Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
17#include <linux/mm.h>
18#include <linux/sched/mm.h>
19#include <linux/module.h>
20#include <linux/gfp.h>
21#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
26#include <linux/vmpressure.h>
27#include <linux/vmstat.h>
28#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
34#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
39#include <linux/compaction.h>
40#include <linux/notifier.h>
41#include <linux/rwsem.h>
42#include <linux/delay.h>
43#include <linux/kthread.h>
44#include <linux/freezer.h>
45#include <linux/memcontrol.h>
46#include <linux/delayacct.h>
47#include <linux/sysctl.h>
48#include <linux/oom.h>
49#include <linux/pagevec.h>
50#include <linux/prefetch.h>
51#include <linux/printk.h>
52#include <linux/dax.h>
53#include <linux/psi.h>
54
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
59#include <linux/balloon_compaction.h>
60
61#include "internal.h"
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
66struct scan_control {
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
70 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
75
76 /*
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
81
82 /*
83 * Scan pressure balancing between anon and file LRUs
84 */
85 unsigned long anon_cost;
86 unsigned long file_cost;
87
88 /* Can active pages be deactivated as part of reclaim? */
89#define DEACTIVATE_ANON 1
90#define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
94
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage:1;
97
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
100
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
103
104 /*
105 * Cgroups are not reclaimed below their configured memory.low,
106 * unless we threaten to OOM. If any cgroups are skipped due to
107 * memory.low and nothing was reclaimed, go back for memory.low.
108 */
109 unsigned int memcg_low_reclaim:1;
110 unsigned int memcg_low_skipped:1;
111
112 unsigned int hibernation_mode:1;
113
114 /* One of the zones is ready for compaction */
115 unsigned int compaction_ready:1;
116
117 /* There is easily reclaimable cold cache in the current node */
118 unsigned int cache_trim_mode:1;
119
120 /* The file pages on the current node are dangerously low */
121 unsigned int file_is_tiny:1;
122
123 /* Allocation order */
124 s8 order;
125
126 /* Scan (total_size >> priority) pages at once */
127 s8 priority;
128
129 /* The highest zone to isolate pages for reclaim from */
130 s8 reclaim_idx;
131
132 /* This context's GFP mask */
133 gfp_t gfp_mask;
134
135 /* Incremented by the number of inactive pages that were scanned */
136 unsigned long nr_scanned;
137
138 /* Number of pages freed so far during a call to shrink_zones() */
139 unsigned long nr_reclaimed;
140
141 struct {
142 unsigned int dirty;
143 unsigned int unqueued_dirty;
144 unsigned int congested;
145 unsigned int writeback;
146 unsigned int immediate;
147 unsigned int file_taken;
148 unsigned int taken;
149 } nr;
150
151 /* for recording the reclaimed slab by now */
152 struct reclaim_state reclaim_state;
153};
154
155#ifdef ARCH_HAS_PREFETCHW
156#define prefetchw_prev_lru_page(_page, _base, _field) \
157 do { \
158 if ((_page)->lru.prev != _base) { \
159 struct page *prev; \
160 \
161 prev = lru_to_page(&(_page->lru)); \
162 prefetchw(&prev->_field); \
163 } \
164 } while (0)
165#else
166#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
167#endif
168
169/*
170 * From 0 .. 200. Higher means more swappy.
171 */
172int vm_swappiness = 60;
173
174static void set_task_reclaim_state(struct task_struct *task,
175 struct reclaim_state *rs)
176{
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs && task->reclaim_state);
179
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs && !task->reclaim_state);
182
183 task->reclaim_state = rs;
184}
185
186static LIST_HEAD(shrinker_list);
187static DECLARE_RWSEM(shrinker_rwsem);
188
189#ifdef CONFIG_MEMCG
190/*
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
197 *
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
200 */
201#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202
203static DEFINE_IDR(shrinker_idr);
204static int shrinker_nr_max;
205
206static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207{
208 int id, ret = -ENOMEM;
209
210 down_write(&shrinker_rwsem);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
213 if (id < 0)
214 goto unlock;
215
216 if (id >= shrinker_nr_max) {
217 if (memcg_expand_shrinker_maps(id)) {
218 idr_remove(&shrinker_idr, id);
219 goto unlock;
220 }
221
222 shrinker_nr_max = id + 1;
223 }
224 shrinker->id = id;
225 ret = 0;
226unlock:
227 up_write(&shrinker_rwsem);
228 return ret;
229}
230
231static void unregister_memcg_shrinker(struct shrinker *shrinker)
232{
233 int id = shrinker->id;
234
235 BUG_ON(id < 0);
236
237 down_write(&shrinker_rwsem);
238 idr_remove(&shrinker_idr, id);
239 up_write(&shrinker_rwsem);
240}
241
242static bool cgroup_reclaim(struct scan_control *sc)
243{
244 return sc->target_mem_cgroup;
245}
246
247/**
248 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
249 * @sc: scan_control in question
250 *
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
256 *
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
259 */
260static bool writeback_throttling_sane(struct scan_control *sc)
261{
262 if (!cgroup_reclaim(sc))
263 return true;
264#ifdef CONFIG_CGROUP_WRITEBACK
265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
266 return true;
267#endif
268 return false;
269}
270#else
271static int prealloc_memcg_shrinker(struct shrinker *shrinker)
272{
273 return 0;
274}
275
276static void unregister_memcg_shrinker(struct shrinker *shrinker)
277{
278}
279
280static bool cgroup_reclaim(struct scan_control *sc)
281{
282 return false;
283}
284
285static bool writeback_throttling_sane(struct scan_control *sc)
286{
287 return true;
288}
289#endif
290
291/*
292 * This misses isolated pages which are not accounted for to save counters.
293 * As the data only determines if reclaim or compaction continues, it is
294 * not expected that isolated pages will be a dominating factor.
295 */
296unsigned long zone_reclaimable_pages(struct zone *zone)
297{
298 unsigned long nr;
299
300 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
301 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
302 if (get_nr_swap_pages() > 0)
303 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
305
306 return nr;
307}
308
309/**
310 * lruvec_lru_size - Returns the number of pages on the given LRU list.
311 * @lruvec: lru vector
312 * @lru: lru to use
313 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
314 */
315unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
316{
317 unsigned long size = 0;
318 int zid;
319
320 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
321 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
322
323 if (!managed_zone(zone))
324 continue;
325
326 if (!mem_cgroup_disabled())
327 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
328 else
329 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
330 }
331 return size;
332}
333
334/*
335 * Add a shrinker callback to be called from the vm.
336 */
337int prealloc_shrinker(struct shrinker *shrinker)
338{
339 unsigned int size = sizeof(*shrinker->nr_deferred);
340
341 if (shrinker->flags & SHRINKER_NUMA_AWARE)
342 size *= nr_node_ids;
343
344 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
345 if (!shrinker->nr_deferred)
346 return -ENOMEM;
347
348 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
349 if (prealloc_memcg_shrinker(shrinker))
350 goto free_deferred;
351 }
352
353 return 0;
354
355free_deferred:
356 kfree(shrinker->nr_deferred);
357 shrinker->nr_deferred = NULL;
358 return -ENOMEM;
359}
360
361void free_prealloced_shrinker(struct shrinker *shrinker)
362{
363 if (!shrinker->nr_deferred)
364 return;
365
366 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
367 unregister_memcg_shrinker(shrinker);
368
369 kfree(shrinker->nr_deferred);
370 shrinker->nr_deferred = NULL;
371}
372
373void register_shrinker_prepared(struct shrinker *shrinker)
374{
375 down_write(&shrinker_rwsem);
376 list_add_tail(&shrinker->list, &shrinker_list);
377#ifdef CONFIG_MEMCG
378 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
379 idr_replace(&shrinker_idr, shrinker, shrinker->id);
380#endif
381 up_write(&shrinker_rwsem);
382}
383
384int register_shrinker(struct shrinker *shrinker)
385{
386 int err = prealloc_shrinker(shrinker);
387
388 if (err)
389 return err;
390 register_shrinker_prepared(shrinker);
391 return 0;
392}
393EXPORT_SYMBOL(register_shrinker);
394
395/*
396 * Remove one
397 */
398void unregister_shrinker(struct shrinker *shrinker)
399{
400 if (!shrinker->nr_deferred)
401 return;
402 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
403 unregister_memcg_shrinker(shrinker);
404 down_write(&shrinker_rwsem);
405 list_del(&shrinker->list);
406 up_write(&shrinker_rwsem);
407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
409}
410EXPORT_SYMBOL(unregister_shrinker);
411
412#define SHRINK_BATCH 128
413
414static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
415 struct shrinker *shrinker, int priority)
416{
417 unsigned long freed = 0;
418 unsigned long long delta;
419 long total_scan;
420 long freeable;
421 long nr;
422 long new_nr;
423 int nid = shrinkctl->nid;
424 long batch_size = shrinker->batch ? shrinker->batch
425 : SHRINK_BATCH;
426 long scanned = 0, next_deferred;
427
428 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
429 nid = 0;
430
431 freeable = shrinker->count_objects(shrinker, shrinkctl);
432 if (freeable == 0 || freeable == SHRINK_EMPTY)
433 return freeable;
434
435 /*
436 * copy the current shrinker scan count into a local variable
437 * and zero it so that other concurrent shrinker invocations
438 * don't also do this scanning work.
439 */
440 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
441
442 total_scan = nr;
443 if (shrinker->seeks) {
444 delta = freeable >> priority;
445 delta *= 4;
446 do_div(delta, shrinker->seeks);
447 } else {
448 /*
449 * These objects don't require any IO to create. Trim
450 * them aggressively under memory pressure to keep
451 * them from causing refetches in the IO caches.
452 */
453 delta = freeable / 2;
454 }
455
456 total_scan += delta;
457 if (total_scan < 0) {
458 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
459 shrinker->scan_objects, total_scan);
460 total_scan = freeable;
461 next_deferred = nr;
462 } else
463 next_deferred = total_scan;
464
465 /*
466 * We need to avoid excessive windup on filesystem shrinkers
467 * due to large numbers of GFP_NOFS allocations causing the
468 * shrinkers to return -1 all the time. This results in a large
469 * nr being built up so when a shrink that can do some work
470 * comes along it empties the entire cache due to nr >>>
471 * freeable. This is bad for sustaining a working set in
472 * memory.
473 *
474 * Hence only allow the shrinker to scan the entire cache when
475 * a large delta change is calculated directly.
476 */
477 if (delta < freeable / 4)
478 total_scan = min(total_scan, freeable / 2);
479
480 /*
481 * Avoid risking looping forever due to too large nr value:
482 * never try to free more than twice the estimate number of
483 * freeable entries.
484 */
485 if (total_scan > freeable * 2)
486 total_scan = freeable * 2;
487
488 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
489 freeable, delta, total_scan, priority);
490
491 /*
492 * Normally, we should not scan less than batch_size objects in one
493 * pass to avoid too frequent shrinker calls, but if the slab has less
494 * than batch_size objects in total and we are really tight on memory,
495 * we will try to reclaim all available objects, otherwise we can end
496 * up failing allocations although there are plenty of reclaimable
497 * objects spread over several slabs with usage less than the
498 * batch_size.
499 *
500 * We detect the "tight on memory" situations by looking at the total
501 * number of objects we want to scan (total_scan). If it is greater
502 * than the total number of objects on slab (freeable), we must be
503 * scanning at high prio and therefore should try to reclaim as much as
504 * possible.
505 */
506 while (total_scan >= batch_size ||
507 total_scan >= freeable) {
508 unsigned long ret;
509 unsigned long nr_to_scan = min(batch_size, total_scan);
510
511 shrinkctl->nr_to_scan = nr_to_scan;
512 shrinkctl->nr_scanned = nr_to_scan;
513 ret = shrinker->scan_objects(shrinker, shrinkctl);
514 if (ret == SHRINK_STOP)
515 break;
516 freed += ret;
517
518 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
519 total_scan -= shrinkctl->nr_scanned;
520 scanned += shrinkctl->nr_scanned;
521
522 cond_resched();
523 }
524
525 if (next_deferred >= scanned)
526 next_deferred -= scanned;
527 else
528 next_deferred = 0;
529 /*
530 * move the unused scan count back into the shrinker in a
531 * manner that handles concurrent updates. If we exhausted the
532 * scan, there is no need to do an update.
533 */
534 if (next_deferred > 0)
535 new_nr = atomic_long_add_return(next_deferred,
536 &shrinker->nr_deferred[nid]);
537 else
538 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
539
540 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
541 return freed;
542}
543
544#ifdef CONFIG_MEMCG
545static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
546 struct mem_cgroup *memcg, int priority)
547{
548 struct memcg_shrinker_map *map;
549 unsigned long ret, freed = 0;
550 int i;
551
552 if (!mem_cgroup_online(memcg))
553 return 0;
554
555 if (!down_read_trylock(&shrinker_rwsem))
556 return 0;
557
558 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
559 true);
560 if (unlikely(!map))
561 goto unlock;
562
563 for_each_set_bit(i, map->map, shrinker_nr_max) {
564 struct shrink_control sc = {
565 .gfp_mask = gfp_mask,
566 .nid = nid,
567 .memcg = memcg,
568 };
569 struct shrinker *shrinker;
570
571 shrinker = idr_find(&shrinker_idr, i);
572 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
573 if (!shrinker)
574 clear_bit(i, map->map);
575 continue;
576 }
577
578 /* Call non-slab shrinkers even though kmem is disabled */
579 if (!memcg_kmem_enabled() &&
580 !(shrinker->flags & SHRINKER_NONSLAB))
581 continue;
582
583 ret = do_shrink_slab(&sc, shrinker, priority);
584 if (ret == SHRINK_EMPTY) {
585 clear_bit(i, map->map);
586 /*
587 * After the shrinker reported that it had no objects to
588 * free, but before we cleared the corresponding bit in
589 * the memcg shrinker map, a new object might have been
590 * added. To make sure, we have the bit set in this
591 * case, we invoke the shrinker one more time and reset
592 * the bit if it reports that it is not empty anymore.
593 * The memory barrier here pairs with the barrier in
594 * memcg_set_shrinker_bit():
595 *
596 * list_lru_add() shrink_slab_memcg()
597 * list_add_tail() clear_bit()
598 * <MB> <MB>
599 * set_bit() do_shrink_slab()
600 */
601 smp_mb__after_atomic();
602 ret = do_shrink_slab(&sc, shrinker, priority);
603 if (ret == SHRINK_EMPTY)
604 ret = 0;
605 else
606 memcg_set_shrinker_bit(memcg, nid, i);
607 }
608 freed += ret;
609
610 if (rwsem_is_contended(&shrinker_rwsem)) {
611 freed = freed ? : 1;
612 break;
613 }
614 }
615unlock:
616 up_read(&shrinker_rwsem);
617 return freed;
618}
619#else /* CONFIG_MEMCG */
620static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
621 struct mem_cgroup *memcg, int priority)
622{
623 return 0;
624}
625#endif /* CONFIG_MEMCG */
626
627/**
628 * shrink_slab - shrink slab caches
629 * @gfp_mask: allocation context
630 * @nid: node whose slab caches to target
631 * @memcg: memory cgroup whose slab caches to target
632 * @priority: the reclaim priority
633 *
634 * Call the shrink functions to age shrinkable caches.
635 *
636 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
637 * unaware shrinkers will receive a node id of 0 instead.
638 *
639 * @memcg specifies the memory cgroup to target. Unaware shrinkers
640 * are called only if it is the root cgroup.
641 *
642 * @priority is sc->priority, we take the number of objects and >> by priority
643 * in order to get the scan target.
644 *
645 * Returns the number of reclaimed slab objects.
646 */
647static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
648 struct mem_cgroup *memcg,
649 int priority)
650{
651 unsigned long ret, freed = 0;
652 struct shrinker *shrinker;
653
654 /*
655 * The root memcg might be allocated even though memcg is disabled
656 * via "cgroup_disable=memory" boot parameter. This could make
657 * mem_cgroup_is_root() return false, then just run memcg slab
658 * shrink, but skip global shrink. This may result in premature
659 * oom.
660 */
661 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
662 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
663
664 if (!down_read_trylock(&shrinker_rwsem))
665 goto out;
666
667 list_for_each_entry(shrinker, &shrinker_list, list) {
668 struct shrink_control sc = {
669 .gfp_mask = gfp_mask,
670 .nid = nid,
671 .memcg = memcg,
672 };
673
674 ret = do_shrink_slab(&sc, shrinker, priority);
675 if (ret == SHRINK_EMPTY)
676 ret = 0;
677 freed += ret;
678 /*
679 * Bail out if someone want to register a new shrinker to
680 * prevent the registration from being stalled for long periods
681 * by parallel ongoing shrinking.
682 */
683 if (rwsem_is_contended(&shrinker_rwsem)) {
684 freed = freed ? : 1;
685 break;
686 }
687 }
688
689 up_read(&shrinker_rwsem);
690out:
691 cond_resched();
692 return freed;
693}
694
695void drop_slab_node(int nid)
696{
697 unsigned long freed;
698
699 do {
700 struct mem_cgroup *memcg = NULL;
701
702 freed = 0;
703 memcg = mem_cgroup_iter(NULL, NULL, NULL);
704 do {
705 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
706 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
707 } while (freed > 10);
708}
709
710void drop_slab(void)
711{
712 int nid;
713
714 for_each_online_node(nid)
715 drop_slab_node(nid);
716}
717
718static inline int is_page_cache_freeable(struct page *page)
719{
720 /*
721 * A freeable page cache page is referenced only by the caller
722 * that isolated the page, the page cache and optional buffer
723 * heads at page->private.
724 */
725 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
726 HPAGE_PMD_NR : 1;
727 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
728}
729
730static int may_write_to_inode(struct inode *inode)
731{
732 if (current->flags & PF_SWAPWRITE)
733 return 1;
734 if (!inode_write_congested(inode))
735 return 1;
736 if (inode_to_bdi(inode) == current->backing_dev_info)
737 return 1;
738 return 0;
739}
740
741/*
742 * We detected a synchronous write error writing a page out. Probably
743 * -ENOSPC. We need to propagate that into the address_space for a subsequent
744 * fsync(), msync() or close().
745 *
746 * The tricky part is that after writepage we cannot touch the mapping: nothing
747 * prevents it from being freed up. But we have a ref on the page and once
748 * that page is locked, the mapping is pinned.
749 *
750 * We're allowed to run sleeping lock_page() here because we know the caller has
751 * __GFP_FS.
752 */
753static void handle_write_error(struct address_space *mapping,
754 struct page *page, int error)
755{
756 lock_page(page);
757 if (page_mapping(page) == mapping)
758 mapping_set_error(mapping, error);
759 unlock_page(page);
760}
761
762/* possible outcome of pageout() */
763typedef enum {
764 /* failed to write page out, page is locked */
765 PAGE_KEEP,
766 /* move page to the active list, page is locked */
767 PAGE_ACTIVATE,
768 /* page has been sent to the disk successfully, page is unlocked */
769 PAGE_SUCCESS,
770 /* page is clean and locked */
771 PAGE_CLEAN,
772} pageout_t;
773
774/*
775 * pageout is called by shrink_page_list() for each dirty page.
776 * Calls ->writepage().
777 */
778static pageout_t pageout(struct page *page, struct address_space *mapping)
779{
780 /*
781 * If the page is dirty, only perform writeback if that write
782 * will be non-blocking. To prevent this allocation from being
783 * stalled by pagecache activity. But note that there may be
784 * stalls if we need to run get_block(). We could test
785 * PagePrivate for that.
786 *
787 * If this process is currently in __generic_file_write_iter() against
788 * this page's queue, we can perform writeback even if that
789 * will block.
790 *
791 * If the page is swapcache, write it back even if that would
792 * block, for some throttling. This happens by accident, because
793 * swap_backing_dev_info is bust: it doesn't reflect the
794 * congestion state of the swapdevs. Easy to fix, if needed.
795 */
796 if (!is_page_cache_freeable(page))
797 return PAGE_KEEP;
798 if (!mapping) {
799 /*
800 * Some data journaling orphaned pages can have
801 * page->mapping == NULL while being dirty with clean buffers.
802 */
803 if (page_has_private(page)) {
804 if (try_to_free_buffers(page)) {
805 ClearPageDirty(page);
806 pr_info("%s: orphaned page\n", __func__);
807 return PAGE_CLEAN;
808 }
809 }
810 return PAGE_KEEP;
811 }
812 if (mapping->a_ops->writepage == NULL)
813 return PAGE_ACTIVATE;
814 if (!may_write_to_inode(mapping->host))
815 return PAGE_KEEP;
816
817 if (clear_page_dirty_for_io(page)) {
818 int res;
819 struct writeback_control wbc = {
820 .sync_mode = WB_SYNC_NONE,
821 .nr_to_write = SWAP_CLUSTER_MAX,
822 .range_start = 0,
823 .range_end = LLONG_MAX,
824 .for_reclaim = 1,
825 };
826
827 SetPageReclaim(page);
828 res = mapping->a_ops->writepage(page, &wbc);
829 if (res < 0)
830 handle_write_error(mapping, page, res);
831 if (res == AOP_WRITEPAGE_ACTIVATE) {
832 ClearPageReclaim(page);
833 return PAGE_ACTIVATE;
834 }
835
836 if (!PageWriteback(page)) {
837 /* synchronous write or broken a_ops? */
838 ClearPageReclaim(page);
839 }
840 trace_mm_vmscan_writepage(page);
841 inc_node_page_state(page, NR_VMSCAN_WRITE);
842 return PAGE_SUCCESS;
843 }
844
845 return PAGE_CLEAN;
846}
847
848/*
849 * Same as remove_mapping, but if the page is removed from the mapping, it
850 * gets returned with a refcount of 0.
851 */
852static int __remove_mapping(struct address_space *mapping, struct page *page,
853 bool reclaimed, struct mem_cgroup *target_memcg)
854{
855 unsigned long flags;
856 int refcount;
857 void *shadow = NULL;
858
859 BUG_ON(!PageLocked(page));
860 BUG_ON(mapping != page_mapping(page));
861
862 xa_lock_irqsave(&mapping->i_pages, flags);
863 /*
864 * The non racy check for a busy page.
865 *
866 * Must be careful with the order of the tests. When someone has
867 * a ref to the page, it may be possible that they dirty it then
868 * drop the reference. So if PageDirty is tested before page_count
869 * here, then the following race may occur:
870 *
871 * get_user_pages(&page);
872 * [user mapping goes away]
873 * write_to(page);
874 * !PageDirty(page) [good]
875 * SetPageDirty(page);
876 * put_page(page);
877 * !page_count(page) [good, discard it]
878 *
879 * [oops, our write_to data is lost]
880 *
881 * Reversing the order of the tests ensures such a situation cannot
882 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
883 * load is not satisfied before that of page->_refcount.
884 *
885 * Note that if SetPageDirty is always performed via set_page_dirty,
886 * and thus under the i_pages lock, then this ordering is not required.
887 */
888 refcount = 1 + compound_nr(page);
889 if (!page_ref_freeze(page, refcount))
890 goto cannot_free;
891 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
892 if (unlikely(PageDirty(page))) {
893 page_ref_unfreeze(page, refcount);
894 goto cannot_free;
895 }
896
897 if (PageSwapCache(page)) {
898 swp_entry_t swap = { .val = page_private(page) };
899 mem_cgroup_swapout(page, swap);
900 if (reclaimed && !mapping_exiting(mapping))
901 shadow = workingset_eviction(page, target_memcg);
902 __delete_from_swap_cache(page, swap, shadow);
903 xa_unlock_irqrestore(&mapping->i_pages, flags);
904 put_swap_page(page, swap);
905 } else {
906 void (*freepage)(struct page *);
907
908 freepage = mapping->a_ops->freepage;
909 /*
910 * Remember a shadow entry for reclaimed file cache in
911 * order to detect refaults, thus thrashing, later on.
912 *
913 * But don't store shadows in an address space that is
914 * already exiting. This is not just an optimization,
915 * inode reclaim needs to empty out the radix tree or
916 * the nodes are lost. Don't plant shadows behind its
917 * back.
918 *
919 * We also don't store shadows for DAX mappings because the
920 * only page cache pages found in these are zero pages
921 * covering holes, and because we don't want to mix DAX
922 * exceptional entries and shadow exceptional entries in the
923 * same address_space.
924 */
925 if (reclaimed && page_is_file_lru(page) &&
926 !mapping_exiting(mapping) && !dax_mapping(mapping))
927 shadow = workingset_eviction(page, target_memcg);
928 __delete_from_page_cache(page, shadow);
929 xa_unlock_irqrestore(&mapping->i_pages, flags);
930
931 if (freepage != NULL)
932 freepage(page);
933 }
934
935 return 1;
936
937cannot_free:
938 xa_unlock_irqrestore(&mapping->i_pages, flags);
939 return 0;
940}
941
942/*
943 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
944 * someone else has a ref on the page, abort and return 0. If it was
945 * successfully detached, return 1. Assumes the caller has a single ref on
946 * this page.
947 */
948int remove_mapping(struct address_space *mapping, struct page *page)
949{
950 if (__remove_mapping(mapping, page, false, NULL)) {
951 /*
952 * Unfreezing the refcount with 1 rather than 2 effectively
953 * drops the pagecache ref for us without requiring another
954 * atomic operation.
955 */
956 page_ref_unfreeze(page, 1);
957 return 1;
958 }
959 return 0;
960}
961
962/**
963 * putback_lru_page - put previously isolated page onto appropriate LRU list
964 * @page: page to be put back to appropriate lru list
965 *
966 * Add previously isolated @page to appropriate LRU list.
967 * Page may still be unevictable for other reasons.
968 *
969 * lru_lock must not be held, interrupts must be enabled.
970 */
971void putback_lru_page(struct page *page)
972{
973 lru_cache_add(page);
974 put_page(page); /* drop ref from isolate */
975}
976
977enum page_references {
978 PAGEREF_RECLAIM,
979 PAGEREF_RECLAIM_CLEAN,
980 PAGEREF_KEEP,
981 PAGEREF_ACTIVATE,
982};
983
984static enum page_references page_check_references(struct page *page,
985 struct scan_control *sc)
986{
987 int referenced_ptes, referenced_page;
988 unsigned long vm_flags;
989
990 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
991 &vm_flags);
992 referenced_page = TestClearPageReferenced(page);
993
994 /*
995 * Mlock lost the isolation race with us. Let try_to_unmap()
996 * move the page to the unevictable list.
997 */
998 if (vm_flags & VM_LOCKED)
999 return PAGEREF_RECLAIM;
1000
1001 if (referenced_ptes) {
1002 /*
1003 * All mapped pages start out with page table
1004 * references from the instantiating fault, so we need
1005 * to look twice if a mapped file page is used more
1006 * than once.
1007 *
1008 * Mark it and spare it for another trip around the
1009 * inactive list. Another page table reference will
1010 * lead to its activation.
1011 *
1012 * Note: the mark is set for activated pages as well
1013 * so that recently deactivated but used pages are
1014 * quickly recovered.
1015 */
1016 SetPageReferenced(page);
1017
1018 if (referenced_page || referenced_ptes > 1)
1019 return PAGEREF_ACTIVATE;
1020
1021 /*
1022 * Activate file-backed executable pages after first usage.
1023 */
1024 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1025 return PAGEREF_ACTIVATE;
1026
1027 return PAGEREF_KEEP;
1028 }
1029
1030 /* Reclaim if clean, defer dirty pages to writeback */
1031 if (referenced_page && !PageSwapBacked(page))
1032 return PAGEREF_RECLAIM_CLEAN;
1033
1034 return PAGEREF_RECLAIM;
1035}
1036
1037/* Check if a page is dirty or under writeback */
1038static void page_check_dirty_writeback(struct page *page,
1039 bool *dirty, bool *writeback)
1040{
1041 struct address_space *mapping;
1042
1043 /*
1044 * Anonymous pages are not handled by flushers and must be written
1045 * from reclaim context. Do not stall reclaim based on them
1046 */
1047 if (!page_is_file_lru(page) ||
1048 (PageAnon(page) && !PageSwapBacked(page))) {
1049 *dirty = false;
1050 *writeback = false;
1051 return;
1052 }
1053
1054 /* By default assume that the page flags are accurate */
1055 *dirty = PageDirty(page);
1056 *writeback = PageWriteback(page);
1057
1058 /* Verify dirty/writeback state if the filesystem supports it */
1059 if (!page_has_private(page))
1060 return;
1061
1062 mapping = page_mapping(page);
1063 if (mapping && mapping->a_ops->is_dirty_writeback)
1064 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1065}
1066
1067/*
1068 * shrink_page_list() returns the number of reclaimed pages
1069 */
1070static unsigned int shrink_page_list(struct list_head *page_list,
1071 struct pglist_data *pgdat,
1072 struct scan_control *sc,
1073 enum ttu_flags ttu_flags,
1074 struct reclaim_stat *stat,
1075 bool ignore_references)
1076{
1077 LIST_HEAD(ret_pages);
1078 LIST_HEAD(free_pages);
1079 unsigned int nr_reclaimed = 0;
1080 unsigned int pgactivate = 0;
1081
1082 memset(stat, 0, sizeof(*stat));
1083 cond_resched();
1084
1085 while (!list_empty(page_list)) {
1086 struct address_space *mapping;
1087 struct page *page;
1088 enum page_references references = PAGEREF_RECLAIM;
1089 bool dirty, writeback, may_enter_fs;
1090 unsigned int nr_pages;
1091
1092 cond_resched();
1093
1094 page = lru_to_page(page_list);
1095 list_del(&page->lru);
1096
1097 if (!trylock_page(page))
1098 goto keep;
1099
1100 VM_BUG_ON_PAGE(PageActive(page), page);
1101
1102 nr_pages = compound_nr(page);
1103
1104 /* Account the number of base pages even though THP */
1105 sc->nr_scanned += nr_pages;
1106
1107 if (unlikely(!page_evictable(page)))
1108 goto activate_locked;
1109
1110 if (!sc->may_unmap && page_mapped(page))
1111 goto keep_locked;
1112
1113 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1114 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1115
1116 /*
1117 * The number of dirty pages determines if a node is marked
1118 * reclaim_congested which affects wait_iff_congested. kswapd
1119 * will stall and start writing pages if the tail of the LRU
1120 * is all dirty unqueued pages.
1121 */
1122 page_check_dirty_writeback(page, &dirty, &writeback);
1123 if (dirty || writeback)
1124 stat->nr_dirty++;
1125
1126 if (dirty && !writeback)
1127 stat->nr_unqueued_dirty++;
1128
1129 /*
1130 * Treat this page as congested if the underlying BDI is or if
1131 * pages are cycling through the LRU so quickly that the
1132 * pages marked for immediate reclaim are making it to the
1133 * end of the LRU a second time.
1134 */
1135 mapping = page_mapping(page);
1136 if (((dirty || writeback) && mapping &&
1137 inode_write_congested(mapping->host)) ||
1138 (writeback && PageReclaim(page)))
1139 stat->nr_congested++;
1140
1141 /*
1142 * If a page at the tail of the LRU is under writeback, there
1143 * are three cases to consider.
1144 *
1145 * 1) If reclaim is encountering an excessive number of pages
1146 * under writeback and this page is both under writeback and
1147 * PageReclaim then it indicates that pages are being queued
1148 * for IO but are being recycled through the LRU before the
1149 * IO can complete. Waiting on the page itself risks an
1150 * indefinite stall if it is impossible to writeback the
1151 * page due to IO error or disconnected storage so instead
1152 * note that the LRU is being scanned too quickly and the
1153 * caller can stall after page list has been processed.
1154 *
1155 * 2) Global or new memcg reclaim encounters a page that is
1156 * not marked for immediate reclaim, or the caller does not
1157 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1158 * not to fs). In this case mark the page for immediate
1159 * reclaim and continue scanning.
1160 *
1161 * Require may_enter_fs because we would wait on fs, which
1162 * may not have submitted IO yet. And the loop driver might
1163 * enter reclaim, and deadlock if it waits on a page for
1164 * which it is needed to do the write (loop masks off
1165 * __GFP_IO|__GFP_FS for this reason); but more thought
1166 * would probably show more reasons.
1167 *
1168 * 3) Legacy memcg encounters a page that is already marked
1169 * PageReclaim. memcg does not have any dirty pages
1170 * throttling so we could easily OOM just because too many
1171 * pages are in writeback and there is nothing else to
1172 * reclaim. Wait for the writeback to complete.
1173 *
1174 * In cases 1) and 2) we activate the pages to get them out of
1175 * the way while we continue scanning for clean pages on the
1176 * inactive list and refilling from the active list. The
1177 * observation here is that waiting for disk writes is more
1178 * expensive than potentially causing reloads down the line.
1179 * Since they're marked for immediate reclaim, they won't put
1180 * memory pressure on the cache working set any longer than it
1181 * takes to write them to disk.
1182 */
1183 if (PageWriteback(page)) {
1184 /* Case 1 above */
1185 if (current_is_kswapd() &&
1186 PageReclaim(page) &&
1187 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1188 stat->nr_immediate++;
1189 goto activate_locked;
1190
1191 /* Case 2 above */
1192 } else if (writeback_throttling_sane(sc) ||
1193 !PageReclaim(page) || !may_enter_fs) {
1194 /*
1195 * This is slightly racy - end_page_writeback()
1196 * might have just cleared PageReclaim, then
1197 * setting PageReclaim here end up interpreted
1198 * as PageReadahead - but that does not matter
1199 * enough to care. What we do want is for this
1200 * page to have PageReclaim set next time memcg
1201 * reclaim reaches the tests above, so it will
1202 * then wait_on_page_writeback() to avoid OOM;
1203 * and it's also appropriate in global reclaim.
1204 */
1205 SetPageReclaim(page);
1206 stat->nr_writeback++;
1207 goto activate_locked;
1208
1209 /* Case 3 above */
1210 } else {
1211 unlock_page(page);
1212 wait_on_page_writeback(page);
1213 /* then go back and try same page again */
1214 list_add_tail(&page->lru, page_list);
1215 continue;
1216 }
1217 }
1218
1219 if (!ignore_references)
1220 references = page_check_references(page, sc);
1221
1222 switch (references) {
1223 case PAGEREF_ACTIVATE:
1224 goto activate_locked;
1225 case PAGEREF_KEEP:
1226 stat->nr_ref_keep += nr_pages;
1227 goto keep_locked;
1228 case PAGEREF_RECLAIM:
1229 case PAGEREF_RECLAIM_CLEAN:
1230 ; /* try to reclaim the page below */
1231 }
1232
1233 /*
1234 * Anonymous process memory has backing store?
1235 * Try to allocate it some swap space here.
1236 * Lazyfree page could be freed directly
1237 */
1238 if (PageAnon(page) && PageSwapBacked(page)) {
1239 if (!PageSwapCache(page)) {
1240 if (!(sc->gfp_mask & __GFP_IO))
1241 goto keep_locked;
1242 if (PageTransHuge(page)) {
1243 /* cannot split THP, skip it */
1244 if (!can_split_huge_page(page, NULL))
1245 goto activate_locked;
1246 /*
1247 * Split pages without a PMD map right
1248 * away. Chances are some or all of the
1249 * tail pages can be freed without IO.
1250 */
1251 if (!compound_mapcount(page) &&
1252 split_huge_page_to_list(page,
1253 page_list))
1254 goto activate_locked;
1255 }
1256 if (!add_to_swap(page)) {
1257 if (!PageTransHuge(page))
1258 goto activate_locked_split;
1259 /* Fallback to swap normal pages */
1260 if (split_huge_page_to_list(page,
1261 page_list))
1262 goto activate_locked;
1263#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1264 count_vm_event(THP_SWPOUT_FALLBACK);
1265#endif
1266 if (!add_to_swap(page))
1267 goto activate_locked_split;
1268 }
1269
1270 may_enter_fs = true;
1271
1272 /* Adding to swap updated mapping */
1273 mapping = page_mapping(page);
1274 }
1275 } else if (unlikely(PageTransHuge(page))) {
1276 /* Split file THP */
1277 if (split_huge_page_to_list(page, page_list))
1278 goto keep_locked;
1279 }
1280
1281 /*
1282 * THP may get split above, need minus tail pages and update
1283 * nr_pages to avoid accounting tail pages twice.
1284 *
1285 * The tail pages that are added into swap cache successfully
1286 * reach here.
1287 */
1288 if ((nr_pages > 1) && !PageTransHuge(page)) {
1289 sc->nr_scanned -= (nr_pages - 1);
1290 nr_pages = 1;
1291 }
1292
1293 /*
1294 * The page is mapped into the page tables of one or more
1295 * processes. Try to unmap it here.
1296 */
1297 if (page_mapped(page)) {
1298 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1299 bool was_swapbacked = PageSwapBacked(page);
1300
1301 if (unlikely(PageTransHuge(page)))
1302 flags |= TTU_SPLIT_HUGE_PMD;
1303
1304 if (!try_to_unmap(page, flags)) {
1305 stat->nr_unmap_fail += nr_pages;
1306 if (!was_swapbacked && PageSwapBacked(page))
1307 stat->nr_lazyfree_fail += nr_pages;
1308 goto activate_locked;
1309 }
1310 }
1311
1312 if (PageDirty(page)) {
1313 /*
1314 * Only kswapd can writeback filesystem pages
1315 * to avoid risk of stack overflow. But avoid
1316 * injecting inefficient single-page IO into
1317 * flusher writeback as much as possible: only
1318 * write pages when we've encountered many
1319 * dirty pages, and when we've already scanned
1320 * the rest of the LRU for clean pages and see
1321 * the same dirty pages again (PageReclaim).
1322 */
1323 if (page_is_file_lru(page) &&
1324 (!current_is_kswapd() || !PageReclaim(page) ||
1325 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1326 /*
1327 * Immediately reclaim when written back.
1328 * Similar in principal to deactivate_page()
1329 * except we already have the page isolated
1330 * and know it's dirty
1331 */
1332 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1333 SetPageReclaim(page);
1334
1335 goto activate_locked;
1336 }
1337
1338 if (references == PAGEREF_RECLAIM_CLEAN)
1339 goto keep_locked;
1340 if (!may_enter_fs)
1341 goto keep_locked;
1342 if (!sc->may_writepage)
1343 goto keep_locked;
1344
1345 /*
1346 * Page is dirty. Flush the TLB if a writable entry
1347 * potentially exists to avoid CPU writes after IO
1348 * starts and then write it out here.
1349 */
1350 try_to_unmap_flush_dirty();
1351 switch (pageout(page, mapping)) {
1352 case PAGE_KEEP:
1353 goto keep_locked;
1354 case PAGE_ACTIVATE:
1355 goto activate_locked;
1356 case PAGE_SUCCESS:
1357 stat->nr_pageout += thp_nr_pages(page);
1358
1359 if (PageWriteback(page))
1360 goto keep;
1361 if (PageDirty(page))
1362 goto keep;
1363
1364 /*
1365 * A synchronous write - probably a ramdisk. Go
1366 * ahead and try to reclaim the page.
1367 */
1368 if (!trylock_page(page))
1369 goto keep;
1370 if (PageDirty(page) || PageWriteback(page))
1371 goto keep_locked;
1372 mapping = page_mapping(page);
1373 case PAGE_CLEAN:
1374 ; /* try to free the page below */
1375 }
1376 }
1377
1378 /*
1379 * If the page has buffers, try to free the buffer mappings
1380 * associated with this page. If we succeed we try to free
1381 * the page as well.
1382 *
1383 * We do this even if the page is PageDirty().
1384 * try_to_release_page() does not perform I/O, but it is
1385 * possible for a page to have PageDirty set, but it is actually
1386 * clean (all its buffers are clean). This happens if the
1387 * buffers were written out directly, with submit_bh(). ext3
1388 * will do this, as well as the blockdev mapping.
1389 * try_to_release_page() will discover that cleanness and will
1390 * drop the buffers and mark the page clean - it can be freed.
1391 *
1392 * Rarely, pages can have buffers and no ->mapping. These are
1393 * the pages which were not successfully invalidated in
1394 * truncate_complete_page(). We try to drop those buffers here
1395 * and if that worked, and the page is no longer mapped into
1396 * process address space (page_count == 1) it can be freed.
1397 * Otherwise, leave the page on the LRU so it is swappable.
1398 */
1399 if (page_has_private(page)) {
1400 if (!try_to_release_page(page, sc->gfp_mask))
1401 goto activate_locked;
1402 if (!mapping && page_count(page) == 1) {
1403 unlock_page(page);
1404 if (put_page_testzero(page))
1405 goto free_it;
1406 else {
1407 /*
1408 * rare race with speculative reference.
1409 * the speculative reference will free
1410 * this page shortly, so we may
1411 * increment nr_reclaimed here (and
1412 * leave it off the LRU).
1413 */
1414 nr_reclaimed++;
1415 continue;
1416 }
1417 }
1418 }
1419
1420 if (PageAnon(page) && !PageSwapBacked(page)) {
1421 /* follow __remove_mapping for reference */
1422 if (!page_ref_freeze(page, 1))
1423 goto keep_locked;
1424 if (PageDirty(page)) {
1425 page_ref_unfreeze(page, 1);
1426 goto keep_locked;
1427 }
1428
1429 count_vm_event(PGLAZYFREED);
1430 count_memcg_page_event(page, PGLAZYFREED);
1431 } else if (!mapping || !__remove_mapping(mapping, page, true,
1432 sc->target_mem_cgroup))
1433 goto keep_locked;
1434
1435 unlock_page(page);
1436free_it:
1437 /*
1438 * THP may get swapped out in a whole, need account
1439 * all base pages.
1440 */
1441 nr_reclaimed += nr_pages;
1442
1443 /*
1444 * Is there need to periodically free_page_list? It would
1445 * appear not as the counts should be low
1446 */
1447 if (unlikely(PageTransHuge(page)))
1448 destroy_compound_page(page);
1449 else
1450 list_add(&page->lru, &free_pages);
1451 continue;
1452
1453activate_locked_split:
1454 /*
1455 * The tail pages that are failed to add into swap cache
1456 * reach here. Fixup nr_scanned and nr_pages.
1457 */
1458 if (nr_pages > 1) {
1459 sc->nr_scanned -= (nr_pages - 1);
1460 nr_pages = 1;
1461 }
1462activate_locked:
1463 /* Not a candidate for swapping, so reclaim swap space. */
1464 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1465 PageMlocked(page)))
1466 try_to_free_swap(page);
1467 VM_BUG_ON_PAGE(PageActive(page), page);
1468 if (!PageMlocked(page)) {
1469 int type = page_is_file_lru(page);
1470 SetPageActive(page);
1471 stat->nr_activate[type] += nr_pages;
1472 count_memcg_page_event(page, PGACTIVATE);
1473 }
1474keep_locked:
1475 unlock_page(page);
1476keep:
1477 list_add(&page->lru, &ret_pages);
1478 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1479 }
1480
1481 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1482
1483 mem_cgroup_uncharge_list(&free_pages);
1484 try_to_unmap_flush();
1485 free_unref_page_list(&free_pages);
1486
1487 list_splice(&ret_pages, page_list);
1488 count_vm_events(PGACTIVATE, pgactivate);
1489
1490 return nr_reclaimed;
1491}
1492
1493unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1494 struct list_head *page_list)
1495{
1496 struct scan_control sc = {
1497 .gfp_mask = GFP_KERNEL,
1498 .priority = DEF_PRIORITY,
1499 .may_unmap = 1,
1500 };
1501 struct reclaim_stat stat;
1502 unsigned int nr_reclaimed;
1503 struct page *page, *next;
1504 LIST_HEAD(clean_pages);
1505
1506 list_for_each_entry_safe(page, next, page_list, lru) {
1507 if (page_is_file_lru(page) && !PageDirty(page) &&
1508 !__PageMovable(page) && !PageUnevictable(page)) {
1509 ClearPageActive(page);
1510 list_move(&page->lru, &clean_pages);
1511 }
1512 }
1513
1514 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1515 TTU_IGNORE_ACCESS, &stat, true);
1516 list_splice(&clean_pages, page_list);
1517 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -nr_reclaimed);
1518 /*
1519 * Since lazyfree pages are isolated from file LRU from the beginning,
1520 * they will rotate back to anonymous LRU in the end if it failed to
1521 * discard so isolated count will be mismatched.
1522 * Compensate the isolated count for both LRU lists.
1523 */
1524 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1525 stat.nr_lazyfree_fail);
1526 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1527 -stat.nr_lazyfree_fail);
1528 return nr_reclaimed;
1529}
1530
1531/*
1532 * Attempt to remove the specified page from its LRU. Only take this page
1533 * if it is of the appropriate PageActive status. Pages which are being
1534 * freed elsewhere are also ignored.
1535 *
1536 * page: page to consider
1537 * mode: one of the LRU isolation modes defined above
1538 *
1539 * returns 0 on success, -ve errno on failure.
1540 */
1541int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1542{
1543 int ret = -EINVAL;
1544
1545 /* Only take pages on the LRU. */
1546 if (!PageLRU(page))
1547 return ret;
1548
1549 /* Compaction should not handle unevictable pages but CMA can do so */
1550 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1551 return ret;
1552
1553 ret = -EBUSY;
1554
1555 /*
1556 * To minimise LRU disruption, the caller can indicate that it only
1557 * wants to isolate pages it will be able to operate on without
1558 * blocking - clean pages for the most part.
1559 *
1560 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1561 * that it is possible to migrate without blocking
1562 */
1563 if (mode & ISOLATE_ASYNC_MIGRATE) {
1564 /* All the caller can do on PageWriteback is block */
1565 if (PageWriteback(page))
1566 return ret;
1567
1568 if (PageDirty(page)) {
1569 struct address_space *mapping;
1570 bool migrate_dirty;
1571
1572 /*
1573 * Only pages without mappings or that have a
1574 * ->migratepage callback are possible to migrate
1575 * without blocking. However, we can be racing with
1576 * truncation so it's necessary to lock the page
1577 * to stabilise the mapping as truncation holds
1578 * the page lock until after the page is removed
1579 * from the page cache.
1580 */
1581 if (!trylock_page(page))
1582 return ret;
1583
1584 mapping = page_mapping(page);
1585 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1586 unlock_page(page);
1587 if (!migrate_dirty)
1588 return ret;
1589 }
1590 }
1591
1592 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1593 return ret;
1594
1595 if (likely(get_page_unless_zero(page))) {
1596 /*
1597 * Be careful not to clear PageLRU until after we're
1598 * sure the page is not being freed elsewhere -- the
1599 * page release code relies on it.
1600 */
1601 ClearPageLRU(page);
1602 ret = 0;
1603 }
1604
1605 return ret;
1606}
1607
1608
1609/*
1610 * Update LRU sizes after isolating pages. The LRU size updates must
1611 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1612 */
1613static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1614 enum lru_list lru, unsigned long *nr_zone_taken)
1615{
1616 int zid;
1617
1618 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1619 if (!nr_zone_taken[zid])
1620 continue;
1621
1622 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1623 }
1624
1625}
1626
1627/**
1628 * pgdat->lru_lock is heavily contended. Some of the functions that
1629 * shrink the lists perform better by taking out a batch of pages
1630 * and working on them outside the LRU lock.
1631 *
1632 * For pagecache intensive workloads, this function is the hottest
1633 * spot in the kernel (apart from copy_*_user functions).
1634 *
1635 * Appropriate locks must be held before calling this function.
1636 *
1637 * @nr_to_scan: The number of eligible pages to look through on the list.
1638 * @lruvec: The LRU vector to pull pages from.
1639 * @dst: The temp list to put pages on to.
1640 * @nr_scanned: The number of pages that were scanned.
1641 * @sc: The scan_control struct for this reclaim session
1642 * @lru: LRU list id for isolating
1643 *
1644 * returns how many pages were moved onto *@dst.
1645 */
1646static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1647 struct lruvec *lruvec, struct list_head *dst,
1648 unsigned long *nr_scanned, struct scan_control *sc,
1649 enum lru_list lru)
1650{
1651 struct list_head *src = &lruvec->lists[lru];
1652 unsigned long nr_taken = 0;
1653 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1654 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1655 unsigned long skipped = 0;
1656 unsigned long scan, total_scan, nr_pages;
1657 LIST_HEAD(pages_skipped);
1658 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1659
1660 total_scan = 0;
1661 scan = 0;
1662 while (scan < nr_to_scan && !list_empty(src)) {
1663 struct page *page;
1664
1665 page = lru_to_page(src);
1666 prefetchw_prev_lru_page(page, src, flags);
1667
1668 VM_BUG_ON_PAGE(!PageLRU(page), page);
1669
1670 nr_pages = compound_nr(page);
1671 total_scan += nr_pages;
1672
1673 if (page_zonenum(page) > sc->reclaim_idx) {
1674 list_move(&page->lru, &pages_skipped);
1675 nr_skipped[page_zonenum(page)] += nr_pages;
1676 continue;
1677 }
1678
1679 /*
1680 * Do not count skipped pages because that makes the function
1681 * return with no isolated pages if the LRU mostly contains
1682 * ineligible pages. This causes the VM to not reclaim any
1683 * pages, triggering a premature OOM.
1684 *
1685 * Account all tail pages of THP. This would not cause
1686 * premature OOM since __isolate_lru_page() returns -EBUSY
1687 * only when the page is being freed somewhere else.
1688 */
1689 scan += nr_pages;
1690 switch (__isolate_lru_page(page, mode)) {
1691 case 0:
1692 nr_taken += nr_pages;
1693 nr_zone_taken[page_zonenum(page)] += nr_pages;
1694 list_move(&page->lru, dst);
1695 break;
1696
1697 case -EBUSY:
1698 /* else it is being freed elsewhere */
1699 list_move(&page->lru, src);
1700 continue;
1701
1702 default:
1703 BUG();
1704 }
1705 }
1706
1707 /*
1708 * Splice any skipped pages to the start of the LRU list. Note that
1709 * this disrupts the LRU order when reclaiming for lower zones but
1710 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1711 * scanning would soon rescan the same pages to skip and put the
1712 * system at risk of premature OOM.
1713 */
1714 if (!list_empty(&pages_skipped)) {
1715 int zid;
1716
1717 list_splice(&pages_skipped, src);
1718 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1719 if (!nr_skipped[zid])
1720 continue;
1721
1722 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1723 skipped += nr_skipped[zid];
1724 }
1725 }
1726 *nr_scanned = total_scan;
1727 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1728 total_scan, skipped, nr_taken, mode, lru);
1729 update_lru_sizes(lruvec, lru, nr_zone_taken);
1730 return nr_taken;
1731}
1732
1733/**
1734 * isolate_lru_page - tries to isolate a page from its LRU list
1735 * @page: page to isolate from its LRU list
1736 *
1737 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1738 * vmstat statistic corresponding to whatever LRU list the page was on.
1739 *
1740 * Returns 0 if the page was removed from an LRU list.
1741 * Returns -EBUSY if the page was not on an LRU list.
1742 *
1743 * The returned page will have PageLRU() cleared. If it was found on
1744 * the active list, it will have PageActive set. If it was found on
1745 * the unevictable list, it will have the PageUnevictable bit set. That flag
1746 * may need to be cleared by the caller before letting the page go.
1747 *
1748 * The vmstat statistic corresponding to the list on which the page was
1749 * found will be decremented.
1750 *
1751 * Restrictions:
1752 *
1753 * (1) Must be called with an elevated refcount on the page. This is a
1754 * fundamentnal difference from isolate_lru_pages (which is called
1755 * without a stable reference).
1756 * (2) the lru_lock must not be held.
1757 * (3) interrupts must be enabled.
1758 */
1759int isolate_lru_page(struct page *page)
1760{
1761 int ret = -EBUSY;
1762
1763 VM_BUG_ON_PAGE(!page_count(page), page);
1764 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1765
1766 if (PageLRU(page)) {
1767 pg_data_t *pgdat = page_pgdat(page);
1768 struct lruvec *lruvec;
1769
1770 spin_lock_irq(&pgdat->lru_lock);
1771 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1772 if (PageLRU(page)) {
1773 int lru = page_lru(page);
1774 get_page(page);
1775 ClearPageLRU(page);
1776 del_page_from_lru_list(page, lruvec, lru);
1777 ret = 0;
1778 }
1779 spin_unlock_irq(&pgdat->lru_lock);
1780 }
1781 return ret;
1782}
1783
1784/*
1785 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1786 * then get rescheduled. When there are massive number of tasks doing page
1787 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1788 * the LRU list will go small and be scanned faster than necessary, leading to
1789 * unnecessary swapping, thrashing and OOM.
1790 */
1791static int too_many_isolated(struct pglist_data *pgdat, int file,
1792 struct scan_control *sc)
1793{
1794 unsigned long inactive, isolated;
1795
1796 if (current_is_kswapd())
1797 return 0;
1798
1799 if (!writeback_throttling_sane(sc))
1800 return 0;
1801
1802 if (file) {
1803 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1804 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1805 } else {
1806 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1807 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1808 }
1809
1810 /*
1811 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1812 * won't get blocked by normal direct-reclaimers, forming a circular
1813 * deadlock.
1814 */
1815 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1816 inactive >>= 3;
1817
1818 return isolated > inactive;
1819}
1820
1821/*
1822 * This moves pages from @list to corresponding LRU list.
1823 *
1824 * We move them the other way if the page is referenced by one or more
1825 * processes, from rmap.
1826 *
1827 * If the pages are mostly unmapped, the processing is fast and it is
1828 * appropriate to hold zone_lru_lock across the whole operation. But if
1829 * the pages are mapped, the processing is slow (page_referenced()) so we
1830 * should drop zone_lru_lock around each page. It's impossible to balance
1831 * this, so instead we remove the pages from the LRU while processing them.
1832 * It is safe to rely on PG_active against the non-LRU pages in here because
1833 * nobody will play with that bit on a non-LRU page.
1834 *
1835 * The downside is that we have to touch page->_refcount against each page.
1836 * But we had to alter page->flags anyway.
1837 *
1838 * Returns the number of pages moved to the given lruvec.
1839 */
1840
1841static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1842 struct list_head *list)
1843{
1844 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1845 int nr_pages, nr_moved = 0;
1846 LIST_HEAD(pages_to_free);
1847 struct page *page;
1848 enum lru_list lru;
1849
1850 while (!list_empty(list)) {
1851 page = lru_to_page(list);
1852 VM_BUG_ON_PAGE(PageLRU(page), page);
1853 if (unlikely(!page_evictable(page))) {
1854 list_del(&page->lru);
1855 spin_unlock_irq(&pgdat->lru_lock);
1856 putback_lru_page(page);
1857 spin_lock_irq(&pgdat->lru_lock);
1858 continue;
1859 }
1860 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1861
1862 SetPageLRU(page);
1863 lru = page_lru(page);
1864
1865 nr_pages = thp_nr_pages(page);
1866 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1867 list_move(&page->lru, &lruvec->lists[lru]);
1868
1869 if (put_page_testzero(page)) {
1870 __ClearPageLRU(page);
1871 __ClearPageActive(page);
1872 del_page_from_lru_list(page, lruvec, lru);
1873
1874 if (unlikely(PageCompound(page))) {
1875 spin_unlock_irq(&pgdat->lru_lock);
1876 destroy_compound_page(page);
1877 spin_lock_irq(&pgdat->lru_lock);
1878 } else
1879 list_add(&page->lru, &pages_to_free);
1880 } else {
1881 nr_moved += nr_pages;
1882 if (PageActive(page))
1883 workingset_age_nonresident(lruvec, nr_pages);
1884 }
1885 }
1886
1887 /*
1888 * To save our caller's stack, now use input list for pages to free.
1889 */
1890 list_splice(&pages_to_free, list);
1891
1892 return nr_moved;
1893}
1894
1895/*
1896 * If a kernel thread (such as nfsd for loop-back mounts) services
1897 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1898 * In that case we should only throttle if the backing device it is
1899 * writing to is congested. In other cases it is safe to throttle.
1900 */
1901static int current_may_throttle(void)
1902{
1903 return !(current->flags & PF_LOCAL_THROTTLE) ||
1904 current->backing_dev_info == NULL ||
1905 bdi_write_congested(current->backing_dev_info);
1906}
1907
1908/*
1909 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1910 * of reclaimed pages
1911 */
1912static noinline_for_stack unsigned long
1913shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1914 struct scan_control *sc, enum lru_list lru)
1915{
1916 LIST_HEAD(page_list);
1917 unsigned long nr_scanned;
1918 unsigned int nr_reclaimed = 0;
1919 unsigned long nr_taken;
1920 struct reclaim_stat stat;
1921 bool file = is_file_lru(lru);
1922 enum vm_event_item item;
1923 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1924 bool stalled = false;
1925
1926 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1927 if (stalled)
1928 return 0;
1929
1930 /* wait a bit for the reclaimer. */
1931 msleep(100);
1932 stalled = true;
1933
1934 /* We are about to die and free our memory. Return now. */
1935 if (fatal_signal_pending(current))
1936 return SWAP_CLUSTER_MAX;
1937 }
1938
1939 lru_add_drain();
1940
1941 spin_lock_irq(&pgdat->lru_lock);
1942
1943 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1944 &nr_scanned, sc, lru);
1945
1946 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1947 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1948 if (!cgroup_reclaim(sc))
1949 __count_vm_events(item, nr_scanned);
1950 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1951 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1952
1953 spin_unlock_irq(&pgdat->lru_lock);
1954
1955 if (nr_taken == 0)
1956 return 0;
1957
1958 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1959 &stat, false);
1960
1961 spin_lock_irq(&pgdat->lru_lock);
1962
1963 move_pages_to_lru(lruvec, &page_list);
1964
1965 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1966 lru_note_cost(lruvec, file, stat.nr_pageout);
1967 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1968 if (!cgroup_reclaim(sc))
1969 __count_vm_events(item, nr_reclaimed);
1970 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1971 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1972
1973 spin_unlock_irq(&pgdat->lru_lock);
1974
1975 mem_cgroup_uncharge_list(&page_list);
1976 free_unref_page_list(&page_list);
1977
1978 /*
1979 * If dirty pages are scanned that are not queued for IO, it
1980 * implies that flushers are not doing their job. This can
1981 * happen when memory pressure pushes dirty pages to the end of
1982 * the LRU before the dirty limits are breached and the dirty
1983 * data has expired. It can also happen when the proportion of
1984 * dirty pages grows not through writes but through memory
1985 * pressure reclaiming all the clean cache. And in some cases,
1986 * the flushers simply cannot keep up with the allocation
1987 * rate. Nudge the flusher threads in case they are asleep.
1988 */
1989 if (stat.nr_unqueued_dirty == nr_taken)
1990 wakeup_flusher_threads(WB_REASON_VMSCAN);
1991
1992 sc->nr.dirty += stat.nr_dirty;
1993 sc->nr.congested += stat.nr_congested;
1994 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1995 sc->nr.writeback += stat.nr_writeback;
1996 sc->nr.immediate += stat.nr_immediate;
1997 sc->nr.taken += nr_taken;
1998 if (file)
1999 sc->nr.file_taken += nr_taken;
2000
2001 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2002 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2003 return nr_reclaimed;
2004}
2005
2006static void shrink_active_list(unsigned long nr_to_scan,
2007 struct lruvec *lruvec,
2008 struct scan_control *sc,
2009 enum lru_list lru)
2010{
2011 unsigned long nr_taken;
2012 unsigned long nr_scanned;
2013 unsigned long vm_flags;
2014 LIST_HEAD(l_hold); /* The pages which were snipped off */
2015 LIST_HEAD(l_active);
2016 LIST_HEAD(l_inactive);
2017 struct page *page;
2018 unsigned nr_deactivate, nr_activate;
2019 unsigned nr_rotated = 0;
2020 int file = is_file_lru(lru);
2021 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2022
2023 lru_add_drain();
2024
2025 spin_lock_irq(&pgdat->lru_lock);
2026
2027 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2028 &nr_scanned, sc, lru);
2029
2030 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2031
2032 if (!cgroup_reclaim(sc))
2033 __count_vm_events(PGREFILL, nr_scanned);
2034 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2035
2036 spin_unlock_irq(&pgdat->lru_lock);
2037
2038 while (!list_empty(&l_hold)) {
2039 cond_resched();
2040 page = lru_to_page(&l_hold);
2041 list_del(&page->lru);
2042
2043 if (unlikely(!page_evictable(page))) {
2044 putback_lru_page(page);
2045 continue;
2046 }
2047
2048 if (unlikely(buffer_heads_over_limit)) {
2049 if (page_has_private(page) && trylock_page(page)) {
2050 if (page_has_private(page))
2051 try_to_release_page(page, 0);
2052 unlock_page(page);
2053 }
2054 }
2055
2056 if (page_referenced(page, 0, sc->target_mem_cgroup,
2057 &vm_flags)) {
2058 /*
2059 * Identify referenced, file-backed active pages and
2060 * give them one more trip around the active list. So
2061 * that executable code get better chances to stay in
2062 * memory under moderate memory pressure. Anon pages
2063 * are not likely to be evicted by use-once streaming
2064 * IO, plus JVM can create lots of anon VM_EXEC pages,
2065 * so we ignore them here.
2066 */
2067 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2068 nr_rotated += thp_nr_pages(page);
2069 list_add(&page->lru, &l_active);
2070 continue;
2071 }
2072 }
2073
2074 ClearPageActive(page); /* we are de-activating */
2075 SetPageWorkingset(page);
2076 list_add(&page->lru, &l_inactive);
2077 }
2078
2079 /*
2080 * Move pages back to the lru list.
2081 */
2082 spin_lock_irq(&pgdat->lru_lock);
2083
2084 nr_activate = move_pages_to_lru(lruvec, &l_active);
2085 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2086 /* Keep all free pages in l_active list */
2087 list_splice(&l_inactive, &l_active);
2088
2089 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2090 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2091
2092 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2093 spin_unlock_irq(&pgdat->lru_lock);
2094
2095 mem_cgroup_uncharge_list(&l_active);
2096 free_unref_page_list(&l_active);
2097 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2098 nr_deactivate, nr_rotated, sc->priority, file);
2099}
2100
2101unsigned long reclaim_pages(struct list_head *page_list)
2102{
2103 int nid = NUMA_NO_NODE;
2104 unsigned int nr_reclaimed = 0;
2105 LIST_HEAD(node_page_list);
2106 struct reclaim_stat dummy_stat;
2107 struct page *page;
2108 struct scan_control sc = {
2109 .gfp_mask = GFP_KERNEL,
2110 .priority = DEF_PRIORITY,
2111 .may_writepage = 1,
2112 .may_unmap = 1,
2113 .may_swap = 1,
2114 };
2115
2116 while (!list_empty(page_list)) {
2117 page = lru_to_page(page_list);
2118 if (nid == NUMA_NO_NODE) {
2119 nid = page_to_nid(page);
2120 INIT_LIST_HEAD(&node_page_list);
2121 }
2122
2123 if (nid == page_to_nid(page)) {
2124 ClearPageActive(page);
2125 list_move(&page->lru, &node_page_list);
2126 continue;
2127 }
2128
2129 nr_reclaimed += shrink_page_list(&node_page_list,
2130 NODE_DATA(nid),
2131 &sc, 0,
2132 &dummy_stat, false);
2133 while (!list_empty(&node_page_list)) {
2134 page = lru_to_page(&node_page_list);
2135 list_del(&page->lru);
2136 putback_lru_page(page);
2137 }
2138
2139 nid = NUMA_NO_NODE;
2140 }
2141
2142 if (!list_empty(&node_page_list)) {
2143 nr_reclaimed += shrink_page_list(&node_page_list,
2144 NODE_DATA(nid),
2145 &sc, 0,
2146 &dummy_stat, false);
2147 while (!list_empty(&node_page_list)) {
2148 page = lru_to_page(&node_page_list);
2149 list_del(&page->lru);
2150 putback_lru_page(page);
2151 }
2152 }
2153
2154 return nr_reclaimed;
2155}
2156
2157static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2158 struct lruvec *lruvec, struct scan_control *sc)
2159{
2160 if (is_active_lru(lru)) {
2161 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2162 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2163 else
2164 sc->skipped_deactivate = 1;
2165 return 0;
2166 }
2167
2168 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2169}
2170
2171/*
2172 * The inactive anon list should be small enough that the VM never has
2173 * to do too much work.
2174 *
2175 * The inactive file list should be small enough to leave most memory
2176 * to the established workingset on the scan-resistant active list,
2177 * but large enough to avoid thrashing the aggregate readahead window.
2178 *
2179 * Both inactive lists should also be large enough that each inactive
2180 * page has a chance to be referenced again before it is reclaimed.
2181 *
2182 * If that fails and refaulting is observed, the inactive list grows.
2183 *
2184 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2185 * on this LRU, maintained by the pageout code. An inactive_ratio
2186 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2187 *
2188 * total target max
2189 * memory ratio inactive
2190 * -------------------------------------
2191 * 10MB 1 5MB
2192 * 100MB 1 50MB
2193 * 1GB 3 250MB
2194 * 10GB 10 0.9GB
2195 * 100GB 31 3GB
2196 * 1TB 101 10GB
2197 * 10TB 320 32GB
2198 */
2199static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2200{
2201 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2202 unsigned long inactive, active;
2203 unsigned long inactive_ratio;
2204 unsigned long gb;
2205
2206 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2207 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2208
2209 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2210 if (gb)
2211 inactive_ratio = int_sqrt(10 * gb);
2212 else
2213 inactive_ratio = 1;
2214
2215 return inactive * inactive_ratio < active;
2216}
2217
2218enum scan_balance {
2219 SCAN_EQUAL,
2220 SCAN_FRACT,
2221 SCAN_ANON,
2222 SCAN_FILE,
2223};
2224
2225/*
2226 * Determine how aggressively the anon and file LRU lists should be
2227 * scanned. The relative value of each set of LRU lists is determined
2228 * by looking at the fraction of the pages scanned we did rotate back
2229 * onto the active list instead of evict.
2230 *
2231 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2232 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2233 */
2234static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2235 unsigned long *nr)
2236{
2237 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2238 unsigned long anon_cost, file_cost, total_cost;
2239 int swappiness = mem_cgroup_swappiness(memcg);
2240 u64 fraction[2];
2241 u64 denominator = 0; /* gcc */
2242 enum scan_balance scan_balance;
2243 unsigned long ap, fp;
2244 enum lru_list lru;
2245
2246 /* If we have no swap space, do not bother scanning anon pages. */
2247 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2248 scan_balance = SCAN_FILE;
2249 goto out;
2250 }
2251
2252 /*
2253 * Global reclaim will swap to prevent OOM even with no
2254 * swappiness, but memcg users want to use this knob to
2255 * disable swapping for individual groups completely when
2256 * using the memory controller's swap limit feature would be
2257 * too expensive.
2258 */
2259 if (cgroup_reclaim(sc) && !swappiness) {
2260 scan_balance = SCAN_FILE;
2261 goto out;
2262 }
2263
2264 /*
2265 * Do not apply any pressure balancing cleverness when the
2266 * system is close to OOM, scan both anon and file equally
2267 * (unless the swappiness setting disagrees with swapping).
2268 */
2269 if (!sc->priority && swappiness) {
2270 scan_balance = SCAN_EQUAL;
2271 goto out;
2272 }
2273
2274 /*
2275 * If the system is almost out of file pages, force-scan anon.
2276 */
2277 if (sc->file_is_tiny) {
2278 scan_balance = SCAN_ANON;
2279 goto out;
2280 }
2281
2282 /*
2283 * If there is enough inactive page cache, we do not reclaim
2284 * anything from the anonymous working right now.
2285 */
2286 if (sc->cache_trim_mode) {
2287 scan_balance = SCAN_FILE;
2288 goto out;
2289 }
2290
2291 scan_balance = SCAN_FRACT;
2292 /*
2293 * Calculate the pressure balance between anon and file pages.
2294 *
2295 * The amount of pressure we put on each LRU is inversely
2296 * proportional to the cost of reclaiming each list, as
2297 * determined by the share of pages that are refaulting, times
2298 * the relative IO cost of bringing back a swapped out
2299 * anonymous page vs reloading a filesystem page (swappiness).
2300 *
2301 * Although we limit that influence to ensure no list gets
2302 * left behind completely: at least a third of the pressure is
2303 * applied, before swappiness.
2304 *
2305 * With swappiness at 100, anon and file have equal IO cost.
2306 */
2307 total_cost = sc->anon_cost + sc->file_cost;
2308 anon_cost = total_cost + sc->anon_cost;
2309 file_cost = total_cost + sc->file_cost;
2310 total_cost = anon_cost + file_cost;
2311
2312 ap = swappiness * (total_cost + 1);
2313 ap /= anon_cost + 1;
2314
2315 fp = (200 - swappiness) * (total_cost + 1);
2316 fp /= file_cost + 1;
2317
2318 fraction[0] = ap;
2319 fraction[1] = fp;
2320 denominator = ap + fp;
2321out:
2322 for_each_evictable_lru(lru) {
2323 int file = is_file_lru(lru);
2324 unsigned long lruvec_size;
2325 unsigned long scan;
2326 unsigned long protection;
2327
2328 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2329 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2330 memcg,
2331 sc->memcg_low_reclaim);
2332
2333 if (protection) {
2334 /*
2335 * Scale a cgroup's reclaim pressure by proportioning
2336 * its current usage to its memory.low or memory.min
2337 * setting.
2338 *
2339 * This is important, as otherwise scanning aggression
2340 * becomes extremely binary -- from nothing as we
2341 * approach the memory protection threshold, to totally
2342 * nominal as we exceed it. This results in requiring
2343 * setting extremely liberal protection thresholds. It
2344 * also means we simply get no protection at all if we
2345 * set it too low, which is not ideal.
2346 *
2347 * If there is any protection in place, we reduce scan
2348 * pressure by how much of the total memory used is
2349 * within protection thresholds.
2350 *
2351 * There is one special case: in the first reclaim pass,
2352 * we skip over all groups that are within their low
2353 * protection. If that fails to reclaim enough pages to
2354 * satisfy the reclaim goal, we come back and override
2355 * the best-effort low protection. However, we still
2356 * ideally want to honor how well-behaved groups are in
2357 * that case instead of simply punishing them all
2358 * equally. As such, we reclaim them based on how much
2359 * memory they are using, reducing the scan pressure
2360 * again by how much of the total memory used is under
2361 * hard protection.
2362 */
2363 unsigned long cgroup_size = mem_cgroup_size(memcg);
2364
2365 /* Avoid TOCTOU with earlier protection check */
2366 cgroup_size = max(cgroup_size, protection);
2367
2368 scan = lruvec_size - lruvec_size * protection /
2369 cgroup_size;
2370
2371 /*
2372 * Minimally target SWAP_CLUSTER_MAX pages to keep
2373 * reclaim moving forwards, avoiding decrementing
2374 * sc->priority further than desirable.
2375 */
2376 scan = max(scan, SWAP_CLUSTER_MAX);
2377 } else {
2378 scan = lruvec_size;
2379 }
2380
2381 scan >>= sc->priority;
2382
2383 /*
2384 * If the cgroup's already been deleted, make sure to
2385 * scrape out the remaining cache.
2386 */
2387 if (!scan && !mem_cgroup_online(memcg))
2388 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2389
2390 switch (scan_balance) {
2391 case SCAN_EQUAL:
2392 /* Scan lists relative to size */
2393 break;
2394 case SCAN_FRACT:
2395 /*
2396 * Scan types proportional to swappiness and
2397 * their relative recent reclaim efficiency.
2398 * Make sure we don't miss the last page on
2399 * the offlined memory cgroups because of a
2400 * round-off error.
2401 */
2402 scan = mem_cgroup_online(memcg) ?
2403 div64_u64(scan * fraction[file], denominator) :
2404 DIV64_U64_ROUND_UP(scan * fraction[file],
2405 denominator);
2406 break;
2407 case SCAN_FILE:
2408 case SCAN_ANON:
2409 /* Scan one type exclusively */
2410 if ((scan_balance == SCAN_FILE) != file)
2411 scan = 0;
2412 break;
2413 default:
2414 /* Look ma, no brain */
2415 BUG();
2416 }
2417
2418 nr[lru] = scan;
2419 }
2420}
2421
2422static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2423{
2424 unsigned long nr[NR_LRU_LISTS];
2425 unsigned long targets[NR_LRU_LISTS];
2426 unsigned long nr_to_scan;
2427 enum lru_list lru;
2428 unsigned long nr_reclaimed = 0;
2429 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2430 struct blk_plug plug;
2431 bool scan_adjusted;
2432
2433 get_scan_count(lruvec, sc, nr);
2434
2435 /* Record the original scan target for proportional adjustments later */
2436 memcpy(targets, nr, sizeof(nr));
2437
2438 /*
2439 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2440 * event that can occur when there is little memory pressure e.g.
2441 * multiple streaming readers/writers. Hence, we do not abort scanning
2442 * when the requested number of pages are reclaimed when scanning at
2443 * DEF_PRIORITY on the assumption that the fact we are direct
2444 * reclaiming implies that kswapd is not keeping up and it is best to
2445 * do a batch of work at once. For memcg reclaim one check is made to
2446 * abort proportional reclaim if either the file or anon lru has already
2447 * dropped to zero at the first pass.
2448 */
2449 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2450 sc->priority == DEF_PRIORITY);
2451
2452 blk_start_plug(&plug);
2453 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2454 nr[LRU_INACTIVE_FILE]) {
2455 unsigned long nr_anon, nr_file, percentage;
2456 unsigned long nr_scanned;
2457
2458 for_each_evictable_lru(lru) {
2459 if (nr[lru]) {
2460 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2461 nr[lru] -= nr_to_scan;
2462
2463 nr_reclaimed += shrink_list(lru, nr_to_scan,
2464 lruvec, sc);
2465 }
2466 }
2467
2468 cond_resched();
2469
2470 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2471 continue;
2472
2473 /*
2474 * For kswapd and memcg, reclaim at least the number of pages
2475 * requested. Ensure that the anon and file LRUs are scanned
2476 * proportionally what was requested by get_scan_count(). We
2477 * stop reclaiming one LRU and reduce the amount scanning
2478 * proportional to the original scan target.
2479 */
2480 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2481 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2482
2483 /*
2484 * It's just vindictive to attack the larger once the smaller
2485 * has gone to zero. And given the way we stop scanning the
2486 * smaller below, this makes sure that we only make one nudge
2487 * towards proportionality once we've got nr_to_reclaim.
2488 */
2489 if (!nr_file || !nr_anon)
2490 break;
2491
2492 if (nr_file > nr_anon) {
2493 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2494 targets[LRU_ACTIVE_ANON] + 1;
2495 lru = LRU_BASE;
2496 percentage = nr_anon * 100 / scan_target;
2497 } else {
2498 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2499 targets[LRU_ACTIVE_FILE] + 1;
2500 lru = LRU_FILE;
2501 percentage = nr_file * 100 / scan_target;
2502 }
2503
2504 /* Stop scanning the smaller of the LRU */
2505 nr[lru] = 0;
2506 nr[lru + LRU_ACTIVE] = 0;
2507
2508 /*
2509 * Recalculate the other LRU scan count based on its original
2510 * scan target and the percentage scanning already complete
2511 */
2512 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2513 nr_scanned = targets[lru] - nr[lru];
2514 nr[lru] = targets[lru] * (100 - percentage) / 100;
2515 nr[lru] -= min(nr[lru], nr_scanned);
2516
2517 lru += LRU_ACTIVE;
2518 nr_scanned = targets[lru] - nr[lru];
2519 nr[lru] = targets[lru] * (100 - percentage) / 100;
2520 nr[lru] -= min(nr[lru], nr_scanned);
2521
2522 scan_adjusted = true;
2523 }
2524 blk_finish_plug(&plug);
2525 sc->nr_reclaimed += nr_reclaimed;
2526
2527 /*
2528 * Even if we did not try to evict anon pages at all, we want to
2529 * rebalance the anon lru active/inactive ratio.
2530 */
2531 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2532 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2533 sc, LRU_ACTIVE_ANON);
2534}
2535
2536/* Use reclaim/compaction for costly allocs or under memory pressure */
2537static bool in_reclaim_compaction(struct scan_control *sc)
2538{
2539 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2540 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2541 sc->priority < DEF_PRIORITY - 2))
2542 return true;
2543
2544 return false;
2545}
2546
2547/*
2548 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2549 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2550 * true if more pages should be reclaimed such that when the page allocator
2551 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2552 * It will give up earlier than that if there is difficulty reclaiming pages.
2553 */
2554static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2555 unsigned long nr_reclaimed,
2556 struct scan_control *sc)
2557{
2558 unsigned long pages_for_compaction;
2559 unsigned long inactive_lru_pages;
2560 int z;
2561
2562 /* If not in reclaim/compaction mode, stop */
2563 if (!in_reclaim_compaction(sc))
2564 return false;
2565
2566 /*
2567 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2568 * number of pages that were scanned. This will return to the caller
2569 * with the risk reclaim/compaction and the resulting allocation attempt
2570 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2571 * allocations through requiring that the full LRU list has been scanned
2572 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2573 * scan, but that approximation was wrong, and there were corner cases
2574 * where always a non-zero amount of pages were scanned.
2575 */
2576 if (!nr_reclaimed)
2577 return false;
2578
2579 /* If compaction would go ahead or the allocation would succeed, stop */
2580 for (z = 0; z <= sc->reclaim_idx; z++) {
2581 struct zone *zone = &pgdat->node_zones[z];
2582 if (!managed_zone(zone))
2583 continue;
2584
2585 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2586 case COMPACT_SUCCESS:
2587 case COMPACT_CONTINUE:
2588 return false;
2589 default:
2590 /* check next zone */
2591 ;
2592 }
2593 }
2594
2595 /*
2596 * If we have not reclaimed enough pages for compaction and the
2597 * inactive lists are large enough, continue reclaiming
2598 */
2599 pages_for_compaction = compact_gap(sc->order);
2600 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2601 if (get_nr_swap_pages() > 0)
2602 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2603
2604 return inactive_lru_pages > pages_for_compaction;
2605}
2606
2607static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2608{
2609 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2610 struct mem_cgroup *memcg;
2611
2612 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2613 do {
2614 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2615 unsigned long reclaimed;
2616 unsigned long scanned;
2617
2618 /*
2619 * This loop can become CPU-bound when target memcgs
2620 * aren't eligible for reclaim - either because they
2621 * don't have any reclaimable pages, or because their
2622 * memory is explicitly protected. Avoid soft lockups.
2623 */
2624 cond_resched();
2625
2626 mem_cgroup_calculate_protection(target_memcg, memcg);
2627
2628 if (mem_cgroup_below_min(memcg)) {
2629 /*
2630 * Hard protection.
2631 * If there is no reclaimable memory, OOM.
2632 */
2633 continue;
2634 } else if (mem_cgroup_below_low(memcg)) {
2635 /*
2636 * Soft protection.
2637 * Respect the protection only as long as
2638 * there is an unprotected supply
2639 * of reclaimable memory from other cgroups.
2640 */
2641 if (!sc->memcg_low_reclaim) {
2642 sc->memcg_low_skipped = 1;
2643 continue;
2644 }
2645 memcg_memory_event(memcg, MEMCG_LOW);
2646 }
2647
2648 reclaimed = sc->nr_reclaimed;
2649 scanned = sc->nr_scanned;
2650
2651 shrink_lruvec(lruvec, sc);
2652
2653 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2654 sc->priority);
2655
2656 /* Record the group's reclaim efficiency */
2657 vmpressure(sc->gfp_mask, memcg, false,
2658 sc->nr_scanned - scanned,
2659 sc->nr_reclaimed - reclaimed);
2660
2661 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2662}
2663
2664static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2665{
2666 struct reclaim_state *reclaim_state = current->reclaim_state;
2667 unsigned long nr_reclaimed, nr_scanned;
2668 struct lruvec *target_lruvec;
2669 bool reclaimable = false;
2670 unsigned long file;
2671
2672 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2673
2674again:
2675 memset(&sc->nr, 0, sizeof(sc->nr));
2676
2677 nr_reclaimed = sc->nr_reclaimed;
2678 nr_scanned = sc->nr_scanned;
2679
2680 /*
2681 * Determine the scan balance between anon and file LRUs.
2682 */
2683 spin_lock_irq(&pgdat->lru_lock);
2684 sc->anon_cost = target_lruvec->anon_cost;
2685 sc->file_cost = target_lruvec->file_cost;
2686 spin_unlock_irq(&pgdat->lru_lock);
2687
2688 /*
2689 * Target desirable inactive:active list ratios for the anon
2690 * and file LRU lists.
2691 */
2692 if (!sc->force_deactivate) {
2693 unsigned long refaults;
2694
2695 refaults = lruvec_page_state(target_lruvec,
2696 WORKINGSET_ACTIVATE_ANON);
2697 if (refaults != target_lruvec->refaults[0] ||
2698 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2699 sc->may_deactivate |= DEACTIVATE_ANON;
2700 else
2701 sc->may_deactivate &= ~DEACTIVATE_ANON;
2702
2703 /*
2704 * When refaults are being observed, it means a new
2705 * workingset is being established. Deactivate to get
2706 * rid of any stale active pages quickly.
2707 */
2708 refaults = lruvec_page_state(target_lruvec,
2709 WORKINGSET_ACTIVATE_FILE);
2710 if (refaults != target_lruvec->refaults[1] ||
2711 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2712 sc->may_deactivate |= DEACTIVATE_FILE;
2713 else
2714 sc->may_deactivate &= ~DEACTIVATE_FILE;
2715 } else
2716 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2717
2718 /*
2719 * If we have plenty of inactive file pages that aren't
2720 * thrashing, try to reclaim those first before touching
2721 * anonymous pages.
2722 */
2723 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2724 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2725 sc->cache_trim_mode = 1;
2726 else
2727 sc->cache_trim_mode = 0;
2728
2729 /*
2730 * Prevent the reclaimer from falling into the cache trap: as
2731 * cache pages start out inactive, every cache fault will tip
2732 * the scan balance towards the file LRU. And as the file LRU
2733 * shrinks, so does the window for rotation from references.
2734 * This means we have a runaway feedback loop where a tiny
2735 * thrashing file LRU becomes infinitely more attractive than
2736 * anon pages. Try to detect this based on file LRU size.
2737 */
2738 if (!cgroup_reclaim(sc)) {
2739 unsigned long total_high_wmark = 0;
2740 unsigned long free, anon;
2741 int z;
2742
2743 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2744 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2745 node_page_state(pgdat, NR_INACTIVE_FILE);
2746
2747 for (z = 0; z < MAX_NR_ZONES; z++) {
2748 struct zone *zone = &pgdat->node_zones[z];
2749 if (!managed_zone(zone))
2750 continue;
2751
2752 total_high_wmark += high_wmark_pages(zone);
2753 }
2754
2755 /*
2756 * Consider anon: if that's low too, this isn't a
2757 * runaway file reclaim problem, but rather just
2758 * extreme pressure. Reclaim as per usual then.
2759 */
2760 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2761
2762 sc->file_is_tiny =
2763 file + free <= total_high_wmark &&
2764 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2765 anon >> sc->priority;
2766 }
2767
2768 shrink_node_memcgs(pgdat, sc);
2769
2770 if (reclaim_state) {
2771 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2772 reclaim_state->reclaimed_slab = 0;
2773 }
2774
2775 /* Record the subtree's reclaim efficiency */
2776 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2777 sc->nr_scanned - nr_scanned,
2778 sc->nr_reclaimed - nr_reclaimed);
2779
2780 if (sc->nr_reclaimed - nr_reclaimed)
2781 reclaimable = true;
2782
2783 if (current_is_kswapd()) {
2784 /*
2785 * If reclaim is isolating dirty pages under writeback,
2786 * it implies that the long-lived page allocation rate
2787 * is exceeding the page laundering rate. Either the
2788 * global limits are not being effective at throttling
2789 * processes due to the page distribution throughout
2790 * zones or there is heavy usage of a slow backing
2791 * device. The only option is to throttle from reclaim
2792 * context which is not ideal as there is no guarantee
2793 * the dirtying process is throttled in the same way
2794 * balance_dirty_pages() manages.
2795 *
2796 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2797 * count the number of pages under pages flagged for
2798 * immediate reclaim and stall if any are encountered
2799 * in the nr_immediate check below.
2800 */
2801 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2802 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2803
2804 /* Allow kswapd to start writing pages during reclaim.*/
2805 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2806 set_bit(PGDAT_DIRTY, &pgdat->flags);
2807
2808 /*
2809 * If kswapd scans pages marked for immediate
2810 * reclaim and under writeback (nr_immediate), it
2811 * implies that pages are cycling through the LRU
2812 * faster than they are written so also forcibly stall.
2813 */
2814 if (sc->nr.immediate)
2815 congestion_wait(BLK_RW_ASYNC, HZ/10);
2816 }
2817
2818 /*
2819 * Tag a node/memcg as congested if all the dirty pages
2820 * scanned were backed by a congested BDI and
2821 * wait_iff_congested will stall.
2822 *
2823 * Legacy memcg will stall in page writeback so avoid forcibly
2824 * stalling in wait_iff_congested().
2825 */
2826 if ((current_is_kswapd() ||
2827 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2828 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2829 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2830
2831 /*
2832 * Stall direct reclaim for IO completions if underlying BDIs
2833 * and node is congested. Allow kswapd to continue until it
2834 * starts encountering unqueued dirty pages or cycling through
2835 * the LRU too quickly.
2836 */
2837 if (!current_is_kswapd() && current_may_throttle() &&
2838 !sc->hibernation_mode &&
2839 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2840 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2841
2842 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2843 sc))
2844 goto again;
2845
2846 /*
2847 * Kswapd gives up on balancing particular nodes after too
2848 * many failures to reclaim anything from them and goes to
2849 * sleep. On reclaim progress, reset the failure counter. A
2850 * successful direct reclaim run will revive a dormant kswapd.
2851 */
2852 if (reclaimable)
2853 pgdat->kswapd_failures = 0;
2854}
2855
2856/*
2857 * Returns true if compaction should go ahead for a costly-order request, or
2858 * the allocation would already succeed without compaction. Return false if we
2859 * should reclaim first.
2860 */
2861static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2862{
2863 unsigned long watermark;
2864 enum compact_result suitable;
2865
2866 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2867 if (suitable == COMPACT_SUCCESS)
2868 /* Allocation should succeed already. Don't reclaim. */
2869 return true;
2870 if (suitable == COMPACT_SKIPPED)
2871 /* Compaction cannot yet proceed. Do reclaim. */
2872 return false;
2873
2874 /*
2875 * Compaction is already possible, but it takes time to run and there
2876 * are potentially other callers using the pages just freed. So proceed
2877 * with reclaim to make a buffer of free pages available to give
2878 * compaction a reasonable chance of completing and allocating the page.
2879 * Note that we won't actually reclaim the whole buffer in one attempt
2880 * as the target watermark in should_continue_reclaim() is lower. But if
2881 * we are already above the high+gap watermark, don't reclaim at all.
2882 */
2883 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2884
2885 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2886}
2887
2888/*
2889 * This is the direct reclaim path, for page-allocating processes. We only
2890 * try to reclaim pages from zones which will satisfy the caller's allocation
2891 * request.
2892 *
2893 * If a zone is deemed to be full of pinned pages then just give it a light
2894 * scan then give up on it.
2895 */
2896static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2897{
2898 struct zoneref *z;
2899 struct zone *zone;
2900 unsigned long nr_soft_reclaimed;
2901 unsigned long nr_soft_scanned;
2902 gfp_t orig_mask;
2903 pg_data_t *last_pgdat = NULL;
2904
2905 /*
2906 * If the number of buffer_heads in the machine exceeds the maximum
2907 * allowed level, force direct reclaim to scan the highmem zone as
2908 * highmem pages could be pinning lowmem pages storing buffer_heads
2909 */
2910 orig_mask = sc->gfp_mask;
2911 if (buffer_heads_over_limit) {
2912 sc->gfp_mask |= __GFP_HIGHMEM;
2913 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2914 }
2915
2916 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2917 sc->reclaim_idx, sc->nodemask) {
2918 /*
2919 * Take care memory controller reclaiming has small influence
2920 * to global LRU.
2921 */
2922 if (!cgroup_reclaim(sc)) {
2923 if (!cpuset_zone_allowed(zone,
2924 GFP_KERNEL | __GFP_HARDWALL))
2925 continue;
2926
2927 /*
2928 * If we already have plenty of memory free for
2929 * compaction in this zone, don't free any more.
2930 * Even though compaction is invoked for any
2931 * non-zero order, only frequent costly order
2932 * reclamation is disruptive enough to become a
2933 * noticeable problem, like transparent huge
2934 * page allocations.
2935 */
2936 if (IS_ENABLED(CONFIG_COMPACTION) &&
2937 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2938 compaction_ready(zone, sc)) {
2939 sc->compaction_ready = true;
2940 continue;
2941 }
2942
2943 /*
2944 * Shrink each node in the zonelist once. If the
2945 * zonelist is ordered by zone (not the default) then a
2946 * node may be shrunk multiple times but in that case
2947 * the user prefers lower zones being preserved.
2948 */
2949 if (zone->zone_pgdat == last_pgdat)
2950 continue;
2951
2952 /*
2953 * This steals pages from memory cgroups over softlimit
2954 * and returns the number of reclaimed pages and
2955 * scanned pages. This works for global memory pressure
2956 * and balancing, not for a memcg's limit.
2957 */
2958 nr_soft_scanned = 0;
2959 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2960 sc->order, sc->gfp_mask,
2961 &nr_soft_scanned);
2962 sc->nr_reclaimed += nr_soft_reclaimed;
2963 sc->nr_scanned += nr_soft_scanned;
2964 /* need some check for avoid more shrink_zone() */
2965 }
2966
2967 /* See comment about same check for global reclaim above */
2968 if (zone->zone_pgdat == last_pgdat)
2969 continue;
2970 last_pgdat = zone->zone_pgdat;
2971 shrink_node(zone->zone_pgdat, sc);
2972 }
2973
2974 /*
2975 * Restore to original mask to avoid the impact on the caller if we
2976 * promoted it to __GFP_HIGHMEM.
2977 */
2978 sc->gfp_mask = orig_mask;
2979}
2980
2981static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2982{
2983 struct lruvec *target_lruvec;
2984 unsigned long refaults;
2985
2986 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2987 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2988 target_lruvec->refaults[0] = refaults;
2989 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
2990 target_lruvec->refaults[1] = refaults;
2991}
2992
2993/*
2994 * This is the main entry point to direct page reclaim.
2995 *
2996 * If a full scan of the inactive list fails to free enough memory then we
2997 * are "out of memory" and something needs to be killed.
2998 *
2999 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3000 * high - the zone may be full of dirty or under-writeback pages, which this
3001 * caller can't do much about. We kick the writeback threads and take explicit
3002 * naps in the hope that some of these pages can be written. But if the
3003 * allocating task holds filesystem locks which prevent writeout this might not
3004 * work, and the allocation attempt will fail.
3005 *
3006 * returns: 0, if no pages reclaimed
3007 * else, the number of pages reclaimed
3008 */
3009static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3010 struct scan_control *sc)
3011{
3012 int initial_priority = sc->priority;
3013 pg_data_t *last_pgdat;
3014 struct zoneref *z;
3015 struct zone *zone;
3016retry:
3017 delayacct_freepages_start();
3018
3019 if (!cgroup_reclaim(sc))
3020 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3021
3022 do {
3023 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3024 sc->priority);
3025 sc->nr_scanned = 0;
3026 shrink_zones(zonelist, sc);
3027
3028 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3029 break;
3030
3031 if (sc->compaction_ready)
3032 break;
3033
3034 /*
3035 * If we're getting trouble reclaiming, start doing
3036 * writepage even in laptop mode.
3037 */
3038 if (sc->priority < DEF_PRIORITY - 2)
3039 sc->may_writepage = 1;
3040 } while (--sc->priority >= 0);
3041
3042 last_pgdat = NULL;
3043 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3044 sc->nodemask) {
3045 if (zone->zone_pgdat == last_pgdat)
3046 continue;
3047 last_pgdat = zone->zone_pgdat;
3048
3049 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3050
3051 if (cgroup_reclaim(sc)) {
3052 struct lruvec *lruvec;
3053
3054 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3055 zone->zone_pgdat);
3056 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3057 }
3058 }
3059
3060 delayacct_freepages_end();
3061
3062 if (sc->nr_reclaimed)
3063 return sc->nr_reclaimed;
3064
3065 /* Aborted reclaim to try compaction? don't OOM, then */
3066 if (sc->compaction_ready)
3067 return 1;
3068
3069 /*
3070 * We make inactive:active ratio decisions based on the node's
3071 * composition of memory, but a restrictive reclaim_idx or a
3072 * memory.low cgroup setting can exempt large amounts of
3073 * memory from reclaim. Neither of which are very common, so
3074 * instead of doing costly eligibility calculations of the
3075 * entire cgroup subtree up front, we assume the estimates are
3076 * good, and retry with forcible deactivation if that fails.
3077 */
3078 if (sc->skipped_deactivate) {
3079 sc->priority = initial_priority;
3080 sc->force_deactivate = 1;
3081 sc->skipped_deactivate = 0;
3082 goto retry;
3083 }
3084
3085 /* Untapped cgroup reserves? Don't OOM, retry. */
3086 if (sc->memcg_low_skipped) {
3087 sc->priority = initial_priority;
3088 sc->force_deactivate = 0;
3089 sc->memcg_low_reclaim = 1;
3090 sc->memcg_low_skipped = 0;
3091 goto retry;
3092 }
3093
3094 return 0;
3095}
3096
3097static bool allow_direct_reclaim(pg_data_t *pgdat)
3098{
3099 struct zone *zone;
3100 unsigned long pfmemalloc_reserve = 0;
3101 unsigned long free_pages = 0;
3102 int i;
3103 bool wmark_ok;
3104
3105 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3106 return true;
3107
3108 for (i = 0; i <= ZONE_NORMAL; i++) {
3109 zone = &pgdat->node_zones[i];
3110 if (!managed_zone(zone))
3111 continue;
3112
3113 if (!zone_reclaimable_pages(zone))
3114 continue;
3115
3116 pfmemalloc_reserve += min_wmark_pages(zone);
3117 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3118 }
3119
3120 /* If there are no reserves (unexpected config) then do not throttle */
3121 if (!pfmemalloc_reserve)
3122 return true;
3123
3124 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3125
3126 /* kswapd must be awake if processes are being throttled */
3127 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3128 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3129 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3130
3131 wake_up_interruptible(&pgdat->kswapd_wait);
3132 }
3133
3134 return wmark_ok;
3135}
3136
3137/*
3138 * Throttle direct reclaimers if backing storage is backed by the network
3139 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3140 * depleted. kswapd will continue to make progress and wake the processes
3141 * when the low watermark is reached.
3142 *
3143 * Returns true if a fatal signal was delivered during throttling. If this
3144 * happens, the page allocator should not consider triggering the OOM killer.
3145 */
3146static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3147 nodemask_t *nodemask)
3148{
3149 struct zoneref *z;
3150 struct zone *zone;
3151 pg_data_t *pgdat = NULL;
3152
3153 /*
3154 * Kernel threads should not be throttled as they may be indirectly
3155 * responsible for cleaning pages necessary for reclaim to make forward
3156 * progress. kjournald for example may enter direct reclaim while
3157 * committing a transaction where throttling it could forcing other
3158 * processes to block on log_wait_commit().
3159 */
3160 if (current->flags & PF_KTHREAD)
3161 goto out;
3162
3163 /*
3164 * If a fatal signal is pending, this process should not throttle.
3165 * It should return quickly so it can exit and free its memory
3166 */
3167 if (fatal_signal_pending(current))
3168 goto out;
3169
3170 /*
3171 * Check if the pfmemalloc reserves are ok by finding the first node
3172 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3173 * GFP_KERNEL will be required for allocating network buffers when
3174 * swapping over the network so ZONE_HIGHMEM is unusable.
3175 *
3176 * Throttling is based on the first usable node and throttled processes
3177 * wait on a queue until kswapd makes progress and wakes them. There
3178 * is an affinity then between processes waking up and where reclaim
3179 * progress has been made assuming the process wakes on the same node.
3180 * More importantly, processes running on remote nodes will not compete
3181 * for remote pfmemalloc reserves and processes on different nodes
3182 * should make reasonable progress.
3183 */
3184 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3185 gfp_zone(gfp_mask), nodemask) {
3186 if (zone_idx(zone) > ZONE_NORMAL)
3187 continue;
3188
3189 /* Throttle based on the first usable node */
3190 pgdat = zone->zone_pgdat;
3191 if (allow_direct_reclaim(pgdat))
3192 goto out;
3193 break;
3194 }
3195
3196 /* If no zone was usable by the allocation flags then do not throttle */
3197 if (!pgdat)
3198 goto out;
3199
3200 /* Account for the throttling */
3201 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3202
3203 /*
3204 * If the caller cannot enter the filesystem, it's possible that it
3205 * is due to the caller holding an FS lock or performing a journal
3206 * transaction in the case of a filesystem like ext[3|4]. In this case,
3207 * it is not safe to block on pfmemalloc_wait as kswapd could be
3208 * blocked waiting on the same lock. Instead, throttle for up to a
3209 * second before continuing.
3210 */
3211 if (!(gfp_mask & __GFP_FS)) {
3212 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3213 allow_direct_reclaim(pgdat), HZ);
3214
3215 goto check_pending;
3216 }
3217
3218 /* Throttle until kswapd wakes the process */
3219 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3220 allow_direct_reclaim(pgdat));
3221
3222check_pending:
3223 if (fatal_signal_pending(current))
3224 return true;
3225
3226out:
3227 return false;
3228}
3229
3230unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3231 gfp_t gfp_mask, nodemask_t *nodemask)
3232{
3233 unsigned long nr_reclaimed;
3234 struct scan_control sc = {
3235 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3236 .gfp_mask = current_gfp_context(gfp_mask),
3237 .reclaim_idx = gfp_zone(gfp_mask),
3238 .order = order,
3239 .nodemask = nodemask,
3240 .priority = DEF_PRIORITY,
3241 .may_writepage = !laptop_mode,
3242 .may_unmap = 1,
3243 .may_swap = 1,
3244 };
3245
3246 /*
3247 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3248 * Confirm they are large enough for max values.
3249 */
3250 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3251 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3252 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3253
3254 /*
3255 * Do not enter reclaim if fatal signal was delivered while throttled.
3256 * 1 is returned so that the page allocator does not OOM kill at this
3257 * point.
3258 */
3259 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3260 return 1;
3261
3262 set_task_reclaim_state(current, &sc.reclaim_state);
3263 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3264
3265 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3266
3267 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3268 set_task_reclaim_state(current, NULL);
3269
3270 return nr_reclaimed;
3271}
3272
3273#ifdef CONFIG_MEMCG
3274
3275/* Only used by soft limit reclaim. Do not reuse for anything else. */
3276unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3277 gfp_t gfp_mask, bool noswap,
3278 pg_data_t *pgdat,
3279 unsigned long *nr_scanned)
3280{
3281 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3282 struct scan_control sc = {
3283 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3284 .target_mem_cgroup = memcg,
3285 .may_writepage = !laptop_mode,
3286 .may_unmap = 1,
3287 .reclaim_idx = MAX_NR_ZONES - 1,
3288 .may_swap = !noswap,
3289 };
3290
3291 WARN_ON_ONCE(!current->reclaim_state);
3292
3293 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3294 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3295
3296 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3297 sc.gfp_mask);
3298
3299 /*
3300 * NOTE: Although we can get the priority field, using it
3301 * here is not a good idea, since it limits the pages we can scan.
3302 * if we don't reclaim here, the shrink_node from balance_pgdat
3303 * will pick up pages from other mem cgroup's as well. We hack
3304 * the priority and make it zero.
3305 */
3306 shrink_lruvec(lruvec, &sc);
3307
3308 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3309
3310 *nr_scanned = sc.nr_scanned;
3311
3312 return sc.nr_reclaimed;
3313}
3314
3315unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3316 unsigned long nr_pages,
3317 gfp_t gfp_mask,
3318 bool may_swap)
3319{
3320 unsigned long nr_reclaimed;
3321 unsigned int noreclaim_flag;
3322 struct scan_control sc = {
3323 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3324 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3325 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3326 .reclaim_idx = MAX_NR_ZONES - 1,
3327 .target_mem_cgroup = memcg,
3328 .priority = DEF_PRIORITY,
3329 .may_writepage = !laptop_mode,
3330 .may_unmap = 1,
3331 .may_swap = may_swap,
3332 };
3333 /*
3334 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3335 * equal pressure on all the nodes. This is based on the assumption that
3336 * the reclaim does not bail out early.
3337 */
3338 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3339
3340 set_task_reclaim_state(current, &sc.reclaim_state);
3341 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3342 noreclaim_flag = memalloc_noreclaim_save();
3343
3344 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3345
3346 memalloc_noreclaim_restore(noreclaim_flag);
3347 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3348 set_task_reclaim_state(current, NULL);
3349
3350 return nr_reclaimed;
3351}
3352#endif
3353
3354static void age_active_anon(struct pglist_data *pgdat,
3355 struct scan_control *sc)
3356{
3357 struct mem_cgroup *memcg;
3358 struct lruvec *lruvec;
3359
3360 if (!total_swap_pages)
3361 return;
3362
3363 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3364 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3365 return;
3366
3367 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3368 do {
3369 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3370 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3371 sc, LRU_ACTIVE_ANON);
3372 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3373 } while (memcg);
3374}
3375
3376static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3377{
3378 int i;
3379 struct zone *zone;
3380
3381 /*
3382 * Check for watermark boosts top-down as the higher zones
3383 * are more likely to be boosted. Both watermarks and boosts
3384 * should not be checked at the same time as reclaim would
3385 * start prematurely when there is no boosting and a lower
3386 * zone is balanced.
3387 */
3388 for (i = highest_zoneidx; i >= 0; i--) {
3389 zone = pgdat->node_zones + i;
3390 if (!managed_zone(zone))
3391 continue;
3392
3393 if (zone->watermark_boost)
3394 return true;
3395 }
3396
3397 return false;
3398}
3399
3400/*
3401 * Returns true if there is an eligible zone balanced for the request order
3402 * and highest_zoneidx
3403 */
3404static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3405{
3406 int i;
3407 unsigned long mark = -1;
3408 struct zone *zone;
3409
3410 /*
3411 * Check watermarks bottom-up as lower zones are more likely to
3412 * meet watermarks.
3413 */
3414 for (i = 0; i <= highest_zoneidx; i++) {
3415 zone = pgdat->node_zones + i;
3416
3417 if (!managed_zone(zone))
3418 continue;
3419
3420 mark = high_wmark_pages(zone);
3421 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3422 return true;
3423 }
3424
3425 /*
3426 * If a node has no populated zone within highest_zoneidx, it does not
3427 * need balancing by definition. This can happen if a zone-restricted
3428 * allocation tries to wake a remote kswapd.
3429 */
3430 if (mark == -1)
3431 return true;
3432
3433 return false;
3434}
3435
3436/* Clear pgdat state for congested, dirty or under writeback. */
3437static void clear_pgdat_congested(pg_data_t *pgdat)
3438{
3439 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3440
3441 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3442 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3443 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3444}
3445
3446/*
3447 * Prepare kswapd for sleeping. This verifies that there are no processes
3448 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3449 *
3450 * Returns true if kswapd is ready to sleep
3451 */
3452static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3453 int highest_zoneidx)
3454{
3455 /*
3456 * The throttled processes are normally woken up in balance_pgdat() as
3457 * soon as allow_direct_reclaim() is true. But there is a potential
3458 * race between when kswapd checks the watermarks and a process gets
3459 * throttled. There is also a potential race if processes get
3460 * throttled, kswapd wakes, a large process exits thereby balancing the
3461 * zones, which causes kswapd to exit balance_pgdat() before reaching
3462 * the wake up checks. If kswapd is going to sleep, no process should
3463 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3464 * the wake up is premature, processes will wake kswapd and get
3465 * throttled again. The difference from wake ups in balance_pgdat() is
3466 * that here we are under prepare_to_wait().
3467 */
3468 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3469 wake_up_all(&pgdat->pfmemalloc_wait);
3470
3471 /* Hopeless node, leave it to direct reclaim */
3472 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3473 return true;
3474
3475 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3476 clear_pgdat_congested(pgdat);
3477 return true;
3478 }
3479
3480 return false;
3481}
3482
3483/*
3484 * kswapd shrinks a node of pages that are at or below the highest usable
3485 * zone that is currently unbalanced.
3486 *
3487 * Returns true if kswapd scanned at least the requested number of pages to
3488 * reclaim or if the lack of progress was due to pages under writeback.
3489 * This is used to determine if the scanning priority needs to be raised.
3490 */
3491static bool kswapd_shrink_node(pg_data_t *pgdat,
3492 struct scan_control *sc)
3493{
3494 struct zone *zone;
3495 int z;
3496
3497 /* Reclaim a number of pages proportional to the number of zones */
3498 sc->nr_to_reclaim = 0;
3499 for (z = 0; z <= sc->reclaim_idx; z++) {
3500 zone = pgdat->node_zones + z;
3501 if (!managed_zone(zone))
3502 continue;
3503
3504 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3505 }
3506
3507 /*
3508 * Historically care was taken to put equal pressure on all zones but
3509 * now pressure is applied based on node LRU order.
3510 */
3511 shrink_node(pgdat, sc);
3512
3513 /*
3514 * Fragmentation may mean that the system cannot be rebalanced for
3515 * high-order allocations. If twice the allocation size has been
3516 * reclaimed then recheck watermarks only at order-0 to prevent
3517 * excessive reclaim. Assume that a process requested a high-order
3518 * can direct reclaim/compact.
3519 */
3520 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3521 sc->order = 0;
3522
3523 return sc->nr_scanned >= sc->nr_to_reclaim;
3524}
3525
3526/*
3527 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3528 * that are eligible for use by the caller until at least one zone is
3529 * balanced.
3530 *
3531 * Returns the order kswapd finished reclaiming at.
3532 *
3533 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3534 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3535 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3536 * or lower is eligible for reclaim until at least one usable zone is
3537 * balanced.
3538 */
3539static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3540{
3541 int i;
3542 unsigned long nr_soft_reclaimed;
3543 unsigned long nr_soft_scanned;
3544 unsigned long pflags;
3545 unsigned long nr_boost_reclaim;
3546 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3547 bool boosted;
3548 struct zone *zone;
3549 struct scan_control sc = {
3550 .gfp_mask = GFP_KERNEL,
3551 .order = order,
3552 .may_unmap = 1,
3553 };
3554
3555 set_task_reclaim_state(current, &sc.reclaim_state);
3556 psi_memstall_enter(&pflags);
3557 __fs_reclaim_acquire();
3558
3559 count_vm_event(PAGEOUTRUN);
3560
3561 /*
3562 * Account for the reclaim boost. Note that the zone boost is left in
3563 * place so that parallel allocations that are near the watermark will
3564 * stall or direct reclaim until kswapd is finished.
3565 */
3566 nr_boost_reclaim = 0;
3567 for (i = 0; i <= highest_zoneidx; i++) {
3568 zone = pgdat->node_zones + i;
3569 if (!managed_zone(zone))
3570 continue;
3571
3572 nr_boost_reclaim += zone->watermark_boost;
3573 zone_boosts[i] = zone->watermark_boost;
3574 }
3575 boosted = nr_boost_reclaim;
3576
3577restart:
3578 sc.priority = DEF_PRIORITY;
3579 do {
3580 unsigned long nr_reclaimed = sc.nr_reclaimed;
3581 bool raise_priority = true;
3582 bool balanced;
3583 bool ret;
3584
3585 sc.reclaim_idx = highest_zoneidx;
3586
3587 /*
3588 * If the number of buffer_heads exceeds the maximum allowed
3589 * then consider reclaiming from all zones. This has a dual
3590 * purpose -- on 64-bit systems it is expected that
3591 * buffer_heads are stripped during active rotation. On 32-bit
3592 * systems, highmem pages can pin lowmem memory and shrinking
3593 * buffers can relieve lowmem pressure. Reclaim may still not
3594 * go ahead if all eligible zones for the original allocation
3595 * request are balanced to avoid excessive reclaim from kswapd.
3596 */
3597 if (buffer_heads_over_limit) {
3598 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3599 zone = pgdat->node_zones + i;
3600 if (!managed_zone(zone))
3601 continue;
3602
3603 sc.reclaim_idx = i;
3604 break;
3605 }
3606 }
3607
3608 /*
3609 * If the pgdat is imbalanced then ignore boosting and preserve
3610 * the watermarks for a later time and restart. Note that the
3611 * zone watermarks will be still reset at the end of balancing
3612 * on the grounds that the normal reclaim should be enough to
3613 * re-evaluate if boosting is required when kswapd next wakes.
3614 */
3615 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3616 if (!balanced && nr_boost_reclaim) {
3617 nr_boost_reclaim = 0;
3618 goto restart;
3619 }
3620
3621 /*
3622 * If boosting is not active then only reclaim if there are no
3623 * eligible zones. Note that sc.reclaim_idx is not used as
3624 * buffer_heads_over_limit may have adjusted it.
3625 */
3626 if (!nr_boost_reclaim && balanced)
3627 goto out;
3628
3629 /* Limit the priority of boosting to avoid reclaim writeback */
3630 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3631 raise_priority = false;
3632
3633 /*
3634 * Do not writeback or swap pages for boosted reclaim. The
3635 * intent is to relieve pressure not issue sub-optimal IO
3636 * from reclaim context. If no pages are reclaimed, the
3637 * reclaim will be aborted.
3638 */
3639 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3640 sc.may_swap = !nr_boost_reclaim;
3641
3642 /*
3643 * Do some background aging of the anon list, to give
3644 * pages a chance to be referenced before reclaiming. All
3645 * pages are rotated regardless of classzone as this is
3646 * about consistent aging.
3647 */
3648 age_active_anon(pgdat, &sc);
3649
3650 /*
3651 * If we're getting trouble reclaiming, start doing writepage
3652 * even in laptop mode.
3653 */
3654 if (sc.priority < DEF_PRIORITY - 2)
3655 sc.may_writepage = 1;
3656
3657 /* Call soft limit reclaim before calling shrink_node. */
3658 sc.nr_scanned = 0;
3659 nr_soft_scanned = 0;
3660 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3661 sc.gfp_mask, &nr_soft_scanned);
3662 sc.nr_reclaimed += nr_soft_reclaimed;
3663
3664 /*
3665 * There should be no need to raise the scanning priority if
3666 * enough pages are already being scanned that that high
3667 * watermark would be met at 100% efficiency.
3668 */
3669 if (kswapd_shrink_node(pgdat, &sc))
3670 raise_priority = false;
3671
3672 /*
3673 * If the low watermark is met there is no need for processes
3674 * to be throttled on pfmemalloc_wait as they should not be
3675 * able to safely make forward progress. Wake them
3676 */
3677 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3678 allow_direct_reclaim(pgdat))
3679 wake_up_all(&pgdat->pfmemalloc_wait);
3680
3681 /* Check if kswapd should be suspending */
3682 __fs_reclaim_release();
3683 ret = try_to_freeze();
3684 __fs_reclaim_acquire();
3685 if (ret || kthread_should_stop())
3686 break;
3687
3688 /*
3689 * Raise priority if scanning rate is too low or there was no
3690 * progress in reclaiming pages
3691 */
3692 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3693 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3694
3695 /*
3696 * If reclaim made no progress for a boost, stop reclaim as
3697 * IO cannot be queued and it could be an infinite loop in
3698 * extreme circumstances.
3699 */
3700 if (nr_boost_reclaim && !nr_reclaimed)
3701 break;
3702
3703 if (raise_priority || !nr_reclaimed)
3704 sc.priority--;
3705 } while (sc.priority >= 1);
3706
3707 if (!sc.nr_reclaimed)
3708 pgdat->kswapd_failures++;
3709
3710out:
3711 /* If reclaim was boosted, account for the reclaim done in this pass */
3712 if (boosted) {
3713 unsigned long flags;
3714
3715 for (i = 0; i <= highest_zoneidx; i++) {
3716 if (!zone_boosts[i])
3717 continue;
3718
3719 /* Increments are under the zone lock */
3720 zone = pgdat->node_zones + i;
3721 spin_lock_irqsave(&zone->lock, flags);
3722 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3723 spin_unlock_irqrestore(&zone->lock, flags);
3724 }
3725
3726 /*
3727 * As there is now likely space, wakeup kcompact to defragment
3728 * pageblocks.
3729 */
3730 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3731 }
3732
3733 snapshot_refaults(NULL, pgdat);
3734 __fs_reclaim_release();
3735 psi_memstall_leave(&pflags);
3736 set_task_reclaim_state(current, NULL);
3737
3738 /*
3739 * Return the order kswapd stopped reclaiming at as
3740 * prepare_kswapd_sleep() takes it into account. If another caller
3741 * entered the allocator slow path while kswapd was awake, order will
3742 * remain at the higher level.
3743 */
3744 return sc.order;
3745}
3746
3747/*
3748 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3749 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3750 * not a valid index then either kswapd runs for first time or kswapd couldn't
3751 * sleep after previous reclaim attempt (node is still unbalanced). In that
3752 * case return the zone index of the previous kswapd reclaim cycle.
3753 */
3754static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3755 enum zone_type prev_highest_zoneidx)
3756{
3757 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3758
3759 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3760}
3761
3762static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3763 unsigned int highest_zoneidx)
3764{
3765 long remaining = 0;
3766 DEFINE_WAIT(wait);
3767
3768 if (freezing(current) || kthread_should_stop())
3769 return;
3770
3771 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3772
3773 /*
3774 * Try to sleep for a short interval. Note that kcompactd will only be
3775 * woken if it is possible to sleep for a short interval. This is
3776 * deliberate on the assumption that if reclaim cannot keep an
3777 * eligible zone balanced that it's also unlikely that compaction will
3778 * succeed.
3779 */
3780 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3781 /*
3782 * Compaction records what page blocks it recently failed to
3783 * isolate pages from and skips them in the future scanning.
3784 * When kswapd is going to sleep, it is reasonable to assume
3785 * that pages and compaction may succeed so reset the cache.
3786 */
3787 reset_isolation_suitable(pgdat);
3788
3789 /*
3790 * We have freed the memory, now we should compact it to make
3791 * allocation of the requested order possible.
3792 */
3793 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3794
3795 remaining = schedule_timeout(HZ/10);
3796
3797 /*
3798 * If woken prematurely then reset kswapd_highest_zoneidx and
3799 * order. The values will either be from a wakeup request or
3800 * the previous request that slept prematurely.
3801 */
3802 if (remaining) {
3803 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3804 kswapd_highest_zoneidx(pgdat,
3805 highest_zoneidx));
3806
3807 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3808 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3809 }
3810
3811 finish_wait(&pgdat->kswapd_wait, &wait);
3812 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3813 }
3814
3815 /*
3816 * After a short sleep, check if it was a premature sleep. If not, then
3817 * go fully to sleep until explicitly woken up.
3818 */
3819 if (!remaining &&
3820 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3821 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3822
3823 /*
3824 * vmstat counters are not perfectly accurate and the estimated
3825 * value for counters such as NR_FREE_PAGES can deviate from the
3826 * true value by nr_online_cpus * threshold. To avoid the zone
3827 * watermarks being breached while under pressure, we reduce the
3828 * per-cpu vmstat threshold while kswapd is awake and restore
3829 * them before going back to sleep.
3830 */
3831 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3832
3833 if (!kthread_should_stop())
3834 schedule();
3835
3836 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3837 } else {
3838 if (remaining)
3839 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3840 else
3841 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3842 }
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3844}
3845
3846/*
3847 * The background pageout daemon, started as a kernel thread
3848 * from the init process.
3849 *
3850 * This basically trickles out pages so that we have _some_
3851 * free memory available even if there is no other activity
3852 * that frees anything up. This is needed for things like routing
3853 * etc, where we otherwise might have all activity going on in
3854 * asynchronous contexts that cannot page things out.
3855 *
3856 * If there are applications that are active memory-allocators
3857 * (most normal use), this basically shouldn't matter.
3858 */
3859static int kswapd(void *p)
3860{
3861 unsigned int alloc_order, reclaim_order;
3862 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3863 pg_data_t *pgdat = (pg_data_t*)p;
3864 struct task_struct *tsk = current;
3865 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3866
3867 if (!cpumask_empty(cpumask))
3868 set_cpus_allowed_ptr(tsk, cpumask);
3869
3870 /*
3871 * Tell the memory management that we're a "memory allocator",
3872 * and that if we need more memory we should get access to it
3873 * regardless (see "__alloc_pages()"). "kswapd" should
3874 * never get caught in the normal page freeing logic.
3875 *
3876 * (Kswapd normally doesn't need memory anyway, but sometimes
3877 * you need a small amount of memory in order to be able to
3878 * page out something else, and this flag essentially protects
3879 * us from recursively trying to free more memory as we're
3880 * trying to free the first piece of memory in the first place).
3881 */
3882 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3883 set_freezable();
3884
3885 WRITE_ONCE(pgdat->kswapd_order, 0);
3886 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3887 for ( ; ; ) {
3888 bool ret;
3889
3890 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3891 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3892 highest_zoneidx);
3893
3894kswapd_try_sleep:
3895 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3896 highest_zoneidx);
3897
3898 /* Read the new order and highest_zoneidx */
3899 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3900 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3901 highest_zoneidx);
3902 WRITE_ONCE(pgdat->kswapd_order, 0);
3903 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3904
3905 ret = try_to_freeze();
3906 if (kthread_should_stop())
3907 break;
3908
3909 /*
3910 * We can speed up thawing tasks if we don't call balance_pgdat
3911 * after returning from the refrigerator
3912 */
3913 if (ret)
3914 continue;
3915
3916 /*
3917 * Reclaim begins at the requested order but if a high-order
3918 * reclaim fails then kswapd falls back to reclaiming for
3919 * order-0. If that happens, kswapd will consider sleeping
3920 * for the order it finished reclaiming at (reclaim_order)
3921 * but kcompactd is woken to compact for the original
3922 * request (alloc_order).
3923 */
3924 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3925 alloc_order);
3926 reclaim_order = balance_pgdat(pgdat, alloc_order,
3927 highest_zoneidx);
3928 if (reclaim_order < alloc_order)
3929 goto kswapd_try_sleep;
3930 }
3931
3932 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3933
3934 return 0;
3935}
3936
3937/*
3938 * A zone is low on free memory or too fragmented for high-order memory. If
3939 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3940 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3941 * has failed or is not needed, still wake up kcompactd if only compaction is
3942 * needed.
3943 */
3944void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3945 enum zone_type highest_zoneidx)
3946{
3947 pg_data_t *pgdat;
3948 enum zone_type curr_idx;
3949
3950 if (!managed_zone(zone))
3951 return;
3952
3953 if (!cpuset_zone_allowed(zone, gfp_flags))
3954 return;
3955
3956 pgdat = zone->zone_pgdat;
3957 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3958
3959 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3960 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3961
3962 if (READ_ONCE(pgdat->kswapd_order) < order)
3963 WRITE_ONCE(pgdat->kswapd_order, order);
3964
3965 if (!waitqueue_active(&pgdat->kswapd_wait))
3966 return;
3967
3968 /* Hopeless node, leave it to direct reclaim if possible */
3969 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3970 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3971 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3972 /*
3973 * There may be plenty of free memory available, but it's too
3974 * fragmented for high-order allocations. Wake up kcompactd
3975 * and rely on compaction_suitable() to determine if it's
3976 * needed. If it fails, it will defer subsequent attempts to
3977 * ratelimit its work.
3978 */
3979 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3980 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3981 return;
3982 }
3983
3984 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3985 gfp_flags);
3986 wake_up_interruptible(&pgdat->kswapd_wait);
3987}
3988
3989#ifdef CONFIG_HIBERNATION
3990/*
3991 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3992 * freed pages.
3993 *
3994 * Rather than trying to age LRUs the aim is to preserve the overall
3995 * LRU order by reclaiming preferentially
3996 * inactive > active > active referenced > active mapped
3997 */
3998unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3999{
4000 struct scan_control sc = {
4001 .nr_to_reclaim = nr_to_reclaim,
4002 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4003 .reclaim_idx = MAX_NR_ZONES - 1,
4004 .priority = DEF_PRIORITY,
4005 .may_writepage = 1,
4006 .may_unmap = 1,
4007 .may_swap = 1,
4008 .hibernation_mode = 1,
4009 };
4010 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4011 unsigned long nr_reclaimed;
4012 unsigned int noreclaim_flag;
4013
4014 fs_reclaim_acquire(sc.gfp_mask);
4015 noreclaim_flag = memalloc_noreclaim_save();
4016 set_task_reclaim_state(current, &sc.reclaim_state);
4017
4018 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4019
4020 set_task_reclaim_state(current, NULL);
4021 memalloc_noreclaim_restore(noreclaim_flag);
4022 fs_reclaim_release(sc.gfp_mask);
4023
4024 return nr_reclaimed;
4025}
4026#endif /* CONFIG_HIBERNATION */
4027
4028/*
4029 * This kswapd start function will be called by init and node-hot-add.
4030 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4031 */
4032int kswapd_run(int nid)
4033{
4034 pg_data_t *pgdat = NODE_DATA(nid);
4035 int ret = 0;
4036
4037 if (pgdat->kswapd)
4038 return 0;
4039
4040 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4041 if (IS_ERR(pgdat->kswapd)) {
4042 /* failure at boot is fatal */
4043 BUG_ON(system_state < SYSTEM_RUNNING);
4044 pr_err("Failed to start kswapd on node %d\n", nid);
4045 ret = PTR_ERR(pgdat->kswapd);
4046 pgdat->kswapd = NULL;
4047 }
4048 return ret;
4049}
4050
4051/*
4052 * Called by memory hotplug when all memory in a node is offlined. Caller must
4053 * hold mem_hotplug_begin/end().
4054 */
4055void kswapd_stop(int nid)
4056{
4057 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4058
4059 if (kswapd) {
4060 kthread_stop(kswapd);
4061 NODE_DATA(nid)->kswapd = NULL;
4062 }
4063}
4064
4065static int __init kswapd_init(void)
4066{
4067 int nid;
4068
4069 swap_setup();
4070 for_each_node_state(nid, N_MEMORY)
4071 kswapd_run(nid);
4072 return 0;
4073}
4074
4075module_init(kswapd_init)
4076
4077#ifdef CONFIG_NUMA
4078/*
4079 * Node reclaim mode
4080 *
4081 * If non-zero call node_reclaim when the number of free pages falls below
4082 * the watermarks.
4083 */
4084int node_reclaim_mode __read_mostly;
4085
4086#define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4087#define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4088
4089/*
4090 * Priority for NODE_RECLAIM. This determines the fraction of pages
4091 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4092 * a zone.
4093 */
4094#define NODE_RECLAIM_PRIORITY 4
4095
4096/*
4097 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4098 * occur.
4099 */
4100int sysctl_min_unmapped_ratio = 1;
4101
4102/*
4103 * If the number of slab pages in a zone grows beyond this percentage then
4104 * slab reclaim needs to occur.
4105 */
4106int sysctl_min_slab_ratio = 5;
4107
4108static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4109{
4110 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4111 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4112 node_page_state(pgdat, NR_ACTIVE_FILE);
4113
4114 /*
4115 * It's possible for there to be more file mapped pages than
4116 * accounted for by the pages on the file LRU lists because
4117 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4118 */
4119 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4120}
4121
4122/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4123static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4124{
4125 unsigned long nr_pagecache_reclaimable;
4126 unsigned long delta = 0;
4127
4128 /*
4129 * If RECLAIM_UNMAP is set, then all file pages are considered
4130 * potentially reclaimable. Otherwise, we have to worry about
4131 * pages like swapcache and node_unmapped_file_pages() provides
4132 * a better estimate
4133 */
4134 if (node_reclaim_mode & RECLAIM_UNMAP)
4135 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4136 else
4137 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4138
4139 /* If we can't clean pages, remove dirty pages from consideration */
4140 if (!(node_reclaim_mode & RECLAIM_WRITE))
4141 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4142
4143 /* Watch for any possible underflows due to delta */
4144 if (unlikely(delta > nr_pagecache_reclaimable))
4145 delta = nr_pagecache_reclaimable;
4146
4147 return nr_pagecache_reclaimable - delta;
4148}
4149
4150/*
4151 * Try to free up some pages from this node through reclaim.
4152 */
4153static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4154{
4155 /* Minimum pages needed in order to stay on node */
4156 const unsigned long nr_pages = 1 << order;
4157 struct task_struct *p = current;
4158 unsigned int noreclaim_flag;
4159 struct scan_control sc = {
4160 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4161 .gfp_mask = current_gfp_context(gfp_mask),
4162 .order = order,
4163 .priority = NODE_RECLAIM_PRIORITY,
4164 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4165 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4166 .may_swap = 1,
4167 .reclaim_idx = gfp_zone(gfp_mask),
4168 };
4169
4170 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4171 sc.gfp_mask);
4172
4173 cond_resched();
4174 fs_reclaim_acquire(sc.gfp_mask);
4175 /*
4176 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4177 * and we also need to be able to write out pages for RECLAIM_WRITE
4178 * and RECLAIM_UNMAP.
4179 */
4180 noreclaim_flag = memalloc_noreclaim_save();
4181 p->flags |= PF_SWAPWRITE;
4182 set_task_reclaim_state(p, &sc.reclaim_state);
4183
4184 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4185 /*
4186 * Free memory by calling shrink node with increasing
4187 * priorities until we have enough memory freed.
4188 */
4189 do {
4190 shrink_node(pgdat, &sc);
4191 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4192 }
4193
4194 set_task_reclaim_state(p, NULL);
4195 current->flags &= ~PF_SWAPWRITE;
4196 memalloc_noreclaim_restore(noreclaim_flag);
4197 fs_reclaim_release(sc.gfp_mask);
4198
4199 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4200
4201 return sc.nr_reclaimed >= nr_pages;
4202}
4203
4204int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4205{
4206 int ret;
4207
4208 /*
4209 * Node reclaim reclaims unmapped file backed pages and
4210 * slab pages if we are over the defined limits.
4211 *
4212 * A small portion of unmapped file backed pages is needed for
4213 * file I/O otherwise pages read by file I/O will be immediately
4214 * thrown out if the node is overallocated. So we do not reclaim
4215 * if less than a specified percentage of the node is used by
4216 * unmapped file backed pages.
4217 */
4218 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4219 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4220 pgdat->min_slab_pages)
4221 return NODE_RECLAIM_FULL;
4222
4223 /*
4224 * Do not scan if the allocation should not be delayed.
4225 */
4226 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4227 return NODE_RECLAIM_NOSCAN;
4228
4229 /*
4230 * Only run node reclaim on the local node or on nodes that do not
4231 * have associated processors. This will favor the local processor
4232 * over remote processors and spread off node memory allocations
4233 * as wide as possible.
4234 */
4235 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4236 return NODE_RECLAIM_NOSCAN;
4237
4238 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4239 return NODE_RECLAIM_NOSCAN;
4240
4241 ret = __node_reclaim(pgdat, gfp_mask, order);
4242 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4243
4244 if (!ret)
4245 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4246
4247 return ret;
4248}
4249#endif
4250
4251/**
4252 * check_move_unevictable_pages - check pages for evictability and move to
4253 * appropriate zone lru list
4254 * @pvec: pagevec with lru pages to check
4255 *
4256 * Checks pages for evictability, if an evictable page is in the unevictable
4257 * lru list, moves it to the appropriate evictable lru list. This function
4258 * should be only used for lru pages.
4259 */
4260void check_move_unevictable_pages(struct pagevec *pvec)
4261{
4262 struct lruvec *lruvec;
4263 struct pglist_data *pgdat = NULL;
4264 int pgscanned = 0;
4265 int pgrescued = 0;
4266 int i;
4267
4268 for (i = 0; i < pvec->nr; i++) {
4269 struct page *page = pvec->pages[i];
4270 struct pglist_data *pagepgdat = page_pgdat(page);
4271 int nr_pages;
4272
4273 if (PageTransTail(page))
4274 continue;
4275
4276 nr_pages = thp_nr_pages(page);
4277 pgscanned += nr_pages;
4278
4279 if (pagepgdat != pgdat) {
4280 if (pgdat)
4281 spin_unlock_irq(&pgdat->lru_lock);
4282 pgdat = pagepgdat;
4283 spin_lock_irq(&pgdat->lru_lock);
4284 }
4285 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4286
4287 if (!PageLRU(page) || !PageUnevictable(page))
4288 continue;
4289
4290 if (page_evictable(page)) {
4291 enum lru_list lru = page_lru_base_type(page);
4292
4293 VM_BUG_ON_PAGE(PageActive(page), page);
4294 ClearPageUnevictable(page);
4295 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4296 add_page_to_lru_list(page, lruvec, lru);
4297 pgrescued += nr_pages;
4298 }
4299 }
4300
4301 if (pgdat) {
4302 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4303 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4304 spin_unlock_irq(&pgdat->lru_lock);
4305 }
4306}
4307EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4 *
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
11 */
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/mm.h>
16#include <linux/sched/mm.h>
17#include <linux/module.h>
18#include <linux/gfp.h>
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
24#include <linux/vmpressure.h>
25#include <linux/vmstat.h>
26#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
37#include <linux/compaction.h>
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
40#include <linux/delay.h>
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43#include <linux/memcontrol.h>
44#include <linux/delayacct.h>
45#include <linux/sysctl.h>
46#include <linux/oom.h>
47#include <linux/pagevec.h>
48#include <linux/prefetch.h>
49#include <linux/printk.h>
50#include <linux/dax.h>
51#include <linux/psi.h>
52
53#include <asm/tlbflush.h>
54#include <asm/div64.h>
55
56#include <linux/swapops.h>
57#include <linux/balloon_compaction.h>
58
59#include "internal.h"
60
61#define CREATE_TRACE_POINTS
62#include <trace/events/vmscan.h>
63
64struct scan_control {
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
68 /*
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
70 * are scanned.
71 */
72 nodemask_t *nodemask;
73
74 /*
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
77 */
78 struct mem_cgroup *target_mem_cgroup;
79
80 /*
81 * Scan pressure balancing between anon and file LRUs
82 */
83 unsigned long anon_cost;
84 unsigned long file_cost;
85
86 /* Can active pages be deactivated as part of reclaim? */
87#define DEACTIVATE_ANON 1
88#define DEACTIVATE_FILE 2
89 unsigned int may_deactivate:2;
90 unsigned int force_deactivate:1;
91 unsigned int skipped_deactivate:1;
92
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage:1;
95
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap:1;
98
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap:1;
101
102 /*
103 * Cgroup memory below memory.low is protected as long as we
104 * don't threaten to OOM. If any cgroup is reclaimed at
105 * reduced force or passed over entirely due to its memory.low
106 * setting (memcg_low_skipped), and nothing is reclaimed as a
107 * result, then go back for one more cycle that reclaims the protected
108 * memory (memcg_low_reclaim) to avert OOM.
109 */
110 unsigned int memcg_low_reclaim:1;
111 unsigned int memcg_low_skipped:1;
112
113 unsigned int hibernation_mode:1;
114
115 /* One of the zones is ready for compaction */
116 unsigned int compaction_ready:1;
117
118 /* There is easily reclaimable cold cache in the current node */
119 unsigned int cache_trim_mode:1;
120
121 /* The file pages on the current node are dangerously low */
122 unsigned int file_is_tiny:1;
123
124 /* Allocation order */
125 s8 order;
126
127 /* Scan (total_size >> priority) pages at once */
128 s8 priority;
129
130 /* The highest zone to isolate pages for reclaim from */
131 s8 reclaim_idx;
132
133 /* This context's GFP mask */
134 gfp_t gfp_mask;
135
136 /* Incremented by the number of inactive pages that were scanned */
137 unsigned long nr_scanned;
138
139 /* Number of pages freed so far during a call to shrink_zones() */
140 unsigned long nr_reclaimed;
141
142 struct {
143 unsigned int dirty;
144 unsigned int unqueued_dirty;
145 unsigned int congested;
146 unsigned int writeback;
147 unsigned int immediate;
148 unsigned int file_taken;
149 unsigned int taken;
150 } nr;
151
152 /* for recording the reclaimed slab by now */
153 struct reclaim_state reclaim_state;
154};
155
156#ifdef ARCH_HAS_PREFETCHW
157#define prefetchw_prev_lru_page(_page, _base, _field) \
158 do { \
159 if ((_page)->lru.prev != _base) { \
160 struct page *prev; \
161 \
162 prev = lru_to_page(&(_page->lru)); \
163 prefetchw(&prev->_field); \
164 } \
165 } while (0)
166#else
167#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168#endif
169
170/*
171 * From 0 .. 200. Higher means more swappy.
172 */
173int vm_swappiness = 60;
174
175static void set_task_reclaim_state(struct task_struct *task,
176 struct reclaim_state *rs)
177{
178 /* Check for an overwrite */
179 WARN_ON_ONCE(rs && task->reclaim_state);
180
181 /* Check for the nulling of an already-nulled member */
182 WARN_ON_ONCE(!rs && !task->reclaim_state);
183
184 task->reclaim_state = rs;
185}
186
187static LIST_HEAD(shrinker_list);
188static DECLARE_RWSEM(shrinker_rwsem);
189
190#ifdef CONFIG_MEMCG
191static int shrinker_nr_max;
192
193/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
194static inline int shrinker_map_size(int nr_items)
195{
196 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
197}
198
199static inline int shrinker_defer_size(int nr_items)
200{
201 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
202}
203
204static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
205 int nid)
206{
207 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
208 lockdep_is_held(&shrinker_rwsem));
209}
210
211static int expand_one_shrinker_info(struct mem_cgroup *memcg,
212 int map_size, int defer_size,
213 int old_map_size, int old_defer_size)
214{
215 struct shrinker_info *new, *old;
216 struct mem_cgroup_per_node *pn;
217 int nid;
218 int size = map_size + defer_size;
219
220 for_each_node(nid) {
221 pn = memcg->nodeinfo[nid];
222 old = shrinker_info_protected(memcg, nid);
223 /* Not yet online memcg */
224 if (!old)
225 return 0;
226
227 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
228 if (!new)
229 return -ENOMEM;
230
231 new->nr_deferred = (atomic_long_t *)(new + 1);
232 new->map = (void *)new->nr_deferred + defer_size;
233
234 /* map: set all old bits, clear all new bits */
235 memset(new->map, (int)0xff, old_map_size);
236 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
237 /* nr_deferred: copy old values, clear all new values */
238 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
239 memset((void *)new->nr_deferred + old_defer_size, 0,
240 defer_size - old_defer_size);
241
242 rcu_assign_pointer(pn->shrinker_info, new);
243 kvfree_rcu(old, rcu);
244 }
245
246 return 0;
247}
248
249void free_shrinker_info(struct mem_cgroup *memcg)
250{
251 struct mem_cgroup_per_node *pn;
252 struct shrinker_info *info;
253 int nid;
254
255 for_each_node(nid) {
256 pn = memcg->nodeinfo[nid];
257 info = rcu_dereference_protected(pn->shrinker_info, true);
258 kvfree(info);
259 rcu_assign_pointer(pn->shrinker_info, NULL);
260 }
261}
262
263int alloc_shrinker_info(struct mem_cgroup *memcg)
264{
265 struct shrinker_info *info;
266 int nid, size, ret = 0;
267 int map_size, defer_size = 0;
268
269 down_write(&shrinker_rwsem);
270 map_size = shrinker_map_size(shrinker_nr_max);
271 defer_size = shrinker_defer_size(shrinker_nr_max);
272 size = map_size + defer_size;
273 for_each_node(nid) {
274 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
275 if (!info) {
276 free_shrinker_info(memcg);
277 ret = -ENOMEM;
278 break;
279 }
280 info->nr_deferred = (atomic_long_t *)(info + 1);
281 info->map = (void *)info->nr_deferred + defer_size;
282 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
283 }
284 up_write(&shrinker_rwsem);
285
286 return ret;
287}
288
289static inline bool need_expand(int nr_max)
290{
291 return round_up(nr_max, BITS_PER_LONG) >
292 round_up(shrinker_nr_max, BITS_PER_LONG);
293}
294
295static int expand_shrinker_info(int new_id)
296{
297 int ret = 0;
298 int new_nr_max = new_id + 1;
299 int map_size, defer_size = 0;
300 int old_map_size, old_defer_size = 0;
301 struct mem_cgroup *memcg;
302
303 if (!need_expand(new_nr_max))
304 goto out;
305
306 if (!root_mem_cgroup)
307 goto out;
308
309 lockdep_assert_held(&shrinker_rwsem);
310
311 map_size = shrinker_map_size(new_nr_max);
312 defer_size = shrinker_defer_size(new_nr_max);
313 old_map_size = shrinker_map_size(shrinker_nr_max);
314 old_defer_size = shrinker_defer_size(shrinker_nr_max);
315
316 memcg = mem_cgroup_iter(NULL, NULL, NULL);
317 do {
318 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
319 old_map_size, old_defer_size);
320 if (ret) {
321 mem_cgroup_iter_break(NULL, memcg);
322 goto out;
323 }
324 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
325out:
326 if (!ret)
327 shrinker_nr_max = new_nr_max;
328
329 return ret;
330}
331
332void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
333{
334 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
335 struct shrinker_info *info;
336
337 rcu_read_lock();
338 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
339 /* Pairs with smp mb in shrink_slab() */
340 smp_mb__before_atomic();
341 set_bit(shrinker_id, info->map);
342 rcu_read_unlock();
343 }
344}
345
346static DEFINE_IDR(shrinker_idr);
347
348static int prealloc_memcg_shrinker(struct shrinker *shrinker)
349{
350 int id, ret = -ENOMEM;
351
352 if (mem_cgroup_disabled())
353 return -ENOSYS;
354
355 down_write(&shrinker_rwsem);
356 /* This may call shrinker, so it must use down_read_trylock() */
357 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
358 if (id < 0)
359 goto unlock;
360
361 if (id >= shrinker_nr_max) {
362 if (expand_shrinker_info(id)) {
363 idr_remove(&shrinker_idr, id);
364 goto unlock;
365 }
366 }
367 shrinker->id = id;
368 ret = 0;
369unlock:
370 up_write(&shrinker_rwsem);
371 return ret;
372}
373
374static void unregister_memcg_shrinker(struct shrinker *shrinker)
375{
376 int id = shrinker->id;
377
378 BUG_ON(id < 0);
379
380 lockdep_assert_held(&shrinker_rwsem);
381
382 idr_remove(&shrinker_idr, id);
383}
384
385static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
386 struct mem_cgroup *memcg)
387{
388 struct shrinker_info *info;
389
390 info = shrinker_info_protected(memcg, nid);
391 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
392}
393
394static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
395 struct mem_cgroup *memcg)
396{
397 struct shrinker_info *info;
398
399 info = shrinker_info_protected(memcg, nid);
400 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
401}
402
403void reparent_shrinker_deferred(struct mem_cgroup *memcg)
404{
405 int i, nid;
406 long nr;
407 struct mem_cgroup *parent;
408 struct shrinker_info *child_info, *parent_info;
409
410 parent = parent_mem_cgroup(memcg);
411 if (!parent)
412 parent = root_mem_cgroup;
413
414 /* Prevent from concurrent shrinker_info expand */
415 down_read(&shrinker_rwsem);
416 for_each_node(nid) {
417 child_info = shrinker_info_protected(memcg, nid);
418 parent_info = shrinker_info_protected(parent, nid);
419 for (i = 0; i < shrinker_nr_max; i++) {
420 nr = atomic_long_read(&child_info->nr_deferred[i]);
421 atomic_long_add(nr, &parent_info->nr_deferred[i]);
422 }
423 }
424 up_read(&shrinker_rwsem);
425}
426
427static bool cgroup_reclaim(struct scan_control *sc)
428{
429 return sc->target_mem_cgroup;
430}
431
432/**
433 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
434 * @sc: scan_control in question
435 *
436 * The normal page dirty throttling mechanism in balance_dirty_pages() is
437 * completely broken with the legacy memcg and direct stalling in
438 * shrink_page_list() is used for throttling instead, which lacks all the
439 * niceties such as fairness, adaptive pausing, bandwidth proportional
440 * allocation and configurability.
441 *
442 * This function tests whether the vmscan currently in progress can assume
443 * that the normal dirty throttling mechanism is operational.
444 */
445static bool writeback_throttling_sane(struct scan_control *sc)
446{
447 if (!cgroup_reclaim(sc))
448 return true;
449#ifdef CONFIG_CGROUP_WRITEBACK
450 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
451 return true;
452#endif
453 return false;
454}
455#else
456static int prealloc_memcg_shrinker(struct shrinker *shrinker)
457{
458 return -ENOSYS;
459}
460
461static void unregister_memcg_shrinker(struct shrinker *shrinker)
462{
463}
464
465static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
466 struct mem_cgroup *memcg)
467{
468 return 0;
469}
470
471static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
472 struct mem_cgroup *memcg)
473{
474 return 0;
475}
476
477static bool cgroup_reclaim(struct scan_control *sc)
478{
479 return false;
480}
481
482static bool writeback_throttling_sane(struct scan_control *sc)
483{
484 return true;
485}
486#endif
487
488static long xchg_nr_deferred(struct shrinker *shrinker,
489 struct shrink_control *sc)
490{
491 int nid = sc->nid;
492
493 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
494 nid = 0;
495
496 if (sc->memcg &&
497 (shrinker->flags & SHRINKER_MEMCG_AWARE))
498 return xchg_nr_deferred_memcg(nid, shrinker,
499 sc->memcg);
500
501 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
502}
503
504
505static long add_nr_deferred(long nr, struct shrinker *shrinker,
506 struct shrink_control *sc)
507{
508 int nid = sc->nid;
509
510 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
511 nid = 0;
512
513 if (sc->memcg &&
514 (shrinker->flags & SHRINKER_MEMCG_AWARE))
515 return add_nr_deferred_memcg(nr, nid, shrinker,
516 sc->memcg);
517
518 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
519}
520
521/*
522 * This misses isolated pages which are not accounted for to save counters.
523 * As the data only determines if reclaim or compaction continues, it is
524 * not expected that isolated pages will be a dominating factor.
525 */
526unsigned long zone_reclaimable_pages(struct zone *zone)
527{
528 unsigned long nr;
529
530 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
531 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
532 if (get_nr_swap_pages() > 0)
533 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
534 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
535
536 return nr;
537}
538
539/**
540 * lruvec_lru_size - Returns the number of pages on the given LRU list.
541 * @lruvec: lru vector
542 * @lru: lru to use
543 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
544 */
545static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
546 int zone_idx)
547{
548 unsigned long size = 0;
549 int zid;
550
551 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
552 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
553
554 if (!managed_zone(zone))
555 continue;
556
557 if (!mem_cgroup_disabled())
558 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
559 else
560 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
561 }
562 return size;
563}
564
565/*
566 * Add a shrinker callback to be called from the vm.
567 */
568int prealloc_shrinker(struct shrinker *shrinker)
569{
570 unsigned int size;
571 int err;
572
573 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
574 err = prealloc_memcg_shrinker(shrinker);
575 if (err != -ENOSYS)
576 return err;
577
578 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
579 }
580
581 size = sizeof(*shrinker->nr_deferred);
582 if (shrinker->flags & SHRINKER_NUMA_AWARE)
583 size *= nr_node_ids;
584
585 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
586 if (!shrinker->nr_deferred)
587 return -ENOMEM;
588
589 return 0;
590}
591
592void free_prealloced_shrinker(struct shrinker *shrinker)
593{
594 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
595 down_write(&shrinker_rwsem);
596 unregister_memcg_shrinker(shrinker);
597 up_write(&shrinker_rwsem);
598 return;
599 }
600
601 kfree(shrinker->nr_deferred);
602 shrinker->nr_deferred = NULL;
603}
604
605void register_shrinker_prepared(struct shrinker *shrinker)
606{
607 down_write(&shrinker_rwsem);
608 list_add_tail(&shrinker->list, &shrinker_list);
609 shrinker->flags |= SHRINKER_REGISTERED;
610 up_write(&shrinker_rwsem);
611}
612
613int register_shrinker(struct shrinker *shrinker)
614{
615 int err = prealloc_shrinker(shrinker);
616
617 if (err)
618 return err;
619 register_shrinker_prepared(shrinker);
620 return 0;
621}
622EXPORT_SYMBOL(register_shrinker);
623
624/*
625 * Remove one
626 */
627void unregister_shrinker(struct shrinker *shrinker)
628{
629 if (!(shrinker->flags & SHRINKER_REGISTERED))
630 return;
631
632 down_write(&shrinker_rwsem);
633 list_del(&shrinker->list);
634 shrinker->flags &= ~SHRINKER_REGISTERED;
635 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
636 unregister_memcg_shrinker(shrinker);
637 up_write(&shrinker_rwsem);
638
639 kfree(shrinker->nr_deferred);
640 shrinker->nr_deferred = NULL;
641}
642EXPORT_SYMBOL(unregister_shrinker);
643
644#define SHRINK_BATCH 128
645
646static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
647 struct shrinker *shrinker, int priority)
648{
649 unsigned long freed = 0;
650 unsigned long long delta;
651 long total_scan;
652 long freeable;
653 long nr;
654 long new_nr;
655 long batch_size = shrinker->batch ? shrinker->batch
656 : SHRINK_BATCH;
657 long scanned = 0, next_deferred;
658
659 freeable = shrinker->count_objects(shrinker, shrinkctl);
660 if (freeable == 0 || freeable == SHRINK_EMPTY)
661 return freeable;
662
663 /*
664 * copy the current shrinker scan count into a local variable
665 * and zero it so that other concurrent shrinker invocations
666 * don't also do this scanning work.
667 */
668 nr = xchg_nr_deferred(shrinker, shrinkctl);
669
670 if (shrinker->seeks) {
671 delta = freeable >> priority;
672 delta *= 4;
673 do_div(delta, shrinker->seeks);
674 } else {
675 /*
676 * These objects don't require any IO to create. Trim
677 * them aggressively under memory pressure to keep
678 * them from causing refetches in the IO caches.
679 */
680 delta = freeable / 2;
681 }
682
683 total_scan = nr >> priority;
684 total_scan += delta;
685 total_scan = min(total_scan, (2 * freeable));
686
687 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
688 freeable, delta, total_scan, priority);
689
690 /*
691 * Normally, we should not scan less than batch_size objects in one
692 * pass to avoid too frequent shrinker calls, but if the slab has less
693 * than batch_size objects in total and we are really tight on memory,
694 * we will try to reclaim all available objects, otherwise we can end
695 * up failing allocations although there are plenty of reclaimable
696 * objects spread over several slabs with usage less than the
697 * batch_size.
698 *
699 * We detect the "tight on memory" situations by looking at the total
700 * number of objects we want to scan (total_scan). If it is greater
701 * than the total number of objects on slab (freeable), we must be
702 * scanning at high prio and therefore should try to reclaim as much as
703 * possible.
704 */
705 while (total_scan >= batch_size ||
706 total_scan >= freeable) {
707 unsigned long ret;
708 unsigned long nr_to_scan = min(batch_size, total_scan);
709
710 shrinkctl->nr_to_scan = nr_to_scan;
711 shrinkctl->nr_scanned = nr_to_scan;
712 ret = shrinker->scan_objects(shrinker, shrinkctl);
713 if (ret == SHRINK_STOP)
714 break;
715 freed += ret;
716
717 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
718 total_scan -= shrinkctl->nr_scanned;
719 scanned += shrinkctl->nr_scanned;
720
721 cond_resched();
722 }
723
724 /*
725 * The deferred work is increased by any new work (delta) that wasn't
726 * done, decreased by old deferred work that was done now.
727 *
728 * And it is capped to two times of the freeable items.
729 */
730 next_deferred = max_t(long, (nr + delta - scanned), 0);
731 next_deferred = min(next_deferred, (2 * freeable));
732
733 /*
734 * move the unused scan count back into the shrinker in a
735 * manner that handles concurrent updates.
736 */
737 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
738
739 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
740 return freed;
741}
742
743#ifdef CONFIG_MEMCG
744static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
745 struct mem_cgroup *memcg, int priority)
746{
747 struct shrinker_info *info;
748 unsigned long ret, freed = 0;
749 int i;
750
751 if (!mem_cgroup_online(memcg))
752 return 0;
753
754 if (!down_read_trylock(&shrinker_rwsem))
755 return 0;
756
757 info = shrinker_info_protected(memcg, nid);
758 if (unlikely(!info))
759 goto unlock;
760
761 for_each_set_bit(i, info->map, shrinker_nr_max) {
762 struct shrink_control sc = {
763 .gfp_mask = gfp_mask,
764 .nid = nid,
765 .memcg = memcg,
766 };
767 struct shrinker *shrinker;
768
769 shrinker = idr_find(&shrinker_idr, i);
770 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
771 if (!shrinker)
772 clear_bit(i, info->map);
773 continue;
774 }
775
776 /* Call non-slab shrinkers even though kmem is disabled */
777 if (!memcg_kmem_enabled() &&
778 !(shrinker->flags & SHRINKER_NONSLAB))
779 continue;
780
781 ret = do_shrink_slab(&sc, shrinker, priority);
782 if (ret == SHRINK_EMPTY) {
783 clear_bit(i, info->map);
784 /*
785 * After the shrinker reported that it had no objects to
786 * free, but before we cleared the corresponding bit in
787 * the memcg shrinker map, a new object might have been
788 * added. To make sure, we have the bit set in this
789 * case, we invoke the shrinker one more time and reset
790 * the bit if it reports that it is not empty anymore.
791 * The memory barrier here pairs with the barrier in
792 * set_shrinker_bit():
793 *
794 * list_lru_add() shrink_slab_memcg()
795 * list_add_tail() clear_bit()
796 * <MB> <MB>
797 * set_bit() do_shrink_slab()
798 */
799 smp_mb__after_atomic();
800 ret = do_shrink_slab(&sc, shrinker, priority);
801 if (ret == SHRINK_EMPTY)
802 ret = 0;
803 else
804 set_shrinker_bit(memcg, nid, i);
805 }
806 freed += ret;
807
808 if (rwsem_is_contended(&shrinker_rwsem)) {
809 freed = freed ? : 1;
810 break;
811 }
812 }
813unlock:
814 up_read(&shrinker_rwsem);
815 return freed;
816}
817#else /* CONFIG_MEMCG */
818static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
819 struct mem_cgroup *memcg, int priority)
820{
821 return 0;
822}
823#endif /* CONFIG_MEMCG */
824
825/**
826 * shrink_slab - shrink slab caches
827 * @gfp_mask: allocation context
828 * @nid: node whose slab caches to target
829 * @memcg: memory cgroup whose slab caches to target
830 * @priority: the reclaim priority
831 *
832 * Call the shrink functions to age shrinkable caches.
833 *
834 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
835 * unaware shrinkers will receive a node id of 0 instead.
836 *
837 * @memcg specifies the memory cgroup to target. Unaware shrinkers
838 * are called only if it is the root cgroup.
839 *
840 * @priority is sc->priority, we take the number of objects and >> by priority
841 * in order to get the scan target.
842 *
843 * Returns the number of reclaimed slab objects.
844 */
845static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
846 struct mem_cgroup *memcg,
847 int priority)
848{
849 unsigned long ret, freed = 0;
850 struct shrinker *shrinker;
851
852 /*
853 * The root memcg might be allocated even though memcg is disabled
854 * via "cgroup_disable=memory" boot parameter. This could make
855 * mem_cgroup_is_root() return false, then just run memcg slab
856 * shrink, but skip global shrink. This may result in premature
857 * oom.
858 */
859 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
860 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
861
862 if (!down_read_trylock(&shrinker_rwsem))
863 goto out;
864
865 list_for_each_entry(shrinker, &shrinker_list, list) {
866 struct shrink_control sc = {
867 .gfp_mask = gfp_mask,
868 .nid = nid,
869 .memcg = memcg,
870 };
871
872 ret = do_shrink_slab(&sc, shrinker, priority);
873 if (ret == SHRINK_EMPTY)
874 ret = 0;
875 freed += ret;
876 /*
877 * Bail out if someone want to register a new shrinker to
878 * prevent the registration from being stalled for long periods
879 * by parallel ongoing shrinking.
880 */
881 if (rwsem_is_contended(&shrinker_rwsem)) {
882 freed = freed ? : 1;
883 break;
884 }
885 }
886
887 up_read(&shrinker_rwsem);
888out:
889 cond_resched();
890 return freed;
891}
892
893void drop_slab_node(int nid)
894{
895 unsigned long freed;
896
897 do {
898 struct mem_cgroup *memcg = NULL;
899
900 if (fatal_signal_pending(current))
901 return;
902
903 freed = 0;
904 memcg = mem_cgroup_iter(NULL, NULL, NULL);
905 do {
906 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
907 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
908 } while (freed > 10);
909}
910
911void drop_slab(void)
912{
913 int nid;
914
915 for_each_online_node(nid)
916 drop_slab_node(nid);
917}
918
919static inline int is_page_cache_freeable(struct page *page)
920{
921 /*
922 * A freeable page cache page is referenced only by the caller
923 * that isolated the page, the page cache and optional buffer
924 * heads at page->private.
925 */
926 int page_cache_pins = thp_nr_pages(page);
927 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
928}
929
930static int may_write_to_inode(struct inode *inode)
931{
932 if (current->flags & PF_SWAPWRITE)
933 return 1;
934 if (!inode_write_congested(inode))
935 return 1;
936 if (inode_to_bdi(inode) == current->backing_dev_info)
937 return 1;
938 return 0;
939}
940
941/*
942 * We detected a synchronous write error writing a page out. Probably
943 * -ENOSPC. We need to propagate that into the address_space for a subsequent
944 * fsync(), msync() or close().
945 *
946 * The tricky part is that after writepage we cannot touch the mapping: nothing
947 * prevents it from being freed up. But we have a ref on the page and once
948 * that page is locked, the mapping is pinned.
949 *
950 * We're allowed to run sleeping lock_page() here because we know the caller has
951 * __GFP_FS.
952 */
953static void handle_write_error(struct address_space *mapping,
954 struct page *page, int error)
955{
956 lock_page(page);
957 if (page_mapping(page) == mapping)
958 mapping_set_error(mapping, error);
959 unlock_page(page);
960}
961
962/* possible outcome of pageout() */
963typedef enum {
964 /* failed to write page out, page is locked */
965 PAGE_KEEP,
966 /* move page to the active list, page is locked */
967 PAGE_ACTIVATE,
968 /* page has been sent to the disk successfully, page is unlocked */
969 PAGE_SUCCESS,
970 /* page is clean and locked */
971 PAGE_CLEAN,
972} pageout_t;
973
974/*
975 * pageout is called by shrink_page_list() for each dirty page.
976 * Calls ->writepage().
977 */
978static pageout_t pageout(struct page *page, struct address_space *mapping)
979{
980 /*
981 * If the page is dirty, only perform writeback if that write
982 * will be non-blocking. To prevent this allocation from being
983 * stalled by pagecache activity. But note that there may be
984 * stalls if we need to run get_block(). We could test
985 * PagePrivate for that.
986 *
987 * If this process is currently in __generic_file_write_iter() against
988 * this page's queue, we can perform writeback even if that
989 * will block.
990 *
991 * If the page is swapcache, write it back even if that would
992 * block, for some throttling. This happens by accident, because
993 * swap_backing_dev_info is bust: it doesn't reflect the
994 * congestion state of the swapdevs. Easy to fix, if needed.
995 */
996 if (!is_page_cache_freeable(page))
997 return PAGE_KEEP;
998 if (!mapping) {
999 /*
1000 * Some data journaling orphaned pages can have
1001 * page->mapping == NULL while being dirty with clean buffers.
1002 */
1003 if (page_has_private(page)) {
1004 if (try_to_free_buffers(page)) {
1005 ClearPageDirty(page);
1006 pr_info("%s: orphaned page\n", __func__);
1007 return PAGE_CLEAN;
1008 }
1009 }
1010 return PAGE_KEEP;
1011 }
1012 if (mapping->a_ops->writepage == NULL)
1013 return PAGE_ACTIVATE;
1014 if (!may_write_to_inode(mapping->host))
1015 return PAGE_KEEP;
1016
1017 if (clear_page_dirty_for_io(page)) {
1018 int res;
1019 struct writeback_control wbc = {
1020 .sync_mode = WB_SYNC_NONE,
1021 .nr_to_write = SWAP_CLUSTER_MAX,
1022 .range_start = 0,
1023 .range_end = LLONG_MAX,
1024 .for_reclaim = 1,
1025 };
1026
1027 SetPageReclaim(page);
1028 res = mapping->a_ops->writepage(page, &wbc);
1029 if (res < 0)
1030 handle_write_error(mapping, page, res);
1031 if (res == AOP_WRITEPAGE_ACTIVATE) {
1032 ClearPageReclaim(page);
1033 return PAGE_ACTIVATE;
1034 }
1035
1036 if (!PageWriteback(page)) {
1037 /* synchronous write or broken a_ops? */
1038 ClearPageReclaim(page);
1039 }
1040 trace_mm_vmscan_writepage(page);
1041 inc_node_page_state(page, NR_VMSCAN_WRITE);
1042 return PAGE_SUCCESS;
1043 }
1044
1045 return PAGE_CLEAN;
1046}
1047
1048/*
1049 * Same as remove_mapping, but if the page is removed from the mapping, it
1050 * gets returned with a refcount of 0.
1051 */
1052static int __remove_mapping(struct address_space *mapping, struct page *page,
1053 bool reclaimed, struct mem_cgroup *target_memcg)
1054{
1055 unsigned long flags;
1056 int refcount;
1057 void *shadow = NULL;
1058
1059 BUG_ON(!PageLocked(page));
1060 BUG_ON(mapping != page_mapping(page));
1061
1062 xa_lock_irqsave(&mapping->i_pages, flags);
1063 /*
1064 * The non racy check for a busy page.
1065 *
1066 * Must be careful with the order of the tests. When someone has
1067 * a ref to the page, it may be possible that they dirty it then
1068 * drop the reference. So if PageDirty is tested before page_count
1069 * here, then the following race may occur:
1070 *
1071 * get_user_pages(&page);
1072 * [user mapping goes away]
1073 * write_to(page);
1074 * !PageDirty(page) [good]
1075 * SetPageDirty(page);
1076 * put_page(page);
1077 * !page_count(page) [good, discard it]
1078 *
1079 * [oops, our write_to data is lost]
1080 *
1081 * Reversing the order of the tests ensures such a situation cannot
1082 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1083 * load is not satisfied before that of page->_refcount.
1084 *
1085 * Note that if SetPageDirty is always performed via set_page_dirty,
1086 * and thus under the i_pages lock, then this ordering is not required.
1087 */
1088 refcount = 1 + compound_nr(page);
1089 if (!page_ref_freeze(page, refcount))
1090 goto cannot_free;
1091 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1092 if (unlikely(PageDirty(page))) {
1093 page_ref_unfreeze(page, refcount);
1094 goto cannot_free;
1095 }
1096
1097 if (PageSwapCache(page)) {
1098 swp_entry_t swap = { .val = page_private(page) };
1099 mem_cgroup_swapout(page, swap);
1100 if (reclaimed && !mapping_exiting(mapping))
1101 shadow = workingset_eviction(page, target_memcg);
1102 __delete_from_swap_cache(page, swap, shadow);
1103 xa_unlock_irqrestore(&mapping->i_pages, flags);
1104 put_swap_page(page, swap);
1105 } else {
1106 void (*freepage)(struct page *);
1107
1108 freepage = mapping->a_ops->freepage;
1109 /*
1110 * Remember a shadow entry for reclaimed file cache in
1111 * order to detect refaults, thus thrashing, later on.
1112 *
1113 * But don't store shadows in an address space that is
1114 * already exiting. This is not just an optimization,
1115 * inode reclaim needs to empty out the radix tree or
1116 * the nodes are lost. Don't plant shadows behind its
1117 * back.
1118 *
1119 * We also don't store shadows for DAX mappings because the
1120 * only page cache pages found in these are zero pages
1121 * covering holes, and because we don't want to mix DAX
1122 * exceptional entries and shadow exceptional entries in the
1123 * same address_space.
1124 */
1125 if (reclaimed && page_is_file_lru(page) &&
1126 !mapping_exiting(mapping) && !dax_mapping(mapping))
1127 shadow = workingset_eviction(page, target_memcg);
1128 __delete_from_page_cache(page, shadow);
1129 xa_unlock_irqrestore(&mapping->i_pages, flags);
1130
1131 if (freepage != NULL)
1132 freepage(page);
1133 }
1134
1135 return 1;
1136
1137cannot_free:
1138 xa_unlock_irqrestore(&mapping->i_pages, flags);
1139 return 0;
1140}
1141
1142/*
1143 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1144 * someone else has a ref on the page, abort and return 0. If it was
1145 * successfully detached, return 1. Assumes the caller has a single ref on
1146 * this page.
1147 */
1148int remove_mapping(struct address_space *mapping, struct page *page)
1149{
1150 if (__remove_mapping(mapping, page, false, NULL)) {
1151 /*
1152 * Unfreezing the refcount with 1 rather than 2 effectively
1153 * drops the pagecache ref for us without requiring another
1154 * atomic operation.
1155 */
1156 page_ref_unfreeze(page, 1);
1157 return 1;
1158 }
1159 return 0;
1160}
1161
1162/**
1163 * putback_lru_page - put previously isolated page onto appropriate LRU list
1164 * @page: page to be put back to appropriate lru list
1165 *
1166 * Add previously isolated @page to appropriate LRU list.
1167 * Page may still be unevictable for other reasons.
1168 *
1169 * lru_lock must not be held, interrupts must be enabled.
1170 */
1171void putback_lru_page(struct page *page)
1172{
1173 lru_cache_add(page);
1174 put_page(page); /* drop ref from isolate */
1175}
1176
1177enum page_references {
1178 PAGEREF_RECLAIM,
1179 PAGEREF_RECLAIM_CLEAN,
1180 PAGEREF_KEEP,
1181 PAGEREF_ACTIVATE,
1182};
1183
1184static enum page_references page_check_references(struct page *page,
1185 struct scan_control *sc)
1186{
1187 int referenced_ptes, referenced_page;
1188 unsigned long vm_flags;
1189
1190 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1191 &vm_flags);
1192 referenced_page = TestClearPageReferenced(page);
1193
1194 /*
1195 * Mlock lost the isolation race with us. Let try_to_unmap()
1196 * move the page to the unevictable list.
1197 */
1198 if (vm_flags & VM_LOCKED)
1199 return PAGEREF_RECLAIM;
1200
1201 if (referenced_ptes) {
1202 /*
1203 * All mapped pages start out with page table
1204 * references from the instantiating fault, so we need
1205 * to look twice if a mapped file page is used more
1206 * than once.
1207 *
1208 * Mark it and spare it for another trip around the
1209 * inactive list. Another page table reference will
1210 * lead to its activation.
1211 *
1212 * Note: the mark is set for activated pages as well
1213 * so that recently deactivated but used pages are
1214 * quickly recovered.
1215 */
1216 SetPageReferenced(page);
1217
1218 if (referenced_page || referenced_ptes > 1)
1219 return PAGEREF_ACTIVATE;
1220
1221 /*
1222 * Activate file-backed executable pages after first usage.
1223 */
1224 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1225 return PAGEREF_ACTIVATE;
1226
1227 return PAGEREF_KEEP;
1228 }
1229
1230 /* Reclaim if clean, defer dirty pages to writeback */
1231 if (referenced_page && !PageSwapBacked(page))
1232 return PAGEREF_RECLAIM_CLEAN;
1233
1234 return PAGEREF_RECLAIM;
1235}
1236
1237/* Check if a page is dirty or under writeback */
1238static void page_check_dirty_writeback(struct page *page,
1239 bool *dirty, bool *writeback)
1240{
1241 struct address_space *mapping;
1242
1243 /*
1244 * Anonymous pages are not handled by flushers and must be written
1245 * from reclaim context. Do not stall reclaim based on them
1246 */
1247 if (!page_is_file_lru(page) ||
1248 (PageAnon(page) && !PageSwapBacked(page))) {
1249 *dirty = false;
1250 *writeback = false;
1251 return;
1252 }
1253
1254 /* By default assume that the page flags are accurate */
1255 *dirty = PageDirty(page);
1256 *writeback = PageWriteback(page);
1257
1258 /* Verify dirty/writeback state if the filesystem supports it */
1259 if (!page_has_private(page))
1260 return;
1261
1262 mapping = page_mapping(page);
1263 if (mapping && mapping->a_ops->is_dirty_writeback)
1264 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1265}
1266
1267/*
1268 * shrink_page_list() returns the number of reclaimed pages
1269 */
1270static unsigned int shrink_page_list(struct list_head *page_list,
1271 struct pglist_data *pgdat,
1272 struct scan_control *sc,
1273 struct reclaim_stat *stat,
1274 bool ignore_references)
1275{
1276 LIST_HEAD(ret_pages);
1277 LIST_HEAD(free_pages);
1278 unsigned int nr_reclaimed = 0;
1279 unsigned int pgactivate = 0;
1280
1281 memset(stat, 0, sizeof(*stat));
1282 cond_resched();
1283
1284 while (!list_empty(page_list)) {
1285 struct address_space *mapping;
1286 struct page *page;
1287 enum page_references references = PAGEREF_RECLAIM;
1288 bool dirty, writeback, may_enter_fs;
1289 unsigned int nr_pages;
1290
1291 cond_resched();
1292
1293 page = lru_to_page(page_list);
1294 list_del(&page->lru);
1295
1296 if (!trylock_page(page))
1297 goto keep;
1298
1299 VM_BUG_ON_PAGE(PageActive(page), page);
1300
1301 nr_pages = compound_nr(page);
1302
1303 /* Account the number of base pages even though THP */
1304 sc->nr_scanned += nr_pages;
1305
1306 if (unlikely(!page_evictable(page)))
1307 goto activate_locked;
1308
1309 if (!sc->may_unmap && page_mapped(page))
1310 goto keep_locked;
1311
1312 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1313 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1314
1315 /*
1316 * The number of dirty pages determines if a node is marked
1317 * reclaim_congested which affects wait_iff_congested. kswapd
1318 * will stall and start writing pages if the tail of the LRU
1319 * is all dirty unqueued pages.
1320 */
1321 page_check_dirty_writeback(page, &dirty, &writeback);
1322 if (dirty || writeback)
1323 stat->nr_dirty++;
1324
1325 if (dirty && !writeback)
1326 stat->nr_unqueued_dirty++;
1327
1328 /*
1329 * Treat this page as congested if the underlying BDI is or if
1330 * pages are cycling through the LRU so quickly that the
1331 * pages marked for immediate reclaim are making it to the
1332 * end of the LRU a second time.
1333 */
1334 mapping = page_mapping(page);
1335 if (((dirty || writeback) && mapping &&
1336 inode_write_congested(mapping->host)) ||
1337 (writeback && PageReclaim(page)))
1338 stat->nr_congested++;
1339
1340 /*
1341 * If a page at the tail of the LRU is under writeback, there
1342 * are three cases to consider.
1343 *
1344 * 1) If reclaim is encountering an excessive number of pages
1345 * under writeback and this page is both under writeback and
1346 * PageReclaim then it indicates that pages are being queued
1347 * for IO but are being recycled through the LRU before the
1348 * IO can complete. Waiting on the page itself risks an
1349 * indefinite stall if it is impossible to writeback the
1350 * page due to IO error or disconnected storage so instead
1351 * note that the LRU is being scanned too quickly and the
1352 * caller can stall after page list has been processed.
1353 *
1354 * 2) Global or new memcg reclaim encounters a page that is
1355 * not marked for immediate reclaim, or the caller does not
1356 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1357 * not to fs). In this case mark the page for immediate
1358 * reclaim and continue scanning.
1359 *
1360 * Require may_enter_fs because we would wait on fs, which
1361 * may not have submitted IO yet. And the loop driver might
1362 * enter reclaim, and deadlock if it waits on a page for
1363 * which it is needed to do the write (loop masks off
1364 * __GFP_IO|__GFP_FS for this reason); but more thought
1365 * would probably show more reasons.
1366 *
1367 * 3) Legacy memcg encounters a page that is already marked
1368 * PageReclaim. memcg does not have any dirty pages
1369 * throttling so we could easily OOM just because too many
1370 * pages are in writeback and there is nothing else to
1371 * reclaim. Wait for the writeback to complete.
1372 *
1373 * In cases 1) and 2) we activate the pages to get them out of
1374 * the way while we continue scanning for clean pages on the
1375 * inactive list and refilling from the active list. The
1376 * observation here is that waiting for disk writes is more
1377 * expensive than potentially causing reloads down the line.
1378 * Since they're marked for immediate reclaim, they won't put
1379 * memory pressure on the cache working set any longer than it
1380 * takes to write them to disk.
1381 */
1382 if (PageWriteback(page)) {
1383 /* Case 1 above */
1384 if (current_is_kswapd() &&
1385 PageReclaim(page) &&
1386 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1387 stat->nr_immediate++;
1388 goto activate_locked;
1389
1390 /* Case 2 above */
1391 } else if (writeback_throttling_sane(sc) ||
1392 !PageReclaim(page) || !may_enter_fs) {
1393 /*
1394 * This is slightly racy - end_page_writeback()
1395 * might have just cleared PageReclaim, then
1396 * setting PageReclaim here end up interpreted
1397 * as PageReadahead - but that does not matter
1398 * enough to care. What we do want is for this
1399 * page to have PageReclaim set next time memcg
1400 * reclaim reaches the tests above, so it will
1401 * then wait_on_page_writeback() to avoid OOM;
1402 * and it's also appropriate in global reclaim.
1403 */
1404 SetPageReclaim(page);
1405 stat->nr_writeback++;
1406 goto activate_locked;
1407
1408 /* Case 3 above */
1409 } else {
1410 unlock_page(page);
1411 wait_on_page_writeback(page);
1412 /* then go back and try same page again */
1413 list_add_tail(&page->lru, page_list);
1414 continue;
1415 }
1416 }
1417
1418 if (!ignore_references)
1419 references = page_check_references(page, sc);
1420
1421 switch (references) {
1422 case PAGEREF_ACTIVATE:
1423 goto activate_locked;
1424 case PAGEREF_KEEP:
1425 stat->nr_ref_keep += nr_pages;
1426 goto keep_locked;
1427 case PAGEREF_RECLAIM:
1428 case PAGEREF_RECLAIM_CLEAN:
1429 ; /* try to reclaim the page below */
1430 }
1431
1432 /*
1433 * Anonymous process memory has backing store?
1434 * Try to allocate it some swap space here.
1435 * Lazyfree page could be freed directly
1436 */
1437 if (PageAnon(page) && PageSwapBacked(page)) {
1438 if (!PageSwapCache(page)) {
1439 if (!(sc->gfp_mask & __GFP_IO))
1440 goto keep_locked;
1441 if (page_maybe_dma_pinned(page))
1442 goto keep_locked;
1443 if (PageTransHuge(page)) {
1444 /* cannot split THP, skip it */
1445 if (!can_split_huge_page(page, NULL))
1446 goto activate_locked;
1447 /*
1448 * Split pages without a PMD map right
1449 * away. Chances are some or all of the
1450 * tail pages can be freed without IO.
1451 */
1452 if (!compound_mapcount(page) &&
1453 split_huge_page_to_list(page,
1454 page_list))
1455 goto activate_locked;
1456 }
1457 if (!add_to_swap(page)) {
1458 if (!PageTransHuge(page))
1459 goto activate_locked_split;
1460 /* Fallback to swap normal pages */
1461 if (split_huge_page_to_list(page,
1462 page_list))
1463 goto activate_locked;
1464#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1465 count_vm_event(THP_SWPOUT_FALLBACK);
1466#endif
1467 if (!add_to_swap(page))
1468 goto activate_locked_split;
1469 }
1470
1471 may_enter_fs = true;
1472
1473 /* Adding to swap updated mapping */
1474 mapping = page_mapping(page);
1475 }
1476 } else if (unlikely(PageTransHuge(page))) {
1477 /* Split file THP */
1478 if (split_huge_page_to_list(page, page_list))
1479 goto keep_locked;
1480 }
1481
1482 /*
1483 * THP may get split above, need minus tail pages and update
1484 * nr_pages to avoid accounting tail pages twice.
1485 *
1486 * The tail pages that are added into swap cache successfully
1487 * reach here.
1488 */
1489 if ((nr_pages > 1) && !PageTransHuge(page)) {
1490 sc->nr_scanned -= (nr_pages - 1);
1491 nr_pages = 1;
1492 }
1493
1494 /*
1495 * The page is mapped into the page tables of one or more
1496 * processes. Try to unmap it here.
1497 */
1498 if (page_mapped(page)) {
1499 enum ttu_flags flags = TTU_BATCH_FLUSH;
1500 bool was_swapbacked = PageSwapBacked(page);
1501
1502 if (unlikely(PageTransHuge(page)))
1503 flags |= TTU_SPLIT_HUGE_PMD;
1504
1505 try_to_unmap(page, flags);
1506 if (page_mapped(page)) {
1507 stat->nr_unmap_fail += nr_pages;
1508 if (!was_swapbacked && PageSwapBacked(page))
1509 stat->nr_lazyfree_fail += nr_pages;
1510 goto activate_locked;
1511 }
1512 }
1513
1514 if (PageDirty(page)) {
1515 /*
1516 * Only kswapd can writeback filesystem pages
1517 * to avoid risk of stack overflow. But avoid
1518 * injecting inefficient single-page IO into
1519 * flusher writeback as much as possible: only
1520 * write pages when we've encountered many
1521 * dirty pages, and when we've already scanned
1522 * the rest of the LRU for clean pages and see
1523 * the same dirty pages again (PageReclaim).
1524 */
1525 if (page_is_file_lru(page) &&
1526 (!current_is_kswapd() || !PageReclaim(page) ||
1527 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1528 /*
1529 * Immediately reclaim when written back.
1530 * Similar in principal to deactivate_page()
1531 * except we already have the page isolated
1532 * and know it's dirty
1533 */
1534 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1535 SetPageReclaim(page);
1536
1537 goto activate_locked;
1538 }
1539
1540 if (references == PAGEREF_RECLAIM_CLEAN)
1541 goto keep_locked;
1542 if (!may_enter_fs)
1543 goto keep_locked;
1544 if (!sc->may_writepage)
1545 goto keep_locked;
1546
1547 /*
1548 * Page is dirty. Flush the TLB if a writable entry
1549 * potentially exists to avoid CPU writes after IO
1550 * starts and then write it out here.
1551 */
1552 try_to_unmap_flush_dirty();
1553 switch (pageout(page, mapping)) {
1554 case PAGE_KEEP:
1555 goto keep_locked;
1556 case PAGE_ACTIVATE:
1557 goto activate_locked;
1558 case PAGE_SUCCESS:
1559 stat->nr_pageout += thp_nr_pages(page);
1560
1561 if (PageWriteback(page))
1562 goto keep;
1563 if (PageDirty(page))
1564 goto keep;
1565
1566 /*
1567 * A synchronous write - probably a ramdisk. Go
1568 * ahead and try to reclaim the page.
1569 */
1570 if (!trylock_page(page))
1571 goto keep;
1572 if (PageDirty(page) || PageWriteback(page))
1573 goto keep_locked;
1574 mapping = page_mapping(page);
1575 fallthrough;
1576 case PAGE_CLEAN:
1577 ; /* try to free the page below */
1578 }
1579 }
1580
1581 /*
1582 * If the page has buffers, try to free the buffer mappings
1583 * associated with this page. If we succeed we try to free
1584 * the page as well.
1585 *
1586 * We do this even if the page is PageDirty().
1587 * try_to_release_page() does not perform I/O, but it is
1588 * possible for a page to have PageDirty set, but it is actually
1589 * clean (all its buffers are clean). This happens if the
1590 * buffers were written out directly, with submit_bh(). ext3
1591 * will do this, as well as the blockdev mapping.
1592 * try_to_release_page() will discover that cleanness and will
1593 * drop the buffers and mark the page clean - it can be freed.
1594 *
1595 * Rarely, pages can have buffers and no ->mapping. These are
1596 * the pages which were not successfully invalidated in
1597 * truncate_cleanup_page(). We try to drop those buffers here
1598 * and if that worked, and the page is no longer mapped into
1599 * process address space (page_count == 1) it can be freed.
1600 * Otherwise, leave the page on the LRU so it is swappable.
1601 */
1602 if (page_has_private(page)) {
1603 if (!try_to_release_page(page, sc->gfp_mask))
1604 goto activate_locked;
1605 if (!mapping && page_count(page) == 1) {
1606 unlock_page(page);
1607 if (put_page_testzero(page))
1608 goto free_it;
1609 else {
1610 /*
1611 * rare race with speculative reference.
1612 * the speculative reference will free
1613 * this page shortly, so we may
1614 * increment nr_reclaimed here (and
1615 * leave it off the LRU).
1616 */
1617 nr_reclaimed++;
1618 continue;
1619 }
1620 }
1621 }
1622
1623 if (PageAnon(page) && !PageSwapBacked(page)) {
1624 /* follow __remove_mapping for reference */
1625 if (!page_ref_freeze(page, 1))
1626 goto keep_locked;
1627 if (PageDirty(page)) {
1628 page_ref_unfreeze(page, 1);
1629 goto keep_locked;
1630 }
1631
1632 count_vm_event(PGLAZYFREED);
1633 count_memcg_page_event(page, PGLAZYFREED);
1634 } else if (!mapping || !__remove_mapping(mapping, page, true,
1635 sc->target_mem_cgroup))
1636 goto keep_locked;
1637
1638 unlock_page(page);
1639free_it:
1640 /*
1641 * THP may get swapped out in a whole, need account
1642 * all base pages.
1643 */
1644 nr_reclaimed += nr_pages;
1645
1646 /*
1647 * Is there need to periodically free_page_list? It would
1648 * appear not as the counts should be low
1649 */
1650 if (unlikely(PageTransHuge(page)))
1651 destroy_compound_page(page);
1652 else
1653 list_add(&page->lru, &free_pages);
1654 continue;
1655
1656activate_locked_split:
1657 /*
1658 * The tail pages that are failed to add into swap cache
1659 * reach here. Fixup nr_scanned and nr_pages.
1660 */
1661 if (nr_pages > 1) {
1662 sc->nr_scanned -= (nr_pages - 1);
1663 nr_pages = 1;
1664 }
1665activate_locked:
1666 /* Not a candidate for swapping, so reclaim swap space. */
1667 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1668 PageMlocked(page)))
1669 try_to_free_swap(page);
1670 VM_BUG_ON_PAGE(PageActive(page), page);
1671 if (!PageMlocked(page)) {
1672 int type = page_is_file_lru(page);
1673 SetPageActive(page);
1674 stat->nr_activate[type] += nr_pages;
1675 count_memcg_page_event(page, PGACTIVATE);
1676 }
1677keep_locked:
1678 unlock_page(page);
1679keep:
1680 list_add(&page->lru, &ret_pages);
1681 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1682 }
1683
1684 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1685
1686 mem_cgroup_uncharge_list(&free_pages);
1687 try_to_unmap_flush();
1688 free_unref_page_list(&free_pages);
1689
1690 list_splice(&ret_pages, page_list);
1691 count_vm_events(PGACTIVATE, pgactivate);
1692
1693 return nr_reclaimed;
1694}
1695
1696unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1697 struct list_head *page_list)
1698{
1699 struct scan_control sc = {
1700 .gfp_mask = GFP_KERNEL,
1701 .priority = DEF_PRIORITY,
1702 .may_unmap = 1,
1703 };
1704 struct reclaim_stat stat;
1705 unsigned int nr_reclaimed;
1706 struct page *page, *next;
1707 LIST_HEAD(clean_pages);
1708 unsigned int noreclaim_flag;
1709
1710 list_for_each_entry_safe(page, next, page_list, lru) {
1711 if (!PageHuge(page) && page_is_file_lru(page) &&
1712 !PageDirty(page) && !__PageMovable(page) &&
1713 !PageUnevictable(page)) {
1714 ClearPageActive(page);
1715 list_move(&page->lru, &clean_pages);
1716 }
1717 }
1718
1719 /*
1720 * We should be safe here since we are only dealing with file pages and
1721 * we are not kswapd and therefore cannot write dirty file pages. But
1722 * call memalloc_noreclaim_save() anyway, just in case these conditions
1723 * change in the future.
1724 */
1725 noreclaim_flag = memalloc_noreclaim_save();
1726 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1727 &stat, true);
1728 memalloc_noreclaim_restore(noreclaim_flag);
1729
1730 list_splice(&clean_pages, page_list);
1731 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1732 -(long)nr_reclaimed);
1733 /*
1734 * Since lazyfree pages are isolated from file LRU from the beginning,
1735 * they will rotate back to anonymous LRU in the end if it failed to
1736 * discard so isolated count will be mismatched.
1737 * Compensate the isolated count for both LRU lists.
1738 */
1739 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1740 stat.nr_lazyfree_fail);
1741 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1742 -(long)stat.nr_lazyfree_fail);
1743 return nr_reclaimed;
1744}
1745
1746/*
1747 * Attempt to remove the specified page from its LRU. Only take this page
1748 * if it is of the appropriate PageActive status. Pages which are being
1749 * freed elsewhere are also ignored.
1750 *
1751 * page: page to consider
1752 * mode: one of the LRU isolation modes defined above
1753 *
1754 * returns true on success, false on failure.
1755 */
1756bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1757{
1758 /* Only take pages on the LRU. */
1759 if (!PageLRU(page))
1760 return false;
1761
1762 /* Compaction should not handle unevictable pages but CMA can do so */
1763 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1764 return false;
1765
1766 /*
1767 * To minimise LRU disruption, the caller can indicate that it only
1768 * wants to isolate pages it will be able to operate on without
1769 * blocking - clean pages for the most part.
1770 *
1771 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1772 * that it is possible to migrate without blocking
1773 */
1774 if (mode & ISOLATE_ASYNC_MIGRATE) {
1775 /* All the caller can do on PageWriteback is block */
1776 if (PageWriteback(page))
1777 return false;
1778
1779 if (PageDirty(page)) {
1780 struct address_space *mapping;
1781 bool migrate_dirty;
1782
1783 /*
1784 * Only pages without mappings or that have a
1785 * ->migratepage callback are possible to migrate
1786 * without blocking. However, we can be racing with
1787 * truncation so it's necessary to lock the page
1788 * to stabilise the mapping as truncation holds
1789 * the page lock until after the page is removed
1790 * from the page cache.
1791 */
1792 if (!trylock_page(page))
1793 return false;
1794
1795 mapping = page_mapping(page);
1796 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1797 unlock_page(page);
1798 if (!migrate_dirty)
1799 return false;
1800 }
1801 }
1802
1803 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1804 return false;
1805
1806 return true;
1807}
1808
1809/*
1810 * Update LRU sizes after isolating pages. The LRU size updates must
1811 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1812 */
1813static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1814 enum lru_list lru, unsigned long *nr_zone_taken)
1815{
1816 int zid;
1817
1818 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1819 if (!nr_zone_taken[zid])
1820 continue;
1821
1822 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1823 }
1824
1825}
1826
1827/*
1828 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1829 *
1830 * lruvec->lru_lock is heavily contended. Some of the functions that
1831 * shrink the lists perform better by taking out a batch of pages
1832 * and working on them outside the LRU lock.
1833 *
1834 * For pagecache intensive workloads, this function is the hottest
1835 * spot in the kernel (apart from copy_*_user functions).
1836 *
1837 * Lru_lock must be held before calling this function.
1838 *
1839 * @nr_to_scan: The number of eligible pages to look through on the list.
1840 * @lruvec: The LRU vector to pull pages from.
1841 * @dst: The temp list to put pages on to.
1842 * @nr_scanned: The number of pages that were scanned.
1843 * @sc: The scan_control struct for this reclaim session
1844 * @lru: LRU list id for isolating
1845 *
1846 * returns how many pages were moved onto *@dst.
1847 */
1848static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1849 struct lruvec *lruvec, struct list_head *dst,
1850 unsigned long *nr_scanned, struct scan_control *sc,
1851 enum lru_list lru)
1852{
1853 struct list_head *src = &lruvec->lists[lru];
1854 unsigned long nr_taken = 0;
1855 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1856 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1857 unsigned long skipped = 0;
1858 unsigned long scan, total_scan, nr_pages;
1859 LIST_HEAD(pages_skipped);
1860 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1861
1862 total_scan = 0;
1863 scan = 0;
1864 while (scan < nr_to_scan && !list_empty(src)) {
1865 struct page *page;
1866
1867 page = lru_to_page(src);
1868 prefetchw_prev_lru_page(page, src, flags);
1869
1870 nr_pages = compound_nr(page);
1871 total_scan += nr_pages;
1872
1873 if (page_zonenum(page) > sc->reclaim_idx) {
1874 list_move(&page->lru, &pages_skipped);
1875 nr_skipped[page_zonenum(page)] += nr_pages;
1876 continue;
1877 }
1878
1879 /*
1880 * Do not count skipped pages because that makes the function
1881 * return with no isolated pages if the LRU mostly contains
1882 * ineligible pages. This causes the VM to not reclaim any
1883 * pages, triggering a premature OOM.
1884 *
1885 * Account all tail pages of THP. This would not cause
1886 * premature OOM since __isolate_lru_page() returns -EBUSY
1887 * only when the page is being freed somewhere else.
1888 */
1889 scan += nr_pages;
1890 if (!__isolate_lru_page_prepare(page, mode)) {
1891 /* It is being freed elsewhere */
1892 list_move(&page->lru, src);
1893 continue;
1894 }
1895 /*
1896 * Be careful not to clear PageLRU until after we're
1897 * sure the page is not being freed elsewhere -- the
1898 * page release code relies on it.
1899 */
1900 if (unlikely(!get_page_unless_zero(page))) {
1901 list_move(&page->lru, src);
1902 continue;
1903 }
1904
1905 if (!TestClearPageLRU(page)) {
1906 /* Another thread is already isolating this page */
1907 put_page(page);
1908 list_move(&page->lru, src);
1909 continue;
1910 }
1911
1912 nr_taken += nr_pages;
1913 nr_zone_taken[page_zonenum(page)] += nr_pages;
1914 list_move(&page->lru, dst);
1915 }
1916
1917 /*
1918 * Splice any skipped pages to the start of the LRU list. Note that
1919 * this disrupts the LRU order when reclaiming for lower zones but
1920 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1921 * scanning would soon rescan the same pages to skip and put the
1922 * system at risk of premature OOM.
1923 */
1924 if (!list_empty(&pages_skipped)) {
1925 int zid;
1926
1927 list_splice(&pages_skipped, src);
1928 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1929 if (!nr_skipped[zid])
1930 continue;
1931
1932 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1933 skipped += nr_skipped[zid];
1934 }
1935 }
1936 *nr_scanned = total_scan;
1937 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1938 total_scan, skipped, nr_taken, mode, lru);
1939 update_lru_sizes(lruvec, lru, nr_zone_taken);
1940 return nr_taken;
1941}
1942
1943/**
1944 * isolate_lru_page - tries to isolate a page from its LRU list
1945 * @page: page to isolate from its LRU list
1946 *
1947 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1948 * vmstat statistic corresponding to whatever LRU list the page was on.
1949 *
1950 * Returns 0 if the page was removed from an LRU list.
1951 * Returns -EBUSY if the page was not on an LRU list.
1952 *
1953 * The returned page will have PageLRU() cleared. If it was found on
1954 * the active list, it will have PageActive set. If it was found on
1955 * the unevictable list, it will have the PageUnevictable bit set. That flag
1956 * may need to be cleared by the caller before letting the page go.
1957 *
1958 * The vmstat statistic corresponding to the list on which the page was
1959 * found will be decremented.
1960 *
1961 * Restrictions:
1962 *
1963 * (1) Must be called with an elevated refcount on the page. This is a
1964 * fundamental difference from isolate_lru_pages (which is called
1965 * without a stable reference).
1966 * (2) the lru_lock must not be held.
1967 * (3) interrupts must be enabled.
1968 */
1969int isolate_lru_page(struct page *page)
1970{
1971 int ret = -EBUSY;
1972
1973 VM_BUG_ON_PAGE(!page_count(page), page);
1974 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1975
1976 if (TestClearPageLRU(page)) {
1977 struct lruvec *lruvec;
1978
1979 get_page(page);
1980 lruvec = lock_page_lruvec_irq(page);
1981 del_page_from_lru_list(page, lruvec);
1982 unlock_page_lruvec_irq(lruvec);
1983 ret = 0;
1984 }
1985
1986 return ret;
1987}
1988
1989/*
1990 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1991 * then get rescheduled. When there are massive number of tasks doing page
1992 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1993 * the LRU list will go small and be scanned faster than necessary, leading to
1994 * unnecessary swapping, thrashing and OOM.
1995 */
1996static int too_many_isolated(struct pglist_data *pgdat, int file,
1997 struct scan_control *sc)
1998{
1999 unsigned long inactive, isolated;
2000
2001 if (current_is_kswapd())
2002 return 0;
2003
2004 if (!writeback_throttling_sane(sc))
2005 return 0;
2006
2007 if (file) {
2008 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2009 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2010 } else {
2011 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2012 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2013 }
2014
2015 /*
2016 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2017 * won't get blocked by normal direct-reclaimers, forming a circular
2018 * deadlock.
2019 */
2020 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2021 inactive >>= 3;
2022
2023 return isolated > inactive;
2024}
2025
2026/*
2027 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2028 * On return, @list is reused as a list of pages to be freed by the caller.
2029 *
2030 * Returns the number of pages moved to the given lruvec.
2031 */
2032static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2033 struct list_head *list)
2034{
2035 int nr_pages, nr_moved = 0;
2036 LIST_HEAD(pages_to_free);
2037 struct page *page;
2038
2039 while (!list_empty(list)) {
2040 page = lru_to_page(list);
2041 VM_BUG_ON_PAGE(PageLRU(page), page);
2042 list_del(&page->lru);
2043 if (unlikely(!page_evictable(page))) {
2044 spin_unlock_irq(&lruvec->lru_lock);
2045 putback_lru_page(page);
2046 spin_lock_irq(&lruvec->lru_lock);
2047 continue;
2048 }
2049
2050 /*
2051 * The SetPageLRU needs to be kept here for list integrity.
2052 * Otherwise:
2053 * #0 move_pages_to_lru #1 release_pages
2054 * if !put_page_testzero
2055 * if (put_page_testzero())
2056 * !PageLRU //skip lru_lock
2057 * SetPageLRU()
2058 * list_add(&page->lru,)
2059 * list_add(&page->lru,)
2060 */
2061 SetPageLRU(page);
2062
2063 if (unlikely(put_page_testzero(page))) {
2064 __clear_page_lru_flags(page);
2065
2066 if (unlikely(PageCompound(page))) {
2067 spin_unlock_irq(&lruvec->lru_lock);
2068 destroy_compound_page(page);
2069 spin_lock_irq(&lruvec->lru_lock);
2070 } else
2071 list_add(&page->lru, &pages_to_free);
2072
2073 continue;
2074 }
2075
2076 /*
2077 * All pages were isolated from the same lruvec (and isolation
2078 * inhibits memcg migration).
2079 */
2080 VM_BUG_ON_PAGE(!page_matches_lruvec(page, lruvec), page);
2081 add_page_to_lru_list(page, lruvec);
2082 nr_pages = thp_nr_pages(page);
2083 nr_moved += nr_pages;
2084 if (PageActive(page))
2085 workingset_age_nonresident(lruvec, nr_pages);
2086 }
2087
2088 /*
2089 * To save our caller's stack, now use input list for pages to free.
2090 */
2091 list_splice(&pages_to_free, list);
2092
2093 return nr_moved;
2094}
2095
2096/*
2097 * If a kernel thread (such as nfsd for loop-back mounts) services
2098 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2099 * In that case we should only throttle if the backing device it is
2100 * writing to is congested. In other cases it is safe to throttle.
2101 */
2102static int current_may_throttle(void)
2103{
2104 return !(current->flags & PF_LOCAL_THROTTLE) ||
2105 current->backing_dev_info == NULL ||
2106 bdi_write_congested(current->backing_dev_info);
2107}
2108
2109/*
2110 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2111 * of reclaimed pages
2112 */
2113static unsigned long
2114shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2115 struct scan_control *sc, enum lru_list lru)
2116{
2117 LIST_HEAD(page_list);
2118 unsigned long nr_scanned;
2119 unsigned int nr_reclaimed = 0;
2120 unsigned long nr_taken;
2121 struct reclaim_stat stat;
2122 bool file = is_file_lru(lru);
2123 enum vm_event_item item;
2124 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2125 bool stalled = false;
2126
2127 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2128 if (stalled)
2129 return 0;
2130
2131 /* wait a bit for the reclaimer. */
2132 msleep(100);
2133 stalled = true;
2134
2135 /* We are about to die and free our memory. Return now. */
2136 if (fatal_signal_pending(current))
2137 return SWAP_CLUSTER_MAX;
2138 }
2139
2140 lru_add_drain();
2141
2142 spin_lock_irq(&lruvec->lru_lock);
2143
2144 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2145 &nr_scanned, sc, lru);
2146
2147 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2148 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2149 if (!cgroup_reclaim(sc))
2150 __count_vm_events(item, nr_scanned);
2151 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2152 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2153
2154 spin_unlock_irq(&lruvec->lru_lock);
2155
2156 if (nr_taken == 0)
2157 return 0;
2158
2159 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2160
2161 spin_lock_irq(&lruvec->lru_lock);
2162 move_pages_to_lru(lruvec, &page_list);
2163
2164 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2165 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2166 if (!cgroup_reclaim(sc))
2167 __count_vm_events(item, nr_reclaimed);
2168 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2169 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2170 spin_unlock_irq(&lruvec->lru_lock);
2171
2172 lru_note_cost(lruvec, file, stat.nr_pageout);
2173 mem_cgroup_uncharge_list(&page_list);
2174 free_unref_page_list(&page_list);
2175
2176 /*
2177 * If dirty pages are scanned that are not queued for IO, it
2178 * implies that flushers are not doing their job. This can
2179 * happen when memory pressure pushes dirty pages to the end of
2180 * the LRU before the dirty limits are breached and the dirty
2181 * data has expired. It can also happen when the proportion of
2182 * dirty pages grows not through writes but through memory
2183 * pressure reclaiming all the clean cache. And in some cases,
2184 * the flushers simply cannot keep up with the allocation
2185 * rate. Nudge the flusher threads in case they are asleep.
2186 */
2187 if (stat.nr_unqueued_dirty == nr_taken)
2188 wakeup_flusher_threads(WB_REASON_VMSCAN);
2189
2190 sc->nr.dirty += stat.nr_dirty;
2191 sc->nr.congested += stat.nr_congested;
2192 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2193 sc->nr.writeback += stat.nr_writeback;
2194 sc->nr.immediate += stat.nr_immediate;
2195 sc->nr.taken += nr_taken;
2196 if (file)
2197 sc->nr.file_taken += nr_taken;
2198
2199 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2200 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2201 return nr_reclaimed;
2202}
2203
2204/*
2205 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2206 *
2207 * We move them the other way if the page is referenced by one or more
2208 * processes.
2209 *
2210 * If the pages are mostly unmapped, the processing is fast and it is
2211 * appropriate to hold lru_lock across the whole operation. But if
2212 * the pages are mapped, the processing is slow (page_referenced()), so
2213 * we should drop lru_lock around each page. It's impossible to balance
2214 * this, so instead we remove the pages from the LRU while processing them.
2215 * It is safe to rely on PG_active against the non-LRU pages in here because
2216 * nobody will play with that bit on a non-LRU page.
2217 *
2218 * The downside is that we have to touch page->_refcount against each page.
2219 * But we had to alter page->flags anyway.
2220 */
2221static void shrink_active_list(unsigned long nr_to_scan,
2222 struct lruvec *lruvec,
2223 struct scan_control *sc,
2224 enum lru_list lru)
2225{
2226 unsigned long nr_taken;
2227 unsigned long nr_scanned;
2228 unsigned long vm_flags;
2229 LIST_HEAD(l_hold); /* The pages which were snipped off */
2230 LIST_HEAD(l_active);
2231 LIST_HEAD(l_inactive);
2232 struct page *page;
2233 unsigned nr_deactivate, nr_activate;
2234 unsigned nr_rotated = 0;
2235 int file = is_file_lru(lru);
2236 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2237
2238 lru_add_drain();
2239
2240 spin_lock_irq(&lruvec->lru_lock);
2241
2242 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2243 &nr_scanned, sc, lru);
2244
2245 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2246
2247 if (!cgroup_reclaim(sc))
2248 __count_vm_events(PGREFILL, nr_scanned);
2249 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2250
2251 spin_unlock_irq(&lruvec->lru_lock);
2252
2253 while (!list_empty(&l_hold)) {
2254 cond_resched();
2255 page = lru_to_page(&l_hold);
2256 list_del(&page->lru);
2257
2258 if (unlikely(!page_evictable(page))) {
2259 putback_lru_page(page);
2260 continue;
2261 }
2262
2263 if (unlikely(buffer_heads_over_limit)) {
2264 if (page_has_private(page) && trylock_page(page)) {
2265 if (page_has_private(page))
2266 try_to_release_page(page, 0);
2267 unlock_page(page);
2268 }
2269 }
2270
2271 if (page_referenced(page, 0, sc->target_mem_cgroup,
2272 &vm_flags)) {
2273 /*
2274 * Identify referenced, file-backed active pages and
2275 * give them one more trip around the active list. So
2276 * that executable code get better chances to stay in
2277 * memory under moderate memory pressure. Anon pages
2278 * are not likely to be evicted by use-once streaming
2279 * IO, plus JVM can create lots of anon VM_EXEC pages,
2280 * so we ignore them here.
2281 */
2282 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2283 nr_rotated += thp_nr_pages(page);
2284 list_add(&page->lru, &l_active);
2285 continue;
2286 }
2287 }
2288
2289 ClearPageActive(page); /* we are de-activating */
2290 SetPageWorkingset(page);
2291 list_add(&page->lru, &l_inactive);
2292 }
2293
2294 /*
2295 * Move pages back to the lru list.
2296 */
2297 spin_lock_irq(&lruvec->lru_lock);
2298
2299 nr_activate = move_pages_to_lru(lruvec, &l_active);
2300 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2301 /* Keep all free pages in l_active list */
2302 list_splice(&l_inactive, &l_active);
2303
2304 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2305 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2306
2307 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2308 spin_unlock_irq(&lruvec->lru_lock);
2309
2310 mem_cgroup_uncharge_list(&l_active);
2311 free_unref_page_list(&l_active);
2312 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2313 nr_deactivate, nr_rotated, sc->priority, file);
2314}
2315
2316unsigned long reclaim_pages(struct list_head *page_list)
2317{
2318 int nid = NUMA_NO_NODE;
2319 unsigned int nr_reclaimed = 0;
2320 LIST_HEAD(node_page_list);
2321 struct reclaim_stat dummy_stat;
2322 struct page *page;
2323 unsigned int noreclaim_flag;
2324 struct scan_control sc = {
2325 .gfp_mask = GFP_KERNEL,
2326 .priority = DEF_PRIORITY,
2327 .may_writepage = 1,
2328 .may_unmap = 1,
2329 .may_swap = 1,
2330 };
2331
2332 noreclaim_flag = memalloc_noreclaim_save();
2333
2334 while (!list_empty(page_list)) {
2335 page = lru_to_page(page_list);
2336 if (nid == NUMA_NO_NODE) {
2337 nid = page_to_nid(page);
2338 INIT_LIST_HEAD(&node_page_list);
2339 }
2340
2341 if (nid == page_to_nid(page)) {
2342 ClearPageActive(page);
2343 list_move(&page->lru, &node_page_list);
2344 continue;
2345 }
2346
2347 nr_reclaimed += shrink_page_list(&node_page_list,
2348 NODE_DATA(nid),
2349 &sc, &dummy_stat, false);
2350 while (!list_empty(&node_page_list)) {
2351 page = lru_to_page(&node_page_list);
2352 list_del(&page->lru);
2353 putback_lru_page(page);
2354 }
2355
2356 nid = NUMA_NO_NODE;
2357 }
2358
2359 if (!list_empty(&node_page_list)) {
2360 nr_reclaimed += shrink_page_list(&node_page_list,
2361 NODE_DATA(nid),
2362 &sc, &dummy_stat, false);
2363 while (!list_empty(&node_page_list)) {
2364 page = lru_to_page(&node_page_list);
2365 list_del(&page->lru);
2366 putback_lru_page(page);
2367 }
2368 }
2369
2370 memalloc_noreclaim_restore(noreclaim_flag);
2371
2372 return nr_reclaimed;
2373}
2374
2375static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2376 struct lruvec *lruvec, struct scan_control *sc)
2377{
2378 if (is_active_lru(lru)) {
2379 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2380 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2381 else
2382 sc->skipped_deactivate = 1;
2383 return 0;
2384 }
2385
2386 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2387}
2388
2389/*
2390 * The inactive anon list should be small enough that the VM never has
2391 * to do too much work.
2392 *
2393 * The inactive file list should be small enough to leave most memory
2394 * to the established workingset on the scan-resistant active list,
2395 * but large enough to avoid thrashing the aggregate readahead window.
2396 *
2397 * Both inactive lists should also be large enough that each inactive
2398 * page has a chance to be referenced again before it is reclaimed.
2399 *
2400 * If that fails and refaulting is observed, the inactive list grows.
2401 *
2402 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2403 * on this LRU, maintained by the pageout code. An inactive_ratio
2404 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2405 *
2406 * total target max
2407 * memory ratio inactive
2408 * -------------------------------------
2409 * 10MB 1 5MB
2410 * 100MB 1 50MB
2411 * 1GB 3 250MB
2412 * 10GB 10 0.9GB
2413 * 100GB 31 3GB
2414 * 1TB 101 10GB
2415 * 10TB 320 32GB
2416 */
2417static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2418{
2419 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2420 unsigned long inactive, active;
2421 unsigned long inactive_ratio;
2422 unsigned long gb;
2423
2424 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2425 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2426
2427 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2428 if (gb)
2429 inactive_ratio = int_sqrt(10 * gb);
2430 else
2431 inactive_ratio = 1;
2432
2433 return inactive * inactive_ratio < active;
2434}
2435
2436enum scan_balance {
2437 SCAN_EQUAL,
2438 SCAN_FRACT,
2439 SCAN_ANON,
2440 SCAN_FILE,
2441};
2442
2443/*
2444 * Determine how aggressively the anon and file LRU lists should be
2445 * scanned. The relative value of each set of LRU lists is determined
2446 * by looking at the fraction of the pages scanned we did rotate back
2447 * onto the active list instead of evict.
2448 *
2449 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2450 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2451 */
2452static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2453 unsigned long *nr)
2454{
2455 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2456 unsigned long anon_cost, file_cost, total_cost;
2457 int swappiness = mem_cgroup_swappiness(memcg);
2458 u64 fraction[ANON_AND_FILE];
2459 u64 denominator = 0; /* gcc */
2460 enum scan_balance scan_balance;
2461 unsigned long ap, fp;
2462 enum lru_list lru;
2463
2464 /* If we have no swap space, do not bother scanning anon pages. */
2465 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2466 scan_balance = SCAN_FILE;
2467 goto out;
2468 }
2469
2470 /*
2471 * Global reclaim will swap to prevent OOM even with no
2472 * swappiness, but memcg users want to use this knob to
2473 * disable swapping for individual groups completely when
2474 * using the memory controller's swap limit feature would be
2475 * too expensive.
2476 */
2477 if (cgroup_reclaim(sc) && !swappiness) {
2478 scan_balance = SCAN_FILE;
2479 goto out;
2480 }
2481
2482 /*
2483 * Do not apply any pressure balancing cleverness when the
2484 * system is close to OOM, scan both anon and file equally
2485 * (unless the swappiness setting disagrees with swapping).
2486 */
2487 if (!sc->priority && swappiness) {
2488 scan_balance = SCAN_EQUAL;
2489 goto out;
2490 }
2491
2492 /*
2493 * If the system is almost out of file pages, force-scan anon.
2494 */
2495 if (sc->file_is_tiny) {
2496 scan_balance = SCAN_ANON;
2497 goto out;
2498 }
2499
2500 /*
2501 * If there is enough inactive page cache, we do not reclaim
2502 * anything from the anonymous working right now.
2503 */
2504 if (sc->cache_trim_mode) {
2505 scan_balance = SCAN_FILE;
2506 goto out;
2507 }
2508
2509 scan_balance = SCAN_FRACT;
2510 /*
2511 * Calculate the pressure balance between anon and file pages.
2512 *
2513 * The amount of pressure we put on each LRU is inversely
2514 * proportional to the cost of reclaiming each list, as
2515 * determined by the share of pages that are refaulting, times
2516 * the relative IO cost of bringing back a swapped out
2517 * anonymous page vs reloading a filesystem page (swappiness).
2518 *
2519 * Although we limit that influence to ensure no list gets
2520 * left behind completely: at least a third of the pressure is
2521 * applied, before swappiness.
2522 *
2523 * With swappiness at 100, anon and file have equal IO cost.
2524 */
2525 total_cost = sc->anon_cost + sc->file_cost;
2526 anon_cost = total_cost + sc->anon_cost;
2527 file_cost = total_cost + sc->file_cost;
2528 total_cost = anon_cost + file_cost;
2529
2530 ap = swappiness * (total_cost + 1);
2531 ap /= anon_cost + 1;
2532
2533 fp = (200 - swappiness) * (total_cost + 1);
2534 fp /= file_cost + 1;
2535
2536 fraction[0] = ap;
2537 fraction[1] = fp;
2538 denominator = ap + fp;
2539out:
2540 for_each_evictable_lru(lru) {
2541 int file = is_file_lru(lru);
2542 unsigned long lruvec_size;
2543 unsigned long low, min;
2544 unsigned long scan;
2545
2546 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2547 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2548 &min, &low);
2549
2550 if (min || low) {
2551 /*
2552 * Scale a cgroup's reclaim pressure by proportioning
2553 * its current usage to its memory.low or memory.min
2554 * setting.
2555 *
2556 * This is important, as otherwise scanning aggression
2557 * becomes extremely binary -- from nothing as we
2558 * approach the memory protection threshold, to totally
2559 * nominal as we exceed it. This results in requiring
2560 * setting extremely liberal protection thresholds. It
2561 * also means we simply get no protection at all if we
2562 * set it too low, which is not ideal.
2563 *
2564 * If there is any protection in place, we reduce scan
2565 * pressure by how much of the total memory used is
2566 * within protection thresholds.
2567 *
2568 * There is one special case: in the first reclaim pass,
2569 * we skip over all groups that are within their low
2570 * protection. If that fails to reclaim enough pages to
2571 * satisfy the reclaim goal, we come back and override
2572 * the best-effort low protection. However, we still
2573 * ideally want to honor how well-behaved groups are in
2574 * that case instead of simply punishing them all
2575 * equally. As such, we reclaim them based on how much
2576 * memory they are using, reducing the scan pressure
2577 * again by how much of the total memory used is under
2578 * hard protection.
2579 */
2580 unsigned long cgroup_size = mem_cgroup_size(memcg);
2581 unsigned long protection;
2582
2583 /* memory.low scaling, make sure we retry before OOM */
2584 if (!sc->memcg_low_reclaim && low > min) {
2585 protection = low;
2586 sc->memcg_low_skipped = 1;
2587 } else {
2588 protection = min;
2589 }
2590
2591 /* Avoid TOCTOU with earlier protection check */
2592 cgroup_size = max(cgroup_size, protection);
2593
2594 scan = lruvec_size - lruvec_size * protection /
2595 (cgroup_size + 1);
2596
2597 /*
2598 * Minimally target SWAP_CLUSTER_MAX pages to keep
2599 * reclaim moving forwards, avoiding decrementing
2600 * sc->priority further than desirable.
2601 */
2602 scan = max(scan, SWAP_CLUSTER_MAX);
2603 } else {
2604 scan = lruvec_size;
2605 }
2606
2607 scan >>= sc->priority;
2608
2609 /*
2610 * If the cgroup's already been deleted, make sure to
2611 * scrape out the remaining cache.
2612 */
2613 if (!scan && !mem_cgroup_online(memcg))
2614 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2615
2616 switch (scan_balance) {
2617 case SCAN_EQUAL:
2618 /* Scan lists relative to size */
2619 break;
2620 case SCAN_FRACT:
2621 /*
2622 * Scan types proportional to swappiness and
2623 * their relative recent reclaim efficiency.
2624 * Make sure we don't miss the last page on
2625 * the offlined memory cgroups because of a
2626 * round-off error.
2627 */
2628 scan = mem_cgroup_online(memcg) ?
2629 div64_u64(scan * fraction[file], denominator) :
2630 DIV64_U64_ROUND_UP(scan * fraction[file],
2631 denominator);
2632 break;
2633 case SCAN_FILE:
2634 case SCAN_ANON:
2635 /* Scan one type exclusively */
2636 if ((scan_balance == SCAN_FILE) != file)
2637 scan = 0;
2638 break;
2639 default:
2640 /* Look ma, no brain */
2641 BUG();
2642 }
2643
2644 nr[lru] = scan;
2645 }
2646}
2647
2648static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2649{
2650 unsigned long nr[NR_LRU_LISTS];
2651 unsigned long targets[NR_LRU_LISTS];
2652 unsigned long nr_to_scan;
2653 enum lru_list lru;
2654 unsigned long nr_reclaimed = 0;
2655 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2656 struct blk_plug plug;
2657 bool scan_adjusted;
2658
2659 get_scan_count(lruvec, sc, nr);
2660
2661 /* Record the original scan target for proportional adjustments later */
2662 memcpy(targets, nr, sizeof(nr));
2663
2664 /*
2665 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2666 * event that can occur when there is little memory pressure e.g.
2667 * multiple streaming readers/writers. Hence, we do not abort scanning
2668 * when the requested number of pages are reclaimed when scanning at
2669 * DEF_PRIORITY on the assumption that the fact we are direct
2670 * reclaiming implies that kswapd is not keeping up and it is best to
2671 * do a batch of work at once. For memcg reclaim one check is made to
2672 * abort proportional reclaim if either the file or anon lru has already
2673 * dropped to zero at the first pass.
2674 */
2675 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2676 sc->priority == DEF_PRIORITY);
2677
2678 blk_start_plug(&plug);
2679 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2680 nr[LRU_INACTIVE_FILE]) {
2681 unsigned long nr_anon, nr_file, percentage;
2682 unsigned long nr_scanned;
2683
2684 for_each_evictable_lru(lru) {
2685 if (nr[lru]) {
2686 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2687 nr[lru] -= nr_to_scan;
2688
2689 nr_reclaimed += shrink_list(lru, nr_to_scan,
2690 lruvec, sc);
2691 }
2692 }
2693
2694 cond_resched();
2695
2696 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2697 continue;
2698
2699 /*
2700 * For kswapd and memcg, reclaim at least the number of pages
2701 * requested. Ensure that the anon and file LRUs are scanned
2702 * proportionally what was requested by get_scan_count(). We
2703 * stop reclaiming one LRU and reduce the amount scanning
2704 * proportional to the original scan target.
2705 */
2706 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2707 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2708
2709 /*
2710 * It's just vindictive to attack the larger once the smaller
2711 * has gone to zero. And given the way we stop scanning the
2712 * smaller below, this makes sure that we only make one nudge
2713 * towards proportionality once we've got nr_to_reclaim.
2714 */
2715 if (!nr_file || !nr_anon)
2716 break;
2717
2718 if (nr_file > nr_anon) {
2719 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2720 targets[LRU_ACTIVE_ANON] + 1;
2721 lru = LRU_BASE;
2722 percentage = nr_anon * 100 / scan_target;
2723 } else {
2724 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2725 targets[LRU_ACTIVE_FILE] + 1;
2726 lru = LRU_FILE;
2727 percentage = nr_file * 100 / scan_target;
2728 }
2729
2730 /* Stop scanning the smaller of the LRU */
2731 nr[lru] = 0;
2732 nr[lru + LRU_ACTIVE] = 0;
2733
2734 /*
2735 * Recalculate the other LRU scan count based on its original
2736 * scan target and the percentage scanning already complete
2737 */
2738 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2739 nr_scanned = targets[lru] - nr[lru];
2740 nr[lru] = targets[lru] * (100 - percentage) / 100;
2741 nr[lru] -= min(nr[lru], nr_scanned);
2742
2743 lru += LRU_ACTIVE;
2744 nr_scanned = targets[lru] - nr[lru];
2745 nr[lru] = targets[lru] * (100 - percentage) / 100;
2746 nr[lru] -= min(nr[lru], nr_scanned);
2747
2748 scan_adjusted = true;
2749 }
2750 blk_finish_plug(&plug);
2751 sc->nr_reclaimed += nr_reclaimed;
2752
2753 /*
2754 * Even if we did not try to evict anon pages at all, we want to
2755 * rebalance the anon lru active/inactive ratio.
2756 */
2757 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2758 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2759 sc, LRU_ACTIVE_ANON);
2760}
2761
2762/* Use reclaim/compaction for costly allocs or under memory pressure */
2763static bool in_reclaim_compaction(struct scan_control *sc)
2764{
2765 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2766 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2767 sc->priority < DEF_PRIORITY - 2))
2768 return true;
2769
2770 return false;
2771}
2772
2773/*
2774 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2775 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2776 * true if more pages should be reclaimed such that when the page allocator
2777 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2778 * It will give up earlier than that if there is difficulty reclaiming pages.
2779 */
2780static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2781 unsigned long nr_reclaimed,
2782 struct scan_control *sc)
2783{
2784 unsigned long pages_for_compaction;
2785 unsigned long inactive_lru_pages;
2786 int z;
2787
2788 /* If not in reclaim/compaction mode, stop */
2789 if (!in_reclaim_compaction(sc))
2790 return false;
2791
2792 /*
2793 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2794 * number of pages that were scanned. This will return to the caller
2795 * with the risk reclaim/compaction and the resulting allocation attempt
2796 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2797 * allocations through requiring that the full LRU list has been scanned
2798 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2799 * scan, but that approximation was wrong, and there were corner cases
2800 * where always a non-zero amount of pages were scanned.
2801 */
2802 if (!nr_reclaimed)
2803 return false;
2804
2805 /* If compaction would go ahead or the allocation would succeed, stop */
2806 for (z = 0; z <= sc->reclaim_idx; z++) {
2807 struct zone *zone = &pgdat->node_zones[z];
2808 if (!managed_zone(zone))
2809 continue;
2810
2811 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2812 case COMPACT_SUCCESS:
2813 case COMPACT_CONTINUE:
2814 return false;
2815 default:
2816 /* check next zone */
2817 ;
2818 }
2819 }
2820
2821 /*
2822 * If we have not reclaimed enough pages for compaction and the
2823 * inactive lists are large enough, continue reclaiming
2824 */
2825 pages_for_compaction = compact_gap(sc->order);
2826 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2827 if (get_nr_swap_pages() > 0)
2828 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2829
2830 return inactive_lru_pages > pages_for_compaction;
2831}
2832
2833static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2834{
2835 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2836 struct mem_cgroup *memcg;
2837
2838 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2839 do {
2840 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2841 unsigned long reclaimed;
2842 unsigned long scanned;
2843
2844 /*
2845 * This loop can become CPU-bound when target memcgs
2846 * aren't eligible for reclaim - either because they
2847 * don't have any reclaimable pages, or because their
2848 * memory is explicitly protected. Avoid soft lockups.
2849 */
2850 cond_resched();
2851
2852 mem_cgroup_calculate_protection(target_memcg, memcg);
2853
2854 if (mem_cgroup_below_min(memcg)) {
2855 /*
2856 * Hard protection.
2857 * If there is no reclaimable memory, OOM.
2858 */
2859 continue;
2860 } else if (mem_cgroup_below_low(memcg)) {
2861 /*
2862 * Soft protection.
2863 * Respect the protection only as long as
2864 * there is an unprotected supply
2865 * of reclaimable memory from other cgroups.
2866 */
2867 if (!sc->memcg_low_reclaim) {
2868 sc->memcg_low_skipped = 1;
2869 continue;
2870 }
2871 memcg_memory_event(memcg, MEMCG_LOW);
2872 }
2873
2874 reclaimed = sc->nr_reclaimed;
2875 scanned = sc->nr_scanned;
2876
2877 shrink_lruvec(lruvec, sc);
2878
2879 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2880 sc->priority);
2881
2882 /* Record the group's reclaim efficiency */
2883 vmpressure(sc->gfp_mask, memcg, false,
2884 sc->nr_scanned - scanned,
2885 sc->nr_reclaimed - reclaimed);
2886
2887 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2888}
2889
2890static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2891{
2892 struct reclaim_state *reclaim_state = current->reclaim_state;
2893 unsigned long nr_reclaimed, nr_scanned;
2894 struct lruvec *target_lruvec;
2895 bool reclaimable = false;
2896 unsigned long file;
2897
2898 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2899
2900again:
2901 memset(&sc->nr, 0, sizeof(sc->nr));
2902
2903 nr_reclaimed = sc->nr_reclaimed;
2904 nr_scanned = sc->nr_scanned;
2905
2906 /*
2907 * Determine the scan balance between anon and file LRUs.
2908 */
2909 spin_lock_irq(&target_lruvec->lru_lock);
2910 sc->anon_cost = target_lruvec->anon_cost;
2911 sc->file_cost = target_lruvec->file_cost;
2912 spin_unlock_irq(&target_lruvec->lru_lock);
2913
2914 /*
2915 * Target desirable inactive:active list ratios for the anon
2916 * and file LRU lists.
2917 */
2918 if (!sc->force_deactivate) {
2919 unsigned long refaults;
2920
2921 refaults = lruvec_page_state(target_lruvec,
2922 WORKINGSET_ACTIVATE_ANON);
2923 if (refaults != target_lruvec->refaults[0] ||
2924 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2925 sc->may_deactivate |= DEACTIVATE_ANON;
2926 else
2927 sc->may_deactivate &= ~DEACTIVATE_ANON;
2928
2929 /*
2930 * When refaults are being observed, it means a new
2931 * workingset is being established. Deactivate to get
2932 * rid of any stale active pages quickly.
2933 */
2934 refaults = lruvec_page_state(target_lruvec,
2935 WORKINGSET_ACTIVATE_FILE);
2936 if (refaults != target_lruvec->refaults[1] ||
2937 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2938 sc->may_deactivate |= DEACTIVATE_FILE;
2939 else
2940 sc->may_deactivate &= ~DEACTIVATE_FILE;
2941 } else
2942 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2943
2944 /*
2945 * If we have plenty of inactive file pages that aren't
2946 * thrashing, try to reclaim those first before touching
2947 * anonymous pages.
2948 */
2949 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2950 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2951 sc->cache_trim_mode = 1;
2952 else
2953 sc->cache_trim_mode = 0;
2954
2955 /*
2956 * Prevent the reclaimer from falling into the cache trap: as
2957 * cache pages start out inactive, every cache fault will tip
2958 * the scan balance towards the file LRU. And as the file LRU
2959 * shrinks, so does the window for rotation from references.
2960 * This means we have a runaway feedback loop where a tiny
2961 * thrashing file LRU becomes infinitely more attractive than
2962 * anon pages. Try to detect this based on file LRU size.
2963 */
2964 if (!cgroup_reclaim(sc)) {
2965 unsigned long total_high_wmark = 0;
2966 unsigned long free, anon;
2967 int z;
2968
2969 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2970 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2971 node_page_state(pgdat, NR_INACTIVE_FILE);
2972
2973 for (z = 0; z < MAX_NR_ZONES; z++) {
2974 struct zone *zone = &pgdat->node_zones[z];
2975 if (!managed_zone(zone))
2976 continue;
2977
2978 total_high_wmark += high_wmark_pages(zone);
2979 }
2980
2981 /*
2982 * Consider anon: if that's low too, this isn't a
2983 * runaway file reclaim problem, but rather just
2984 * extreme pressure. Reclaim as per usual then.
2985 */
2986 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2987
2988 sc->file_is_tiny =
2989 file + free <= total_high_wmark &&
2990 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2991 anon >> sc->priority;
2992 }
2993
2994 shrink_node_memcgs(pgdat, sc);
2995
2996 if (reclaim_state) {
2997 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2998 reclaim_state->reclaimed_slab = 0;
2999 }
3000
3001 /* Record the subtree's reclaim efficiency */
3002 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3003 sc->nr_scanned - nr_scanned,
3004 sc->nr_reclaimed - nr_reclaimed);
3005
3006 if (sc->nr_reclaimed - nr_reclaimed)
3007 reclaimable = true;
3008
3009 if (current_is_kswapd()) {
3010 /*
3011 * If reclaim is isolating dirty pages under writeback,
3012 * it implies that the long-lived page allocation rate
3013 * is exceeding the page laundering rate. Either the
3014 * global limits are not being effective at throttling
3015 * processes due to the page distribution throughout
3016 * zones or there is heavy usage of a slow backing
3017 * device. The only option is to throttle from reclaim
3018 * context which is not ideal as there is no guarantee
3019 * the dirtying process is throttled in the same way
3020 * balance_dirty_pages() manages.
3021 *
3022 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3023 * count the number of pages under pages flagged for
3024 * immediate reclaim and stall if any are encountered
3025 * in the nr_immediate check below.
3026 */
3027 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3028 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3029
3030 /* Allow kswapd to start writing pages during reclaim.*/
3031 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3032 set_bit(PGDAT_DIRTY, &pgdat->flags);
3033
3034 /*
3035 * If kswapd scans pages marked for immediate
3036 * reclaim and under writeback (nr_immediate), it
3037 * implies that pages are cycling through the LRU
3038 * faster than they are written so also forcibly stall.
3039 */
3040 if (sc->nr.immediate)
3041 congestion_wait(BLK_RW_ASYNC, HZ/10);
3042 }
3043
3044 /*
3045 * Tag a node/memcg as congested if all the dirty pages
3046 * scanned were backed by a congested BDI and
3047 * wait_iff_congested will stall.
3048 *
3049 * Legacy memcg will stall in page writeback so avoid forcibly
3050 * stalling in wait_iff_congested().
3051 */
3052 if ((current_is_kswapd() ||
3053 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3054 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3055 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3056
3057 /*
3058 * Stall direct reclaim for IO completions if underlying BDIs
3059 * and node is congested. Allow kswapd to continue until it
3060 * starts encountering unqueued dirty pages or cycling through
3061 * the LRU too quickly.
3062 */
3063 if (!current_is_kswapd() && current_may_throttle() &&
3064 !sc->hibernation_mode &&
3065 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3066 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
3067
3068 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3069 sc))
3070 goto again;
3071
3072 /*
3073 * Kswapd gives up on balancing particular nodes after too
3074 * many failures to reclaim anything from them and goes to
3075 * sleep. On reclaim progress, reset the failure counter. A
3076 * successful direct reclaim run will revive a dormant kswapd.
3077 */
3078 if (reclaimable)
3079 pgdat->kswapd_failures = 0;
3080}
3081
3082/*
3083 * Returns true if compaction should go ahead for a costly-order request, or
3084 * the allocation would already succeed without compaction. Return false if we
3085 * should reclaim first.
3086 */
3087static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3088{
3089 unsigned long watermark;
3090 enum compact_result suitable;
3091
3092 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3093 if (suitable == COMPACT_SUCCESS)
3094 /* Allocation should succeed already. Don't reclaim. */
3095 return true;
3096 if (suitable == COMPACT_SKIPPED)
3097 /* Compaction cannot yet proceed. Do reclaim. */
3098 return false;
3099
3100 /*
3101 * Compaction is already possible, but it takes time to run and there
3102 * are potentially other callers using the pages just freed. So proceed
3103 * with reclaim to make a buffer of free pages available to give
3104 * compaction a reasonable chance of completing and allocating the page.
3105 * Note that we won't actually reclaim the whole buffer in one attempt
3106 * as the target watermark in should_continue_reclaim() is lower. But if
3107 * we are already above the high+gap watermark, don't reclaim at all.
3108 */
3109 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3110
3111 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3112}
3113
3114/*
3115 * This is the direct reclaim path, for page-allocating processes. We only
3116 * try to reclaim pages from zones which will satisfy the caller's allocation
3117 * request.
3118 *
3119 * If a zone is deemed to be full of pinned pages then just give it a light
3120 * scan then give up on it.
3121 */
3122static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3123{
3124 struct zoneref *z;
3125 struct zone *zone;
3126 unsigned long nr_soft_reclaimed;
3127 unsigned long nr_soft_scanned;
3128 gfp_t orig_mask;
3129 pg_data_t *last_pgdat = NULL;
3130
3131 /*
3132 * If the number of buffer_heads in the machine exceeds the maximum
3133 * allowed level, force direct reclaim to scan the highmem zone as
3134 * highmem pages could be pinning lowmem pages storing buffer_heads
3135 */
3136 orig_mask = sc->gfp_mask;
3137 if (buffer_heads_over_limit) {
3138 sc->gfp_mask |= __GFP_HIGHMEM;
3139 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3140 }
3141
3142 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3143 sc->reclaim_idx, sc->nodemask) {
3144 /*
3145 * Take care memory controller reclaiming has small influence
3146 * to global LRU.
3147 */
3148 if (!cgroup_reclaim(sc)) {
3149 if (!cpuset_zone_allowed(zone,
3150 GFP_KERNEL | __GFP_HARDWALL))
3151 continue;
3152
3153 /*
3154 * If we already have plenty of memory free for
3155 * compaction in this zone, don't free any more.
3156 * Even though compaction is invoked for any
3157 * non-zero order, only frequent costly order
3158 * reclamation is disruptive enough to become a
3159 * noticeable problem, like transparent huge
3160 * page allocations.
3161 */
3162 if (IS_ENABLED(CONFIG_COMPACTION) &&
3163 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3164 compaction_ready(zone, sc)) {
3165 sc->compaction_ready = true;
3166 continue;
3167 }
3168
3169 /*
3170 * Shrink each node in the zonelist once. If the
3171 * zonelist is ordered by zone (not the default) then a
3172 * node may be shrunk multiple times but in that case
3173 * the user prefers lower zones being preserved.
3174 */
3175 if (zone->zone_pgdat == last_pgdat)
3176 continue;
3177
3178 /*
3179 * This steals pages from memory cgroups over softlimit
3180 * and returns the number of reclaimed pages and
3181 * scanned pages. This works for global memory pressure
3182 * and balancing, not for a memcg's limit.
3183 */
3184 nr_soft_scanned = 0;
3185 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3186 sc->order, sc->gfp_mask,
3187 &nr_soft_scanned);
3188 sc->nr_reclaimed += nr_soft_reclaimed;
3189 sc->nr_scanned += nr_soft_scanned;
3190 /* need some check for avoid more shrink_zone() */
3191 }
3192
3193 /* See comment about same check for global reclaim above */
3194 if (zone->zone_pgdat == last_pgdat)
3195 continue;
3196 last_pgdat = zone->zone_pgdat;
3197 shrink_node(zone->zone_pgdat, sc);
3198 }
3199
3200 /*
3201 * Restore to original mask to avoid the impact on the caller if we
3202 * promoted it to __GFP_HIGHMEM.
3203 */
3204 sc->gfp_mask = orig_mask;
3205}
3206
3207static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3208{
3209 struct lruvec *target_lruvec;
3210 unsigned long refaults;
3211
3212 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3213 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3214 target_lruvec->refaults[0] = refaults;
3215 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3216 target_lruvec->refaults[1] = refaults;
3217}
3218
3219/*
3220 * This is the main entry point to direct page reclaim.
3221 *
3222 * If a full scan of the inactive list fails to free enough memory then we
3223 * are "out of memory" and something needs to be killed.
3224 *
3225 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3226 * high - the zone may be full of dirty or under-writeback pages, which this
3227 * caller can't do much about. We kick the writeback threads and take explicit
3228 * naps in the hope that some of these pages can be written. But if the
3229 * allocating task holds filesystem locks which prevent writeout this might not
3230 * work, and the allocation attempt will fail.
3231 *
3232 * returns: 0, if no pages reclaimed
3233 * else, the number of pages reclaimed
3234 */
3235static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3236 struct scan_control *sc)
3237{
3238 int initial_priority = sc->priority;
3239 pg_data_t *last_pgdat;
3240 struct zoneref *z;
3241 struct zone *zone;
3242retry:
3243 delayacct_freepages_start();
3244
3245 if (!cgroup_reclaim(sc))
3246 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3247
3248 do {
3249 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3250 sc->priority);
3251 sc->nr_scanned = 0;
3252 shrink_zones(zonelist, sc);
3253
3254 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3255 break;
3256
3257 if (sc->compaction_ready)
3258 break;
3259
3260 /*
3261 * If we're getting trouble reclaiming, start doing
3262 * writepage even in laptop mode.
3263 */
3264 if (sc->priority < DEF_PRIORITY - 2)
3265 sc->may_writepage = 1;
3266 } while (--sc->priority >= 0);
3267
3268 last_pgdat = NULL;
3269 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3270 sc->nodemask) {
3271 if (zone->zone_pgdat == last_pgdat)
3272 continue;
3273 last_pgdat = zone->zone_pgdat;
3274
3275 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3276
3277 if (cgroup_reclaim(sc)) {
3278 struct lruvec *lruvec;
3279
3280 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3281 zone->zone_pgdat);
3282 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3283 }
3284 }
3285
3286 delayacct_freepages_end();
3287
3288 if (sc->nr_reclaimed)
3289 return sc->nr_reclaimed;
3290
3291 /* Aborted reclaim to try compaction? don't OOM, then */
3292 if (sc->compaction_ready)
3293 return 1;
3294
3295 /*
3296 * We make inactive:active ratio decisions based on the node's
3297 * composition of memory, but a restrictive reclaim_idx or a
3298 * memory.low cgroup setting can exempt large amounts of
3299 * memory from reclaim. Neither of which are very common, so
3300 * instead of doing costly eligibility calculations of the
3301 * entire cgroup subtree up front, we assume the estimates are
3302 * good, and retry with forcible deactivation if that fails.
3303 */
3304 if (sc->skipped_deactivate) {
3305 sc->priority = initial_priority;
3306 sc->force_deactivate = 1;
3307 sc->skipped_deactivate = 0;
3308 goto retry;
3309 }
3310
3311 /* Untapped cgroup reserves? Don't OOM, retry. */
3312 if (sc->memcg_low_skipped) {
3313 sc->priority = initial_priority;
3314 sc->force_deactivate = 0;
3315 sc->memcg_low_reclaim = 1;
3316 sc->memcg_low_skipped = 0;
3317 goto retry;
3318 }
3319
3320 return 0;
3321}
3322
3323static bool allow_direct_reclaim(pg_data_t *pgdat)
3324{
3325 struct zone *zone;
3326 unsigned long pfmemalloc_reserve = 0;
3327 unsigned long free_pages = 0;
3328 int i;
3329 bool wmark_ok;
3330
3331 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3332 return true;
3333
3334 for (i = 0; i <= ZONE_NORMAL; i++) {
3335 zone = &pgdat->node_zones[i];
3336 if (!managed_zone(zone))
3337 continue;
3338
3339 if (!zone_reclaimable_pages(zone))
3340 continue;
3341
3342 pfmemalloc_reserve += min_wmark_pages(zone);
3343 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3344 }
3345
3346 /* If there are no reserves (unexpected config) then do not throttle */
3347 if (!pfmemalloc_reserve)
3348 return true;
3349
3350 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3351
3352 /* kswapd must be awake if processes are being throttled */
3353 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3354 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3355 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3356
3357 wake_up_interruptible(&pgdat->kswapd_wait);
3358 }
3359
3360 return wmark_ok;
3361}
3362
3363/*
3364 * Throttle direct reclaimers if backing storage is backed by the network
3365 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3366 * depleted. kswapd will continue to make progress and wake the processes
3367 * when the low watermark is reached.
3368 *
3369 * Returns true if a fatal signal was delivered during throttling. If this
3370 * happens, the page allocator should not consider triggering the OOM killer.
3371 */
3372static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3373 nodemask_t *nodemask)
3374{
3375 struct zoneref *z;
3376 struct zone *zone;
3377 pg_data_t *pgdat = NULL;
3378
3379 /*
3380 * Kernel threads should not be throttled as they may be indirectly
3381 * responsible for cleaning pages necessary for reclaim to make forward
3382 * progress. kjournald for example may enter direct reclaim while
3383 * committing a transaction where throttling it could forcing other
3384 * processes to block on log_wait_commit().
3385 */
3386 if (current->flags & PF_KTHREAD)
3387 goto out;
3388
3389 /*
3390 * If a fatal signal is pending, this process should not throttle.
3391 * It should return quickly so it can exit and free its memory
3392 */
3393 if (fatal_signal_pending(current))
3394 goto out;
3395
3396 /*
3397 * Check if the pfmemalloc reserves are ok by finding the first node
3398 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3399 * GFP_KERNEL will be required for allocating network buffers when
3400 * swapping over the network so ZONE_HIGHMEM is unusable.
3401 *
3402 * Throttling is based on the first usable node and throttled processes
3403 * wait on a queue until kswapd makes progress and wakes them. There
3404 * is an affinity then between processes waking up and where reclaim
3405 * progress has been made assuming the process wakes on the same node.
3406 * More importantly, processes running on remote nodes will not compete
3407 * for remote pfmemalloc reserves and processes on different nodes
3408 * should make reasonable progress.
3409 */
3410 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3411 gfp_zone(gfp_mask), nodemask) {
3412 if (zone_idx(zone) > ZONE_NORMAL)
3413 continue;
3414
3415 /* Throttle based on the first usable node */
3416 pgdat = zone->zone_pgdat;
3417 if (allow_direct_reclaim(pgdat))
3418 goto out;
3419 break;
3420 }
3421
3422 /* If no zone was usable by the allocation flags then do not throttle */
3423 if (!pgdat)
3424 goto out;
3425
3426 /* Account for the throttling */
3427 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3428
3429 /*
3430 * If the caller cannot enter the filesystem, it's possible that it
3431 * is due to the caller holding an FS lock or performing a journal
3432 * transaction in the case of a filesystem like ext[3|4]. In this case,
3433 * it is not safe to block on pfmemalloc_wait as kswapd could be
3434 * blocked waiting on the same lock. Instead, throttle for up to a
3435 * second before continuing.
3436 */
3437 if (!(gfp_mask & __GFP_FS)) {
3438 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3439 allow_direct_reclaim(pgdat), HZ);
3440
3441 goto check_pending;
3442 }
3443
3444 /* Throttle until kswapd wakes the process */
3445 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3446 allow_direct_reclaim(pgdat));
3447
3448check_pending:
3449 if (fatal_signal_pending(current))
3450 return true;
3451
3452out:
3453 return false;
3454}
3455
3456unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3457 gfp_t gfp_mask, nodemask_t *nodemask)
3458{
3459 unsigned long nr_reclaimed;
3460 struct scan_control sc = {
3461 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3462 .gfp_mask = current_gfp_context(gfp_mask),
3463 .reclaim_idx = gfp_zone(gfp_mask),
3464 .order = order,
3465 .nodemask = nodemask,
3466 .priority = DEF_PRIORITY,
3467 .may_writepage = !laptop_mode,
3468 .may_unmap = 1,
3469 .may_swap = 1,
3470 };
3471
3472 /*
3473 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3474 * Confirm they are large enough for max values.
3475 */
3476 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3477 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3478 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3479
3480 /*
3481 * Do not enter reclaim if fatal signal was delivered while throttled.
3482 * 1 is returned so that the page allocator does not OOM kill at this
3483 * point.
3484 */
3485 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3486 return 1;
3487
3488 set_task_reclaim_state(current, &sc.reclaim_state);
3489 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3490
3491 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3492
3493 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3494 set_task_reclaim_state(current, NULL);
3495
3496 return nr_reclaimed;
3497}
3498
3499#ifdef CONFIG_MEMCG
3500
3501/* Only used by soft limit reclaim. Do not reuse for anything else. */
3502unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3503 gfp_t gfp_mask, bool noswap,
3504 pg_data_t *pgdat,
3505 unsigned long *nr_scanned)
3506{
3507 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3508 struct scan_control sc = {
3509 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3510 .target_mem_cgroup = memcg,
3511 .may_writepage = !laptop_mode,
3512 .may_unmap = 1,
3513 .reclaim_idx = MAX_NR_ZONES - 1,
3514 .may_swap = !noswap,
3515 };
3516
3517 WARN_ON_ONCE(!current->reclaim_state);
3518
3519 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3520 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3521
3522 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3523 sc.gfp_mask);
3524
3525 /*
3526 * NOTE: Although we can get the priority field, using it
3527 * here is not a good idea, since it limits the pages we can scan.
3528 * if we don't reclaim here, the shrink_node from balance_pgdat
3529 * will pick up pages from other mem cgroup's as well. We hack
3530 * the priority and make it zero.
3531 */
3532 shrink_lruvec(lruvec, &sc);
3533
3534 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3535
3536 *nr_scanned = sc.nr_scanned;
3537
3538 return sc.nr_reclaimed;
3539}
3540
3541unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3542 unsigned long nr_pages,
3543 gfp_t gfp_mask,
3544 bool may_swap)
3545{
3546 unsigned long nr_reclaimed;
3547 unsigned int noreclaim_flag;
3548 struct scan_control sc = {
3549 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3550 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3551 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3552 .reclaim_idx = MAX_NR_ZONES - 1,
3553 .target_mem_cgroup = memcg,
3554 .priority = DEF_PRIORITY,
3555 .may_writepage = !laptop_mode,
3556 .may_unmap = 1,
3557 .may_swap = may_swap,
3558 };
3559 /*
3560 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3561 * equal pressure on all the nodes. This is based on the assumption that
3562 * the reclaim does not bail out early.
3563 */
3564 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3565
3566 set_task_reclaim_state(current, &sc.reclaim_state);
3567 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3568 noreclaim_flag = memalloc_noreclaim_save();
3569
3570 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3571
3572 memalloc_noreclaim_restore(noreclaim_flag);
3573 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3574 set_task_reclaim_state(current, NULL);
3575
3576 return nr_reclaimed;
3577}
3578#endif
3579
3580static void age_active_anon(struct pglist_data *pgdat,
3581 struct scan_control *sc)
3582{
3583 struct mem_cgroup *memcg;
3584 struct lruvec *lruvec;
3585
3586 if (!total_swap_pages)
3587 return;
3588
3589 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3590 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3591 return;
3592
3593 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3594 do {
3595 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3596 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3597 sc, LRU_ACTIVE_ANON);
3598 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3599 } while (memcg);
3600}
3601
3602static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3603{
3604 int i;
3605 struct zone *zone;
3606
3607 /*
3608 * Check for watermark boosts top-down as the higher zones
3609 * are more likely to be boosted. Both watermarks and boosts
3610 * should not be checked at the same time as reclaim would
3611 * start prematurely when there is no boosting and a lower
3612 * zone is balanced.
3613 */
3614 for (i = highest_zoneidx; i >= 0; i--) {
3615 zone = pgdat->node_zones + i;
3616 if (!managed_zone(zone))
3617 continue;
3618
3619 if (zone->watermark_boost)
3620 return true;
3621 }
3622
3623 return false;
3624}
3625
3626/*
3627 * Returns true if there is an eligible zone balanced for the request order
3628 * and highest_zoneidx
3629 */
3630static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3631{
3632 int i;
3633 unsigned long mark = -1;
3634 struct zone *zone;
3635
3636 /*
3637 * Check watermarks bottom-up as lower zones are more likely to
3638 * meet watermarks.
3639 */
3640 for (i = 0; i <= highest_zoneidx; i++) {
3641 zone = pgdat->node_zones + i;
3642
3643 if (!managed_zone(zone))
3644 continue;
3645
3646 mark = high_wmark_pages(zone);
3647 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3648 return true;
3649 }
3650
3651 /*
3652 * If a node has no populated zone within highest_zoneidx, it does not
3653 * need balancing by definition. This can happen if a zone-restricted
3654 * allocation tries to wake a remote kswapd.
3655 */
3656 if (mark == -1)
3657 return true;
3658
3659 return false;
3660}
3661
3662/* Clear pgdat state for congested, dirty or under writeback. */
3663static void clear_pgdat_congested(pg_data_t *pgdat)
3664{
3665 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3666
3667 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3668 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3669 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3670}
3671
3672/*
3673 * Prepare kswapd for sleeping. This verifies that there are no processes
3674 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3675 *
3676 * Returns true if kswapd is ready to sleep
3677 */
3678static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3679 int highest_zoneidx)
3680{
3681 /*
3682 * The throttled processes are normally woken up in balance_pgdat() as
3683 * soon as allow_direct_reclaim() is true. But there is a potential
3684 * race between when kswapd checks the watermarks and a process gets
3685 * throttled. There is also a potential race if processes get
3686 * throttled, kswapd wakes, a large process exits thereby balancing the
3687 * zones, which causes kswapd to exit balance_pgdat() before reaching
3688 * the wake up checks. If kswapd is going to sleep, no process should
3689 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3690 * the wake up is premature, processes will wake kswapd and get
3691 * throttled again. The difference from wake ups in balance_pgdat() is
3692 * that here we are under prepare_to_wait().
3693 */
3694 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3695 wake_up_all(&pgdat->pfmemalloc_wait);
3696
3697 /* Hopeless node, leave it to direct reclaim */
3698 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3699 return true;
3700
3701 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3702 clear_pgdat_congested(pgdat);
3703 return true;
3704 }
3705
3706 return false;
3707}
3708
3709/*
3710 * kswapd shrinks a node of pages that are at or below the highest usable
3711 * zone that is currently unbalanced.
3712 *
3713 * Returns true if kswapd scanned at least the requested number of pages to
3714 * reclaim or if the lack of progress was due to pages under writeback.
3715 * This is used to determine if the scanning priority needs to be raised.
3716 */
3717static bool kswapd_shrink_node(pg_data_t *pgdat,
3718 struct scan_control *sc)
3719{
3720 struct zone *zone;
3721 int z;
3722
3723 /* Reclaim a number of pages proportional to the number of zones */
3724 sc->nr_to_reclaim = 0;
3725 for (z = 0; z <= sc->reclaim_idx; z++) {
3726 zone = pgdat->node_zones + z;
3727 if (!managed_zone(zone))
3728 continue;
3729
3730 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3731 }
3732
3733 /*
3734 * Historically care was taken to put equal pressure on all zones but
3735 * now pressure is applied based on node LRU order.
3736 */
3737 shrink_node(pgdat, sc);
3738
3739 /*
3740 * Fragmentation may mean that the system cannot be rebalanced for
3741 * high-order allocations. If twice the allocation size has been
3742 * reclaimed then recheck watermarks only at order-0 to prevent
3743 * excessive reclaim. Assume that a process requested a high-order
3744 * can direct reclaim/compact.
3745 */
3746 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3747 sc->order = 0;
3748
3749 return sc->nr_scanned >= sc->nr_to_reclaim;
3750}
3751
3752/* Page allocator PCP high watermark is lowered if reclaim is active. */
3753static inline void
3754update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
3755{
3756 int i;
3757 struct zone *zone;
3758
3759 for (i = 0; i <= highest_zoneidx; i++) {
3760 zone = pgdat->node_zones + i;
3761
3762 if (!managed_zone(zone))
3763 continue;
3764
3765 if (active)
3766 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3767 else
3768 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3769 }
3770}
3771
3772static inline void
3773set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3774{
3775 update_reclaim_active(pgdat, highest_zoneidx, true);
3776}
3777
3778static inline void
3779clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3780{
3781 update_reclaim_active(pgdat, highest_zoneidx, false);
3782}
3783
3784/*
3785 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3786 * that are eligible for use by the caller until at least one zone is
3787 * balanced.
3788 *
3789 * Returns the order kswapd finished reclaiming at.
3790 *
3791 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3792 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3793 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3794 * or lower is eligible for reclaim until at least one usable zone is
3795 * balanced.
3796 */
3797static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3798{
3799 int i;
3800 unsigned long nr_soft_reclaimed;
3801 unsigned long nr_soft_scanned;
3802 unsigned long pflags;
3803 unsigned long nr_boost_reclaim;
3804 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3805 bool boosted;
3806 struct zone *zone;
3807 struct scan_control sc = {
3808 .gfp_mask = GFP_KERNEL,
3809 .order = order,
3810 .may_unmap = 1,
3811 };
3812
3813 set_task_reclaim_state(current, &sc.reclaim_state);
3814 psi_memstall_enter(&pflags);
3815 __fs_reclaim_acquire();
3816
3817 count_vm_event(PAGEOUTRUN);
3818
3819 /*
3820 * Account for the reclaim boost. Note that the zone boost is left in
3821 * place so that parallel allocations that are near the watermark will
3822 * stall or direct reclaim until kswapd is finished.
3823 */
3824 nr_boost_reclaim = 0;
3825 for (i = 0; i <= highest_zoneidx; i++) {
3826 zone = pgdat->node_zones + i;
3827 if (!managed_zone(zone))
3828 continue;
3829
3830 nr_boost_reclaim += zone->watermark_boost;
3831 zone_boosts[i] = zone->watermark_boost;
3832 }
3833 boosted = nr_boost_reclaim;
3834
3835restart:
3836 set_reclaim_active(pgdat, highest_zoneidx);
3837 sc.priority = DEF_PRIORITY;
3838 do {
3839 unsigned long nr_reclaimed = sc.nr_reclaimed;
3840 bool raise_priority = true;
3841 bool balanced;
3842 bool ret;
3843
3844 sc.reclaim_idx = highest_zoneidx;
3845
3846 /*
3847 * If the number of buffer_heads exceeds the maximum allowed
3848 * then consider reclaiming from all zones. This has a dual
3849 * purpose -- on 64-bit systems it is expected that
3850 * buffer_heads are stripped during active rotation. On 32-bit
3851 * systems, highmem pages can pin lowmem memory and shrinking
3852 * buffers can relieve lowmem pressure. Reclaim may still not
3853 * go ahead if all eligible zones for the original allocation
3854 * request are balanced to avoid excessive reclaim from kswapd.
3855 */
3856 if (buffer_heads_over_limit) {
3857 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3858 zone = pgdat->node_zones + i;
3859 if (!managed_zone(zone))
3860 continue;
3861
3862 sc.reclaim_idx = i;
3863 break;
3864 }
3865 }
3866
3867 /*
3868 * If the pgdat is imbalanced then ignore boosting and preserve
3869 * the watermarks for a later time and restart. Note that the
3870 * zone watermarks will be still reset at the end of balancing
3871 * on the grounds that the normal reclaim should be enough to
3872 * re-evaluate if boosting is required when kswapd next wakes.
3873 */
3874 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3875 if (!balanced && nr_boost_reclaim) {
3876 nr_boost_reclaim = 0;
3877 goto restart;
3878 }
3879
3880 /*
3881 * If boosting is not active then only reclaim if there are no
3882 * eligible zones. Note that sc.reclaim_idx is not used as
3883 * buffer_heads_over_limit may have adjusted it.
3884 */
3885 if (!nr_boost_reclaim && balanced)
3886 goto out;
3887
3888 /* Limit the priority of boosting to avoid reclaim writeback */
3889 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3890 raise_priority = false;
3891
3892 /*
3893 * Do not writeback or swap pages for boosted reclaim. The
3894 * intent is to relieve pressure not issue sub-optimal IO
3895 * from reclaim context. If no pages are reclaimed, the
3896 * reclaim will be aborted.
3897 */
3898 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3899 sc.may_swap = !nr_boost_reclaim;
3900
3901 /*
3902 * Do some background aging of the anon list, to give
3903 * pages a chance to be referenced before reclaiming. All
3904 * pages are rotated regardless of classzone as this is
3905 * about consistent aging.
3906 */
3907 age_active_anon(pgdat, &sc);
3908
3909 /*
3910 * If we're getting trouble reclaiming, start doing writepage
3911 * even in laptop mode.
3912 */
3913 if (sc.priority < DEF_PRIORITY - 2)
3914 sc.may_writepage = 1;
3915
3916 /* Call soft limit reclaim before calling shrink_node. */
3917 sc.nr_scanned = 0;
3918 nr_soft_scanned = 0;
3919 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3920 sc.gfp_mask, &nr_soft_scanned);
3921 sc.nr_reclaimed += nr_soft_reclaimed;
3922
3923 /*
3924 * There should be no need to raise the scanning priority if
3925 * enough pages are already being scanned that that high
3926 * watermark would be met at 100% efficiency.
3927 */
3928 if (kswapd_shrink_node(pgdat, &sc))
3929 raise_priority = false;
3930
3931 /*
3932 * If the low watermark is met there is no need for processes
3933 * to be throttled on pfmemalloc_wait as they should not be
3934 * able to safely make forward progress. Wake them
3935 */
3936 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3937 allow_direct_reclaim(pgdat))
3938 wake_up_all(&pgdat->pfmemalloc_wait);
3939
3940 /* Check if kswapd should be suspending */
3941 __fs_reclaim_release();
3942 ret = try_to_freeze();
3943 __fs_reclaim_acquire();
3944 if (ret || kthread_should_stop())
3945 break;
3946
3947 /*
3948 * Raise priority if scanning rate is too low or there was no
3949 * progress in reclaiming pages
3950 */
3951 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3952 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3953
3954 /*
3955 * If reclaim made no progress for a boost, stop reclaim as
3956 * IO cannot be queued and it could be an infinite loop in
3957 * extreme circumstances.
3958 */
3959 if (nr_boost_reclaim && !nr_reclaimed)
3960 break;
3961
3962 if (raise_priority || !nr_reclaimed)
3963 sc.priority--;
3964 } while (sc.priority >= 1);
3965
3966 if (!sc.nr_reclaimed)
3967 pgdat->kswapd_failures++;
3968
3969out:
3970 clear_reclaim_active(pgdat, highest_zoneidx);
3971
3972 /* If reclaim was boosted, account for the reclaim done in this pass */
3973 if (boosted) {
3974 unsigned long flags;
3975
3976 for (i = 0; i <= highest_zoneidx; i++) {
3977 if (!zone_boosts[i])
3978 continue;
3979
3980 /* Increments are under the zone lock */
3981 zone = pgdat->node_zones + i;
3982 spin_lock_irqsave(&zone->lock, flags);
3983 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3984 spin_unlock_irqrestore(&zone->lock, flags);
3985 }
3986
3987 /*
3988 * As there is now likely space, wakeup kcompact to defragment
3989 * pageblocks.
3990 */
3991 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3992 }
3993
3994 snapshot_refaults(NULL, pgdat);
3995 __fs_reclaim_release();
3996 psi_memstall_leave(&pflags);
3997 set_task_reclaim_state(current, NULL);
3998
3999 /*
4000 * Return the order kswapd stopped reclaiming at as
4001 * prepare_kswapd_sleep() takes it into account. If another caller
4002 * entered the allocator slow path while kswapd was awake, order will
4003 * remain at the higher level.
4004 */
4005 return sc.order;
4006}
4007
4008/*
4009 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4010 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4011 * not a valid index then either kswapd runs for first time or kswapd couldn't
4012 * sleep after previous reclaim attempt (node is still unbalanced). In that
4013 * case return the zone index of the previous kswapd reclaim cycle.
4014 */
4015static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4016 enum zone_type prev_highest_zoneidx)
4017{
4018 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4019
4020 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4021}
4022
4023static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4024 unsigned int highest_zoneidx)
4025{
4026 long remaining = 0;
4027 DEFINE_WAIT(wait);
4028
4029 if (freezing(current) || kthread_should_stop())
4030 return;
4031
4032 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4033
4034 /*
4035 * Try to sleep for a short interval. Note that kcompactd will only be
4036 * woken if it is possible to sleep for a short interval. This is
4037 * deliberate on the assumption that if reclaim cannot keep an
4038 * eligible zone balanced that it's also unlikely that compaction will
4039 * succeed.
4040 */
4041 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4042 /*
4043 * Compaction records what page blocks it recently failed to
4044 * isolate pages from and skips them in the future scanning.
4045 * When kswapd is going to sleep, it is reasonable to assume
4046 * that pages and compaction may succeed so reset the cache.
4047 */
4048 reset_isolation_suitable(pgdat);
4049
4050 /*
4051 * We have freed the memory, now we should compact it to make
4052 * allocation of the requested order possible.
4053 */
4054 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4055
4056 remaining = schedule_timeout(HZ/10);
4057
4058 /*
4059 * If woken prematurely then reset kswapd_highest_zoneidx and
4060 * order. The values will either be from a wakeup request or
4061 * the previous request that slept prematurely.
4062 */
4063 if (remaining) {
4064 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4065 kswapd_highest_zoneidx(pgdat,
4066 highest_zoneidx));
4067
4068 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4069 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4070 }
4071
4072 finish_wait(&pgdat->kswapd_wait, &wait);
4073 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4074 }
4075
4076 /*
4077 * After a short sleep, check if it was a premature sleep. If not, then
4078 * go fully to sleep until explicitly woken up.
4079 */
4080 if (!remaining &&
4081 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4082 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4083
4084 /*
4085 * vmstat counters are not perfectly accurate and the estimated
4086 * value for counters such as NR_FREE_PAGES can deviate from the
4087 * true value by nr_online_cpus * threshold. To avoid the zone
4088 * watermarks being breached while under pressure, we reduce the
4089 * per-cpu vmstat threshold while kswapd is awake and restore
4090 * them before going back to sleep.
4091 */
4092 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4093
4094 if (!kthread_should_stop())
4095 schedule();
4096
4097 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4098 } else {
4099 if (remaining)
4100 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4101 else
4102 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4103 }
4104 finish_wait(&pgdat->kswapd_wait, &wait);
4105}
4106
4107/*
4108 * The background pageout daemon, started as a kernel thread
4109 * from the init process.
4110 *
4111 * This basically trickles out pages so that we have _some_
4112 * free memory available even if there is no other activity
4113 * that frees anything up. This is needed for things like routing
4114 * etc, where we otherwise might have all activity going on in
4115 * asynchronous contexts that cannot page things out.
4116 *
4117 * If there are applications that are active memory-allocators
4118 * (most normal use), this basically shouldn't matter.
4119 */
4120static int kswapd(void *p)
4121{
4122 unsigned int alloc_order, reclaim_order;
4123 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4124 pg_data_t *pgdat = (pg_data_t *)p;
4125 struct task_struct *tsk = current;
4126 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4127
4128 if (!cpumask_empty(cpumask))
4129 set_cpus_allowed_ptr(tsk, cpumask);
4130
4131 /*
4132 * Tell the memory management that we're a "memory allocator",
4133 * and that if we need more memory we should get access to it
4134 * regardless (see "__alloc_pages()"). "kswapd" should
4135 * never get caught in the normal page freeing logic.
4136 *
4137 * (Kswapd normally doesn't need memory anyway, but sometimes
4138 * you need a small amount of memory in order to be able to
4139 * page out something else, and this flag essentially protects
4140 * us from recursively trying to free more memory as we're
4141 * trying to free the first piece of memory in the first place).
4142 */
4143 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4144 set_freezable();
4145
4146 WRITE_ONCE(pgdat->kswapd_order, 0);
4147 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4148 for ( ; ; ) {
4149 bool ret;
4150
4151 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4152 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4153 highest_zoneidx);
4154
4155kswapd_try_sleep:
4156 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4157 highest_zoneidx);
4158
4159 /* Read the new order and highest_zoneidx */
4160 alloc_order = READ_ONCE(pgdat->kswapd_order);
4161 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4162 highest_zoneidx);
4163 WRITE_ONCE(pgdat->kswapd_order, 0);
4164 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4165
4166 ret = try_to_freeze();
4167 if (kthread_should_stop())
4168 break;
4169
4170 /*
4171 * We can speed up thawing tasks if we don't call balance_pgdat
4172 * after returning from the refrigerator
4173 */
4174 if (ret)
4175 continue;
4176
4177 /*
4178 * Reclaim begins at the requested order but if a high-order
4179 * reclaim fails then kswapd falls back to reclaiming for
4180 * order-0. If that happens, kswapd will consider sleeping
4181 * for the order it finished reclaiming at (reclaim_order)
4182 * but kcompactd is woken to compact for the original
4183 * request (alloc_order).
4184 */
4185 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4186 alloc_order);
4187 reclaim_order = balance_pgdat(pgdat, alloc_order,
4188 highest_zoneidx);
4189 if (reclaim_order < alloc_order)
4190 goto kswapd_try_sleep;
4191 }
4192
4193 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4194
4195 return 0;
4196}
4197
4198/*
4199 * A zone is low on free memory or too fragmented for high-order memory. If
4200 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4201 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4202 * has failed or is not needed, still wake up kcompactd if only compaction is
4203 * needed.
4204 */
4205void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4206 enum zone_type highest_zoneidx)
4207{
4208 pg_data_t *pgdat;
4209 enum zone_type curr_idx;
4210
4211 if (!managed_zone(zone))
4212 return;
4213
4214 if (!cpuset_zone_allowed(zone, gfp_flags))
4215 return;
4216
4217 pgdat = zone->zone_pgdat;
4218 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4219
4220 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4221 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4222
4223 if (READ_ONCE(pgdat->kswapd_order) < order)
4224 WRITE_ONCE(pgdat->kswapd_order, order);
4225
4226 if (!waitqueue_active(&pgdat->kswapd_wait))
4227 return;
4228
4229 /* Hopeless node, leave it to direct reclaim if possible */
4230 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4231 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4232 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4233 /*
4234 * There may be plenty of free memory available, but it's too
4235 * fragmented for high-order allocations. Wake up kcompactd
4236 * and rely on compaction_suitable() to determine if it's
4237 * needed. If it fails, it will defer subsequent attempts to
4238 * ratelimit its work.
4239 */
4240 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4241 wakeup_kcompactd(pgdat, order, highest_zoneidx);
4242 return;
4243 }
4244
4245 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4246 gfp_flags);
4247 wake_up_interruptible(&pgdat->kswapd_wait);
4248}
4249
4250#ifdef CONFIG_HIBERNATION
4251/*
4252 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4253 * freed pages.
4254 *
4255 * Rather than trying to age LRUs the aim is to preserve the overall
4256 * LRU order by reclaiming preferentially
4257 * inactive > active > active referenced > active mapped
4258 */
4259unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4260{
4261 struct scan_control sc = {
4262 .nr_to_reclaim = nr_to_reclaim,
4263 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4264 .reclaim_idx = MAX_NR_ZONES - 1,
4265 .priority = DEF_PRIORITY,
4266 .may_writepage = 1,
4267 .may_unmap = 1,
4268 .may_swap = 1,
4269 .hibernation_mode = 1,
4270 };
4271 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4272 unsigned long nr_reclaimed;
4273 unsigned int noreclaim_flag;
4274
4275 fs_reclaim_acquire(sc.gfp_mask);
4276 noreclaim_flag = memalloc_noreclaim_save();
4277 set_task_reclaim_state(current, &sc.reclaim_state);
4278
4279 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4280
4281 set_task_reclaim_state(current, NULL);
4282 memalloc_noreclaim_restore(noreclaim_flag);
4283 fs_reclaim_release(sc.gfp_mask);
4284
4285 return nr_reclaimed;
4286}
4287#endif /* CONFIG_HIBERNATION */
4288
4289/*
4290 * This kswapd start function will be called by init and node-hot-add.
4291 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4292 */
4293int kswapd_run(int nid)
4294{
4295 pg_data_t *pgdat = NODE_DATA(nid);
4296 int ret = 0;
4297
4298 if (pgdat->kswapd)
4299 return 0;
4300
4301 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4302 if (IS_ERR(pgdat->kswapd)) {
4303 /* failure at boot is fatal */
4304 BUG_ON(system_state < SYSTEM_RUNNING);
4305 pr_err("Failed to start kswapd on node %d\n", nid);
4306 ret = PTR_ERR(pgdat->kswapd);
4307 pgdat->kswapd = NULL;
4308 }
4309 return ret;
4310}
4311
4312/*
4313 * Called by memory hotplug when all memory in a node is offlined. Caller must
4314 * hold mem_hotplug_begin/end().
4315 */
4316void kswapd_stop(int nid)
4317{
4318 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4319
4320 if (kswapd) {
4321 kthread_stop(kswapd);
4322 NODE_DATA(nid)->kswapd = NULL;
4323 }
4324}
4325
4326static int __init kswapd_init(void)
4327{
4328 int nid;
4329
4330 swap_setup();
4331 for_each_node_state(nid, N_MEMORY)
4332 kswapd_run(nid);
4333 return 0;
4334}
4335
4336module_init(kswapd_init)
4337
4338#ifdef CONFIG_NUMA
4339/*
4340 * Node reclaim mode
4341 *
4342 * If non-zero call node_reclaim when the number of free pages falls below
4343 * the watermarks.
4344 */
4345int node_reclaim_mode __read_mostly;
4346
4347/*
4348 * Priority for NODE_RECLAIM. This determines the fraction of pages
4349 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4350 * a zone.
4351 */
4352#define NODE_RECLAIM_PRIORITY 4
4353
4354/*
4355 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4356 * occur.
4357 */
4358int sysctl_min_unmapped_ratio = 1;
4359
4360/*
4361 * If the number of slab pages in a zone grows beyond this percentage then
4362 * slab reclaim needs to occur.
4363 */
4364int sysctl_min_slab_ratio = 5;
4365
4366static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4367{
4368 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4369 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4370 node_page_state(pgdat, NR_ACTIVE_FILE);
4371
4372 /*
4373 * It's possible for there to be more file mapped pages than
4374 * accounted for by the pages on the file LRU lists because
4375 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4376 */
4377 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4378}
4379
4380/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4381static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4382{
4383 unsigned long nr_pagecache_reclaimable;
4384 unsigned long delta = 0;
4385
4386 /*
4387 * If RECLAIM_UNMAP is set, then all file pages are considered
4388 * potentially reclaimable. Otherwise, we have to worry about
4389 * pages like swapcache and node_unmapped_file_pages() provides
4390 * a better estimate
4391 */
4392 if (node_reclaim_mode & RECLAIM_UNMAP)
4393 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4394 else
4395 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4396
4397 /* If we can't clean pages, remove dirty pages from consideration */
4398 if (!(node_reclaim_mode & RECLAIM_WRITE))
4399 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4400
4401 /* Watch for any possible underflows due to delta */
4402 if (unlikely(delta > nr_pagecache_reclaimable))
4403 delta = nr_pagecache_reclaimable;
4404
4405 return nr_pagecache_reclaimable - delta;
4406}
4407
4408/*
4409 * Try to free up some pages from this node through reclaim.
4410 */
4411static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4412{
4413 /* Minimum pages needed in order to stay on node */
4414 const unsigned long nr_pages = 1 << order;
4415 struct task_struct *p = current;
4416 unsigned int noreclaim_flag;
4417 struct scan_control sc = {
4418 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4419 .gfp_mask = current_gfp_context(gfp_mask),
4420 .order = order,
4421 .priority = NODE_RECLAIM_PRIORITY,
4422 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4423 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4424 .may_swap = 1,
4425 .reclaim_idx = gfp_zone(gfp_mask),
4426 };
4427 unsigned long pflags;
4428
4429 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4430 sc.gfp_mask);
4431
4432 cond_resched();
4433 psi_memstall_enter(&pflags);
4434 fs_reclaim_acquire(sc.gfp_mask);
4435 /*
4436 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4437 * and we also need to be able to write out pages for RECLAIM_WRITE
4438 * and RECLAIM_UNMAP.
4439 */
4440 noreclaim_flag = memalloc_noreclaim_save();
4441 p->flags |= PF_SWAPWRITE;
4442 set_task_reclaim_state(p, &sc.reclaim_state);
4443
4444 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4445 /*
4446 * Free memory by calling shrink node with increasing
4447 * priorities until we have enough memory freed.
4448 */
4449 do {
4450 shrink_node(pgdat, &sc);
4451 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4452 }
4453
4454 set_task_reclaim_state(p, NULL);
4455 current->flags &= ~PF_SWAPWRITE;
4456 memalloc_noreclaim_restore(noreclaim_flag);
4457 fs_reclaim_release(sc.gfp_mask);
4458 psi_memstall_leave(&pflags);
4459
4460 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4461
4462 return sc.nr_reclaimed >= nr_pages;
4463}
4464
4465int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4466{
4467 int ret;
4468
4469 /*
4470 * Node reclaim reclaims unmapped file backed pages and
4471 * slab pages if we are over the defined limits.
4472 *
4473 * A small portion of unmapped file backed pages is needed for
4474 * file I/O otherwise pages read by file I/O will be immediately
4475 * thrown out if the node is overallocated. So we do not reclaim
4476 * if less than a specified percentage of the node is used by
4477 * unmapped file backed pages.
4478 */
4479 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4480 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4481 pgdat->min_slab_pages)
4482 return NODE_RECLAIM_FULL;
4483
4484 /*
4485 * Do not scan if the allocation should not be delayed.
4486 */
4487 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4488 return NODE_RECLAIM_NOSCAN;
4489
4490 /*
4491 * Only run node reclaim on the local node or on nodes that do not
4492 * have associated processors. This will favor the local processor
4493 * over remote processors and spread off node memory allocations
4494 * as wide as possible.
4495 */
4496 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4497 return NODE_RECLAIM_NOSCAN;
4498
4499 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4500 return NODE_RECLAIM_NOSCAN;
4501
4502 ret = __node_reclaim(pgdat, gfp_mask, order);
4503 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4504
4505 if (!ret)
4506 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4507
4508 return ret;
4509}
4510#endif
4511
4512/**
4513 * check_move_unevictable_pages - check pages for evictability and move to
4514 * appropriate zone lru list
4515 * @pvec: pagevec with lru pages to check
4516 *
4517 * Checks pages for evictability, if an evictable page is in the unevictable
4518 * lru list, moves it to the appropriate evictable lru list. This function
4519 * should be only used for lru pages.
4520 */
4521void check_move_unevictable_pages(struct pagevec *pvec)
4522{
4523 struct lruvec *lruvec = NULL;
4524 int pgscanned = 0;
4525 int pgrescued = 0;
4526 int i;
4527
4528 for (i = 0; i < pvec->nr; i++) {
4529 struct page *page = pvec->pages[i];
4530 int nr_pages;
4531
4532 if (PageTransTail(page))
4533 continue;
4534
4535 nr_pages = thp_nr_pages(page);
4536 pgscanned += nr_pages;
4537
4538 /* block memcg migration during page moving between lru */
4539 if (!TestClearPageLRU(page))
4540 continue;
4541
4542 lruvec = relock_page_lruvec_irq(page, lruvec);
4543 if (page_evictable(page) && PageUnevictable(page)) {
4544 del_page_from_lru_list(page, lruvec);
4545 ClearPageUnevictable(page);
4546 add_page_to_lru_list(page, lruvec);
4547 pgrescued += nr_pages;
4548 }
4549 SetPageLRU(page);
4550 }
4551
4552 if (lruvec) {
4553 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4554 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4555 unlock_page_lruvec_irq(lruvec);
4556 } else if (pgscanned) {
4557 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4558 }
4559}
4560EXPORT_SYMBOL_GPL(check_move_unevictable_pages);