
1393 * potentially exists to avoid CPU writes after IO
1394 * starts and then write it out here.
1395 */
1396 try_to_unmap_flush_dirty();
1397 switch (pageout(page, mapping, sc)) {
1398 case PAGE_KEEP:
1399 goto keep_locked;
1400 case PAGE_ACTIVATE:
1401 goto activate_locked;
1402 case PAGE_SUCCESS:
1403 if (PageWriteback(page))
1404 goto keep;
1405 if (PageDirty(page))
1406 goto keep;
1407
1408 /*
1409 * A synchronous write - probably a ramdisk. Go
1410 * ahead and try to reclaim the page.
1411 */
1412 if (!trylock_page(page))
1413 goto keep;
1414 if (PageDirty(page) || PageWriteback(page))
1415 goto keep_locked;
1416 mapping = page_mapping(page);
1417 case PAGE_CLEAN:
1418 ; /* try to free the page below */
1419 }
1420 }
1421
1422 /*
1423 * If the page has buffers, try to free the buffer mappings
1424 * associated with this page. If we succeed we try to free
1425 * the page as well.
1426 *
1427 * We do this even if the page is PageDirty().
1428 * try_to_release_page() does not perform I/O, but it is
1429 * possible for a page to have PageDirty set, but it is actually
1430 * clean (all its buffers are clean). This happens if the
1431 * buffers were written out directly, with submit_bh(). ext3
1432 * will do this, as well as the blockdev mapping.
1433 * try_to_release_page() will discover that cleanness and will
1434 * drop the buffers and mark the page clean - it can be freed.
1435 *
1436 * Rarely, pages can have buffers and no ->mapping. These are
1437 * the pages which were not successfully invalidated in
1438 * truncate_complete_page(). We try to drop those buffers here
1439 * and if that worked, and the page is no longer mapped into
1440 * process address space (page_count == 1) it can be freed.
1441 * Otherwise, leave the page on the LRU so it is swappable.
1442 */
1443 if (page_has_private(page)) {
1444 if (!try_to_release_page(page, sc->gfp_mask))
1445 goto activate_locked;
1446 if (!mapping && page_count(page) == 1) {
1447 unlock_page(page);
1448 if (put_page_testzero(page))
1449 goto free_it;
1450 else {
1451 /*
1452 * rare race with speculative reference.
1453 * the speculative reference will free
1454 * this page shortly, so we may
1455 * increment nr_reclaimed here (and
1456 * leave it off the LRU).
1457 */
1458 nr_reclaimed++;
1459 continue;
1460 }
1461 }
1462 }
1463
1464 if (PageAnon(page) && !PageSwapBacked(page)) {
1465 /* follow __remove_mapping for reference */
1466 if (!page_ref_freeze(page, 1))
1467 goto keep_locked;
1468 if (PageDirty(page)) {
1469 page_ref_unfreeze(page, 1);
1470 goto keep_locked;
1471 }
1472
1473 count_vm_event(PGLAZYFREED);
1474 count_memcg_page_event(page, PGLAZYFREED);
1475 } else if (!mapping || !__remove_mapping(mapping, page, true))
1476 goto keep_locked;
1477
1478 unlock_page(page);
1479free_it:
1480 /*
1481 * THP may get swapped out in a whole, need account
1482 * all base pages.
1483 */
1484 nr_reclaimed += nr_pages;
1485
1486 /*
1487 * Is there need to periodically free_page_list? It would
1488 * appear not as the counts should be low
1489 */
1490 if (unlikely(PageTransHuge(page)))
1491 (*get_compound_page_dtor(page))(page);
1492 else
1493 list_add(&page->lru, &free_pages);
1494 continue;
1495
1496activate_locked_split:
1497 /*
1498 * The tail pages that are failed to add into swap cache
1499 * reach here. Fixup nr_scanned and nr_pages.
1500 */
1501 if (nr_pages > 1) {
1502 sc->nr_scanned -= (nr_pages - 1);
1503 nr_pages = 1;
1504 }
1505activate_locked:
1506 /* Not a candidate for swapping, so reclaim swap space. */
1507 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508 PageMlocked(page)))
1509 try_to_free_swap(page);
1510 VM_BUG_ON_PAGE(PageActive(page), page);
1511 if (!PageMlocked(page)) {
1512 int type = page_is_file_cache(page);
1513 SetPageActive(page);
1514 stat->nr_activate[type] += nr_pages;
1515 count_memcg_page_event(page, PGACTIVATE);
1516 }
1517keep_locked:
1518 unlock_page(page);
1519keep:
1520 list_add(&page->lru, &ret_pages);
1521 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1522 }
1523
1524 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525
1526 mem_cgroup_uncharge_list(&free_pages);
1527 try_to_unmap_flush();
1528 free_unref_page_list(&free_pages);
1529
1530 list_splice(&ret_pages, page_list);
1531 count_vm_events(PGACTIVATE, pgactivate);
1532
1533 return nr_reclaimed;
1534}
1535
1536unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537 struct list_head *page_list)
1538{
1539 struct scan_control sc = {
1540 .gfp_mask = GFP_KERNEL,
1541 .priority = DEF_PRIORITY,
1542 .may_unmap = 1,
1543 };
1544 struct reclaim_stat dummy_stat;
1545 unsigned long ret;
1546 struct page *page, *next;
1547 LIST_HEAD(clean_pages);
1548
1549 list_for_each_entry_safe(page, next, page_list, lru) {
1550 if (page_is_file_cache(page) && !PageDirty(page) &&
1551 !__PageMovable(page) && !PageUnevictable(page)) {
1552 ClearPageActive(page);
1553 list_move(&page->lru, &clean_pages);
1554 }
1555 }
1556
1557 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558 TTU_IGNORE_ACCESS, &dummy_stat, true);
1559 list_splice(&clean_pages, page_list);
1560 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1561 return ret;
1562}
1563
1564/*
1565 * Attempt to remove the specified page from its LRU. Only take this page
1566 * if it is of the appropriate PageActive status. Pages which are being
1567 * freed elsewhere are also ignored.
1568 *
1569 * page: page to consider
1570 * mode: one of the LRU isolation modes defined above
1571 *
1572 * returns 0 on success, -ve errno on failure.
1573 */
1574int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1575{
1576 int ret = -EINVAL;
1577
1578 /* Only take pages on the LRU. */
1579 if (!PageLRU(page))
1580 return ret;
1581
1582 /* Compaction should not handle unevictable pages but CMA can do so */
1583 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1584 return ret;
1585
1586 ret = -EBUSY;
1587
1588 /*
1589 * To minimise LRU disruption, the caller can indicate that it only
1590 * wants to isolate pages it will be able to operate on without
1591 * blocking - clean pages for the most part.
1592 *
1593 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 * that it is possible to migrate without blocking
1595 */
1596 if (mode & ISOLATE_ASYNC_MIGRATE) {
1597 /* All the caller can do on PageWriteback is block */
1598 if (PageWriteback(page))
1599 return ret;
1600
1601 if (PageDirty(page)) {
1602 struct address_space *mapping;
1603 bool migrate_dirty;
1604
1605 /*
1606 * Only pages without mappings or that have a
1607 * ->migratepage callback are possible to migrate
1608 * without blocking. However, we can be racing with
1609 * truncation so it's necessary to lock the page
1610 * to stabilise the mapping as truncation holds
1611 * the page lock until after the page is removed
1612 * from the page cache.
1613 */
1614 if (!trylock_page(page))
1615 return ret;
1616
1617 mapping = page_mapping(page);
1618 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1619 unlock_page(page);
1620 if (!migrate_dirty)
1621 return ret;
1622 }
1623 }
1624
1625 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1626 return ret;
1627
1628 if (likely(get_page_unless_zero(page))) {
1629 /*
1630 * Be careful not to clear PageLRU until after we're
1631 * sure the page is not being freed elsewhere -- the
1632 * page release code relies on it.
1633 */
1634 ClearPageLRU(page);
1635 ret = 0;
1636 }
1637
1638 return ret;
1639}
1640
1641
1642/*
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 */
1646static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647 enum lru_list lru, unsigned long *nr_zone_taken)
1648{
1649 int zid;
1650
1651 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652 if (!nr_zone_taken[zid])
1653 continue;
1654
1655 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656#ifdef CONFIG_MEMCG
1657 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1658#endif
1659 }
1660
1661}
1662
1663/**
1664 * pgdat->lru_lock is heavily contended. Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1667 *
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1670 *
1671 * Appropriate locks must be held before calling this function.
1672 *
1673 * @nr_to_scan: The number of eligible pages to look through on the list.
1674 * @lruvec: The LRU vector to pull pages from.
1675 * @dst: The temp list to put pages on to.
1676 * @nr_scanned: The number of pages that were scanned.
1677 * @sc: The scan_control struct for this reclaim session
1678 * @mode: One of the LRU isolation modes
1679 * @lru: LRU list id for isolating
1680 *
1681 * returns how many pages were moved onto *@dst.
1682 */
1683static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684 struct lruvec *lruvec, struct list_head *dst,
1685 unsigned long *nr_scanned, struct scan_control *sc,
1686 enum lru_list lru)
1687{
1688 struct list_head *src = &lruvec->lists[lru];
1689 unsigned long nr_taken = 0;
1690 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692 unsigned long skipped = 0;
1693 unsigned long scan, total_scan, nr_pages;
1694 LIST_HEAD(pages_skipped);
1695 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1696
1697 total_scan = 0;
1698 scan = 0;
1699 while (scan < nr_to_scan && !list_empty(src)) {
1700 struct page *page;
1701
1702 page = lru_to_page(src);
1703 prefetchw_prev_lru_page(page, src, flags);
1704
1705 VM_BUG_ON_PAGE(!PageLRU(page), page);
1706
1707 nr_pages = compound_nr(page);
1708 total_scan += nr_pages;
1709
1710 if (page_zonenum(page) > sc->reclaim_idx) {
1711 list_move(&page->lru, &pages_skipped);
1712 nr_skipped[page_zonenum(page)] += nr_pages;
1713 continue;
1714 }
1715
1716 /*
1717 * Do not count skipped pages because that makes the function
1718 * return with no isolated pages if the LRU mostly contains
1719 * ineligible pages. This causes the VM to not reclaim any
1720 * pages, triggering a premature OOM.
1721 *
1722 * Account all tail pages of THP. This would not cause
1723 * premature OOM since __isolate_lru_page() returns -EBUSY
1724 * only when the page is being freed somewhere else.
1725 */
1726 scan += nr_pages;
1727 switch (__isolate_lru_page(page, mode)) {
1728 case 0:
1729 nr_taken += nr_pages;
1730 nr_zone_taken[page_zonenum(page)] += nr_pages;
1731 list_move(&page->lru, dst);
1732 break;
1733
1734 case -EBUSY:
1735 /* else it is being freed elsewhere */
1736 list_move(&page->lru, src);
1737 continue;
1738
1739 default:
1740 BUG();
1741 }
1742 }
1743
1744 /*
1745 * Splice any skipped pages to the start of the LRU list. Note that
1746 * this disrupts the LRU order when reclaiming for lower zones but
1747 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 * scanning would soon rescan the same pages to skip and put the
1749 * system at risk of premature OOM.
1750 */
1751 if (!list_empty(&pages_skipped)) {
1752 int zid;
1753
1754 list_splice(&pages_skipped, src);
1755 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756 if (!nr_skipped[zid])
1757 continue;
1758
1759 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760 skipped += nr_skipped[zid];
1761 }
1762 }
1763 *nr_scanned = total_scan;
1764 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765 total_scan, skipped, nr_taken, mode, lru);
1766 update_lru_sizes(lruvec, lru, nr_zone_taken);
1767 return nr_taken;
1768}
1769
1770/**
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1773 *
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1776 *
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1779 *
1780 * The returned page will have PageLRU() cleared. If it was found on
1781 * the active list, it will have PageActive set. If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1784 *
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1787 *
1788 * Restrictions:
1789 *
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 * fundamentnal difference from isolate_lru_pages (which is called
1792 * without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1795 */
1796int isolate_lru_page(struct page *page)
1797{
1798 int ret = -EBUSY;
1799
1800 VM_BUG_ON_PAGE(!page_count(page), page);
1801 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802
1803 if (PageLRU(page)) {
1804 pg_data_t *pgdat = page_pgdat(page);
1805 struct lruvec *lruvec;
1806
1807 spin_lock_irq(&pgdat->lru_lock);
1808 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809 if (PageLRU(page)) {
1810 int lru = page_lru(page);
1811 get_page(page);
1812 ClearPageLRU(page);
1813 del_page_from_lru_list(page, lruvec, lru);
1814 ret = 0;
1815 }
1816 spin_unlock_irq(&pgdat->lru_lock);
1817 }
1818 return ret;
1819}
1820
1821/*
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1827 */
1828static int too_many_isolated(struct pglist_data *pgdat, int file,
1829 struct scan_control *sc)
1830{
1831 unsigned long inactive, isolated;
1832
1833 if (current_is_kswapd())
1834 return 0;
1835
1836 if (!sane_reclaim(sc))
1837 return 0;
1838
1839 if (file) {
1840 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842 } else {
1843 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1845 }
1846
1847 /*
1848 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 * won't get blocked by normal direct-reclaimers, forming a circular
1850 * deadlock.
1851 */
1852 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1853 inactive >>= 3;
1854
1855 return isolated > inactive;
1856}
1857
1858/*
1859 * This moves pages from @list to corresponding LRU list.
1860 *
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1863 *
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation. But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page. It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1871 *
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1874 *
1875 * Returns the number of pages moved to the given lruvec.
1876 */
1877
1878static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879 struct list_head *list)
1880{
1881 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882 int nr_pages, nr_moved = 0;
1883 LIST_HEAD(pages_to_free);
1884 struct page *page;
1885 enum lru_list lru;
1886
1887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
1889 VM_BUG_ON_PAGE(PageLRU(page), page);
1890 if (unlikely(!page_evictable(page))) {
1891 list_del(&page->lru);
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 putback_lru_page(page);
1894 spin_lock_irq(&pgdat->lru_lock);
1895 continue;
1896 }
1897 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1898
1899 SetPageLRU(page);
1900 lru = page_lru(page);
1901
1902 nr_pages = hpage_nr_pages(page);
1903 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904 list_move(&page->lru, &lruvec->lists[lru]);
1905
1906 if (put_page_testzero(page)) {
1907 __ClearPageLRU(page);
1908 __ClearPageActive(page);
1909 del_page_from_lru_list(page, lruvec, lru);
1910
1911 if (unlikely(PageCompound(page))) {
1912 spin_unlock_irq(&pgdat->lru_lock);
1913 (*get_compound_page_dtor(page))(page);
1914 spin_lock_irq(&pgdat->lru_lock);
1915 } else
1916 list_add(&page->lru, &pages_to_free);
1917 } else {
1918 nr_moved += nr_pages;
1919 }
1920 }
1921
1922 /*
1923 * To save our caller's stack, now use input list for pages to free.
1924 */
1925 list_splice(&pages_to_free, list);
1926
1927 return nr_moved;
1928}
1929
1930/*
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested. In other cases it is safe to throttle.
1935 */
1936static int current_may_throttle(void)
1937{
1938 return !(current->flags & PF_LESS_THROTTLE) ||
1939 current->backing_dev_info == NULL ||
1940 bdi_write_congested(current->backing_dev_info);
1941}
1942
1943/*
1944 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 * of reclaimed pages
1946 */
1947static noinline_for_stack unsigned long
1948shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949 struct scan_control *sc, enum lru_list lru)
1950{
1951 LIST_HEAD(page_list);
1952 unsigned long nr_scanned;
1953 unsigned long nr_reclaimed = 0;
1954 unsigned long nr_taken;
1955 struct reclaim_stat stat;
1956 int file = is_file_lru(lru);
1957 enum vm_event_item item;
1958 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960 bool stalled = false;
1961
1962 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1963 if (stalled)
1964 return 0;
1965
1966 /* wait a bit for the reclaimer. */
1967 msleep(100);
1968 stalled = true;
1969
1970 /* We are about to die and free our memory. Return now. */
1971 if (fatal_signal_pending(current))
1972 return SWAP_CLUSTER_MAX;
1973 }
1974
1975 lru_add_drain();
1976
1977 spin_lock_irq(&pgdat->lru_lock);
1978
1979 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980 &nr_scanned, sc, lru);
1981
1982 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 reclaim_stat->recent_scanned[file] += nr_taken;
1984
1985 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986 if (global_reclaim(sc))
1987 __count_vm_events(item, nr_scanned);
1988 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1989 spin_unlock_irq(&pgdat->lru_lock);
1990
1991 if (nr_taken == 0)
1992 return 0;
1993
1994 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1995 &stat, false);
1996
1997 spin_lock_irq(&pgdat->lru_lock);
1998
1999 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000 if (global_reclaim(sc))
2001 __count_vm_events(item, nr_reclaimed);
2002 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2005
2006 move_pages_to_lru(lruvec, &page_list);
2007
2008 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009
2010 spin_unlock_irq(&pgdat->lru_lock);
2011
2012 mem_cgroup_uncharge_list(&page_list);
2013 free_unref_page_list(&page_list);
2014
2015 /*
2016 * If dirty pages are scanned that are not queued for IO, it
2017 * implies that flushers are not doing their job. This can
2018 * happen when memory pressure pushes dirty pages to the end of
2019 * the LRU before the dirty limits are breached and the dirty
2020 * data has expired. It can also happen when the proportion of
2021 * dirty pages grows not through writes but through memory
2022 * pressure reclaiming all the clean cache. And in some cases,
2023 * the flushers simply cannot keep up with the allocation
2024 * rate. Nudge the flusher threads in case they are asleep.
2025 */
2026 if (stat.nr_unqueued_dirty == nr_taken)
2027 wakeup_flusher_threads(WB_REASON_VMSCAN);
2028
2029 sc->nr.dirty += stat.nr_dirty;
2030 sc->nr.congested += stat.nr_congested;
2031 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032 sc->nr.writeback += stat.nr_writeback;
2033 sc->nr.immediate += stat.nr_immediate;
2034 sc->nr.taken += nr_taken;
2035 if (file)
2036 sc->nr.file_taken += nr_taken;
2037
2038 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040 return nr_reclaimed;
2041}
2042
2043static void shrink_active_list(unsigned long nr_to_scan,
2044 struct lruvec *lruvec,
2045 struct scan_control *sc,
2046 enum lru_list lru)
2047{
2048 unsigned long nr_taken;
2049 unsigned long nr_scanned;
2050 unsigned long vm_flags;
2051 LIST_HEAD(l_hold); /* The pages which were snipped off */
2052 LIST_HEAD(l_active);
2053 LIST_HEAD(l_inactive);
2054 struct page *page;
2055 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056 unsigned nr_deactivate, nr_activate;
2057 unsigned nr_rotated = 0;
2058 int file = is_file_lru(lru);
2059 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2060
2061 lru_add_drain();
2062
2063 spin_lock_irq(&pgdat->lru_lock);
2064
2065 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066 &nr_scanned, sc, lru);
2067
2068 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069 reclaim_stat->recent_scanned[file] += nr_taken;
2070
2071 __count_vm_events(PGREFILL, nr_scanned);
2072 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073
2074 spin_unlock_irq(&pgdat->lru_lock);
2075
2076 while (!list_empty(&l_hold)) {
2077 cond_resched();
2078 page = lru_to_page(&l_hold);
2079 list_del(&page->lru);
2080
2081 if (unlikely(!page_evictable(page))) {
2082 putback_lru_page(page);
2083 continue;
2084 }
2085
2086 if (unlikely(buffer_heads_over_limit)) {
2087 if (page_has_private(page) && trylock_page(page)) {
2088 if (page_has_private(page))
2089 try_to_release_page(page, 0);
2090 unlock_page(page);
2091 }
2092 }
2093
2094 if (page_referenced(page, 0, sc->target_mem_cgroup,
2095 &vm_flags)) {
2096 nr_rotated += hpage_nr_pages(page);
2097 /*
2098 * Identify referenced, file-backed active pages and
2099 * give them one more trip around the active list. So
2100 * that executable code get better chances to stay in
2101 * memory under moderate memory pressure. Anon pages
2102 * are not likely to be evicted by use-once streaming
2103 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 * so we ignore them here.
2105 */
2106 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107 list_add(&page->lru, &l_active);
2108 continue;
2109 }
2110 }
2111
2112 ClearPageActive(page); /* we are de-activating */
2113 SetPageWorkingset(page);
2114 list_add(&page->lru, &l_inactive);
2115 }
2116
2117 /*
2118 * Move pages back to the lru list.
2119 */
2120 spin_lock_irq(&pgdat->lru_lock);
2121 /*
2122 * Count referenced pages from currently used mappings as rotated,
2123 * even though only some of them are actually re-activated. This
2124 * helps balance scan pressure between file and anonymous pages in
2125 * get_scan_count.
2126 */
2127 reclaim_stat->recent_rotated[file] += nr_rotated;
2128
2129 nr_activate = move_pages_to_lru(lruvec, &l_active);
2130 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131 /* Keep all free pages in l_active list */
2132 list_splice(&l_inactive, &l_active);
2133
2134 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2135 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136
2137 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138 spin_unlock_irq(&pgdat->lru_lock);
2139
2140 mem_cgroup_uncharge_list(&l_active);
2141 free_unref_page_list(&l_active);
2142 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143 nr_deactivate, nr_rotated, sc->priority, file);
2144}
2145
2146unsigned long reclaim_pages(struct list_head *page_list)
2147{
2148 int nid = -1;
2149 unsigned long nr_reclaimed = 0;
2150 LIST_HEAD(node_page_list);
2151 struct reclaim_stat dummy_stat;
2152 struct page *page;
2153 struct scan_control sc = {
2154 .gfp_mask = GFP_KERNEL,
2155 .priority = DEF_PRIORITY,
2156 .may_writepage = 1,
2157 .may_unmap = 1,
2158 .may_swap = 1,
2159 };
2160
2161 while (!list_empty(page_list)) {
2162 page = lru_to_page(page_list);
2163 if (nid == -1) {
2164 nid = page_to_nid(page);
2165 INIT_LIST_HEAD(&node_page_list);
2166 }
2167
2168 if (nid == page_to_nid(page)) {
2169 ClearPageActive(page);
2170 list_move(&page->lru, &node_page_list);
2171 continue;
2172 }
2173
2174 nr_reclaimed += shrink_page_list(&node_page_list,
2175 NODE_DATA(nid),
2176 &sc, 0,
2177 &dummy_stat, false);
2178 while (!list_empty(&node_page_list)) {
2179 page = lru_to_page(&node_page_list);
2180 list_del(&page->lru);
2181 putback_lru_page(page);
2182 }
2183
2184 nid = -1;
2185 }
2186
2187 if (!list_empty(&node_page_list)) {
2188 nr_reclaimed += shrink_page_list(&node_page_list,
2189 NODE_DATA(nid),
2190 &sc, 0,
2191 &dummy_stat, false);
2192 while (!list_empty(&node_page_list)) {
2193 page = lru_to_page(&node_page_list);
2194 list_del(&page->lru);
2195 putback_lru_page(page);
2196 }
2197 }
2198
2199 return nr_reclaimed;
2200}
2201
2202/*
2203 * The inactive anon list should be small enough that the VM never has
2204 * to do too much work.
2205 *
2206 * The inactive file list should be small enough to leave most memory
2207 * to the established workingset on the scan-resistant active list,
2208 * but large enough to avoid thrashing the aggregate readahead window.
2209 *
2210 * Both inactive lists should also be large enough that each inactive
2211 * page has a chance to be referenced again before it is reclaimed.
2212 *
2213 * If that fails and refaulting is observed, the inactive list grows.
2214 *
2215 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 * on this LRU, maintained by the pageout code. An inactive_ratio
2217 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2218 *
2219 * total target max
2220 * memory ratio inactive
2221 * -------------------------------------
2222 * 10MB 1 5MB
2223 * 100MB 1 50MB
2224 * 1GB 3 250MB
2225 * 10GB 10 0.9GB
2226 * 100GB 31 3GB
2227 * 1TB 101 10GB
2228 * 10TB 320 32GB
2229 */
2230static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2231 struct scan_control *sc, bool trace)
2232{
2233 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2234 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2235 enum lru_list inactive_lru = file * LRU_FILE;
2236 unsigned long inactive, active;
2237 unsigned long inactive_ratio;
2238 unsigned long refaults;
2239 unsigned long gb;
2240
2241 /*
2242 * If we don't have swap space, anonymous page deactivation
2243 * is pointless.
2244 */
2245 if (!file && !total_swap_pages)
2246 return false;
2247
2248 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2249 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2250
2251 /*
2252 * When refaults are being observed, it means a new workingset
2253 * is being established. Disable active list protection to get
2254 * rid of the stale workingset quickly.
2255 */
2256 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2257 if (file && lruvec->refaults != refaults) {
2258 inactive_ratio = 0;
2259 } else {
2260 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261 if (gb)
2262 inactive_ratio = int_sqrt(10 * gb);
2263 else
2264 inactive_ratio = 1;
2265 }
2266
2267 if (trace)
2268 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2269 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2270 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2271 inactive_ratio, file);
2272
2273 return inactive * inactive_ratio < active;
2274}
2275
2276static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277 struct lruvec *lruvec, struct scan_control *sc)
2278{
2279 if (is_active_lru(lru)) {
2280 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2281 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282 return 0;
2283 }
2284
2285 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2286}
2287
2288enum scan_balance {
2289 SCAN_EQUAL,
2290 SCAN_FRACT,
2291 SCAN_ANON,
2292 SCAN_FILE,
2293};
2294
2295/*
2296 * Determine how aggressively the anon and file LRU lists should be
2297 * scanned. The relative value of each set of LRU lists is determined
2298 * by looking at the fraction of the pages scanned we did rotate back
2299 * onto the active list instead of evict.
2300 *
2301 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 */
2304static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2305 struct scan_control *sc, unsigned long *nr,
2306 unsigned long *lru_pages)
2307{
2308 int swappiness = mem_cgroup_swappiness(memcg);
2309 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2310 u64 fraction[2];
2311 u64 denominator = 0; /* gcc */
2312 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2313 unsigned long anon_prio, file_prio;
2314 enum scan_balance scan_balance;
2315 unsigned long anon, file;
2316 unsigned long ap, fp;
2317 enum lru_list lru;
2318
2319 /* If we have no swap space, do not bother scanning anon pages. */
2320 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2321 scan_balance = SCAN_FILE;
2322 goto out;
2323 }
2324
2325 /*
2326 * Global reclaim will swap to prevent OOM even with no
2327 * swappiness, but memcg users want to use this knob to
2328 * disable swapping for individual groups completely when
2329 * using the memory controller's swap limit feature would be
2330 * too expensive.
2331 */
2332 if (!global_reclaim(sc) && !swappiness) {
2333 scan_balance = SCAN_FILE;
2334 goto out;
2335 }
2336
2337 /*
2338 * Do not apply any pressure balancing cleverness when the
2339 * system is close to OOM, scan both anon and file equally
2340 * (unless the swappiness setting disagrees with swapping).
2341 */
2342 if (!sc->priority && swappiness) {
2343 scan_balance = SCAN_EQUAL;
2344 goto out;
2345 }
2346
2347 /*
2348 * Prevent the reclaimer from falling into the cache trap: as
2349 * cache pages start out inactive, every cache fault will tip
2350 * the scan balance towards the file LRU. And as the file LRU
2351 * shrinks, so does the window for rotation from references.
2352 * This means we have a runaway feedback loop where a tiny
2353 * thrashing file LRU becomes infinitely more attractive than
2354 * anon pages. Try to detect this based on file LRU size.
2355 */
2356 if (global_reclaim(sc)) {
2357 unsigned long pgdatfile;
2358 unsigned long pgdatfree;
2359 int z;
2360 unsigned long total_high_wmark = 0;
2361
2362 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2363 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2364 node_page_state(pgdat, NR_INACTIVE_FILE);
2365
2366 for (z = 0; z < MAX_NR_ZONES; z++) {
2367 struct zone *zone = &pgdat->node_zones[z];
2368 if (!managed_zone(zone))
2369 continue;
2370
2371 total_high_wmark += high_wmark_pages(zone);
2372 }
2373
2374 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2375 /*
2376 * Force SCAN_ANON if there are enough inactive
2377 * anonymous pages on the LRU in eligible zones.
2378 * Otherwise, the small LRU gets thrashed.
2379 */
2380 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2381 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2382 >> sc->priority) {
2383 scan_balance = SCAN_ANON;
2384 goto out;
2385 }
2386 }
2387 }
2388
2389 /*
2390 * If there is enough inactive page cache, i.e. if the size of the
2391 * inactive list is greater than that of the active list *and* the
2392 * inactive list actually has some pages to scan on this priority, we
2393 * do not reclaim anything from the anonymous working set right now.
2394 * Without the second condition we could end up never scanning an
2395 * lruvec even if it has plenty of old anonymous pages unless the
2396 * system is under heavy pressure.
2397 */
2398 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2399 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2400 scan_balance = SCAN_FILE;
2401 goto out;
2402 }
2403
2404 scan_balance = SCAN_FRACT;
2405
2406 /*
2407 * With swappiness at 100, anonymous and file have the same priority.
2408 * This scanning priority is essentially the inverse of IO cost.
2409 */
2410 anon_prio = swappiness;
2411 file_prio = 200 - anon_prio;
2412
2413 /*
2414 * OK, so we have swap space and a fair amount of page cache
2415 * pages. We use the recently rotated / recently scanned
2416 * ratios to determine how valuable each cache is.
2417 *
2418 * Because workloads change over time (and to avoid overflow)
2419 * we keep these statistics as a floating average, which ends
2420 * up weighing recent references more than old ones.
2421 *
2422 * anon in [0], file in [1]
2423 */
2424
2425 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2426 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2427 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2428 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2429
2430 spin_lock_irq(&pgdat->lru_lock);
2431 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2432 reclaim_stat->recent_scanned[0] /= 2;
2433 reclaim_stat->recent_rotated[0] /= 2;
2434 }
2435
2436 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2437 reclaim_stat->recent_scanned[1] /= 2;
2438 reclaim_stat->recent_rotated[1] /= 2;
2439 }
2440
2441 /*
2442 * The amount of pressure on anon vs file pages is inversely
2443 * proportional to the fraction of recently scanned pages on
2444 * each list that were recently referenced and in active use.
2445 */
2446 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2447 ap /= reclaim_stat->recent_rotated[0] + 1;
2448
2449 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2450 fp /= reclaim_stat->recent_rotated[1] + 1;
2451 spin_unlock_irq(&pgdat->lru_lock);
2452
2453 fraction[0] = ap;
2454 fraction[1] = fp;
2455 denominator = ap + fp + 1;
2456out:
2457 *lru_pages = 0;
2458 for_each_evictable_lru(lru) {
2459 int file = is_file_lru(lru);
2460 unsigned long lruvec_size;
2461 unsigned long scan;
2462 unsigned long protection;
2463
2464 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2465 protection = mem_cgroup_protection(memcg,
2466 sc->memcg_low_reclaim);
2467
2468 if (protection) {
2469 /*
2470 * Scale a cgroup's reclaim pressure by proportioning
2471 * its current usage to its memory.low or memory.min
2472 * setting.
2473 *
2474 * This is important, as otherwise scanning aggression
2475 * becomes extremely binary -- from nothing as we
2476 * approach the memory protection threshold, to totally
2477 * nominal as we exceed it. This results in requiring
2478 * setting extremely liberal protection thresholds. It
2479 * also means we simply get no protection at all if we
2480 * set it too low, which is not ideal.
2481 *
2482 * If there is any protection in place, we reduce scan
2483 * pressure by how much of the total memory used is
2484 * within protection thresholds.
2485 *
2486 * There is one special case: in the first reclaim pass,
2487 * we skip over all groups that are within their low
2488 * protection. If that fails to reclaim enough pages to
2489 * satisfy the reclaim goal, we come back and override
2490 * the best-effort low protection. However, we still
2491 * ideally want to honor how well-behaved groups are in
2492 * that case instead of simply punishing them all
2493 * equally. As such, we reclaim them based on how much
2494 * memory they are using, reducing the scan pressure
2495 * again by how much of the total memory used is under
2496 * hard protection.
2497 */
2498 unsigned long cgroup_size = mem_cgroup_size(memcg);
2499
2500 /* Avoid TOCTOU with earlier protection check */
2501 cgroup_size = max(cgroup_size, protection);
2502
2503 scan = lruvec_size - lruvec_size * protection /
2504 cgroup_size;
2505
2506 /*
2507 * Minimally target SWAP_CLUSTER_MAX pages to keep
2508 * reclaim moving forwards, avoiding decremeting
2509 * sc->priority further than desirable.
2510 */
2511 scan = max(scan, SWAP_CLUSTER_MAX);
2512 } else {
2513 scan = lruvec_size;
2514 }
2515
2516 scan >>= sc->priority;
2517
2518 /*
2519 * If the cgroup's already been deleted, make sure to
2520 * scrape out the remaining cache.
2521 */
2522 if (!scan && !mem_cgroup_online(memcg))
2523 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2524
2525 switch (scan_balance) {
2526 case SCAN_EQUAL:
2527 /* Scan lists relative to size */
2528 break;
2529 case SCAN_FRACT:
2530 /*
2531 * Scan types proportional to swappiness and
2532 * their relative recent reclaim efficiency.
2533 * Make sure we don't miss the last page
2534 * because of a round-off error.
2535 */
2536 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2537 denominator);
2538 break;
2539 case SCAN_FILE:
2540 case SCAN_ANON:
2541 /* Scan one type exclusively */
2542 if ((scan_balance == SCAN_FILE) != file) {
2543 lruvec_size = 0;
2544 scan = 0;
2545 }
2546 break;
2547 default:
2548 /* Look ma, no brain */
2549 BUG();
2550 }
2551
2552 *lru_pages += lruvec_size;
2553 nr[lru] = scan;
2554 }
2555}
2556
2557/*
2558 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2559 */
2560static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2561 struct scan_control *sc, unsigned long *lru_pages)
2562{
2563 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2564 unsigned long nr[NR_LRU_LISTS];
2565 unsigned long targets[NR_LRU_LISTS];
2566 unsigned long nr_to_scan;
2567 enum lru_list lru;
2568 unsigned long nr_reclaimed = 0;
2569 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2570 struct blk_plug plug;
2571 bool scan_adjusted;
2572
2573 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2574
2575 /* Record the original scan target for proportional adjustments later */
2576 memcpy(targets, nr, sizeof(nr));
2577
2578 /*
2579 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2580 * event that can occur when there is little memory pressure e.g.
2581 * multiple streaming readers/writers. Hence, we do not abort scanning
2582 * when the requested number of pages are reclaimed when scanning at
2583 * DEF_PRIORITY on the assumption that the fact we are direct
2584 * reclaiming implies that kswapd is not keeping up and it is best to
2585 * do a batch of work at once. For memcg reclaim one check is made to
2586 * abort proportional reclaim if either the file or anon lru has already
2587 * dropped to zero at the first pass.
2588 */
2589 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2590 sc->priority == DEF_PRIORITY);
2591
2592 blk_start_plug(&plug);
2593 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2594 nr[LRU_INACTIVE_FILE]) {
2595 unsigned long nr_anon, nr_file, percentage;
2596 unsigned long nr_scanned;
2597
2598 for_each_evictable_lru(lru) {
2599 if (nr[lru]) {
2600 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2601 nr[lru] -= nr_to_scan;
2602
2603 nr_reclaimed += shrink_list(lru, nr_to_scan,
2604 lruvec, sc);
2605 }
2606 }
2607
2608 cond_resched();
2609
2610 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2611 continue;
2612
2613 /*
2614 * For kswapd and memcg, reclaim at least the number of pages
2615 * requested. Ensure that the anon and file LRUs are scanned
2616 * proportionally what was requested by get_scan_count(). We
2617 * stop reclaiming one LRU and reduce the amount scanning
2618 * proportional to the original scan target.
2619 */
2620 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2621 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2622
2623 /*
2624 * It's just vindictive to attack the larger once the smaller
2625 * has gone to zero. And given the way we stop scanning the
2626 * smaller below, this makes sure that we only make one nudge
2627 * towards proportionality once we've got nr_to_reclaim.
2628 */
2629 if (!nr_file || !nr_anon)
2630 break;
2631
2632 if (nr_file > nr_anon) {
2633 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2634 targets[LRU_ACTIVE_ANON] + 1;
2635 lru = LRU_BASE;
2636 percentage = nr_anon * 100 / scan_target;
2637 } else {
2638 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2639 targets[LRU_ACTIVE_FILE] + 1;
2640 lru = LRU_FILE;
2641 percentage = nr_file * 100 / scan_target;
2642 }
2643
2644 /* Stop scanning the smaller of the LRU */
2645 nr[lru] = 0;
2646 nr[lru + LRU_ACTIVE] = 0;
2647
2648 /*
2649 * Recalculate the other LRU scan count based on its original
2650 * scan target and the percentage scanning already complete
2651 */
2652 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2653 nr_scanned = targets[lru] - nr[lru];
2654 nr[lru] = targets[lru] * (100 - percentage) / 100;
2655 nr[lru] -= min(nr[lru], nr_scanned);
2656
2657 lru += LRU_ACTIVE;
2658 nr_scanned = targets[lru] - nr[lru];
2659 nr[lru] = targets[lru] * (100 - percentage) / 100;
2660 nr[lru] -= min(nr[lru], nr_scanned);
2661
2662 scan_adjusted = true;
2663 }
2664 blk_finish_plug(&plug);
2665 sc->nr_reclaimed += nr_reclaimed;
2666
2667 /*
2668 * Even if we did not try to evict anon pages at all, we want to
2669 * rebalance the anon lru active/inactive ratio.
2670 */
2671 if (inactive_list_is_low(lruvec, false, sc, true))
2672 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2673 sc, LRU_ACTIVE_ANON);
2674}
2675
2676/* Use reclaim/compaction for costly allocs or under memory pressure */
2677static bool in_reclaim_compaction(struct scan_control *sc)
2678{
2679 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2680 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2681 sc->priority < DEF_PRIORITY - 2))
2682 return true;
2683
2684 return false;
2685}
2686
2687/*
2688 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2689 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2690 * true if more pages should be reclaimed such that when the page allocator
2691 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2692 * It will give up earlier than that if there is difficulty reclaiming pages.
2693 */
2694static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2695 unsigned long nr_reclaimed,
2696 struct scan_control *sc)
2697{
2698 unsigned long pages_for_compaction;
2699 unsigned long inactive_lru_pages;
2700 int z;
2701
2702 /* If not in reclaim/compaction mode, stop */
2703 if (!in_reclaim_compaction(sc))
2704 return false;
2705
2706 /*
2707 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2708 * number of pages that were scanned. This will return to the caller
2709 * with the risk reclaim/compaction and the resulting allocation attempt
2710 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2711 * allocations through requiring that the full LRU list has been scanned
2712 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2713 * scan, but that approximation was wrong, and there were corner cases
2714 * where always a non-zero amount of pages were scanned.
2715 */
2716 if (!nr_reclaimed)
2717 return false;
2718
2719 /* If compaction would go ahead or the allocation would succeed, stop */
2720 for (z = 0; z <= sc->reclaim_idx; z++) {
2721 struct zone *zone = &pgdat->node_zones[z];
2722 if (!managed_zone(zone))
2723 continue;
2724
2725 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2726 case COMPACT_SUCCESS:
2727 case COMPACT_CONTINUE:
2728 return false;
2729 default:
2730 /* check next zone */
2731 ;
2732 }
2733 }
2734
2735 /*
2736 * If we have not reclaimed enough pages for compaction and the
2737 * inactive lists are large enough, continue reclaiming
2738 */
2739 pages_for_compaction = compact_gap(sc->order);
2740 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2741 if (get_nr_swap_pages() > 0)
2742 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2743
2744 return inactive_lru_pages > pages_for_compaction;
2745}
2746
2747static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2748{
2749 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2750 (memcg && memcg_congested(pgdat, memcg));
2751}
2752
2753static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2754{
2755 struct reclaim_state *reclaim_state = current->reclaim_state;
2756 unsigned long nr_reclaimed, nr_scanned;
2757 bool reclaimable = false;
2758
2759 do {
2760 struct mem_cgroup *root = sc->target_mem_cgroup;
2761 unsigned long node_lru_pages = 0;
2762 struct mem_cgroup *memcg;
2763
2764 memset(&sc->nr, 0, sizeof(sc->nr));
2765
2766 nr_reclaimed = sc->nr_reclaimed;
2767 nr_scanned = sc->nr_scanned;
2768
2769 memcg = mem_cgroup_iter(root, NULL, NULL);
2770 do {
2771 unsigned long lru_pages;
2772 unsigned long reclaimed;
2773 unsigned long scanned;
2774
2775 switch (mem_cgroup_protected(root, memcg)) {
2776 case MEMCG_PROT_MIN:
2777 /*
2778 * Hard protection.
2779 * If there is no reclaimable memory, OOM.
2780 */
2781 continue;
2782 case MEMCG_PROT_LOW:
2783 /*
2784 * Soft protection.
2785 * Respect the protection only as long as
2786 * there is an unprotected supply
2787 * of reclaimable memory from other cgroups.
2788 */
2789 if (!sc->memcg_low_reclaim) {
2790 sc->memcg_low_skipped = 1;
2791 continue;
2792 }
2793 memcg_memory_event(memcg, MEMCG_LOW);
2794 break;
2795 case MEMCG_PROT_NONE:
2796 /*
2797 * All protection thresholds breached. We may
2798 * still choose to vary the scan pressure
2799 * applied based on by how much the cgroup in
2800 * question has exceeded its protection
2801 * thresholds (see get_scan_count).
2802 */
2803 break;
2804 }
2805
2806 reclaimed = sc->nr_reclaimed;
2807 scanned = sc->nr_scanned;
2808 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2809 node_lru_pages += lru_pages;
2810
2811 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2812 sc->priority);
2813
2814 /* Record the group's reclaim efficiency */
2815 vmpressure(sc->gfp_mask, memcg, false,
2816 sc->nr_scanned - scanned,
2817 sc->nr_reclaimed - reclaimed);
2818
2819 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2820
2821 if (reclaim_state) {
2822 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823 reclaim_state->reclaimed_slab = 0;
2824 }
2825
2826 /* Record the subtree's reclaim efficiency */
2827 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2828 sc->nr_scanned - nr_scanned,
2829 sc->nr_reclaimed - nr_reclaimed);
2830
2831 if (sc->nr_reclaimed - nr_reclaimed)
2832 reclaimable = true;
2833
2834 if (current_is_kswapd()) {
2835 /*
2836 * If reclaim is isolating dirty pages under writeback,
2837 * it implies that the long-lived page allocation rate
2838 * is exceeding the page laundering rate. Either the
2839 * global limits are not being effective at throttling
2840 * processes due to the page distribution throughout
2841 * zones or there is heavy usage of a slow backing
2842 * device. The only option is to throttle from reclaim
2843 * context which is not ideal as there is no guarantee
2844 * the dirtying process is throttled in the same way
2845 * balance_dirty_pages() manages.
2846 *
2847 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2848 * count the number of pages under pages flagged for
2849 * immediate reclaim and stall if any are encountered
2850 * in the nr_immediate check below.
2851 */
2852 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2853 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2854
2855 /*
2856 * Tag a node as congested if all the dirty pages
2857 * scanned were backed by a congested BDI and
2858 * wait_iff_congested will stall.
2859 */
2860 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2861 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2862
2863 /* Allow kswapd to start writing pages during reclaim.*/
2864 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2865 set_bit(PGDAT_DIRTY, &pgdat->flags);
2866
2867 /*
2868 * If kswapd scans pages marked marked for immediate
2869 * reclaim and under writeback (nr_immediate), it
2870 * implies that pages are cycling through the LRU
2871 * faster than they are written so also forcibly stall.
2872 */
2873 if (sc->nr.immediate)
2874 congestion_wait(BLK_RW_ASYNC, HZ/10);
2875 }
2876
2877 /*
2878 * Legacy memcg will stall in page writeback so avoid forcibly
2879 * stalling in wait_iff_congested().
2880 */
2881 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2882 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2883 set_memcg_congestion(pgdat, root, true);
2884
2885 /*
2886 * Stall direct reclaim for IO completions if underlying BDIs
2887 * and node is congested. Allow kswapd to continue until it
2888 * starts encountering unqueued dirty pages or cycling through
2889 * the LRU too quickly.
2890 */
2891 if (!sc->hibernation_mode && !current_is_kswapd() &&
2892 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2893 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2894
2895 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2896 sc));
2897
2898 /*
2899 * Kswapd gives up on balancing particular nodes after too
2900 * many failures to reclaim anything from them and goes to
2901 * sleep. On reclaim progress, reset the failure counter. A
2902 * successful direct reclaim run will revive a dormant kswapd.
2903 */
2904 if (reclaimable)
2905 pgdat->kswapd_failures = 0;
2906
2907 return reclaimable;
2908}
2909
2910/*
2911 * Returns true if compaction should go ahead for a costly-order request, or
2912 * the allocation would already succeed without compaction. Return false if we
2913 * should reclaim first.
2914 */
2915static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2916{
2917 unsigned long watermark;
2918 enum compact_result suitable;
2919
2920 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2921 if (suitable == COMPACT_SUCCESS)
2922 /* Allocation should succeed already. Don't reclaim. */
2923 return true;
2924 if (suitable == COMPACT_SKIPPED)
2925 /* Compaction cannot yet proceed. Do reclaim. */
2926 return false;
2927
2928 /*
2929 * Compaction is already possible, but it takes time to run and there
2930 * are potentially other callers using the pages just freed. So proceed
2931 * with reclaim to make a buffer of free pages available to give
2932 * compaction a reasonable chance of completing and allocating the page.
2933 * Note that we won't actually reclaim the whole buffer in one attempt
2934 * as the target watermark in should_continue_reclaim() is lower. But if
2935 * we are already above the high+gap watermark, don't reclaim at all.
2936 */
2937 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2938
2939 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2940}
2941
2942/*
2943 * This is the direct reclaim path, for page-allocating processes. We only
2944 * try to reclaim pages from zones which will satisfy the caller's allocation
2945 * request.
2946 *
2947 * If a zone is deemed to be full of pinned pages then just give it a light
2948 * scan then give up on it.
2949 */
2950static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2951{
2952 struct zoneref *z;
2953 struct zone *zone;
2954 unsigned long nr_soft_reclaimed;
2955 unsigned long nr_soft_scanned;
2956 gfp_t orig_mask;
2957 pg_data_t *last_pgdat = NULL;
2958
2959 /*
2960 * If the number of buffer_heads in the machine exceeds the maximum
2961 * allowed level, force direct reclaim to scan the highmem zone as
2962 * highmem pages could be pinning lowmem pages storing buffer_heads
2963 */
2964 orig_mask = sc->gfp_mask;
2965 if (buffer_heads_over_limit) {
2966 sc->gfp_mask |= __GFP_HIGHMEM;
2967 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2968 }
2969
2970 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2971 sc->reclaim_idx, sc->nodemask) {
2972 /*
2973 * Take care memory controller reclaiming has small influence
2974 * to global LRU.
2975 */
2976 if (global_reclaim(sc)) {
2977 if (!cpuset_zone_allowed(zone,
2978 GFP_KERNEL | __GFP_HARDWALL))
2979 continue;
2980
2981 /*
2982 * If we already have plenty of memory free for
2983 * compaction in this zone, don't free any more.
2984 * Even though compaction is invoked for any
2985 * non-zero order, only frequent costly order
2986 * reclamation is disruptive enough to become a
2987 * noticeable problem, like transparent huge
2988 * page allocations.
2989 */
2990 if (IS_ENABLED(CONFIG_COMPACTION) &&
2991 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2992 compaction_ready(zone, sc)) {
2993 sc->compaction_ready = true;
2994 continue;
2995 }
2996
2997 /*
2998 * Shrink each node in the zonelist once. If the
2999 * zonelist is ordered by zone (not the default) then a
3000 * node may be shrunk multiple times but in that case
3001 * the user prefers lower zones being preserved.
3002 */
3003 if (zone->zone_pgdat == last_pgdat)
3004 continue;
3005
3006 /*
3007 * This steals pages from memory cgroups over softlimit
3008 * and returns the number of reclaimed pages and
3009 * scanned pages. This works for global memory pressure
3010 * and balancing, not for a memcg's limit.
3011 */
3012 nr_soft_scanned = 0;
3013 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3014 sc->order, sc->gfp_mask,
3015 &nr_soft_scanned);
3016 sc->nr_reclaimed += nr_soft_reclaimed;
3017 sc->nr_scanned += nr_soft_scanned;
3018 /* need some check for avoid more shrink_zone() */
3019 }
3020
3021 /* See comment about same check for global reclaim above */
3022 if (zone->zone_pgdat == last_pgdat)
3023 continue;
3024 last_pgdat = zone->zone_pgdat;
3025 shrink_node(zone->zone_pgdat, sc);
3026 }
3027
3028 /*
3029 * Restore to original mask to avoid the impact on the caller if we
3030 * promoted it to __GFP_HIGHMEM.
3031 */
3032 sc->gfp_mask = orig_mask;
3033}
3034
3035static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3036{
3037 struct mem_cgroup *memcg;
3038
3039 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3040 do {
3041 unsigned long refaults;
3042 struct lruvec *lruvec;
3043
3044 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3045 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3046 lruvec->refaults = refaults;
3047 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3048}
3049
3050/*
3051 * This is the main entry point to direct page reclaim.
3052 *
3053 * If a full scan of the inactive list fails to free enough memory then we
3054 * are "out of memory" and something needs to be killed.
3055 *
3056 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3057 * high - the zone may be full of dirty or under-writeback pages, which this
3058 * caller can't do much about. We kick the writeback threads and take explicit
3059 * naps in the hope that some of these pages can be written. But if the
3060 * allocating task holds filesystem locks which prevent writeout this might not
3061 * work, and the allocation attempt will fail.
3062 *
3063 * returns: 0, if no pages reclaimed
3064 * else, the number of pages reclaimed
3065 */
3066static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3067 struct scan_control *sc)
3068{
3069 int initial_priority = sc->priority;
3070 pg_data_t *last_pgdat;
3071 struct zoneref *z;
3072 struct zone *zone;
3073retry:
3074 delayacct_freepages_start();
3075
3076 if (global_reclaim(sc))
3077 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3078
3079 do {
3080 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3081 sc->priority);
3082 sc->nr_scanned = 0;
3083 shrink_zones(zonelist, sc);
3084
3085 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3086 break;
3087
3088 if (sc->compaction_ready)
3089 break;
3090
3091 /*
3092 * If we're getting trouble reclaiming, start doing
3093 * writepage even in laptop mode.
3094 */
3095 if (sc->priority < DEF_PRIORITY - 2)
3096 sc->may_writepage = 1;
3097 } while (--sc->priority >= 0);
3098
3099 last_pgdat = NULL;
3100 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3101 sc->nodemask) {
3102 if (zone->zone_pgdat == last_pgdat)
3103 continue;
3104 last_pgdat = zone->zone_pgdat;
3105 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3106 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3107 }
3108
3109 delayacct_freepages_end();
3110
3111 if (sc->nr_reclaimed)
3112 return sc->nr_reclaimed;
3113
3114 /* Aborted reclaim to try compaction? don't OOM, then */
3115 if (sc->compaction_ready)
3116 return 1;
3117
3118 /* Untapped cgroup reserves? Don't OOM, retry. */
3119 if (sc->memcg_low_skipped) {
3120 sc->priority = initial_priority;
3121 sc->memcg_low_reclaim = 1;
3122 sc->memcg_low_skipped = 0;
3123 goto retry;
3124 }
3125
3126 return 0;
3127}
3128
3129static bool allow_direct_reclaim(pg_data_t *pgdat)
3130{
3131 struct zone *zone;
3132 unsigned long pfmemalloc_reserve = 0;
3133 unsigned long free_pages = 0;
3134 int i;
3135 bool wmark_ok;
3136
3137 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3138 return true;
3139
3140 for (i = 0; i <= ZONE_NORMAL; i++) {
3141 zone = &pgdat->node_zones[i];
3142 if (!managed_zone(zone))
3143 continue;
3144
3145 if (!zone_reclaimable_pages(zone))
3146 continue;
3147
3148 pfmemalloc_reserve += min_wmark_pages(zone);
3149 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3150 }
3151
3152 /* If there are no reserves (unexpected config) then do not throttle */
3153 if (!pfmemalloc_reserve)
3154 return true;
3155
3156 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3157
3158 /* kswapd must be awake if processes are being throttled */
3159 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3160 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3161 (enum zone_type)ZONE_NORMAL);
3162 wake_up_interruptible(&pgdat->kswapd_wait);
3163 }
3164
3165 return wmark_ok;
3166}
3167
3168/*
3169 * Throttle direct reclaimers if backing storage is backed by the network
3170 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3171 * depleted. kswapd will continue to make progress and wake the processes
3172 * when the low watermark is reached.
3173 *
3174 * Returns true if a fatal signal was delivered during throttling. If this
3175 * happens, the page allocator should not consider triggering the OOM killer.
3176 */
3177static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3178 nodemask_t *nodemask)
3179{
3180 struct zoneref *z;
3181 struct zone *zone;
3182 pg_data_t *pgdat = NULL;
3183
3184 /*
3185 * Kernel threads should not be throttled as they may be indirectly
3186 * responsible for cleaning pages necessary for reclaim to make forward
3187 * progress. kjournald for example may enter direct reclaim while
3188 * committing a transaction where throttling it could forcing other
3189 * processes to block on log_wait_commit().
3190 */
3191 if (current->flags & PF_KTHREAD)
3192 goto out;
3193
3194 /*
3195 * If a fatal signal is pending, this process should not throttle.
3196 * It should return quickly so it can exit and free its memory
3197 */
3198 if (fatal_signal_pending(current))
3199 goto out;
3200
3201 /*
3202 * Check if the pfmemalloc reserves are ok by finding the first node
3203 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3204 * GFP_KERNEL will be required for allocating network buffers when
3205 * swapping over the network so ZONE_HIGHMEM is unusable.
3206 *
3207 * Throttling is based on the first usable node and throttled processes
3208 * wait on a queue until kswapd makes progress and wakes them. There
3209 * is an affinity then between processes waking up and where reclaim
3210 * progress has been made assuming the process wakes on the same node.
3211 * More importantly, processes running on remote nodes will not compete
3212 * for remote pfmemalloc reserves and processes on different nodes
3213 * should make reasonable progress.
3214 */
3215 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3216 gfp_zone(gfp_mask), nodemask) {
3217 if (zone_idx(zone) > ZONE_NORMAL)
3218 continue;
3219
3220 /* Throttle based on the first usable node */
3221 pgdat = zone->zone_pgdat;
3222 if (allow_direct_reclaim(pgdat))
3223 goto out;
3224 break;
3225 }
3226
3227 /* If no zone was usable by the allocation flags then do not throttle */
3228 if (!pgdat)
3229 goto out;
3230
3231 /* Account for the throttling */
3232 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3233
3234 /*
3235 * If the caller cannot enter the filesystem, it's possible that it
3236 * is due to the caller holding an FS lock or performing a journal
3237 * transaction in the case of a filesystem like ext[3|4]. In this case,
3238 * it is not safe to block on pfmemalloc_wait as kswapd could be
3239 * blocked waiting on the same lock. Instead, throttle for up to a
3240 * second before continuing.
3241 */
3242 if (!(gfp_mask & __GFP_FS)) {
3243 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3244 allow_direct_reclaim(pgdat), HZ);
3245
3246 goto check_pending;
3247 }
3248
3249 /* Throttle until kswapd wakes the process */
3250 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3251 allow_direct_reclaim(pgdat));
3252
3253check_pending:
3254 if (fatal_signal_pending(current))
3255 return true;
3256
3257out:
3258 return false;
3259}
3260
3261unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3262 gfp_t gfp_mask, nodemask_t *nodemask)
3263{
3264 unsigned long nr_reclaimed;
3265 struct scan_control sc = {
3266 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3267 .gfp_mask = current_gfp_context(gfp_mask),
3268 .reclaim_idx = gfp_zone(gfp_mask),
3269 .order = order,
3270 .nodemask = nodemask,
3271 .priority = DEF_PRIORITY,
3272 .may_writepage = !laptop_mode,
3273 .may_unmap = 1,
3274 .may_swap = 1,
3275 };
3276
3277 /*
3278 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3279 * Confirm they are large enough for max values.
3280 */
3281 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3282 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3283 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3284
3285 /*
3286 * Do not enter reclaim if fatal signal was delivered while throttled.
3287 * 1 is returned so that the page allocator does not OOM kill at this
3288 * point.
3289 */
3290 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3291 return 1;
3292
3293 set_task_reclaim_state(current, &sc.reclaim_state);
3294 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3295
3296 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3297
3298 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3299 set_task_reclaim_state(current, NULL);
3300
3301 return nr_reclaimed;
3302}
3303
3304#ifdef CONFIG_MEMCG
3305
3306/* Only used by soft limit reclaim. Do not reuse for anything else. */
3307unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3308 gfp_t gfp_mask, bool noswap,
3309 pg_data_t *pgdat,
3310 unsigned long *nr_scanned)
3311{
3312 struct scan_control sc = {
3313 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3314 .target_mem_cgroup = memcg,
3315 .may_writepage = !laptop_mode,
3316 .may_unmap = 1,
3317 .reclaim_idx = MAX_NR_ZONES - 1,
3318 .may_swap = !noswap,
3319 };
3320 unsigned long lru_pages;
3321
3322 WARN_ON_ONCE(!current->reclaim_state);
3323
3324 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3325 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3326
3327 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3328 sc.gfp_mask);
3329
3330 /*
3331 * NOTE: Although we can get the priority field, using it
3332 * here is not a good idea, since it limits the pages we can scan.
3333 * if we don't reclaim here, the shrink_node from balance_pgdat
3334 * will pick up pages from other mem cgroup's as well. We hack
3335 * the priority and make it zero.
3336 */
3337 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3338
3339 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3340
3341 *nr_scanned = sc.nr_scanned;
3342
3343 return sc.nr_reclaimed;
3344}
3345
3346unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3347 unsigned long nr_pages,
3348 gfp_t gfp_mask,
3349 bool may_swap)
3350{
3351 struct zonelist *zonelist;
3352 unsigned long nr_reclaimed;
3353 unsigned long pflags;
3354 int nid;
3355 unsigned int noreclaim_flag;
3356 struct scan_control sc = {
3357 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3358 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3359 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3360 .reclaim_idx = MAX_NR_ZONES - 1,
3361 .target_mem_cgroup = memcg,
3362 .priority = DEF_PRIORITY,
3363 .may_writepage = !laptop_mode,
3364 .may_unmap = 1,
3365 .may_swap = may_swap,
3366 };
3367
3368 set_task_reclaim_state(current, &sc.reclaim_state);
3369 /*
3370 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3371 * take care of from where we get pages. So the node where we start the
3372 * scan does not need to be the current node.
3373 */
3374 nid = mem_cgroup_select_victim_node(memcg);
3375
3376 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3377
3378 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3379
3380 psi_memstall_enter(&pflags);
3381 noreclaim_flag = memalloc_noreclaim_save();
3382
3383 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3384
3385 memalloc_noreclaim_restore(noreclaim_flag);
3386 psi_memstall_leave(&pflags);
3387
3388 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3389 set_task_reclaim_state(current, NULL);
3390
3391 return nr_reclaimed;
3392}
3393#endif
3394
3395static void age_active_anon(struct pglist_data *pgdat,
3396 struct scan_control *sc)
3397{
3398 struct mem_cgroup *memcg;
3399
3400 if (!total_swap_pages)
3401 return;
3402
3403 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3404 do {
3405 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3406
3407 if (inactive_list_is_low(lruvec, false, sc, true))
3408 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3409 sc, LRU_ACTIVE_ANON);
3410
3411 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3412 } while (memcg);
3413}
3414
3415static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3416{
3417 int i;
3418 struct zone *zone;
3419
3420 /*
3421 * Check for watermark boosts top-down as the higher zones
3422 * are more likely to be boosted. Both watermarks and boosts
3423 * should not be checked at the time time as reclaim would
3424 * start prematurely when there is no boosting and a lower
3425 * zone is balanced.
3426 */
3427 for (i = classzone_idx; i >= 0; i--) {
3428 zone = pgdat->node_zones + i;
3429 if (!managed_zone(zone))
3430 continue;
3431
3432 if (zone->watermark_boost)
3433 return true;
3434 }
3435
3436 return false;
3437}
3438
3439/*
3440 * Returns true if there is an eligible zone balanced for the request order
3441 * and classzone_idx
3442 */
3443static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3444{
3445 int i;
3446 unsigned long mark = -1;
3447 struct zone *zone;
3448
3449 /*
3450 * Check watermarks bottom-up as lower zones are more likely to
3451 * meet watermarks.
3452 */
3453 for (i = 0; i <= classzone_idx; i++) {
3454 zone = pgdat->node_zones + i;
3455
3456 if (!managed_zone(zone))
3457 continue;
3458
3459 mark = high_wmark_pages(zone);
3460 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3461 return true;
3462 }
3463
3464 /*
3465 * If a node has no populated zone within classzone_idx, it does not
3466 * need balancing by definition. This can happen if a zone-restricted
3467 * allocation tries to wake a remote kswapd.
3468 */
3469 if (mark == -1)
3470 return true;
3471
3472 return false;
3473}
3474
3475/* Clear pgdat state for congested, dirty or under writeback. */
3476static void clear_pgdat_congested(pg_data_t *pgdat)
3477{
3478 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3479 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3480 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3481}
3482
3483/*
3484 * Prepare kswapd for sleeping. This verifies that there are no processes
3485 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3486 *
3487 * Returns true if kswapd is ready to sleep
3488 */
3489static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3490{
3491 /*
3492 * The throttled processes are normally woken up in balance_pgdat() as
3493 * soon as allow_direct_reclaim() is true. But there is a potential
3494 * race between when kswapd checks the watermarks and a process gets
3495 * throttled. There is also a potential race if processes get
3496 * throttled, kswapd wakes, a large process exits thereby balancing the
3497 * zones, which causes kswapd to exit balance_pgdat() before reaching
3498 * the wake up checks. If kswapd is going to sleep, no process should
3499 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3500 * the wake up is premature, processes will wake kswapd and get
3501 * throttled again. The difference from wake ups in balance_pgdat() is
3502 * that here we are under prepare_to_wait().
3503 */
3504 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3505 wake_up_all(&pgdat->pfmemalloc_wait);
3506
3507 /* Hopeless node, leave it to direct reclaim */
3508 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3509 return true;
3510
3511 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3512 clear_pgdat_congested(pgdat);
3513 return true;
3514 }
3515
3516 return false;
3517}
3518
3519/*
3520 * kswapd shrinks a node of pages that are at or below the highest usable
3521 * zone that is currently unbalanced.
3522 *
3523 * Returns true if kswapd scanned at least the requested number of pages to
3524 * reclaim or if the lack of progress was due to pages under writeback.
3525 * This is used to determine if the scanning priority needs to be raised.
3526 */
3527static bool kswapd_shrink_node(pg_data_t *pgdat,
3528 struct scan_control *sc)
3529{
3530 struct zone *zone;
3531 int z;
3532
3533 /* Reclaim a number of pages proportional to the number of zones */
3534 sc->nr_to_reclaim = 0;
3535 for (z = 0; z <= sc->reclaim_idx; z++) {
3536 zone = pgdat->node_zones + z;
3537 if (!managed_zone(zone))
3538 continue;
3539
3540 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3541 }
3542
3543 /*
3544 * Historically care was taken to put equal pressure on all zones but
3545 * now pressure is applied based on node LRU order.
3546 */
3547 shrink_node(pgdat, sc);
3548
3549 /*
3550 * Fragmentation may mean that the system cannot be rebalanced for
3551 * high-order allocations. If twice the allocation size has been
3552 * reclaimed then recheck watermarks only at order-0 to prevent
3553 * excessive reclaim. Assume that a process requested a high-order
3554 * can direct reclaim/compact.
3555 */
3556 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3557 sc->order = 0;
3558
3559 return sc->nr_scanned >= sc->nr_to_reclaim;
3560}
3561
3562/*
3563 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3564 * that are eligible for use by the caller until at least one zone is
3565 * balanced.
3566 *
3567 * Returns the order kswapd finished reclaiming at.
3568 *
3569 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3570 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3571 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3572 * or lower is eligible for reclaim until at least one usable zone is
3573 * balanced.
3574 */
3575static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3576{
3577 int i;
3578 unsigned long nr_soft_reclaimed;
3579 unsigned long nr_soft_scanned;
3580 unsigned long pflags;
3581 unsigned long nr_boost_reclaim;
3582 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3583 bool boosted;
3584 struct zone *zone;
3585 struct scan_control sc = {
3586 .gfp_mask = GFP_KERNEL,
3587 .order = order,
3588 .may_unmap = 1,
3589 };
3590
3591 set_task_reclaim_state(current, &sc.reclaim_state);
3592 psi_memstall_enter(&pflags);
3593 __fs_reclaim_acquire();
3594
3595 count_vm_event(PAGEOUTRUN);
3596
3597 /*
3598 * Account for the reclaim boost. Note that the zone boost is left in
3599 * place so that parallel allocations that are near the watermark will
3600 * stall or direct reclaim until kswapd is finished.
3601 */
3602 nr_boost_reclaim = 0;
3603 for (i = 0; i <= classzone_idx; i++) {
3604 zone = pgdat->node_zones + i;
3605 if (!managed_zone(zone))
3606 continue;
3607
3608 nr_boost_reclaim += zone->watermark_boost;
3609 zone_boosts[i] = zone->watermark_boost;
3610 }
3611 boosted = nr_boost_reclaim;
3612
3613restart:
3614 sc.priority = DEF_PRIORITY;
3615 do {
3616 unsigned long nr_reclaimed = sc.nr_reclaimed;
3617 bool raise_priority = true;
3618 bool balanced;
3619 bool ret;
3620
3621 sc.reclaim_idx = classzone_idx;
3622
3623 /*
3624 * If the number of buffer_heads exceeds the maximum allowed
3625 * then consider reclaiming from all zones. This has a dual
3626 * purpose -- on 64-bit systems it is expected that
3627 * buffer_heads are stripped during active rotation. On 32-bit
3628 * systems, highmem pages can pin lowmem memory and shrinking
3629 * buffers can relieve lowmem pressure. Reclaim may still not
3630 * go ahead if all eligible zones for the original allocation
3631 * request are balanced to avoid excessive reclaim from kswapd.
3632 */
3633 if (buffer_heads_over_limit) {
3634 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3635 zone = pgdat->node_zones + i;
3636 if (!managed_zone(zone))
3637 continue;
3638
3639 sc.reclaim_idx = i;
3640 break;
3641 }
3642 }
3643
3644 /*
3645 * If the pgdat is imbalanced then ignore boosting and preserve
3646 * the watermarks for a later time and restart. Note that the
3647 * zone watermarks will be still reset at the end of balancing
3648 * on the grounds that the normal reclaim should be enough to
3649 * re-evaluate if boosting is required when kswapd next wakes.
3650 */
3651 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3652 if (!balanced && nr_boost_reclaim) {
3653 nr_boost_reclaim = 0;
3654 goto restart;
3655 }
3656
3657 /*
3658 * If boosting is not active then only reclaim if there are no
3659 * eligible zones. Note that sc.reclaim_idx is not used as
3660 * buffer_heads_over_limit may have adjusted it.
3661 */
3662 if (!nr_boost_reclaim && balanced)
3663 goto out;
3664
3665 /* Limit the priority of boosting to avoid reclaim writeback */
3666 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3667 raise_priority = false;
3668
3669 /*
3670 * Do not writeback or swap pages for boosted reclaim. The
3671 * intent is to relieve pressure not issue sub-optimal IO
3672 * from reclaim context. If no pages are reclaimed, the
3673 * reclaim will be aborted.
3674 */
3675 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3676 sc.may_swap = !nr_boost_reclaim;
3677
3678 /*
3679 * Do some background aging of the anon list, to give
3680 * pages a chance to be referenced before reclaiming. All
3681 * pages are rotated regardless of classzone as this is
3682 * about consistent aging.
3683 */
3684 age_active_anon(pgdat, &sc);
3685
3686 /*
3687 * If we're getting trouble reclaiming, start doing writepage
3688 * even in laptop mode.
3689 */
3690 if (sc.priority < DEF_PRIORITY - 2)
3691 sc.may_writepage = 1;
3692
3693 /* Call soft limit reclaim before calling shrink_node. */
3694 sc.nr_scanned = 0;
3695 nr_soft_scanned = 0;
3696 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3697 sc.gfp_mask, &nr_soft_scanned);
3698 sc.nr_reclaimed += nr_soft_reclaimed;
3699
3700 /*
3701 * There should be no need to raise the scanning priority if
3702 * enough pages are already being scanned that that high
3703 * watermark would be met at 100% efficiency.
3704 */
3705 if (kswapd_shrink_node(pgdat, &sc))
3706 raise_priority = false;
3707
3708 /*
3709 * If the low watermark is met there is no need for processes
3710 * to be throttled on pfmemalloc_wait as they should not be
3711 * able to safely make forward progress. Wake them
3712 */
3713 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3714 allow_direct_reclaim(pgdat))
3715 wake_up_all(&pgdat->pfmemalloc_wait);
3716
3717 /* Check if kswapd should be suspending */
3718 __fs_reclaim_release();
3719 ret = try_to_freeze();
3720 __fs_reclaim_acquire();
3721 if (ret || kthread_should_stop())
3722 break;
3723
3724 /*
3725 * Raise priority if scanning rate is too low or there was no
3726 * progress in reclaiming pages
3727 */
3728 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3729 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3730
3731 /*
3732 * If reclaim made no progress for a boost, stop reclaim as
3733 * IO cannot be queued and it could be an infinite loop in
3734 * extreme circumstances.
3735 */
3736 if (nr_boost_reclaim && !nr_reclaimed)
3737 break;
3738
3739 if (raise_priority || !nr_reclaimed)
3740 sc.priority--;
3741 } while (sc.priority >= 1);
3742
3743 if (!sc.nr_reclaimed)
3744 pgdat->kswapd_failures++;
3745
3746out:
3747 /* If reclaim was boosted, account for the reclaim done in this pass */
3748 if (boosted) {
3749 unsigned long flags;
3750
3751 for (i = 0; i <= classzone_idx; i++) {
3752 if (!zone_boosts[i])
3753 continue;
3754
3755 /* Increments are under the zone lock */
3756 zone = pgdat->node_zones + i;
3757 spin_lock_irqsave(&zone->lock, flags);
3758 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3759 spin_unlock_irqrestore(&zone->lock, flags);
3760 }
3761
3762 /*
3763 * As there is now likely space, wakeup kcompact to defragment
3764 * pageblocks.
3765 */
3766 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3767 }
3768
3769 snapshot_refaults(NULL, pgdat);
3770 __fs_reclaim_release();
3771 psi_memstall_leave(&pflags);
3772 set_task_reclaim_state(current, NULL);
3773
3774 /*
3775 * Return the order kswapd stopped reclaiming at as
3776 * prepare_kswapd_sleep() takes it into account. If another caller
3777 * entered the allocator slow path while kswapd was awake, order will
3778 * remain at the higher level.
3779 */
3780 return sc.order;
3781}
3782
3783/*
3784 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3785 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3786 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3787 * after previous reclaim attempt (node is still unbalanced). In that case
3788 * return the zone index of the previous kswapd reclaim cycle.
3789 */
3790static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3791 enum zone_type prev_classzone_idx)
3792{
3793 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3794 return prev_classzone_idx;
3795 return pgdat->kswapd_classzone_idx;
3796}
3797
3798static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3799 unsigned int classzone_idx)
3800{
3801 long remaining = 0;
3802 DEFINE_WAIT(wait);
3803
3804 if (freezing(current) || kthread_should_stop())
3805 return;
3806
3807 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3808
3809 /*
3810 * Try to sleep for a short interval. Note that kcompactd will only be
3811 * woken if it is possible to sleep for a short interval. This is
3812 * deliberate on the assumption that if reclaim cannot keep an
3813 * eligible zone balanced that it's also unlikely that compaction will
3814 * succeed.
3815 */
3816 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3817 /*
3818 * Compaction records what page blocks it recently failed to
3819 * isolate pages from and skips them in the future scanning.
3820 * When kswapd is going to sleep, it is reasonable to assume
3821 * that pages and compaction may succeed so reset the cache.
3822 */
3823 reset_isolation_suitable(pgdat);
3824
3825 /*
3826 * We have freed the memory, now we should compact it to make
3827 * allocation of the requested order possible.
3828 */
3829 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3830
3831 remaining = schedule_timeout(HZ/10);
3832
3833 /*
3834 * If woken prematurely then reset kswapd_classzone_idx and
3835 * order. The values will either be from a wakeup request or
3836 * the previous request that slept prematurely.
3837 */
3838 if (remaining) {
3839 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3840 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3841 }
3842
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3844 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3845 }
3846
3847 /*
3848 * After a short sleep, check if it was a premature sleep. If not, then
3849 * go fully to sleep until explicitly woken up.
3850 */
3851 if (!remaining &&
3852 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3853 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3854
3855 /*
3856 * vmstat counters are not perfectly accurate and the estimated
3857 * value for counters such as NR_FREE_PAGES can deviate from the
3858 * true value by nr_online_cpus * threshold. To avoid the zone
3859 * watermarks being breached while under pressure, we reduce the
3860 * per-cpu vmstat threshold while kswapd is awake and restore
3861 * them before going back to sleep.
3862 */
3863 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3864
3865 if (!kthread_should_stop())
3866 schedule();
3867
3868 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3869 } else {
3870 if (remaining)
3871 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3872 else
3873 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3874 }
3875 finish_wait(&pgdat->kswapd_wait, &wait);
3876}
3877
3878/*
3879 * The background pageout daemon, started as a kernel thread
3880 * from the init process.
3881 *
3882 * This basically trickles out pages so that we have _some_
3883 * free memory available even if there is no other activity
3884 * that frees anything up. This is needed for things like routing
3885 * etc, where we otherwise might have all activity going on in
3886 * asynchronous contexts that cannot page things out.
3887 *
3888 * If there are applications that are active memory-allocators
3889 * (most normal use), this basically shouldn't matter.
3890 */
3891static int kswapd(void *p)
3892{
3893 unsigned int alloc_order, reclaim_order;
3894 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3895 pg_data_t *pgdat = (pg_data_t*)p;
3896 struct task_struct *tsk = current;
3897 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3898
3899 if (!cpumask_empty(cpumask))
3900 set_cpus_allowed_ptr(tsk, cpumask);
3901
3902 /*
3903 * Tell the memory management that we're a "memory allocator",
3904 * and that if we need more memory we should get access to it
3905 * regardless (see "__alloc_pages()"). "kswapd" should
3906 * never get caught in the normal page freeing logic.
3907 *
3908 * (Kswapd normally doesn't need memory anyway, but sometimes
3909 * you need a small amount of memory in order to be able to
3910 * page out something else, and this flag essentially protects
3911 * us from recursively trying to free more memory as we're
3912 * trying to free the first piece of memory in the first place).
3913 */
3914 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3915 set_freezable();
3916
3917 pgdat->kswapd_order = 0;
3918 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3919 for ( ; ; ) {
3920 bool ret;
3921
3922 alloc_order = reclaim_order = pgdat->kswapd_order;
3923 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3924
3925kswapd_try_sleep:
3926 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3927 classzone_idx);
3928
3929 /* Read the new order and classzone_idx */
3930 alloc_order = reclaim_order = pgdat->kswapd_order;
3931 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3932 pgdat->kswapd_order = 0;
3933 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3934
3935 ret = try_to_freeze();
3936 if (kthread_should_stop())
3937 break;
3938
3939 /*
3940 * We can speed up thawing tasks if we don't call balance_pgdat
3941 * after returning from the refrigerator
3942 */
3943 if (ret)
3944 continue;
3945
3946 /*
3947 * Reclaim begins at the requested order but if a high-order
3948 * reclaim fails then kswapd falls back to reclaiming for
3949 * order-0. If that happens, kswapd will consider sleeping
3950 * for the order it finished reclaiming at (reclaim_order)
3951 * but kcompactd is woken to compact for the original
3952 * request (alloc_order).
3953 */
3954 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3955 alloc_order);
3956 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3957 if (reclaim_order < alloc_order)
3958 goto kswapd_try_sleep;
3959 }
3960
3961 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3962
3963 return 0;
3964}
3965
3966/*
3967 * A zone is low on free memory or too fragmented for high-order memory. If
3968 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3969 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3970 * has failed or is not needed, still wake up kcompactd if only compaction is
3971 * needed.
3972 */
3973void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3974 enum zone_type classzone_idx)
3975{
3976 pg_data_t *pgdat;
3977
3978 if (!managed_zone(zone))
3979 return;
3980
3981 if (!cpuset_zone_allowed(zone, gfp_flags))
3982 return;
3983 pgdat = zone->zone_pgdat;
3984
3985 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3986 pgdat->kswapd_classzone_idx = classzone_idx;
3987 else
3988 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3989 classzone_idx);
3990 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3991 if (!waitqueue_active(&pgdat->kswapd_wait))
3992 return;
3993
3994 /* Hopeless node, leave it to direct reclaim if possible */
3995 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3996 (pgdat_balanced(pgdat, order, classzone_idx) &&
3997 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3998 /*
3999 * There may be plenty of free memory available, but it's too
4000 * fragmented for high-order allocations. Wake up kcompactd
4001 * and rely on compaction_suitable() to determine if it's
4002 * needed. If it fails, it will defer subsequent attempts to
4003 * ratelimit its work.
4004 */
4005 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4006 wakeup_kcompactd(pgdat, order, classzone_idx);
4007 return;
4008 }
4009
4010 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4011 gfp_flags);
4012 wake_up_interruptible(&pgdat->kswapd_wait);
4013}
4014
4015#ifdef CONFIG_HIBERNATION
4016/*
4017 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4018 * freed pages.
4019 *
4020 * Rather than trying to age LRUs the aim is to preserve the overall
4021 * LRU order by reclaiming preferentially
4022 * inactive > active > active referenced > active mapped
4023 */
4024unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4025{
4026 struct scan_control sc = {
4027 .nr_to_reclaim = nr_to_reclaim,
4028 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4029 .reclaim_idx = MAX_NR_ZONES - 1,
4030 .priority = DEF_PRIORITY,
4031 .may_writepage = 1,
4032 .may_unmap = 1,
4033 .may_swap = 1,
4034 .hibernation_mode = 1,
4035 };
4036 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4037 unsigned long nr_reclaimed;
4038 unsigned int noreclaim_flag;
4039
4040 fs_reclaim_acquire(sc.gfp_mask);
4041 noreclaim_flag = memalloc_noreclaim_save();
4042 set_task_reclaim_state(current, &sc.reclaim_state);
4043
4044 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4045
4046 set_task_reclaim_state(current, NULL);
4047 memalloc_noreclaim_restore(noreclaim_flag);
4048 fs_reclaim_release(sc.gfp_mask);
4049
4050 return nr_reclaimed;
4051}
4052#endif /* CONFIG_HIBERNATION */
4053
4054/* It's optimal to keep kswapds on the same CPUs as their memory, but
4055 not required for correctness. So if the last cpu in a node goes
4056 away, we get changed to run anywhere: as the first one comes back,
4057 restore their cpu bindings. */
4058static int kswapd_cpu_online(unsigned int cpu)
4059{
4060 int nid;
4061
4062 for_each_node_state(nid, N_MEMORY) {
4063 pg_data_t *pgdat = NODE_DATA(nid);
4064 const struct cpumask *mask;
4065
4066 mask = cpumask_of_node(pgdat->node_id);
4067
4068 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4069 /* One of our CPUs online: restore mask */
4070 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4071 }
4072 return 0;
4073}
4074
4075/*
4076 * This kswapd start function will be called by init and node-hot-add.
4077 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4078 */
4079int kswapd_run(int nid)
4080{
4081 pg_data_t *pgdat = NODE_DATA(nid);
4082 int ret = 0;
4083
4084 if (pgdat->kswapd)
4085 return 0;
4086
4087 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4088 if (IS_ERR(pgdat->kswapd)) {
4089 /* failure at boot is fatal */
4090 BUG_ON(system_state < SYSTEM_RUNNING);
4091 pr_err("Failed to start kswapd on node %d\n", nid);
4092 ret = PTR_ERR(pgdat->kswapd);
4093 pgdat->kswapd = NULL;
4094 }
4095 return ret;
4096}
4097
4098/*
4099 * Called by memory hotplug when all memory in a node is offlined. Caller must
4100 * hold mem_hotplug_begin/end().
4101 */
4102void kswapd_stop(int nid)
4103{
4104 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4105
4106 if (kswapd) {
4107 kthread_stop(kswapd);
4108 NODE_DATA(nid)->kswapd = NULL;
4109 }
4110}
4111
4112static int __init kswapd_init(void)
4113{
4114 int nid, ret;
4115
4116 swap_setup();
4117 for_each_node_state(nid, N_MEMORY)
4118 kswapd_run(nid);
4119 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4120 "mm/vmscan:online", kswapd_cpu_online,
4121 NULL);
4122 WARN_ON(ret < 0);
4123 return 0;
4124}
4125
4126module_init(kswapd_init)
4127
4128#ifdef CONFIG_NUMA
4129/*
4130 * Node reclaim mode
4131 *
4132 * If non-zero call node_reclaim when the number of free pages falls below
4133 * the watermarks.
4134 */
4135int node_reclaim_mode __read_mostly;
4136
4137#define RECLAIM_OFF 0
4138#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4139#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4140#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4141
4142/*
4143 * Priority for NODE_RECLAIM. This determines the fraction of pages
4144 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4145 * a zone.
4146 */
4147#define NODE_RECLAIM_PRIORITY 4
4148
4149/*
4150 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4151 * occur.
4152 */
4153int sysctl_min_unmapped_ratio = 1;
4154
4155/*
4156 * If the number of slab pages in a zone grows beyond this percentage then
4157 * slab reclaim needs to occur.
4158 */
4159int sysctl_min_slab_ratio = 5;
4160
4161static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4162{
4163 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4164 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4165 node_page_state(pgdat, NR_ACTIVE_FILE);
4166
4167 /*
4168 * It's possible for there to be more file mapped pages than
4169 * accounted for by the pages on the file LRU lists because
4170 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4171 */
4172 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4173}
4174
4175/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4176static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4177{
4178 unsigned long nr_pagecache_reclaimable;
4179 unsigned long delta = 0;
4180
4181 /*
4182 * If RECLAIM_UNMAP is set, then all file pages are considered
4183 * potentially reclaimable. Otherwise, we have to worry about
4184 * pages like swapcache and node_unmapped_file_pages() provides
4185 * a better estimate
4186 */
4187 if (node_reclaim_mode & RECLAIM_UNMAP)
4188 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4189 else
4190 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4191
4192 /* If we can't clean pages, remove dirty pages from consideration */
4193 if (!(node_reclaim_mode & RECLAIM_WRITE))
4194 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4195
4196 /* Watch for any possible underflows due to delta */
4197 if (unlikely(delta > nr_pagecache_reclaimable))
4198 delta = nr_pagecache_reclaimable;
4199
4200 return nr_pagecache_reclaimable - delta;
4201}
4202
4203/*
4204 * Try to free up some pages from this node through reclaim.
4205 */
4206static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4207{
4208 /* Minimum pages needed in order to stay on node */
4209 const unsigned long nr_pages = 1 << order;
4210 struct task_struct *p = current;
4211 unsigned int noreclaim_flag;
4212 struct scan_control sc = {
4213 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4214 .gfp_mask = current_gfp_context(gfp_mask),
4215 .order = order,
4216 .priority = NODE_RECLAIM_PRIORITY,
4217 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4218 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4219 .may_swap = 1,
4220 .reclaim_idx = gfp_zone(gfp_mask),
4221 };
4222
4223 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4224 sc.gfp_mask);
4225
4226 cond_resched();
4227 fs_reclaim_acquire(sc.gfp_mask);
4228 /*
4229 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4230 * and we also need to be able to write out pages for RECLAIM_WRITE
4231 * and RECLAIM_UNMAP.
4232 */
4233 noreclaim_flag = memalloc_noreclaim_save();
4234 p->flags |= PF_SWAPWRITE;
4235 set_task_reclaim_state(p, &sc.reclaim_state);
4236
4237 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4238 /*
4239 * Free memory by calling shrink node with increasing
4240 * priorities until we have enough memory freed.
4241 */
4242 do {
4243 shrink_node(pgdat, &sc);
4244 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4245 }
4246
4247 set_task_reclaim_state(p, NULL);
4248 current->flags &= ~PF_SWAPWRITE;
4249 memalloc_noreclaim_restore(noreclaim_flag);
4250 fs_reclaim_release(sc.gfp_mask);
4251
4252 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4253
4254 return sc.nr_reclaimed >= nr_pages;
4255}
4256
4257int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4258{
4259 int ret;
4260
4261 /*
4262 * Node reclaim reclaims unmapped file backed pages and
4263 * slab pages if we are over the defined limits.
4264 *
4265 * A small portion of unmapped file backed pages is needed for
4266 * file I/O otherwise pages read by file I/O will be immediately
4267 * thrown out if the node is overallocated. So we do not reclaim
4268 * if less than a specified percentage of the node is used by
4269 * unmapped file backed pages.
4270 */
4271 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4272 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4273 return NODE_RECLAIM_FULL;
4274
4275 /*
4276 * Do not scan if the allocation should not be delayed.
4277 */
4278 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4279 return NODE_RECLAIM_NOSCAN;
4280
4281 /*
4282 * Only run node reclaim on the local node or on nodes that do not
4283 * have associated processors. This will favor the local processor
4284 * over remote processors and spread off node memory allocations
4285 * as wide as possible.
4286 */
4287 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4288 return NODE_RECLAIM_NOSCAN;
4289
4290 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4291 return NODE_RECLAIM_NOSCAN;
4292
4293 ret = __node_reclaim(pgdat, gfp_mask, order);
4294 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4295
4296 if (!ret)
4297 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4298
4299 return ret;
4300}
4301#endif
4302
4303/*
4304 * page_evictable - test whether a page is evictable
4305 * @page: the page to test
4306 *
4307 * Test whether page is evictable--i.e., should be placed on active/inactive
4308 * lists vs unevictable list.
4309 *
4310 * Reasons page might not be evictable:
4311 * (1) page's mapping marked unevictable
4312 * (2) page is part of an mlocked VMA
4313 *
4314 */
4315int page_evictable(struct page *page)
4316{
4317 int ret;
4318
4319 /* Prevent address_space of inode and swap cache from being freed */
4320 rcu_read_lock();
4321 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4322 rcu_read_unlock();
4323 return ret;
4324}
4325
4326/**
4327 * check_move_unevictable_pages - check pages for evictability and move to
4328 * appropriate zone lru list
4329 * @pvec: pagevec with lru pages to check
4330 *
4331 * Checks pages for evictability, if an evictable page is in the unevictable
4332 * lru list, moves it to the appropriate evictable lru list. This function
4333 * should be only used for lru pages.
4334 */
4335void check_move_unevictable_pages(struct pagevec *pvec)
4336{
4337 struct lruvec *lruvec;
4338 struct pglist_data *pgdat = NULL;
4339 int pgscanned = 0;
4340 int pgrescued = 0;
4341 int i;
4342
4343 for (i = 0; i < pvec->nr; i++) {
4344 struct page *page = pvec->pages[i];
4345 struct pglist_data *pagepgdat = page_pgdat(page);
4346
4347 pgscanned++;
4348 if (pagepgdat != pgdat) {
4349 if (pgdat)
4350 spin_unlock_irq(&pgdat->lru_lock);
4351 pgdat = pagepgdat;
4352 spin_lock_irq(&pgdat->lru_lock);
4353 }
4354 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4355
4356 if (!PageLRU(page) || !PageUnevictable(page))
4357 continue;
4358
4359 if (page_evictable(page)) {
4360 enum lru_list lru = page_lru_base_type(page);
4361
4362 VM_BUG_ON_PAGE(PageActive(page), page);
4363 ClearPageUnevictable(page);
4364 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4365 add_page_to_lru_list(page, lruvec, lru);
4366 pgrescued++;
4367 }
4368 }
4369
4370 if (pgdat) {
4371 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4372 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4373 spin_unlock_irq(&pgdat->lru_lock);
4374 }
4375}
4376EXPORT_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