mm: workingset: tell cache transitions from workingset thrashing
[muen/linux.git] / mm / vmscan.c
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/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.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
64 struct 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         /* Writepage batching in laptop mode; RECLAIM_WRITE */
81         unsigned int may_writepage:1;
82
83         /* Can mapped pages be reclaimed? */
84         unsigned int may_unmap:1;
85
86         /* Can pages be swapped as part of reclaim? */
87         unsigned int may_swap:1;
88
89         /*
90          * Cgroups are not reclaimed below their configured memory.low,
91          * unless we threaten to OOM. If any cgroups are skipped due to
92          * memory.low and nothing was reclaimed, go back for memory.low.
93          */
94         unsigned int memcg_low_reclaim:1;
95         unsigned int memcg_low_skipped:1;
96
97         unsigned int hibernation_mode:1;
98
99         /* One of the zones is ready for compaction */
100         unsigned int compaction_ready:1;
101
102         /* Allocation order */
103         s8 order;
104
105         /* Scan (total_size >> priority) pages at once */
106         s8 priority;
107
108         /* The highest zone to isolate pages for reclaim from */
109         s8 reclaim_idx;
110
111         /* This context's GFP mask */
112         gfp_t gfp_mask;
113
114         /* Incremented by the number of inactive pages that were scanned */
115         unsigned long nr_scanned;
116
117         /* Number of pages freed so far during a call to shrink_zones() */
118         unsigned long nr_reclaimed;
119
120         struct {
121                 unsigned int dirty;
122                 unsigned int unqueued_dirty;
123                 unsigned int congested;
124                 unsigned int writeback;
125                 unsigned int immediate;
126                 unsigned int file_taken;
127                 unsigned int taken;
128         } nr;
129 };
130
131 #ifdef ARCH_HAS_PREFETCH
132 #define prefetch_prev_lru_page(_page, _base, _field)                    \
133         do {                                                            \
134                 if ((_page)->lru.prev != _base) {                       \
135                         struct page *prev;                              \
136                                                                         \
137                         prev = lru_to_page(&(_page->lru));              \
138                         prefetch(&prev->_field);                        \
139                 }                                                       \
140         } while (0)
141 #else
142 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
143 #endif
144
145 #ifdef ARCH_HAS_PREFETCHW
146 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
147         do {                                                            \
148                 if ((_page)->lru.prev != _base) {                       \
149                         struct page *prev;                              \
150                                                                         \
151                         prev = lru_to_page(&(_page->lru));              \
152                         prefetchw(&prev->_field);                       \
153                 }                                                       \
154         } while (0)
155 #else
156 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
157 #endif
158
159 /*
160  * From 0 .. 100.  Higher means more swappy.
161  */
162 int vm_swappiness = 60;
163 /*
164  * The total number of pages which are beyond the high watermark within all
165  * zones.
166  */
167 unsigned long vm_total_pages;
168
169 static LIST_HEAD(shrinker_list);
170 static DECLARE_RWSEM(shrinker_rwsem);
171
172 #ifdef CONFIG_MEMCG_KMEM
173
174 /*
175  * We allow subsystems to populate their shrinker-related
176  * LRU lists before register_shrinker_prepared() is called
177  * for the shrinker, since we don't want to impose
178  * restrictions on their internal registration order.
179  * In this case shrink_slab_memcg() may find corresponding
180  * bit is set in the shrinkers map.
181  *
182  * This value is used by the function to detect registering
183  * shrinkers and to skip do_shrink_slab() calls for them.
184  */
185 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
186
187 static DEFINE_IDR(shrinker_idr);
188 static int shrinker_nr_max;
189
190 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
191 {
192         int id, ret = -ENOMEM;
193
194         down_write(&shrinker_rwsem);
195         /* This may call shrinker, so it must use down_read_trylock() */
196         id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
197         if (id < 0)
198                 goto unlock;
199
200         if (id >= shrinker_nr_max) {
201                 if (memcg_expand_shrinker_maps(id)) {
202                         idr_remove(&shrinker_idr, id);
203                         goto unlock;
204                 }
205
206                 shrinker_nr_max = id + 1;
207         }
208         shrinker->id = id;
209         ret = 0;
210 unlock:
211         up_write(&shrinker_rwsem);
212         return ret;
213 }
214
215 static void unregister_memcg_shrinker(struct shrinker *shrinker)
216 {
217         int id = shrinker->id;
218
219         BUG_ON(id < 0);
220
221         down_write(&shrinker_rwsem);
222         idr_remove(&shrinker_idr, id);
223         up_write(&shrinker_rwsem);
224 }
225 #else /* CONFIG_MEMCG_KMEM */
226 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
227 {
228         return 0;
229 }
230
231 static void unregister_memcg_shrinker(struct shrinker *shrinker)
232 {
233 }
234 #endif /* CONFIG_MEMCG_KMEM */
235
236 #ifdef CONFIG_MEMCG
237 static bool global_reclaim(struct scan_control *sc)
238 {
239         return !sc->target_mem_cgroup;
240 }
241
242 /**
243  * sane_reclaim - is the usual dirty throttling mechanism operational?
244  * @sc: scan_control in question
245  *
246  * The normal page dirty throttling mechanism in balance_dirty_pages() is
247  * completely broken with the legacy memcg and direct stalling in
248  * shrink_page_list() is used for throttling instead, which lacks all the
249  * niceties such as fairness, adaptive pausing, bandwidth proportional
250  * allocation and configurability.
251  *
252  * This function tests whether the vmscan currently in progress can assume
253  * that the normal dirty throttling mechanism is operational.
254  */
255 static bool sane_reclaim(struct scan_control *sc)
256 {
257         struct mem_cgroup *memcg = sc->target_mem_cgroup;
258
259         if (!memcg)
260                 return true;
261 #ifdef CONFIG_CGROUP_WRITEBACK
262         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
263                 return true;
264 #endif
265         return false;
266 }
267
268 static void set_memcg_congestion(pg_data_t *pgdat,
269                                 struct mem_cgroup *memcg,
270                                 bool congested)
271 {
272         struct mem_cgroup_per_node *mn;
273
274         if (!memcg)
275                 return;
276
277         mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
278         WRITE_ONCE(mn->congested, congested);
279 }
280
281 static bool memcg_congested(pg_data_t *pgdat,
282                         struct mem_cgroup *memcg)
283 {
284         struct mem_cgroup_per_node *mn;
285
286         mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
287         return READ_ONCE(mn->congested);
288
289 }
290 #else
291 static bool global_reclaim(struct scan_control *sc)
292 {
293         return true;
294 }
295
296 static bool sane_reclaim(struct scan_control *sc)
297 {
298         return true;
299 }
300
301 static inline void set_memcg_congestion(struct pglist_data *pgdat,
302                                 struct mem_cgroup *memcg, bool congested)
303 {
304 }
305
306 static inline bool memcg_congested(struct pglist_data *pgdat,
307                         struct mem_cgroup *memcg)
308 {
309         return false;
310
311 }
312 #endif
313
314 /*
315  * This misses isolated pages which are not accounted for to save counters.
316  * As the data only determines if reclaim or compaction continues, it is
317  * not expected that isolated pages will be a dominating factor.
318  */
319 unsigned long zone_reclaimable_pages(struct zone *zone)
320 {
321         unsigned long nr;
322
323         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
324                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
325         if (get_nr_swap_pages() > 0)
326                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
327                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
328
329         return nr;
330 }
331
332 /**
333  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
334  * @lruvec: lru vector
335  * @lru: lru to use
336  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
337  */
338 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
339 {
340         unsigned long lru_size;
341         int zid;
342
343         if (!mem_cgroup_disabled())
344                 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
345         else
346                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
347
348         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
349                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
350                 unsigned long size;
351
352                 if (!managed_zone(zone))
353                         continue;
354
355                 if (!mem_cgroup_disabled())
356                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
357                 else
358                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
359                                        NR_ZONE_LRU_BASE + lru);
360                 lru_size -= min(size, lru_size);
361         }
362
363         return lru_size;
364
365 }
366
367 /*
368  * Add a shrinker callback to be called from the vm.
369  */
370 int prealloc_shrinker(struct shrinker *shrinker)
371 {
372         size_t size = sizeof(*shrinker->nr_deferred);
373
374         if (shrinker->flags & SHRINKER_NUMA_AWARE)
375                 size *= nr_node_ids;
376
377         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
378         if (!shrinker->nr_deferred)
379                 return -ENOMEM;
380
381         if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
382                 if (prealloc_memcg_shrinker(shrinker))
383                         goto free_deferred;
384         }
385
386         return 0;
387
388 free_deferred:
389         kfree(shrinker->nr_deferred);
390         shrinker->nr_deferred = NULL;
391         return -ENOMEM;
392 }
393
394 void free_prealloced_shrinker(struct shrinker *shrinker)
395 {
396         if (!shrinker->nr_deferred)
397                 return;
398
399         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
400                 unregister_memcg_shrinker(shrinker);
401
402         kfree(shrinker->nr_deferred);
403         shrinker->nr_deferred = NULL;
404 }
405
406 void register_shrinker_prepared(struct shrinker *shrinker)
407 {
408         down_write(&shrinker_rwsem);
409         list_add_tail(&shrinker->list, &shrinker_list);
410 #ifdef CONFIG_MEMCG_KMEM
411         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
412                 idr_replace(&shrinker_idr, shrinker, shrinker->id);
413 #endif
414         up_write(&shrinker_rwsem);
415 }
416
417 int register_shrinker(struct shrinker *shrinker)
418 {
419         int err = prealloc_shrinker(shrinker);
420
421         if (err)
422                 return err;
423         register_shrinker_prepared(shrinker);
424         return 0;
425 }
426 EXPORT_SYMBOL(register_shrinker);
427
428 /*
429  * Remove one
430  */
431 void unregister_shrinker(struct shrinker *shrinker)
432 {
433         if (!shrinker->nr_deferred)
434                 return;
435         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
436                 unregister_memcg_shrinker(shrinker);
437         down_write(&shrinker_rwsem);
438         list_del(&shrinker->list);
439         up_write(&shrinker_rwsem);
440         kfree(shrinker->nr_deferred);
441         shrinker->nr_deferred = NULL;
442 }
443 EXPORT_SYMBOL(unregister_shrinker);
444
445 #define SHRINK_BATCH 128
446
447 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
448                                     struct shrinker *shrinker, int priority)
449 {
450         unsigned long freed = 0;
451         unsigned long long delta;
452         long total_scan;
453         long freeable;
454         long nr;
455         long new_nr;
456         int nid = shrinkctl->nid;
457         long batch_size = shrinker->batch ? shrinker->batch
458                                           : SHRINK_BATCH;
459         long scanned = 0, next_deferred;
460
461         if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
462                 nid = 0;
463
464         freeable = shrinker->count_objects(shrinker, shrinkctl);
465         if (freeable == 0 || freeable == SHRINK_EMPTY)
466                 return freeable;
467
468         /*
469          * copy the current shrinker scan count into a local variable
470          * and zero it so that other concurrent shrinker invocations
471          * don't also do this scanning work.
472          */
473         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
474
475         total_scan = nr;
476         delta = freeable >> priority;
477         delta *= 4;
478         do_div(delta, shrinker->seeks);
479
480         /*
481          * Make sure we apply some minimal pressure on default priority
482          * even on small cgroups. Stale objects are not only consuming memory
483          * by themselves, but can also hold a reference to a dying cgroup,
484          * preventing it from being reclaimed. A dying cgroup with all
485          * corresponding structures like per-cpu stats and kmem caches
486          * can be really big, so it may lead to a significant waste of memory.
487          */
488         delta = max_t(unsigned long long, delta, min(freeable, batch_size));
489
490         total_scan += delta;
491         if (total_scan < 0) {
492                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
493                        shrinker->scan_objects, total_scan);
494                 total_scan = freeable;
495                 next_deferred = nr;
496         } else
497                 next_deferred = total_scan;
498
499         /*
500          * We need to avoid excessive windup on filesystem shrinkers
501          * due to large numbers of GFP_NOFS allocations causing the
502          * shrinkers to return -1 all the time. This results in a large
503          * nr being built up so when a shrink that can do some work
504          * comes along it empties the entire cache due to nr >>>
505          * freeable. This is bad for sustaining a working set in
506          * memory.
507          *
508          * Hence only allow the shrinker to scan the entire cache when
509          * a large delta change is calculated directly.
510          */
511         if (delta < freeable / 4)
512                 total_scan = min(total_scan, freeable / 2);
513
514         /*
515          * Avoid risking looping forever due to too large nr value:
516          * never try to free more than twice the estimate number of
517          * freeable entries.
518          */
519         if (total_scan > freeable * 2)
520                 total_scan = freeable * 2;
521
522         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
523                                    freeable, delta, total_scan, priority);
524
525         /*
526          * Normally, we should not scan less than batch_size objects in one
527          * pass to avoid too frequent shrinker calls, but if the slab has less
528          * than batch_size objects in total and we are really tight on memory,
529          * we will try to reclaim all available objects, otherwise we can end
530          * up failing allocations although there are plenty of reclaimable
531          * objects spread over several slabs with usage less than the
532          * batch_size.
533          *
534          * We detect the "tight on memory" situations by looking at the total
535          * number of objects we want to scan (total_scan). If it is greater
536          * than the total number of objects on slab (freeable), we must be
537          * scanning at high prio and therefore should try to reclaim as much as
538          * possible.
539          */
540         while (total_scan >= batch_size ||
541                total_scan >= freeable) {
542                 unsigned long ret;
543                 unsigned long nr_to_scan = min(batch_size, total_scan);
544
545                 shrinkctl->nr_to_scan = nr_to_scan;
546                 shrinkctl->nr_scanned = nr_to_scan;
547                 ret = shrinker->scan_objects(shrinker, shrinkctl);
548                 if (ret == SHRINK_STOP)
549                         break;
550                 freed += ret;
551
552                 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
553                 total_scan -= shrinkctl->nr_scanned;
554                 scanned += shrinkctl->nr_scanned;
555
556                 cond_resched();
557         }
558
559         if (next_deferred >= scanned)
560                 next_deferred -= scanned;
561         else
562                 next_deferred = 0;
563         /*
564          * move the unused scan count back into the shrinker in a
565          * manner that handles concurrent updates. If we exhausted the
566          * scan, there is no need to do an update.
567          */
568         if (next_deferred > 0)
569                 new_nr = atomic_long_add_return(next_deferred,
570                                                 &shrinker->nr_deferred[nid]);
571         else
572                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
573
574         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
575         return freed;
576 }
577
578 #ifdef CONFIG_MEMCG_KMEM
579 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
580                         struct mem_cgroup *memcg, int priority)
581 {
582         struct memcg_shrinker_map *map;
583         unsigned long ret, freed = 0;
584         int i;
585
586         if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
587                 return 0;
588
589         if (!down_read_trylock(&shrinker_rwsem))
590                 return 0;
591
592         map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
593                                         true);
594         if (unlikely(!map))
595                 goto unlock;
596
597         for_each_set_bit(i, map->map, shrinker_nr_max) {
598                 struct shrink_control sc = {
599                         .gfp_mask = gfp_mask,
600                         .nid = nid,
601                         .memcg = memcg,
602                 };
603                 struct shrinker *shrinker;
604
605                 shrinker = idr_find(&shrinker_idr, i);
606                 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
607                         if (!shrinker)
608                                 clear_bit(i, map->map);
609                         continue;
610                 }
611
612                 ret = do_shrink_slab(&sc, shrinker, priority);
613                 if (ret == SHRINK_EMPTY) {
614                         clear_bit(i, map->map);
615                         /*
616                          * After the shrinker reported that it had no objects to
617                          * free, but before we cleared the corresponding bit in
618                          * the memcg shrinker map, a new object might have been
619                          * added. To make sure, we have the bit set in this
620                          * case, we invoke the shrinker one more time and reset
621                          * the bit if it reports that it is not empty anymore.
622                          * The memory barrier here pairs with the barrier in
623                          * memcg_set_shrinker_bit():
624                          *
625                          * list_lru_add()     shrink_slab_memcg()
626                          *   list_add_tail()    clear_bit()
627                          *   <MB>               <MB>
628                          *   set_bit()          do_shrink_slab()
629                          */
630                         smp_mb__after_atomic();
631                         ret = do_shrink_slab(&sc, shrinker, priority);
632                         if (ret == SHRINK_EMPTY)
633                                 ret = 0;
634                         else
635                                 memcg_set_shrinker_bit(memcg, nid, i);
636                 }
637                 freed += ret;
638
639                 if (rwsem_is_contended(&shrinker_rwsem)) {
640                         freed = freed ? : 1;
641                         break;
642                 }
643         }
644 unlock:
645         up_read(&shrinker_rwsem);
646         return freed;
647 }
648 #else /* CONFIG_MEMCG_KMEM */
649 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
650                         struct mem_cgroup *memcg, int priority)
651 {
652         return 0;
653 }
654 #endif /* CONFIG_MEMCG_KMEM */
655
656 /**
657  * shrink_slab - shrink slab caches
658  * @gfp_mask: allocation context
659  * @nid: node whose slab caches to target
660  * @memcg: memory cgroup whose slab caches to target
661  * @priority: the reclaim priority
662  *
663  * Call the shrink functions to age shrinkable caches.
664  *
665  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
666  * unaware shrinkers will receive a node id of 0 instead.
667  *
668  * @memcg specifies the memory cgroup to target. Unaware shrinkers
669  * are called only if it is the root cgroup.
670  *
671  * @priority is sc->priority, we take the number of objects and >> by priority
672  * in order to get the scan target.
673  *
674  * Returns the number of reclaimed slab objects.
675  */
676 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
677                                  struct mem_cgroup *memcg,
678                                  int priority)
679 {
680         unsigned long ret, freed = 0;
681         struct shrinker *shrinker;
682
683         if (!mem_cgroup_is_root(memcg))
684                 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
685
686         if (!down_read_trylock(&shrinker_rwsem))
687                 goto out;
688
689         list_for_each_entry(shrinker, &shrinker_list, list) {
690                 struct shrink_control sc = {
691                         .gfp_mask = gfp_mask,
692                         .nid = nid,
693                         .memcg = memcg,
694                 };
695
696                 ret = do_shrink_slab(&sc, shrinker, priority);
697                 if (ret == SHRINK_EMPTY)
698                         ret = 0;
699                 freed += ret;
700                 /*
701                  * Bail out if someone want to register a new shrinker to
702                  * prevent the regsitration from being stalled for long periods
703                  * by parallel ongoing shrinking.
704                  */
705                 if (rwsem_is_contended(&shrinker_rwsem)) {
706                         freed = freed ? : 1;
707                         break;
708                 }
709         }
710
711         up_read(&shrinker_rwsem);
712 out:
713         cond_resched();
714         return freed;
715 }
716
717 void drop_slab_node(int nid)
718 {
719         unsigned long freed;
720
721         do {
722                 struct mem_cgroup *memcg = NULL;
723
724                 freed = 0;
725                 memcg = mem_cgroup_iter(NULL, NULL, NULL);
726                 do {
727                         freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
728                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
729         } while (freed > 10);
730 }
731
732 void drop_slab(void)
733 {
734         int nid;
735
736         for_each_online_node(nid)
737                 drop_slab_node(nid);
738 }
739
740 static inline int is_page_cache_freeable(struct page *page)
741 {
742         /*
743          * A freeable page cache page is referenced only by the caller
744          * that isolated the page, the page cache radix tree and
745          * optional buffer heads at page->private.
746          */
747         int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
748                 HPAGE_PMD_NR : 1;
749         return page_count(page) - page_has_private(page) == 1 + radix_pins;
750 }
751
752 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
753 {
754         if (current->flags & PF_SWAPWRITE)
755                 return 1;
756         if (!inode_write_congested(inode))
757                 return 1;
758         if (inode_to_bdi(inode) == current->backing_dev_info)
759                 return 1;
760         return 0;
761 }
762
763 /*
764  * We detected a synchronous write error writing a page out.  Probably
765  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
766  * fsync(), msync() or close().
767  *
768  * The tricky part is that after writepage we cannot touch the mapping: nothing
769  * prevents it from being freed up.  But we have a ref on the page and once
770  * that page is locked, the mapping is pinned.
771  *
772  * We're allowed to run sleeping lock_page() here because we know the caller has
773  * __GFP_FS.
774  */
775 static void handle_write_error(struct address_space *mapping,
776                                 struct page *page, int error)
777 {
778         lock_page(page);
779         if (page_mapping(page) == mapping)
780                 mapping_set_error(mapping, error);
781         unlock_page(page);
782 }
783
784 /* possible outcome of pageout() */
785 typedef enum {
786         /* failed to write page out, page is locked */
787         PAGE_KEEP,
788         /* move page to the active list, page is locked */
789         PAGE_ACTIVATE,
790         /* page has been sent to the disk successfully, page is unlocked */
791         PAGE_SUCCESS,
792         /* page is clean and locked */
793         PAGE_CLEAN,
794 } pageout_t;
795
796 /*
797  * pageout is called by shrink_page_list() for each dirty page.
798  * Calls ->writepage().
799  */
800 static pageout_t pageout(struct page *page, struct address_space *mapping,
801                          struct scan_control *sc)
802 {
803         /*
804          * If the page is dirty, only perform writeback if that write
805          * will be non-blocking.  To prevent this allocation from being
806          * stalled by pagecache activity.  But note that there may be
807          * stalls if we need to run get_block().  We could test
808          * PagePrivate for that.
809          *
810          * If this process is currently in __generic_file_write_iter() against
811          * this page's queue, we can perform writeback even if that
812          * will block.
813          *
814          * If the page is swapcache, write it back even if that would
815          * block, for some throttling. This happens by accident, because
816          * swap_backing_dev_info is bust: it doesn't reflect the
817          * congestion state of the swapdevs.  Easy to fix, if needed.
818          */
819         if (!is_page_cache_freeable(page))
820                 return PAGE_KEEP;
821         if (!mapping) {
822                 /*
823                  * Some data journaling orphaned pages can have
824                  * page->mapping == NULL while being dirty with clean buffers.
825                  */
826                 if (page_has_private(page)) {
827                         if (try_to_free_buffers(page)) {
828                                 ClearPageDirty(page);
829                                 pr_info("%s: orphaned page\n", __func__);
830                                 return PAGE_CLEAN;
831                         }
832                 }
833                 return PAGE_KEEP;
834         }
835         if (mapping->a_ops->writepage == NULL)
836                 return PAGE_ACTIVATE;
837         if (!may_write_to_inode(mapping->host, sc))
838                 return PAGE_KEEP;
839
840         if (clear_page_dirty_for_io(page)) {
841                 int res;
842                 struct writeback_control wbc = {
843                         .sync_mode = WB_SYNC_NONE,
844                         .nr_to_write = SWAP_CLUSTER_MAX,
845                         .range_start = 0,
846                         .range_end = LLONG_MAX,
847                         .for_reclaim = 1,
848                 };
849
850                 SetPageReclaim(page);
851                 res = mapping->a_ops->writepage(page, &wbc);
852                 if (res < 0)
853                         handle_write_error(mapping, page, res);
854                 if (res == AOP_WRITEPAGE_ACTIVATE) {
855                         ClearPageReclaim(page);
856                         return PAGE_ACTIVATE;
857                 }
858
859                 if (!PageWriteback(page)) {
860                         /* synchronous write or broken a_ops? */
861                         ClearPageReclaim(page);
862                 }
863                 trace_mm_vmscan_writepage(page);
864                 inc_node_page_state(page, NR_VMSCAN_WRITE);
865                 return PAGE_SUCCESS;
866         }
867
868         return PAGE_CLEAN;
869 }
870
871 /*
872  * Same as remove_mapping, but if the page is removed from the mapping, it
873  * gets returned with a refcount of 0.
874  */
875 static int __remove_mapping(struct address_space *mapping, struct page *page,
876                             bool reclaimed)
877 {
878         unsigned long flags;
879         int refcount;
880
881         BUG_ON(!PageLocked(page));
882         BUG_ON(mapping != page_mapping(page));
883
884         xa_lock_irqsave(&mapping->i_pages, flags);
885         /*
886          * The non racy check for a busy page.
887          *
888          * Must be careful with the order of the tests. When someone has
889          * a ref to the page, it may be possible that they dirty it then
890          * drop the reference. So if PageDirty is tested before page_count
891          * here, then the following race may occur:
892          *
893          * get_user_pages(&page);
894          * [user mapping goes away]
895          * write_to(page);
896          *                              !PageDirty(page)    [good]
897          * SetPageDirty(page);
898          * put_page(page);
899          *                              !page_count(page)   [good, discard it]
900          *
901          * [oops, our write_to data is lost]
902          *
903          * Reversing the order of the tests ensures such a situation cannot
904          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
905          * load is not satisfied before that of page->_refcount.
906          *
907          * Note that if SetPageDirty is always performed via set_page_dirty,
908          * and thus under the i_pages lock, then this ordering is not required.
909          */
910         if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
911                 refcount = 1 + HPAGE_PMD_NR;
912         else
913                 refcount = 2;
914         if (!page_ref_freeze(page, refcount))
915                 goto cannot_free;
916         /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
917         if (unlikely(PageDirty(page))) {
918                 page_ref_unfreeze(page, refcount);
919                 goto cannot_free;
920         }
921
922         if (PageSwapCache(page)) {
923                 swp_entry_t swap = { .val = page_private(page) };
924                 mem_cgroup_swapout(page, swap);
925                 __delete_from_swap_cache(page);
926                 xa_unlock_irqrestore(&mapping->i_pages, flags);
927                 put_swap_page(page, swap);
928         } else {
929                 void (*freepage)(struct page *);
930                 void *shadow = NULL;
931
932                 freepage = mapping->a_ops->freepage;
933                 /*
934                  * Remember a shadow entry for reclaimed file cache in
935                  * order to detect refaults, thus thrashing, later on.
936                  *
937                  * But don't store shadows in an address space that is
938                  * already exiting.  This is not just an optizimation,
939                  * inode reclaim needs to empty out the radix tree or
940                  * the nodes are lost.  Don't plant shadows behind its
941                  * back.
942                  *
943                  * We also don't store shadows for DAX mappings because the
944                  * only page cache pages found in these are zero pages
945                  * covering holes, and because we don't want to mix DAX
946                  * exceptional entries and shadow exceptional entries in the
947                  * same address_space.
948                  */
949                 if (reclaimed && page_is_file_cache(page) &&
950                     !mapping_exiting(mapping) && !dax_mapping(mapping))
951                         shadow = workingset_eviction(mapping, page);
952                 __delete_from_page_cache(page, shadow);
953                 xa_unlock_irqrestore(&mapping->i_pages, flags);
954
955                 if (freepage != NULL)
956                         freepage(page);
957         }
958
959         return 1;
960
961 cannot_free:
962         xa_unlock_irqrestore(&mapping->i_pages, flags);
963         return 0;
964 }
965
966 /*
967  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
968  * someone else has a ref on the page, abort and return 0.  If it was
969  * successfully detached, return 1.  Assumes the caller has a single ref on
970  * this page.
971  */
972 int remove_mapping(struct address_space *mapping, struct page *page)
973 {
974         if (__remove_mapping(mapping, page, false)) {
975                 /*
976                  * Unfreezing the refcount with 1 rather than 2 effectively
977                  * drops the pagecache ref for us without requiring another
978                  * atomic operation.
979                  */
980                 page_ref_unfreeze(page, 1);
981                 return 1;
982         }
983         return 0;
984 }
985
986 /**
987  * putback_lru_page - put previously isolated page onto appropriate LRU list
988  * @page: page to be put back to appropriate lru list
989  *
990  * Add previously isolated @page to appropriate LRU list.
991  * Page may still be unevictable for other reasons.
992  *
993  * lru_lock must not be held, interrupts must be enabled.
994  */
995 void putback_lru_page(struct page *page)
996 {
997         lru_cache_add(page);
998         put_page(page);         /* drop ref from isolate */
999 }
1000
1001 enum page_references {
1002         PAGEREF_RECLAIM,
1003         PAGEREF_RECLAIM_CLEAN,
1004         PAGEREF_KEEP,
1005         PAGEREF_ACTIVATE,
1006 };
1007
1008 static enum page_references page_check_references(struct page *page,
1009                                                   struct scan_control *sc)
1010 {
1011         int referenced_ptes, referenced_page;
1012         unsigned long vm_flags;
1013
1014         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1015                                           &vm_flags);
1016         referenced_page = TestClearPageReferenced(page);
1017
1018         /*
1019          * Mlock lost the isolation race with us.  Let try_to_unmap()
1020          * move the page to the unevictable list.
1021          */
1022         if (vm_flags & VM_LOCKED)
1023                 return PAGEREF_RECLAIM;
1024
1025         if (referenced_ptes) {
1026                 if (PageSwapBacked(page))
1027                         return PAGEREF_ACTIVATE;
1028                 /*
1029                  * All mapped pages start out with page table
1030                  * references from the instantiating fault, so we need
1031                  * to look twice if a mapped file page is used more
1032                  * than once.
1033                  *
1034                  * Mark it and spare it for another trip around the
1035                  * inactive list.  Another page table reference will
1036                  * lead to its activation.
1037                  *
1038                  * Note: the mark is set for activated pages as well
1039                  * so that recently deactivated but used pages are
1040                  * quickly recovered.
1041                  */
1042                 SetPageReferenced(page);
1043
1044                 if (referenced_page || referenced_ptes > 1)
1045                         return PAGEREF_ACTIVATE;
1046
1047                 /*
1048                  * Activate file-backed executable pages after first usage.
1049                  */
1050                 if (vm_flags & VM_EXEC)
1051                         return PAGEREF_ACTIVATE;
1052
1053                 return PAGEREF_KEEP;
1054         }
1055
1056         /* Reclaim if clean, defer dirty pages to writeback */
1057         if (referenced_page && !PageSwapBacked(page))
1058                 return PAGEREF_RECLAIM_CLEAN;
1059
1060         return PAGEREF_RECLAIM;
1061 }
1062
1063 /* Check if a page is dirty or under writeback */
1064 static void page_check_dirty_writeback(struct page *page,
1065                                        bool *dirty, bool *writeback)
1066 {
1067         struct address_space *mapping;
1068
1069         /*
1070          * Anonymous pages are not handled by flushers and must be written
1071          * from reclaim context. Do not stall reclaim based on them
1072          */
1073         if (!page_is_file_cache(page) ||
1074             (PageAnon(page) && !PageSwapBacked(page))) {
1075                 *dirty = false;
1076                 *writeback = false;
1077                 return;
1078         }
1079
1080         /* By default assume that the page flags are accurate */
1081         *dirty = PageDirty(page);
1082         *writeback = PageWriteback(page);
1083
1084         /* Verify dirty/writeback state if the filesystem supports it */
1085         if (!page_has_private(page))
1086                 return;
1087
1088         mapping = page_mapping(page);
1089         if (mapping && mapping->a_ops->is_dirty_writeback)
1090                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1091 }
1092
1093 /*
1094  * shrink_page_list() returns the number of reclaimed pages
1095  */
1096 static unsigned long shrink_page_list(struct list_head *page_list,
1097                                       struct pglist_data *pgdat,
1098                                       struct scan_control *sc,
1099                                       enum ttu_flags ttu_flags,
1100                                       struct reclaim_stat *stat,
1101                                       bool force_reclaim)
1102 {
1103         LIST_HEAD(ret_pages);
1104         LIST_HEAD(free_pages);
1105         int pgactivate = 0;
1106         unsigned nr_unqueued_dirty = 0;
1107         unsigned nr_dirty = 0;
1108         unsigned nr_congested = 0;
1109         unsigned nr_reclaimed = 0;
1110         unsigned nr_writeback = 0;
1111         unsigned nr_immediate = 0;
1112         unsigned nr_ref_keep = 0;
1113         unsigned nr_unmap_fail = 0;
1114
1115         cond_resched();
1116
1117         while (!list_empty(page_list)) {
1118                 struct address_space *mapping;
1119                 struct page *page;
1120                 int may_enter_fs;
1121                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1122                 bool dirty, writeback;
1123
1124                 cond_resched();
1125
1126                 page = lru_to_page(page_list);
1127                 list_del(&page->lru);
1128
1129                 if (!trylock_page(page))
1130                         goto keep;
1131
1132                 VM_BUG_ON_PAGE(PageActive(page), page);
1133
1134                 sc->nr_scanned++;
1135
1136                 if (unlikely(!page_evictable(page)))
1137                         goto activate_locked;
1138
1139                 if (!sc->may_unmap && page_mapped(page))
1140                         goto keep_locked;
1141
1142                 /* Double the slab pressure for mapped and swapcache pages */
1143                 if ((page_mapped(page) || PageSwapCache(page)) &&
1144                     !(PageAnon(page) && !PageSwapBacked(page)))
1145                         sc->nr_scanned++;
1146
1147                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1148                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1149
1150                 /*
1151                  * The number of dirty pages determines if a node is marked
1152                  * reclaim_congested which affects wait_iff_congested. kswapd
1153                  * will stall and start writing pages if the tail of the LRU
1154                  * is all dirty unqueued pages.
1155                  */
1156                 page_check_dirty_writeback(page, &dirty, &writeback);
1157                 if (dirty || writeback)
1158                         nr_dirty++;
1159
1160                 if (dirty && !writeback)
1161                         nr_unqueued_dirty++;
1162
1163                 /*
1164                  * Treat this page as congested if the underlying BDI is or if
1165                  * pages are cycling through the LRU so quickly that the
1166                  * pages marked for immediate reclaim are making it to the
1167                  * end of the LRU a second time.
1168                  */
1169                 mapping = page_mapping(page);
1170                 if (((dirty || writeback) && mapping &&
1171                      inode_write_congested(mapping->host)) ||
1172                     (writeback && PageReclaim(page)))
1173                         nr_congested++;
1174
1175                 /*
1176                  * If a page at the tail of the LRU is under writeback, there
1177                  * are three cases to consider.
1178                  *
1179                  * 1) If reclaim is encountering an excessive number of pages
1180                  *    under writeback and this page is both under writeback and
1181                  *    PageReclaim then it indicates that pages are being queued
1182                  *    for IO but are being recycled through the LRU before the
1183                  *    IO can complete. Waiting on the page itself risks an
1184                  *    indefinite stall if it is impossible to writeback the
1185                  *    page due to IO error or disconnected storage so instead
1186                  *    note that the LRU is being scanned too quickly and the
1187                  *    caller can stall after page list has been processed.
1188                  *
1189                  * 2) Global or new memcg reclaim encounters a page that is
1190                  *    not marked for immediate reclaim, or the caller does not
1191                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1192                  *    not to fs). In this case mark the page for immediate
1193                  *    reclaim and continue scanning.
1194                  *
1195                  *    Require may_enter_fs because we would wait on fs, which
1196                  *    may not have submitted IO yet. And the loop driver might
1197                  *    enter reclaim, and deadlock if it waits on a page for
1198                  *    which it is needed to do the write (loop masks off
1199                  *    __GFP_IO|__GFP_FS for this reason); but more thought
1200                  *    would probably show more reasons.
1201                  *
1202                  * 3) Legacy memcg encounters a page that is already marked
1203                  *    PageReclaim. memcg does not have any dirty pages
1204                  *    throttling so we could easily OOM just because too many
1205                  *    pages are in writeback and there is nothing else to
1206                  *    reclaim. Wait for the writeback to complete.
1207                  *
1208                  * In cases 1) and 2) we activate the pages to get them out of
1209                  * the way while we continue scanning for clean pages on the
1210                  * inactive list and refilling from the active list. The
1211                  * observation here is that waiting for disk writes is more
1212                  * expensive than potentially causing reloads down the line.
1213                  * Since they're marked for immediate reclaim, they won't put
1214                  * memory pressure on the cache working set any longer than it
1215                  * takes to write them to disk.
1216                  */
1217                 if (PageWriteback(page)) {
1218                         /* Case 1 above */
1219                         if (current_is_kswapd() &&
1220                             PageReclaim(page) &&
1221                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1222                                 nr_immediate++;
1223                                 goto activate_locked;
1224
1225                         /* Case 2 above */
1226                         } else if (sane_reclaim(sc) ||
1227                             !PageReclaim(page) || !may_enter_fs) {
1228                                 /*
1229                                  * This is slightly racy - end_page_writeback()
1230                                  * might have just cleared PageReclaim, then
1231                                  * setting PageReclaim here end up interpreted
1232                                  * as PageReadahead - but that does not matter
1233                                  * enough to care.  What we do want is for this
1234                                  * page to have PageReclaim set next time memcg
1235                                  * reclaim reaches the tests above, so it will
1236                                  * then wait_on_page_writeback() to avoid OOM;
1237                                  * and it's also appropriate in global reclaim.
1238                                  */
1239                                 SetPageReclaim(page);
1240                                 nr_writeback++;
1241                                 goto activate_locked;
1242
1243                         /* Case 3 above */
1244                         } else {
1245                                 unlock_page(page);
1246                                 wait_on_page_writeback(page);
1247                                 /* then go back and try same page again */
1248                                 list_add_tail(&page->lru, page_list);
1249                                 continue;
1250                         }
1251                 }
1252
1253                 if (!force_reclaim)
1254                         references = page_check_references(page, sc);
1255
1256                 switch (references) {
1257                 case PAGEREF_ACTIVATE:
1258                         goto activate_locked;
1259                 case PAGEREF_KEEP:
1260                         nr_ref_keep++;
1261                         goto keep_locked;
1262                 case PAGEREF_RECLAIM:
1263                 case PAGEREF_RECLAIM_CLEAN:
1264                         ; /* try to reclaim the page below */
1265                 }
1266
1267                 /*
1268                  * Anonymous process memory has backing store?
1269                  * Try to allocate it some swap space here.
1270                  * Lazyfree page could be freed directly
1271                  */
1272                 if (PageAnon(page) && PageSwapBacked(page)) {
1273                         if (!PageSwapCache(page)) {
1274                                 if (!(sc->gfp_mask & __GFP_IO))
1275                                         goto keep_locked;
1276                                 if (PageTransHuge(page)) {
1277                                         /* cannot split THP, skip it */
1278                                         if (!can_split_huge_page(page, NULL))
1279                                                 goto activate_locked;
1280                                         /*
1281                                          * Split pages without a PMD map right
1282                                          * away. Chances are some or all of the
1283                                          * tail pages can be freed without IO.
1284                                          */
1285                                         if (!compound_mapcount(page) &&
1286                                             split_huge_page_to_list(page,
1287                                                                     page_list))
1288                                                 goto activate_locked;
1289                                 }
1290                                 if (!add_to_swap(page)) {
1291                                         if (!PageTransHuge(page))
1292                                                 goto activate_locked;
1293                                         /* Fallback to swap normal pages */
1294                                         if (split_huge_page_to_list(page,
1295                                                                     page_list))
1296                                                 goto activate_locked;
1297 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1298                                         count_vm_event(THP_SWPOUT_FALLBACK);
1299 #endif
1300                                         if (!add_to_swap(page))
1301                                                 goto activate_locked;
1302                                 }
1303
1304                                 may_enter_fs = 1;
1305
1306                                 /* Adding to swap updated mapping */
1307                                 mapping = page_mapping(page);
1308                         }
1309                 } else if (unlikely(PageTransHuge(page))) {
1310                         /* Split file THP */
1311                         if (split_huge_page_to_list(page, page_list))
1312                                 goto keep_locked;
1313                 }
1314
1315                 /*
1316                  * The page is mapped into the page tables of one or more
1317                  * processes. Try to unmap it here.
1318                  */
1319                 if (page_mapped(page)) {
1320                         enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1321
1322                         if (unlikely(PageTransHuge(page)))
1323                                 flags |= TTU_SPLIT_HUGE_PMD;
1324                         if (!try_to_unmap(page, flags)) {
1325                                 nr_unmap_fail++;
1326                                 goto activate_locked;
1327                         }
1328                 }
1329
1330                 if (PageDirty(page)) {
1331                         /*
1332                          * Only kswapd can writeback filesystem pages
1333                          * to avoid risk of stack overflow. But avoid
1334                          * injecting inefficient single-page IO into
1335                          * flusher writeback as much as possible: only
1336                          * write pages when we've encountered many
1337                          * dirty pages, and when we've already scanned
1338                          * the rest of the LRU for clean pages and see
1339                          * the same dirty pages again (PageReclaim).
1340                          */
1341                         if (page_is_file_cache(page) &&
1342                             (!current_is_kswapd() || !PageReclaim(page) ||
1343                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1344                                 /*
1345                                  * Immediately reclaim when written back.
1346                                  * Similar in principal to deactivate_page()
1347                                  * except we already have the page isolated
1348                                  * and know it's dirty
1349                                  */
1350                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1351                                 SetPageReclaim(page);
1352
1353                                 goto activate_locked;
1354                         }
1355
1356                         if (references == PAGEREF_RECLAIM_CLEAN)
1357                                 goto keep_locked;
1358                         if (!may_enter_fs)
1359                                 goto keep_locked;
1360                         if (!sc->may_writepage)
1361                                 goto keep_locked;
1362
1363                         /*
1364                          * Page is dirty. Flush the TLB if a writable entry
1365                          * potentially exists to avoid CPU writes after IO
1366                          * starts and then write it out here.
1367                          */
1368                         try_to_unmap_flush_dirty();
1369                         switch (pageout(page, mapping, sc)) {
1370                         case PAGE_KEEP:
1371                                 goto keep_locked;
1372                         case PAGE_ACTIVATE:
1373                                 goto activate_locked;
1374                         case PAGE_SUCCESS:
1375                                 if (PageWriteback(page))
1376                                         goto keep;
1377                                 if (PageDirty(page))
1378                                         goto keep;
1379
1380                                 /*
1381                                  * A synchronous write - probably a ramdisk.  Go
1382                                  * ahead and try to reclaim the page.
1383                                  */
1384                                 if (!trylock_page(page))
1385                                         goto keep;
1386                                 if (PageDirty(page) || PageWriteback(page))
1387                                         goto keep_locked;
1388                                 mapping = page_mapping(page);
1389                         case PAGE_CLEAN:
1390                                 ; /* try to free the page below */
1391                         }
1392                 }
1393
1394                 /*
1395                  * If the page has buffers, try to free the buffer mappings
1396                  * associated with this page. If we succeed we try to free
1397                  * the page as well.
1398                  *
1399                  * We do this even if the page is PageDirty().
1400                  * try_to_release_page() does not perform I/O, but it is
1401                  * possible for a page to have PageDirty set, but it is actually
1402                  * clean (all its buffers are clean).  This happens if the
1403                  * buffers were written out directly, with submit_bh(). ext3
1404                  * will do this, as well as the blockdev mapping.
1405                  * try_to_release_page() will discover that cleanness and will
1406                  * drop the buffers and mark the page clean - it can be freed.
1407                  *
1408                  * Rarely, pages can have buffers and no ->mapping.  These are
1409                  * the pages which were not successfully invalidated in
1410                  * truncate_complete_page().  We try to drop those buffers here
1411                  * and if that worked, and the page is no longer mapped into
1412                  * process address space (page_count == 1) it can be freed.
1413                  * Otherwise, leave the page on the LRU so it is swappable.
1414                  */
1415                 if (page_has_private(page)) {
1416                         if (!try_to_release_page(page, sc->gfp_mask))
1417                                 goto activate_locked;
1418                         if (!mapping && page_count(page) == 1) {
1419                                 unlock_page(page);
1420                                 if (put_page_testzero(page))
1421                                         goto free_it;
1422                                 else {
1423                                         /*
1424                                          * rare race with speculative reference.
1425                                          * the speculative reference will free
1426                                          * this page shortly, so we may
1427                                          * increment nr_reclaimed here (and
1428                                          * leave it off the LRU).
1429                                          */
1430                                         nr_reclaimed++;
1431                                         continue;
1432                                 }
1433                         }
1434                 }
1435
1436                 if (PageAnon(page) && !PageSwapBacked(page)) {
1437                         /* follow __remove_mapping for reference */
1438                         if (!page_ref_freeze(page, 1))
1439                                 goto keep_locked;
1440                         if (PageDirty(page)) {
1441                                 page_ref_unfreeze(page, 1);
1442                                 goto keep_locked;
1443                         }
1444
1445                         count_vm_event(PGLAZYFREED);
1446                         count_memcg_page_event(page, PGLAZYFREED);
1447                 } else if (!mapping || !__remove_mapping(mapping, page, true))
1448                         goto keep_locked;
1449                 /*
1450                  * At this point, we have no other references and there is
1451                  * no way to pick any more up (removed from LRU, removed
1452                  * from pagecache). Can use non-atomic bitops now (and
1453                  * we obviously don't have to worry about waking up a process
1454                  * waiting on the page lock, because there are no references.
1455                  */
1456                 __ClearPageLocked(page);
1457 free_it:
1458                 nr_reclaimed++;
1459
1460                 /*
1461                  * Is there need to periodically free_page_list? It would
1462                  * appear not as the counts should be low
1463                  */
1464                 if (unlikely(PageTransHuge(page))) {
1465                         mem_cgroup_uncharge(page);
1466                         (*get_compound_page_dtor(page))(page);
1467                 } else
1468                         list_add(&page->lru, &free_pages);
1469                 continue;
1470
1471 activate_locked:
1472                 /* Not a candidate for swapping, so reclaim swap space. */
1473                 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1474                                                 PageMlocked(page)))
1475                         try_to_free_swap(page);
1476                 VM_BUG_ON_PAGE(PageActive(page), page);
1477                 if (!PageMlocked(page)) {
1478                         SetPageActive(page);
1479                         pgactivate++;
1480                         count_memcg_page_event(page, PGACTIVATE);
1481                 }
1482 keep_locked:
1483                 unlock_page(page);
1484 keep:
1485                 list_add(&page->lru, &ret_pages);
1486                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1487         }
1488
1489         mem_cgroup_uncharge_list(&free_pages);
1490         try_to_unmap_flush();
1491         free_unref_page_list(&free_pages);
1492
1493         list_splice(&ret_pages, page_list);
1494         count_vm_events(PGACTIVATE, pgactivate);
1495
1496         if (stat) {
1497                 stat->nr_dirty = nr_dirty;
1498                 stat->nr_congested = nr_congested;
1499                 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1500                 stat->nr_writeback = nr_writeback;
1501                 stat->nr_immediate = nr_immediate;
1502                 stat->nr_activate = pgactivate;
1503                 stat->nr_ref_keep = nr_ref_keep;
1504                 stat->nr_unmap_fail = nr_unmap_fail;
1505         }
1506         return nr_reclaimed;
1507 }
1508
1509 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1510                                             struct list_head *page_list)
1511 {
1512         struct scan_control sc = {
1513                 .gfp_mask = GFP_KERNEL,
1514                 .priority = DEF_PRIORITY,
1515                 .may_unmap = 1,
1516         };
1517         unsigned long ret;
1518         struct page *page, *next;
1519         LIST_HEAD(clean_pages);
1520
1521         list_for_each_entry_safe(page, next, page_list, lru) {
1522                 if (page_is_file_cache(page) && !PageDirty(page) &&
1523                     !__PageMovable(page)) {
1524                         ClearPageActive(page);
1525                         list_move(&page->lru, &clean_pages);
1526                 }
1527         }
1528
1529         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1530                         TTU_IGNORE_ACCESS, NULL, true);
1531         list_splice(&clean_pages, page_list);
1532         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1533         return ret;
1534 }
1535
1536 /*
1537  * Attempt to remove the specified page from its LRU.  Only take this page
1538  * if it is of the appropriate PageActive status.  Pages which are being
1539  * freed elsewhere are also ignored.
1540  *
1541  * page:        page to consider
1542  * mode:        one of the LRU isolation modes defined above
1543  *
1544  * returns 0 on success, -ve errno on failure.
1545  */
1546 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1547 {
1548         int ret = -EINVAL;
1549
1550         /* Only take pages on the LRU. */
1551         if (!PageLRU(page))
1552                 return ret;
1553
1554         /* Compaction should not handle unevictable pages but CMA can do so */
1555         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1556                 return ret;
1557
1558         ret = -EBUSY;
1559
1560         /*
1561          * To minimise LRU disruption, the caller can indicate that it only
1562          * wants to isolate pages it will be able to operate on without
1563          * blocking - clean pages for the most part.
1564          *
1565          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1566          * that it is possible to migrate without blocking
1567          */
1568         if (mode & ISOLATE_ASYNC_MIGRATE) {
1569                 /* All the caller can do on PageWriteback is block */
1570                 if (PageWriteback(page))
1571                         return ret;
1572
1573                 if (PageDirty(page)) {
1574                         struct address_space *mapping;
1575                         bool migrate_dirty;
1576
1577                         /*
1578                          * Only pages without mappings or that have a
1579                          * ->migratepage callback are possible to migrate
1580                          * without blocking. However, we can be racing with
1581                          * truncation so it's necessary to lock the page
1582                          * to stabilise the mapping as truncation holds
1583                          * the page lock until after the page is removed
1584                          * from the page cache.
1585                          */
1586                         if (!trylock_page(page))
1587                                 return ret;
1588
1589                         mapping = page_mapping(page);
1590                         migrate_dirty = !mapping || mapping->a_ops->migratepage;
1591                         unlock_page(page);
1592                         if (!migrate_dirty)
1593                                 return ret;
1594                 }
1595         }
1596
1597         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1598                 return ret;
1599
1600         if (likely(get_page_unless_zero(page))) {
1601                 /*
1602                  * Be careful not to clear PageLRU until after we're
1603                  * sure the page is not being freed elsewhere -- the
1604                  * page release code relies on it.
1605                  */
1606                 ClearPageLRU(page);
1607                 ret = 0;
1608         }
1609
1610         return ret;
1611 }
1612
1613
1614 /*
1615  * Update LRU sizes after isolating pages. The LRU size updates must
1616  * be complete before mem_cgroup_update_lru_size due to a santity check.
1617  */
1618 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1619                         enum lru_list lru, unsigned long *nr_zone_taken)
1620 {
1621         int zid;
1622
1623         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1624                 if (!nr_zone_taken[zid])
1625                         continue;
1626
1627                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1628 #ifdef CONFIG_MEMCG
1629                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1630 #endif
1631         }
1632
1633 }
1634
1635 /*
1636  * zone_lru_lock is heavily contended.  Some of the functions that
1637  * shrink the lists perform better by taking out a batch of pages
1638  * and working on them outside the LRU lock.
1639  *
1640  * For pagecache intensive workloads, this function is the hottest
1641  * spot in the kernel (apart from copy_*_user functions).
1642  *
1643  * Appropriate locks must be held before calling this function.
1644  *
1645  * @nr_to_scan: The number of eligible pages to look through on the list.
1646  * @lruvec:     The LRU vector to pull pages from.
1647  * @dst:        The temp list to put pages on to.
1648  * @nr_scanned: The number of pages that were scanned.
1649  * @sc:         The scan_control struct for this reclaim session
1650  * @mode:       One of the LRU isolation modes
1651  * @lru:        LRU list id for isolating
1652  *
1653  * returns how many pages were moved onto *@dst.
1654  */
1655 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1656                 struct lruvec *lruvec, struct list_head *dst,
1657                 unsigned long *nr_scanned, struct scan_control *sc,
1658                 isolate_mode_t mode, enum lru_list lru)
1659 {
1660         struct list_head *src = &lruvec->lists[lru];
1661         unsigned long nr_taken = 0;
1662         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1663         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1664         unsigned long skipped = 0;
1665         unsigned long scan, total_scan, nr_pages;
1666         LIST_HEAD(pages_skipped);
1667
1668         scan = 0;
1669         for (total_scan = 0;
1670              scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1671              total_scan++) {
1672                 struct page *page;
1673
1674                 page = lru_to_page(src);
1675                 prefetchw_prev_lru_page(page, src, flags);
1676
1677                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1678
1679                 if (page_zonenum(page) > sc->reclaim_idx) {
1680                         list_move(&page->lru, &pages_skipped);
1681                         nr_skipped[page_zonenum(page)]++;
1682                         continue;
1683                 }
1684
1685                 /*
1686                  * Do not count skipped pages because that makes the function
1687                  * return with no isolated pages if the LRU mostly contains
1688                  * ineligible pages.  This causes the VM to not reclaim any
1689                  * pages, triggering a premature OOM.
1690                  */
1691                 scan++;
1692                 switch (__isolate_lru_page(page, mode)) {
1693                 case 0:
1694                         nr_pages = hpage_nr_pages(page);
1695                         nr_taken += nr_pages;
1696                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1697                         list_move(&page->lru, dst);
1698                         break;
1699
1700                 case -EBUSY:
1701                         /* else it is being freed elsewhere */
1702                         list_move(&page->lru, src);
1703                         continue;
1704
1705                 default:
1706                         BUG();
1707                 }
1708         }
1709
1710         /*
1711          * Splice any skipped pages to the start of the LRU list. Note that
1712          * this disrupts the LRU order when reclaiming for lower zones but
1713          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1714          * scanning would soon rescan the same pages to skip and put the
1715          * system at risk of premature OOM.
1716          */
1717         if (!list_empty(&pages_skipped)) {
1718                 int zid;
1719
1720                 list_splice(&pages_skipped, src);
1721                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1722                         if (!nr_skipped[zid])
1723                                 continue;
1724
1725                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1726                         skipped += nr_skipped[zid];
1727                 }
1728         }
1729         *nr_scanned = total_scan;
1730         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1731                                     total_scan, skipped, nr_taken, mode, lru);
1732         update_lru_sizes(lruvec, lru, nr_zone_taken);
1733         return nr_taken;
1734 }
1735
1736 /**
1737  * isolate_lru_page - tries to isolate a page from its LRU list
1738  * @page: page to isolate from its LRU list
1739  *
1740  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1741  * vmstat statistic corresponding to whatever LRU list the page was on.
1742  *
1743  * Returns 0 if the page was removed from an LRU list.
1744  * Returns -EBUSY if the page was not on an LRU list.
1745  *
1746  * The returned page will have PageLRU() cleared.  If it was found on
1747  * the active list, it will have PageActive set.  If it was found on
1748  * the unevictable list, it will have the PageUnevictable bit set. That flag
1749  * may need to be cleared by the caller before letting the page go.
1750  *
1751  * The vmstat statistic corresponding to the list on which the page was
1752  * found will be decremented.
1753  *
1754  * Restrictions:
1755  *
1756  * (1) Must be called with an elevated refcount on the page. This is a
1757  *     fundamentnal difference from isolate_lru_pages (which is called
1758  *     without a stable reference).
1759  * (2) the lru_lock must not be held.
1760  * (3) interrupts must be enabled.
1761  */
1762 int isolate_lru_page(struct page *page)
1763 {
1764         int ret = -EBUSY;
1765
1766         VM_BUG_ON_PAGE(!page_count(page), page);
1767         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1768
1769         if (PageLRU(page)) {
1770                 struct zone *zone = page_zone(page);
1771                 struct lruvec *lruvec;
1772
1773                 spin_lock_irq(zone_lru_lock(zone));
1774                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1775                 if (PageLRU(page)) {
1776                         int lru = page_lru(page);
1777                         get_page(page);
1778                         ClearPageLRU(page);
1779                         del_page_from_lru_list(page, lruvec, lru);
1780                         ret = 0;
1781                 }
1782                 spin_unlock_irq(zone_lru_lock(zone));
1783         }
1784         return ret;
1785 }
1786
1787 /*
1788  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1789  * then get resheduled. When there are massive number of tasks doing page
1790  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1791  * the LRU list will go small and be scanned faster than necessary, leading to
1792  * unnecessary swapping, thrashing and OOM.
1793  */
1794 static int too_many_isolated(struct pglist_data *pgdat, int file,
1795                 struct scan_control *sc)
1796 {
1797         unsigned long inactive, isolated;
1798
1799         if (current_is_kswapd())
1800                 return 0;
1801
1802         if (!sane_reclaim(sc))
1803                 return 0;
1804
1805         if (file) {
1806                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1807                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1808         } else {
1809                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1810                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1811         }
1812
1813         /*
1814          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1815          * won't get blocked by normal direct-reclaimers, forming a circular
1816          * deadlock.
1817          */
1818         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1819                 inactive >>= 3;
1820
1821         return isolated > inactive;
1822 }
1823
1824 static noinline_for_stack void
1825 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1826 {
1827         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1828         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1829         LIST_HEAD(pages_to_free);
1830
1831         /*
1832          * Put back any unfreeable pages.
1833          */
1834         while (!list_empty(page_list)) {
1835                 struct page *page = lru_to_page(page_list);
1836                 int lru;
1837
1838                 VM_BUG_ON_PAGE(PageLRU(page), page);
1839                 list_del(&page->lru);
1840                 if (unlikely(!page_evictable(page))) {
1841                         spin_unlock_irq(&pgdat->lru_lock);
1842                         putback_lru_page(page);
1843                         spin_lock_irq(&pgdat->lru_lock);
1844                         continue;
1845                 }
1846
1847                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1848
1849                 SetPageLRU(page);
1850                 lru = page_lru(page);
1851                 add_page_to_lru_list(page, lruvec, lru);
1852
1853                 if (is_active_lru(lru)) {
1854                         int file = is_file_lru(lru);
1855                         int numpages = hpage_nr_pages(page);
1856                         reclaim_stat->recent_rotated[file] += numpages;
1857                 }
1858                 if (put_page_testzero(page)) {
1859                         __ClearPageLRU(page);
1860                         __ClearPageActive(page);
1861                         del_page_from_lru_list(page, lruvec, lru);
1862
1863                         if (unlikely(PageCompound(page))) {
1864                                 spin_unlock_irq(&pgdat->lru_lock);
1865                                 mem_cgroup_uncharge(page);
1866                                 (*get_compound_page_dtor(page))(page);
1867                                 spin_lock_irq(&pgdat->lru_lock);
1868                         } else
1869                                 list_add(&page->lru, &pages_to_free);
1870                 }
1871         }
1872
1873         /*
1874          * To save our caller's stack, now use input list for pages to free.
1875          */
1876         list_splice(&pages_to_free, page_list);
1877 }
1878
1879 /*
1880  * If a kernel thread (such as nfsd for loop-back mounts) services
1881  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1882  * In that case we should only throttle if the backing device it is
1883  * writing to is congested.  In other cases it is safe to throttle.
1884  */
1885 static int current_may_throttle(void)
1886 {
1887         return !(current->flags & PF_LESS_THROTTLE) ||
1888                 current->backing_dev_info == NULL ||
1889                 bdi_write_congested(current->backing_dev_info);
1890 }
1891
1892 /*
1893  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1894  * of reclaimed pages
1895  */
1896 static noinline_for_stack unsigned long
1897 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1898                      struct scan_control *sc, enum lru_list lru)
1899 {
1900         LIST_HEAD(page_list);
1901         unsigned long nr_scanned;
1902         unsigned long nr_reclaimed = 0;
1903         unsigned long nr_taken;
1904         struct reclaim_stat stat = {};
1905         isolate_mode_t isolate_mode = 0;
1906         int file = is_file_lru(lru);
1907         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1908         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1909         bool stalled = false;
1910
1911         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1912                 if (stalled)
1913                         return 0;
1914
1915                 /* wait a bit for the reclaimer. */
1916                 msleep(100);
1917                 stalled = true;
1918
1919                 /* We are about to die and free our memory. Return now. */
1920                 if (fatal_signal_pending(current))
1921                         return SWAP_CLUSTER_MAX;
1922         }
1923
1924         lru_add_drain();
1925
1926         if (!sc->may_unmap)
1927                 isolate_mode |= ISOLATE_UNMAPPED;
1928
1929         spin_lock_irq(&pgdat->lru_lock);
1930
1931         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1932                                      &nr_scanned, sc, isolate_mode, lru);
1933
1934         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1935         reclaim_stat->recent_scanned[file] += nr_taken;
1936
1937         if (current_is_kswapd()) {
1938                 if (global_reclaim(sc))
1939                         __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1940                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1941                                    nr_scanned);
1942         } else {
1943                 if (global_reclaim(sc))
1944                         __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1945                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1946                                    nr_scanned);
1947         }
1948         spin_unlock_irq(&pgdat->lru_lock);
1949
1950         if (nr_taken == 0)
1951                 return 0;
1952
1953         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1954                                 &stat, false);
1955
1956         spin_lock_irq(&pgdat->lru_lock);
1957
1958         if (current_is_kswapd()) {
1959                 if (global_reclaim(sc))
1960                         __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1961                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1962                                    nr_reclaimed);
1963         } else {
1964                 if (global_reclaim(sc))
1965                         __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1966                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1967                                    nr_reclaimed);
1968         }
1969
1970         putback_inactive_pages(lruvec, &page_list);
1971
1972         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1973
1974         spin_unlock_irq(&pgdat->lru_lock);
1975
1976         mem_cgroup_uncharge_list(&page_list);
1977         free_unref_page_list(&page_list);
1978
1979         /*
1980          * If dirty pages are scanned that are not queued for IO, it
1981          * implies that flushers are not doing their job. This can
1982          * happen when memory pressure pushes dirty pages to the end of
1983          * the LRU before the dirty limits are breached and the dirty
1984          * data has expired. It can also happen when the proportion of
1985          * dirty pages grows not through writes but through memory
1986          * pressure reclaiming all the clean cache. And in some cases,
1987          * the flushers simply cannot keep up with the allocation
1988          * rate. Nudge the flusher threads in case they are asleep.
1989          */
1990         if (stat.nr_unqueued_dirty == nr_taken)
1991                 wakeup_flusher_threads(WB_REASON_VMSCAN);
1992
1993         sc->nr.dirty += stat.nr_dirty;
1994         sc->nr.congested += stat.nr_congested;
1995         sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1996         sc->nr.writeback += stat.nr_writeback;
1997         sc->nr.immediate += stat.nr_immediate;
1998         sc->nr.taken += nr_taken;
1999         if (file)
2000                 sc->nr.file_taken += nr_taken;
2001
2002         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2003                         nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2004         return nr_reclaimed;
2005 }
2006
2007 /*
2008  * This moves pages from the active list to the inactive list.
2009  *
2010  * We move them the other way if the page is referenced by one or more
2011  * processes, from rmap.
2012  *
2013  * If the pages are mostly unmapped, the processing is fast and it is
2014  * appropriate to hold zone_lru_lock across the whole operation.  But if
2015  * the pages are mapped, the processing is slow (page_referenced()) so we
2016  * should drop zone_lru_lock around each page.  It's impossible to balance
2017  * this, so instead we remove the pages from the LRU while processing them.
2018  * It is safe to rely on PG_active against the non-LRU pages in here because
2019  * nobody will play with that bit on a non-LRU page.
2020  *
2021  * The downside is that we have to touch page->_refcount against each page.
2022  * But we had to alter page->flags anyway.
2023  *
2024  * Returns the number of pages moved to the given lru.
2025  */
2026
2027 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
2028                                      struct list_head *list,
2029                                      struct list_head *pages_to_free,
2030                                      enum lru_list lru)
2031 {
2032         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2033         struct page *page;
2034         int nr_pages;
2035         int nr_moved = 0;
2036
2037         while (!list_empty(list)) {
2038                 page = lru_to_page(list);
2039                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2040
2041                 VM_BUG_ON_PAGE(PageLRU(page), page);
2042                 SetPageLRU(page);
2043
2044                 nr_pages = hpage_nr_pages(page);
2045                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
2046                 list_move(&page->lru, &lruvec->lists[lru]);
2047
2048                 if (put_page_testzero(page)) {
2049                         __ClearPageLRU(page);
2050                         __ClearPageActive(page);
2051                         del_page_from_lru_list(page, lruvec, lru);
2052
2053                         if (unlikely(PageCompound(page))) {
2054                                 spin_unlock_irq(&pgdat->lru_lock);
2055                                 mem_cgroup_uncharge(page);
2056                                 (*get_compound_page_dtor(page))(page);
2057                                 spin_lock_irq(&pgdat->lru_lock);
2058                         } else
2059                                 list_add(&page->lru, pages_to_free);
2060                 } else {
2061                         nr_moved += nr_pages;
2062                 }
2063         }
2064
2065         if (!is_active_lru(lru)) {
2066                 __count_vm_events(PGDEACTIVATE, nr_moved);
2067                 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
2068                                    nr_moved);
2069         }
2070
2071         return nr_moved;
2072 }
2073
2074 static void shrink_active_list(unsigned long nr_to_scan,
2075                                struct lruvec *lruvec,
2076                                struct scan_control *sc,
2077                                enum lru_list lru)
2078 {
2079         unsigned long nr_taken;
2080         unsigned long nr_scanned;
2081         unsigned long vm_flags;
2082         LIST_HEAD(l_hold);      /* The pages which were snipped off */
2083         LIST_HEAD(l_active);
2084         LIST_HEAD(l_inactive);
2085         struct page *page;
2086         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2087         unsigned nr_deactivate, nr_activate;
2088         unsigned nr_rotated = 0;
2089         isolate_mode_t isolate_mode = 0;
2090         int file = is_file_lru(lru);
2091         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2092
2093         lru_add_drain();
2094
2095         if (!sc->may_unmap)
2096                 isolate_mode |= ISOLATE_UNMAPPED;
2097
2098         spin_lock_irq(&pgdat->lru_lock);
2099
2100         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2101                                      &nr_scanned, sc, isolate_mode, lru);
2102
2103         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2104         reclaim_stat->recent_scanned[file] += nr_taken;
2105
2106         __count_vm_events(PGREFILL, nr_scanned);
2107         count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2108
2109         spin_unlock_irq(&pgdat->lru_lock);
2110
2111         while (!list_empty(&l_hold)) {
2112                 cond_resched();
2113                 page = lru_to_page(&l_hold);
2114                 list_del(&page->lru);
2115
2116                 if (unlikely(!page_evictable(page))) {
2117                         putback_lru_page(page);
2118                         continue;
2119                 }
2120
2121                 if (unlikely(buffer_heads_over_limit)) {
2122                         if (page_has_private(page) && trylock_page(page)) {
2123                                 if (page_has_private(page))
2124                                         try_to_release_page(page, 0);
2125                                 unlock_page(page);
2126                         }
2127                 }
2128
2129                 if (page_referenced(page, 0, sc->target_mem_cgroup,
2130                                     &vm_flags)) {
2131                         nr_rotated += hpage_nr_pages(page);
2132                         /*
2133                          * Identify referenced, file-backed active pages and
2134                          * give them one more trip around the active list. So
2135                          * that executable code get better chances to stay in
2136                          * memory under moderate memory pressure.  Anon pages
2137                          * are not likely to be evicted by use-once streaming
2138                          * IO, plus JVM can create lots of anon VM_EXEC pages,
2139                          * so we ignore them here.
2140                          */
2141                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2142                                 list_add(&page->lru, &l_active);
2143                                 continue;
2144                         }
2145                 }
2146
2147                 ClearPageActive(page);  /* we are de-activating */
2148                 SetPageWorkingset(page);
2149                 list_add(&page->lru, &l_inactive);
2150         }
2151
2152         /*
2153          * Move pages back to the lru list.
2154          */
2155         spin_lock_irq(&pgdat->lru_lock);
2156         /*
2157          * Count referenced pages from currently used mappings as rotated,
2158          * even though only some of them are actually re-activated.  This
2159          * helps balance scan pressure between file and anonymous pages in
2160          * get_scan_count.
2161          */
2162         reclaim_stat->recent_rotated[file] += nr_rotated;
2163
2164         nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2165         nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2166         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2167         spin_unlock_irq(&pgdat->lru_lock);
2168
2169         mem_cgroup_uncharge_list(&l_hold);
2170         free_unref_page_list(&l_hold);
2171         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2172                         nr_deactivate, nr_rotated, sc->priority, file);
2173 }
2174
2175 /*
2176  * The inactive anon list should be small enough that the VM never has
2177  * to do too much work.
2178  *
2179  * The inactive file list should be small enough to leave most memory
2180  * to the established workingset on the scan-resistant active list,
2181  * but large enough to avoid thrashing the aggregate readahead window.
2182  *
2183  * Both inactive lists should also be large enough that each inactive
2184  * page has a chance to be referenced again before it is reclaimed.
2185  *
2186  * If that fails and refaulting is observed, the inactive list grows.
2187  *
2188  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2189  * on this LRU, maintained by the pageout code. An inactive_ratio
2190  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2191  *
2192  * total     target    max
2193  * memory    ratio     inactive
2194  * -------------------------------------
2195  *   10MB       1         5MB
2196  *  100MB       1        50MB
2197  *    1GB       3       250MB
2198  *   10GB      10       0.9GB
2199  *  100GB      31         3GB
2200  *    1TB     101        10GB
2201  *   10TB     320        32GB
2202  */
2203 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2204                                  struct mem_cgroup *memcg,
2205                                  struct scan_control *sc, bool actual_reclaim)
2206 {
2207         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2208         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2209         enum lru_list inactive_lru = file * LRU_FILE;
2210         unsigned long inactive, active;
2211         unsigned long inactive_ratio;
2212         unsigned long refaults;
2213         unsigned long gb;
2214
2215         /*
2216          * If we don't have swap space, anonymous page deactivation
2217          * is pointless.
2218          */
2219         if (!file && !total_swap_pages)
2220                 return false;
2221
2222         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2223         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2224
2225         if (memcg)
2226                 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2227         else
2228                 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2229
2230         /*
2231          * When refaults are being observed, it means a new workingset
2232          * is being established. Disable active list protection to get
2233          * rid of the stale workingset quickly.
2234          */
2235         if (file && actual_reclaim && lruvec->refaults != refaults) {
2236                 inactive_ratio = 0;
2237         } else {
2238                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2239                 if (gb)
2240                         inactive_ratio = int_sqrt(10 * gb);
2241                 else
2242                         inactive_ratio = 1;
2243         }
2244
2245         if (actual_reclaim)
2246                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2247                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2248                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2249                         inactive_ratio, file);
2250
2251         return inactive * inactive_ratio < active;
2252 }
2253
2254 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2255                                  struct lruvec *lruvec, struct mem_cgroup *memcg,
2256                                  struct scan_control *sc)
2257 {
2258         if (is_active_lru(lru)) {
2259                 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2260                                          memcg, sc, true))
2261                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2262                 return 0;
2263         }
2264
2265         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2266 }
2267
2268 enum scan_balance {
2269         SCAN_EQUAL,
2270         SCAN_FRACT,
2271         SCAN_ANON,
2272         SCAN_FILE,
2273 };
2274
2275 /*
2276  * Determine how aggressively the anon and file LRU lists should be
2277  * scanned.  The relative value of each set of LRU lists is determined
2278  * by looking at the fraction of the pages scanned we did rotate back
2279  * onto the active list instead of evict.
2280  *
2281  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2282  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2283  */
2284 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2285                            struct scan_control *sc, unsigned long *nr,
2286                            unsigned long *lru_pages)
2287 {
2288         int swappiness = mem_cgroup_swappiness(memcg);
2289         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2290         u64 fraction[2];
2291         u64 denominator = 0;    /* gcc */
2292         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2293         unsigned long anon_prio, file_prio;
2294         enum scan_balance scan_balance;
2295         unsigned long anon, file;
2296         unsigned long ap, fp;
2297         enum lru_list lru;
2298
2299         /* If we have no swap space, do not bother scanning anon pages. */
2300         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2301                 scan_balance = SCAN_FILE;
2302                 goto out;
2303         }
2304
2305         /*
2306          * Global reclaim will swap to prevent OOM even with no
2307          * swappiness, but memcg users want to use this knob to
2308          * disable swapping for individual groups completely when
2309          * using the memory controller's swap limit feature would be
2310          * too expensive.
2311          */
2312         if (!global_reclaim(sc) && !swappiness) {
2313                 scan_balance = SCAN_FILE;
2314                 goto out;
2315         }
2316
2317         /*
2318          * Do not apply any pressure balancing cleverness when the
2319          * system is close to OOM, scan both anon and file equally
2320          * (unless the swappiness setting disagrees with swapping).
2321          */
2322         if (!sc->priority && swappiness) {
2323                 scan_balance = SCAN_EQUAL;
2324                 goto out;
2325         }
2326
2327         /*
2328          * Prevent the reclaimer from falling into the cache trap: as
2329          * cache pages start out inactive, every cache fault will tip
2330          * the scan balance towards the file LRU.  And as the file LRU
2331          * shrinks, so does the window for rotation from references.
2332          * This means we have a runaway feedback loop where a tiny
2333          * thrashing file LRU becomes infinitely more attractive than
2334          * anon pages.  Try to detect this based on file LRU size.
2335          */
2336         if (global_reclaim(sc)) {
2337                 unsigned long pgdatfile;
2338                 unsigned long pgdatfree;
2339                 int z;
2340                 unsigned long total_high_wmark = 0;
2341
2342                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2343                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2344                            node_page_state(pgdat, NR_INACTIVE_FILE);
2345
2346                 for (z = 0; z < MAX_NR_ZONES; z++) {
2347                         struct zone *zone = &pgdat->node_zones[z];
2348                         if (!managed_zone(zone))
2349                                 continue;
2350
2351                         total_high_wmark += high_wmark_pages(zone);
2352                 }
2353
2354                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2355                         /*
2356                          * Force SCAN_ANON if there are enough inactive
2357                          * anonymous pages on the LRU in eligible zones.
2358                          * Otherwise, the small LRU gets thrashed.
2359                          */
2360                         if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2361                             lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2362                                         >> sc->priority) {
2363                                 scan_balance = SCAN_ANON;
2364                                 goto out;
2365                         }
2366                 }
2367         }
2368
2369         /*
2370          * If there is enough inactive page cache, i.e. if the size of the
2371          * inactive list is greater than that of the active list *and* the
2372          * inactive list actually has some pages to scan on this priority, we
2373          * do not reclaim anything from the anonymous working set right now.
2374          * Without the second condition we could end up never scanning an
2375          * lruvec even if it has plenty of old anonymous pages unless the
2376          * system is under heavy pressure.
2377          */
2378         if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2379             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2380                 scan_balance = SCAN_FILE;
2381                 goto out;
2382         }
2383
2384         scan_balance = SCAN_FRACT;
2385
2386         /*
2387          * With swappiness at 100, anonymous and file have the same priority.
2388          * This scanning priority is essentially the inverse of IO cost.
2389          */
2390         anon_prio = swappiness;
2391         file_prio = 200 - anon_prio;
2392
2393         /*
2394          * OK, so we have swap space and a fair amount of page cache
2395          * pages.  We use the recently rotated / recently scanned
2396          * ratios to determine how valuable each cache is.
2397          *
2398          * Because workloads change over time (and to avoid overflow)
2399          * we keep these statistics as a floating average, which ends
2400          * up weighing recent references more than old ones.
2401          *
2402          * anon in [0], file in [1]
2403          */
2404
2405         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2406                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2407         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2408                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2409
2410         spin_lock_irq(&pgdat->lru_lock);
2411         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2412                 reclaim_stat->recent_scanned[0] /= 2;
2413                 reclaim_stat->recent_rotated[0] /= 2;
2414         }
2415
2416         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2417                 reclaim_stat->recent_scanned[1] /= 2;
2418                 reclaim_stat->recent_rotated[1] /= 2;
2419         }
2420
2421         /*
2422          * The amount of pressure on anon vs file pages is inversely
2423          * proportional to the fraction of recently scanned pages on
2424          * each list that were recently referenced and in active use.
2425          */
2426         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2427         ap /= reclaim_stat->recent_rotated[0] + 1;
2428
2429         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2430         fp /= reclaim_stat->recent_rotated[1] + 1;
2431         spin_unlock_irq(&pgdat->lru_lock);
2432
2433         fraction[0] = ap;
2434         fraction[1] = fp;
2435         denominator = ap + fp + 1;
2436 out:
2437         *lru_pages = 0;
2438         for_each_evictable_lru(lru) {
2439                 int file = is_file_lru(lru);
2440                 unsigned long size;
2441                 unsigned long scan;
2442
2443                 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2444                 scan = size >> sc->priority;
2445                 /*
2446                  * If the cgroup's already been deleted, make sure to
2447                  * scrape out the remaining cache.
2448                  */
2449                 if (!scan && !mem_cgroup_online(memcg))
2450                         scan = min(size, SWAP_CLUSTER_MAX);
2451
2452                 switch (scan_balance) {
2453                 case SCAN_EQUAL:
2454                         /* Scan lists relative to size */
2455                         break;
2456                 case SCAN_FRACT:
2457                         /*
2458                          * Scan types proportional to swappiness and
2459                          * their relative recent reclaim efficiency.
2460                          * Make sure we don't miss the last page
2461                          * because of a round-off error.
2462                          */
2463                         scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2464                                                   denominator);
2465                         break;
2466                 case SCAN_FILE:
2467                 case SCAN_ANON:
2468                         /* Scan one type exclusively */
2469                         if ((scan_balance == SCAN_FILE) != file) {
2470                                 size = 0;
2471                                 scan = 0;
2472                         }
2473                         break;
2474                 default:
2475                         /* Look ma, no brain */
2476                         BUG();
2477                 }
2478
2479                 *lru_pages += size;
2480                 nr[lru] = scan;
2481         }
2482 }
2483
2484 /*
2485  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2486  */
2487 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2488                               struct scan_control *sc, unsigned long *lru_pages)
2489 {
2490         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2491         unsigned long nr[NR_LRU_LISTS];
2492         unsigned long targets[NR_LRU_LISTS];
2493         unsigned long nr_to_scan;
2494         enum lru_list lru;
2495         unsigned long nr_reclaimed = 0;
2496         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2497         struct blk_plug plug;
2498         bool scan_adjusted;
2499
2500         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2501
2502         /* Record the original scan target for proportional adjustments later */
2503         memcpy(targets, nr, sizeof(nr));
2504
2505         /*
2506          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2507          * event that can occur when there is little memory pressure e.g.
2508          * multiple streaming readers/writers. Hence, we do not abort scanning
2509          * when the requested number of pages are reclaimed when scanning at
2510          * DEF_PRIORITY on the assumption that the fact we are direct
2511          * reclaiming implies that kswapd is not keeping up and it is best to
2512          * do a batch of work at once. For memcg reclaim one check is made to
2513          * abort proportional reclaim if either the file or anon lru has already
2514          * dropped to zero at the first pass.
2515          */
2516         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2517                          sc->priority == DEF_PRIORITY);
2518
2519         blk_start_plug(&plug);
2520         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2521                                         nr[LRU_INACTIVE_FILE]) {
2522                 unsigned long nr_anon, nr_file, percentage;
2523                 unsigned long nr_scanned;
2524
2525                 for_each_evictable_lru(lru) {
2526                         if (nr[lru]) {
2527                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2528                                 nr[lru] -= nr_to_scan;
2529
2530                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2531                                                             lruvec, memcg, sc);
2532                         }
2533                 }
2534
2535                 cond_resched();
2536
2537                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2538                         continue;
2539
2540                 /*
2541                  * For kswapd and memcg, reclaim at least the number of pages
2542                  * requested. Ensure that the anon and file LRUs are scanned
2543                  * proportionally what was requested by get_scan_count(). We
2544                  * stop reclaiming one LRU and reduce the amount scanning
2545                  * proportional to the original scan target.
2546                  */
2547                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2548                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2549
2550                 /*
2551                  * It's just vindictive to attack the larger once the smaller
2552                  * has gone to zero.  And given the way we stop scanning the
2553                  * smaller below, this makes sure that we only make one nudge
2554                  * towards proportionality once we've got nr_to_reclaim.
2555                  */
2556                 if (!nr_file || !nr_anon)
2557                         break;
2558
2559                 if (nr_file > nr_anon) {
2560                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2561                                                 targets[LRU_ACTIVE_ANON] + 1;
2562                         lru = LRU_BASE;
2563                         percentage = nr_anon * 100 / scan_target;
2564                 } else {
2565                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2566                                                 targets[LRU_ACTIVE_FILE] + 1;
2567                         lru = LRU_FILE;
2568                         percentage = nr_file * 100 / scan_target;
2569                 }
2570
2571                 /* Stop scanning the smaller of the LRU */
2572                 nr[lru] = 0;
2573                 nr[lru + LRU_ACTIVE] = 0;
2574
2575                 /*
2576                  * Recalculate the other LRU scan count based on its original
2577                  * scan target and the percentage scanning already complete
2578                  */
2579                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2580                 nr_scanned = targets[lru] - nr[lru];
2581                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2582                 nr[lru] -= min(nr[lru], nr_scanned);
2583
2584                 lru += LRU_ACTIVE;
2585                 nr_scanned = targets[lru] - nr[lru];
2586                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2587                 nr[lru] -= min(nr[lru], nr_scanned);
2588
2589                 scan_adjusted = true;
2590         }
2591         blk_finish_plug(&plug);
2592         sc->nr_reclaimed += nr_reclaimed;
2593
2594         /*
2595          * Even if we did not try to evict anon pages at all, we want to
2596          * rebalance the anon lru active/inactive ratio.
2597          */
2598         if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2599                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2600                                    sc, LRU_ACTIVE_ANON);
2601 }
2602
2603 /* Use reclaim/compaction for costly allocs or under memory pressure */
2604 static bool in_reclaim_compaction(struct scan_control *sc)
2605 {
2606         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2607                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2608                          sc->priority < DEF_PRIORITY - 2))
2609                 return true;
2610
2611         return false;
2612 }
2613
2614 /*
2615  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2616  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2617  * true if more pages should be reclaimed such that when the page allocator
2618  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2619  * It will give up earlier than that if there is difficulty reclaiming pages.
2620  */
2621 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2622                                         unsigned long nr_reclaimed,
2623                                         unsigned long nr_scanned,
2624                                         struct scan_control *sc)
2625 {
2626         unsigned long pages_for_compaction;
2627         unsigned long inactive_lru_pages;
2628         int z;
2629
2630         /* If not in reclaim/compaction mode, stop */
2631         if (!in_reclaim_compaction(sc))
2632                 return false;
2633
2634         /* Consider stopping depending on scan and reclaim activity */
2635         if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2636                 /*
2637                  * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2638                  * full LRU list has been scanned and we are still failing
2639                  * to reclaim pages. This full LRU scan is potentially
2640                  * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2641                  */
2642                 if (!nr_reclaimed && !nr_scanned)
2643                         return false;
2644         } else {
2645                 /*
2646                  * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2647                  * fail without consequence, stop if we failed to reclaim
2648                  * any pages from the last SWAP_CLUSTER_MAX number of
2649                  * pages that were scanned. This will return to the
2650                  * caller faster at the risk reclaim/compaction and
2651                  * the resulting allocation attempt fails
2652                  */
2653                 if (!nr_reclaimed)
2654                         return false;
2655         }
2656
2657         /*
2658          * If we have not reclaimed enough pages for compaction and the
2659          * inactive lists are large enough, continue reclaiming
2660          */
2661         pages_for_compaction = compact_gap(sc->order);
2662         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2663         if (get_nr_swap_pages() > 0)
2664                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2665         if (sc->nr_reclaimed < pages_for_compaction &&
2666                         inactive_lru_pages > pages_for_compaction)
2667                 return true;
2668
2669         /* If compaction would go ahead or the allocation would succeed, stop */
2670         for (z = 0; z <= sc->reclaim_idx; z++) {
2671                 struct zone *zone = &pgdat->node_zones[z];
2672                 if (!managed_zone(zone))
2673                         continue;
2674
2675                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2676                 case COMPACT_SUCCESS:
2677                 case COMPACT_CONTINUE:
2678                         return false;
2679                 default:
2680                         /* check next zone */
2681                         ;
2682                 }
2683         }
2684         return true;
2685 }
2686
2687 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2688 {
2689         return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2690                 (memcg && memcg_congested(pgdat, memcg));
2691 }
2692
2693 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2694 {
2695         struct reclaim_state *reclaim_state = current->reclaim_state;
2696         unsigned long nr_reclaimed, nr_scanned;
2697         bool reclaimable = false;
2698
2699         do {
2700                 struct mem_cgroup *root = sc->target_mem_cgroup;
2701                 struct mem_cgroup_reclaim_cookie reclaim = {
2702                         .pgdat = pgdat,
2703                         .priority = sc->priority,
2704                 };
2705                 unsigned long node_lru_pages = 0;
2706                 struct mem_cgroup *memcg;
2707
2708                 memset(&sc->nr, 0, sizeof(sc->nr));
2709
2710                 nr_reclaimed = sc->nr_reclaimed;
2711                 nr_scanned = sc->nr_scanned;
2712
2713                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2714                 do {
2715                         unsigned long lru_pages;
2716                         unsigned long reclaimed;
2717                         unsigned long scanned;
2718
2719                         switch (mem_cgroup_protected(root, memcg)) {
2720                         case MEMCG_PROT_MIN:
2721                                 /*
2722                                  * Hard protection.
2723                                  * If there is no reclaimable memory, OOM.
2724                                  */
2725                                 continue;
2726                         case MEMCG_PROT_LOW:
2727                                 /*
2728                                  * Soft protection.
2729                                  * Respect the protection only as long as
2730                                  * there is an unprotected supply
2731                                  * of reclaimable memory from other cgroups.
2732                                  */
2733                                 if (!sc->memcg_low_reclaim) {
2734                                         sc->memcg_low_skipped = 1;
2735                                         continue;
2736                                 }
2737                                 memcg_memory_event(memcg, MEMCG_LOW);
2738                                 break;
2739                         case MEMCG_PROT_NONE:
2740                                 break;
2741                         }
2742
2743                         reclaimed = sc->nr_reclaimed;
2744                         scanned = sc->nr_scanned;
2745                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2746                         node_lru_pages += lru_pages;
2747
2748                         shrink_slab(sc->gfp_mask, pgdat->node_id,
2749                                     memcg, sc->priority);
2750
2751                         /* Record the group's reclaim efficiency */
2752                         vmpressure(sc->gfp_mask, memcg, false,
2753                                    sc->nr_scanned - scanned,
2754                                    sc->nr_reclaimed - reclaimed);
2755
2756                         /*
2757                          * Direct reclaim and kswapd have to scan all memory
2758                          * cgroups to fulfill the overall scan target for the
2759                          * node.
2760                          *
2761                          * Limit reclaim, on the other hand, only cares about
2762                          * nr_to_reclaim pages to be reclaimed and it will
2763                          * retry with decreasing priority if one round over the
2764                          * whole hierarchy is not sufficient.
2765                          */
2766                         if (!global_reclaim(sc) &&
2767                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2768                                 mem_cgroup_iter_break(root, memcg);
2769                                 break;
2770                         }
2771                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2772
2773                 if (reclaim_state) {
2774                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2775                         reclaim_state->reclaimed_slab = 0;
2776                 }
2777
2778                 /* Record the subtree's reclaim efficiency */
2779                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2780                            sc->nr_scanned - nr_scanned,
2781                            sc->nr_reclaimed - nr_reclaimed);
2782
2783                 if (sc->nr_reclaimed - nr_reclaimed)
2784                         reclaimable = true;
2785
2786                 if (current_is_kswapd()) {
2787                         /*
2788                          * If reclaim is isolating dirty pages under writeback,
2789                          * it implies that the long-lived page allocation rate
2790                          * is exceeding the page laundering rate. Either the
2791                          * global limits are not being effective at throttling
2792                          * processes due to the page distribution throughout
2793                          * zones or there is heavy usage of a slow backing
2794                          * device. The only option is to throttle from reclaim
2795                          * context which is not ideal as there is no guarantee
2796                          * the dirtying process is throttled in the same way
2797                          * balance_dirty_pages() manages.
2798                          *
2799                          * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2800                          * count the number of pages under pages flagged for
2801                          * immediate reclaim and stall if any are encountered
2802                          * in the nr_immediate check below.
2803                          */
2804                         if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2805                                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2806
2807                         /*
2808                          * Tag a node as congested if all the dirty pages
2809                          * scanned were backed by a congested BDI and
2810                          * wait_iff_congested will stall.
2811                          */
2812                         if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2813                                 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2814
2815                         /* Allow kswapd to start writing pages during reclaim.*/
2816                         if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2817                                 set_bit(PGDAT_DIRTY, &pgdat->flags);
2818
2819                         /*
2820                          * If kswapd scans pages marked marked for immediate
2821                          * reclaim and under writeback (nr_immediate), it
2822                          * implies that pages are cycling through the LRU
2823                          * faster than they are written so also forcibly stall.
2824                          */
2825                         if (sc->nr.immediate)
2826                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2827                 }
2828
2829                 /*
2830                  * Legacy memcg will stall in page writeback so avoid forcibly
2831                  * stalling in wait_iff_congested().
2832                  */
2833                 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2834                     sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2835                         set_memcg_congestion(pgdat, root, true);
2836
2837                 /*
2838                  * Stall direct reclaim for IO completions if underlying BDIs
2839                  * and node is congested. Allow kswapd to continue until it
2840                  * starts encountering unqueued dirty pages or cycling through
2841                  * the LRU too quickly.
2842                  */
2843                 if (!sc->hibernation_mode && !current_is_kswapd() &&
2844                    current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2845                         wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2846
2847         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2848                                          sc->nr_scanned - nr_scanned, sc));
2849
2850         /*
2851          * Kswapd gives up on balancing particular nodes after too
2852          * many failures to reclaim anything from them and goes to
2853          * sleep. On reclaim progress, reset the failure counter. A
2854          * successful direct reclaim run will revive a dormant kswapd.
2855          */
2856         if (reclaimable)
2857                 pgdat->kswapd_failures = 0;
2858
2859         return reclaimable;
2860 }
2861
2862 /*
2863  * Returns true if compaction should go ahead for a costly-order request, or
2864  * the allocation would already succeed without compaction. Return false if we
2865  * should reclaim first.
2866  */
2867 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2868 {
2869         unsigned long watermark;
2870         enum compact_result suitable;
2871
2872         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2873         if (suitable == COMPACT_SUCCESS)
2874                 /* Allocation should succeed already. Don't reclaim. */
2875                 return true;
2876         if (suitable == COMPACT_SKIPPED)
2877                 /* Compaction cannot yet proceed. Do reclaim. */
2878                 return false;
2879
2880         /*
2881          * Compaction is already possible, but it takes time to run and there
2882          * are potentially other callers using the pages just freed. So proceed
2883          * with reclaim to make a buffer of free pages available to give
2884          * compaction a reasonable chance of completing and allocating the page.
2885          * Note that we won't actually reclaim the whole buffer in one attempt
2886          * as the target watermark in should_continue_reclaim() is lower. But if
2887          * we are already above the high+gap watermark, don't reclaim at all.
2888          */
2889         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2890
2891         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2892 }
2893
2894 /*
2895  * This is the direct reclaim path, for page-allocating processes.  We only
2896  * try to reclaim pages from zones which will satisfy the caller's allocation
2897  * request.
2898  *
2899  * If a zone is deemed to be full of pinned pages then just give it a light
2900  * scan then give up on it.
2901  */
2902 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2903 {
2904         struct zoneref *z;
2905         struct zone *zone;
2906         unsigned long nr_soft_reclaimed;
2907         unsigned long nr_soft_scanned;
2908         gfp_t orig_mask;
2909         pg_data_t *last_pgdat = NULL;
2910
2911         /*
2912          * If the number of buffer_heads in the machine exceeds the maximum
2913          * allowed level, force direct reclaim to scan the highmem zone as
2914          * highmem pages could be pinning lowmem pages storing buffer_heads
2915          */
2916         orig_mask = sc->gfp_mask;
2917         if (buffer_heads_over_limit) {
2918                 sc->gfp_mask |= __GFP_HIGHMEM;
2919                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2920         }
2921
2922         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2923                                         sc->reclaim_idx, sc->nodemask) {
2924                 /*
2925                  * Take care memory controller reclaiming has small influence
2926                  * to global LRU.
2927                  */
2928                 if (global_reclaim(sc)) {
2929                         if (!cpuset_zone_allowed(zone,
2930                                                  GFP_KERNEL | __GFP_HARDWALL))
2931                                 continue;
2932
2933                         /*
2934                          * If we already have plenty of memory free for
2935                          * compaction in this zone, don't free any more.
2936                          * Even though compaction is invoked for any
2937                          * non-zero order, only frequent costly order
2938                          * reclamation is disruptive enough to become a
2939                          * noticeable problem, like transparent huge
2940                          * page allocations.
2941                          */
2942                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2943                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2944                             compaction_ready(zone, sc)) {
2945                                 sc->compaction_ready = true;
2946                                 continue;
2947                         }
2948
2949                         /*
2950                          * Shrink each node in the zonelist once. If the
2951                          * zonelist is ordered by zone (not the default) then a
2952                          * node may be shrunk multiple times but in that case
2953                          * the user prefers lower zones being preserved.
2954                          */
2955                         if (zone->zone_pgdat == last_pgdat)
2956                                 continue;
2957
2958                         /*
2959                          * This steals pages from memory cgroups over softlimit
2960                          * and returns the number of reclaimed pages and
2961                          * scanned pages. This works for global memory pressure
2962                          * and balancing, not for a memcg's limit.
2963                          */
2964                         nr_soft_scanned = 0;
2965                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2966                                                 sc->order, sc->gfp_mask,
2967                                                 &nr_soft_scanned);
2968                         sc->nr_reclaimed += nr_soft_reclaimed;
2969                         sc->nr_scanned += nr_soft_scanned;
2970                         /* need some check for avoid more shrink_zone() */
2971                 }
2972
2973                 /* See comment about same check for global reclaim above */
2974                 if (zone->zone_pgdat == last_pgdat)
2975                         continue;
2976                 last_pgdat = zone->zone_pgdat;
2977                 shrink_node(zone->zone_pgdat, sc);
2978         }
2979
2980         /*
2981          * Restore to original mask to avoid the impact on the caller if we
2982          * promoted it to __GFP_HIGHMEM.
2983          */
2984         sc->gfp_mask = orig_mask;
2985 }
2986
2987 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2988 {
2989         struct mem_cgroup *memcg;
2990
2991         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2992         do {
2993                 unsigned long refaults;
2994                 struct lruvec *lruvec;
2995
2996                 if (memcg)
2997                         refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2998                 else
2999                         refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
3000
3001                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3002                 lruvec->refaults = refaults;
3003         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3004 }
3005
3006 /*
3007  * This is the main entry point to direct page reclaim.
3008  *
3009  * If a full scan of the inactive list fails to free enough memory then we
3010  * are "out of memory" and something needs to be killed.
3011  *
3012  * If the caller is !__GFP_FS then the probability of a failure is reasonably
3013  * high - the zone may be full of dirty or under-writeback pages, which this
3014  * caller can't do much about.  We kick the writeback threads and take explicit
3015  * naps in the hope that some of these pages can be written.  But if the
3016  * allocating task holds filesystem locks which prevent writeout this might not
3017  * work, and the allocation attempt will fail.
3018  *
3019  * returns:     0, if no pages reclaimed
3020  *              else, the number of pages reclaimed
3021  */
3022 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3023                                           struct scan_control *sc)
3024 {
3025         int initial_priority = sc->priority;
3026         pg_data_t *last_pgdat;
3027         struct zoneref *z;
3028         struct zone *zone;
3029 retry:
3030         delayacct_freepages_start();
3031
3032         if (global_reclaim(sc))
3033                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3034
3035         do {
3036                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3037                                 sc->priority);
3038                 sc->nr_scanned = 0;
3039                 shrink_zones(zonelist, sc);
3040
3041                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3042                         break;
3043
3044                 if (sc->compaction_ready)
3045                         break;
3046
3047                 /*
3048                  * If we're getting trouble reclaiming, start doing
3049                  * writepage even in laptop mode.
3050                  */
3051                 if (sc->priority < DEF_PRIORITY - 2)
3052                         sc->may_writepage = 1;
3053         } while (--sc->priority >= 0);
3054
3055         last_pgdat = NULL;
3056         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3057                                         sc->nodemask) {
3058                 if (zone->zone_pgdat == last_pgdat)
3059                         continue;
3060                 last_pgdat = zone->zone_pgdat;
3061                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3062                 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3063         }
3064
3065         delayacct_freepages_end();
3066
3067         if (sc->nr_reclaimed)
3068                 return sc->nr_reclaimed;
3069
3070         /* Aborted reclaim to try compaction? don't OOM, then */
3071         if (sc->compaction_ready)
3072                 return 1;
3073
3074         /* Untapped cgroup reserves?  Don't OOM, retry. */
3075         if (sc->memcg_low_skipped) {
3076                 sc->priority = initial_priority;
3077                 sc->memcg_low_reclaim = 1;
3078                 sc->memcg_low_skipped = 0;
3079                 goto retry;
3080         }
3081
3082         return 0;
3083 }
3084
3085 static bool allow_direct_reclaim(pg_data_t *pgdat)
3086 {
3087         struct zone *zone;
3088         unsigned long pfmemalloc_reserve = 0;
3089         unsigned long free_pages = 0;
3090         int i;
3091         bool wmark_ok;
3092
3093         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3094                 return true;
3095
3096         for (i = 0; i <= ZONE_NORMAL; i++) {
3097                 zone = &pgdat->node_zones[i];
3098                 if (!managed_zone(zone))
3099                         continue;
3100
3101                 if (!zone_reclaimable_pages(zone))
3102                         continue;
3103
3104                 pfmemalloc_reserve += min_wmark_pages(zone);
3105                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3106         }
3107
3108         /* If there are no reserves (unexpected config) then do not throttle */
3109         if (!pfmemalloc_reserve)
3110                 return true;
3111
3112         wmark_ok = free_pages > pfmemalloc_reserve / 2;
3113
3114         /* kswapd must be awake if processes are being throttled */
3115         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3116                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3117                                                 (enum zone_type)ZONE_NORMAL);
3118                 wake_up_interruptible(&pgdat->kswapd_wait);
3119         }
3120
3121         return wmark_ok;
3122 }
3123
3124 /*
3125  * Throttle direct reclaimers if backing storage is backed by the network
3126  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3127  * depleted. kswapd will continue to make progress and wake the processes
3128  * when the low watermark is reached.
3129  *
3130  * Returns true if a fatal signal was delivered during throttling. If this
3131  * happens, the page allocator should not consider triggering the OOM killer.
3132  */
3133 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3134                                         nodemask_t *nodemask)
3135 {
3136         struct zoneref *z;
3137         struct zone *zone;
3138         pg_data_t *pgdat = NULL;
3139
3140         /*
3141          * Kernel threads should not be throttled as they may be indirectly
3142          * responsible for cleaning pages necessary for reclaim to make forward
3143          * progress. kjournald for example may enter direct reclaim while
3144          * committing a transaction where throttling it could forcing other
3145          * processes to block on log_wait_commit().
3146          */
3147         if (current->flags & PF_KTHREAD)
3148                 goto out;
3149
3150         /*
3151          * If a fatal signal is pending, this process should not throttle.
3152          * It should return quickly so it can exit and free its memory
3153          */
3154         if (fatal_signal_pending(current))
3155                 goto out;
3156
3157         /*
3158          * Check if the pfmemalloc reserves are ok by finding the first node
3159          * with a usable ZONE_NORMAL or lower zone. The expectation is that
3160          * GFP_KERNEL will be required for allocating network buffers when
3161          * swapping over the network so ZONE_HIGHMEM is unusable.
3162          *
3163          * Throttling is based on the first usable node and throttled processes
3164          * wait on a queue until kswapd makes progress and wakes them. There
3165          * is an affinity then between processes waking up and where reclaim
3166          * progress has been made assuming the process wakes on the same node.
3167          * More importantly, processes running on remote nodes will not compete
3168          * for remote pfmemalloc reserves and processes on different nodes
3169          * should make reasonable progress.
3170          */
3171         for_each_zone_zonelist_nodemask(zone, z, zonelist,
3172                                         gfp_zone(gfp_mask), nodemask) {
3173                 if (zone_idx(zone) > ZONE_NORMAL)
3174                         continue;
3175
3176                 /* Throttle based on the first usable node */
3177                 pgdat = zone->zone_pgdat;
3178                 if (allow_direct_reclaim(pgdat))
3179                         goto out;
3180                 break;
3181         }
3182
3183         /* If no zone was usable by the allocation flags then do not throttle */
3184         if (!pgdat)
3185                 goto out;
3186
3187         /* Account for the throttling */
3188         count_vm_event(PGSCAN_DIRECT_THROTTLE);
3189
3190         /*
3191          * If the caller cannot enter the filesystem, it's possible that it
3192          * is due to the caller holding an FS lock or performing a journal
3193          * transaction in the case of a filesystem like ext[3|4]. In this case,
3194          * it is not safe to block on pfmemalloc_wait as kswapd could be
3195          * blocked waiting on the same lock. Instead, throttle for up to a
3196          * second before continuing.
3197          */
3198         if (!(gfp_mask & __GFP_FS)) {
3199                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3200                         allow_direct_reclaim(pgdat), HZ);
3201
3202                 goto check_pending;
3203         }
3204
3205         /* Throttle until kswapd wakes the process */
3206         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3207                 allow_direct_reclaim(pgdat));
3208
3209 check_pending:
3210         if (fatal_signal_pending(current))
3211                 return true;
3212
3213 out:
3214         return false;
3215 }
3216
3217 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3218                                 gfp_t gfp_mask, nodemask_t *nodemask)
3219 {
3220         unsigned long nr_reclaimed;
3221         struct scan_control sc = {
3222                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3223                 .gfp_mask = current_gfp_context(gfp_mask),
3224                 .reclaim_idx = gfp_zone(gfp_mask),
3225                 .order = order,
3226                 .nodemask = nodemask,
3227                 .priority = DEF_PRIORITY,
3228                 .may_writepage = !laptop_mode,
3229                 .may_unmap = 1,
3230                 .may_swap = 1,
3231         };
3232
3233         /*
3234          * scan_control uses s8 fields for order, priority, and reclaim_idx.
3235          * Confirm they are large enough for max values.
3236          */
3237         BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3238         BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3239         BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3240
3241         /*
3242          * Do not enter reclaim if fatal signal was delivered while throttled.
3243          * 1 is returned so that the page allocator does not OOM kill at this
3244          * point.
3245          */
3246         if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3247                 return 1;
3248
3249         trace_mm_vmscan_direct_reclaim_begin(order,
3250                                 sc.may_writepage,
3251                                 sc.gfp_mask,
3252                                 sc.reclaim_idx);
3253
3254         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3255
3256         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3257
3258         return nr_reclaimed;
3259 }
3260
3261 #ifdef CONFIG_MEMCG
3262
3263 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3264                                                 gfp_t gfp_mask, bool noswap,
3265                                                 pg_data_t *pgdat,
3266                                                 unsigned long *nr_scanned)
3267 {
3268         struct scan_control sc = {
3269                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3270                 .target_mem_cgroup = memcg,
3271                 .may_writepage = !laptop_mode,
3272                 .may_unmap = 1,
3273                 .reclaim_idx = MAX_NR_ZONES - 1,
3274                 .may_swap = !noswap,
3275         };
3276         unsigned long lru_pages;
3277
3278         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3279                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3280
3281         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3282                                                       sc.may_writepage,
3283                                                       sc.gfp_mask,
3284                                                       sc.reclaim_idx);
3285
3286         /*
3287          * NOTE: Although we can get the priority field, using it
3288          * here is not a good idea, since it limits the pages we can scan.
3289          * if we don't reclaim here, the shrink_node from balance_pgdat
3290          * will pick up pages from other mem cgroup's as well. We hack
3291          * the priority and make it zero.
3292          */
3293         shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3294
3295         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3296
3297         *nr_scanned = sc.nr_scanned;
3298         return sc.nr_reclaimed;
3299 }
3300
3301 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3302                                            unsigned long nr_pages,
3303                                            gfp_t gfp_mask,
3304                                            bool may_swap)
3305 {
3306         struct zonelist *zonelist;
3307         unsigned long nr_reclaimed;
3308         int nid;
3309         unsigned int noreclaim_flag;
3310         struct scan_control sc = {
3311                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3312                 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3313                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3314                 .reclaim_idx = MAX_NR_ZONES - 1,
3315                 .target_mem_cgroup = memcg,
3316                 .priority = DEF_PRIORITY,
3317                 .may_writepage = !laptop_mode,
3318                 .may_unmap = 1,
3319                 .may_swap = may_swap,
3320         };
3321
3322         /*
3323          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3324          * take care of from where we get pages. So the node where we&