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