mm, thp: copying user pages must schedule on collapse
[sfrench/cifs-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16 #include <linux/mm.h>
17 #include <linux/sched/mm.h>
18 #include <linux/module.h>
19 #include <linux/gfp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/vmpressure.h>
26 #include <linux/vmstat.h>
27 #include <linux/file.h>
28 #include <linux/writeback.h>
29 #include <linux/blkdev.h>
30 #include <linux/buffer_head.h>  /* for try_to_release_page(),
31                                         buffer_heads_over_limit */
32 #include <linux/mm_inline.h>
33 #include <linux/backing-dev.h>
34 #include <linux/rmap.h>
35 #include <linux/topology.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/compaction.h>
39 #include <linux/notifier.h>
40 #include <linux/rwsem.h>
41 #include <linux/delay.h>
42 #include <linux/kthread.h>
43 #include <linux/freezer.h>
44 #include <linux/memcontrol.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
51
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
54
55 #include <linux/swapops.h>
56 #include <linux/balloon_compaction.h>
57
58 #include "internal.h"
59
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/vmscan.h>
62
63 struct scan_control {
64         /* How many pages shrink_list() should reclaim */
65         unsigned long nr_to_reclaim;
66
67         /* This context's GFP mask */
68         gfp_t gfp_mask;
69
70         /* Allocation order */
71         int order;
72
73         /*
74          * Nodemask of nodes allowed by the caller. If NULL, all nodes
75          * are scanned.
76          */
77         nodemask_t      *nodemask;
78
79         /*
80          * The memory cgroup that hit its limit and as a result is the
81          * primary target of this reclaim invocation.
82          */
83         struct mem_cgroup *target_mem_cgroup;
84
85         /* Scan (total_size >> priority) pages at once */
86         int priority;
87
88         /* The highest zone to isolate pages for reclaim from */
89         enum zone_type reclaim_idx;
90
91         /* Writepage batching in laptop mode; RECLAIM_WRITE */
92         unsigned int may_writepage:1;
93
94         /* Can mapped pages be reclaimed? */
95         unsigned int may_unmap:1;
96
97         /* Can pages be swapped as part of reclaim? */
98         unsigned int may_swap:1;
99
100         /*
101          * Cgroups are not reclaimed below their configured memory.low,
102          * unless we threaten to OOM. If any cgroups are skipped due to
103          * memory.low and nothing was reclaimed, go back for memory.low.
104          */
105         unsigned int memcg_low_reclaim:1;
106         unsigned int memcg_low_skipped:1;
107
108         unsigned int hibernation_mode:1;
109
110         /* One of the zones is ready for compaction */
111         unsigned int compaction_ready:1;
112
113         /* Incremented by the number of inactive pages that were scanned */
114         unsigned long nr_scanned;
115
116         /* Number of pages freed so far during a call to shrink_zones() */
117         unsigned long nr_reclaimed;
118 };
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 /*
153  * The total number of pages which are beyond the high watermark within all
154  * zones.
155  */
156 unsigned long vm_total_pages;
157
158 static LIST_HEAD(shrinker_list);
159 static DECLARE_RWSEM(shrinker_rwsem);
160
161 #ifdef CONFIG_MEMCG
162 static bool global_reclaim(struct scan_control *sc)
163 {
164         return !sc->target_mem_cgroup;
165 }
166
167 /**
168  * sane_reclaim - is the usual dirty throttling mechanism operational?
169  * @sc: scan_control in question
170  *
171  * The normal page dirty throttling mechanism in balance_dirty_pages() is
172  * completely broken with the legacy memcg and direct stalling in
173  * shrink_page_list() is used for throttling instead, which lacks all the
174  * niceties such as fairness, adaptive pausing, bandwidth proportional
175  * allocation and configurability.
176  *
177  * This function tests whether the vmscan currently in progress can assume
178  * that the normal dirty throttling mechanism is operational.
179  */
180 static bool sane_reclaim(struct scan_control *sc)
181 {
182         struct mem_cgroup *memcg = sc->target_mem_cgroup;
183
184         if (!memcg)
185                 return true;
186 #ifdef CONFIG_CGROUP_WRITEBACK
187         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
188                 return true;
189 #endif
190         return false;
191 }
192 #else
193 static bool global_reclaim(struct scan_control *sc)
194 {
195         return true;
196 }
197
198 static bool sane_reclaim(struct scan_control *sc)
199 {
200         return true;
201 }
202 #endif
203
204 /*
205  * This misses isolated pages which are not accounted for to save counters.
206  * As the data only determines if reclaim or compaction continues, it is
207  * not expected that isolated pages will be a dominating factor.
208  */
209 unsigned long zone_reclaimable_pages(struct zone *zone)
210 {
211         unsigned long nr;
212
213         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
214                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
215         if (get_nr_swap_pages() > 0)
216                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
217                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
218
219         return nr;
220 }
221
222 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
223 {
224         unsigned long nr;
225
226         nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
227              node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
228              node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
229
230         if (get_nr_swap_pages() > 0)
231                 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
232                       node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
233                       node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
234
235         return nr;
236 }
237
238 /**
239  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
240  * @lruvec: lru vector
241  * @lru: lru to use
242  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
243  */
244 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
245 {
246         unsigned long lru_size;
247         int zid;
248
249         if (!mem_cgroup_disabled())
250                 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
251         else
252                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
253
254         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
255                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
256                 unsigned long size;
257
258                 if (!managed_zone(zone))
259                         continue;
260
261                 if (!mem_cgroup_disabled())
262                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
263                 else
264                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
265                                        NR_ZONE_LRU_BASE + lru);
266                 lru_size -= min(size, lru_size);
267         }
268
269         return lru_size;
270
271 }
272
273 /*
274  * Add a shrinker callback to be called from the vm.
275  */
276 int register_shrinker(struct shrinker *shrinker)
277 {
278         size_t size = sizeof(*shrinker->nr_deferred);
279
280         if (shrinker->flags & SHRINKER_NUMA_AWARE)
281                 size *= nr_node_ids;
282
283         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
284         if (!shrinker->nr_deferred)
285                 return -ENOMEM;
286
287         down_write(&shrinker_rwsem);
288         list_add_tail(&shrinker->list, &shrinker_list);
289         up_write(&shrinker_rwsem);
290         return 0;
291 }
292 EXPORT_SYMBOL(register_shrinker);
293
294 /*
295  * Remove one
296  */
297 void unregister_shrinker(struct shrinker *shrinker)
298 {
299         down_write(&shrinker_rwsem);
300         list_del(&shrinker->list);
301         up_write(&shrinker_rwsem);
302         kfree(shrinker->nr_deferred);
303 }
304 EXPORT_SYMBOL(unregister_shrinker);
305
306 #define SHRINK_BATCH 128
307
308 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
309                                     struct shrinker *shrinker,
310                                     unsigned long nr_scanned,
311                                     unsigned long nr_eligible)
312 {
313         unsigned long freed = 0;
314         unsigned long long delta;
315         long total_scan;
316         long freeable;
317         long nr;
318         long new_nr;
319         int nid = shrinkctl->nid;
320         long batch_size = shrinker->batch ? shrinker->batch
321                                           : SHRINK_BATCH;
322         long scanned = 0, next_deferred;
323
324         freeable = shrinker->count_objects(shrinker, shrinkctl);
325         if (freeable == 0)
326                 return 0;
327
328         /*
329          * copy the current shrinker scan count into a local variable
330          * and zero it so that other concurrent shrinker invocations
331          * don't also do this scanning work.
332          */
333         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
334
335         total_scan = nr;
336         delta = (4 * nr_scanned) / shrinker->seeks;
337         delta *= freeable;
338         do_div(delta, nr_eligible + 1);
339         total_scan += delta;
340         if (total_scan < 0) {
341                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342                        shrinker->scan_objects, total_scan);
343                 total_scan = freeable;
344                 next_deferred = nr;
345         } else
346                 next_deferred = total_scan;
347
348         /*
349          * We need to avoid excessive windup on filesystem shrinkers
350          * due to large numbers of GFP_NOFS allocations causing the
351          * shrinkers to return -1 all the time. This results in a large
352          * nr being built up so when a shrink that can do some work
353          * comes along it empties the entire cache due to nr >>>
354          * freeable. This is bad for sustaining a working set in
355          * memory.
356          *
357          * Hence only allow the shrinker to scan the entire cache when
358          * a large delta change is calculated directly.
359          */
360         if (delta < freeable / 4)
361                 total_scan = min(total_scan, freeable / 2);
362
363         /*
364          * Avoid risking looping forever due to too large nr value:
365          * never try to free more than twice the estimate number of
366          * freeable entries.
367          */
368         if (total_scan > freeable * 2)
369                 total_scan = freeable * 2;
370
371         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
372                                    nr_scanned, nr_eligible,
373                                    freeable, delta, total_scan);
374
375         /*
376          * Normally, we should not scan less than batch_size objects in one
377          * pass to avoid too frequent shrinker calls, but if the slab has less
378          * than batch_size objects in total and we are really tight on memory,
379          * we will try to reclaim all available objects, otherwise we can end
380          * up failing allocations although there are plenty of reclaimable
381          * objects spread over several slabs with usage less than the
382          * batch_size.
383          *
384          * We detect the "tight on memory" situations by looking at the total
385          * number of objects we want to scan (total_scan). If it is greater
386          * than the total number of objects on slab (freeable), we must be
387          * scanning at high prio and therefore should try to reclaim as much as
388          * possible.
389          */
390         while (total_scan >= batch_size ||
391                total_scan >= freeable) {
392                 unsigned long ret;
393                 unsigned long nr_to_scan = min(batch_size, total_scan);
394
395                 shrinkctl->nr_to_scan = nr_to_scan;
396                 ret = shrinker->scan_objects(shrinker, shrinkctl);
397                 if (ret == SHRINK_STOP)
398                         break;
399                 freed += ret;
400
401                 count_vm_events(SLABS_SCANNED, nr_to_scan);
402                 total_scan -= nr_to_scan;
403                 scanned += nr_to_scan;
404
405                 cond_resched();
406         }
407
408         if (next_deferred >= scanned)
409                 next_deferred -= scanned;
410         else
411                 next_deferred = 0;
412         /*
413          * move the unused scan count back into the shrinker in a
414          * manner that handles concurrent updates. If we exhausted the
415          * scan, there is no need to do an update.
416          */
417         if (next_deferred > 0)
418                 new_nr = atomic_long_add_return(next_deferred,
419                                                 &shrinker->nr_deferred[nid]);
420         else
421                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
422
423         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
424         return freed;
425 }
426
427 /**
428  * shrink_slab - shrink slab caches
429  * @gfp_mask: allocation context
430  * @nid: node whose slab caches to target
431  * @memcg: memory cgroup whose slab caches to target
432  * @nr_scanned: pressure numerator
433  * @nr_eligible: pressure denominator
434  *
435  * Call the shrink functions to age shrinkable caches.
436  *
437  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
438  * unaware shrinkers will receive a node id of 0 instead.
439  *
440  * @memcg specifies the memory cgroup to target. If it is not NULL,
441  * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
442  * objects from the memory cgroup specified. Otherwise, only unaware
443  * shrinkers are called.
444  *
445  * @nr_scanned and @nr_eligible form a ratio that indicate how much of
446  * the available objects should be scanned.  Page reclaim for example
447  * passes the number of pages scanned and the number of pages on the
448  * LRU lists that it considered on @nid, plus a bias in @nr_scanned
449  * when it encountered mapped pages.  The ratio is further biased by
450  * the ->seeks setting of the shrink function, which indicates the
451  * cost to recreate an object relative to that of an LRU page.
452  *
453  * Returns the number of reclaimed slab objects.
454  */
455 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
456                                  struct mem_cgroup *memcg,
457                                  unsigned long nr_scanned,
458                                  unsigned long nr_eligible)
459 {
460         struct shrinker *shrinker;
461         unsigned long freed = 0;
462
463         if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
464                 return 0;
465
466         if (nr_scanned == 0)
467                 nr_scanned = SWAP_CLUSTER_MAX;
468
469         if (!down_read_trylock(&shrinker_rwsem)) {
470                 /*
471                  * If we would return 0, our callers would understand that we
472                  * have nothing else to shrink and give up trying. By returning
473                  * 1 we keep it going and assume we'll be able to shrink next
474                  * time.
475                  */
476                 freed = 1;
477                 goto out;
478         }
479
480         list_for_each_entry(shrinker, &shrinker_list, list) {
481                 struct shrink_control sc = {
482                         .gfp_mask = gfp_mask,
483                         .nid = nid,
484                         .memcg = memcg,
485                 };
486
487                 /*
488                  * If kernel memory accounting is disabled, we ignore
489                  * SHRINKER_MEMCG_AWARE flag and call all shrinkers
490                  * passing NULL for memcg.
491                  */
492                 if (memcg_kmem_enabled() &&
493                     !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
494                         continue;
495
496                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
497                         sc.nid = 0;
498
499                 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
500         }
501
502         up_read(&shrinker_rwsem);
503 out:
504         cond_resched();
505         return freed;
506 }
507
508 void drop_slab_node(int nid)
509 {
510         unsigned long freed;
511
512         do {
513                 struct mem_cgroup *memcg = NULL;
514
515                 freed = 0;
516                 do {
517                         freed += shrink_slab(GFP_KERNEL, nid, memcg,
518                                              1000, 1000);
519                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
520         } while (freed > 10);
521 }
522
523 void drop_slab(void)
524 {
525         int nid;
526
527         for_each_online_node(nid)
528                 drop_slab_node(nid);
529 }
530
531 static inline int is_page_cache_freeable(struct page *page)
532 {
533         /*
534          * A freeable page cache page is referenced only by the caller
535          * that isolated the page, the page cache radix tree and
536          * optional buffer heads at page->private.
537          */
538         return page_count(page) - page_has_private(page) == 2;
539 }
540
541 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
542 {
543         if (current->flags & PF_SWAPWRITE)
544                 return 1;
545         if (!inode_write_congested(inode))
546                 return 1;
547         if (inode_to_bdi(inode) == current->backing_dev_info)
548                 return 1;
549         return 0;
550 }
551
552 /*
553  * We detected a synchronous write error writing a page out.  Probably
554  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
555  * fsync(), msync() or close().
556  *
557  * The tricky part is that after writepage we cannot touch the mapping: nothing
558  * prevents it from being freed up.  But we have a ref on the page and once
559  * that page is locked, the mapping is pinned.
560  *
561  * We're allowed to run sleeping lock_page() here because we know the caller has
562  * __GFP_FS.
563  */
564 static void handle_write_error(struct address_space *mapping,
565                                 struct page *page, int error)
566 {
567         lock_page(page);
568         if (page_mapping(page) == mapping)
569                 mapping_set_error(mapping, error);
570         unlock_page(page);
571 }
572
573 /* possible outcome of pageout() */
574 typedef enum {
575         /* failed to write page out, page is locked */
576         PAGE_KEEP,
577         /* move page to the active list, page is locked */
578         PAGE_ACTIVATE,
579         /* page has been sent to the disk successfully, page is unlocked */
580         PAGE_SUCCESS,
581         /* page is clean and locked */
582         PAGE_CLEAN,
583 } pageout_t;
584
585 /*
586  * pageout is called by shrink_page_list() for each dirty page.
587  * Calls ->writepage().
588  */
589 static pageout_t pageout(struct page *page, struct address_space *mapping,
590                          struct scan_control *sc)
591 {
592         /*
593          * If the page is dirty, only perform writeback if that write
594          * will be non-blocking.  To prevent this allocation from being
595          * stalled by pagecache activity.  But note that there may be
596          * stalls if we need to run get_block().  We could test
597          * PagePrivate for that.
598          *
599          * If this process is currently in __generic_file_write_iter() against
600          * this page's queue, we can perform writeback even if that
601          * will block.
602          *
603          * If the page is swapcache, write it back even if that would
604          * block, for some throttling. This happens by accident, because
605          * swap_backing_dev_info is bust: it doesn't reflect the
606          * congestion state of the swapdevs.  Easy to fix, if needed.
607          */
608         if (!is_page_cache_freeable(page))
609                 return PAGE_KEEP;
610         if (!mapping) {
611                 /*
612                  * Some data journaling orphaned pages can have
613                  * page->mapping == NULL while being dirty with clean buffers.
614                  */
615                 if (page_has_private(page)) {
616                         if (try_to_free_buffers(page)) {
617                                 ClearPageDirty(page);
618                                 pr_info("%s: orphaned page\n", __func__);
619                                 return PAGE_CLEAN;
620                         }
621                 }
622                 return PAGE_KEEP;
623         }
624         if (mapping->a_ops->writepage == NULL)
625                 return PAGE_ACTIVATE;
626         if (!may_write_to_inode(mapping->host, sc))
627                 return PAGE_KEEP;
628
629         if (clear_page_dirty_for_io(page)) {
630                 int res;
631                 struct writeback_control wbc = {
632                         .sync_mode = WB_SYNC_NONE,
633                         .nr_to_write = SWAP_CLUSTER_MAX,
634                         .range_start = 0,
635                         .range_end = LLONG_MAX,
636                         .for_reclaim = 1,
637                 };
638
639                 SetPageReclaim(page);
640                 res = mapping->a_ops->writepage(page, &wbc);
641                 if (res < 0)
642                         handle_write_error(mapping, page, res);
643                 if (res == AOP_WRITEPAGE_ACTIVATE) {
644                         ClearPageReclaim(page);
645                         return PAGE_ACTIVATE;
646                 }
647
648                 if (!PageWriteback(page)) {
649                         /* synchronous write or broken a_ops? */
650                         ClearPageReclaim(page);
651                 }
652                 trace_mm_vmscan_writepage(page);
653                 inc_node_page_state(page, NR_VMSCAN_WRITE);
654                 return PAGE_SUCCESS;
655         }
656
657         return PAGE_CLEAN;
658 }
659
660 /*
661  * Same as remove_mapping, but if the page is removed from the mapping, it
662  * gets returned with a refcount of 0.
663  */
664 static int __remove_mapping(struct address_space *mapping, struct page *page,
665                             bool reclaimed)
666 {
667         unsigned long flags;
668
669         BUG_ON(!PageLocked(page));
670         BUG_ON(mapping != page_mapping(page));
671
672         spin_lock_irqsave(&mapping->tree_lock, flags);
673         /*
674          * The non racy check for a busy page.
675          *
676          * Must be careful with the order of the tests. When someone has
677          * a ref to the page, it may be possible that they dirty it then
678          * drop the reference. So if PageDirty is tested before page_count
679          * here, then the following race may occur:
680          *
681          * get_user_pages(&page);
682          * [user mapping goes away]
683          * write_to(page);
684          *                              !PageDirty(page)    [good]
685          * SetPageDirty(page);
686          * put_page(page);
687          *                              !page_count(page)   [good, discard it]
688          *
689          * [oops, our write_to data is lost]
690          *
691          * Reversing the order of the tests ensures such a situation cannot
692          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
693          * load is not satisfied before that of page->_refcount.
694          *
695          * Note that if SetPageDirty is always performed via set_page_dirty,
696          * and thus under tree_lock, then this ordering is not required.
697          */
698         if (!page_ref_freeze(page, 2))
699                 goto cannot_free;
700         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
701         if (unlikely(PageDirty(page))) {
702                 page_ref_unfreeze(page, 2);
703                 goto cannot_free;
704         }
705
706         if (PageSwapCache(page)) {
707                 swp_entry_t swap = { .val = page_private(page) };
708                 mem_cgroup_swapout(page, swap);
709                 __delete_from_swap_cache(page);
710                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
711                 swapcache_free(swap);
712         } else {
713                 void (*freepage)(struct page *);
714                 void *shadow = NULL;
715
716                 freepage = mapping->a_ops->freepage;
717                 /*
718                  * Remember a shadow entry for reclaimed file cache in
719                  * order to detect refaults, thus thrashing, later on.
720                  *
721                  * But don't store shadows in an address space that is
722                  * already exiting.  This is not just an optizimation,
723                  * inode reclaim needs to empty out the radix tree or
724                  * the nodes are lost.  Don't plant shadows behind its
725                  * back.
726                  *
727                  * We also don't store shadows for DAX mappings because the
728                  * only page cache pages found in these are zero pages
729                  * covering holes, and because we don't want to mix DAX
730                  * exceptional entries and shadow exceptional entries in the
731                  * same page_tree.
732                  */
733                 if (reclaimed && page_is_file_cache(page) &&
734                     !mapping_exiting(mapping) && !dax_mapping(mapping))
735                         shadow = workingset_eviction(mapping, page);
736                 __delete_from_page_cache(page, shadow);
737                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
738
739                 if (freepage != NULL)
740                         freepage(page);
741         }
742
743         return 1;
744
745 cannot_free:
746         spin_unlock_irqrestore(&mapping->tree_lock, flags);
747         return 0;
748 }
749
750 /*
751  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
752  * someone else has a ref on the page, abort and return 0.  If it was
753  * successfully detached, return 1.  Assumes the caller has a single ref on
754  * this page.
755  */
756 int remove_mapping(struct address_space *mapping, struct page *page)
757 {
758         if (__remove_mapping(mapping, page, false)) {
759                 /*
760                  * Unfreezing the refcount with 1 rather than 2 effectively
761                  * drops the pagecache ref for us without requiring another
762                  * atomic operation.
763                  */
764                 page_ref_unfreeze(page, 1);
765                 return 1;
766         }
767         return 0;
768 }
769
770 /**
771  * putback_lru_page - put previously isolated page onto appropriate LRU list
772  * @page: page to be put back to appropriate lru list
773  *
774  * Add previously isolated @page to appropriate LRU list.
775  * Page may still be unevictable for other reasons.
776  *
777  * lru_lock must not be held, interrupts must be enabled.
778  */
779 void putback_lru_page(struct page *page)
780 {
781         bool is_unevictable;
782         int was_unevictable = PageUnevictable(page);
783
784         VM_BUG_ON_PAGE(PageLRU(page), page);
785
786 redo:
787         ClearPageUnevictable(page);
788
789         if (page_evictable(page)) {
790                 /*
791                  * For evictable pages, we can use the cache.
792                  * In event of a race, worst case is we end up with an
793                  * unevictable page on [in]active list.
794                  * We know how to handle that.
795                  */
796                 is_unevictable = false;
797                 lru_cache_add(page);
798         } else {
799                 /*
800                  * Put unevictable pages directly on zone's unevictable
801                  * list.
802                  */
803                 is_unevictable = true;
804                 add_page_to_unevictable_list(page);
805                 /*
806                  * When racing with an mlock or AS_UNEVICTABLE clearing
807                  * (page is unlocked) make sure that if the other thread
808                  * does not observe our setting of PG_lru and fails
809                  * isolation/check_move_unevictable_pages,
810                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
811                  * the page back to the evictable list.
812                  *
813                  * The other side is TestClearPageMlocked() or shmem_lock().
814                  */
815                 smp_mb();
816         }
817
818         /*
819          * page's status can change while we move it among lru. If an evictable
820          * page is on unevictable list, it never be freed. To avoid that,
821          * check after we added it to the list, again.
822          */
823         if (is_unevictable && page_evictable(page)) {
824                 if (!isolate_lru_page(page)) {
825                         put_page(page);
826                         goto redo;
827                 }
828                 /* This means someone else dropped this page from LRU
829                  * So, it will be freed or putback to LRU again. There is
830                  * nothing to do here.
831                  */
832         }
833
834         if (was_unevictable && !is_unevictable)
835                 count_vm_event(UNEVICTABLE_PGRESCUED);
836         else if (!was_unevictable && is_unevictable)
837                 count_vm_event(UNEVICTABLE_PGCULLED);
838
839         put_page(page);         /* drop ref from isolate */
840 }
841
842 enum page_references {
843         PAGEREF_RECLAIM,
844         PAGEREF_RECLAIM_CLEAN,
845         PAGEREF_KEEP,
846         PAGEREF_ACTIVATE,
847 };
848
849 static enum page_references page_check_references(struct page *page,
850                                                   struct scan_control *sc)
851 {
852         int referenced_ptes, referenced_page;
853         unsigned long vm_flags;
854
855         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
856                                           &vm_flags);
857         referenced_page = TestClearPageReferenced(page);
858
859         /*
860          * Mlock lost the isolation race with us.  Let try_to_unmap()
861          * move the page to the unevictable list.
862          */
863         if (vm_flags & VM_LOCKED)
864                 return PAGEREF_RECLAIM;
865
866         if (referenced_ptes) {
867                 if (PageSwapBacked(page))
868                         return PAGEREF_ACTIVATE;
869                 /*
870                  * All mapped pages start out with page table
871                  * references from the instantiating fault, so we need
872                  * to look twice if a mapped file page is used more
873                  * than once.
874                  *
875                  * Mark it and spare it for another trip around the
876                  * inactive list.  Another page table reference will
877                  * lead to its activation.
878                  *
879                  * Note: the mark is set for activated pages as well
880                  * so that recently deactivated but used pages are
881                  * quickly recovered.
882                  */
883                 SetPageReferenced(page);
884
885                 if (referenced_page || referenced_ptes > 1)
886                         return PAGEREF_ACTIVATE;
887
888                 /*
889                  * Activate file-backed executable pages after first usage.
890                  */
891                 if (vm_flags & VM_EXEC)
892                         return PAGEREF_ACTIVATE;
893
894                 return PAGEREF_KEEP;
895         }
896
897         /* Reclaim if clean, defer dirty pages to writeback */
898         if (referenced_page && !PageSwapBacked(page))
899                 return PAGEREF_RECLAIM_CLEAN;
900
901         return PAGEREF_RECLAIM;
902 }
903
904 /* Check if a page is dirty or under writeback */
905 static void page_check_dirty_writeback(struct page *page,
906                                        bool *dirty, bool *writeback)
907 {
908         struct address_space *mapping;
909
910         /*
911          * Anonymous pages are not handled by flushers and must be written
912          * from reclaim context. Do not stall reclaim based on them
913          */
914         if (!page_is_file_cache(page) ||
915             (PageAnon(page) && !PageSwapBacked(page))) {
916                 *dirty = false;
917                 *writeback = false;
918                 return;
919         }
920
921         /* By default assume that the page flags are accurate */
922         *dirty = PageDirty(page);
923         *writeback = PageWriteback(page);
924
925         /* Verify dirty/writeback state if the filesystem supports it */
926         if (!page_has_private(page))
927                 return;
928
929         mapping = page_mapping(page);
930         if (mapping && mapping->a_ops->is_dirty_writeback)
931                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
932 }
933
934 struct reclaim_stat {
935         unsigned nr_dirty;
936         unsigned nr_unqueued_dirty;
937         unsigned nr_congested;
938         unsigned nr_writeback;
939         unsigned nr_immediate;
940         unsigned nr_activate;
941         unsigned nr_ref_keep;
942         unsigned nr_unmap_fail;
943 };
944
945 /*
946  * shrink_page_list() returns the number of reclaimed pages
947  */
948 static unsigned long shrink_page_list(struct list_head *page_list,
949                                       struct pglist_data *pgdat,
950                                       struct scan_control *sc,
951                                       enum ttu_flags ttu_flags,
952                                       struct reclaim_stat *stat,
953                                       bool force_reclaim)
954 {
955         LIST_HEAD(ret_pages);
956         LIST_HEAD(free_pages);
957         int pgactivate = 0;
958         unsigned nr_unqueued_dirty = 0;
959         unsigned nr_dirty = 0;
960         unsigned nr_congested = 0;
961         unsigned nr_reclaimed = 0;
962         unsigned nr_writeback = 0;
963         unsigned nr_immediate = 0;
964         unsigned nr_ref_keep = 0;
965         unsigned nr_unmap_fail = 0;
966
967         cond_resched();
968
969         while (!list_empty(page_list)) {
970                 struct address_space *mapping;
971                 struct page *page;
972                 int may_enter_fs;
973                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
974                 bool dirty, writeback;
975
976                 cond_resched();
977
978                 page = lru_to_page(page_list);
979                 list_del(&page->lru);
980
981                 if (!trylock_page(page))
982                         goto keep;
983
984                 VM_BUG_ON_PAGE(PageActive(page), page);
985
986                 sc->nr_scanned++;
987
988                 if (unlikely(!page_evictable(page)))
989                         goto activate_locked;
990
991                 if (!sc->may_unmap && page_mapped(page))
992                         goto keep_locked;
993
994                 /* Double the slab pressure for mapped and swapcache pages */
995                 if ((page_mapped(page) || PageSwapCache(page)) &&
996                     !(PageAnon(page) && !PageSwapBacked(page)))
997                         sc->nr_scanned++;
998
999                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1000                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1001
1002                 /*
1003                  * The number of dirty pages determines if a zone is marked
1004                  * reclaim_congested which affects wait_iff_congested. kswapd
1005                  * will stall and start writing pages if the tail of the LRU
1006                  * is all dirty unqueued pages.
1007                  */
1008                 page_check_dirty_writeback(page, &dirty, &writeback);
1009                 if (dirty || writeback)
1010                         nr_dirty++;
1011
1012                 if (dirty && !writeback)
1013                         nr_unqueued_dirty++;
1014
1015                 /*
1016                  * Treat this page as congested if the underlying BDI is or if
1017                  * pages are cycling through the LRU so quickly that the
1018                  * pages marked for immediate reclaim are making it to the
1019                  * end of the LRU a second time.
1020                  */
1021                 mapping = page_mapping(page);
1022                 if (((dirty || writeback) && mapping &&
1023                      inode_write_congested(mapping->host)) ||
1024                     (writeback && PageReclaim(page)))
1025                         nr_congested++;
1026
1027                 /*
1028                  * If a page at the tail of the LRU is under writeback, there
1029                  * are three cases to consider.
1030                  *
1031                  * 1) If reclaim is encountering an excessive number of pages
1032                  *    under writeback and this page is both under writeback and
1033                  *    PageReclaim then it indicates that pages are being queued
1034                  *    for IO but are being recycled through the LRU before the
1035                  *    IO can complete. Waiting on the page itself risks an
1036                  *    indefinite stall if it is impossible to writeback the
1037                  *    page due to IO error or disconnected storage so instead
1038                  *    note that the LRU is being scanned too quickly and the
1039                  *    caller can stall after page list has been processed.
1040                  *
1041                  * 2) Global or new memcg reclaim encounters a page that is
1042                  *    not marked for immediate reclaim, or the caller does not
1043                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1044                  *    not to fs). In this case mark the page for immediate
1045                  *    reclaim and continue scanning.
1046                  *
1047                  *    Require may_enter_fs because we would wait on fs, which
1048                  *    may not have submitted IO yet. And the loop driver might
1049                  *    enter reclaim, and deadlock if it waits on a page for
1050                  *    which it is needed to do the write (loop masks off
1051                  *    __GFP_IO|__GFP_FS for this reason); but more thought
1052                  *    would probably show more reasons.
1053                  *
1054                  * 3) Legacy memcg encounters a page that is already marked
1055                  *    PageReclaim. memcg does not have any dirty pages
1056                  *    throttling so we could easily OOM just because too many
1057                  *    pages are in writeback and there is nothing else to
1058                  *    reclaim. Wait for the writeback to complete.
1059                  *
1060                  * In cases 1) and 2) we activate the pages to get them out of
1061                  * the way while we continue scanning for clean pages on the
1062                  * inactive list and refilling from the active list. The
1063                  * observation here is that waiting for disk writes is more
1064                  * expensive than potentially causing reloads down the line.
1065                  * Since they're marked for immediate reclaim, they won't put
1066                  * memory pressure on the cache working set any longer than it
1067                  * takes to write them to disk.
1068                  */
1069                 if (PageWriteback(page)) {
1070                         /* Case 1 above */
1071                         if (current_is_kswapd() &&
1072                             PageReclaim(page) &&
1073                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1074                                 nr_immediate++;
1075                                 goto activate_locked;
1076
1077                         /* Case 2 above */
1078                         } else if (sane_reclaim(sc) ||
1079                             !PageReclaim(page) || !may_enter_fs) {
1080                                 /*
1081                                  * This is slightly racy - end_page_writeback()
1082                                  * might have just cleared PageReclaim, then
1083                                  * setting PageReclaim here end up interpreted
1084                                  * as PageReadahead - but that does not matter
1085                                  * enough to care.  What we do want is for this
1086                                  * page to have PageReclaim set next time memcg
1087                                  * reclaim reaches the tests above, so it will
1088                                  * then wait_on_page_writeback() to avoid OOM;
1089                                  * and it's also appropriate in global reclaim.
1090                                  */
1091                                 SetPageReclaim(page);
1092                                 nr_writeback++;
1093                                 goto activate_locked;
1094
1095                         /* Case 3 above */
1096                         } else {
1097                                 unlock_page(page);
1098                                 wait_on_page_writeback(page);
1099                                 /* then go back and try same page again */
1100                                 list_add_tail(&page->lru, page_list);
1101                                 continue;
1102                         }
1103                 }
1104
1105                 if (!force_reclaim)
1106                         references = page_check_references(page, sc);
1107
1108                 switch (references) {
1109                 case PAGEREF_ACTIVATE:
1110                         goto activate_locked;
1111                 case PAGEREF_KEEP:
1112                         nr_ref_keep++;
1113                         goto keep_locked;
1114                 case PAGEREF_RECLAIM:
1115                 case PAGEREF_RECLAIM_CLEAN:
1116                         ; /* try to reclaim the page below */
1117                 }
1118
1119                 /*
1120                  * Anonymous process memory has backing store?
1121                  * Try to allocate it some swap space here.
1122                  * Lazyfree page could be freed directly
1123                  */
1124                 if (PageAnon(page) && PageSwapBacked(page) &&
1125                     !PageSwapCache(page)) {
1126                         if (!(sc->gfp_mask & __GFP_IO))
1127                                 goto keep_locked;
1128                         if (!add_to_swap(page, page_list))
1129                                 goto activate_locked;
1130                         may_enter_fs = 1;
1131
1132                         /* Adding to swap updated mapping */
1133                         mapping = page_mapping(page);
1134                 } else if (unlikely(PageTransHuge(page))) {
1135                         /* Split file THP */
1136                         if (split_huge_page_to_list(page, page_list))
1137                                 goto keep_locked;
1138                 }
1139
1140                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1141
1142                 /*
1143                  * The page is mapped into the page tables of one or more
1144                  * processes. Try to unmap it here.
1145                  */
1146                 if (page_mapped(page)) {
1147                         if (!try_to_unmap(page, ttu_flags | TTU_BATCH_FLUSH)) {
1148                                 nr_unmap_fail++;
1149                                 goto activate_locked;
1150                         }
1151                 }
1152
1153                 if (PageDirty(page)) {
1154                         /*
1155                          * Only kswapd can writeback filesystem pages
1156                          * to avoid risk of stack overflow. But avoid
1157                          * injecting inefficient single-page IO into
1158                          * flusher writeback as much as possible: only
1159                          * write pages when we've encountered many
1160                          * dirty pages, and when we've already scanned
1161                          * the rest of the LRU for clean pages and see
1162                          * the same dirty pages again (PageReclaim).
1163                          */
1164                         if (page_is_file_cache(page) &&
1165                             (!current_is_kswapd() || !PageReclaim(page) ||
1166                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1167                                 /*
1168                                  * Immediately reclaim when written back.
1169                                  * Similar in principal to deactivate_page()
1170                                  * except we already have the page isolated
1171                                  * and know it's dirty
1172                                  */
1173                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1174                                 SetPageReclaim(page);
1175
1176                                 goto activate_locked;
1177                         }
1178
1179                         if (references == PAGEREF_RECLAIM_CLEAN)
1180                                 goto keep_locked;
1181                         if (!may_enter_fs)
1182                                 goto keep_locked;
1183                         if (!sc->may_writepage)
1184                                 goto keep_locked;
1185
1186                         /*
1187                          * Page is dirty. Flush the TLB if a writable entry
1188                          * potentially exists to avoid CPU writes after IO
1189                          * starts and then write it out here.
1190                          */
1191                         try_to_unmap_flush_dirty();
1192                         switch (pageout(page, mapping, sc)) {
1193                         case PAGE_KEEP:
1194                                 goto keep_locked;
1195                         case PAGE_ACTIVATE:
1196                                 goto activate_locked;
1197                         case PAGE_SUCCESS:
1198                                 if (PageWriteback(page))
1199                                         goto keep;
1200                                 if (PageDirty(page))
1201                                         goto keep;
1202
1203                                 /*
1204                                  * A synchronous write - probably a ramdisk.  Go
1205                                  * ahead and try to reclaim the page.
1206                                  */
1207                                 if (!trylock_page(page))
1208                                         goto keep;
1209                                 if (PageDirty(page) || PageWriteback(page))
1210                                         goto keep_locked;
1211                                 mapping = page_mapping(page);
1212                         case PAGE_CLEAN:
1213                                 ; /* try to free the page below */
1214                         }
1215                 }
1216
1217                 /*
1218                  * If the page has buffers, try to free the buffer mappings
1219                  * associated with this page. If we succeed we try to free
1220                  * the page as well.
1221                  *
1222                  * We do this even if the page is PageDirty().
1223                  * try_to_release_page() does not perform I/O, but it is
1224                  * possible for a page to have PageDirty set, but it is actually
1225                  * clean (all its buffers are clean).  This happens if the
1226                  * buffers were written out directly, with submit_bh(). ext3
1227                  * will do this, as well as the blockdev mapping.
1228                  * try_to_release_page() will discover that cleanness and will
1229                  * drop the buffers and mark the page clean - it can be freed.
1230                  *
1231                  * Rarely, pages can have buffers and no ->mapping.  These are
1232                  * the pages which were not successfully invalidated in
1233                  * truncate_complete_page().  We try to drop those buffers here
1234                  * and if that worked, and the page is no longer mapped into
1235                  * process address space (page_count == 1) it can be freed.
1236                  * Otherwise, leave the page on the LRU so it is swappable.
1237                  */
1238                 if (page_has_private(page)) {
1239                         if (!try_to_release_page(page, sc->gfp_mask))
1240                                 goto activate_locked;
1241                         if (!mapping && page_count(page) == 1) {
1242                                 unlock_page(page);
1243                                 if (put_page_testzero(page))
1244                                         goto free_it;
1245                                 else {
1246                                         /*
1247                                          * rare race with speculative reference.
1248                                          * the speculative reference will free
1249                                          * this page shortly, so we may
1250                                          * increment nr_reclaimed here (and
1251                                          * leave it off the LRU).
1252                                          */
1253                                         nr_reclaimed++;
1254                                         continue;
1255                                 }
1256                         }
1257                 }
1258
1259                 if (PageAnon(page) && !PageSwapBacked(page)) {
1260                         /* follow __remove_mapping for reference */
1261                         if (!page_ref_freeze(page, 1))
1262                                 goto keep_locked;
1263                         if (PageDirty(page)) {
1264                                 page_ref_unfreeze(page, 1);
1265                                 goto keep_locked;
1266                         }
1267
1268                         count_vm_event(PGLAZYFREED);
1269                 } else if (!mapping || !__remove_mapping(mapping, page, true))
1270                         goto keep_locked;
1271                 /*
1272                  * At this point, we have no other references and there is
1273                  * no way to pick any more up (removed from LRU, removed
1274                  * from pagecache). Can use non-atomic bitops now (and
1275                  * we obviously don't have to worry about waking up a process
1276                  * waiting on the page lock, because there are no references.
1277                  */
1278                 __ClearPageLocked(page);
1279 free_it:
1280                 nr_reclaimed++;
1281
1282                 /*
1283                  * Is there need to periodically free_page_list? It would
1284                  * appear not as the counts should be low
1285                  */
1286                 list_add(&page->lru, &free_pages);
1287                 continue;
1288
1289 activate_locked:
1290                 /* Not a candidate for swapping, so reclaim swap space. */
1291                 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1292                                                 PageMlocked(page)))
1293                         try_to_free_swap(page);
1294                 VM_BUG_ON_PAGE(PageActive(page), page);
1295                 if (!PageMlocked(page)) {
1296                         SetPageActive(page);
1297                         pgactivate++;
1298                 }
1299 keep_locked:
1300                 unlock_page(page);
1301 keep:
1302                 list_add(&page->lru, &ret_pages);
1303                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1304         }
1305
1306         mem_cgroup_uncharge_list(&free_pages);
1307         try_to_unmap_flush();
1308         free_hot_cold_page_list(&free_pages, true);
1309
1310         list_splice(&ret_pages, page_list);
1311         count_vm_events(PGACTIVATE, pgactivate);
1312
1313         if (stat) {
1314                 stat->nr_dirty = nr_dirty;
1315                 stat->nr_congested = nr_congested;
1316                 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1317                 stat->nr_writeback = nr_writeback;
1318                 stat->nr_immediate = nr_immediate;
1319                 stat->nr_activate = pgactivate;
1320                 stat->nr_ref_keep = nr_ref_keep;
1321                 stat->nr_unmap_fail = nr_unmap_fail;
1322         }
1323         return nr_reclaimed;
1324 }
1325
1326 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1327                                             struct list_head *page_list)
1328 {
1329         struct scan_control sc = {
1330                 .gfp_mask = GFP_KERNEL,
1331                 .priority = DEF_PRIORITY,
1332                 .may_unmap = 1,
1333         };
1334         unsigned long ret;
1335         struct page *page, *next;
1336         LIST_HEAD(clean_pages);
1337
1338         list_for_each_entry_safe(page, next, page_list, lru) {
1339                 if (page_is_file_cache(page) && !PageDirty(page) &&
1340                     !__PageMovable(page)) {
1341                         ClearPageActive(page);
1342                         list_move(&page->lru, &clean_pages);
1343                 }
1344         }
1345
1346         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1347                         TTU_IGNORE_ACCESS, NULL, true);
1348         list_splice(&clean_pages, page_list);
1349         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1350         return ret;
1351 }
1352
1353 /*
1354  * Attempt to remove the specified page from its LRU.  Only take this page
1355  * if it is of the appropriate PageActive status.  Pages which are being
1356  * freed elsewhere are also ignored.
1357  *
1358  * page:        page to consider
1359  * mode:        one of the LRU isolation modes defined above
1360  *
1361  * returns 0 on success, -ve errno on failure.
1362  */
1363 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1364 {
1365         int ret = -EINVAL;
1366
1367         /* Only take pages on the LRU. */
1368         if (!PageLRU(page))
1369                 return ret;
1370
1371         /* Compaction should not handle unevictable pages but CMA can do so */
1372         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1373                 return ret;
1374
1375         ret = -EBUSY;
1376
1377         /*
1378          * To minimise LRU disruption, the caller can indicate that it only
1379          * wants to isolate pages it will be able to operate on without
1380          * blocking - clean pages for the most part.
1381          *
1382          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1383          * that it is possible to migrate without blocking
1384          */
1385         if (mode & ISOLATE_ASYNC_MIGRATE) {
1386                 /* All the caller can do on PageWriteback is block */
1387                 if (PageWriteback(page))
1388                         return ret;
1389
1390                 if (PageDirty(page)) {
1391                         struct address_space *mapping;
1392
1393                         /*
1394                          * Only pages without mappings or that have a
1395                          * ->migratepage callback are possible to migrate
1396                          * without blocking
1397                          */
1398                         mapping = page_mapping(page);
1399                         if (mapping && !mapping->a_ops->migratepage)
1400                                 return ret;
1401                 }
1402         }
1403
1404         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1405                 return ret;
1406
1407         if (likely(get_page_unless_zero(page))) {
1408                 /*
1409                  * Be careful not to clear PageLRU until after we're
1410                  * sure the page is not being freed elsewhere -- the
1411                  * page release code relies on it.
1412                  */
1413                 ClearPageLRU(page);
1414                 ret = 0;
1415         }
1416
1417         return ret;
1418 }
1419
1420
1421 /*
1422  * Update LRU sizes after isolating pages. The LRU size updates must
1423  * be complete before mem_cgroup_update_lru_size due to a santity check.
1424  */
1425 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1426                         enum lru_list lru, unsigned long *nr_zone_taken)
1427 {
1428         int zid;
1429
1430         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1431                 if (!nr_zone_taken[zid])
1432                         continue;
1433
1434                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1435 #ifdef CONFIG_MEMCG
1436                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1437 #endif
1438         }
1439
1440 }
1441
1442 /*
1443  * zone_lru_lock is heavily contended.  Some of the functions that
1444  * shrink the lists perform better by taking out a batch of pages
1445  * and working on them outside the LRU lock.
1446  *
1447  * For pagecache intensive workloads, this function is the hottest
1448  * spot in the kernel (apart from copy_*_user functions).
1449  *
1450  * Appropriate locks must be held before calling this function.
1451  *
1452  * @nr_to_scan: The number of pages to look through on the list.
1453  * @lruvec:     The LRU vector to pull pages from.
1454  * @dst:        The temp list to put pages on to.
1455  * @nr_scanned: The number of pages that were scanned.
1456  * @sc:         The scan_control struct for this reclaim session
1457  * @mode:       One of the LRU isolation modes
1458  * @lru:        LRU list id for isolating
1459  *
1460  * returns how many pages were moved onto *@dst.
1461  */
1462 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1463                 struct lruvec *lruvec, struct list_head *dst,
1464                 unsigned long *nr_scanned, struct scan_control *sc,
1465                 isolate_mode_t mode, enum lru_list lru)
1466 {
1467         struct list_head *src = &lruvec->lists[lru];
1468         unsigned long nr_taken = 0;
1469         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1470         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1471         unsigned long skipped = 0;
1472         unsigned long scan, nr_pages;
1473         LIST_HEAD(pages_skipped);
1474
1475         for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1476                                         !list_empty(src); scan++) {
1477                 struct page *page;
1478
1479                 page = lru_to_page(src);
1480                 prefetchw_prev_lru_page(page, src, flags);
1481
1482                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1483
1484                 if (page_zonenum(page) > sc->reclaim_idx) {
1485                         list_move(&page->lru, &pages_skipped);
1486                         nr_skipped[page_zonenum(page)]++;
1487                         continue;
1488                 }
1489
1490                 switch (__isolate_lru_page(page, mode)) {
1491                 case 0:
1492                         nr_pages = hpage_nr_pages(page);
1493                         nr_taken += nr_pages;
1494                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1495                         list_move(&page->lru, dst);
1496                         break;
1497
1498                 case -EBUSY:
1499                         /* else it is being freed elsewhere */
1500                         list_move(&page->lru, src);
1501                         continue;
1502
1503                 default:
1504                         BUG();
1505                 }
1506         }
1507
1508         /*
1509          * Splice any skipped pages to the start of the LRU list. Note that
1510          * this disrupts the LRU order when reclaiming for lower zones but
1511          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1512          * scanning would soon rescan the same pages to skip and put the
1513          * system at risk of premature OOM.
1514          */
1515         if (!list_empty(&pages_skipped)) {
1516                 int zid;
1517
1518                 list_splice(&pages_skipped, src);
1519                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1520                         if (!nr_skipped[zid])
1521                                 continue;
1522
1523                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1524                         skipped += nr_skipped[zid];
1525                 }
1526         }
1527         *nr_scanned = scan;
1528         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1529                                     scan, skipped, nr_taken, mode, lru);
1530         update_lru_sizes(lruvec, lru, nr_zone_taken);
1531         return nr_taken;
1532 }
1533
1534 /**
1535  * isolate_lru_page - tries to isolate a page from its LRU list
1536  * @page: page to isolate from its LRU list
1537  *
1538  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1539  * vmstat statistic corresponding to whatever LRU list the page was on.
1540  *
1541  * Returns 0 if the page was removed from an LRU list.
1542  * Returns -EBUSY if the page was not on an LRU list.
1543  *
1544  * The returned page will have PageLRU() cleared.  If it was found on
1545  * the active list, it will have PageActive set.  If it was found on
1546  * the unevictable list, it will have the PageUnevictable bit set. That flag
1547  * may need to be cleared by the caller before letting the page go.
1548  *
1549  * The vmstat statistic corresponding to the list on which the page was
1550  * found will be decremented.
1551  *
1552  * Restrictions:
1553  * (1) Must be called with an elevated refcount on the page. This is a
1554  *     fundamentnal difference from isolate_lru_pages (which is called
1555  *     without a stable reference).
1556  * (2) the lru_lock must not be held.
1557  * (3) interrupts must be enabled.
1558  */
1559 int isolate_lru_page(struct page *page)
1560 {
1561         int ret = -EBUSY;
1562
1563         VM_BUG_ON_PAGE(!page_count(page), page);
1564         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1565
1566         if (PageLRU(page)) {
1567                 struct zone *zone = page_zone(page);
1568                 struct lruvec *lruvec;
1569
1570                 spin_lock_irq(zone_lru_lock(zone));
1571                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1572                 if (PageLRU(page)) {
1573                         int lru = page_lru(page);
1574                         get_page(page);
1575                         ClearPageLRU(page);
1576                         del_page_from_lru_list(page, lruvec, lru);
1577                         ret = 0;
1578                 }
1579                 spin_unlock_irq(zone_lru_lock(zone));
1580         }
1581         return ret;
1582 }
1583
1584 /*
1585  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1586  * then get resheduled. When there are massive number of tasks doing page
1587  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1588  * the LRU list will go small and be scanned faster than necessary, leading to
1589  * unnecessary swapping, thrashing and OOM.
1590  */
1591 static int too_many_isolated(struct pglist_data *pgdat, int file,
1592                 struct scan_control *sc)
1593 {
1594         unsigned long inactive, isolated;
1595
1596         if (current_is_kswapd())
1597                 return 0;
1598
1599         if (!sane_reclaim(sc))
1600                 return 0;
1601
1602         if (file) {
1603                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1604                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1605         } else {
1606                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1607                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1608         }
1609
1610         /*
1611          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1612          * won't get blocked by normal direct-reclaimers, forming a circular
1613          * deadlock.
1614          */
1615         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1616                 inactive >>= 3;
1617
1618         return isolated > inactive;
1619 }
1620
1621 static noinline_for_stack void
1622 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1623 {
1624         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1625         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1626         LIST_HEAD(pages_to_free);
1627
1628         /*
1629          * Put back any unfreeable pages.
1630          */
1631         while (!list_empty(page_list)) {
1632                 struct page *page = lru_to_page(page_list);
1633                 int lru;
1634
1635                 VM_BUG_ON_PAGE(PageLRU(page), page);
1636                 list_del(&page->lru);
1637                 if (unlikely(!page_evictable(page))) {
1638                         spin_unlock_irq(&pgdat->lru_lock);
1639                         putback_lru_page(page);
1640                         spin_lock_irq(&pgdat->lru_lock);
1641                         continue;
1642                 }
1643
1644                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1645
1646                 SetPageLRU(page);
1647                 lru = page_lru(page);
1648                 add_page_to_lru_list(page, lruvec, lru);
1649
1650                 if (is_active_lru(lru)) {
1651                         int file = is_file_lru(lru);
1652                         int numpages = hpage_nr_pages(page);
1653                         reclaim_stat->recent_rotated[file] += numpages;
1654                 }
1655                 if (put_page_testzero(page)) {
1656                         __ClearPageLRU(page);
1657                         __ClearPageActive(page);
1658                         del_page_from_lru_list(page, lruvec, lru);
1659
1660                         if (unlikely(PageCompound(page))) {
1661                                 spin_unlock_irq(&pgdat->lru_lock);
1662                                 mem_cgroup_uncharge(page);
1663                                 (*get_compound_page_dtor(page))(page);
1664                                 spin_lock_irq(&pgdat->lru_lock);
1665                         } else
1666                                 list_add(&page->lru, &pages_to_free);
1667                 }
1668         }
1669
1670         /*
1671          * To save our caller's stack, now use input list for pages to free.
1672          */
1673         list_splice(&pages_to_free, page_list);
1674 }
1675
1676 /*
1677  * If a kernel thread (such as nfsd for loop-back mounts) services
1678  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1679  * In that case we should only throttle if the backing device it is
1680  * writing to is congested.  In other cases it is safe to throttle.
1681  */
1682 static int current_may_throttle(void)
1683 {
1684         return !(current->flags & PF_LESS_THROTTLE) ||
1685                 current->backing_dev_info == NULL ||
1686                 bdi_write_congested(current->backing_dev_info);
1687 }
1688
1689 /*
1690  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1691  * of reclaimed pages
1692  */
1693 static noinline_for_stack unsigned long
1694 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1695                      struct scan_control *sc, enum lru_list lru)
1696 {
1697         LIST_HEAD(page_list);
1698         unsigned long nr_scanned;
1699         unsigned long nr_reclaimed = 0;
1700         unsigned long nr_taken;
1701         struct reclaim_stat stat = {};
1702         isolate_mode_t isolate_mode = 0;
1703         int file = is_file_lru(lru);
1704         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1705         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1706
1707         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1708                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1709
1710                 /* We are about to die and free our memory. Return now. */
1711                 if (fatal_signal_pending(current))
1712                         return SWAP_CLUSTER_MAX;
1713         }
1714
1715         lru_add_drain();
1716
1717         if (!sc->may_unmap)
1718                 isolate_mode |= ISOLATE_UNMAPPED;
1719
1720         spin_lock_irq(&pgdat->lru_lock);
1721
1722         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1723                                      &nr_scanned, sc, isolate_mode, lru);
1724
1725         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1726         reclaim_stat->recent_scanned[file] += nr_taken;
1727
1728         if (global_reclaim(sc)) {
1729                 if (current_is_kswapd())
1730                         __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1731                 else
1732                         __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1733         }
1734         spin_unlock_irq(&pgdat->lru_lock);
1735
1736         if (nr_taken == 0)
1737                 return 0;
1738
1739         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1740                                 &stat, false);
1741
1742         spin_lock_irq(&pgdat->lru_lock);
1743
1744         if (global_reclaim(sc)) {
1745                 if (current_is_kswapd())
1746                         __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1747                 else
1748                         __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1749         }
1750
1751         putback_inactive_pages(lruvec, &page_list);
1752
1753         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1754
1755         spin_unlock_irq(&pgdat->lru_lock);
1756
1757         mem_cgroup_uncharge_list(&page_list);
1758         free_hot_cold_page_list(&page_list, true);
1759
1760         /*
1761          * If reclaim is isolating dirty pages under writeback, it implies
1762          * that the long-lived page allocation rate is exceeding the page
1763          * laundering rate. Either the global limits are not being effective
1764          * at throttling processes due to the page distribution throughout
1765          * zones or there is heavy usage of a slow backing device. The
1766          * only option is to throttle from reclaim context which is not ideal
1767          * as there is no guarantee the dirtying process is throttled in the
1768          * same way balance_dirty_pages() manages.
1769          *
1770          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1771          * of pages under pages flagged for immediate reclaim and stall if any
1772          * are encountered in the nr_immediate check below.
1773          */
1774         if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1775                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1776
1777         /*
1778          * Legacy memcg will stall in page writeback so avoid forcibly
1779          * stalling here.
1780          */
1781         if (sane_reclaim(sc)) {
1782                 /*
1783                  * Tag a zone as congested if all the dirty pages scanned were
1784                  * backed by a congested BDI and wait_iff_congested will stall.
1785                  */
1786                 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1787                         set_bit(PGDAT_CONGESTED, &pgdat->flags);
1788
1789                 /*
1790                  * If dirty pages are scanned that are not queued for IO, it
1791                  * implies that flushers are not doing their job. This can
1792                  * happen when memory pressure pushes dirty pages to the end of
1793                  * the LRU before the dirty limits are breached and the dirty
1794                  * data has expired. It can also happen when the proportion of
1795                  * dirty pages grows not through writes but through memory
1796                  * pressure reclaiming all the clean cache. And in some cases,
1797                  * the flushers simply cannot keep up with the allocation
1798                  * rate. Nudge the flusher threads in case they are asleep, but
1799                  * also allow kswapd to start writing pages during reclaim.
1800                  */
1801                 if (stat.nr_unqueued_dirty == nr_taken) {
1802                         wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1803                         set_bit(PGDAT_DIRTY, &pgdat->flags);
1804                 }
1805
1806                 /*
1807                  * If kswapd scans pages marked marked for immediate
1808                  * reclaim and under writeback (nr_immediate), it implies
1809                  * that pages are cycling through the LRU faster than
1810                  * they are written so also forcibly stall.
1811                  */
1812                 if (stat.nr_immediate && current_may_throttle())
1813                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1814         }
1815
1816         /*
1817          * Stall direct reclaim for IO completions if underlying BDIs or zone
1818          * is congested. Allow kswapd to continue until it starts encountering
1819          * unqueued dirty pages or cycling through the LRU too quickly.
1820          */
1821         if (!sc->hibernation_mode && !current_is_kswapd() &&
1822             current_may_throttle())
1823                 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1824
1825         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1826                         nr_scanned, nr_reclaimed,
1827                         stat.nr_dirty,  stat.nr_writeback,
1828                         stat.nr_congested, stat.nr_immediate,
1829                         stat.nr_activate, stat.nr_ref_keep,
1830                         stat.nr_unmap_fail,
1831                         sc->priority, file);
1832         return nr_reclaimed;
1833 }
1834
1835 /*
1836  * This moves pages from the active list to the inactive list.
1837  *
1838  * We move them the other way if the page is referenced by one or more
1839  * processes, from rmap.
1840  *
1841  * If the pages are mostly unmapped, the processing is fast and it is
1842  * appropriate to hold zone_lru_lock across the whole operation.  But if
1843  * the pages are mapped, the processing is slow (page_referenced()) so we
1844  * should drop zone_lru_lock around each page.  It's impossible to balance
1845  * this, so instead we remove the pages from the LRU while processing them.
1846  * It is safe to rely on PG_active against the non-LRU pages in here because
1847  * nobody will play with that bit on a non-LRU page.
1848  *
1849  * The downside is that we have to touch page->_refcount against each page.
1850  * But we had to alter page->flags anyway.
1851  *
1852  * Returns the number of pages moved to the given lru.
1853  */
1854
1855 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1856                                      struct list_head *list,
1857                                      struct list_head *pages_to_free,
1858                                      enum lru_list lru)
1859 {
1860         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1861         struct page *page;
1862         int nr_pages;
1863         int nr_moved = 0;
1864
1865         while (!list_empty(list)) {
1866                 page = lru_to_page(list);
1867                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1868
1869                 VM_BUG_ON_PAGE(PageLRU(page), page);
1870                 SetPageLRU(page);
1871
1872                 nr_pages = hpage_nr_pages(page);
1873                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1874                 list_move(&page->lru, &lruvec->lists[lru]);
1875
1876                 if (put_page_testzero(page)) {
1877                         __ClearPageLRU(page);
1878                         __ClearPageActive(page);
1879                         del_page_from_lru_list(page, lruvec, lru);
1880
1881                         if (unlikely(PageCompound(page))) {
1882                                 spin_unlock_irq(&pgdat->lru_lock);
1883                                 mem_cgroup_uncharge(page);
1884                                 (*get_compound_page_dtor(page))(page);
1885                                 spin_lock_irq(&pgdat->lru_lock);
1886                         } else
1887                                 list_add(&page->lru, pages_to_free);
1888                 } else {
1889                         nr_moved += nr_pages;
1890                 }
1891         }
1892
1893         if (!is_active_lru(lru))
1894                 __count_vm_events(PGDEACTIVATE, nr_moved);
1895
1896         return nr_moved;
1897 }
1898
1899 static void shrink_active_list(unsigned long nr_to_scan,
1900                                struct lruvec *lruvec,
1901                                struct scan_control *sc,
1902                                enum lru_list lru)
1903 {
1904         unsigned long nr_taken;
1905         unsigned long nr_scanned;
1906         unsigned long vm_flags;
1907         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1908         LIST_HEAD(l_active);
1909         LIST_HEAD(l_inactive);
1910         struct page *page;
1911         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1912         unsigned nr_deactivate, nr_activate;
1913         unsigned nr_rotated = 0;
1914         isolate_mode_t isolate_mode = 0;
1915         int file = is_file_lru(lru);
1916         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1917
1918         lru_add_drain();
1919
1920         if (!sc->may_unmap)
1921                 isolate_mode |= ISOLATE_UNMAPPED;
1922
1923         spin_lock_irq(&pgdat->lru_lock);
1924
1925         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1926                                      &nr_scanned, sc, isolate_mode, lru);
1927
1928         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1929         reclaim_stat->recent_scanned[file] += nr_taken;
1930
1931         __count_vm_events(PGREFILL, nr_scanned);
1932
1933         spin_unlock_irq(&pgdat->lru_lock);
1934
1935         while (!list_empty(&l_hold)) {
1936                 cond_resched();
1937                 page = lru_to_page(&l_hold);
1938                 list_del(&page->lru);
1939
1940                 if (unlikely(!page_evictable(page))) {
1941                         putback_lru_page(page);
1942                         continue;
1943                 }
1944
1945                 if (unlikely(buffer_heads_over_limit)) {
1946                         if (page_has_private(page) && trylock_page(page)) {
1947                                 if (page_has_private(page))
1948                                         try_to_release_page(page, 0);
1949                                 unlock_page(page);
1950                         }
1951                 }
1952
1953                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1954                                     &vm_flags)) {
1955                         nr_rotated += hpage_nr_pages(page);
1956                         /*
1957                          * Identify referenced, file-backed active pages and
1958                          * give them one more trip around the active list. So
1959                          * that executable code get better chances to stay in
1960                          * memory under moderate memory pressure.  Anon pages
1961                          * are not likely to be evicted by use-once streaming
1962                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1963                          * so we ignore them here.
1964                          */
1965                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1966                                 list_add(&page->lru, &l_active);
1967                                 continue;
1968                         }
1969                 }
1970
1971                 ClearPageActive(page);  /* we are de-activating */
1972                 list_add(&page->lru, &l_inactive);
1973         }
1974
1975         /*
1976          * Move pages back to the lru list.
1977          */
1978         spin_lock_irq(&pgdat->lru_lock);
1979         /*
1980          * Count referenced pages from currently used mappings as rotated,
1981          * even though only some of them are actually re-activated.  This
1982          * helps balance scan pressure between file and anonymous pages in
1983          * get_scan_count.
1984          */
1985         reclaim_stat->recent_rotated[file] += nr_rotated;
1986
1987         nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1988         nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1989         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1990         spin_unlock_irq(&pgdat->lru_lock);
1991
1992         mem_cgroup_uncharge_list(&l_hold);
1993         free_hot_cold_page_list(&l_hold, true);
1994         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
1995                         nr_deactivate, nr_rotated, sc->priority, file);
1996 }
1997
1998 /*
1999  * The inactive anon list should be small enough that the VM never has
2000  * to do too much work.
2001  *
2002  * The inactive file list should be small enough to leave most memory
2003  * to the established workingset on the scan-resistant active list,
2004  * but large enough to avoid thrashing the aggregate readahead window.
2005  *
2006  * Both inactive lists should also be large enough that each inactive
2007  * page has a chance to be referenced again before it is reclaimed.
2008  *
2009  * If that fails and refaulting is observed, the inactive list grows.
2010  *
2011  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2012  * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2013  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2014  *
2015  * total     target    max
2016  * memory    ratio     inactive
2017  * -------------------------------------
2018  *   10MB       1         5MB
2019  *  100MB       1        50MB
2020  *    1GB       3       250MB
2021  *   10GB      10       0.9GB
2022  *  100GB      31         3GB
2023  *    1TB     101        10GB
2024  *   10TB     320        32GB
2025  */
2026 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2027                                  struct mem_cgroup *memcg,
2028                                  struct scan_control *sc, bool actual_reclaim)
2029 {
2030         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2031         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2032         enum lru_list inactive_lru = file * LRU_FILE;
2033         unsigned long inactive, active;
2034         unsigned long inactive_ratio;
2035         unsigned long refaults;
2036         unsigned long gb;
2037
2038         /*
2039          * If we don't have swap space, anonymous page deactivation
2040          * is pointless.
2041          */
2042         if (!file && !total_swap_pages)
2043                 return false;
2044
2045         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2046         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2047
2048         if (memcg)
2049                 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2050         else
2051                 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2052
2053         /*
2054          * When refaults are being observed, it means a new workingset
2055          * is being established. Disable active list protection to get
2056          * rid of the stale workingset quickly.
2057          */
2058         if (file && actual_reclaim && lruvec->refaults != refaults) {
2059                 inactive_ratio = 0;
2060         } else {
2061                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2062                 if (gb)
2063                         inactive_ratio = int_sqrt(10 * gb);
2064                 else
2065                         inactive_ratio = 1;
2066         }
2067
2068         if (actual_reclaim)
2069                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2070                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2071                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2072                         inactive_ratio, file);
2073
2074         return inactive * inactive_ratio < active;
2075 }
2076
2077 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2078                                  struct lruvec *lruvec, struct mem_cgroup *memcg,
2079                                  struct scan_control *sc)
2080 {
2081         if (is_active_lru(lru)) {
2082                 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2083                                          memcg, sc, true))
2084                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2085                 return 0;
2086         }
2087
2088         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2089 }
2090
2091 enum scan_balance {
2092         SCAN_EQUAL,
2093         SCAN_FRACT,
2094         SCAN_ANON,
2095         SCAN_FILE,
2096 };
2097
2098 /*
2099  * Determine how aggressively the anon and file LRU lists should be
2100  * scanned.  The relative value of each set of LRU lists is determined
2101  * by looking at the fraction of the pages scanned we did rotate back
2102  * onto the active list instead of evict.
2103  *
2104  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2105  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2106  */
2107 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2108                            struct scan_control *sc, unsigned long *nr,
2109                            unsigned long *lru_pages)
2110 {
2111         int swappiness = mem_cgroup_swappiness(memcg);
2112         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2113         u64 fraction[2];
2114         u64 denominator = 0;    /* gcc */
2115         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2116         unsigned long anon_prio, file_prio;
2117         enum scan_balance scan_balance;
2118         unsigned long anon, file;
2119         unsigned long ap, fp;
2120         enum lru_list lru;
2121
2122         /* If we have no swap space, do not bother scanning anon pages. */
2123         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2124                 scan_balance = SCAN_FILE;
2125                 goto out;
2126         }
2127
2128         /*
2129          * Global reclaim will swap to prevent OOM even with no
2130          * swappiness, but memcg users want to use this knob to
2131          * disable swapping for individual groups completely when
2132          * using the memory controller's swap limit feature would be
2133          * too expensive.
2134          */
2135         if (!global_reclaim(sc) && !swappiness) {
2136                 scan_balance = SCAN_FILE;
2137                 goto out;
2138         }
2139
2140         /*
2141          * Do not apply any pressure balancing cleverness when the
2142          * system is close to OOM, scan both anon and file equally
2143          * (unless the swappiness setting disagrees with swapping).
2144          */
2145         if (!sc->priority && swappiness) {
2146                 scan_balance = SCAN_EQUAL;
2147                 goto out;
2148         }
2149
2150         /*
2151          * Prevent the reclaimer from falling into the cache trap: as
2152          * cache pages start out inactive, every cache fault will tip
2153          * the scan balance towards the file LRU.  And as the file LRU
2154          * shrinks, so does the window for rotation from references.
2155          * This means we have a runaway feedback loop where a tiny
2156          * thrashing file LRU becomes infinitely more attractive than
2157          * anon pages.  Try to detect this based on file LRU size.
2158          */
2159         if (global_reclaim(sc)) {
2160                 unsigned long pgdatfile;
2161                 unsigned long pgdatfree;
2162                 int z;
2163                 unsigned long total_high_wmark = 0;
2164
2165                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2166                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2167                            node_page_state(pgdat, NR_INACTIVE_FILE);
2168
2169                 for (z = 0; z < MAX_NR_ZONES; z++) {
2170                         struct zone *zone = &pgdat->node_zones[z];
2171                         if (!managed_zone(zone))
2172                                 continue;
2173
2174                         total_high_wmark += high_wmark_pages(zone);
2175                 }
2176
2177                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2178                         scan_balance = SCAN_ANON;
2179                         goto out;
2180                 }
2181         }
2182
2183         /*
2184          * If there is enough inactive page cache, i.e. if the size of the
2185          * inactive list is greater than that of the active list *and* the
2186          * inactive list actually has some pages to scan on this priority, we
2187          * do not reclaim anything from the anonymous working set right now.
2188          * Without the second condition we could end up never scanning an
2189          * lruvec even if it has plenty of old anonymous pages unless the
2190          * system is under heavy pressure.
2191          */
2192         if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2193             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2194                 scan_balance = SCAN_FILE;
2195                 goto out;
2196         }
2197
2198         scan_balance = SCAN_FRACT;
2199
2200         /*
2201          * With swappiness at 100, anonymous and file have the same priority.
2202          * This scanning priority is essentially the inverse of IO cost.
2203          */
2204         anon_prio = swappiness;
2205         file_prio = 200 - anon_prio;
2206
2207         /*
2208          * OK, so we have swap space and a fair amount of page cache
2209          * pages.  We use the recently rotated / recently scanned
2210          * ratios to determine how valuable each cache is.
2211          *
2212          * Because workloads change over time (and to avoid overflow)
2213          * we keep these statistics as a floating average, which ends
2214          * up weighing recent references more than old ones.
2215          *
2216          * anon in [0], file in [1]
2217          */
2218
2219         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2220                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2221         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2222                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2223
2224         spin_lock_irq(&pgdat->lru_lock);
2225         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2226                 reclaim_stat->recent_scanned[0] /= 2;
2227                 reclaim_stat->recent_rotated[0] /= 2;
2228         }
2229
2230         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2231                 reclaim_stat->recent_scanned[1] /= 2;
2232                 reclaim_stat->recent_rotated[1] /= 2;
2233         }
2234
2235         /*
2236          * The amount of pressure on anon vs file pages is inversely
2237          * proportional to the fraction of recently scanned pages on
2238          * each list that were recently referenced and in active use.
2239          */
2240         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2241         ap /= reclaim_stat->recent_rotated[0] + 1;
2242
2243         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2244         fp /= reclaim_stat->recent_rotated[1] + 1;
2245         spin_unlock_irq(&pgdat->lru_lock);
2246
2247         fraction[0] = ap;
2248         fraction[1] = fp;
2249         denominator = ap + fp + 1;
2250 out:
2251         *lru_pages = 0;
2252         for_each_evictable_lru(lru) {
2253                 int file = is_file_lru(lru);
2254                 unsigned long size;
2255                 unsigned long scan;
2256
2257                 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2258                 scan = size >> sc->priority;
2259                 /*
2260                  * If the cgroup's already been deleted, make sure to
2261                  * scrape out the remaining cache.
2262                  */
2263                 if (!scan && !mem_cgroup_online(memcg))
2264                         scan = min(size, SWAP_CLUSTER_MAX);
2265
2266                 switch (scan_balance) {
2267                 case SCAN_EQUAL:
2268                         /* Scan lists relative to size */
2269                         break;
2270                 case SCAN_FRACT:
2271                         /*
2272                          * Scan types proportional to swappiness and
2273                          * their relative recent reclaim efficiency.
2274                          */
2275                         scan = div64_u64(scan * fraction[file],
2276                                          denominator);
2277                         break;
2278                 case SCAN_FILE:
2279                 case SCAN_ANON:
2280                         /* Scan one type exclusively */
2281                         if ((scan_balance == SCAN_FILE) != file) {
2282                                 size = 0;
2283                                 scan = 0;
2284                         }
2285                         break;
2286                 default:
2287                         /* Look ma, no brain */
2288                         BUG();
2289                 }
2290
2291                 *lru_pages += size;
2292                 nr[lru] = scan;
2293         }
2294 }
2295
2296 /*
2297  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2298  */
2299 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2300                               struct scan_control *sc, unsigned long *lru_pages)
2301 {
2302         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2303         unsigned long nr[NR_LRU_LISTS];
2304         unsigned long targets[NR_LRU_LISTS];
2305         unsigned long nr_to_scan;
2306         enum lru_list lru;
2307         unsigned long nr_reclaimed = 0;
2308         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2309         struct blk_plug plug;
2310         bool scan_adjusted;
2311
2312         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2313
2314         /* Record the original scan target for proportional adjustments later */
2315         memcpy(targets, nr, sizeof(nr));
2316
2317         /*
2318          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2319          * event that can occur when there is little memory pressure e.g.
2320          * multiple streaming readers/writers. Hence, we do not abort scanning
2321          * when the requested number of pages are reclaimed when scanning at
2322          * DEF_PRIORITY on the assumption that the fact we are direct
2323          * reclaiming implies that kswapd is not keeping up and it is best to
2324          * do a batch of work at once. For memcg reclaim one check is made to
2325          * abort proportional reclaim if either the file or anon lru has already
2326          * dropped to zero at the first pass.
2327          */
2328         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2329                          sc->priority == DEF_PRIORITY);
2330
2331         blk_start_plug(&plug);
2332         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2333                                         nr[LRU_INACTIVE_FILE]) {
2334                 unsigned long nr_anon, nr_file, percentage;
2335                 unsigned long nr_scanned;
2336
2337                 for_each_evictable_lru(lru) {
2338                         if (nr[lru]) {
2339                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2340                                 nr[lru] -= nr_to_scan;
2341
2342                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2343                                                             lruvec, memcg, sc);
2344                         }
2345                 }
2346
2347                 cond_resched();
2348
2349                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2350                         continue;
2351
2352                 /*
2353                  * For kswapd and memcg, reclaim at least the number of pages
2354                  * requested. Ensure that the anon and file LRUs are scanned
2355                  * proportionally what was requested by get_scan_count(). We
2356                  * stop reclaiming one LRU and reduce the amount scanning
2357                  * proportional to the original scan target.
2358                  */
2359                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2360                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2361
2362                 /*
2363                  * It's just vindictive to attack the larger once the smaller
2364                  * has gone to zero.  And given the way we stop scanning the
2365                  * smaller below, this makes sure that we only make one nudge
2366                  * towards proportionality once we've got nr_to_reclaim.
2367                  */
2368                 if (!nr_file || !nr_anon)
2369                         break;
2370
2371                 if (nr_file > nr_anon) {
2372                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2373                                                 targets[LRU_ACTIVE_ANON] + 1;
2374                         lru = LRU_BASE;
2375                         percentage = nr_anon * 100 / scan_target;
2376                 } else {
2377                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2378                                                 targets[LRU_ACTIVE_FILE] + 1;
2379                         lru = LRU_FILE;
2380                         percentage = nr_file * 100 / scan_target;
2381                 }
2382
2383                 /* Stop scanning the smaller of the LRU */
2384                 nr[lru] = 0;
2385                 nr[lru + LRU_ACTIVE] = 0;
2386
2387                 /*
2388                  * Recalculate the other LRU scan count based on its original
2389                  * scan target and the percentage scanning already complete
2390                  */
2391                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2392                 nr_scanned = targets[lru] - nr[lru];
2393                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2394                 nr[lru] -= min(nr[lru], nr_scanned);
2395
2396                 lru += LRU_ACTIVE;
2397                 nr_scanned = targets[lru] - nr[lru];
2398                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2399                 nr[lru] -= min(nr[lru], nr_scanned);
2400
2401                 scan_adjusted = true;
2402         }
2403         blk_finish_plug(&plug);
2404         sc->nr_reclaimed += nr_reclaimed;
2405
2406         /*
2407          * Even if we did not try to evict anon pages at all, we want to
2408          * rebalance the anon lru active/inactive ratio.
2409          */
2410         if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2411                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2412                                    sc, LRU_ACTIVE_ANON);
2413 }
2414
2415 /* Use reclaim/compaction for costly allocs or under memory pressure */
2416 static bool in_reclaim_compaction(struct scan_control *sc)
2417 {
2418         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2419                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2420                          sc->priority < DEF_PRIORITY - 2))
2421                 return true;
2422
2423         return false;
2424 }
2425
2426 /*
2427  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2428  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2429  * true if more pages should be reclaimed such that when the page allocator
2430  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2431  * It will give up earlier than that if there is difficulty reclaiming pages.
2432  */
2433 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2434                                         unsigned long nr_reclaimed,
2435                                         unsigned long nr_scanned,
2436                                         struct scan_control *sc)
2437 {
2438         unsigned long pages_for_compaction;
2439         unsigned long inactive_lru_pages;
2440         int z;
2441
2442         /* If not in reclaim/compaction mode, stop */
2443         if (!in_reclaim_compaction(sc))
2444                 return false;
2445
2446         /* Consider stopping depending on scan and reclaim activity */
2447         if (sc->gfp_mask & __GFP_REPEAT) {
2448                 /*
2449                  * For __GFP_REPEAT allocations, stop reclaiming if the
2450                  * full LRU list has been scanned and we are still failing
2451                  * to reclaim pages. This full LRU scan is potentially
2452                  * expensive but a __GFP_REPEAT caller really wants to succeed
2453                  */
2454                 if (!nr_reclaimed && !nr_scanned)
2455                         return false;
2456         } else {
2457                 /*
2458                  * For non-__GFP_REPEAT allocations which can presumably
2459                  * fail without consequence, stop if we failed to reclaim
2460                  * any pages from the last SWAP_CLUSTER_MAX number of
2461                  * pages that were scanned. This will return to the
2462                  * caller faster at the risk reclaim/compaction and
2463                  * the resulting allocation attempt fails
2464                  */
2465                 if (!nr_reclaimed)
2466                         return false;
2467         }
2468
2469         /*
2470          * If we have not reclaimed enough pages for compaction and the
2471          * inactive lists are large enough, continue reclaiming
2472          */
2473         pages_for_compaction = compact_gap(sc->order);
2474         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2475         if (get_nr_swap_pages() > 0)
2476                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2477         if (sc->nr_reclaimed < pages_for_compaction &&
2478                         inactive_lru_pages > pages_for_compaction)
2479                 return true;
2480
2481         /* If compaction would go ahead or the allocation would succeed, stop */
2482         for (z = 0; z <= sc->reclaim_idx; z++) {
2483                 struct zone *zone = &pgdat->node_zones[z];
2484                 if (!managed_zone(zone))
2485                         continue;
2486
2487                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2488                 case COMPACT_SUCCESS:
2489                 case COMPACT_CONTINUE:
2490                         return false;
2491                 default:
2492                         /* check next zone */
2493                         ;
2494                 }
2495         }
2496         return true;
2497 }
2498
2499 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2500 {
2501         struct reclaim_state *reclaim_state = current->reclaim_state;
2502         unsigned long nr_reclaimed, nr_scanned;
2503         bool reclaimable = false;
2504
2505         do {
2506                 struct mem_cgroup *root = sc->target_mem_cgroup;
2507                 struct mem_cgroup_reclaim_cookie reclaim = {
2508                         .pgdat = pgdat,
2509                         .priority = sc->priority,
2510                 };
2511                 unsigned long node_lru_pages = 0;
2512                 struct mem_cgroup *memcg;
2513
2514                 nr_reclaimed = sc->nr_reclaimed;
2515                 nr_scanned = sc->nr_scanned;
2516
2517                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2518                 do {
2519                         unsigned long lru_pages;
2520                         unsigned long reclaimed;
2521                         unsigned long scanned;
2522
2523                         if (mem_cgroup_low(root, memcg)) {
2524                                 if (!sc->memcg_low_reclaim) {
2525                                         sc->memcg_low_skipped = 1;
2526                                         continue;
2527                                 }
2528                                 mem_cgroup_event(memcg, MEMCG_LOW);
2529                         }
2530
2531                         reclaimed = sc->nr_reclaimed;
2532                         scanned = sc->nr_scanned;
2533
2534                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2535                         node_lru_pages += lru_pages;
2536
2537                         if (memcg)
2538                                 shrink_slab(sc->gfp_mask, pgdat->node_id,
2539                                             memcg, sc->nr_scanned - scanned,
2540                                             lru_pages);
2541
2542                         /* Record the group's reclaim efficiency */
2543                         vmpressure(sc->gfp_mask, memcg, false,
2544                                    sc->nr_scanned - scanned,
2545                                    sc->nr_reclaimed - reclaimed);
2546
2547                         /*
2548                          * Direct reclaim and kswapd have to scan all memory
2549                          * cgroups to fulfill the overall scan target for the
2550                          * node.
2551                          *
2552                          * Limit reclaim, on the other hand, only cares about
2553                          * nr_to_reclaim pages to be reclaimed and it will
2554                          * retry with decreasing priority if one round over the
2555                          * whole hierarchy is not sufficient.
2556                          */
2557                         if (!global_reclaim(sc) &&
2558                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2559                                 mem_cgroup_iter_break(root, memcg);
2560                                 break;
2561                         }
2562                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2563
2564                 /*
2565                  * Shrink the slab caches in the same proportion that
2566                  * the eligible LRU pages were scanned.
2567                  */
2568                 if (global_reclaim(sc))
2569                         shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2570                                     sc->nr_scanned - nr_scanned,
2571                                     node_lru_pages);
2572
2573                 if (reclaim_state) {
2574                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2575                         reclaim_state->reclaimed_slab = 0;
2576                 }
2577
2578                 /* Record the subtree's reclaim efficiency */
2579                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2580                            sc->nr_scanned - nr_scanned,
2581                            sc->nr_reclaimed - nr_reclaimed);
2582
2583                 if (sc->nr_reclaimed - nr_reclaimed)
2584                         reclaimable = true;
2585
2586         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2587                                          sc->nr_scanned - nr_scanned, sc));
2588
2589         /*
2590          * Kswapd gives up on balancing particular nodes after too
2591          * many failures to reclaim anything from them and goes to
2592          * sleep. On reclaim progress, reset the failure counter. A
2593          * successful direct reclaim run will revive a dormant kswapd.
2594          */
2595         if (reclaimable)
2596                 pgdat->kswapd_failures = 0;
2597
2598         return reclaimable;
2599 }
2600
2601 /*
2602  * Returns true if compaction should go ahead for a costly-order request, or
2603  * the allocation would already succeed without compaction. Return false if we
2604  * should reclaim first.
2605  */
2606 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2607 {
2608         unsigned long watermark;
2609         enum compact_result suitable;
2610
2611         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2612         if (suitable == COMPACT_SUCCESS)
2613                 /* Allocation should succeed already. Don't reclaim. */
2614                 return true;
2615         if (suitable == COMPACT_SKIPPED)
2616                 /* Compaction cannot yet proceed. Do reclaim. */
2617                 return false;
2618
2619         /*
2620          * Compaction is already possible, but it takes time to run and there
2621          * are potentially other callers using the pages just freed. So proceed
2622          * with reclaim to make a buffer of free pages available to give
2623          * compaction a reasonable chance of completing and allocating the page.
2624          * Note that we won't actually reclaim the whole buffer in one attempt
2625          * as the target watermark in should_continue_reclaim() is lower. But if
2626          * we are already above the high+gap watermark, don't reclaim at all.
2627          */
2628         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2629
2630         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2631 }
2632
2633 /*
2634  * This is the direct reclaim path, for page-allocating processes.  We only
2635  * try to reclaim pages from zones which will satisfy the caller's allocation
2636  * request.
2637  *
2638  * If a zone is deemed to be full of pinned pages then just give it a light
2639  * scan then give up on it.
2640  */
2641 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2642 {
2643         struct zoneref *z;
2644         struct zone *zone;
2645         unsigned long nr_soft_reclaimed;
2646         unsigned long nr_soft_scanned;
2647         gfp_t orig_mask;
2648         pg_data_t *last_pgdat = NULL;
2649
2650         /*
2651          * If the number of buffer_heads in the machine exceeds the maximum
2652          * allowed level, force direct reclaim to scan the highmem zone as
2653          * highmem pages could be pinning lowmem pages storing buffer_heads
2654          */
2655         orig_mask = sc->gfp_mask;
2656         if (buffer_heads_over_limit) {
2657                 sc->gfp_mask |= __GFP_HIGHMEM;
2658                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2659         }
2660
2661         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2662                                         sc->reclaim_idx, sc->nodemask) {
2663                 /*
2664                  * Take care memory controller reclaiming has small influence
2665                  * to global LRU.
2666                  */
2667                 if (global_reclaim(sc)) {
2668                         if (!cpuset_zone_allowed(zone,
2669                                                  GFP_KERNEL | __GFP_HARDWALL))
2670                                 continue;
2671
2672                         /*
2673                          * If we already have plenty of memory free for
2674                          * compaction in this zone, don't free any more.
2675                          * Even though compaction is invoked for any
2676                          * non-zero order, only frequent costly order
2677                          * reclamation is disruptive enough to become a
2678                          * noticeable problem, like transparent huge
2679                          * page allocations.
2680                          */
2681                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2682                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2683                             compaction_ready(zone, sc)) {
2684                                 sc->compaction_ready = true;
2685                                 continue;
2686                         }
2687
2688                         /*
2689                          * Shrink each node in the zonelist once. If the
2690                          * zonelist is ordered by zone (not the default) then a
2691                          * node may be shrunk multiple times but in that case
2692                          * the user prefers lower zones being preserved.
2693                          */
2694                         if (zone->zone_pgdat == last_pgdat)
2695                                 continue;
2696
2697                         /*
2698                          * This steals pages from memory cgroups over softlimit
2699                          * and returns the number of reclaimed pages and
2700                          * scanned pages. This works for global memory pressure
2701                          * and balancing, not for a memcg's limit.
2702                          */
2703                         nr_soft_scanned = 0;
2704                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2705                                                 sc->order, sc->gfp_mask,
2706                                                 &nr_soft_scanned);
2707                         sc->nr_reclaimed += nr_soft_reclaimed;
2708                         sc->nr_scanned += nr_soft_scanned;
2709                         /* need some check for avoid more shrink_zone() */
2710                 }
2711
2712                 /* See comment about same check for global reclaim above */
2713                 if (zone->zone_pgdat == last_pgdat)
2714                         continue;
2715                 last_pgdat = zone->zone_pgdat;
2716                 shrink_node(zone->zone_pgdat, sc);
2717         }
2718
2719         /*
2720          * Restore to original mask to avoid the impact on the caller if we
2721          * promoted it to __GFP_HIGHMEM.
2722          */
2723         sc->gfp_mask = orig_mask;
2724 }
2725
2726 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2727 {
2728         struct mem_cgroup *memcg;
2729
2730         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2731         do {
2732                 unsigned long refaults;
2733                 struct lruvec *lruvec;
2734
2735                 if (memcg)
2736                         refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2737                 else
2738                         refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2739
2740                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2741                 lruvec->refaults = refaults;
2742         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2743 }
2744
2745 /*
2746  * This is the main entry point to direct page reclaim.
2747  *
2748  * If a full scan of the inactive list fails to free enough memory then we
2749  * are "out of memory" and something needs to be killed.
2750  *
2751  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2752  * high - the zone may be full of dirty or under-writeback pages, which this
2753  * caller can't do much about.  We kick the writeback threads and take explicit
2754  * naps in the hope that some of these pages can be written.  But if the
2755  * allocating task holds filesystem locks which prevent writeout this might not
2756  * work, and the allocation attempt will fail.
2757  *
2758  * returns:     0, if no pages reclaimed
2759  *              else, the number of pages reclaimed
2760  */
2761 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2762                                           struct scan_control *sc)
2763 {
2764         int initial_priority = sc->priority;
2765         pg_data_t *last_pgdat;
2766         struct zoneref *z;
2767         struct zone *zone;
2768 retry:
2769         delayacct_freepages_start();
2770
2771         if (global_reclaim(sc))
2772                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2773
2774         do {
2775                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2776                                 sc->priority);
2777                 sc->nr_scanned = 0;
2778                 shrink_zones(zonelist, sc);
2779
2780                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2781                         break;
2782
2783                 if (sc->compaction_ready)
2784                         break;
2785
2786                 /*
2787                  * If we're getting trouble reclaiming, start doing
2788                  * writepage even in laptop mode.
2789                  */
2790                 if (sc->priority < DEF_PRIORITY - 2)
2791                         sc->may_writepage = 1;
2792         } while (--sc->priority >= 0);
2793
2794         last_pgdat = NULL;
2795         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2796                                         sc->nodemask) {
2797                 if (zone->zone_pgdat == last_pgdat)
2798                         continue;
2799                 last_pgdat = zone->zone_pgdat;
2800                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2801         }
2802
2803         delayacct_freepages_end();
2804
2805         if (sc->nr_reclaimed)
2806                 return sc->nr_reclaimed;
2807
2808         /* Aborted reclaim to try compaction? don't OOM, then */
2809         if (sc->compaction_ready)
2810                 return 1;
2811
2812         /* Untapped cgroup reserves?  Don't OOM, retry. */
2813         if (sc->memcg_low_skipped) {
2814                 sc->priority = initial_priority;
2815                 sc->memcg_low_reclaim = 1;
2816                 sc->memcg_low_skipped = 0;
2817                 goto retry;
2818         }
2819
2820         return 0;
2821 }
2822
2823 static bool allow_direct_reclaim(pg_data_t *pgdat)
2824 {
2825         struct zone *zone;
2826         unsigned long pfmemalloc_reserve = 0;
2827         unsigned long free_pages = 0;
2828         int i;
2829         bool wmark_ok;
2830
2831         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2832                 return true;
2833
2834         for (i = 0; i <= ZONE_NORMAL; i++) {
2835                 zone = &pgdat->node_zones[i];
2836                 if (!managed_zone(zone))
2837                         continue;
2838
2839                 if (!zone_reclaimable_pages(zone))
2840                         continue;
2841
2842                 pfmemalloc_reserve += min_wmark_pages(zone);
2843                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2844         }
2845
2846         /* If there are no reserves (unexpected config) then do not throttle */
2847         if (!pfmemalloc_reserve)
2848                 return true;
2849
2850         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2851
2852         /* kswapd must be awake if processes are being throttled */
2853         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2854                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2855                                                 (enum zone_type)ZONE_NORMAL);
2856                 wake_up_interruptible(&pgdat->kswapd_wait);
2857         }
2858
2859         return wmark_ok;
2860 }
2861
2862 /*
2863  * Throttle direct reclaimers if backing storage is backed by the network
2864  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2865  * depleted. kswapd will continue to make progress and wake the processes
2866  * when the low watermark is reached.
2867  *
2868  * Returns true if a fatal signal was delivered during throttling. If this
2869  * happens, the page allocator should not consider triggering the OOM killer.
2870  */
2871 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2872                                         nodemask_t *nodemask)
2873 {
2874         struct zoneref *z;
2875         struct zone *zone;
2876         pg_data_t *pgdat = NULL;
2877
2878         /*
2879          * Kernel threads should not be throttled as they may be indirectly
2880          * responsible for cleaning pages necessary for reclaim to make forward
2881          * progress. kjournald for example may enter direct reclaim while
2882          * committing a transaction where throttling it could forcing other
2883          * processes to block on log_wait_commit().
2884          */
2885         if (current->flags & PF_KTHREAD)
2886                 goto out;
2887
2888         /*
2889          * If a fatal signal is pending, this process should not throttle.
2890          * It should return quickly so it can exit and free its memory
2891          */
2892         if (fatal_signal_pending(current))
2893                 goto out;
2894
2895         /*
2896          * Check if the pfmemalloc reserves are ok by finding the first node
2897          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2898          * GFP_KERNEL will be required for allocating network buffers when
2899          * swapping over the network so ZONE_HIGHMEM is unusable.
2900          *
2901          * Throttling is based on the first usable node and throttled processes
2902          * wait on a queue until kswapd makes progress and wakes them. There
2903          * is an affinity then between processes waking up and where reclaim
2904          * progress has been made assuming the process wakes on the same node.
2905          * More importantly, processes running on remote nodes will not compete
2906          * for remote pfmemalloc reserves and processes on different nodes
2907          * should make reasonable progress.
2908          */
2909         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2910                                         gfp_zone(gfp_mask), nodemask) {
2911                 if (zone_idx(zone) > ZONE_NORMAL)
2912                         continue;
2913
2914                 /* Throttle based on the first usable node */
2915                 pgdat = zone->zone_pgdat;
2916                 if (allow_direct_reclaim(pgdat))
2917                         goto out;
2918                 break;
2919         }
2920
2921         /* If no zone was usable by the allocation flags then do not throttle */
2922         if (!pgdat)
2923                 goto out;
2924
2925         /* Account for the throttling */
2926         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2927
2928         /*
2929          * If the caller cannot enter the filesystem, it's possible that it
2930          * is due to the caller holding an FS lock or performing a journal
2931          * transaction in the case of a filesystem like ext[3|4]. In this case,
2932          * it is not safe to block on pfmemalloc_wait as kswapd could be
2933          * blocked waiting on the same lock. Instead, throttle for up to a
2934          * second before continuing.
2935          */
2936         if (!(gfp_mask & __GFP_FS)) {
2937                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2938                         allow_direct_reclaim(pgdat), HZ);
2939
2940                 goto check_pending;
2941         }
2942
2943         /* Throttle until kswapd wakes the process */
2944         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2945                 allow_direct_reclaim(pgdat));
2946
2947 check_pending:
2948         if (fatal_signal_pending(current))
2949                 return true;
2950
2951 out:
2952         return false;
2953 }
2954
2955 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2956                                 gfp_t gfp_mask, nodemask_t *nodemask)
2957 {
2958         unsigned long nr_reclaimed;
2959         struct scan_control sc = {
2960                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2961                 .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)),
2962                 .reclaim_idx = gfp_zone(gfp_mask),
2963                 .order = order,
2964                 .nodemask = nodemask,
2965                 .priority = DEF_PRIORITY,
2966                 .may_writepage = !laptop_mode,
2967                 .may_unmap = 1,
2968                 .may_swap = 1,
2969         };
2970
2971         /*
2972          * Do not enter reclaim if fatal signal was delivered while throttled.
2973          * 1 is returned so that the page allocator does not OOM kill at this
2974          * point.
2975          */
2976         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2977                 return 1;
2978
2979         trace_mm_vmscan_direct_reclaim_begin(order,
2980                                 sc.may_writepage,
2981                                 gfp_mask,
2982                                 sc.reclaim_idx);
2983
2984         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2985
2986         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2987
2988         return nr_reclaimed;
2989 }
2990
2991 #ifdef CONFIG_MEMCG
2992
2993 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2994                                                 gfp_t gfp_mask, bool noswap,
2995                                                 pg_data_t *pgdat,
2996                                                 unsigned long *nr_scanned)
2997 {
2998         struct scan_control sc = {
2999                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3000                 .target_mem_cgroup = memcg,
3001                 .may_writepage = !laptop_mode,
3002                 .may_unmap = 1,
3003                 .reclaim_idx = MAX_NR_ZONES - 1,
3004                 .may_swap = !noswap,
3005         };
3006         unsigned long lru_pages;
3007
3008         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3009                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3010
3011         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3012                                                       sc.may_writepage,
3013                                                       sc.gfp_mask,
3014                                                       sc.reclaim_idx);
3015
3016         /*
3017          * NOTE: Although we can get the priority field, using it
3018          * here is not a good idea, since it limits the pages we can scan.
3019          * if we don't reclaim here, the shrink_node from balance_pgdat
3020          * will pick up pages from other mem cgroup's as well. We hack
3021          * the priority and make it zero.
3022          */
3023         shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3024
3025         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3026
3027         *nr_scanned = sc.nr_scanned;
3028         return sc.nr_reclaimed;
3029 }
3030
3031 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3032                                            unsigned long nr_pages,
3033                                            gfp_t gfp_mask,
3034                                            bool may_swap)
3035 {
3036         struct zonelist *zonelist;
3037         unsigned long nr_reclaimed;
3038         int nid;
3039         unsigned int noreclaim_flag;
3040         struct scan_control sc = {
3041                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3042                 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3043                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3044                 .reclaim_idx = MAX_NR_ZONES - 1,
3045                 .target_mem_cgroup = memcg,
3046                 .priority = DEF_PRIORITY,
3047                 .may_writepage = !laptop_mode,
3048                 .may_unmap = 1,
3049                 .may_swap = may_swap,
3050         };
3051
3052         /*
3053          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3054          * take care of from where we get pages. So the node where we start the
3055          * scan does not need to be the current node.
3056          */
3057         nid = mem_cgroup_select_victim_node(memcg);
3058
3059         zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3060
3061         trace_mm_vmscan_memcg_reclaim_begin(0,
3062                                             sc.may_writepage,
3063                                             sc.gfp_mask,
3064                                             sc.reclaim_idx);
3065
3066         noreclaim_flag = memalloc_noreclaim_save();
3067         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3068         memalloc_noreclaim_restore(noreclaim_flag);
3069
3070         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3071
3072         return nr_reclaimed;
3073 }
3074 #endif
3075
3076 static void age_active_anon(struct pglist_data *pgdat,
3077                                 struct scan_control *sc)
3078 {
3079         struct mem_cgroup *memcg;
3080
3081         if (!total_swap_pages)
3082                 return;
3083
3084         memcg = mem_cgroup_iter(NULL, NULL, NULL);
3085         do {
3086                 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3087
3088                 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3089                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3090                                            sc, LRU_ACTIVE_ANON);
3091
3092                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3093         } while (memcg);
3094 }
3095
3096 /*
3097  * Returns true if there is an eligible zone balanced for the request order
3098  * and classzone_idx
3099  */
3100 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3101 {
3102         int i;
3103         unsigned long mark = -1;
3104         struct zone *zone;
3105
3106         for (i = 0; i <= classzone_idx; i++) {
3107                 zone = pgdat->node_zones + i;
3108
3109                 if (!managed_zone(zone))
3110                         continue;
3111
3112                 mark = high_wmark_pages(zone);
3113                 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3114                         return true;
3115         }
3116
3117         /*
3118          * If a node has no populated zone within classzone_idx, it does not
3119          * need balancing by definition. This can happen if a zone-restricted
3120          * allocation tries to wake a remote kswapd.
3121          */
3122         if (mark == -1)
3123                 return true;
3124
3125         return false;
3126 }
3127
3128 /* Clear pgdat state for congested, dirty or under writeback. */
3129 static void clear_pgdat_congested(pg_data_t *pgdat)
3130 {
3131         clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3132         clear_bit(PGDAT_DIRTY, &pgdat->flags);
3133         clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3134 }
3135
3136 /*
3137  * Prepare kswapd for sleeping. This verifies that there are no processes
3138  * waiting in throttle_direct_reclaim() and that watermarks have been met.
3139  *
3140  * Returns true if kswapd is ready to sleep
3141  */
3142 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3143 {
3144         /*
3145          * The throttled processes are normally woken up in balance_pgdat() as
3146          * soon as allow_direct_reclaim() is true. But there is a potential
3147          * race between when kswapd checks the watermarks and a process gets
3148          * throttled. There is also a potential race if processes get
3149          * throttled, kswapd wakes, a large process exits thereby balancing the
3150          * zones, which causes kswapd to exit balance_pgdat() before reaching
3151          * the wake up checks. If kswapd is going to sleep, no process should
3152          * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3153          * the wake up is premature, processes will wake kswapd and get
3154          * throttled again. The difference from wake ups in balance_pgdat() is
3155          * that here we are under prepare_to_wait().
3156          */
3157         if (waitqueue_active(&pgdat->pfmemalloc_wait))
3158                 wake_up_all(&pgdat->pfmemalloc_wait);
3159
3160         /* Hopeless node, leave it to direct reclaim */
3161         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3162                 return true;
3163
3164         if (pgdat_balanced(pgdat, order, classzone_idx)) {
3165                 clear_pgdat_congested(pgdat);
3166                 return true;
3167         }
3168
3169         return false;
3170 }
3171
3172 /*
3173  * kswapd shrinks a node of pages that are at or below the highest usable
3174  * zone that is currently unbalanced.
3175  *
3176  * Returns true if kswapd scanned at least the requested number of pages to
3177  * reclaim or if the lack of progress was due to pages under writeback.
3178  * This is used to determine if the scanning priority needs to be raised.
3179  */
3180 static bool kswapd_shrink_node(pg_data_t *pgdat,
3181                                struct scan_control *sc)
3182 {
3183         struct zone *zone;
3184         int z;