treewide: Replace GPLv2 boilerplate/reference with SPDX - rule 110
[sfrench/cifs-2.6.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/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
54
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
57
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
60
61 #include "internal.h"
62
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
65
66 struct scan_control {
67         /* How many pages shrink_list() should reclaim */
68         unsigned long nr_to_reclaim;
69
70         /*
71          * Nodemask of nodes allowed by the caller. If NULL, all nodes
72          * are scanned.
73          */
74         nodemask_t      *nodemask;
75
76         /*
77          * The memory cgroup that hit its limit and as a result is the
78          * primary target of this reclaim invocation.
79          */
80         struct mem_cgroup *target_mem_cgroup;
81
82         /* Writepage batching in laptop mode; RECLAIM_WRITE */
83         unsigned int may_writepage:1;
84
85         /* Can mapped pages be reclaimed? */
86         unsigned int may_unmap:1;
87
88         /* Can pages be swapped as part of reclaim? */
89         unsigned int may_swap:1;
90
91         /* e.g. boosted watermark reclaim leaves slabs alone */
92         unsigned int may_shrinkslab:1;
93
94         /*
95          * Cgroups are not reclaimed below their configured memory.low,
96          * unless we threaten to OOM. If any cgroups are skipped due to
97          * memory.low and nothing was reclaimed, go back for memory.low.
98          */
99         unsigned int memcg_low_reclaim:1;
100         unsigned int memcg_low_skipped:1;
101
102         unsigned int hibernation_mode:1;
103
104         /* One of the zones is ready for compaction */
105         unsigned int compaction_ready:1;
106
107         /* Allocation order */
108         s8 order;
109
110         /* Scan (total_size >> priority) pages at once */
111         s8 priority;
112
113         /* The highest zone to isolate pages for reclaim from */
114         s8 reclaim_idx;
115
116         /* This context's GFP mask */
117         gfp_t gfp_mask;
118
119         /* Incremented by the number of inactive pages that were scanned */
120         unsigned long nr_scanned;
121
122         /* Number of pages freed so far during a call to shrink_zones() */
123         unsigned long nr_reclaimed;
124
125         struct {
126                 unsigned int dirty;
127                 unsigned int unqueued_dirty;
128                 unsigned int congested;
129                 unsigned int writeback;
130                 unsigned int immediate;
131                 unsigned int file_taken;
132                 unsigned int taken;
133         } nr;
134 };
135
136 #ifdef ARCH_HAS_PREFETCH
137 #define prefetch_prev_lru_page(_page, _base, _field)                    \
138         do {                                                            \
139                 if ((_page)->lru.prev != _base) {                       \
140                         struct page *prev;                              \
141                                                                         \
142                         prev = lru_to_page(&(_page->lru));              \
143                         prefetch(&prev->_field);                        \
144                 }                                                       \
145         } while (0)
146 #else
147 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148 #endif
149
150 #ifdef ARCH_HAS_PREFETCHW
151 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
152         do {                                                            \
153                 if ((_page)->lru.prev != _base) {                       \
154                         struct page *prev;                              \
155                                                                         \
156                         prev = lru_to_page(&(_page->lru));              \
157                         prefetchw(&prev->_field);                       \
158                 }                                                       \
159         } while (0)
160 #else
161 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162 #endif
163
164 /*
165  * From 0 .. 100.  Higher means more swappy.
166  */
167 int vm_swappiness = 60;
168 /*
169  * The total number of pages which are beyond the high watermark within all
170  * zones.
171  */
172 unsigned long vm_total_pages;
173
174 static LIST_HEAD(shrinker_list);
175 static DECLARE_RWSEM(shrinker_rwsem);
176
177 #ifdef CONFIG_MEMCG_KMEM
178
179 /*
180  * We allow subsystems to populate their shrinker-related
181  * LRU lists before register_shrinker_prepared() is called
182  * for the shrinker, since we don't want to impose
183  * restrictions on their internal registration order.
184  * In this case shrink_slab_memcg() may find corresponding
185  * bit is set in the shrinkers map.
186  *
187  * This value is used by the function to detect registering
188  * shrinkers and to skip do_shrink_slab() calls for them.
189  */
190 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
191
192 static DEFINE_IDR(shrinker_idr);
193 static int shrinker_nr_max;
194
195 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
196 {
197         int id, ret = -ENOMEM;
198
199         down_write(&shrinker_rwsem);
200         /* This may call shrinker, so it must use down_read_trylock() */
201         id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
202         if (id < 0)
203                 goto unlock;
204
205         if (id >= shrinker_nr_max) {
206                 if (memcg_expand_shrinker_maps(id)) {
207                         idr_remove(&shrinker_idr, id);
208                         goto unlock;
209                 }
210
211                 shrinker_nr_max = id + 1;
212         }
213         shrinker->id = id;
214         ret = 0;
215 unlock:
216         up_write(&shrinker_rwsem);
217         return ret;
218 }
219
220 static void unregister_memcg_shrinker(struct shrinker *shrinker)
221 {
222         int id = shrinker->id;
223
224         BUG_ON(id < 0);
225
226         down_write(&shrinker_rwsem);
227         idr_remove(&shrinker_idr, id);
228         up_write(&shrinker_rwsem);
229 }
230 #else /* CONFIG_MEMCG_KMEM */
231 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
232 {
233         return 0;
234 }
235
236 static void unregister_memcg_shrinker(struct shrinker *shrinker)
237 {
238 }
239 #endif /* CONFIG_MEMCG_KMEM */
240
241 #ifdef CONFIG_MEMCG
242 static bool global_reclaim(struct scan_control *sc)
243 {
244         return !sc->target_mem_cgroup;
245 }
246
247 /**
248  * sane_reclaim - is the usual dirty throttling mechanism operational?
249  * @sc: scan_control in question
250  *
251  * The normal page dirty throttling mechanism in balance_dirty_pages() is
252  * completely broken with the legacy memcg and direct stalling in
253  * shrink_page_list() is used for throttling instead, which lacks all the
254  * niceties such as fairness, adaptive pausing, bandwidth proportional
255  * allocation and configurability.
256  *
257  * This function tests whether the vmscan currently in progress can assume
258  * that the normal dirty throttling mechanism is operational.
259  */
260 static bool sane_reclaim(struct scan_control *sc)
261 {
262         struct mem_cgroup *memcg = sc->target_mem_cgroup;
263
264         if (!memcg)
265                 return true;
266 #ifdef CONFIG_CGROUP_WRITEBACK
267         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
268                 return true;
269 #endif
270         return false;
271 }
272
273 static void set_memcg_congestion(pg_data_t *pgdat,
274                                 struct mem_cgroup *memcg,
275                                 bool congested)
276 {
277         struct mem_cgroup_per_node *mn;
278
279         if (!memcg)
280                 return;
281
282         mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283         WRITE_ONCE(mn->congested, congested);
284 }
285
286 static bool memcg_congested(pg_data_t *pgdat,
287                         struct mem_cgroup *memcg)
288 {
289         struct mem_cgroup_per_node *mn;
290
291         mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292         return READ_ONCE(mn->congested);
293
294 }
295 #else
296 static bool global_reclaim(struct scan_control *sc)
297 {
298         return true;
299 }
300
301 static bool sane_reclaim(struct scan_control *sc)
302 {
303         return true;
304 }
305
306 static inline void set_memcg_congestion(struct pglist_data *pgdat,
307                                 struct mem_cgroup *memcg, bool congested)
308 {
309 }
310
311 static inline bool memcg_congested(struct pglist_data *pgdat,
312                         struct mem_cgroup *memcg)
313 {
314         return false;
315
316 }
317 #endif
318
319 /*
320  * This misses isolated pages which are not accounted for to save counters.
321  * As the data only determines if reclaim or compaction continues, it is
322  * not expected that isolated pages will be a dominating factor.
323  */
324 unsigned long zone_reclaimable_pages(struct zone *zone)
325 {
326         unsigned long nr;
327
328         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
329                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
330         if (get_nr_swap_pages() > 0)
331                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
332                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
333
334         return nr;
335 }
336
337 /**
338  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
339  * @lruvec: lru vector
340  * @lru: lru to use
341  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
342  */
343 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
344 {
345         unsigned long lru_size;
346         int zid;
347
348         if (!mem_cgroup_disabled())
349                 lru_size = lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
350         else
351                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
352
353         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
354                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
355                 unsigned long size;
356
357                 if (!managed_zone(zone))
358                         continue;
359
360                 if (!mem_cgroup_disabled())
361                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
362                 else
363                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
364                                        NR_ZONE_LRU_BASE + lru);
365                 lru_size -= min(size, lru_size);
366         }
367
368         return lru_size;
369
370 }
371
372 /*
373  * Add a shrinker callback to be called from the vm.
374  */
375 int prealloc_shrinker(struct shrinker *shrinker)
376 {
377         unsigned int size = sizeof(*shrinker->nr_deferred);
378
379         if (shrinker->flags & SHRINKER_NUMA_AWARE)
380                 size *= nr_node_ids;
381
382         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
383         if (!shrinker->nr_deferred)
384                 return -ENOMEM;
385
386         if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
387                 if (prealloc_memcg_shrinker(shrinker))
388                         goto free_deferred;
389         }
390
391         return 0;
392
393 free_deferred:
394         kfree(shrinker->nr_deferred);
395         shrinker->nr_deferred = NULL;
396         return -ENOMEM;
397 }
398
399 void free_prealloced_shrinker(struct shrinker *shrinker)
400 {
401         if (!shrinker->nr_deferred)
402                 return;
403
404         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
405                 unregister_memcg_shrinker(shrinker);
406
407         kfree(shrinker->nr_deferred);
408         shrinker->nr_deferred = NULL;
409 }
410
411 void register_shrinker_prepared(struct shrinker *shrinker)
412 {
413         down_write(&shrinker_rwsem);
414         list_add_tail(&shrinker->list, &shrinker_list);
415 #ifdef CONFIG_MEMCG_KMEM
416         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
417                 idr_replace(&shrinker_idr, shrinker, shrinker->id);
418 #endif
419         up_write(&shrinker_rwsem);
420 }
421
422 int register_shrinker(struct shrinker *shrinker)
423 {
424         int err = prealloc_shrinker(shrinker);
425
426         if (err)
427                 return err;
428         register_shrinker_prepared(shrinker);
429         return 0;
430 }
431 EXPORT_SYMBOL(register_shrinker);
432
433 /*
434  * Remove one
435  */
436 void unregister_shrinker(struct shrinker *shrinker)
437 {
438         if (!shrinker->nr_deferred)
439                 return;
440         if (shrinker->flags & SHRINKER_MEMCG_AWARE)
441                 unregister_memcg_shrinker(shrinker);
442         down_write(&shrinker_rwsem);
443         list_del(&shrinker->list);
444         up_write(&shrinker_rwsem);
445         kfree(shrinker->nr_deferred);
446         shrinker->nr_deferred = NULL;
447 }
448 EXPORT_SYMBOL(unregister_shrinker);
449
450 #define SHRINK_BATCH 128
451
452 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
453                                     struct shrinker *shrinker, int priority)
454 {
455         unsigned long freed = 0;
456         unsigned long long delta;
457         long total_scan;
458         long freeable;
459         long nr;
460         long new_nr;
461         int nid = shrinkctl->nid;
462         long batch_size = shrinker->batch ? shrinker->batch
463                                           : SHRINK_BATCH;
464         long scanned = 0, next_deferred;
465
466         if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
467                 nid = 0;
468
469         freeable = shrinker->count_objects(shrinker, shrinkctl);
470         if (freeable == 0 || freeable == SHRINK_EMPTY)
471                 return freeable;
472
473         /*
474          * copy the current shrinker scan count into a local variable
475          * and zero it so that other concurrent shrinker invocations
476          * don't also do this scanning work.
477          */
478         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
479
480         total_scan = nr;
481         if (shrinker->seeks) {
482                 delta = freeable >> priority;
483                 delta *= 4;
484                 do_div(delta, shrinker->seeks);
485         } else {
486                 /*
487                  * These objects don't require any IO to create. Trim
488                  * them aggressively under memory pressure to keep
489                  * them from causing refetches in the IO caches.
490                  */
491                 delta = freeable / 2;
492         }
493
494         total_scan += delta;
495         if (total_scan < 0) {
496                 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
497                        shrinker->scan_objects, total_scan);
498                 total_scan = freeable;
499                 next_deferred = nr;
500         } else
501                 next_deferred = total_scan;
502
503         /*
504          * We need to avoid excessive windup on filesystem shrinkers
505          * due to large numbers of GFP_NOFS allocations causing the
506          * shrinkers to return -1 all the time. This results in a large
507          * nr being built up so when a shrink that can do some work
508          * comes along it empties the entire cache due to nr >>>
509          * freeable. This is bad for sustaining a working set in
510          * memory.
511          *
512          * Hence only allow the shrinker to scan the entire cache when
513          * a large delta change is calculated directly.
514          */
515         if (delta < freeable / 4)
516                 total_scan = min(total_scan, freeable / 2);
517
518         /*
519          * Avoid risking looping forever due to too large nr value:
520          * never try to free more than twice the estimate number of
521          * freeable entries.
522          */
523         if (total_scan > freeable * 2)
524                 total_scan = freeable * 2;
525
526         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
527                                    freeable, delta, total_scan, priority);
528
529         /*
530          * Normally, we should not scan less than batch_size objects in one
531          * pass to avoid too frequent shrinker calls, but if the slab has less
532          * than batch_size objects in total and we are really tight on memory,
533          * we will try to reclaim all available objects, otherwise we can end
534          * up failing allocations although there are plenty of reclaimable
535          * objects spread over several slabs with usage less than the
536          * batch_size.
537          *
538          * We detect the "tight on memory" situations by looking at the total
539          * number of objects we want to scan (total_scan). If it is greater
540          * than the total number of objects on slab (freeable), we must be
541          * scanning at high prio and therefore should try to reclaim as much as
542          * possible.
543          */
544         while (total_scan >= batch_size ||
545                total_scan >= freeable) {
546                 unsigned long ret;
547                 unsigned long nr_to_scan = min(batch_size, total_scan);
548
549                 shrinkctl->nr_to_scan = nr_to_scan;
550                 shrinkctl->nr_scanned = nr_to_scan;
551                 ret = shrinker->scan_objects(shrinker, shrinkctl);
552                 if (ret == SHRINK_STOP)
553                         break;
554                 freed += ret;
555
556                 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
557                 total_scan -= shrinkctl->nr_scanned;
558                 scanned += shrinkctl->nr_scanned;
559
560                 cond_resched();
561         }
562
563         if (next_deferred >= scanned)
564                 next_deferred -= scanned;
565         else
566                 next_deferred = 0;
567         /*
568          * move the unused scan count back into the shrinker in a
569          * manner that handles concurrent updates. If we exhausted the
570          * scan, there is no need to do an update.
571          */
572         if (next_deferred > 0)
573                 new_nr = atomic_long_add_return(next_deferred,
574                                                 &shrinker->nr_deferred[nid]);
575         else
576                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
577
578         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
579         return freed;
580 }
581
582 #ifdef CONFIG_MEMCG_KMEM
583 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
584                         struct mem_cgroup *memcg, int priority)
585 {
586         struct memcg_shrinker_map *map;
587         unsigned long ret, freed = 0;
588         int i;
589
590         if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
591                 return 0;
592
593         if (!down_read_trylock(&shrinker_rwsem))
594                 return 0;
595
596         map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
597                                         true);
598         if (unlikely(!map))
599                 goto unlock;
600
601         for_each_set_bit(i, map->map, shrinker_nr_max) {
602                 struct shrink_control sc = {
603                         .gfp_mask = gfp_mask,
604                         .nid = nid,
605                         .memcg = memcg,
606                 };
607                 struct shrinker *shrinker;
608
609                 shrinker = idr_find(&shrinker_idr, i);
610                 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
611                         if (!shrinker)
612                                 clear_bit(i, map->map);
613                         continue;
614                 }
615
616                 ret = do_shrink_slab(&sc, shrinker, priority);
617                 if (ret == SHRINK_EMPTY) {
618                         clear_bit(i, map->map);
619                         /*
620                          * After the shrinker reported that it had no objects to
621                          * free, but before we cleared the corresponding bit in
622                          * the memcg shrinker map, a new object might have been
623                          * added. To make sure, we have the bit set in this
624                          * case, we invoke the shrinker one more time and reset
625                          * the bit if it reports that it is not empty anymore.
626                          * The memory barrier here pairs with the barrier in
627                          * memcg_set_shrinker_bit():
628                          *
629                          * list_lru_add()     shrink_slab_memcg()
630                          *   list_add_tail()    clear_bit()
631                          *   <MB>               <MB>
632                          *   set_bit()          do_shrink_slab()
633                          */
634                         smp_mb__after_atomic();
635                         ret = do_shrink_slab(&sc, shrinker, priority);
636                         if (ret == SHRINK_EMPTY)
637                                 ret = 0;
638                         else
639                                 memcg_set_shrinker_bit(memcg, nid, i);
640                 }
641                 freed += ret;
642
643                 if (rwsem_is_contended(&shrinker_rwsem)) {
644                         freed = freed ? : 1;
645                         break;
646                 }
647         }
648 unlock:
649         up_read(&shrinker_rwsem);
650         return freed;
651 }
652 #else /* CONFIG_MEMCG_KMEM */
653 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
654                         struct mem_cgroup *memcg, int priority)
655 {
656         return 0;
657 }
658 #endif /* CONFIG_MEMCG_KMEM */
659
660 /**
661  * shrink_slab - shrink slab caches
662  * @gfp_mask: allocation context
663  * @nid: node whose slab caches to target
664  * @memcg: memory cgroup whose slab caches to target
665  * @priority: the reclaim priority
666  *
667  * Call the shrink functions to age shrinkable caches.
668  *
669  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
670  * unaware shrinkers will receive a node id of 0 instead.
671  *
672  * @memcg specifies the memory cgroup to target. Unaware shrinkers
673  * are called only if it is the root cgroup.
674  *
675  * @priority is sc->priority, we take the number of objects and >> by priority
676  * in order to get the scan target.
677  *
678  * Returns the number of reclaimed slab objects.
679  */
680 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
681                                  struct mem_cgroup *memcg,
682                                  int priority)
683 {
684         unsigned long ret, freed = 0;
685         struct shrinker *shrinker;
686
687         if (!mem_cgroup_is_root(memcg))
688                 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
689
690         if (!down_read_trylock(&shrinker_rwsem))
691                 goto out;
692
693         list_for_each_entry(shrinker, &shrinker_list, list) {
694                 struct shrink_control sc = {
695                         .gfp_mask = gfp_mask,
696                         .nid = nid,
697                         .memcg = memcg,
698                 };
699
700                 ret = do_shrink_slab(&sc, shrinker, priority);
701                 if (ret == SHRINK_EMPTY)
702                         ret = 0;
703                 freed += ret;
704                 /*
705                  * Bail out if someone want to register a new shrinker to
706                  * prevent the regsitration from being stalled for long periods
707                  * by parallel ongoing shrinking.
708                  */
709                 if (rwsem_is_contended(&shrinker_rwsem)) {
710                         freed = freed ? : 1;
711                         break;
712                 }
713         }
714
715         up_read(&shrinker_rwsem);
716 out:
717         cond_resched();
718         return freed;
719 }
720
721 void drop_slab_node(int nid)
722 {
723         unsigned long freed;
724
725         do {
726                 struct mem_cgroup *memcg = NULL;
727
728                 freed = 0;
729                 memcg = mem_cgroup_iter(NULL, NULL, NULL);
730                 do {
731                         freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
732                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
733         } while (freed > 10);
734 }
735
736 void drop_slab(void)
737 {
738         int nid;
739
740         for_each_online_node(nid)
741                 drop_slab_node(nid);
742 }
743
744 static inline int is_page_cache_freeable(struct page *page)
745 {
746         /*
747          * A freeable page cache page is referenced only by the caller
748          * that isolated the page, the page cache and optional buffer
749          * heads at page->private.
750          */
751         int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
752                 HPAGE_PMD_NR : 1;
753         return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
754 }
755
756 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
757 {
758         if (current->flags & PF_SWAPWRITE)
759                 return 1;
760         if (!inode_write_congested(inode))
761                 return 1;
762         if (inode_to_bdi(inode) == current->backing_dev_info)
763                 return 1;
764         return 0;
765 }
766
767 /*
768  * We detected a synchronous write error writing a page out.  Probably
769  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
770  * fsync(), msync() or close().
771  *
772  * The tricky part is that after writepage we cannot touch the mapping: nothing
773  * prevents it from being freed up.  But we have a ref on the page and once
774  * that page is locked, the mapping is pinned.
775  *
776  * We're allowed to run sleeping lock_page() here because we know the caller has
777  * __GFP_FS.
778  */
779 static void handle_write_error(struct address_space *mapping,
780                                 struct page *page, int error)
781 {
782         lock_page(page);
783         if (page_mapping(page) == mapping)
784                 mapping_set_error(mapping, error);
785         unlock_page(page);
786 }
787
788 /* possible outcome of pageout() */
789 typedef enum {
790         /* failed to write page out, page is locked */
791         PAGE_KEEP,
792         /* move page to the active list, page is locked */
793         PAGE_ACTIVATE,
794         /* page has been sent to the disk successfully, page is unlocked */
795         PAGE_SUCCESS,
796         /* page is clean and locked */
797         PAGE_CLEAN,
798 } pageout_t;
799
800 /*
801  * pageout is called by shrink_page_list() for each dirty page.
802  * Calls ->writepage().
803  */
804 static pageout_t pageout(struct page *page, struct address_space *mapping,
805                          struct scan_control *sc)
806 {
807         /*
808          * If the page is dirty, only perform writeback if that write
809          * will be non-blocking.  To prevent this allocation from being
810          * stalled by pagecache activity.  But note that there may be
811          * stalls if we need to run get_block().  We could test
812          * PagePrivate for that.
813          *
814          * If this process is currently in __generic_file_write_iter() against
815          * this page's queue, we can perform writeback even if that
816          * will block.
817          *
818          * If the page is swapcache, write it back even if that would
819          * block, for some throttling. This happens by accident, because
820          * swap_backing_dev_info is bust: it doesn't reflect the
821          * congestion state of the swapdevs.  Easy to fix, if needed.
822          */
823         if (!is_page_cache_freeable(page))
824                 return PAGE_KEEP;
825         if (!mapping) {
826                 /*
827                  * Some data journaling orphaned pages can have
828                  * page->mapping == NULL while being dirty with clean buffers.
829                  */
830                 if (page_has_private(page)) {
831                         if (try_to_free_buffers(page)) {
832                                 ClearPageDirty(page);
833                                 pr_info("%s: orphaned page\n", __func__);
834                                 return PAGE_CLEAN;
835                         }
836                 }
837                 return PAGE_KEEP;
838         }
839         if (mapping->a_ops->writepage == NULL)
840                 return PAGE_ACTIVATE;
841         if (!may_write_to_inode(mapping->host, sc))
842                 return PAGE_KEEP;
843
844         if (clear_page_dirty_for_io(page)) {
845                 int res;
846                 struct writeback_control wbc = {
847                         .sync_mode = WB_SYNC_NONE,
848                         .nr_to_write = SWAP_CLUSTER_MAX,
849                         .range_start = 0,
850                         .range_end = LLONG_MAX,
851                         .for_reclaim = 1,
852                 };
853
854                 SetPageReclaim(page);
855                 res = mapping->a_ops->writepage(page, &wbc);
856                 if (res < 0)
857                         handle_write_error(mapping, page, res);
858                 if (res == AOP_WRITEPAGE_ACTIVATE) {
859                         ClearPageReclaim(page);
860                         return PAGE_ACTIVATE;
861                 }
862
863                 if (!PageWriteback(page)) {
864                         /* synchronous write or broken a_ops? */
865                         ClearPageReclaim(page);
866                 }
867                 trace_mm_vmscan_writepage(page);
868                 inc_node_page_state(page, NR_VMSCAN_WRITE);
869                 return PAGE_SUCCESS;
870         }
871
872         return PAGE_CLEAN;
873 }
874
875 /*
876  * Same as remove_mapping, but if the page is removed from the mapping, it
877  * gets returned with a refcount of 0.
878  */
879 static int __remove_mapping(struct address_space *mapping, struct page *page,
880                             bool reclaimed)
881 {
882         unsigned long flags;
883         int refcount;
884
885         BUG_ON(!PageLocked(page));
886         BUG_ON(mapping != page_mapping(page));
887
888         xa_lock_irqsave(&mapping->i_pages, flags);
889         /*
890          * The non racy check for a busy page.
891          *
892          * Must be careful with the order of the tests. When someone has
893          * a ref to the page, it may be possible that they dirty it then
894          * drop the reference. So if PageDirty is tested before page_count
895          * here, then the following race may occur:
896          *
897          * get_user_pages(&page);
898          * [user mapping goes away]
899          * write_to(page);
900          *                              !PageDirty(page)    [good]
901          * SetPageDirty(page);
902          * put_page(page);
903          *                              !page_count(page)   [good, discard it]
904          *
905          * [oops, our write_to data is lost]
906          *
907          * Reversing the order of the tests ensures such a situation cannot
908          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
909          * load is not satisfied before that of page->_refcount.
910          *
911          * Note that if SetPageDirty is always performed via set_page_dirty,
912          * and thus under the i_pages lock, then this ordering is not required.
913          */
914         if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
915                 refcount = 1 + HPAGE_PMD_NR;
916         else
917                 refcount = 2;
918         if (!page_ref_freeze(page, refcount))
919                 goto cannot_free;
920         /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
921         if (unlikely(PageDirty(page))) {
922                 page_ref_unfreeze(page, refcount);
923                 goto cannot_free;
924         }
925
926         if (PageSwapCache(page)) {
927                 swp_entry_t swap = { .val = page_private(page) };
928                 mem_cgroup_swapout(page, swap);
929                 __delete_from_swap_cache(page, swap);
930                 xa_unlock_irqrestore(&mapping->i_pages, flags);
931                 put_swap_page(page, swap);
932         } else {
933                 void (*freepage)(struct page *);
934                 void *shadow = NULL;
935
936                 freepage = mapping->a_ops->freepage;
937                 /*
938                  * Remember a shadow entry for reclaimed file cache in
939                  * order to detect refaults, thus thrashing, later on.
940                  *
941                  * But don't store shadows in an address space that is
942                  * already exiting.  This is not just an optizimation,
943                  * inode reclaim needs to empty out the radix tree or
944                  * the nodes are lost.  Don't plant shadows behind its
945                  * back.
946                  *
947                  * We also don't store shadows for DAX mappings because the
948                  * only page cache pages found in these are zero pages
949                  * covering holes, and because we don't want to mix DAX
950                  * exceptional entries and shadow exceptional entries in the
951                  * same address_space.
952                  */
953                 if (reclaimed && page_is_file_cache(page) &&
954                     !mapping_exiting(mapping) && !dax_mapping(mapping))
955                         shadow = workingset_eviction(page);
956                 __delete_from_page_cache(page, shadow);
957                 xa_unlock_irqrestore(&mapping->i_pages, flags);
958
959                 if (freepage != NULL)
960                         freepage(page);
961         }
962
963         return 1;
964
965 cannot_free:
966         xa_unlock_irqrestore(&mapping->i_pages, flags);
967         return 0;
968 }
969
970 /*
971  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
972  * someone else has a ref on the page, abort and return 0.  If it was
973  * successfully detached, return 1.  Assumes the caller has a single ref on
974  * this page.
975  */
976 int remove_mapping(struct address_space *mapping, struct page *page)
977 {
978         if (__remove_mapping(mapping, page, false)) {
979                 /*
980                  * Unfreezing the refcount with 1 rather than 2 effectively
981                  * drops the pagecache ref for us without requiring another
982                  * atomic operation.
983                  */
984                 page_ref_unfreeze(page, 1);
985                 return 1;
986         }
987         return 0;
988 }
989
990 /**
991  * putback_lru_page - put previously isolated page onto appropriate LRU list
992  * @page: page to be put back to appropriate lru list
993  *
994  * Add previously isolated @page to appropriate LRU list.
995  * Page may still be unevictable for other reasons.
996  *
997  * lru_lock must not be held, interrupts must be enabled.
998  */
999 void putback_lru_page(struct page *page)
1000 {
1001         lru_cache_add(page);
1002         put_page(page);         /* drop ref from isolate */
1003 }
1004
1005 enum page_references {
1006         PAGEREF_RECLAIM,
1007         PAGEREF_RECLAIM_CLEAN,
1008         PAGEREF_KEEP,
1009         PAGEREF_ACTIVATE,
1010 };
1011
1012 static enum page_references page_check_references(struct page *page,
1013                                                   struct scan_control *sc)
1014 {
1015         int referenced_ptes, referenced_page;
1016         unsigned long vm_flags;
1017
1018         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1019                                           &vm_flags);
1020         referenced_page = TestClearPageReferenced(page);
1021
1022         /*
1023          * Mlock lost the isolation race with us.  Let try_to_unmap()
1024          * move the page to the unevictable list.
1025          */
1026         if (vm_flags & VM_LOCKED)
1027                 return PAGEREF_RECLAIM;
1028
1029         if (referenced_ptes) {
1030                 if (PageSwapBacked(page))
1031                         return PAGEREF_ACTIVATE;
1032                 /*
1033                  * All mapped pages start out with page table
1034                  * references from the instantiating fault, so we need
1035                  * to look twice if a mapped file page is used more
1036                  * than once.
1037                  *
1038                  * Mark it and spare it for another trip around the
1039                  * inactive list.  Another page table reference will
1040                  * lead to its activation.
1041                  *
1042                  * Note: the mark is set for activated pages as well
1043                  * so that recently deactivated but used pages are
1044                  * quickly recovered.
1045                  */
1046                 SetPageReferenced(page);
1047
1048                 if (referenced_page || referenced_ptes > 1)
1049                         return PAGEREF_ACTIVATE;
1050
1051                 /*
1052                  * Activate file-backed executable pages after first usage.
1053                  */
1054                 if (vm_flags & VM_EXEC)
1055                         return PAGEREF_ACTIVATE;
1056
1057                 return PAGEREF_KEEP;
1058         }
1059
1060         /* Reclaim if clean, defer dirty pages to writeback */
1061         if (referenced_page && !PageSwapBacked(page))
1062                 return PAGEREF_RECLAIM_CLEAN;
1063
1064         return PAGEREF_RECLAIM;
1065 }
1066
1067 /* Check if a page is dirty or under writeback */
1068 static void page_check_dirty_writeback(struct page *page,
1069                                        bool *dirty, bool *writeback)
1070 {
1071         struct address_space *mapping;
1072
1073         /*
1074          * Anonymous pages are not handled by flushers and must be written
1075          * from reclaim context. Do not stall reclaim based on them
1076          */
1077         if (!page_is_file_cache(page) ||
1078             (PageAnon(page) && !PageSwapBacked(page))) {
1079                 *dirty = false;
1080                 *writeback = false;
1081                 return;
1082         }
1083
1084         /* By default assume that the page flags are accurate */
1085         *dirty = PageDirty(page);
1086         *writeback = PageWriteback(page);
1087
1088         /* Verify dirty/writeback state if the filesystem supports it */
1089         if (!page_has_private(page))
1090                 return;
1091
1092         mapping = page_mapping(page);
1093         if (mapping && mapping->a_ops->is_dirty_writeback)
1094                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1095 }
1096
1097 /*
1098  * shrink_page_list() returns the number of reclaimed pages
1099  */
1100 static unsigned long shrink_page_list(struct list_head *page_list,
1101                                       struct pglist_data *pgdat,
1102                                       struct scan_control *sc,
1103                                       enum ttu_flags ttu_flags,
1104                                       struct reclaim_stat *stat,
1105                                       bool force_reclaim)
1106 {
1107         LIST_HEAD(ret_pages);
1108         LIST_HEAD(free_pages);
1109         unsigned nr_reclaimed = 0;
1110         unsigned pgactivate = 0;
1111
1112         memset(stat, 0, sizeof(*stat));
1113         cond_resched();
1114
1115         while (!list_empty(page_list)) {
1116                 struct address_space *mapping;
1117                 struct page *page;
1118                 int may_enter_fs;
1119                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1120                 bool dirty, writeback;
1121
1122                 cond_resched();
1123
1124                 page = lru_to_page(page_list);
1125                 list_del(&page->lru);
1126
1127                 if (!trylock_page(page))
1128                         goto keep;
1129
1130                 VM_BUG_ON_PAGE(PageActive(page), page);
1131
1132                 sc->nr_scanned++;
1133
1134                 if (unlikely(!page_evictable(page)))
1135                         goto activate_locked;
1136
1137                 if (!sc->may_unmap && page_mapped(page))
1138                         goto keep_locked;
1139
1140                 /* Double the slab pressure for mapped and swapcache pages */
1141                 if ((page_mapped(page) || PageSwapCache(page)) &&
1142                     !(PageAnon(page) && !PageSwapBacked(page)))
1143                         sc->nr_scanned++;
1144
1145                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1146                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1147
1148                 /*
1149                  * The number of dirty pages determines if a node is marked
1150                  * reclaim_congested which affects wait_iff_congested. kswapd
1151                  * will stall and start writing pages if the tail of the LRU
1152                  * is all dirty unqueued pages.
1153                  */
1154                 page_check_dirty_writeback(page, &dirty, &writeback);
1155                 if (dirty || writeback)
1156                         stat->nr_dirty++;
1157
1158                 if (dirty && !writeback)
1159                         stat->nr_unqueued_dirty++;
1160
1161                 /*
1162                  * Treat this page as congested if the underlying BDI is or if
1163                  * pages are cycling through the LRU so quickly that the
1164                  * pages marked for immediate reclaim are making it to the
1165                  * end of the LRU a second time.
1166                  */
1167                 mapping = page_mapping(page);
1168                 if (((dirty || writeback) && mapping &&
1169                      inode_write_congested(mapping->host)) ||
1170                     (writeback && PageReclaim(page)))
1171                         stat->nr_congested++;
1172
1173                 /*
1174                  * If a page at the tail of the LRU is under writeback, there
1175                  * are three cases to consider.
1176                  *
1177                  * 1) If reclaim is encountering an excessive number of pages
1178                  *    under writeback and this page is both under writeback and
1179                  *    PageReclaim then it indicates that pages are being queued
1180                  *    for IO but are being recycled through the LRU before the
1181                  *    IO can complete. Waiting on the page itself risks an
1182                  *    indefinite stall if it is impossible to writeback the
1183                  *    page due to IO error or disconnected storage so instead
1184                  *    note that the LRU is being scanned too quickly and the
1185                  *    caller can stall after page list has been processed.
1186                  *
1187                  * 2) Global or new memcg reclaim encounters a page that is
1188                  *    not marked for immediate reclaim, or the caller does not
1189                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1190                  *    not to fs). In this case mark the page for immediate
1191                  *    reclaim and continue scanning.
1192                  *
1193                  *    Require may_enter_fs because we would wait on fs, which
1194                  *    may not have submitted IO yet. And the loop driver might
1195                  *    enter reclaim, and deadlock if it waits on a page for
1196                  *    which it is needed to do the write (loop masks off
1197                  *    __GFP_IO|__GFP_FS for this reason); but more thought
1198                  *    would probably show more reasons.
1199                  *
1200                  * 3) Legacy memcg encounters a page that is already marked
1201                  *    PageReclaim. memcg does not have any dirty pages
1202                  *    throttling so we could easily OOM just because too many
1203                  *    pages are in writeback and there is nothing else to
1204                  *    reclaim. Wait for the writeback to complete.
1205                  *
1206                  * In cases 1) and 2) we activate the pages to get them out of
1207                  * the way while we continue scanning for clean pages on the
1208                  * inactive list and refilling from the active list. The
1209                  * observation here is that waiting for disk writes is more
1210                  * expensive than potentially causing reloads down the line.
1211                  * Since they're marked for immediate reclaim, they won't put
1212                  * memory pressure on the cache working set any longer than it
1213                  * takes to write them to disk.
1214                  */
1215                 if (PageWriteback(page)) {
1216                         /* Case 1 above */
1217                         if (current_is_kswapd() &&
1218                             PageReclaim(page) &&
1219                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1220                                 stat->nr_immediate++;
1221                                 goto activate_locked;
1222
1223                         /* Case 2 above */
1224                         } else if (sane_reclaim(sc) ||
1225                             !PageReclaim(page) || !may_enter_fs) {
1226                                 /*
1227                                  * This is slightly racy - end_page_writeback()
1228                                  * might have just cleared PageReclaim, then
1229                                  * setting PageReclaim here end up interpreted
1230                                  * as PageReadahead - but that does not matter
1231                                  * enough to care.  What we do want is for this
1232                                  * page to have PageReclaim set next time memcg
1233                                  * reclaim reaches the tests above, so it will
1234                                  * then wait_on_page_writeback() to avoid OOM;
1235                                  * and it's also appropriate in global reclaim.
1236                                  */
1237                                 SetPageReclaim(page);
1238                                 stat->nr_writeback++;
1239                                 goto activate_locked;
1240
1241                         /* Case 3 above */
1242                         } else {
1243                                 unlock_page(page);
1244                                 wait_on_page_writeback(page);
1245                                 /* then go back and try same page again */
1246                                 list_add_tail(&page->lru, page_list);
1247                                 continue;
1248                         }
1249                 }
1250
1251                 if (!force_reclaim)
1252                         references = page_check_references(page, sc);
1253
1254                 switch (references) {
1255                 case PAGEREF_ACTIVATE:
1256                         goto activate_locked;
1257                 case PAGEREF_KEEP:
1258                         stat->nr_ref_keep++;
1259                         goto keep_locked;
1260                 case PAGEREF_RECLAIM:
1261                 case PAGEREF_RECLAIM_CLEAN:
1262                         ; /* try to reclaim the page below */
1263                 }
1264
1265                 /*
1266                  * Anonymous process memory has backing store?
1267                  * Try to allocate it some swap space here.
1268                  * Lazyfree page could be freed directly
1269                  */
1270                 if (PageAnon(page) && PageSwapBacked(page)) {
1271                         if (!PageSwapCache(page)) {
1272                                 if (!(sc->gfp_mask & __GFP_IO))
1273                                         goto keep_locked;
1274                                 if (PageTransHuge(page)) {
1275                                         /* cannot split THP, skip it */
1276                                         if (!can_split_huge_page(page, NULL))
1277                                                 goto activate_locked;
1278                                         /*
1279                                          * Split pages without a PMD map right
1280                                          * away. Chances are some or all of the
1281                                          * tail pages can be freed without IO.
1282                                          */
1283                                         if (!compound_mapcount(page) &&
1284                                             split_huge_page_to_list(page,
1285                                                                     page_list))
1286                                                 goto activate_locked;
1287                                 }
1288                                 if (!add_to_swap(page)) {
1289                                         if (!PageTransHuge(page))
1290                                                 goto activate_locked;
1291                                         /* Fallback to swap normal pages */
1292                                         if (split_huge_page_to_list(page,
1293                                                                     page_list))
1294                                                 goto activate_locked;
1295 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1296                                         count_vm_event(THP_SWPOUT_FALLBACK);
1297 #endif
1298                                         if (!add_to_swap(page))
1299                                                 goto activate_locked;
1300                                 }
1301
1302                                 may_enter_fs = 1;
1303
1304                                 /* Adding to swap updated mapping */
1305                                 mapping = page_mapping(page);
1306                         }
1307                 } else if (unlikely(PageTransHuge(page))) {
1308                         /* Split file THP */
1309                         if (split_huge_page_to_list(page, page_list))
1310                                 goto keep_locked;
1311                 }
1312
1313                 /*
1314                  * The page is mapped into the page tables of one or more
1315                  * processes. Try to unmap it here.
1316                  */
1317                 if (page_mapped(page)) {
1318                         enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1319
1320                         if (unlikely(PageTransHuge(page)))
1321                                 flags |= TTU_SPLIT_HUGE_PMD;
1322                         if (!try_to_unmap(page, flags)) {
1323                                 stat->nr_unmap_fail++;
1324                                 goto activate_locked;
1325                         }
1326                 }
1327
1328                 if (PageDirty(page)) {
1329                         /*
1330                          * Only kswapd can writeback filesystem pages
1331                          * to avoid risk of stack overflow. But avoid
1332                          * injecting inefficient single-page IO into
1333                          * flusher writeback as much as possible: only
1334                          * write pages when we've encountered many
1335                          * dirty pages, and when we've already scanned
1336                          * the rest of the LRU for clean pages and see
1337                          * the same dirty pages again (PageReclaim).
1338                          */
1339                         if (page_is_file_cache(page) &&
1340                             (!current_is_kswapd() || !PageReclaim(page) ||
1341                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1342                                 /*
1343                                  * Immediately reclaim when written back.
1344                                  * Similar in principal to deactivate_page()
1345                                  * except we already have the page isolated
1346                                  * and know it's dirty
1347                                  */
1348                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1349                                 SetPageReclaim(page);
1350
1351                                 goto activate_locked;
1352                         }
1353
1354                         if (references == PAGEREF_RECLAIM_CLEAN)
1355                                 goto keep_locked;
1356                         if (!may_enter_fs)
1357                                 goto keep_locked;
1358                         if (!sc->may_writepage)
1359                                 goto keep_locked;
1360
1361                         /*
1362                          * Page is dirty. Flush the TLB if a writable entry
1363                          * potentially exists to avoid CPU writes after IO
1364                          * starts and then write it out here.
1365                          */
1366                         try_to_unmap_flush_dirty();
1367                         switch (pageout(page, mapping, sc)) {
1368                         case PAGE_KEEP:
1369                                 goto keep_locked;
1370                         case PAGE_ACTIVATE:
1371                                 goto activate_locked;
1372                         case PAGE_SUCCESS:
1373                                 if (PageWriteback(page))
1374                                         goto keep;
1375                                 if (PageDirty(page))
1376                                         goto keep;
1377
1378                                 /*
1379                                  * A synchronous write - probably a ramdisk.  Go
1380                                  * ahead and try to reclaim the page.
1381                                  */
1382                                 if (!trylock_page(page))
1383                                         goto keep;
1384                                 if (PageDirty(page) || PageWriteback(page))
1385                                         goto keep_locked;
1386                                 mapping = page_mapping(page);
1387                         case PAGE_CLEAN:
1388                                 ; /* try to free the page below */
1389                         }
1390                 }
1391
1392                 /*
1393                  * If the page has buffers, try to free the buffer mappings
1394                  * associated with this page. If we succeed we try to free
1395                  * the page as well.
1396                  *
1397                  * We do this even if the page is PageDirty().
1398                  * try_to_release_page() does not perform I/O, but it is
1399                  * possible for a page to have PageDirty set, but it is actually
1400                  * clean (all its buffers are clean).  This happens if the
1401                  * buffers were written out directly, with submit_bh(). ext3
1402                  * will do this, as well as the blockdev mapping.
1403                  * try_to_release_page() will discover that cleanness and will
1404                  * drop the buffers and mark the page clean - it can be freed.
1405                  *
1406                  * Rarely, pages can have buffers and no ->mapping.  These are
1407                  * the pages which were not successfully invalidated in
1408                  * truncate_complete_page().  We try to drop those buffers here
1409                  * and if that worked, and the page is no longer mapped into
1410                  * process address space (page_count == 1) it can be freed.
1411                  * Otherwise, leave the page on the LRU so it is swappable.
1412                  */
1413                 if (page_has_private(page)) {
1414                         if (!try_to_release_page(page, sc->gfp_mask))
1415                                 goto activate_locked;
1416                         if (!mapping && page_count(page) == 1) {
1417                                 unlock_page(page);
1418                                 if (put_page_testzero(page))
1419                                         goto free_it;
1420                                 else {
1421                                         /*
1422                                          * rare race with speculative reference.
1423                                          * the speculative reference will free
1424                                          * this page shortly, so we may
1425                                          * increment nr_reclaimed here (and
1426                                          * leave it off the LRU).
1427                                          */
1428                                         nr_reclaimed++;
1429                                         continue;
1430                                 }
1431                         }
1432                 }
1433
1434                 if (PageAnon(page) && !PageSwapBacked(page)) {
1435                         /* follow __remove_mapping for reference */
1436                         if (!page_ref_freeze(page, 1))
1437                                 goto keep_locked;
1438                         if (PageDirty(page)) {
1439                                 page_ref_unfreeze(page, 1);
1440                                 goto keep_locked;
1441                         }
1442
1443                         count_vm_event(PGLAZYFREED);
1444                         count_memcg_page_event(page, PGLAZYFREED);
1445                 } else if (!mapping || !__remove_mapping(mapping, page, true))
1446                         goto keep_locked;
1447
1448                 unlock_page(page);
1449 free_it:
1450                 nr_reclaimed++;
1451
1452                 /*
1453                  * Is there need to periodically free_page_list? It would
1454                  * appear not as the counts should be low
1455                  */
1456                 if (unlikely(PageTransHuge(page))) {
1457                         mem_cgroup_uncharge(page);
1458                         (*get_compound_page_dtor(page))(page);
1459                 } else
1460                         list_add(&page->lru, &free_pages);
1461                 continue;
1462
1463 activate_locked:
1464                 /* Not a candidate for swapping, so reclaim swap space. */
1465                 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1466                                                 PageMlocked(page)))
1467                         try_to_free_swap(page);
1468                 VM_BUG_ON_PAGE(PageActive(page), page);
1469                 if (!PageMlocked(page)) {
1470                         int type = page_is_file_cache(page);
1471                         SetPageActive(page);
1472                         pgactivate++;
1473                         stat->nr_activate[type] += hpage_nr_pages(page);
1474                         count_memcg_page_event(page, PGACTIVATE);
1475                 }
1476 keep_locked:
1477                 unlock_page(page);
1478 keep:
1479                 list_add(&page->lru, &ret_pages);
1480                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1481         }
1482
1483         mem_cgroup_uncharge_list(&free_pages);
1484         try_to_unmap_flush();
1485         free_unref_page_list(&free_pages);
1486
1487         list_splice(&ret_pages, page_list);
1488         count_vm_events(PGACTIVATE, pgactivate);
1489
1490         return nr_reclaimed;
1491 }
1492
1493 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1494                                             struct list_head *page_list)
1495 {
1496         struct scan_control sc = {
1497                 .gfp_mask = GFP_KERNEL,
1498                 .priority = DEF_PRIORITY,
1499                 .may_unmap = 1,
1500         };
1501         struct reclaim_stat dummy_stat;
1502         unsigned long ret;
1503         struct page *page, *next;
1504         LIST_HEAD(clean_pages);
1505
1506         list_for_each_entry_safe(page, next, page_list, lru) {
1507                 if (page_is_file_cache(page) && !PageDirty(page) &&
1508                     !__PageMovable(page)) {
1509                         ClearPageActive(page);
1510                         list_move(&page->lru, &clean_pages);
1511                 }
1512         }
1513
1514         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1515                         TTU_IGNORE_ACCESS, &dummy_stat, true);
1516         list_splice(&clean_pages, page_list);
1517         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1518         return ret;
1519 }
1520
1521 /*
1522  * Attempt to remove the specified page from its LRU.  Only take this page
1523  * if it is of the appropriate PageActive status.  Pages which are being
1524  * freed elsewhere are also ignored.
1525  *
1526  * page:        page to consider
1527  * mode:        one of the LRU isolation modes defined above
1528  *
1529  * returns 0 on success, -ve errno on failure.
1530  */
1531 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1532 {
1533         int ret = -EINVAL;
1534
1535         /* Only take pages on the LRU. */
1536         if (!PageLRU(page))
1537                 return ret;
1538
1539         /* Compaction should not handle unevictable pages but CMA can do so */
1540         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1541                 return ret;
1542
1543         ret = -EBUSY;
1544
1545         /*
1546          * To minimise LRU disruption, the caller can indicate that it only
1547          * wants to isolate pages it will be able to operate on without
1548          * blocking - clean pages for the most part.
1549          *
1550          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1551          * that it is possible to migrate without blocking
1552          */
1553         if (mode & ISOLATE_ASYNC_MIGRATE) {
1554                 /* All the caller can do on PageWriteback is block */
1555                 if (PageWriteback(page))
1556                         return ret;
1557
1558                 if (PageDirty(page)) {
1559                         struct address_space *mapping;
1560                         bool migrate_dirty;
1561
1562                         /*
1563                          * Only pages without mappings or that have a
1564                          * ->migratepage callback are possible to migrate
1565                          * without blocking. However, we can be racing with
1566                          * truncation so it's necessary to lock the page
1567                          * to stabilise the mapping as truncation holds
1568                          * the page lock until after the page is removed
1569                          * from the page cache.
1570                          */
1571                         if (!trylock_page(page))
1572                                 return ret;
1573
1574                         mapping = page_mapping(page);
1575                         migrate_dirty = !mapping || mapping->a_ops->migratepage;
1576                         unlock_page(page);
1577                         if (!migrate_dirty)
1578                                 return ret;
1579                 }
1580         }
1581
1582         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1583                 return ret;
1584
1585         if (likely(get_page_unless_zero(page))) {
1586                 /*
1587                  * Be careful not to clear PageLRU until after we're
1588                  * sure the page is not being freed elsewhere -- the
1589                  * page release code relies on it.
1590                  */
1591                 ClearPageLRU(page);
1592                 ret = 0;
1593         }
1594
1595         return ret;
1596 }
1597
1598
1599 /*
1600  * Update LRU sizes after isolating pages. The LRU size updates must
1601  * be complete before mem_cgroup_update_lru_size due to a santity check.
1602  */
1603 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1604                         enum lru_list lru, unsigned long *nr_zone_taken)
1605 {
1606         int zid;
1607
1608         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1609                 if (!nr_zone_taken[zid])
1610                         continue;
1611
1612                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1613 #ifdef CONFIG_MEMCG
1614                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1615 #endif
1616         }
1617
1618 }
1619
1620 /**
1621  * pgdat->lru_lock is heavily contended.  Some of the functions that
1622  * shrink the lists perform better by taking out a batch of pages
1623  * and working on them outside the LRU lock.
1624  *
1625  * For pagecache intensive workloads, this function is the hottest
1626  * spot in the kernel (apart from copy_*_user functions).
1627  *
1628  * Appropriate locks must be held before calling this function.
1629  *
1630  * @nr_to_scan: The number of eligible pages to look through on the list.
1631  * @lruvec:     The LRU vector to pull pages from.
1632  * @dst:        The temp list to put pages on to.
1633  * @nr_scanned: The number of pages that were scanned.
1634  * @sc:         The scan_control struct for this reclaim session
1635  * @mode:       One of the LRU isolation modes
1636  * @lru:        LRU list id for isolating
1637  *
1638  * returns how many pages were moved onto *@dst.
1639  */
1640 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1641                 struct lruvec *lruvec, struct list_head *dst,
1642                 unsigned long *nr_scanned, struct scan_control *sc,
1643                 enum lru_list lru)
1644 {
1645         struct list_head *src = &lruvec->lists[lru];
1646         unsigned long nr_taken = 0;
1647         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1648         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1649         unsigned long skipped = 0;
1650         unsigned long scan, total_scan, nr_pages;
1651         LIST_HEAD(pages_skipped);
1652         isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1653
1654         scan = 0;
1655         for (total_scan = 0;
1656              scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1657              total_scan++) {
1658                 struct page *page;
1659
1660                 page = lru_to_page(src);
1661                 prefetchw_prev_lru_page(page, src, flags);
1662
1663                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1664
1665                 if (page_zonenum(page) > sc->reclaim_idx) {
1666                         list_move(&page->lru, &pages_skipped);
1667                         nr_skipped[page_zonenum(page)]++;
1668                         continue;
1669                 }
1670
1671                 /*
1672                  * Do not count skipped pages because that makes the function
1673                  * return with no isolated pages if the LRU mostly contains
1674                  * ineligible pages.  This causes the VM to not reclaim any
1675                  * pages, triggering a premature OOM.
1676                  */
1677                 scan++;
1678                 switch (__isolate_lru_page(page, mode)) {
1679                 case 0:
1680                         nr_pages = hpage_nr_pages(page);
1681                         nr_taken += nr_pages;
1682                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1683                         list_move(&page->lru, dst);
1684                         break;
1685
1686                 case -EBUSY:
1687                         /* else it is being freed elsewhere */
1688                         list_move(&page->lru, src);
1689                         continue;
1690
1691                 default:
1692                         BUG();
1693                 }
1694         }
1695
1696         /*
1697          * Splice any skipped pages to the start of the LRU list. Note that
1698          * this disrupts the LRU order when reclaiming for lower zones but
1699          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1700          * scanning would soon rescan the same pages to skip and put the
1701          * system at risk of premature OOM.
1702          */
1703         if (!list_empty(&pages_skipped)) {
1704                 int zid;
1705
1706                 list_splice(&pages_skipped, src);
1707                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1708                         if (!nr_skipped[zid])
1709                                 continue;
1710
1711                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1712                         skipped += nr_skipped[zid];
1713                 }
1714         }
1715         *nr_scanned = total_scan;
1716         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1717                                     total_scan, skipped, nr_taken, mode, lru);
1718         update_lru_sizes(lruvec, lru, nr_zone_taken);
1719         return nr_taken;
1720 }
1721
1722 /**
1723  * isolate_lru_page - tries to isolate a page from its LRU list
1724  * @page: page to isolate from its LRU list
1725  *
1726  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1727  * vmstat statistic corresponding to whatever LRU list the page was on.
1728  *
1729  * Returns 0 if the page was removed from an LRU list.
1730  * Returns -EBUSY if the page was not on an LRU list.
1731  *
1732  * The returned page will have PageLRU() cleared.  If it was found on
1733  * the active list, it will have PageActive set.  If it was found on
1734  * the unevictable list, it will have the PageUnevictable bit set. That flag
1735  * may need to be cleared by the caller before letting the page go.
1736  *
1737  * The vmstat statistic corresponding to the list on which the page was
1738  * found will be decremented.
1739  *
1740  * Restrictions:
1741  *
1742  * (1) Must be called with an elevated refcount on the page. This is a
1743  *     fundamentnal difference from isolate_lru_pages (which is called
1744  *     without a stable reference).
1745  * (2) the lru_lock must not be held.
1746  * (3) interrupts must be enabled.
1747  */
1748 int isolate_lru_page(struct page *page)
1749 {
1750         int ret = -EBUSY;
1751
1752         VM_BUG_ON_PAGE(!page_count(page), page);
1753         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1754
1755         if (PageLRU(page)) {
1756                 pg_data_t *pgdat = page_pgdat(page);
1757                 struct lruvec *lruvec;
1758
1759                 spin_lock_irq(&pgdat->lru_lock);
1760                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1761                 if (PageLRU(page)) {
1762                         int lru = page_lru(page);
1763                         get_page(page);
1764                         ClearPageLRU(page);
1765                         del_page_from_lru_list(page, lruvec, lru);
1766                         ret = 0;
1767                 }
1768                 spin_unlock_irq(&pgdat->lru_lock);
1769         }
1770         return ret;
1771 }
1772
1773 /*
1774  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1775  * then get resheduled. When there are massive number of tasks doing page
1776  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1777  * the LRU list will go small and be scanned faster than necessary, leading to
1778  * unnecessary swapping, thrashing and OOM.
1779  */
1780 static int too_many_isolated(struct pglist_data *pgdat, int file,
1781                 struct scan_control *sc)
1782 {
1783         unsigned long inactive, isolated;
1784
1785         if (current_is_kswapd())
1786                 return 0;
1787
1788         if (!sane_reclaim(sc))
1789                 return 0;
1790
1791         if (file) {
1792                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1793                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1794         } else {
1795                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1796                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1797         }
1798
1799         /*
1800          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1801          * won't get blocked by normal direct-reclaimers, forming a circular
1802          * deadlock.
1803          */
1804         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1805                 inactive >>= 3;
1806
1807         return isolated > inactive;
1808 }
1809
1810 /*
1811  * This moves pages from @list to corresponding LRU list.
1812  *
1813  * We move them the other way if the page is referenced by one or more
1814  * processes, from rmap.
1815  *
1816  * If the pages are mostly unmapped, the processing is fast and it is
1817  * appropriate to hold zone_lru_lock across the whole operation.  But if
1818  * the pages are mapped, the processing is slow (page_referenced()) so we
1819  * should drop zone_lru_lock around each page.  It's impossible to balance
1820  * this, so instead we remove the pages from the LRU while processing them.
1821  * It is safe to rely on PG_active against the non-LRU pages in here because
1822  * nobody will play with that bit on a non-LRU page.
1823  *
1824  * The downside is that we have to touch page->_refcount against each page.
1825  * But we had to alter page->flags anyway.
1826  *
1827  * Returns the number of pages moved to the given lruvec.
1828  */
1829
1830 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1831                                                      struct list_head *list)
1832 {
1833         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1834         int nr_pages, nr_moved = 0;
1835         LIST_HEAD(pages_to_free);
1836         struct page *page;
1837         enum lru_list lru;
1838
1839         while (!list_empty(list)) {
1840                 page = lru_to_page(list);
1841                 VM_BUG_ON_PAGE(PageLRU(page), page);
1842                 if (unlikely(!page_evictable(page))) {
1843                         list_del(&page->lru);
1844                         spin_unlock_irq(&pgdat->lru_lock);
1845                         putback_lru_page(page);
1846                         spin_lock_irq(&pgdat->lru_lock);
1847                         continue;
1848                 }
1849                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1850
1851                 SetPageLRU(page);
1852                 lru = page_lru(page);
1853
1854                 nr_pages = hpage_nr_pages(page);
1855                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1856                 list_move(&page->lru, &lruvec->lists[lru]);
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                 } else {
1871                         nr_moved += nr_pages;
1872                 }
1873         }
1874
1875         /*
1876          * To save our caller's stack, now use input list for pages to free.
1877          */
1878         list_splice(&pages_to_free, list);
1879
1880         return nr_moved;
1881 }
1882
1883 /*
1884  * If a kernel thread (such as nfsd for loop-back mounts) services
1885  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1886  * In that case we should only throttle if the backing device it is
1887  * writing to is congested.  In other cases it is safe to throttle.
1888  */
1889 static int current_may_throttle(void)
1890 {
1891         return !(current->flags & PF_LESS_THROTTLE) ||
1892                 current->backing_dev_info == NULL ||
1893                 bdi_write_congested(current->backing_dev_info);
1894 }
1895
1896 /*
1897  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1898  * of reclaimed pages
1899  */
1900 static noinline_for_stack unsigned long
1901 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1902                      struct scan_control *sc, enum lru_list lru)
1903 {
1904         LIST_HEAD(page_list);
1905         unsigned long nr_scanned;
1906         unsigned long nr_reclaimed = 0;
1907         unsigned long nr_taken;
1908         struct reclaim_stat stat;
1909         int file = is_file_lru(lru);
1910         enum vm_event_item item;
1911         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1912         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1913         bool stalled = false;
1914
1915         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1916                 if (stalled)
1917                         return 0;
1918
1919                 /* wait a bit for the reclaimer. */
1920                 msleep(100);
1921                 stalled = true;
1922
1923                 /* We are about to die and free our memory. Return now. */
1924                 if (fatal_signal_pending(current))
1925                         return SWAP_CLUSTER_MAX;
1926         }
1927
1928         lru_add_drain();
1929
1930         spin_lock_irq(&pgdat->lru_lock);
1931
1932         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1933                                      &nr_scanned, sc, lru);
1934
1935         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1936         reclaim_stat->recent_scanned[file] += nr_taken;
1937
1938         item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1939         if (global_reclaim(sc))
1940                 __count_vm_events(item, nr_scanned);
1941         __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1942         spin_unlock_irq(&pgdat->lru_lock);
1943
1944         if (nr_taken == 0)
1945                 return 0;
1946
1947         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1948                                 &stat, false);
1949
1950         spin_lock_irq(&pgdat->lru_lock);
1951
1952         item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1953         if (global_reclaim(sc))
1954                 __count_vm_events(item, nr_reclaimed);
1955         __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1956         reclaim_stat->recent_rotated[0] = stat.nr_activate[0];
1957         reclaim_stat->recent_rotated[1] = stat.nr_activate[1];
1958
1959         move_pages_to_lru(lruvec, &page_list);
1960
1961         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1962
1963         spin_unlock_irq(&pgdat->lru_lock);
1964
1965         mem_cgroup_uncharge_list(&page_list);
1966         free_unref_page_list(&page_list);
1967
1968         /*
1969          * If dirty pages are scanned that are not queued for IO, it
1970          * implies that flushers are not doing their job. This can
1971          * happen when memory pressure pushes dirty pages to the end of
1972          * the LRU before the dirty limits are breached and the dirty
1973          * data has expired. It can also happen when the proportion of
1974          * dirty pages grows not through writes but through memory
1975          * pressure reclaiming all the clean cache. And in some cases,
1976          * the flushers simply cannot keep up with the allocation
1977          * rate. Nudge the flusher threads in case they are asleep.
1978          */
1979         if (stat.nr_unqueued_dirty == nr_taken)
1980                 wakeup_flusher_threads(WB_REASON_VMSCAN);
1981
1982         sc->nr.dirty += stat.nr_dirty;
1983         sc->nr.congested += stat.nr_congested;
1984         sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1985         sc->nr.writeback += stat.nr_writeback;
1986         sc->nr.immediate += stat.nr_immediate;
1987         sc->nr.taken += nr_taken;
1988         if (file)
1989                 sc->nr.file_taken += nr_taken;
1990
1991         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1992                         nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1993         return nr_reclaimed;
1994 }
1995
1996 static void shrink_active_list(unsigned long nr_to_scan,
1997                                struct lruvec *lruvec,
1998                                struct scan_control *sc,
1999                                enum lru_list lru)
2000 {
2001         unsigned long nr_taken;
2002         unsigned long nr_scanned;
2003         unsigned long vm_flags;
2004         LIST_HEAD(l_hold);      /* The pages which were snipped off */
2005         LIST_HEAD(l_active);
2006         LIST_HEAD(l_inactive);
2007         struct page *page;
2008         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2009         unsigned nr_deactivate, nr_activate;
2010         unsigned nr_rotated = 0;
2011         int file = is_file_lru(lru);
2012         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2013
2014         lru_add_drain();
2015
2016         spin_lock_irq(&pgdat->lru_lock);
2017
2018         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2019                                      &nr_scanned, sc, lru);
2020
2021         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2022         reclaim_stat->recent_scanned[file] += nr_taken;
2023
2024         __count_vm_events(PGREFILL, nr_scanned);
2025         __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2026
2027         spin_unlock_irq(&pgdat->lru_lock);
2028
2029         while (!list_empty(&l_hold)) {
2030                 cond_resched();
2031                 page = lru_to_page(&l_hold);
2032                 list_del(&page->lru);
2033
2034                 if (unlikely(!page_evictable(page))) {
2035                         putback_lru_page(page);
2036                         continue;
2037                 }
2038
2039                 if (unlikely(buffer_heads_over_limit)) {
2040                         if (page_has_private(page) && trylock_page(page)) {
2041                                 if (page_has_private(page))
2042                                         try_to_release_page(page, 0);
2043                                 unlock_page(page);
2044                         }
2045                 }
2046
2047                 if (page_referenced(page, 0, sc->target_mem_cgroup,
2048                                     &vm_flags)) {
2049                         nr_rotated += hpage_nr_pages(page);
2050                         /*
2051                          * Identify referenced, file-backed active pages and
2052                          * give them one more trip around the active list. So
2053                          * that executable code get better chances to stay in
2054                          * memory under moderate memory pressure.  Anon pages
2055                          * are not likely to be evicted by use-once streaming
2056                          * IO, plus JVM can create lots of anon VM_EXEC pages,
2057                          * so we ignore them here.
2058                          */
2059                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2060                                 list_add(&page->lru, &l_active);
2061                                 continue;
2062                         }
2063                 }
2064
2065                 ClearPageActive(page);  /* we are de-activating */
2066                 SetPageWorkingset(page);
2067                 list_add(&page->lru, &l_inactive);
2068         }
2069
2070         /*
2071          * Move pages back to the lru list.
2072          */
2073         spin_lock_irq(&pgdat->lru_lock);
2074         /*
2075          * Count referenced pages from currently used mappings as rotated,
2076          * even though only some of them are actually re-activated.  This
2077          * helps balance scan pressure between file and anonymous pages in
2078          * get_scan_count.
2079          */
2080         reclaim_stat->recent_rotated[file] += nr_rotated;
2081
2082         nr_activate = move_pages_to_lru(lruvec, &l_active);
2083         nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2084         /* Keep all free pages in l_active list */
2085         list_splice(&l_inactive, &l_active);
2086
2087         __count_vm_events(PGDEACTIVATE, nr_deactivate);
2088         __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2089
2090         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2091         spin_unlock_irq(&pgdat->lru_lock);
2092
2093         mem_cgroup_uncharge_list(&l_active);
2094         free_unref_page_list(&l_active);
2095         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2096                         nr_deactivate, nr_rotated, sc->priority, file);
2097 }
2098
2099 /*
2100  * The inactive anon list should be small enough that the VM never has
2101  * to do too much work.
2102  *
2103  * The inactive file list should be small enough to leave most memory
2104  * to the established workingset on the scan-resistant active list,
2105  * but large enough to avoid thrashing the aggregate readahead window.
2106  *
2107  * Both inactive lists should also be large enough that each inactive
2108  * page has a chance to be referenced again before it is reclaimed.
2109  *
2110  * If that fails and refaulting is observed, the inactive list grows.
2111  *
2112  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2113  * on this LRU, maintained by the pageout code. An inactive_ratio
2114  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2115  *
2116  * total     target    max
2117  * memory    ratio     inactive
2118  * -------------------------------------
2119  *   10MB       1         5MB
2120  *  100MB       1        50MB
2121  *    1GB       3       250MB
2122  *   10GB      10       0.9GB
2123  *  100GB      31         3GB
2124  *    1TB     101        10GB
2125  *   10TB     320        32GB
2126  */
2127 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2128                                  struct scan_control *sc, bool actual_reclaim)
2129 {
2130         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2131         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2132         enum lru_list inactive_lru = file * LRU_FILE;
2133         unsigned long inactive, active;
2134         unsigned long inactive_ratio;
2135         unsigned long refaults;
2136         unsigned long gb;
2137
2138         /*
2139          * If we don't have swap space, anonymous page deactivation
2140          * is pointless.
2141          */
2142         if (!file && !total_swap_pages)
2143                 return false;
2144
2145         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2146         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2147
2148         /*
2149          * When refaults are being observed, it means a new workingset
2150          * is being established. Disable active list protection to get
2151          * rid of the stale workingset quickly.
2152          */
2153         refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2154         if (file && actual_reclaim && lruvec->refaults != refaults) {
2155                 inactive_ratio = 0;
2156         } else {
2157                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2158                 if (gb)
2159                         inactive_ratio = int_sqrt(10 * gb);
2160                 else
2161                         inactive_ratio = 1;
2162         }
2163
2164         if (actual_reclaim)
2165                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2166                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2167                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2168                         inactive_ratio, file);
2169
2170         return inactive * inactive_ratio < active;
2171 }
2172
2173 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2174                                  struct lruvec *lruvec, struct scan_control *sc)
2175 {
2176         if (is_active_lru(lru)) {
2177                 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2178                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2179                 return 0;
2180         }
2181
2182         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2183 }
2184
2185 enum scan_balance {
2186         SCAN_EQUAL,
2187         SCAN_FRACT,
2188         SCAN_ANON,
2189         SCAN_FILE,
2190 };
2191
2192 /*
2193  * Determine how aggressively the anon and file LRU lists should be
2194  * scanned.  The relative value of each set of LRU lists is determined
2195  * by looking at the fraction of the pages scanned we did rotate back
2196  * onto the active list instead of evict.
2197  *
2198  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2199  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2200  */
2201 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2202                            struct scan_control *sc, unsigned long *nr,
2203                            unsigned long *lru_pages)
2204 {
2205         int swappiness = mem_cgroup_swappiness(memcg);
2206         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2207         u64 fraction[2];
2208         u64 denominator = 0;    /* gcc */
2209         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2210         unsigned long anon_prio, file_prio;
2211         enum scan_balance scan_balance;
2212         unsigned long anon, file;
2213         unsigned long ap, fp;
2214         enum lru_list lru;
2215
2216         /* If we have no swap space, do not bother scanning anon pages. */
2217         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2218                 scan_balance = SCAN_FILE;
2219                 goto out;
2220         }
2221
2222         /*
2223          * Global reclaim will swap to prevent OOM even with no
2224          * swappiness, but memcg users want to use this knob to
2225          * disable swapping for individual groups completely when
2226          * using the memory controller's swap limit feature would be
2227          * too expensive.
2228          */
2229         if (!global_reclaim(sc) && !swappiness) {
2230                 scan_balance = SCAN_FILE;
2231                 goto out;
2232         }
2233
2234         /*
2235          * Do not apply any pressure balancing cleverness when the
2236          * system is close to OOM, scan both anon and file equally
2237          * (unless the swappiness setting disagrees with swapping).
2238          */
2239         if (!sc->priority && swappiness) {
2240                 scan_balance = SCAN_EQUAL;
2241                 goto out;
2242         }
2243
2244         /*
2245          * Prevent the reclaimer from falling into the cache trap: as
2246          * cache pages start out inactive, every cache fault will tip
2247          * the scan balance towards the file LRU.  And as the file LRU
2248          * shrinks, so does the window for rotation from references.
2249          * This means we have a runaway feedback loop where a tiny
2250          * thrashing file LRU becomes infinitely more attractive than
2251          * anon pages.  Try to detect this based on file LRU size.
2252          */
2253         if (global_reclaim(sc)) {
2254                 unsigned long pgdatfile;
2255                 unsigned long pgdatfree;
2256                 int z;
2257                 unsigned long total_high_wmark = 0;
2258
2259                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2260                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2261                            node_page_state(pgdat, NR_INACTIVE_FILE);
2262
2263                 for (z = 0; z < MAX_NR_ZONES; z++) {
2264                         struct zone *zone = &pgdat->node_zones[z];
2265                         if (!managed_zone(zone))
2266                                 continue;
2267
2268                         total_high_wmark += high_wmark_pages(zone);
2269                 }
2270
2271                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2272                         /*
2273                          * Force SCAN_ANON if there are enough inactive
2274                          * anonymous pages on the LRU in eligible zones.
2275                          * Otherwise, the small LRU gets thrashed.
2276                          */
2277                         if (!inactive_list_is_low(lruvec, false, sc, false) &&
2278                             lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2279                                         >> sc->priority) {
2280                                 scan_balance = SCAN_ANON;
2281                                 goto out;
2282                         }
2283                 }
2284         }
2285
2286         /*
2287          * If there is enough inactive page cache, i.e. if the size of the
2288          * inactive list is greater than that of the active list *and* the
2289          * inactive list actually has some pages to scan on this priority, we
2290          * do not reclaim anything from the anonymous working set right now.
2291          * Without the second condition we could end up never scanning an
2292          * lruvec even if it has plenty of old anonymous pages unless the
2293          * system is under heavy pressure.
2294          */
2295         if (!inactive_list_is_low(lruvec, true, sc, false) &&
2296             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2297                 scan_balance = SCAN_FILE;
2298                 goto out;
2299         }
2300
2301         scan_balance = SCAN_FRACT;
2302
2303         /*
2304          * With swappiness at 100, anonymous and file have the same priority.
2305          * This scanning priority is essentially the inverse of IO cost.
2306          */
2307         anon_prio = swappiness;
2308         file_prio = 200 - anon_prio;
2309
2310         /*
2311          * OK, so we have swap space and a fair amount of page cache
2312          * pages.  We use the recently rotated / recently scanned
2313          * ratios to determine how valuable each cache is.
2314          *
2315          * Because workloads change over time (and to avoid overflow)
2316          * we keep these statistics as a floating average, which ends
2317          * up weighing recent references more than old ones.
2318          *
2319          * anon in [0], file in [1]
2320          */
2321
2322         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2323                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2324         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2325                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2326
2327         spin_lock_irq(&pgdat->lru_lock);
2328         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2329                 reclaim_stat->recent_scanned[0] /= 2;
2330                 reclaim_stat->recent_rotated[0] /= 2;
2331         }
2332
2333         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2334                 reclaim_stat->recent_scanned[1] /= 2;
2335                 reclaim_stat->recent_rotated[1] /= 2;
2336         }
2337
2338         /*
2339          * The amount of pressure on anon vs file pages is inversely
2340          * proportional to the fraction of recently scanned pages on
2341          * each list that were recently referenced and in active use.
2342          */
2343         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2344         ap /= reclaim_stat->recent_rotated[0] + 1;
2345
2346         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2347         fp /= reclaim_stat->recent_rotated[1] + 1;
2348         spin_unlock_irq(&pgdat->lru_lock);
2349
2350         fraction[0] = ap;
2351         fraction[1] = fp;
2352         denominator = ap + fp + 1;
2353 out:
2354         *lru_pages = 0;
2355         for_each_evictable_lru(lru) {
2356                 int file = is_file_lru(lru);
2357                 unsigned long size;
2358                 unsigned long scan;
2359
2360                 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2361                 scan = size >> sc->priority;
2362                 /*
2363                  * If the cgroup's already been deleted, make sure to
2364                  * scrape out the remaining cache.
2365                  */
2366                 if (!scan && !mem_cgroup_online(memcg))
2367                         scan = min(size, SWAP_CLUSTER_MAX);
2368
2369                 switch (scan_balance) {
2370                 case SCAN_EQUAL:
2371                         /* Scan lists relative to size */
2372                         break;
2373                 case SCAN_FRACT:
2374                         /*
2375                          * Scan types proportional to swappiness and
2376                          * their relative recent reclaim efficiency.
2377                          * Make sure we don't miss the last page
2378                          * because of a round-off error.
2379                          */
2380                         scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2381                                                   denominator);
2382                         break;
2383                 case SCAN_FILE:
2384                 case SCAN_ANON:
2385                         /* Scan one type exclusively */
2386                         if ((scan_balance == SCAN_FILE) != file) {
2387                                 size = 0;
2388                                 scan = 0;
2389                         }
2390                         break;
2391                 default:
2392                         /* Look ma, no brain */
2393                         BUG();
2394                 }
2395
2396                 *lru_pages += size;
2397                 nr[lru] = scan;
2398         }
2399 }
2400
2401 /*
2402  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2403  */
2404 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2405                               struct scan_control *sc, unsigned long *lru_pages)
2406 {
2407         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2408         unsigned long nr[NR_LRU_LISTS];
2409         unsigned long targets[NR_LRU_LISTS];
2410         unsigned long nr_to_scan;
2411         enum lru_list lru;
2412         unsigned long nr_reclaimed = 0;
2413         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2414         struct blk_plug plug;
2415         bool scan_adjusted;
2416
2417         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2418
2419         /* Record the original scan target for proportional adjustments later */
2420         memcpy(targets, nr, sizeof(nr));
2421
2422         /*
2423          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2424          * event that can occur when there is little memory pressure e.g.
2425          * multiple streaming readers/writers. Hence, we do not abort scanning
2426          * when the requested number of pages are reclaimed when scanning at
2427          * DEF_PRIORITY on the assumption that the fact we are direct
2428          * reclaiming implies that kswapd is not keeping up and it is best to
2429          * do a batch of work at once. For memcg reclaim one check is made to
2430          * abort proportional reclaim if either the file or anon lru has already
2431          * dropped to zero at the first pass.
2432          */
2433         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2434                          sc->priority == DEF_PRIORITY);
2435
2436         blk_start_plug(&plug);
2437         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2438                                         nr[LRU_INACTIVE_FILE]) {
2439                 unsigned long nr_anon, nr_file, percentage;
2440                 unsigned long nr_scanned;
2441
2442                 for_each_evictable_lru(lru) {
2443                         if (nr[lru]) {
2444                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2445                                 nr[lru] -= nr_to_scan;
2446
2447                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2448                                                             lruvec, sc);
2449                         }
2450                 }
2451
2452                 cond_resched();
2453
2454                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2455                         continue;
2456
2457                 /*
2458                  * For kswapd and memcg, reclaim at least the number of pages
2459                  * requested. Ensure that the anon and file LRUs are scanned
2460                  * proportionally what was requested by get_scan_count(). We
2461                  * stop reclaiming one LRU and reduce the amount scanning
2462                  * proportional to the original scan target.
2463                  */
2464                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2465                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2466
2467                 /*
2468                  * It's just vindictive to attack the larger once the smaller
2469                  * has gone to zero.  And given the way we stop scanning the
2470                  * smaller below, this makes sure that we only make one nudge
2471                  * towards proportionality once we've got nr_to_reclaim.
2472                  */
2473                 if (!nr_file || !nr_anon)
2474                         break;
2475
2476                 if (nr_file > nr_anon) {
2477                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2478                                                 targets[LRU_ACTIVE_ANON] + 1;
2479                         lru = LRU_BASE;
2480                         percentage = nr_anon * 100 / scan_target;
2481                 } else {
2482                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2483                                                 targets[LRU_ACTIVE_FILE] + 1;
2484                         lru = LRU_FILE;
2485                         percentage = nr_file * 100 / scan_target;
2486                 }
2487
2488                 /* Stop scanning the smaller of the LRU */
2489                 nr[lru] = 0;
2490                 nr[lru + LRU_ACTIVE] = 0;
2491
2492                 /*
2493                  * Recalculate the other LRU scan count based on its original
2494                  * scan target and the percentage scanning already complete
2495                  */
2496                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2497                 nr_scanned = targets[lru] - nr[lru];
2498                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2499                 nr[lru] -= min(nr[lru], nr_scanned);
2500
2501                 lru += LRU_ACTIVE;
2502                 nr_scanned = targets[lru] - nr[lru];
2503                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2504                 nr[lru] -= min(nr[lru], nr_scanned);
2505
2506                 scan_adjusted = true;
2507         }
2508         blk_finish_plug(&plug);
2509         sc->nr_reclaimed += nr_reclaimed;
2510
2511         /*
2512          * Even if we did not try to evict anon pages at all, we want to
2513          * rebalance the anon lru active/inactive ratio.
2514          */
2515         if (inactive_list_is_low(lruvec, false, sc, true))
2516                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2517                                    sc, LRU_ACTIVE_ANON);
2518 }
2519
2520 /* Use reclaim/compaction for costly allocs or under memory pressure */
2521 static bool in_reclaim_compaction(struct scan_control *sc)
2522 {
2523         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2524                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2525                          sc->priority < DEF_PRIORITY - 2))
2526                 return true;
2527
2528         return false;
2529 }
2530
2531 /*
2532  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2533  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2534  * true if more pages should be reclaimed such that when the page allocator
2535  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2536  * It will give up earlier than that if there is difficulty reclaiming pages.
2537  */
2538 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2539                                         unsigned long nr_reclaimed,
2540                                         unsigned long nr_scanned,
2541                                         struct scan_control *sc)
2542 {
2543         unsigned long pages_for_compaction;
2544         unsigned long inactive_lru_pages;
2545         int z;
2546
2547         /* If not in reclaim/compaction mode, stop */
2548         if (!in_reclaim_compaction(sc))
2549                 return false;
2550
2551         /* Consider stopping depending on scan and reclaim activity */
2552         if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2553                 /*
2554                  * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2555                  * full LRU list has been scanned and we are still failing
2556                  * to reclaim pages. This full LRU scan is potentially
2557                  * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2558                  */
2559                 if (!nr_reclaimed && !nr_scanned)
2560                         return false;
2561         } else {
2562                 /*
2563                  * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2564                  * fail without consequence, stop if we failed to reclaim
2565                  * any pages from the last SWAP_CLUSTER_MAX number of
2566                  * pages that were scanned. This will return to the
2567                  * caller faster at the risk reclaim/compaction and
2568                  * the resulting allocation attempt fails
2569                  */
2570                 if (!nr_reclaimed)
2571                         return false;
2572         }
2573
2574         /*
2575          * If we have not reclaimed enough pages for compaction and the
2576          * inactive lists are large enough, continue reclaiming
2577          */
2578         pages_for_compaction = compact_gap(sc->order);
2579         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2580         if (get_nr_swap_pages() > 0)
2581                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2582         if (sc->nr_reclaimed < pages_for_compaction &&
2583                         inactive_lru_pages > pages_for_compaction)
2584                 return true;
2585
2586         /* If compaction would go ahead or the allocation would succeed, stop */
2587         for (z = 0; z <= sc->reclaim_idx; z++) {
2588                 struct zone *zone = &pgdat->node_zones[z];
2589                 if (!managed_zone(zone))
2590                         continue;
2591
2592                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2593                 case COMPACT_SUCCESS:
2594                 case COMPACT_CONTINUE:
2595                         return false;
2596                 default:
2597                         /* check next zone */
2598                         ;
2599                 }
2600         }
2601         return true;
2602 }
2603
2604 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2605 {
2606         return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2607                 (memcg && memcg_congested(pgdat, memcg));
2608 }
2609
2610 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2611 {
2612         struct reclaim_state *reclaim_state = current->reclaim_state;
2613         unsigned long nr_reclaimed, nr_scanned;
2614         bool reclaimable = false;
2615
2616         do {
2617                 struct mem_cgroup *root = sc->target_mem_cgroup;
2618                 struct mem_cgroup_reclaim_cookie reclaim = {
2619                         .pgdat = pgdat,
2620                         .priority = sc->priority,
2621                 };
2622                 unsigned long node_lru_pages = 0;
2623                 struct mem_cgroup *memcg;
2624
2625                 memset(&sc->nr, 0, sizeof(sc->nr));
2626
2627                 nr_reclaimed = sc->nr_reclaimed;
2628                 nr_scanned = sc->nr_scanned;
2629
2630                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2631                 do {
2632                         unsigned long lru_pages;
2633                         unsigned long reclaimed;
2634                         unsigned long scanned;
2635
2636                         switch (mem_cgroup_protected(root, memcg)) {
2637                         case MEMCG_PROT_MIN:
2638                                 /*
2639                                  * Hard protection.
2640                                  * If there is no reclaimable memory, OOM.
2641                                  */
2642                                 continue;
2643                         case MEMCG_PROT_LOW:
2644                                 /*
2645                                  * Soft protection.
2646                                  * Respect the protection only as long as
2647                                  * there is an unprotected supply
2648                                  * of reclaimable memory from other cgroups.
2649                                  */
2650                                 if (!sc->memcg_low_reclaim) {
2651                                         sc->memcg_low_skipped = 1;
2652                                         continue;
2653                                 }
2654                                 memcg_memory_event(memcg, MEMCG_LOW);
2655                                 break;
2656                         case MEMCG_PROT_NONE:
2657                                 break;
2658                         }
2659
2660                         reclaimed = sc->nr_reclaimed;
2661                         scanned = sc->nr_scanned;
2662                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2663                         node_lru_pages += lru_pages;
2664
2665                         if (sc->may_shrinkslab) {
2666                                 shrink_slab(sc->gfp_mask, pgdat->node_id,
2667                                     memcg, sc->priority);
2668                         }
2669
2670                         /* Record the group's reclaim efficiency */
2671                         vmpressure(sc->gfp_mask, memcg, false,
2672                                    sc->nr_scanned - scanned,
2673                                    sc->nr_reclaimed - reclaimed);
2674
2675                         /*
2676                          * Kswapd have to scan all memory cgroups to fulfill
2677                          * the overall scan target for the node.
2678                          *
2679                          * Limit reclaim, on the other hand, only cares about
2680                          * nr_to_reclaim pages to be reclaimed and it will
2681                          * retry with decreasing priority if one round over the
2682                          * whole hierarchy is not sufficient.
2683                          */
2684                         if (!current_is_kswapd() &&
2685                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2686                                 mem_cgroup_iter_break(root, memcg);
2687                                 break;
2688                         }
2689                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2690
2691                 if (reclaim_state) {
2692                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2693                         reclaim_state->reclaimed_slab = 0;
2694                 }
2695
2696                 /* Record the subtree's reclaim efficiency */
2697                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2698                            sc->nr_scanned - nr_scanned,
2699                            sc->nr_reclaimed - nr_reclaimed);
2700
2701                 if (sc->nr_reclaimed - nr_reclaimed)
2702                         reclaimable = true;
2703
2704                 if (current_is_kswapd()) {
2705                         /*
2706                          * If reclaim is isolating dirty pages under writeback,
2707                          * it implies that the long-lived page allocation rate
2708                          * is exceeding the page laundering rate. Either the
2709                          * global limits are not being effective at throttling
2710                          * processes due to the page distribution throughout
2711                          * zones or there is heavy usage of a slow backing
2712                          * device. The only option is to throttle from reclaim
2713                          * context which is not ideal as there is no guarantee
2714                          * the dirtying process is throttled in the same way
2715                          * balance_dirty_pages() manages.
2716                          *
2717                          * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2718                          * count the number of pages under pages flagged for
2719                          * immediate reclaim and stall if any are encountered
2720                          * in the nr_immediate check below.
2721                          */
2722                         if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2723                                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2724
2725                         /*
2726                          * Tag a node as congested if all the dirty pages
2727                          * scanned were backed by a congested BDI and
2728                          * wait_iff_congested will stall.
2729                          */
2730                         if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2731                                 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2732
2733                         /* Allow kswapd to start writing pages during reclaim.*/
2734                         if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2735                                 set_bit(PGDAT_DIRTY, &pgdat->flags);
2736
2737                         /*
2738                          * If kswapd scans pages marked marked for immediate
2739                          * reclaim and under writeback (nr_immediate), it
2740                          * implies that pages are cycling through the LRU
2741                          * faster than they are written so also forcibly stall.
2742                          */
2743                         if (sc->nr.immediate)
2744                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2745                 }
2746
2747                 /*
2748                  * Legacy memcg will stall in page writeback so avoid forcibly
2749                  * stalling in wait_iff_congested().
2750                  */
2751                 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2752                     sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2753                         set_memcg_congestion(pgdat, root, true);
2754
2755                 /*
2756                  * Stall direct reclaim for IO completions if underlying BDIs
2757                  * and node is congested. Allow kswapd to continue until it
2758                  * starts encountering unqueued dirty pages or cycling through
2759                  * the LRU too quickly.
2760                  */
2761                 if (!sc->hibernation_mode && !current_is_kswapd() &&
2762                    current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2763                         wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2764
2765         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2766                                          sc->nr_scanned - nr_scanned, sc));
2767
2768         /*
2769          * Kswapd gives up on balancing particular nodes after too
2770          * many failures to reclaim anything from them and goes to
2771          * sleep. On reclaim progress, reset the failure counter. A
2772          * successful direct reclaim run will revive a dormant kswapd.
2773          */
2774         if (reclaimable)
2775                 pgdat->kswapd_failures = 0;
2776
2777         return reclaimable;
2778 }
2779
2780 /*
2781  * Returns true if compaction should go ahead for a costly-order request, or
2782  * the allocation would already succeed without compaction. Return false if we
2783  * should reclaim first.
2784  */
2785 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2786 {
2787         unsigned long watermark;
2788         enum compact_result suitable;
2789
2790         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2791         if (suitable == COMPACT_SUCCESS)
2792                 /* Allocation should succeed already. Don't reclaim. */
2793                 return true;
2794         if (suitable == COMPACT_SKIPPED)
2795                 /* Compaction cannot yet proceed. Do reclaim. */
2796                 return false;
2797
2798         /*
2799          * Compaction is already possible, but it takes time to run and there
2800          * are potentially other callers using the pages just freed. So proceed
2801          * with reclaim to make a buffer of free pages available to give
2802          * compaction a reasonable chance of completing and allocating the page.
2803          * Note that we won't actually reclaim the whole buffer in one attempt
2804          * as the target watermark in should_continue_reclaim() is lower. But if
2805          * we are already above the high+gap watermark, don't reclaim at all.
2806          */
2807         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2808
2809         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2810 }
2811
2812 /*
2813  * This is the direct reclaim path, for page-allocating processes.  We only
2814  * try to reclaim pages from zones which will satisfy the caller's allocation
2815  * request.
2816  *
2817  * If a zone is deemed to be full of pinned pages then just give it a light
2818  * scan then give up on it.
2819  */
2820 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2821 {
2822         struct zoneref *z;
2823         struct zone *zone;
2824         unsigned long nr_soft_reclaimed;
2825         unsigned long nr_soft_scanned;
2826         gfp_t orig_mask;
2827         pg_data_t *last_pgdat = NULL;
2828
2829         /*
2830          * If the number of buffer_heads in the machine exceeds the maximum
2831          * allowed level, force direct reclaim to scan the highmem zone as
2832          * highmem pages could be pinning lowmem pages storing buffer_heads
2833          */
2834         orig_mask = sc->gfp_mask;
2835         if (buffer_heads_over_limit) {
2836                 sc->gfp_mask |= __GFP_HIGHMEM;
2837                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2838         }
2839
2840         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2841                                         sc->reclaim_idx, sc->nodemask) {
2842                 /*
2843                  * Take care memory controller reclaiming has small influence
2844                  * to global LRU.
2845                  */
2846                 if (global_reclaim(sc)) {
2847                         if (!cpuset_zone_allowed(zone,
2848                                                  GFP_KERNEL | __GFP_HARDWALL))
2849                                 continue;
2850
2851                         /*
2852                          * If we already have plenty of memory free for
2853                          * compaction in this zone, don't free any more.
2854                          * Even though compaction is invoked for any
2855                          * non-zero order, only frequent costly order
2856                          * reclamation is disruptive enough to become a
2857                          * noticeable problem, like transparent huge
2858                          * page allocations.
2859                          */
2860                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2861                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2862                             compaction_ready(zone, sc)) {
2863                                 sc->compaction_ready = true;
2864                                 continue;
2865                         }
2866
2867                         /*
2868                          * Shrink each node in the zonelist once. If the
2869                          * zonelist is ordered by zone (not the default) then a
2870                          * node may be shrunk multiple times but in that case
2871                          * the user prefers lower zones being preserved.
2872                          */
2873                         if (zone->zone_pgdat == last_pgdat)
2874                                 continue;
2875
2876                         /*
2877                          * This steals pages from memory cgroups over softlimit
2878                          * and returns the number of reclaimed pages and
2879                          * scanned pages. This works for global memory pressure
2880                          * and balancing, not for a memcg's limit.
2881                          */
2882                         nr_soft_scanned = 0;
2883                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2884                                                 sc->order, sc->gfp_mask,
2885                                                 &nr_soft_scanned);
2886                         sc->nr_reclaimed += nr_soft_reclaimed;
2887                         sc->nr_scanned += nr_soft_scanned;
2888                         /* need some check for avoid more shrink_zone() */
2889                 }
2890
2891                 /* See comment about same check for global reclaim above */
2892                 if (zone->zone_pgdat == last_pgdat)
2893                         continue;
2894                 last_pgdat = zone->zone_pgdat;
2895                 shrink_node(zone->zone_pgdat, sc);
2896         }
2897
2898         /*
2899          * Restore to original mask to avoid the impact on the caller if we
2900          * promoted it to __GFP_HIGHMEM.
2901          */
2902         sc->gfp_mask = orig_mask;
2903 }
2904
2905 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2906 {
2907         struct mem_cgroup *memcg;
2908
2909         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2910         do {
2911                 unsigned long refaults;
2912                 struct lruvec *lruvec;
2913
2914                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2915                 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2916                 lruvec->refaults = refaults;
2917         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2918 }
2919
2920 /*
2921  * This is the main entry point to direct page reclaim.
2922  *
2923  * If a full scan of the inactive list fails to free enough memory then we
2924  * are "out of memory" and something needs to be killed.
2925  *
2926  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2927  * high - the zone may be full of dirty or under-writeback pages, which this
2928  * caller can't do much about.  We kick the writeback threads and take explicit
2929  * naps in the hope that some of these pages can be written.  But if the
2930  * allocating task holds filesystem locks which prevent writeout this might not
2931  * work, and the allocation attempt will fail.
2932  *
2933  * returns:     0, if no pages reclaimed
2934  *              else, the number of pages reclaimed
2935  */
2936 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2937                                           struct scan_control *sc)
2938 {
2939         int initial_priority = sc->priority;
2940         pg_data_t *last_pgdat;
2941         struct zoneref *z;
2942         struct zone *zone;
2943 retry:
2944         delayacct_freepages_start();
2945
2946         if (global_reclaim(sc))
2947                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2948
2949         do {
2950                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2951                                 sc->priority);
2952                 sc->nr_scanned = 0;
2953                 shrink_zones(zonelist, sc);
2954
2955                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2956                         break;
2957
2958                 if (sc->compaction_ready)
2959                         break;
2960
2961                 /*
2962                  * If we're getting trouble reclaiming, start doing
2963                  * writepage even in laptop mode.
2964                  */
2965                 if (sc->priority < DEF_PRIORITY - 2)
2966                         sc->may_writepage = 1;
2967         } while (--sc->priority >= 0);
2968
2969         last_pgdat = NULL;
2970         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2971                                         sc->nodemask) {
2972                 if (zone->zone_pgdat == last_pgdat)
2973                         continue;
2974                 last_pgdat = zone->zone_pgdat;
2975                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2976                 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
2977         }
2978
2979         delayacct_freepages_end();
2980
2981         if (sc->nr_reclaimed)
2982                 return sc->nr_reclaimed;
2983
2984         /* Aborted reclaim to try compaction? don't OOM, then */
2985         if (sc->compaction_ready)
2986                 return 1;
2987
2988         /* Untapped cgroup reserves?  Don't OOM, retry. */
2989         if (sc->memcg_low_skipped) {
2990                 sc->priority = initial_priority;
2991                 sc->memcg_low_reclaim = 1;
2992                 sc->memcg_low_skipped = 0;
2993                 goto retry;
2994         }
2995
2996         return 0;
2997 }
2998
2999 static bool allow_direct_reclaim(pg_data_t *pgdat)
3000 {
3001         struct zone *zone;
3002         unsigned long pfmemalloc_reserve = 0;
3003         unsigned long free_pages = 0;
3004         int i;
3005         bool wmark_ok;
3006
3007         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3008                 return true;
3009
3010         for (i = 0; i <= ZONE_NORMAL; i++) {
3011                 zone = &pgdat->node_zones[i];
3012                 if (!managed_zone(zone))
3013                         continue;
3014
3015                 if (!zone_reclaimable_pages(zone))
3016                         continue;
3017
3018                 pfmemalloc_reserve += min_wmark_pages(zone);
3019                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3020         }
3021
3022         /* If there are no reserves (unexpected config) then do not throttle */
3023         if (!pfmemalloc_reserve)
3024                 return true;
3025
3026         wmark_ok = free_pages > pfmemalloc_reserve / 2;
3027
3028         /* kswapd must be awake if processes are being throttled */
3029         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3030                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3031                                                 (enum zone_type)ZONE_NORMAL);
3032                 wake_up_interruptible(&pgdat->kswapd_wait);
3033         }
3034
3035         return wmark_ok;
3036 }
3037
3038 /*
3039  * Throttle direct reclaimers if backing storage is backed by the network
3040  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3041  * depleted. kswapd will continue to make progress and wake the processes
3042  * when the low watermark is reached.
3043  *
3044  * Returns true if a fatal signal was delivered during throttling. If this
3045  * happens, the page allocator should not consider triggering the OOM killer.
3046  */
3047 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3048                                         nodemask_t *nodemask)
3049 {
3050         struct zoneref *z;
3051         struct zone *zone;
3052         pg_data_t *pgdat = NULL;
3053
3054         /*
3055          * Kernel threads should not be throttled as they may be indirectly
3056          * responsible for cleaning pages necessary for reclaim to make forward
3057          * progress. kjournald for example may enter direct reclaim while
3058          * committing a transaction where throttling it could forcing other
3059          * processes to block on log_wait_commit().
3060          */
3061         if (current->flags & PF_KTHREAD)
3062                 goto out;
3063
3064         /*
3065          * If a fatal signal is pending, this process should not throttle.
3066          * It should return quickly so it can exit and free its memory
3067          */
3068         if (fatal_signal_pending(current))
3069                 goto out;
3070
3071         /*
3072          * Check if the pfmemalloc reserves are ok by finding the first node
3073          * with a usable ZONE_NORMAL or lower zone. The expectation is that
3074          * GFP_KERNEL will be required for allocating network buffers when
3075          * swapping over the network so ZONE_HIGHMEM is unusable.
3076          *
3077          * Throttling is based on the first usable node and throttled processes
3078          * wait on a queue until kswapd makes progress and wakes them. There
3079          * is an affinity then between processes waking up and where reclaim
3080          * progress has been made assuming the process wakes on the same node.
3081          * More importantly, processes running on remote nodes will not compete
3082          * for remote pfmemalloc reserves and processes on different nodes
3083          * should make reasonable progress.
3084          */
3085         for_each_zone_zonelist_nodemask(zone, z, zonelist,
3086                                         gfp_zone(gfp_mask), nodemask) {
3087                 if (zone_idx(zone) > ZONE_NORMAL)
3088                         continue;
3089
3090                 /* Throttle based on the first usable node */
3091                 pgdat = zone->zone_pgdat;
3092                 if (allow_direct_reclaim(pgdat))
3093                         goto out;
3094                 break;
3095         }
3096
3097         /* If no zone was usable by the allocation flags then do not throttle */
3098         if (!pgdat)
3099                 goto out;
3100
3101         /* Account for the throttling */
3102         count_vm_event(PGSCAN_DIRECT_THROTTLE);
3103
3104         /*
3105          * If the caller cannot enter the filesystem, it's possible that it
3106          * is due to the caller holding an FS lock or performing a journal
3107          * transaction in the case of a filesystem like ext[3|4]. In this case,
3108          * it is not safe to block on pfmemalloc_wait as kswapd could be
3109          * blocked waiting on the same lock. Instead, throttle for up to a
3110          * second before continuing.
3111          */
3112         if (!(gfp_mask & __GFP_FS)) {
3113                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3114                         allow_direct_reclaim(pgdat), HZ);
3115
3116                 goto check_pending;
3117         }
3118
3119         /* Throttle until kswapd wakes the process */
3120         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3121                 allow_direct_reclaim(pgdat));
3122
3123 check_pending:
3124         if (fatal_signal_pending(current))
3125                 return true;
3126
3127 out:
3128         return false;
3129 }
3130
3131 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3132                                 gfp_t gfp_mask, nodemask_t *nodemask)
3133 {
3134         unsigned long nr_reclaimed;
3135         struct scan_control sc = {
3136                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3137                 .gfp_mask = current_gfp_context(gfp_mask),
3138                 .reclaim_idx = gfp_zone(gfp_mask),
3139                 .order = order,
3140                 .nodemask = nodemask,
3141                 .priority = DEF_PRIORITY,
3142                 .may_writepage = !laptop_mode,
3143                 .may_unmap = 1,
3144                 .may_swap = 1,
3145                 .may_shrinkslab = 1,
3146         };
3147
3148         /*
3149          * scan_control uses s8 fields for order, priority, and reclaim_idx.
3150          * Confirm they are large enough for max values.
3151          */
3152         BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3153         BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3154         BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3155
3156         /*
3157          * Do not enter reclaim if fatal signal was delivered while throttled.
3158          * 1 is returned so that the page allocator does not OOM kill at this
3159          * point.
3160          */
3161         if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3162                 return 1;
3163
3164         trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3165
3166         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3167
3168         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3169