mm: remove cold parameter from free_hot_cold_page*
[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/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
55
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
58
59 #include "internal.h"
60
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
63
64 struct scan_control {
65         /* How many pages shrink_list() should reclaim */
66         unsigned long nr_to_reclaim;
67
68         /* This context's GFP mask */
69         gfp_t gfp_mask;
70
71         /* Allocation order */
72         int order;
73
74         /*
75          * Nodemask of nodes allowed by the caller. If NULL, all nodes
76          * are scanned.
77          */
78         nodemask_t      *nodemask;
79
80         /*
81          * The memory cgroup that hit its limit and as a result is the
82          * primary target of this reclaim invocation.
83          */
84         struct mem_cgroup *target_mem_cgroup;
85
86         /* Scan (total_size >> priority) pages at once */
87         int priority;
88
89         /* The highest zone to isolate pages for reclaim from */
90         enum zone_type reclaim_idx;
91
92         /* Writepage batching in laptop mode; RECLAIM_WRITE */
93         unsigned int may_writepage:1;
94
95         /* Can mapped pages be reclaimed? */
96         unsigned int may_unmap:1;
97
98         /* Can pages be swapped as part of reclaim? */
99         unsigned int may_swap:1;
100
101         /*
102          * Cgroups are not reclaimed below their configured memory.low,
103          * unless we threaten to OOM. If any cgroups are skipped due to
104          * memory.low and nothing was reclaimed, go back for memory.low.
105          */
106         unsigned int memcg_low_reclaim:1;
107         unsigned int memcg_low_skipped:1;
108
109         unsigned int hibernation_mode:1;
110
111         /* One of the zones is ready for compaction */
112         unsigned int compaction_ready:1;
113
114         /* Incremented by the number of inactive pages that were scanned */
115         unsigned long nr_scanned;
116
117         /* Number of pages freed so far during a call to shrink_zones() */
118         unsigned long nr_reclaimed;
119 };
120
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field)                    \
123         do {                                                            \
124                 if ((_page)->lru.prev != _base) {                       \
125                         struct page *prev;                              \
126                                                                         \
127                         prev = lru_to_page(&(_page->lru));              \
128                         prefetch(&prev->_field);                        \
129                 }                                                       \
130         } while (0)
131 #else
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
133 #endif
134
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
137         do {                                                            \
138                 if ((_page)->lru.prev != _base) {                       \
139                         struct page *prev;                              \
140                                                                         \
141                         prev = lru_to_page(&(_page->lru));              \
142                         prefetchw(&prev->_field);                       \
143                 }                                                       \
144         } while (0)
145 #else
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 #endif
148
149 /*
150  * From 0 .. 100.  Higher means more swappy.
151  */
152 int vm_swappiness = 60;
153 /*
154  * The total number of pages which are beyond the high watermark within all
155  * zones.
156  */
157 unsigned long vm_total_pages;
158
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
161
162 #ifdef CONFIG_MEMCG
163 static bool global_reclaim(struct scan_control *sc)
164 {
165         return !sc->target_mem_cgroup;
166 }
167
168 /**
169  * sane_reclaim - is the usual dirty throttling mechanism operational?
170  * @sc: scan_control in question
171  *
172  * The normal page dirty throttling mechanism in balance_dirty_pages() is
173  * completely broken with the legacy memcg and direct stalling in
174  * shrink_page_list() is used for throttling instead, which lacks all the
175  * niceties such as fairness, adaptive pausing, bandwidth proportional
176  * allocation and configurability.
177  *
178  * This function tests whether the vmscan currently in progress can assume
179  * that the normal dirty throttling mechanism is operational.
180  */
181 static bool sane_reclaim(struct scan_control *sc)
182 {
183         struct mem_cgroup *memcg = sc->target_mem_cgroup;
184
185         if (!memcg)
186                 return true;
187 #ifdef CONFIG_CGROUP_WRITEBACK
188         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
189                 return true;
190 #endif
191         return false;
192 }
193 #else
194 static bool global_reclaim(struct scan_control *sc)
195 {
196         return true;
197 }
198
199 static bool sane_reclaim(struct scan_control *sc)
200 {
201         return true;
202 }
203 #endif
204
205 /*
206  * This misses isolated pages which are not accounted for to save counters.
207  * As the data only determines if reclaim or compaction continues, it is
208  * not expected that isolated pages will be a dominating factor.
209  */
210 unsigned long zone_reclaimable_pages(struct zone *zone)
211 {
212         unsigned long nr;
213
214         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
215                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
216         if (get_nr_swap_pages() > 0)
217                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
218                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
219
220         return nr;
221 }
222
223 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
224 {
225         unsigned long nr;
226
227         nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
228              node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
229              node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
230
231         if (get_nr_swap_pages() > 0)
232                 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
233                       node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
234                       node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
235
236         return nr;
237 }
238
239 /**
240  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
241  * @lruvec: lru vector
242  * @lru: lru to use
243  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244  */
245 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
246 {
247         unsigned long lru_size;
248         int zid;
249
250         if (!mem_cgroup_disabled())
251                 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
252         else
253                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
254
255         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
256                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
257                 unsigned long size;
258
259                 if (!managed_zone(zone))
260                         continue;
261
262                 if (!mem_cgroup_disabled())
263                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
264                 else
265                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
266                                        NR_ZONE_LRU_BASE + lru);
267                 lru_size -= min(size, lru_size);
268         }
269
270         return lru_size;
271
272 }
273
274 /*
275  * Add a shrinker callback to be called from the vm.
276  */
277 int register_shrinker(struct shrinker *shrinker)
278 {
279         size_t size = sizeof(*shrinker->nr_deferred);
280
281         if (shrinker->flags & SHRINKER_NUMA_AWARE)
282                 size *= nr_node_ids;
283
284         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
285         if (!shrinker->nr_deferred)
286                 return -ENOMEM;
287
288         down_write(&shrinker_rwsem);
289         list_add_tail(&shrinker->list, &shrinker_list);
290         up_write(&shrinker_rwsem);
291         return 0;
292 }
293 EXPORT_SYMBOL(register_shrinker);
294
295 /*
296  * Remove one
297  */
298 void unregister_shrinker(struct shrinker *shrinker)
299 {
300         down_write(&shrinker_rwsem);
301         list_del(&shrinker->list);
302         up_write(&shrinker_rwsem);
303         kfree(shrinker->nr_deferred);
304 }
305 EXPORT_SYMBOL(unregister_shrinker);
306
307 #define SHRINK_BATCH 128
308
309 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
310                                     struct shrinker *shrinker,
311                                     unsigned long nr_scanned,
312                                     unsigned long nr_eligible)
313 {
314         unsigned long freed = 0;
315         unsigned long long delta;
316         long total_scan;
317         long freeable;
318         long nr;
319         long new_nr;
320         int nid = shrinkctl->nid;
321         long batch_size = shrinker->batch ? shrinker->batch
322                                           : SHRINK_BATCH;
323         long scanned = 0, next_deferred;
324
325         freeable = shrinker->count_objects(shrinker, shrinkctl);
326         if (freeable == 0)
327                 return 0;
328
329         /*
330          * copy the current shrinker scan count into a local variable
331          * and zero it so that other concurrent shrinker invocations
332          * don't also do this scanning work.
333          */
334         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
335
336         total_scan = nr;
337         delta = (4 * nr_scanned) / shrinker->seeks;
338         delta *= freeable;
339         do_div(delta, nr_eligible + 1);
340         total_scan += delta;
341         if (total_scan < 0) {
342                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
343                        shrinker->scan_objects, total_scan);
344                 total_scan = freeable;
345                 next_deferred = nr;
346         } else
347                 next_deferred = total_scan;
348
349         /*
350          * We need to avoid excessive windup on filesystem shrinkers
351          * due to large numbers of GFP_NOFS allocations causing the
352          * shrinkers to return -1 all the time. This results in a large
353          * nr being built up so when a shrink that can do some work
354          * comes along it empties the entire cache due to nr >>>
355          * freeable. This is bad for sustaining a working set in
356          * memory.
357          *
358          * Hence only allow the shrinker to scan the entire cache when
359          * a large delta change is calculated directly.
360          */
361         if (delta < freeable / 4)
362                 total_scan = min(total_scan, freeable / 2);
363
364         /*
365          * Avoid risking looping forever due to too large nr value:
366          * never try to free more than twice the estimate number of
367          * freeable entries.
368          */
369         if (total_scan > freeable * 2)
370                 total_scan = freeable * 2;
371
372         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
373                                    nr_scanned, nr_eligible,
374                                    freeable, delta, total_scan);
375
376         /*
377          * Normally, we should not scan less than batch_size objects in one
378          * pass to avoid too frequent shrinker calls, but if the slab has less
379          * than batch_size objects in total and we are really tight on memory,
380          * we will try to reclaim all available objects, otherwise we can end
381          * up failing allocations although there are plenty of reclaimable
382          * objects spread over several slabs with usage less than the
383          * batch_size.
384          *
385          * We detect the "tight on memory" situations by looking at the total
386          * number of objects we want to scan (total_scan). If it is greater
387          * than the total number of objects on slab (freeable), we must be
388          * scanning at high prio and therefore should try to reclaim as much as
389          * possible.
390          */
391         while (total_scan >= batch_size ||
392                total_scan >= freeable) {
393                 unsigned long ret;
394                 unsigned long nr_to_scan = min(batch_size, total_scan);
395
396                 shrinkctl->nr_to_scan = nr_to_scan;
397                 shrinkctl->nr_scanned = nr_to_scan;
398                 ret = shrinker->scan_objects(shrinker, shrinkctl);
399                 if (ret == SHRINK_STOP)
400                         break;
401                 freed += ret;
402
403                 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
404                 total_scan -= shrinkctl->nr_scanned;
405                 scanned += shrinkctl->nr_scanned;
406
407                 cond_resched();
408         }
409
410         if (next_deferred >= scanned)
411                 next_deferred -= scanned;
412         else
413                 next_deferred = 0;
414         /*
415          * move the unused scan count back into the shrinker in a
416          * manner that handles concurrent updates. If we exhausted the
417          * scan, there is no need to do an update.
418          */
419         if (next_deferred > 0)
420                 new_nr = atomic_long_add_return(next_deferred,
421                                                 &shrinker->nr_deferred[nid]);
422         else
423                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
424
425         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
426         return freed;
427 }
428
429 /**
430  * shrink_slab - shrink slab caches
431  * @gfp_mask: allocation context
432  * @nid: node whose slab caches to target
433  * @memcg: memory cgroup whose slab caches to target
434  * @nr_scanned: pressure numerator
435  * @nr_eligible: pressure denominator
436  *
437  * Call the shrink functions to age shrinkable caches.
438  *
439  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
440  * unaware shrinkers will receive a node id of 0 instead.
441  *
442  * @memcg specifies the memory cgroup to target. If it is not NULL,
443  * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
444  * objects from the memory cgroup specified. Otherwise, only unaware
445  * shrinkers are called.
446  *
447  * @nr_scanned and @nr_eligible form a ratio that indicate how much of
448  * the available objects should be scanned.  Page reclaim for example
449  * passes the number of pages scanned and the number of pages on the
450  * LRU lists that it considered on @nid, plus a bias in @nr_scanned
451  * when it encountered mapped pages.  The ratio is further biased by
452  * the ->seeks setting of the shrink function, which indicates the
453  * cost to recreate an object relative to that of an LRU page.
454  *
455  * Returns the number of reclaimed slab objects.
456  */
457 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
458                                  struct mem_cgroup *memcg,
459                                  unsigned long nr_scanned,
460                                  unsigned long nr_eligible)
461 {
462         struct shrinker *shrinker;
463         unsigned long freed = 0;
464
465         if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
466                 return 0;
467
468         if (nr_scanned == 0)
469                 nr_scanned = SWAP_CLUSTER_MAX;
470
471         if (!down_read_trylock(&shrinker_rwsem)) {
472                 /*
473                  * If we would return 0, our callers would understand that we
474                  * have nothing else to shrink and give up trying. By returning
475                  * 1 we keep it going and assume we'll be able to shrink next
476                  * time.
477                  */
478                 freed = 1;
479                 goto out;
480         }
481
482         list_for_each_entry(shrinker, &shrinker_list, list) {
483                 struct shrink_control sc = {
484                         .gfp_mask = gfp_mask,
485                         .nid = nid,
486                         .memcg = memcg,
487                 };
488
489                 /*
490                  * If kernel memory accounting is disabled, we ignore
491                  * SHRINKER_MEMCG_AWARE flag and call all shrinkers
492                  * passing NULL for memcg.
493                  */
494                 if (memcg_kmem_enabled() &&
495                     !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
496                         continue;
497
498                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
499                         sc.nid = 0;
500
501                 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
502         }
503
504         up_read(&shrinker_rwsem);
505 out:
506         cond_resched();
507         return freed;
508 }
509
510 void drop_slab_node(int nid)
511 {
512         unsigned long freed;
513
514         do {
515                 struct mem_cgroup *memcg = NULL;
516
517                 freed = 0;
518                 do {
519                         freed += shrink_slab(GFP_KERNEL, nid, memcg,
520                                              1000, 1000);
521                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
522         } while (freed > 10);
523 }
524
525 void drop_slab(void)
526 {
527         int nid;
528
529         for_each_online_node(nid)
530                 drop_slab_node(nid);
531 }
532
533 static inline int is_page_cache_freeable(struct page *page)
534 {
535         /*
536          * A freeable page cache page is referenced only by the caller
537          * that isolated the page, the page cache radix tree and
538          * optional buffer heads at page->private.
539          */
540         int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
541                 HPAGE_PMD_NR : 1;
542         return page_count(page) - page_has_private(page) == 1 + radix_pins;
543 }
544
545 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
546 {
547         if (current->flags & PF_SWAPWRITE)
548                 return 1;
549         if (!inode_write_congested(inode))
550                 return 1;
551         if (inode_to_bdi(inode) == current->backing_dev_info)
552                 return 1;
553         return 0;
554 }
555
556 /*
557  * We detected a synchronous write error writing a page out.  Probably
558  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
559  * fsync(), msync() or close().
560  *
561  * The tricky part is that after writepage we cannot touch the mapping: nothing
562  * prevents it from being freed up.  But we have a ref on the page and once
563  * that page is locked, the mapping is pinned.
564  *
565  * We're allowed to run sleeping lock_page() here because we know the caller has
566  * __GFP_FS.
567  */
568 static void handle_write_error(struct address_space *mapping,
569                                 struct page *page, int error)
570 {
571         lock_page(page);
572         if (page_mapping(page) == mapping)
573                 mapping_set_error(mapping, error);
574         unlock_page(page);
575 }
576
577 /* possible outcome of pageout() */
578 typedef enum {
579         /* failed to write page out, page is locked */
580         PAGE_KEEP,
581         /* move page to the active list, page is locked */
582         PAGE_ACTIVATE,
583         /* page has been sent to the disk successfully, page is unlocked */
584         PAGE_SUCCESS,
585         /* page is clean and locked */
586         PAGE_CLEAN,
587 } pageout_t;
588
589 /*
590  * pageout is called by shrink_page_list() for each dirty page.
591  * Calls ->writepage().
592  */
593 static pageout_t pageout(struct page *page, struct address_space *mapping,
594                          struct scan_control *sc)
595 {
596         /*
597          * If the page is dirty, only perform writeback if that write
598          * will be non-blocking.  To prevent this allocation from being
599          * stalled by pagecache activity.  But note that there may be
600          * stalls if we need to run get_block().  We could test
601          * PagePrivate for that.
602          *
603          * If this process is currently in __generic_file_write_iter() against
604          * this page's queue, we can perform writeback even if that
605          * will block.
606          *
607          * If the page is swapcache, write it back even if that would
608          * block, for some throttling. This happens by accident, because
609          * swap_backing_dev_info is bust: it doesn't reflect the
610          * congestion state of the swapdevs.  Easy to fix, if needed.
611          */
612         if (!is_page_cache_freeable(page))
613                 return PAGE_KEEP;
614         if (!mapping) {
615                 /*
616                  * Some data journaling orphaned pages can have
617                  * page->mapping == NULL while being dirty with clean buffers.
618                  */
619                 if (page_has_private(page)) {
620                         if (try_to_free_buffers(page)) {
621                                 ClearPageDirty(page);
622                                 pr_info("%s: orphaned page\n", __func__);
623                                 return PAGE_CLEAN;
624                         }
625                 }
626                 return PAGE_KEEP;
627         }
628         if (mapping->a_ops->writepage == NULL)
629                 return PAGE_ACTIVATE;
630         if (!may_write_to_inode(mapping->host, sc))
631                 return PAGE_KEEP;
632
633         if (clear_page_dirty_for_io(page)) {
634                 int res;
635                 struct writeback_control wbc = {
636                         .sync_mode = WB_SYNC_NONE,
637                         .nr_to_write = SWAP_CLUSTER_MAX,
638                         .range_start = 0,
639                         .range_end = LLONG_MAX,
640                         .for_reclaim = 1,
641                 };
642
643                 SetPageReclaim(page);
644                 res = mapping->a_ops->writepage(page, &wbc);
645                 if (res < 0)
646                         handle_write_error(mapping, page, res);
647                 if (res == AOP_WRITEPAGE_ACTIVATE) {
648                         ClearPageReclaim(page);
649                         return PAGE_ACTIVATE;
650                 }
651
652                 if (!PageWriteback(page)) {
653                         /* synchronous write or broken a_ops? */
654                         ClearPageReclaim(page);
655                 }
656                 trace_mm_vmscan_writepage(page);
657                 inc_node_page_state(page, NR_VMSCAN_WRITE);
658                 return PAGE_SUCCESS;
659         }
660
661         return PAGE_CLEAN;
662 }
663
664 /*
665  * Same as remove_mapping, but if the page is removed from the mapping, it
666  * gets returned with a refcount of 0.
667  */
668 static int __remove_mapping(struct address_space *mapping, struct page *page,
669                             bool reclaimed)
670 {
671         unsigned long flags;
672         int refcount;
673
674         BUG_ON(!PageLocked(page));
675         BUG_ON(mapping != page_mapping(page));
676
677         spin_lock_irqsave(&mapping->tree_lock, flags);
678         /*
679          * The non racy check for a busy page.
680          *
681          * Must be careful with the order of the tests. When someone has
682          * a ref to the page, it may be possible that they dirty it then
683          * drop the reference. So if PageDirty is tested before page_count
684          * here, then the following race may occur:
685          *
686          * get_user_pages(&page);
687          * [user mapping goes away]
688          * write_to(page);
689          *                              !PageDirty(page)    [good]
690          * SetPageDirty(page);
691          * put_page(page);
692          *                              !page_count(page)   [good, discard it]
693          *
694          * [oops, our write_to data is lost]
695          *
696          * Reversing the order of the tests ensures such a situation cannot
697          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
698          * load is not satisfied before that of page->_refcount.
699          *
700          * Note that if SetPageDirty is always performed via set_page_dirty,
701          * and thus under tree_lock, then this ordering is not required.
702          */
703         if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
704                 refcount = 1 + HPAGE_PMD_NR;
705         else
706                 refcount = 2;
707         if (!page_ref_freeze(page, refcount))
708                 goto cannot_free;
709         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
710         if (unlikely(PageDirty(page))) {
711                 page_ref_unfreeze(page, refcount);
712                 goto cannot_free;
713         }
714
715         if (PageSwapCache(page)) {
716                 swp_entry_t swap = { .val = page_private(page) };
717                 mem_cgroup_swapout(page, swap);
718                 __delete_from_swap_cache(page);
719                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
720                 put_swap_page(page, swap);
721         } else {
722                 void (*freepage)(struct page *);
723                 void *shadow = NULL;
724
725                 freepage = mapping->a_ops->freepage;
726                 /*
727                  * Remember a shadow entry for reclaimed file cache in
728                  * order to detect refaults, thus thrashing, later on.
729                  *
730                  * But don't store shadows in an address space that is
731                  * already exiting.  This is not just an optizimation,
732                  * inode reclaim needs to empty out the radix tree or
733                  * the nodes are lost.  Don't plant shadows behind its
734                  * back.
735                  *
736                  * We also don't store shadows for DAX mappings because the
737                  * only page cache pages found in these are zero pages
738                  * covering holes, and because we don't want to mix DAX
739                  * exceptional entries and shadow exceptional entries in the
740                  * same page_tree.
741                  */
742                 if (reclaimed && page_is_file_cache(page) &&
743                     !mapping_exiting(mapping) && !dax_mapping(mapping))
744                         shadow = workingset_eviction(mapping, page);
745                 __delete_from_page_cache(page, shadow);
746                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
747
748                 if (freepage != NULL)
749                         freepage(page);
750         }
751
752         return 1;
753
754 cannot_free:
755         spin_unlock_irqrestore(&mapping->tree_lock, flags);
756         return 0;
757 }
758
759 /*
760  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
761  * someone else has a ref on the page, abort and return 0.  If it was
762  * successfully detached, return 1.  Assumes the caller has a single ref on
763  * this page.
764  */
765 int remove_mapping(struct address_space *mapping, struct page *page)
766 {
767         if (__remove_mapping(mapping, page, false)) {
768                 /*
769                  * Unfreezing the refcount with 1 rather than 2 effectively
770                  * drops the pagecache ref for us without requiring another
771                  * atomic operation.
772                  */
773                 page_ref_unfreeze(page, 1);
774                 return 1;
775         }
776         return 0;
777 }
778
779 /**
780  * putback_lru_page - put previously isolated page onto appropriate LRU list
781  * @page: page to be put back to appropriate lru list
782  *
783  * Add previously isolated @page to appropriate LRU list.
784  * Page may still be unevictable for other reasons.
785  *
786  * lru_lock must not be held, interrupts must be enabled.
787  */
788 void putback_lru_page(struct page *page)
789 {
790         bool is_unevictable;
791         int was_unevictable = PageUnevictable(page);
792
793         VM_BUG_ON_PAGE(PageLRU(page), page);
794
795 redo:
796         ClearPageUnevictable(page);
797
798         if (page_evictable(page)) {
799                 /*
800                  * For evictable pages, we can use the cache.
801                  * In event of a race, worst case is we end up with an
802                  * unevictable page on [in]active list.
803                  * We know how to handle that.
804                  */
805                 is_unevictable = false;
806                 lru_cache_add(page);
807         } else {
808                 /*
809                  * Put unevictable pages directly on zone's unevictable
810                  * list.
811                  */
812                 is_unevictable = true;
813                 add_page_to_unevictable_list(page);
814                 /*
815                  * When racing with an mlock or AS_UNEVICTABLE clearing
816                  * (page is unlocked) make sure that if the other thread
817                  * does not observe our setting of PG_lru and fails
818                  * isolation/check_move_unevictable_pages,
819                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
820                  * the page back to the evictable list.
821                  *
822                  * The other side is TestClearPageMlocked() or shmem_lock().
823                  */
824                 smp_mb();
825         }
826
827         /*
828          * page's status can change while we move it among lru. If an evictable
829          * page is on unevictable list, it never be freed. To avoid that,
830          * check after we added it to the list, again.
831          */
832         if (is_unevictable && page_evictable(page)) {
833                 if (!isolate_lru_page(page)) {
834                         put_page(page);
835                         goto redo;
836                 }
837                 /* This means someone else dropped this page from LRU
838                  * So, it will be freed or putback to LRU again. There is
839                  * nothing to do here.
840                  */
841         }
842
843         if (was_unevictable && !is_unevictable)
844                 count_vm_event(UNEVICTABLE_PGRESCUED);
845         else if (!was_unevictable && is_unevictable)
846                 count_vm_event(UNEVICTABLE_PGCULLED);
847
848         put_page(page);         /* drop ref from isolate */
849 }
850
851 enum page_references {
852         PAGEREF_RECLAIM,
853         PAGEREF_RECLAIM_CLEAN,
854         PAGEREF_KEEP,
855         PAGEREF_ACTIVATE,
856 };
857
858 static enum page_references page_check_references(struct page *page,
859                                                   struct scan_control *sc)
860 {
861         int referenced_ptes, referenced_page;
862         unsigned long vm_flags;
863
864         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
865                                           &vm_flags);
866         referenced_page = TestClearPageReferenced(page);
867
868         /*
869          * Mlock lost the isolation race with us.  Let try_to_unmap()
870          * move the page to the unevictable list.
871          */
872         if (vm_flags & VM_LOCKED)
873                 return PAGEREF_RECLAIM;
874
875         if (referenced_ptes) {
876                 if (PageSwapBacked(page))
877                         return PAGEREF_ACTIVATE;
878                 /*
879                  * All mapped pages start out with page table
880                  * references from the instantiating fault, so we need
881                  * to look twice if a mapped file page is used more
882                  * than once.
883                  *
884                  * Mark it and spare it for another trip around the
885                  * inactive list.  Another page table reference will
886                  * lead to its activation.
887                  *
888                  * Note: the mark is set for activated pages as well
889                  * so that recently deactivated but used pages are
890                  * quickly recovered.
891                  */
892                 SetPageReferenced(page);
893
894                 if (referenced_page || referenced_ptes > 1)
895                         return PAGEREF_ACTIVATE;
896
897                 /*
898                  * Activate file-backed executable pages after first usage.
899                  */
900                 if (vm_flags & VM_EXEC)
901                         return PAGEREF_ACTIVATE;
902
903                 return PAGEREF_KEEP;
904         }
905
906         /* Reclaim if clean, defer dirty pages to writeback */
907         if (referenced_page && !PageSwapBacked(page))
908                 return PAGEREF_RECLAIM_CLEAN;
909
910         return PAGEREF_RECLAIM;
911 }
912
913 /* Check if a page is dirty or under writeback */
914 static void page_check_dirty_writeback(struct page *page,
915                                        bool *dirty, bool *writeback)
916 {
917         struct address_space *mapping;
918
919         /*
920          * Anonymous pages are not handled by flushers and must be written
921          * from reclaim context. Do not stall reclaim based on them
922          */
923         if (!page_is_file_cache(page) ||
924             (PageAnon(page) && !PageSwapBacked(page))) {
925                 *dirty = false;
926                 *writeback = false;
927                 return;
928         }
929
930         /* By default assume that the page flags are accurate */
931         *dirty = PageDirty(page);
932         *writeback = PageWriteback(page);
933
934         /* Verify dirty/writeback state if the filesystem supports it */
935         if (!page_has_private(page))
936                 return;
937
938         mapping = page_mapping(page);
939         if (mapping && mapping->a_ops->is_dirty_writeback)
940                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
941 }
942
943 struct reclaim_stat {
944         unsigned nr_dirty;
945         unsigned nr_unqueued_dirty;
946         unsigned nr_congested;
947         unsigned nr_writeback;
948         unsigned nr_immediate;
949         unsigned nr_activate;
950         unsigned nr_ref_keep;
951         unsigned nr_unmap_fail;
952 };
953
954 /*
955  * shrink_page_list() returns the number of reclaimed pages
956  */
957 static unsigned long shrink_page_list(struct list_head *page_list,
958                                       struct pglist_data *pgdat,
959                                       struct scan_control *sc,
960                                       enum ttu_flags ttu_flags,
961                                       struct reclaim_stat *stat,
962                                       bool force_reclaim)
963 {
964         LIST_HEAD(ret_pages);
965         LIST_HEAD(free_pages);
966         int pgactivate = 0;
967         unsigned nr_unqueued_dirty = 0;
968         unsigned nr_dirty = 0;
969         unsigned nr_congested = 0;
970         unsigned nr_reclaimed = 0;
971         unsigned nr_writeback = 0;
972         unsigned nr_immediate = 0;
973         unsigned nr_ref_keep = 0;
974         unsigned nr_unmap_fail = 0;
975
976         cond_resched();
977
978         while (!list_empty(page_list)) {
979                 struct address_space *mapping;
980                 struct page *page;
981                 int may_enter_fs;
982                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
983                 bool dirty, writeback;
984
985                 cond_resched();
986
987                 page = lru_to_page(page_list);
988                 list_del(&page->lru);
989
990                 if (!trylock_page(page))
991                         goto keep;
992
993                 VM_BUG_ON_PAGE(PageActive(page), page);
994
995                 sc->nr_scanned++;
996
997                 if (unlikely(!page_evictable(page)))
998                         goto activate_locked;
999
1000                 if (!sc->may_unmap && page_mapped(page))
1001                         goto keep_locked;
1002
1003                 /* Double the slab pressure for mapped and swapcache pages */
1004                 if ((page_mapped(page) || PageSwapCache(page)) &&
1005                     !(PageAnon(page) && !PageSwapBacked(page)))
1006                         sc->nr_scanned++;
1007
1008                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1009                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1010
1011                 /*
1012                  * The number of dirty pages determines if a zone is marked
1013                  * reclaim_congested which affects wait_iff_congested. kswapd
1014                  * will stall and start writing pages if the tail of the LRU
1015                  * is all dirty unqueued pages.
1016                  */
1017                 page_check_dirty_writeback(page, &dirty, &writeback);
1018                 if (dirty || writeback)
1019                         nr_dirty++;
1020
1021                 if (dirty && !writeback)
1022                         nr_unqueued_dirty++;
1023
1024                 /*
1025                  * Treat this page as congested if the underlying BDI is or if
1026                  * pages are cycling through the LRU so quickly that the
1027                  * pages marked for immediate reclaim are making it to the
1028                  * end of the LRU a second time.
1029                  */
1030                 mapping = page_mapping(page);
1031                 if (((dirty || writeback) && mapping &&
1032                      inode_write_congested(mapping->host)) ||
1033                     (writeback && PageReclaim(page)))
1034                         nr_congested++;
1035
1036                 /*
1037                  * If a page at the tail of the LRU is under writeback, there
1038                  * are three cases to consider.
1039                  *
1040                  * 1) If reclaim is encountering an excessive number of pages
1041                  *    under writeback and this page is both under writeback and
1042                  *    PageReclaim then it indicates that pages are being queued
1043                  *    for IO but are being recycled through the LRU before the
1044                  *    IO can complete. Waiting on the page itself risks an
1045                  *    indefinite stall if it is impossible to writeback the
1046                  *    page due to IO error or disconnected storage so instead
1047                  *    note that the LRU is being scanned too quickly and the
1048                  *    caller can stall after page list has been processed.
1049                  *
1050                  * 2) Global or new memcg reclaim encounters a page that is
1051                  *    not marked for immediate reclaim, or the caller does not
1052                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1053                  *    not to fs). In this case mark the page for immediate
1054                  *    reclaim and continue scanning.
1055                  *
1056                  *    Require may_enter_fs because we would wait on fs, which
1057                  *    may not have submitted IO yet. And the loop driver might
1058                  *    enter reclaim, and deadlock if it waits on a page for
1059                  *    which it is needed to do the write (loop masks off
1060                  *    __GFP_IO|__GFP_FS for this reason); but more thought
1061                  *    would probably show more reasons.
1062                  *
1063                  * 3) Legacy memcg encounters a page that is already marked
1064                  *    PageReclaim. memcg does not have any dirty pages
1065                  *    throttling so we could easily OOM just because too many
1066                  *    pages are in writeback and there is nothing else to
1067                  *    reclaim. Wait for the writeback to complete.
1068                  *
1069                  * In cases 1) and 2) we activate the pages to get them out of
1070                  * the way while we continue scanning for clean pages on the
1071                  * inactive list and refilling from the active list. The
1072                  * observation here is that waiting for disk writes is more
1073                  * expensive than potentially causing reloads down the line.
1074                  * Since they're marked for immediate reclaim, they won't put
1075                  * memory pressure on the cache working set any longer than it
1076                  * takes to write them to disk.
1077                  */
1078                 if (PageWriteback(page)) {
1079                         /* Case 1 above */
1080                         if (current_is_kswapd() &&
1081                             PageReclaim(page) &&
1082                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1083                                 nr_immediate++;
1084                                 goto activate_locked;
1085
1086                         /* Case 2 above */
1087                         } else if (sane_reclaim(sc) ||
1088                             !PageReclaim(page) || !may_enter_fs) {
1089                                 /*
1090                                  * This is slightly racy - end_page_writeback()
1091                                  * might have just cleared PageReclaim, then
1092                                  * setting PageReclaim here end up interpreted
1093                                  * as PageReadahead - but that does not matter
1094                                  * enough to care.  What we do want is for this
1095                                  * page to have PageReclaim set next time memcg
1096                                  * reclaim reaches the tests above, so it will
1097                                  * then wait_on_page_writeback() to avoid OOM;
1098                                  * and it's also appropriate in global reclaim.
1099                                  */
1100                                 SetPageReclaim(page);
1101                                 nr_writeback++;
1102                                 goto activate_locked;
1103
1104                         /* Case 3 above */
1105                         } else {
1106                                 unlock_page(page);
1107                                 wait_on_page_writeback(page);
1108                                 /* then go back and try same page again */
1109                                 list_add_tail(&page->lru, page_list);
1110                                 continue;
1111                         }
1112                 }
1113
1114                 if (!force_reclaim)
1115                         references = page_check_references(page, sc);
1116
1117                 switch (references) {
1118                 case PAGEREF_ACTIVATE:
1119                         goto activate_locked;
1120                 case PAGEREF_KEEP:
1121                         nr_ref_keep++;
1122                         goto keep_locked;
1123                 case PAGEREF_RECLAIM:
1124                 case PAGEREF_RECLAIM_CLEAN:
1125                         ; /* try to reclaim the page below */
1126                 }
1127
1128                 /*
1129                  * Anonymous process memory has backing store?
1130                  * Try to allocate it some swap space here.
1131                  * Lazyfree page could be freed directly
1132                  */
1133                 if (PageAnon(page) && PageSwapBacked(page)) {
1134                         if (!PageSwapCache(page)) {
1135                                 if (!(sc->gfp_mask & __GFP_IO))
1136                                         goto keep_locked;
1137                                 if (PageTransHuge(page)) {
1138                                         /* cannot split THP, skip it */
1139                                         if (!can_split_huge_page(page, NULL))
1140                                                 goto activate_locked;
1141                                         /*
1142                                          * Split pages without a PMD map right
1143                                          * away. Chances are some or all of the
1144                                          * tail pages can be freed without IO.
1145                                          */
1146                                         if (!compound_mapcount(page) &&
1147                                             split_huge_page_to_list(page,
1148                                                                     page_list))
1149                                                 goto activate_locked;
1150                                 }
1151                                 if (!add_to_swap(page)) {
1152                                         if (!PageTransHuge(page))
1153                                                 goto activate_locked;
1154                                         /* Fallback to swap normal pages */
1155                                         if (split_huge_page_to_list(page,
1156                                                                     page_list))
1157                                                 goto activate_locked;
1158 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1159                                         count_vm_event(THP_SWPOUT_FALLBACK);
1160 #endif
1161                                         if (!add_to_swap(page))
1162                                                 goto activate_locked;
1163                                 }
1164
1165                                 may_enter_fs = 1;
1166
1167                                 /* Adding to swap updated mapping */
1168                                 mapping = page_mapping(page);
1169                         }
1170                 } else if (unlikely(PageTransHuge(page))) {
1171                         /* Split file THP */
1172                         if (split_huge_page_to_list(page, page_list))
1173                                 goto keep_locked;
1174                 }
1175
1176                 /*
1177                  * The page is mapped into the page tables of one or more
1178                  * processes. Try to unmap it here.
1179                  */
1180                 if (page_mapped(page)) {
1181                         enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1182
1183                         if (unlikely(PageTransHuge(page)))
1184                                 flags |= TTU_SPLIT_HUGE_PMD;
1185                         if (!try_to_unmap(page, flags)) {
1186                                 nr_unmap_fail++;
1187                                 goto activate_locked;
1188                         }
1189                 }
1190
1191                 if (PageDirty(page)) {
1192                         /*
1193                          * Only kswapd can writeback filesystem pages
1194                          * to avoid risk of stack overflow. But avoid
1195                          * injecting inefficient single-page IO into
1196                          * flusher writeback as much as possible: only
1197                          * write pages when we've encountered many
1198                          * dirty pages, and when we've already scanned
1199                          * the rest of the LRU for clean pages and see
1200                          * the same dirty pages again (PageReclaim).
1201                          */
1202                         if (page_is_file_cache(page) &&
1203                             (!current_is_kswapd() || !PageReclaim(page) ||
1204                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1205                                 /*
1206                                  * Immediately reclaim when written back.
1207                                  * Similar in principal to deactivate_page()
1208                                  * except we already have the page isolated
1209                                  * and know it's dirty
1210                                  */
1211                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1212                                 SetPageReclaim(page);
1213
1214                                 goto activate_locked;
1215                         }
1216
1217                         if (references == PAGEREF_RECLAIM_CLEAN)
1218                                 goto keep_locked;
1219                         if (!may_enter_fs)
1220                                 goto keep_locked;
1221                         if (!sc->may_writepage)
1222                                 goto keep_locked;
1223
1224                         /*
1225                          * Page is dirty. Flush the TLB if a writable entry
1226                          * potentially exists to avoid CPU writes after IO
1227                          * starts and then write it out here.
1228                          */
1229                         try_to_unmap_flush_dirty();
1230                         switch (pageout(page, mapping, sc)) {
1231                         case PAGE_KEEP:
1232                                 goto keep_locked;
1233                         case PAGE_ACTIVATE:
1234                                 goto activate_locked;
1235                         case PAGE_SUCCESS:
1236                                 if (PageWriteback(page))
1237                                         goto keep;
1238                                 if (PageDirty(page))
1239                                         goto keep;
1240
1241                                 /*
1242                                  * A synchronous write - probably a ramdisk.  Go
1243                                  * ahead and try to reclaim the page.
1244                                  */
1245                                 if (!trylock_page(page))
1246                                         goto keep;
1247                                 if (PageDirty(page) || PageWriteback(page))
1248                                         goto keep_locked;
1249                                 mapping = page_mapping(page);
1250                         case PAGE_CLEAN:
1251                                 ; /* try to free the page below */
1252                         }
1253                 }
1254
1255                 /*
1256                  * If the page has buffers, try to free the buffer mappings
1257                  * associated with this page. If we succeed we try to free
1258                  * the page as well.
1259                  *
1260                  * We do this even if the page is PageDirty().
1261                  * try_to_release_page() does not perform I/O, but it is
1262                  * possible for a page to have PageDirty set, but it is actually
1263                  * clean (all its buffers are clean).  This happens if the
1264                  * buffers were written out directly, with submit_bh(). ext3
1265                  * will do this, as well as the blockdev mapping.
1266                  * try_to_release_page() will discover that cleanness and will
1267                  * drop the buffers and mark the page clean - it can be freed.
1268                  *
1269                  * Rarely, pages can have buffers and no ->mapping.  These are
1270                  * the pages which were not successfully invalidated in
1271                  * truncate_complete_page().  We try to drop those buffers here
1272                  * and if that worked, and the page is no longer mapped into
1273                  * process address space (page_count == 1) it can be freed.
1274                  * Otherwise, leave the page on the LRU so it is swappable.
1275                  */
1276                 if (page_has_private(page)) {
1277                         if (!try_to_release_page(page, sc->gfp_mask))
1278                                 goto activate_locked;
1279                         if (!mapping && page_count(page) == 1) {
1280                                 unlock_page(page);
1281                                 if (put_page_testzero(page))
1282                                         goto free_it;
1283                                 else {
1284                                         /*
1285                                          * rare race with speculative reference.
1286                                          * the speculative reference will free
1287                                          * this page shortly, so we may
1288                                          * increment nr_reclaimed here (and
1289                                          * leave it off the LRU).
1290                                          */
1291                                         nr_reclaimed++;
1292                                         continue;
1293                                 }
1294                         }
1295                 }
1296
1297                 if (PageAnon(page) && !PageSwapBacked(page)) {
1298                         /* follow __remove_mapping for reference */
1299                         if (!page_ref_freeze(page, 1))
1300                                 goto keep_locked;
1301                         if (PageDirty(page)) {
1302                                 page_ref_unfreeze(page, 1);
1303                                 goto keep_locked;
1304                         }
1305
1306                         count_vm_event(PGLAZYFREED);
1307                         count_memcg_page_event(page, PGLAZYFREED);
1308                 } else if (!mapping || !__remove_mapping(mapping, page, true))
1309                         goto keep_locked;
1310                 /*
1311                  * At this point, we have no other references and there is
1312                  * no way to pick any more up (removed from LRU, removed
1313                  * from pagecache). Can use non-atomic bitops now (and
1314                  * we obviously don't have to worry about waking up a process
1315                  * waiting on the page lock, because there are no references.
1316                  */
1317                 __ClearPageLocked(page);
1318 free_it:
1319                 nr_reclaimed++;
1320
1321                 /*
1322                  * Is there need to periodically free_page_list? It would
1323                  * appear not as the counts should be low
1324                  */
1325                 if (unlikely(PageTransHuge(page))) {
1326                         mem_cgroup_uncharge(page);
1327                         (*get_compound_page_dtor(page))(page);
1328                 } else
1329                         list_add(&page->lru, &free_pages);
1330                 continue;
1331
1332 activate_locked:
1333                 /* Not a candidate for swapping, so reclaim swap space. */
1334                 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1335                                                 PageMlocked(page)))
1336                         try_to_free_swap(page);
1337                 VM_BUG_ON_PAGE(PageActive(page), page);
1338                 if (!PageMlocked(page)) {
1339                         SetPageActive(page);
1340                         pgactivate++;
1341                         count_memcg_page_event(page, PGACTIVATE);
1342                 }
1343 keep_locked:
1344                 unlock_page(page);
1345 keep:
1346                 list_add(&page->lru, &ret_pages);
1347                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1348         }
1349
1350         mem_cgroup_uncharge_list(&free_pages);
1351         try_to_unmap_flush();
1352         free_unref_page_list(&free_pages);
1353
1354         list_splice(&ret_pages, page_list);
1355         count_vm_events(PGACTIVATE, pgactivate);
1356
1357         if (stat) {
1358                 stat->nr_dirty = nr_dirty;
1359                 stat->nr_congested = nr_congested;
1360                 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1361                 stat->nr_writeback = nr_writeback;
1362                 stat->nr_immediate = nr_immediate;
1363                 stat->nr_activate = pgactivate;
1364                 stat->nr_ref_keep = nr_ref_keep;
1365                 stat->nr_unmap_fail = nr_unmap_fail;
1366         }
1367         return nr_reclaimed;
1368 }
1369
1370 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1371                                             struct list_head *page_list)
1372 {
1373         struct scan_control sc = {
1374                 .gfp_mask = GFP_KERNEL,
1375                 .priority = DEF_PRIORITY,
1376                 .may_unmap = 1,
1377         };
1378         unsigned long ret;
1379         struct page *page, *next;
1380         LIST_HEAD(clean_pages);
1381
1382         list_for_each_entry_safe(page, next, page_list, lru) {
1383                 if (page_is_file_cache(page) && !PageDirty(page) &&
1384                     !__PageMovable(page)) {
1385                         ClearPageActive(page);
1386                         list_move(&page->lru, &clean_pages);
1387                 }
1388         }
1389
1390         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1391                         TTU_IGNORE_ACCESS, NULL, true);
1392         list_splice(&clean_pages, page_list);
1393         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1394         return ret;
1395 }
1396
1397 /*
1398  * Attempt to remove the specified page from its LRU.  Only take this page
1399  * if it is of the appropriate PageActive status.  Pages which are being
1400  * freed elsewhere are also ignored.
1401  *
1402  * page:        page to consider
1403  * mode:        one of the LRU isolation modes defined above
1404  *
1405  * returns 0 on success, -ve errno on failure.
1406  */
1407 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1408 {
1409         int ret = -EINVAL;
1410
1411         /* Only take pages on the LRU. */
1412         if (!PageLRU(page))
1413                 return ret;
1414
1415         /* Compaction should not handle unevictable pages but CMA can do so */
1416         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1417                 return ret;
1418
1419         ret = -EBUSY;
1420
1421         /*
1422          * To minimise LRU disruption, the caller can indicate that it only
1423          * wants to isolate pages it will be able to operate on without
1424          * blocking - clean pages for the most part.
1425          *
1426          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1427          * that it is possible to migrate without blocking
1428          */
1429         if (mode & ISOLATE_ASYNC_MIGRATE) {
1430                 /* All the caller can do on PageWriteback is block */
1431                 if (PageWriteback(page))
1432                         return ret;
1433
1434                 if (PageDirty(page)) {
1435                         struct address_space *mapping;
1436
1437                         /*
1438                          * Only pages without mappings or that have a
1439                          * ->migratepage callback are possible to migrate
1440                          * without blocking
1441                          */
1442                         mapping = page_mapping(page);
1443                         if (mapping && !mapping->a_ops->migratepage)
1444                                 return ret;
1445                 }
1446         }
1447
1448         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1449                 return ret;
1450
1451         if (likely(get_page_unless_zero(page))) {
1452                 /*
1453                  * Be careful not to clear PageLRU until after we're
1454                  * sure the page is not being freed elsewhere -- the
1455                  * page release code relies on it.
1456                  */
1457                 ClearPageLRU(page);
1458                 ret = 0;
1459         }
1460
1461         return ret;
1462 }
1463
1464
1465 /*
1466  * Update LRU sizes after isolating pages. The LRU size updates must
1467  * be complete before mem_cgroup_update_lru_size due to a santity check.
1468  */
1469 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1470                         enum lru_list lru, unsigned long *nr_zone_taken)
1471 {
1472         int zid;
1473
1474         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1475                 if (!nr_zone_taken[zid])
1476                         continue;
1477
1478                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1479 #ifdef CONFIG_MEMCG
1480                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1481 #endif
1482         }
1483
1484 }
1485
1486 /*
1487  * zone_lru_lock is heavily contended.  Some of the functions that
1488  * shrink the lists perform better by taking out a batch of pages
1489  * and working on them outside the LRU lock.
1490  *
1491  * For pagecache intensive workloads, this function is the hottest
1492  * spot in the kernel (apart from copy_*_user functions).
1493  *
1494  * Appropriate locks must be held before calling this function.
1495  *
1496  * @nr_to_scan: The number of eligible pages to look through on the list.
1497  * @lruvec:     The LRU vector to pull pages from.
1498  * @dst:        The temp list to put pages on to.
1499  * @nr_scanned: The number of pages that were scanned.
1500  * @sc:         The scan_control struct for this reclaim session
1501  * @mode:       One of the LRU isolation modes
1502  * @lru:        LRU list id for isolating
1503  *
1504  * returns how many pages were moved onto *@dst.
1505  */
1506 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1507                 struct lruvec *lruvec, struct list_head *dst,
1508                 unsigned long *nr_scanned, struct scan_control *sc,
1509                 isolate_mode_t mode, enum lru_list lru)
1510 {
1511         struct list_head *src = &lruvec->lists[lru];
1512         unsigned long nr_taken = 0;
1513         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1514         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1515         unsigned long skipped = 0;
1516         unsigned long scan, total_scan, nr_pages;
1517         LIST_HEAD(pages_skipped);
1518
1519         scan = 0;
1520         for (total_scan = 0;
1521              scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1522              total_scan++) {
1523                 struct page *page;
1524
1525                 page = lru_to_page(src);
1526                 prefetchw_prev_lru_page(page, src, flags);
1527
1528                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1529
1530                 if (page_zonenum(page) > sc->reclaim_idx) {
1531                         list_move(&page->lru, &pages_skipped);
1532                         nr_skipped[page_zonenum(page)]++;
1533                         continue;
1534                 }
1535
1536                 /*
1537                  * Do not count skipped pages because that makes the function
1538                  * return with no isolated pages if the LRU mostly contains
1539                  * ineligible pages.  This causes the VM to not reclaim any
1540                  * pages, triggering a premature OOM.
1541                  */
1542                 scan++;
1543                 switch (__isolate_lru_page(page, mode)) {
1544                 case 0:
1545                         nr_pages = hpage_nr_pages(page);
1546                         nr_taken += nr_pages;
1547                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1548                         list_move(&page->lru, dst);
1549                         break;
1550
1551                 case -EBUSY:
1552                         /* else it is being freed elsewhere */
1553                         list_move(&page->lru, src);
1554                         continue;
1555
1556                 default:
1557                         BUG();
1558                 }
1559         }
1560
1561         /*
1562          * Splice any skipped pages to the start of the LRU list. Note that
1563          * this disrupts the LRU order when reclaiming for lower zones but
1564          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1565          * scanning would soon rescan the same pages to skip and put the
1566          * system at risk of premature OOM.
1567          */
1568         if (!list_empty(&pages_skipped)) {
1569                 int zid;
1570
1571                 list_splice(&pages_skipped, src);
1572                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1573                         if (!nr_skipped[zid])
1574                                 continue;
1575
1576                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1577                         skipped += nr_skipped[zid];
1578                 }
1579         }
1580         *nr_scanned = total_scan;
1581         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1582                                     total_scan, skipped, nr_taken, mode, lru);
1583         update_lru_sizes(lruvec, lru, nr_zone_taken);
1584         return nr_taken;
1585 }
1586
1587 /**
1588  * isolate_lru_page - tries to isolate a page from its LRU list
1589  * @page: page to isolate from its LRU list
1590  *
1591  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1592  * vmstat statistic corresponding to whatever LRU list the page was on.
1593  *
1594  * Returns 0 if the page was removed from an LRU list.
1595  * Returns -EBUSY if the page was not on an LRU list.
1596  *
1597  * The returned page will have PageLRU() cleared.  If it was found on
1598  * the active list, it will have PageActive set.  If it was found on
1599  * the unevictable list, it will have the PageUnevictable bit set. That flag
1600  * may need to be cleared by the caller before letting the page go.
1601  *
1602  * The vmstat statistic corresponding to the list on which the page was
1603  * found will be decremented.
1604  *
1605  * Restrictions:
1606  * (1) Must be called with an elevated refcount on the page. This is a
1607  *     fundamentnal difference from isolate_lru_pages (which is called
1608  *     without a stable reference).
1609  * (2) the lru_lock must not be held.
1610  * (3) interrupts must be enabled.
1611  */
1612 int isolate_lru_page(struct page *page)
1613 {
1614         int ret = -EBUSY;
1615
1616         VM_BUG_ON_PAGE(!page_count(page), page);
1617         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1618
1619         if (PageLRU(page)) {
1620                 struct zone *zone = page_zone(page);
1621                 struct lruvec *lruvec;
1622
1623                 spin_lock_irq(zone_lru_lock(zone));
1624                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1625                 if (PageLRU(page)) {
1626                         int lru = page_lru(page);
1627                         get_page(page);
1628                         ClearPageLRU(page);
1629                         del_page_from_lru_list(page, lruvec, lru);
1630                         ret = 0;
1631                 }
1632                 spin_unlock_irq(zone_lru_lock(zone));
1633         }
1634         return ret;
1635 }
1636
1637 /*
1638  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1639  * then get resheduled. When there are massive number of tasks doing page
1640  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1641  * the LRU list will go small and be scanned faster than necessary, leading to
1642  * unnecessary swapping, thrashing and OOM.
1643  */
1644 static int too_many_isolated(struct pglist_data *pgdat, int file,
1645                 struct scan_control *sc)
1646 {
1647         unsigned long inactive, isolated;
1648
1649         if (current_is_kswapd())
1650                 return 0;
1651
1652         if (!sane_reclaim(sc))
1653                 return 0;
1654
1655         if (file) {
1656                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1657                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1658         } else {
1659                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1660                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1661         }
1662
1663         /*
1664          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1665          * won't get blocked by normal direct-reclaimers, forming a circular
1666          * deadlock.
1667          */
1668         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1669                 inactive >>= 3;
1670
1671         return isolated > inactive;
1672 }
1673
1674 static noinline_for_stack void
1675 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1676 {
1677         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1678         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1679         LIST_HEAD(pages_to_free);
1680
1681         /*
1682          * Put back any unfreeable pages.
1683          */
1684         while (!list_empty(page_list)) {
1685                 struct page *page = lru_to_page(page_list);
1686                 int lru;
1687
1688                 VM_BUG_ON_PAGE(PageLRU(page), page);
1689                 list_del(&page->lru);
1690                 if (unlikely(!page_evictable(page))) {
1691                         spin_unlock_irq(&pgdat->lru_lock);
1692                         putback_lru_page(page);
1693                         spin_lock_irq(&pgdat->lru_lock);
1694                         continue;
1695                 }
1696
1697                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1698
1699                 SetPageLRU(page);
1700                 lru = page_lru(page);
1701                 add_page_to_lru_list(page, lruvec, lru);
1702
1703                 if (is_active_lru(lru)) {
1704                         int file = is_file_lru(lru);
1705                         int numpages = hpage_nr_pages(page);
1706                         reclaim_stat->recent_rotated[file] += numpages;
1707                 }
1708                 if (put_page_testzero(page)) {
1709                         __ClearPageLRU(page);
1710                         __ClearPageActive(page);
1711                         del_page_from_lru_list(page, lruvec, lru);
1712
1713                         if (unlikely(PageCompound(page))) {
1714                                 spin_unlock_irq(&pgdat->lru_lock);
1715                                 mem_cgroup_uncharge(page);
1716                                 (*get_compound_page_dtor(page))(page);
1717                                 spin_lock_irq(&pgdat->lru_lock);
1718                         } else
1719                                 list_add(&page->lru, &pages_to_free);
1720                 }
1721         }
1722
1723         /*
1724          * To save our caller's stack, now use input list for pages to free.
1725          */
1726         list_splice(&pages_to_free, page_list);
1727 }
1728
1729 /*
1730  * If a kernel thread (such as nfsd for loop-back mounts) services
1731  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1732  * In that case we should only throttle if the backing device it is
1733  * writing to is congested.  In other cases it is safe to throttle.
1734  */
1735 static int current_may_throttle(void)
1736 {
1737         return !(current->flags & PF_LESS_THROTTLE) ||
1738                 current->backing_dev_info == NULL ||
1739                 bdi_write_congested(current->backing_dev_info);
1740 }
1741
1742 /*
1743  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1744  * of reclaimed pages
1745  */
1746 static noinline_for_stack unsigned long
1747 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1748                      struct scan_control *sc, enum lru_list lru)
1749 {
1750         LIST_HEAD(page_list);
1751         unsigned long nr_scanned;
1752         unsigned long nr_reclaimed = 0;
1753         unsigned long nr_taken;
1754         struct reclaim_stat stat = {};
1755         isolate_mode_t isolate_mode = 0;
1756         int file = is_file_lru(lru);
1757         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1758         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1759         bool stalled = false;
1760
1761         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1762                 if (stalled)
1763                         return 0;
1764
1765                 /* wait a bit for the reclaimer. */
1766                 msleep(100);
1767                 stalled = true;
1768
1769                 /* We are about to die and free our memory. Return now. */
1770                 if (fatal_signal_pending(current))
1771                         return SWAP_CLUSTER_MAX;
1772         }
1773
1774         lru_add_drain();
1775
1776         if (!sc->may_unmap)
1777                 isolate_mode |= ISOLATE_UNMAPPED;
1778
1779         spin_lock_irq(&pgdat->lru_lock);
1780
1781         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1782                                      &nr_scanned, sc, isolate_mode, lru);
1783
1784         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1785         reclaim_stat->recent_scanned[file] += nr_taken;
1786
1787         if (current_is_kswapd()) {
1788                 if (global_reclaim(sc))
1789                         __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1790                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1791                                    nr_scanned);
1792         } else {
1793                 if (global_reclaim(sc))
1794                         __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1795                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1796                                    nr_scanned);
1797         }
1798         spin_unlock_irq(&pgdat->lru_lock);
1799
1800         if (nr_taken == 0)
1801                 return 0;
1802
1803         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1804                                 &stat, false);
1805
1806         spin_lock_irq(&pgdat->lru_lock);
1807
1808         if (current_is_kswapd()) {
1809                 if (global_reclaim(sc))
1810                         __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1811                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1812                                    nr_reclaimed);
1813         } else {
1814                 if (global_reclaim(sc))
1815                         __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1816                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1817                                    nr_reclaimed);
1818         }
1819
1820         putback_inactive_pages(lruvec, &page_list);
1821
1822         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1823
1824         spin_unlock_irq(&pgdat->lru_lock);
1825
1826         mem_cgroup_uncharge_list(&page_list);
1827         free_unref_page_list(&page_list);
1828
1829         /*
1830          * If reclaim is isolating dirty pages under writeback, it implies
1831          * that the long-lived page allocation rate is exceeding the page
1832          * laundering rate. Either the global limits are not being effective
1833          * at throttling processes due to the page distribution throughout
1834          * zones or there is heavy usage of a slow backing device. The
1835          * only option is to throttle from reclaim context which is not ideal
1836          * as there is no guarantee the dirtying process is throttled in the
1837          * same way balance_dirty_pages() manages.
1838          *
1839          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1840          * of pages under pages flagged for immediate reclaim and stall if any
1841          * are encountered in the nr_immediate check below.
1842          */
1843         if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1844                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1845
1846         /*
1847          * Legacy memcg will stall in page writeback so avoid forcibly
1848          * stalling here.
1849          */
1850         if (sane_reclaim(sc)) {
1851                 /*
1852                  * Tag a zone as congested if all the dirty pages scanned were
1853                  * backed by a congested BDI and wait_iff_congested will stall.
1854                  */
1855                 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1856                         set_bit(PGDAT_CONGESTED, &pgdat->flags);
1857
1858                 /*
1859                  * If dirty pages are scanned that are not queued for IO, it
1860                  * implies that flushers are not doing their job. This can
1861                  * happen when memory pressure pushes dirty pages to the end of
1862                  * the LRU before the dirty limits are breached and the dirty
1863                  * data has expired. It can also happen when the proportion of
1864                  * dirty pages grows not through writes but through memory
1865                  * pressure reclaiming all the clean cache. And in some cases,
1866                  * the flushers simply cannot keep up with the allocation
1867                  * rate. Nudge the flusher threads in case they are asleep, but
1868                  * also allow kswapd to start writing pages during reclaim.
1869                  */
1870                 if (stat.nr_unqueued_dirty == nr_taken) {
1871                         wakeup_flusher_threads(WB_REASON_VMSCAN);
1872                         set_bit(PGDAT_DIRTY, &pgdat->flags);
1873                 }
1874
1875                 /*
1876                  * If kswapd scans pages marked marked for immediate
1877                  * reclaim and under writeback (nr_immediate), it implies
1878                  * that pages are cycling through the LRU faster than
1879                  * they are written so also forcibly stall.
1880                  */
1881                 if (stat.nr_immediate && current_may_throttle())
1882                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1883         }
1884
1885         /*
1886          * Stall direct reclaim for IO completions if underlying BDIs or zone
1887          * is congested. Allow kswapd to continue until it starts encountering
1888          * unqueued dirty pages or cycling through the LRU too quickly.
1889          */
1890         if (!sc->hibernation_mode && !current_is_kswapd() &&
1891             current_may_throttle())
1892                 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1893
1894         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1895                         nr_scanned, nr_reclaimed,
1896                         stat.nr_dirty,  stat.nr_writeback,
1897                         stat.nr_congested, stat.nr_immediate,
1898                         stat.nr_activate, stat.nr_ref_keep,
1899                         stat.nr_unmap_fail,
1900                         sc->priority, file);
1901         return nr_reclaimed;
1902 }
1903
1904 /*
1905  * This moves pages from the active list to the inactive list.
1906  *
1907  * We move them the other way if the page is referenced by one or more
1908  * processes, from rmap.
1909  *
1910  * If the pages are mostly unmapped, the processing is fast and it is
1911  * appropriate to hold zone_lru_lock across the whole operation.  But if
1912  * the pages are mapped, the processing is slow (page_referenced()) so we
1913  * should drop zone_lru_lock around each page.  It's impossible to balance
1914  * this, so instead we remove the pages from the LRU while processing them.
1915  * It is safe to rely on PG_active against the non-LRU pages in here because
1916  * nobody will play with that bit on a non-LRU page.
1917  *
1918  * The downside is that we have to touch page->_refcount against each page.
1919  * But we had to alter page->flags anyway.
1920  *
1921  * Returns the number of pages moved to the given lru.
1922  */
1923
1924 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1925                                      struct list_head *list,
1926                                      struct list_head *pages_to_free,
1927                                      enum lru_list lru)
1928 {
1929         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1930         struct page *page;
1931         int nr_pages;
1932         int nr_moved = 0;
1933
1934         while (!list_empty(list)) {
1935                 page = lru_to_page(list);
1936                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1937
1938                 VM_BUG_ON_PAGE(PageLRU(page), page);
1939                 SetPageLRU(page);
1940
1941                 nr_pages = hpage_nr_pages(page);
1942                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1943                 list_move(&page->lru, &lruvec->lists[lru]);
1944
1945                 if (put_page_testzero(page)) {
1946                         __ClearPageLRU(page);
1947                         __ClearPageActive(page);
1948                         del_page_from_lru_list(page, lruvec, lru);
1949
1950                         if (unlikely(PageCompound(page))) {
1951                                 spin_unlock_irq(&pgdat->lru_lock);
1952                                 mem_cgroup_uncharge(page);
1953                                 (*get_compound_page_dtor(page))(page);
1954                                 spin_lock_irq(&pgdat->lru_lock);
1955                         } else
1956                                 list_add(&page->lru, pages_to_free);
1957                 } else {
1958                         nr_moved += nr_pages;
1959                 }
1960         }
1961
1962         if (!is_active_lru(lru)) {
1963                 __count_vm_events(PGDEACTIVATE, nr_moved);
1964                 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1965                                    nr_moved);
1966         }
1967
1968         return nr_moved;
1969 }
1970
1971 static void shrink_active_list(unsigned long nr_to_scan,
1972                                struct lruvec *lruvec,
1973                                struct scan_control *sc,
1974                                enum lru_list lru)
1975 {
1976         unsigned long nr_taken;
1977         unsigned long nr_scanned;
1978         unsigned long vm_flags;
1979         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1980         LIST_HEAD(l_active);
1981         LIST_HEAD(l_inactive);
1982         struct page *page;
1983         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1984         unsigned nr_deactivate, nr_activate;
1985         unsigned nr_rotated = 0;
1986         isolate_mode_t isolate_mode = 0;
1987         int file = is_file_lru(lru);
1988         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1989
1990         lru_add_drain();
1991
1992         if (!sc->may_unmap)
1993                 isolate_mode |= ISOLATE_UNMAPPED;
1994
1995         spin_lock_irq(&pgdat->lru_lock);
1996
1997         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1998                                      &nr_scanned, sc, isolate_mode, lru);
1999
2000         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2001         reclaim_stat->recent_scanned[file] += nr_taken;
2002
2003         __count_vm_events(PGREFILL, nr_scanned);
2004         count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2005
2006         spin_unlock_irq(&pgdat->lru_lock);
2007
2008         while (!list_empty(&l_hold)) {
2009                 cond_resched();
2010                 page = lru_to_page(&l_hold);
2011                 list_del(&page->lru);
2012
2013                 if (unlikely(!page_evictable(page))) {
2014                         putback_lru_page(page);
2015                         continue;
2016                 }
2017
2018                 if (unlikely(buffer_heads_over_limit)) {
2019                         if (page_has_private(page) && trylock_page(page)) {
2020                                 if (page_has_private(page))
2021                                         try_to_release_page(page, 0);
2022                                 unlock_page(page);
2023                         }
2024                 }
2025
2026                 if (page_referenced(page, 0, sc->target_mem_cgroup,
2027                                     &vm_flags)) {
2028                         nr_rotated += hpage_nr_pages(page);
2029                         /*
2030                          * Identify referenced, file-backed active pages and
2031                          * give them one more trip around the active list. So
2032                          * that executable code get better chances to stay in
2033                          * memory under moderate memory pressure.  Anon pages
2034                          * are not likely to be evicted by use-once streaming
2035                          * IO, plus JVM can create lots of anon VM_EXEC pages,
2036                          * so we ignore them here.
2037                          */
2038                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2039                                 list_add(&page->lru, &l_active);
2040                                 continue;
2041                         }
2042                 }
2043
2044                 ClearPageActive(page);  /* we are de-activating */
2045                 list_add(&page->lru, &l_inactive);
2046         }
2047
2048         /*
2049          * Move pages back to the lru list.
2050          */
2051         spin_lock_irq(&pgdat->lru_lock);
2052         /*
2053          * Count referenced pages from currently used mappings as rotated,
2054          * even though only some of them are actually re-activated.  This
2055          * helps balance scan pressure between file and anonymous pages in
2056          * get_scan_count.
2057          */
2058         reclaim_stat->recent_rotated[file] += nr_rotated;
2059
2060         nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2061         nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2062         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2063         spin_unlock_irq(&pgdat->lru_lock);
2064
2065         mem_cgroup_uncharge_list(&l_hold);
2066         free_unref_page_list(&l_hold);
2067         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2068                         nr_deactivate, nr_rotated, sc->priority, file);
2069 }
2070
2071 /*
2072  * The inactive anon list should be small enough that the VM never has
2073  * to do too much work.
2074  *
2075  * The inactive file list should be small enough to leave most memory
2076  * to the established workingset on the scan-resistant active list,
2077  * but large enough to avoid thrashing the aggregate readahead window.
2078  *
2079  * Both inactive lists should also be large enough that each inactive
2080  * page has a chance to be referenced again before it is reclaimed.
2081  *
2082  * If that fails and refaulting is observed, the inactive list grows.
2083  *
2084  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2085  * on this LRU, maintained by the pageout code. An inactive_ratio
2086  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2087  *
2088  * total     target    max
2089  * memory    ratio     inactive
2090  * -------------------------------------
2091  *   10MB       1         5MB
2092  *  100MB       1        50MB
2093  *    1GB       3       250MB
2094  *   10GB      10       0.9GB
2095  *  100GB      31         3GB
2096  *    1TB     101        10GB
2097  *   10TB     320        32GB
2098  */
2099 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2100                                  struct mem_cgroup *memcg,
2101                                  struct scan_control *sc, bool actual_reclaim)
2102 {
2103         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2104         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2105         enum lru_list inactive_lru = file * LRU_FILE;
2106         unsigned long inactive, active;
2107         unsigned long inactive_ratio;
2108         unsigned long refaults;
2109         unsigned long gb;
2110
2111         /*
2112          * If we don't have swap space, anonymous page deactivation
2113          * is pointless.
2114          */
2115         if (!file && !total_swap_pages)
2116                 return false;
2117
2118         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2119         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2120
2121         if (memcg)
2122                 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2123         else
2124                 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2125
2126         /*
2127          * When refaults are being observed, it means a new workingset
2128          * is being established. Disable active list protection to get
2129          * rid of the stale workingset quickly.
2130          */
2131         if (file && actual_reclaim && lruvec->refaults != refaults) {
2132                 inactive_ratio = 0;
2133         } else {
2134                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2135                 if (gb)
2136                         inactive_ratio = int_sqrt(10 * gb);
2137                 else
2138                         inactive_ratio = 1;
2139         }
2140
2141         if (actual_reclaim)
2142                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2143                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2144                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2145                         inactive_ratio, file);
2146
2147         return inactive * inactive_ratio < active;
2148 }
2149
2150 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2151                                  struct lruvec *lruvec, struct mem_cgroup *memcg,
2152                                  struct scan_control *sc)
2153 {
2154         if (is_active_lru(lru)) {
2155                 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2156                                          memcg, sc, true))
2157                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2158                 return 0;
2159         }
2160
2161         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2162 }
2163
2164 enum scan_balance {
2165         SCAN_EQUAL,
2166         SCAN_FRACT,
2167         SCAN_ANON,
2168         SCAN_FILE,
2169 };
2170
2171 /*
2172  * Determine how aggressively the anon and file LRU lists should be
2173  * scanned.  The relative value of each set of LRU lists is determined
2174  * by looking at the fraction of the pages scanned we did rotate back
2175  * onto the active list instead of evict.
2176  *
2177  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2178  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2179  */
2180 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2181                            struct scan_control *sc, unsigned long *nr,
2182                            unsigned long *lru_pages)
2183 {
2184         int swappiness = mem_cgroup_swappiness(memcg);
2185         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2186         u64 fraction[2];
2187         u64 denominator = 0;    /* gcc */
2188         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2189         unsigned long anon_prio, file_prio;
2190         enum scan_balance scan_balance;
2191         unsigned long anon, file;
2192         unsigned long ap, fp;
2193         enum lru_list lru;
2194
2195         /* If we have no swap space, do not bother scanning anon pages. */
2196         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2197                 scan_balance = SCAN_FILE;
2198                 goto out;
2199         }
2200
2201         /*
2202          * Global reclaim will swap to prevent OOM even with no
2203          * swappiness, but memcg users want to use this knob to
2204          * disable swapping for individual groups completely when
2205          * using the memory controller's swap limit feature would be
2206          * too expensive.
2207          */
2208         if (!global_reclaim(sc) && !swappiness) {
2209                 scan_balance = SCAN_FILE;
2210                 goto out;
2211         }
2212
2213         /*
2214          * Do not apply any pressure balancing cleverness when the
2215          * system is close to OOM, scan both anon and file equally
2216          * (unless the swappiness setting disagrees with swapping).
2217          */
2218         if (!sc->priority && swappiness) {
2219                 scan_balance = SCAN_EQUAL;
2220                 goto out;
2221         }
2222
2223         /*
2224          * Prevent the reclaimer from falling into the cache trap: as
2225          * cache pages start out inactive, every cache fault will tip
2226          * the scan balance towards the file LRU.  And as the file LRU
2227          * shrinks, so does the window for rotation from references.
2228          * This means we have a runaway feedback loop where a tiny
2229          * thrashing file LRU becomes infinitely more attractive than
2230          * anon pages.  Try to detect this based on file LRU size.
2231          */
2232         if (global_reclaim(sc)) {
2233                 unsigned long pgdatfile;
2234                 unsigned long pgdatfree;
2235                 int z;
2236                 unsigned long total_high_wmark = 0;
2237
2238                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2239                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2240                            node_page_state(pgdat, NR_INACTIVE_FILE);
2241
2242                 for (z = 0; z < MAX_NR_ZONES; z++) {
2243                         struct zone *zone = &pgdat->node_zones[z];
2244                         if (!managed_zone(zone))
2245                                 continue;
2246
2247                         total_high_wmark += high_wmark_pages(zone);
2248                 }
2249
2250                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2251                         /*
2252                          * Force SCAN_ANON if there are enough inactive
2253                          * anonymous pages on the LRU in eligible zones.
2254                          * Otherwise, the small LRU gets thrashed.
2255                          */
2256                         if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2257                             lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2258                                         >> sc->priority) {
2259                                 scan_balance = SCAN_ANON;
2260                                 goto out;
2261                         }
2262                 }
2263         }
2264
2265         /*
2266          * If there is enough inactive page cache, i.e. if the size of the
2267          * inactive list is greater than that of the active list *and* the
2268          * inactive list actually has some pages to scan on this priority, we
2269          * do not reclaim anything from the anonymous working set right now.
2270          * Without the second condition we could end up never scanning an
2271          * lruvec even if it has plenty of old anonymous pages unless the
2272          * system is under heavy pressure.
2273          */
2274         if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2275             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2276                 scan_balance = SCAN_FILE;
2277                 goto out;
2278         }
2279
2280         scan_balance = SCAN_FRACT;
2281
2282         /*
2283          * With swappiness at 100, anonymous and file have the same priority.
2284          * This scanning priority is essentially the inverse of IO cost.
2285          */
2286         anon_prio = swappiness;
2287         file_prio = 200 - anon_prio;
2288
2289         /*
2290          * OK, so we have swap space and a fair amount of page cache
2291          * pages.  We use the recently rotated / recently scanned
2292          * ratios to determine how valuable each cache is.
2293          *
2294          * Because workloads change over time (and to avoid overflow)
2295          * we keep these statistics as a floating average, which ends
2296          * up weighing recent references more than old ones.
2297          *
2298          * anon in [0], file in [1]
2299          */
2300
2301         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2302                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2303         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2304                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2305
2306         spin_lock_irq(&pgdat->lru_lock);
2307         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2308                 reclaim_stat->recent_scanned[0] /= 2;
2309                 reclaim_stat->recent_rotated[0] /= 2;
2310         }
2311
2312         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2313                 reclaim_stat->recent_scanned[1] /= 2;
2314                 reclaim_stat->recent_rotated[1] /= 2;
2315         }
2316
2317         /*
2318          * The amount of pressure on anon vs file pages is inversely
2319          * proportional to the fraction of recently scanned pages on
2320          * each list that were recently referenced and in active use.
2321          */
2322         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2323         ap /= reclaim_stat->recent_rotated[0] + 1;
2324
2325         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2326         fp /= reclaim_stat->recent_rotated[1] + 1;
2327         spin_unlock_irq(&pgdat->lru_lock);
2328
2329         fraction[0] = ap;
2330         fraction[1] = fp;
2331         denominator = ap + fp + 1;
2332 out:
2333         *lru_pages = 0;
2334         for_each_evictable_lru(lru) {
2335                 int file = is_file_lru(lru);
2336                 unsigned long size;
2337                 unsigned long scan;
2338
2339                 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2340                 scan = size >> sc->priority;
2341                 /*
2342                  * If the cgroup's already been deleted, make sure to
2343                  * scrape out the remaining cache.
2344                  */
2345                 if (!scan && !mem_cgroup_online(memcg))
2346                         scan = min(size, SWAP_CLUSTER_MAX);
2347
2348                 switch (scan_balance) {
2349                 case SCAN_EQUAL:
2350                         /* Scan lists relative to size */
2351                         break;
2352                 case SCAN_FRACT:
2353                         /*
2354                          * Scan types proportional to swappiness and
2355                          * their relative recent reclaim efficiency.
2356                          */
2357                         scan = div64_u64(scan * fraction[file],
2358                                          denominator);
2359                         break;
2360                 case SCAN_FILE:
2361                 case SCAN_ANON:
2362                         /* Scan one type exclusively */
2363                         if ((scan_balance == SCAN_FILE) != file) {
2364                                 size = 0;
2365                                 scan = 0;
2366                         }
2367                         break;
2368                 default:
2369                         /* Look ma, no brain */
2370                         BUG();
2371                 }
2372
2373                 *lru_pages += size;
2374                 nr[lru] = scan;
2375         }
2376 }
2377
2378 /*
2379  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2380  */
2381 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2382                               struct scan_control *sc, unsigned long *lru_pages)
2383 {
2384         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2385         unsigned long nr[NR_LRU_LISTS];
2386         unsigned long targets[NR_LRU_LISTS];
2387         unsigned long nr_to_scan;
2388         enum lru_list lru;
2389         unsigned long nr_reclaimed = 0;
2390         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2391         struct blk_plug plug;
2392         bool scan_adjusted;
2393
2394         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2395
2396         /* Record the original scan target for proportional adjustments later */
2397         memcpy(targets, nr, sizeof(nr));
2398
2399         /*
2400          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2401          * event that can occur when there is little memory pressure e.g.
2402          * multiple streaming readers/writers. Hence, we do not abort scanning
2403          * when the requested number of pages are reclaimed when scanning at
2404          * DEF_PRIORITY on the assumption that the fact we are direct
2405          * reclaiming implies that kswapd is not keeping up and it is best to
2406          * do a batch of work at once. For memcg reclaim one check is made to
2407          * abort proportional reclaim if either the file or anon lru has already
2408          * dropped to zero at the first pass.
2409          */
2410         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2411                          sc->priority == DEF_PRIORITY);
2412
2413         blk_start_plug(&plug);
2414         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2415                                         nr[LRU_INACTIVE_FILE]) {
2416                 unsigned long nr_anon, nr_file, percentage;
2417                 unsigned long nr_scanned;
2418
2419                 for_each_evictable_lru(lru) {
2420                         if (nr[lru]) {
2421                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2422                                 nr[lru] -= nr_to_scan;
2423
2424                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2425                                                             lruvec, memcg, sc);
2426                         }
2427                 }
2428
2429                 cond_resched();
2430
2431                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2432                         continue;
2433
2434                 /*
2435                  * For kswapd and memcg, reclaim at least the number of pages
2436                  * requested. Ensure that the anon and file LRUs are scanned
2437                  * proportionally what was requested by get_scan_count(). We
2438                  * stop reclaiming one LRU and reduce the amount scanning
2439                  * proportional to the original scan target.
2440                  */
2441                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2442                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2443
2444                 /*
2445                  * It's just vindictive to attack the larger once the smaller
2446                  * has gone to zero.  And given the way we stop scanning the
2447                  * smaller below, this makes sure that we only make one nudge
2448                  * towards proportionality once we've got nr_to_reclaim.
2449                  */
2450                 if (!nr_file || !nr_anon)
2451                         break;
2452
2453                 if (nr_file > nr_anon) {
2454                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2455                                                 targets[LRU_ACTIVE_ANON] + 1;
2456                         lru = LRU_BASE;
2457                         percentage = nr_anon * 100 / scan_target;
2458                 } else {
2459                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2460                                                 targets[LRU_ACTIVE_FILE] + 1;
2461                         lru = LRU_FILE;
2462                         percentage = nr_file * 100 / scan_target;
2463                 }
2464
2465                 /* Stop scanning the smaller of the LRU */
2466                 nr[lru] = 0;
2467                 nr[lru + LRU_ACTIVE] = 0;
2468
2469                 /*
2470                  * Recalculate the other LRU scan count based on its original
2471                  * scan target and the percentage scanning already complete
2472                  */
2473                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2474                 nr_scanned = targets[lru] - nr[lru];
2475                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2476                 nr[lru] -= min(nr[lru], nr_scanned);
2477
2478                 lru += LRU_ACTIVE;
2479                 nr_scanned = targets[lru] - nr[lru];
2480                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2481                 nr[lru] -= min(nr[lru], nr_scanned);
2482
2483                 scan_adjusted = true;
2484         }
2485         blk_finish_plug(&plug);
2486         sc->nr_reclaimed += nr_reclaimed;
2487
2488         /*
2489          * Even if we did not try to evict anon pages at all, we want to
2490          * rebalance the anon lru active/inactive ratio.
2491          */
2492         if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2493                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2494                                    sc, LRU_ACTIVE_ANON);
2495 }
2496
2497 /* Use reclaim/compaction for costly allocs or under memory pressure */
2498 static bool in_reclaim_compaction(struct scan_control *sc)
2499 {
2500         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2501                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2502                          sc->priority < DEF_PRIORITY - 2))
2503                 return true;
2504
2505         return false;
2506 }
2507
2508 /*
2509  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2510  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2511  * true if more pages should be reclaimed such that when the page allocator
2512  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2513  * It will give up earlier than that if there is difficulty reclaiming pages.
2514  */
2515 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2516                                         unsigned long nr_reclaimed,
2517                                         unsigned long nr_scanned,
2518                                         struct scan_control *sc)
2519 {
2520         unsigned long pages_for_compaction;
2521         unsigned long inactive_lru_pages;
2522         int z;
2523
2524         /* If not in reclaim/compaction mode, stop */
2525         if (!in_reclaim_compaction(sc))
2526                 return false;
2527
2528         /* Consider stopping depending on scan and reclaim activity */
2529         if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2530                 /*
2531                  * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2532                  * full LRU list has been scanned and we are still failing
2533                  * to reclaim pages. This full LRU scan is potentially
2534                  * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2535                  */
2536                 if (!nr_reclaimed && !nr_scanned)
2537                         return false;
2538         } else {
2539                 /*
2540                  * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2541                  * fail without consequence, stop if we failed to reclaim
2542                  * any pages from the last SWAP_CLUSTER_MAX number of
2543                  * pages that were scanned. This will return to the
2544                  * caller faster at the risk reclaim/compaction and
2545                  * the resulting allocation attempt fails
2546                  */
2547                 if (!nr_reclaimed)
2548                         return false;
2549         }
2550
2551         /*
2552          * If we have not reclaimed enough pages for compaction and the
2553          * inactive lists are large enough, continue reclaiming
2554          */
2555         pages_for_compaction = compact_gap(sc->order);
2556         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2557         if (get_nr_swap_pages() > 0)
2558                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2559         if (sc->nr_reclaimed < pages_for_compaction &&
2560                         inactive_lru_pages > pages_for_compaction)
2561                 return true;
2562
2563         /* If compaction would go ahead or the allocation would succeed, stop */
2564         for (z = 0; z <= sc->reclaim_idx; z++) {
2565                 struct zone *zone = &pgdat->node_zones[z];
2566                 if (!managed_zone(zone))
2567                         continue;
2568
2569                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2570                 case COMPACT_SUCCESS:
2571                 case COMPACT_CONTINUE:
2572                         return false;
2573                 default:
2574                         /* check next zone */
2575                         ;
2576                 }
2577         }
2578         return true;
2579 }
2580
2581 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2582 {
2583         struct reclaim_state *reclaim_state = current->reclaim_state;
2584         unsigned long nr_reclaimed, nr_scanned;
2585         bool reclaimable = false;
2586
2587         do {
2588                 struct mem_cgroup *root = sc->target_mem_cgroup;
2589                 struct mem_cgroup_reclaim_cookie reclaim = {
2590                         .pgdat = pgdat,
2591                         .priority = sc->priority,
2592                 };
2593                 unsigned long node_lru_pages = 0;
2594                 struct mem_cgroup *memcg;
2595
2596                 nr_reclaimed = sc->nr_reclaimed;
2597                 nr_scanned = sc->nr_scanned;
2598
2599                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2600                 do {
2601                         unsigned long lru_pages;
2602                         unsigned long reclaimed;
2603                         unsigned long scanned;
2604
2605                         if (mem_cgroup_low(root, memcg)) {
2606                                 if (!sc->memcg_low_reclaim) {
2607                                         sc->memcg_low_skipped = 1;
2608                                         continue;
2609                                 }
2610                                 mem_cgroup_event(memcg, MEMCG_LOW);
2611                         }
2612
2613                         reclaimed = sc->nr_reclaimed;
2614                         scanned = sc->nr_scanned;
2615
2616                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2617                         node_lru_pages += lru_pages;
2618
2619                         if (memcg)
2620                                 shrink_slab(sc->gfp_mask, pgdat->node_id,
2621                                             memcg, sc->nr_scanned - scanned,
2622                                             lru_pages);
2623
2624                         /* Record the group's reclaim efficiency */
2625                         vmpressure(sc->gfp_mask, memcg, false,
2626                                    sc->nr_scanned - scanned,
2627                                    sc->nr_reclaimed - reclaimed);
2628
2629                         /*
2630                          * Direct reclaim and kswapd have to scan all memory
2631                          * cgroups to fulfill the overall scan target for the
2632                          * node.
2633                          *
2634                          * Limit reclaim, on the other hand, only cares about
2635                          * nr_to_reclaim pages to be reclaimed and it will
2636                          * retry with decreasing priority if one round over the
2637                          * whole hierarchy is not sufficient.
2638                          */
2639                         if (!global_reclaim(sc) &&
2640                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2641                                 mem_cgroup_iter_break(root, memcg);
2642                                 break;
2643                         }
2644                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2645
2646                 /*
2647                  * Shrink the slab caches in the same proportion that
2648                  * the eligible LRU pages were scanned.
2649                  */
2650                 if (global_reclaim(sc))
2651                         shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2652                                     sc->nr_scanned - nr_scanned,
2653                                     node_lru_pages);
2654
2655                 if (reclaim_state) {
2656                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2657                         reclaim_state->reclaimed_slab = 0;
2658                 }
2659
2660                 /* Record the subtree's reclaim efficiency */
2661                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2662                            sc->nr_scanned - nr_scanned,
2663                            sc->nr_reclaimed - nr_reclaimed);
2664
2665                 if (sc->nr_reclaimed - nr_reclaimed)
2666                         reclaimable = true;
2667
2668         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2669                                          sc->nr_scanned - nr_scanned, sc));
2670
2671         /*
2672          * Kswapd gives up on balancing particular nodes after too
2673          * many failures to reclaim anything from them and goes to
2674          * sleep. On reclaim progress, reset the failure counter. A
2675          * successful direct reclaim run will revive a dormant kswapd.
2676          */
2677         if (reclaimable)
2678                 pgdat->kswapd_failures = 0;
2679
2680         return reclaimable;
2681 }
2682
2683 /*
2684  * Returns true if compaction should go ahead for a costly-order request, or
2685  * the allocation would already succeed without compaction. Return false if we
2686  * should reclaim first.
2687  */
2688 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2689 {
2690         unsigned long watermark;
2691         enum compact_result suitable;
2692
2693         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2694         if (suitable == COMPACT_SUCCESS)
2695                 /* Allocation should succeed already. Don't reclaim. */
2696                 return true;
2697         if (suitable == COMPACT_SKIPPED)
2698                 /* Compaction cannot yet proceed. Do reclaim. */
2699                 return false;
2700
2701         /*
2702          * Compaction is already possible, but it takes time to run and there
2703          * are potentially other callers using the pages just freed. So proceed
2704          * with reclaim to make a buffer of free pages available to give
2705          * compaction a reasonable chance of completing and allocating the page.
2706          * Note that we won't actually reclaim the whole buffer in one attempt
2707          * as the target watermark in should_continue_reclaim() is lower. But if
2708          * we are already above the high+gap watermark, don't reclaim at all.
2709          */
2710         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2711
2712         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2713 }
2714
2715 /*
2716  * This is the direct reclaim path, for page-allocating processes.  We only
2717  * try to reclaim pages from zones which will satisfy the caller's allocation
2718  * request.
2719  *
2720  * If a zone is deemed to be full of pinned pages then just give it a light
2721  * scan then give up on it.
2722  */
2723 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2724 {
2725         struct zoneref *z;
2726         struct zone *zone;
2727         unsigned long nr_soft_reclaimed;
2728         unsigned long nr_soft_scanned;
2729         gfp_t orig_mask;
2730         pg_data_t *last_pgdat = NULL;
2731
2732         /*
2733          * If the number of buffer_heads in the machine exceeds the maximum
2734          * allowed level, force direct reclaim to scan the highmem zone as
2735          * highmem pages could be pinning lowmem pages storing buffer_heads
2736          */
2737         orig_mask = sc->gfp_mask;
2738         if (buffer_heads_over_limit) {
2739                 sc->gfp_mask |= __GFP_HIGHMEM;
2740                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2741         }
2742
2743         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2744                                         sc->reclaim_idx, sc->nodemask) {
2745                 /*
2746                  * Take care memory controller reclaiming has small influence
2747                  * to global LRU.
2748                  */
2749                 if (global_reclaim(sc)) {
2750                         if (!cpuset_zone_allowed(zone,
2751                                                  GFP_KERNEL | __GFP_HARDWALL))
2752                                 continue;
2753
2754                         /*
2755                          * If we already have plenty of memory free for
2756                          * compaction in this zone, don't free any more.
2757                          * Even though compaction is invoked for any
2758                          * non-zero order, only frequent costly order
2759                          * reclamation is disruptive enough to become a
2760                          * noticeable problem, like transparent huge
2761                          * page allocations.
2762                          */
2763                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2764                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2765                             compaction_ready(zone, sc)) {
2766                                 sc->compaction_ready = true;
2767                                 continue;
2768                         }
2769
2770                         /*
2771                          * Shrink each node in the zonelist once. If the
2772                          * zonelist is ordered by zone (not the default) then a
2773                          * node may be shrunk multiple times but in that case
2774                          * the user prefers lower zones being preserved.
2775                          */
2776                         if (zone->zone_pgdat == last_pgdat)
2777                                 continue;
2778
2779                         /*
2780                          * This steals pages from memory cgroups over softlimit
2781                          * and returns the number of reclaimed pages and
2782                          * scanned pages. This works for global memory pressure
2783                          * and balancing, not for a memcg's limit.
2784                          */
2785                         nr_soft_scanned = 0;
2786                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2787                                                 sc->order, sc->gfp_mask,
2788                                                 &nr_soft_scanned);
2789                         sc->nr_reclaimed += nr_soft_reclaimed;
2790                         sc->nr_scanned += nr_soft_scanned;
2791                         /* need some check for avoid more shrink_zone() */
2792                 }
2793
2794                 /* See comment about same check for global reclaim above */
2795                 if (zone->zone_pgdat == last_pgdat)
2796                         continue;
2797                 last_pgdat = zone->zone_pgdat;
2798                 shrink_node(zone->zone_pgdat, sc);
2799         }
2800
2801         /*
2802          * Restore to original mask to avoid the impact on the caller if we
2803          * promoted it to __GFP_HIGHMEM.
2804          */
2805         sc->gfp_mask = orig_mask;
2806 }
2807
2808 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2809 {
2810         struct mem_cgroup *memcg;
2811
2812         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2813         do {
2814                 unsigned long refaults;
2815                 struct lruvec *lruvec;
2816
2817                 if (memcg)
2818                         refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2819                 else
2820                         refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2821
2822                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2823                 lruvec->refaults = refaults;
2824         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2825 }
2826
2827 /*
2828  * This is the main entry point to direct page reclaim.
2829  *
2830  * If a full scan of the inactive list fails to free enough memory then we
2831  * are "out of memory" and something needs to be killed.
2832  *
2833  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2834  * high - the zone may be full of dirty or under-writeback pages, which this
2835  * caller can't do much about.  We kick the writeback threads and take explicit
2836  * naps in the hope that some of these pages can be written.  But if the
2837  * allocating task holds filesystem locks which prevent writeout this might not
2838  * work, and the allocation attempt will fail.
2839  *
2840  * returns:     0, if no pages reclaimed
2841  *              else, the number of pages reclaimed
2842  */
2843 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2844                                           struct scan_control *sc)
2845 {
2846         int initial_priority = sc->priority;
2847         pg_data_t *last_pgdat;
2848         struct zoneref *z;
2849         struct zone *zone;
2850 retry:
2851         delayacct_freepages_start();
2852
2853         if (global_reclaim(sc))
2854                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2855
2856         do {
2857                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2858                                 sc->priority);
2859                 sc->nr_scanned = 0;
2860                 shrink_zones(zonelist, sc);
2861
2862                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2863                         break;
2864
2865                 if (sc->compaction_ready)
2866                         break;
2867
2868                 /*
2869                  * If we're getting trouble reclaiming, start doing
2870                  * writepage even in laptop mode.
2871                  */
2872                 if (sc->priority < DEF_PRIORITY - 2)
2873                         sc->may_writepage = 1;
2874         } while (--sc->priority >= 0);
2875
2876         last_pgdat = NULL;
2877         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2878                                         sc->nodemask) {
2879                 if (zone->zone_pgdat == last_pgdat)
2880                         continue;
2881                 last_pgdat = zone->zone_pgdat;
2882                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2883         }
2884
2885         delayacct_freepages_end();
2886
2887         if (sc->nr_reclaimed)
2888                 return sc->nr_reclaimed;
2889
2890         /* Aborted reclaim to try compaction? don't OOM, then */
2891         if (sc->compaction_ready)
2892                 return 1;
2893
2894         /* Untapped cgroup reserves?  Don't OOM, retry. */
2895         if (sc->memcg_low_skipped) {
2896                 sc->priority = initial_priority;
2897                 sc->memcg_low_reclaim = 1;
2898                 sc->memcg_low_skipped = 0;
2899                 goto retry;
2900         }
2901
2902         return 0;
2903 }
2904
2905 static bool allow_direct_reclaim(pg_data_t *pgdat)
2906 {
2907         struct zone *zone;
2908         unsigned long pfmemalloc_reserve = 0;
2909         unsigned long free_pages = 0;
2910         int i;
2911         bool wmark_ok;
2912
2913         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2914                 return true;
2915
2916         for (i = 0; i <= ZONE_NORMAL; i++) {
2917                 zone = &pgdat->node_zones[i];
2918                 if (!managed_zone(zone))
2919                         continue;
2920
2921                 if (!zone_reclaimable_pages(zone))
2922                         continue;
2923
2924                 pfmemalloc_reserve += min_wmark_pages(zone);
2925                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2926         }
2927
2928         /* If there are no reserves (unexpected config) then do not throttle */
2929         if (!pfmemalloc_reserve)
2930                 return true;
2931
2932         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2933
2934         /* kswapd must be awake if processes are being throttled */
2935         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2936                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2937                                                 (enum zone_type)ZONE_NORMAL);
2938                 wake_up_interruptible(&pgdat->kswapd_wait);
2939         }
2940
2941         return wmark_ok;
2942 }
2943
2944 /*
2945  * Throttle direct reclaimers if backing storage is backed by the network
2946  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2947  * depleted. kswapd will continue to make progress and wake the processes
2948  * when the low watermark is reached.
2949  *
2950  * Returns true if a fatal signal was delivered during throttling. If this
2951  * happens, the page allocator should not consider triggering the OOM killer.
2952  */
2953 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2954                                         nodemask_t *nodemask)
2955 {
2956         struct zoneref *z;
2957         struct zone *zone;
2958         pg_data_t *pgdat = NULL;
2959
2960         /*
2961          * Kernel threads should not be throttled as they may be indirectly
2962          * responsible for cleaning pages necessary for reclaim to make forward
2963          * progress. kjournald for example may enter direct reclaim while
2964          * committing a transaction where throttling it could forcing other
2965          * processes to block on log_wait_commit().
2966          */
2967         if (current->flags & PF_KTHREAD)
2968                 goto out;
2969
2970         /*
2971          * If a fatal signal is pending, this process should not throttle.
2972          * It should return quickly so it can exit and free its memory
2973          */
2974         if (fatal_signal_pending(current))
2975                 goto out;
2976
2977         /*
2978          * Check if the pfmemalloc reserves are ok by finding the first node
2979          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2980          * GFP_KERNEL will be required for allocating network buffers when
2981          * swapping over the network so ZONE_HIGHMEM is unusable.
2982          *
2983          * Throttling is based on the first usable node and throttled processes
2984          * wait on a queue until kswapd makes progress and wakes them. There
2985          * is an affinity then between processes waking up and where reclaim
2986          * progress has been made assuming the process wakes on the same node.
2987          * More importantly, processes running on remote nodes will not compete
2988          * for remote pfmemalloc reserves and processes on different nodes
2989          * should make reasonable progress.
2990          */
2991         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2992                                         gfp_zone(gfp_mask), nodemask) {
2993                 if (zone_idx(zone) > ZONE_NORMAL)
2994                         continue;
2995
2996                 /* Throttle based on the first usable node */
2997                 pgdat = zone->zone_pgdat;
2998                 if (allow_direct_reclaim(pgdat))
2999                         goto out;
3000                 break;
3001         }
3002
3003         /* If no zone was usable by the allocation flags then do not throttle */
3004         if (!pgdat)
3005                 goto out;
3006
3007         /* Account for the throttling */
3008         count_vm_event(PGSCAN_DIRECT_THROTTLE);
3009
3010         /*
3011          * If the caller cannot enter the filesystem, it's possible that it
3012          * is due to the caller holding an FS lock or performing a journal
3013          * transaction in the case of a filesystem like ext[3|4]. In this case,
3014          * it is not safe to block on pfmemalloc_wait as kswapd could be
3015          * blocked waiting on the same lock. Instead, throttle for up to a
3016          * second before continuing.
3017          */
3018         if (!(gfp_mask & __GFP_FS)) {
3019                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3020                         allow_direct_reclaim(pgdat), HZ);
3021
3022                 goto check_pending;
3023         }
3024
3025         /* Throttle until kswapd wakes the process */
3026         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3027                 allow_direct_reclaim(pgdat));
3028
3029 check_pending:
3030         if (fatal_signal_pending(current))
3031                 return true;
3032
3033 out:
3034         return false;
3035 }
3036
3037 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3038                                 gfp_t gfp_mask, nodemask_t *nodemask)
3039 {
3040         unsigned long nr_reclaimed;
3041         struct scan_control sc = {
3042                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3043                 .gfp_mask = current_gfp_context(gfp_mask),
3044                 .reclaim_idx = gfp_zone(gfp_mask),
3045                 .order = order,
3046                 .nodemask = nodemask,
3047                 .priority = DEF_PRIORITY,
3048                 .may_writepage = !laptop_mode,
3049                 .may_unmap = 1,
3050                 .may_swap = 1,
3051         };
3052
3053         /*
3054          * Do not enter reclaim if fatal signal was delivered while throttled.
3055          * 1 is returned so that the page allocator does not OOM kill at this
3056          * point.
3057          */
3058         if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3059                 return 1;
3060
3061         trace_mm_vmscan_direct_reclaim_begin(order,
3062                                 sc.may_writepage,
3063                                 sc.gfp_mask,
3064                                 sc.reclaim_idx);
3065
3066         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3067
3068         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3069
3070         return nr_reclaimed;
3071 }
3072
3073 #ifdef CONFIG_MEMCG
3074
3075 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3076                                                 gfp_t gfp_mask, bool noswap,
3077                                                 pg_data_t *pgdat,
3078                                                 unsigned long *nr_scanned)
3079 {
3080         struct scan_control sc = {
3081                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3082                 .target_mem_cgroup = memcg,
3083                 .may_writepage = !laptop_mode,
3084                 .may_unmap = 1,
3085                 .reclaim_idx = MAX_NR_ZONES - 1,
3086                 .may_swap = !noswap,
3087         };
3088         unsigned long lru_pages;
3089
3090         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3091                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3092
3093         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3094                                                       sc.may_writepage,
3095                                                       sc.gfp_mask,
3096                                                       sc.reclaim_idx);
3097
3098         /*
3099          * NOTE: Although we can get the priority field, using it
3100          * here is not a good idea, since it limits the pages we can scan.
3101          * if we don't reclaim here, the shrink_node from balance_pgdat
3102          * will pick up pages from other mem cgroup's as well. We hack
3103          * the priority and make it zero.
3104          */
3105         shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3106
3107         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3108
3109         *nr_scanned = sc.nr_scanned;
3110         return sc.nr_reclaimed;
3111 }
3112
3113 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3114                                            unsigned long nr_pages,
3115                                            gfp_t gfp_mask,
3116                                            bool may_swap)
3117 {
3118         struct zonelist *zonelist;
3119         unsigned long nr_reclaimed;
3120         int nid;
3121         unsigned int noreclaim_flag;
3122         struct scan_control sc = {
3123                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3124                 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3125                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3126                 .reclaim_idx = MAX_NR_ZONES - 1,
3127                 .target_mem_cgroup = memcg,
3128                 .priority = DEF_PRIORITY,
3129                 .may_writepage = !laptop_mode,
3130                 .may_unmap = 1,
3131                 .may_swap = may_swap,
3132         };
3133
3134         /*
3135          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3136          * take care of from where we get pages. So the node where we start the
3137          * scan does not need to be the current node.
3138          */
3139         nid = mem_cgroup_select_victim_node(memcg);
3140
3141         zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3142
3143         trace_mm_vmscan_memcg_reclaim_begin(0,
3144                                             sc.may_writepage,
3145                                             sc.gfp_mask,
3146                                             sc.reclaim_idx);
3147
3148         noreclaim_flag = memalloc_noreclaim_save();
3149         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3150         memalloc_noreclaim_restore(noreclaim_flag);
3151
3152         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3153
3154         return nr_reclaimed;