mm,vmscan: don't pretend forward progress upon shrinker_rwsem contention
[muen/linux.git] / mm / vmscan.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/mm/vmscan.c
4  *
5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
6  *
7  *  Swap reorganised 29.12.95, Stephen Tweedie.
8  *  kswapd added: 7.1.96  sct
9  *  Removed kswapd_ctl limits, and swap out as many pages as needed
10  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12  *  Multiqueue VM started 5.8.00, Rik van Riel.
13  */
14
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
17 #include <linux/mm.h>
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h>  /* for try_to_release_page(),
32                                         buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
55
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
58
59 #include "internal.h"
60
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
63
64 struct scan_control {
65         /* How many pages shrink_list() should reclaim */
66         unsigned long nr_to_reclaim;
67
68         /* This context's GFP mask */
69         gfp_t gfp_mask;
70
71         /* Allocation order */
72         int order;
73
74         /*
75          * Nodemask of nodes allowed by the caller. If NULL, all nodes
76          * are scanned.
77          */
78         nodemask_t      *nodemask;
79
80         /*
81          * The memory cgroup that hit its limit and as a result is the
82          * primary target of this reclaim invocation.
83          */
84         struct mem_cgroup *target_mem_cgroup;
85
86         /* Scan (total_size >> priority) pages at once */
87         int priority;
88
89         /* The highest zone to isolate pages for reclaim from */
90         enum zone_type reclaim_idx;
91
92         /* Writepage batching in laptop mode; RECLAIM_WRITE */
93         unsigned int may_writepage:1;
94
95         /* Can mapped pages be reclaimed? */
96         unsigned int may_unmap:1;
97
98         /* Can pages be swapped as part of reclaim? */
99         unsigned int may_swap:1;
100
101         /*
102          * Cgroups are not reclaimed below their configured memory.low,
103          * unless we threaten to OOM. If any cgroups are skipped due to
104          * memory.low and nothing was reclaimed, go back for memory.low.
105          */
106         unsigned int memcg_low_reclaim:1;
107         unsigned int memcg_low_skipped:1;
108
109         unsigned int hibernation_mode:1;
110
111         /* One of the zones is ready for compaction */
112         unsigned int compaction_ready:1;
113
114         /* Incremented by the number of inactive pages that were scanned */
115         unsigned long nr_scanned;
116
117         /* Number of pages freed so far during a call to shrink_zones() */
118         unsigned long nr_reclaimed;
119 };
120
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field)                    \
123         do {                                                            \
124                 if ((_page)->lru.prev != _base) {                       \
125                         struct page *prev;                              \
126                                                                         \
127                         prev = lru_to_page(&(_page->lru));              \
128                         prefetch(&prev->_field);                        \
129                 }                                                       \
130         } while (0)
131 #else
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
133 #endif
134
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
137         do {                                                            \
138                 if ((_page)->lru.prev != _base) {                       \
139                         struct page *prev;                              \
140                                                                         \
141                         prev = lru_to_page(&(_page->lru));              \
142                         prefetchw(&prev->_field);                       \
143                 }                                                       \
144         } while (0)
145 #else
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 #endif
148
149 /*
150  * From 0 .. 100.  Higher means more swappy.
151  */
152 int vm_swappiness = 60;
153 /*
154  * The total number of pages which are beyond the high watermark within all
155  * zones.
156  */
157 unsigned long vm_total_pages;
158
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
161
162 #ifdef CONFIG_MEMCG
163 static bool global_reclaim(struct scan_control *sc)
164 {
165         return !sc->target_mem_cgroup;
166 }
167
168 /**
169  * sane_reclaim - is the usual dirty throttling mechanism operational?
170  * @sc: scan_control in question
171  *
172  * The normal page dirty throttling mechanism in balance_dirty_pages() is
173  * completely broken with the legacy memcg and direct stalling in
174  * shrink_page_list() is used for throttling instead, which lacks all the
175  * niceties such as fairness, adaptive pausing, bandwidth proportional
176  * allocation and configurability.
177  *
178  * This function tests whether the vmscan currently in progress can assume
179  * that the normal dirty throttling mechanism is operational.
180  */
181 static bool sane_reclaim(struct scan_control *sc)
182 {
183         struct mem_cgroup *memcg = sc->target_mem_cgroup;
184
185         if (!memcg)
186                 return true;
187 #ifdef CONFIG_CGROUP_WRITEBACK
188         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
189                 return true;
190 #endif
191         return false;
192 }
193 #else
194 static bool global_reclaim(struct scan_control *sc)
195 {
196         return true;
197 }
198
199 static bool sane_reclaim(struct scan_control *sc)
200 {
201         return true;
202 }
203 #endif
204
205 /*
206  * This misses isolated pages which are not accounted for to save counters.
207  * As the data only determines if reclaim or compaction continues, it is
208  * not expected that isolated pages will be a dominating factor.
209  */
210 unsigned long zone_reclaimable_pages(struct zone *zone)
211 {
212         unsigned long nr;
213
214         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
215                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
216         if (get_nr_swap_pages() > 0)
217                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
218                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
219
220         return nr;
221 }
222
223 /**
224  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
225  * @lruvec: lru vector
226  * @lru: lru to use
227  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
228  */
229 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
230 {
231         unsigned long lru_size;
232         int zid;
233
234         if (!mem_cgroup_disabled())
235                 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
236         else
237                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
238
239         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
240                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
241                 unsigned long size;
242
243                 if (!managed_zone(zone))
244                         continue;
245
246                 if (!mem_cgroup_disabled())
247                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
248                 else
249                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
250                                        NR_ZONE_LRU_BASE + lru);
251                 lru_size -= min(size, lru_size);
252         }
253
254         return lru_size;
255
256 }
257
258 /*
259  * Add a shrinker callback to be called from the vm.
260  */
261 int register_shrinker(struct shrinker *shrinker)
262 {
263         size_t size = sizeof(*shrinker->nr_deferred);
264
265         if (shrinker->flags & SHRINKER_NUMA_AWARE)
266                 size *= nr_node_ids;
267
268         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
269         if (!shrinker->nr_deferred)
270                 return -ENOMEM;
271
272         down_write(&shrinker_rwsem);
273         list_add_tail(&shrinker->list, &shrinker_list);
274         up_write(&shrinker_rwsem);
275         return 0;
276 }
277 EXPORT_SYMBOL(register_shrinker);
278
279 /*
280  * Remove one
281  */
282 void unregister_shrinker(struct shrinker *shrinker)
283 {
284         if (!shrinker->nr_deferred)
285                 return;
286         down_write(&shrinker_rwsem);
287         list_del(&shrinker->list);
288         up_write(&shrinker_rwsem);
289         kfree(shrinker->nr_deferred);
290         shrinker->nr_deferred = NULL;
291 }
292 EXPORT_SYMBOL(unregister_shrinker);
293
294 #define SHRINK_BATCH 128
295
296 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
297                                     struct shrinker *shrinker, int priority)
298 {
299         unsigned long freed = 0;
300         unsigned long long delta;
301         long total_scan;
302         long freeable;
303         long nr;
304         long new_nr;
305         int nid = shrinkctl->nid;
306         long batch_size = shrinker->batch ? shrinker->batch
307                                           : SHRINK_BATCH;
308         long scanned = 0, next_deferred;
309
310         freeable = shrinker->count_objects(shrinker, shrinkctl);
311         if (freeable == 0)
312                 return 0;
313
314         /*
315          * copy the current shrinker scan count into a local variable
316          * and zero it so that other concurrent shrinker invocations
317          * don't also do this scanning work.
318          */
319         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
320
321         total_scan = nr;
322         delta = freeable >> priority;
323         delta *= 4;
324         do_div(delta, shrinker->seeks);
325         total_scan += delta;
326         if (total_scan < 0) {
327                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
328                        shrinker->scan_objects, total_scan);
329                 total_scan = freeable;
330                 next_deferred = nr;
331         } else
332                 next_deferred = total_scan;
333
334         /*
335          * We need to avoid excessive windup on filesystem shrinkers
336          * due to large numbers of GFP_NOFS allocations causing the
337          * shrinkers to return -1 all the time. This results in a large
338          * nr being built up so when a shrink that can do some work
339          * comes along it empties the entire cache due to nr >>>
340          * freeable. This is bad for sustaining a working set in
341          * memory.
342          *
343          * Hence only allow the shrinker to scan the entire cache when
344          * a large delta change is calculated directly.
345          */
346         if (delta < freeable / 4)
347                 total_scan = min(total_scan, freeable / 2);
348
349         /*
350          * Avoid risking looping forever due to too large nr value:
351          * never try to free more than twice the estimate number of
352          * freeable entries.
353          */
354         if (total_scan > freeable * 2)
355                 total_scan = freeable * 2;
356
357         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
358                                    freeable, delta, total_scan, priority);
359
360         /*
361          * Normally, we should not scan less than batch_size objects in one
362          * pass to avoid too frequent shrinker calls, but if the slab has less
363          * than batch_size objects in total and we are really tight on memory,
364          * we will try to reclaim all available objects, otherwise we can end
365          * up failing allocations although there are plenty of reclaimable
366          * objects spread over several slabs with usage less than the
367          * batch_size.
368          *
369          * We detect the "tight on memory" situations by looking at the total
370          * number of objects we want to scan (total_scan). If it is greater
371          * than the total number of objects on slab (freeable), we must be
372          * scanning at high prio and therefore should try to reclaim as much as
373          * possible.
374          */
375         while (total_scan >= batch_size ||
376                total_scan >= freeable) {
377                 unsigned long ret;
378                 unsigned long nr_to_scan = min(batch_size, total_scan);
379
380                 shrinkctl->nr_to_scan = nr_to_scan;
381                 shrinkctl->nr_scanned = nr_to_scan;
382                 ret = shrinker->scan_objects(shrinker, shrinkctl);
383                 if (ret == SHRINK_STOP)
384                         break;
385                 freed += ret;
386
387                 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
388                 total_scan -= shrinkctl->nr_scanned;
389                 scanned += shrinkctl->nr_scanned;
390
391                 cond_resched();
392         }
393
394         if (next_deferred >= scanned)
395                 next_deferred -= scanned;
396         else
397                 next_deferred = 0;
398         /*
399          * move the unused scan count back into the shrinker in a
400          * manner that handles concurrent updates. If we exhausted the
401          * scan, there is no need to do an update.
402          */
403         if (next_deferred > 0)
404                 new_nr = atomic_long_add_return(next_deferred,
405                                                 &shrinker->nr_deferred[nid]);
406         else
407                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
408
409         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
410         return freed;
411 }
412
413 /**
414  * shrink_slab - shrink slab caches
415  * @gfp_mask: allocation context
416  * @nid: node whose slab caches to target
417  * @memcg: memory cgroup whose slab caches to target
418  * @priority: the reclaim priority
419  *
420  * Call the shrink functions to age shrinkable caches.
421  *
422  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
423  * unaware shrinkers will receive a node id of 0 instead.
424  *
425  * @memcg specifies the memory cgroup to target. If it is not NULL,
426  * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
427  * objects from the memory cgroup specified. Otherwise, only unaware
428  * shrinkers are called.
429  *
430  * @priority is sc->priority, we take the number of objects and >> by priority
431  * in order to get the scan target.
432  *
433  * Returns the number of reclaimed slab objects.
434  */
435 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
436                                  struct mem_cgroup *memcg,
437                                  int priority)
438 {
439         struct shrinker *shrinker;
440         unsigned long freed = 0;
441
442         if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
443                 return 0;
444
445         if (!down_read_trylock(&shrinker_rwsem))
446                 goto out;
447
448         list_for_each_entry(shrinker, &shrinker_list, list) {
449                 struct shrink_control sc = {
450                         .gfp_mask = gfp_mask,
451                         .nid = nid,
452                         .memcg = memcg,
453                 };
454
455                 /*
456                  * If kernel memory accounting is disabled, we ignore
457                  * SHRINKER_MEMCG_AWARE flag and call all shrinkers
458                  * passing NULL for memcg.
459                  */
460                 if (memcg_kmem_enabled() &&
461                     !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
462                         continue;
463
464                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
465                         sc.nid = 0;
466
467                 freed += do_shrink_slab(&sc, shrinker, priority);
468                 /*
469                  * Bail out if someone want to register a new shrinker to
470                  * prevent the regsitration from being stalled for long periods
471                  * by parallel ongoing shrinking.
472                  */
473                 if (rwsem_is_contended(&shrinker_rwsem)) {
474                         freed = freed ? : 1;
475                         break;
476                 }
477         }
478
479         up_read(&shrinker_rwsem);
480 out:
481         cond_resched();
482         return freed;
483 }
484
485 void drop_slab_node(int nid)
486 {
487         unsigned long freed;
488
489         do {
490                 struct mem_cgroup *memcg = NULL;
491
492                 freed = 0;
493                 do {
494                         freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
495                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
496         } while (freed > 10);
497 }
498
499 void drop_slab(void)
500 {
501         int nid;
502
503         for_each_online_node(nid)
504                 drop_slab_node(nid);
505 }
506
507 static inline int is_page_cache_freeable(struct page *page)
508 {
509         /*
510          * A freeable page cache page is referenced only by the caller
511          * that isolated the page, the page cache radix tree and
512          * optional buffer heads at page->private.
513          */
514         int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
515                 HPAGE_PMD_NR : 1;
516         return page_count(page) - page_has_private(page) == 1 + radix_pins;
517 }
518
519 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
520 {
521         if (current->flags & PF_SWAPWRITE)
522                 return 1;
523         if (!inode_write_congested(inode))
524                 return 1;
525         if (inode_to_bdi(inode) == current->backing_dev_info)
526                 return 1;
527         return 0;
528 }
529
530 /*
531  * We detected a synchronous write error writing a page out.  Probably
532  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
533  * fsync(), msync() or close().
534  *
535  * The tricky part is that after writepage we cannot touch the mapping: nothing
536  * prevents it from being freed up.  But we have a ref on the page and once
537  * that page is locked, the mapping is pinned.
538  *
539  * We're allowed to run sleeping lock_page() here because we know the caller has
540  * __GFP_FS.
541  */
542 static void handle_write_error(struct address_space *mapping,
543                                 struct page *page, int error)
544 {
545         lock_page(page);
546         if (page_mapping(page) == mapping)
547                 mapping_set_error(mapping, error);
548         unlock_page(page);
549 }
550
551 /* possible outcome of pageout() */
552 typedef enum {
553         /* failed to write page out, page is locked */
554         PAGE_KEEP,
555         /* move page to the active list, page is locked */
556         PAGE_ACTIVATE,
557         /* page has been sent to the disk successfully, page is unlocked */
558         PAGE_SUCCESS,
559         /* page is clean and locked */
560         PAGE_CLEAN,
561 } pageout_t;
562
563 /*
564  * pageout is called by shrink_page_list() for each dirty page.
565  * Calls ->writepage().
566  */
567 static pageout_t pageout(struct page *page, struct address_space *mapping,
568                          struct scan_control *sc)
569 {
570         /*
571          * If the page is dirty, only perform writeback if that write
572          * will be non-blocking.  To prevent this allocation from being
573          * stalled by pagecache activity.  But note that there may be
574          * stalls if we need to run get_block().  We could test
575          * PagePrivate for that.
576          *
577          * If this process is currently in __generic_file_write_iter() against
578          * this page's queue, we can perform writeback even if that
579          * will block.
580          *
581          * If the page is swapcache, write it back even if that would
582          * block, for some throttling. This happens by accident, because
583          * swap_backing_dev_info is bust: it doesn't reflect the
584          * congestion state of the swapdevs.  Easy to fix, if needed.
585          */
586         if (!is_page_cache_freeable(page))
587                 return PAGE_KEEP;
588         if (!mapping) {
589                 /*
590                  * Some data journaling orphaned pages can have
591                  * page->mapping == NULL while being dirty with clean buffers.
592                  */
593                 if (page_has_private(page)) {
594                         if (try_to_free_buffers(page)) {
595                                 ClearPageDirty(page);
596                                 pr_info("%s: orphaned page\n", __func__);
597                                 return PAGE_CLEAN;
598                         }
599                 }
600                 return PAGE_KEEP;
601         }
602         if (mapping->a_ops->writepage == NULL)
603                 return PAGE_ACTIVATE;
604         if (!may_write_to_inode(mapping->host, sc))
605                 return PAGE_KEEP;
606
607         if (clear_page_dirty_for_io(page)) {
608                 int res;
609                 struct writeback_control wbc = {
610                         .sync_mode = WB_SYNC_NONE,
611                         .nr_to_write = SWAP_CLUSTER_MAX,
612                         .range_start = 0,
613                         .range_end = LLONG_MAX,
614                         .for_reclaim = 1,
615                 };
616
617                 SetPageReclaim(page);
618                 res = mapping->a_ops->writepage(page, &wbc);
619                 if (res < 0)
620                         handle_write_error(mapping, page, res);
621                 if (res == AOP_WRITEPAGE_ACTIVATE) {
622                         ClearPageReclaim(page);
623                         return PAGE_ACTIVATE;
624                 }
625
626                 if (!PageWriteback(page)) {
627                         /* synchronous write or broken a_ops? */
628                         ClearPageReclaim(page);
629                 }
630                 trace_mm_vmscan_writepage(page);
631                 inc_node_page_state(page, NR_VMSCAN_WRITE);
632                 return PAGE_SUCCESS;
633         }
634
635         return PAGE_CLEAN;
636 }
637
638 /*
639  * Same as remove_mapping, but if the page is removed from the mapping, it
640  * gets returned with a refcount of 0.
641  */
642 static int __remove_mapping(struct address_space *mapping, struct page *page,
643                             bool reclaimed)
644 {
645         unsigned long flags;
646         int refcount;
647
648         BUG_ON(!PageLocked(page));
649         BUG_ON(mapping != page_mapping(page));
650
651         spin_lock_irqsave(&mapping->tree_lock, flags);
652         /*
653          * The non racy check for a busy page.
654          *
655          * Must be careful with the order of the tests. When someone has
656          * a ref to the page, it may be possible that they dirty it then
657          * drop the reference. So if PageDirty is tested before page_count
658          * here, then the following race may occur:
659          *
660          * get_user_pages(&page);
661          * [user mapping goes away]
662          * write_to(page);
663          *                              !PageDirty(page)    [good]
664          * SetPageDirty(page);
665          * put_page(page);
666          *                              !page_count(page)   [good, discard it]
667          *
668          * [oops, our write_to data is lost]
669          *
670          * Reversing the order of the tests ensures such a situation cannot
671          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
672          * load is not satisfied before that of page->_refcount.
673          *
674          * Note that if SetPageDirty is always performed via set_page_dirty,
675          * and thus under tree_lock, then this ordering is not required.
676          */
677         if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
678                 refcount = 1 + HPAGE_PMD_NR;
679         else
680                 refcount = 2;
681         if (!page_ref_freeze(page, refcount))
682                 goto cannot_free;
683         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
684         if (unlikely(PageDirty(page))) {
685                 page_ref_unfreeze(page, refcount);
686                 goto cannot_free;
687         }
688
689         if (PageSwapCache(page)) {
690                 swp_entry_t swap = { .val = page_private(page) };
691                 mem_cgroup_swapout(page, swap);
692                 __delete_from_swap_cache(page);
693                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
694                 put_swap_page(page, swap);
695         } else {
696                 void (*freepage)(struct page *);
697                 void *shadow = NULL;
698
699                 freepage = mapping->a_ops->freepage;
700                 /*
701                  * Remember a shadow entry for reclaimed file cache in
702                  * order to detect refaults, thus thrashing, later on.
703                  *
704                  * But don't store shadows in an address space that is
705                  * already exiting.  This is not just an optizimation,
706                  * inode reclaim needs to empty out the radix tree or
707                  * the nodes are lost.  Don't plant shadows behind its
708                  * back.
709                  *
710                  * We also don't store shadows for DAX mappings because the
711                  * only page cache pages found in these are zero pages
712                  * covering holes, and because we don't want to mix DAX
713                  * exceptional entries and shadow exceptional entries in the
714                  * same page_tree.
715                  */
716                 if (reclaimed && page_is_file_cache(page) &&
717                     !mapping_exiting(mapping) && !dax_mapping(mapping))
718                         shadow = workingset_eviction(mapping, page);
719                 __delete_from_page_cache(page, shadow);
720                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
721
722                 if (freepage != NULL)
723                         freepage(page);
724         }
725
726         return 1;
727
728 cannot_free:
729         spin_unlock_irqrestore(&mapping->tree_lock, flags);
730         return 0;
731 }
732
733 /*
734  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
735  * someone else has a ref on the page, abort and return 0.  If it was
736  * successfully detached, return 1.  Assumes the caller has a single ref on
737  * this page.
738  */
739 int remove_mapping(struct address_space *mapping, struct page *page)
740 {
741         if (__remove_mapping(mapping, page, false)) {
742                 /*
743                  * Unfreezing the refcount with 1 rather than 2 effectively
744                  * drops the pagecache ref for us without requiring another
745                  * atomic operation.
746                  */
747                 page_ref_unfreeze(page, 1);
748                 return 1;
749         }
750         return 0;
751 }
752
753 /**
754  * putback_lru_page - put previously isolated page onto appropriate LRU list
755  * @page: page to be put back to appropriate lru list
756  *
757  * Add previously isolated @page to appropriate LRU list.
758  * Page may still be unevictable for other reasons.
759  *
760  * lru_lock must not be held, interrupts must be enabled.
761  */
762 void putback_lru_page(struct page *page)
763 {
764         lru_cache_add(page);
765         put_page(page);         /* drop ref from isolate */
766 }
767
768 enum page_references {
769         PAGEREF_RECLAIM,
770         PAGEREF_RECLAIM_CLEAN,
771         PAGEREF_KEEP,
772         PAGEREF_ACTIVATE,
773 };
774
775 static enum page_references page_check_references(struct page *page,
776                                                   struct scan_control *sc)
777 {
778         int referenced_ptes, referenced_page;
779         unsigned long vm_flags;
780
781         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
782                                           &vm_flags);
783         referenced_page = TestClearPageReferenced(page);
784
785         /*
786          * Mlock lost the isolation race with us.  Let try_to_unmap()
787          * move the page to the unevictable list.
788          */
789         if (vm_flags & VM_LOCKED)
790                 return PAGEREF_RECLAIM;
791
792         if (referenced_ptes) {
793                 if (PageSwapBacked(page))
794                         return PAGEREF_ACTIVATE;
795                 /*
796                  * All mapped pages start out with page table
797                  * references from the instantiating fault, so we need
798                  * to look twice if a mapped file page is used more
799                  * than once.
800                  *
801                  * Mark it and spare it for another trip around the
802                  * inactive list.  Another page table reference will
803                  * lead to its activation.
804                  *
805                  * Note: the mark is set for activated pages as well
806                  * so that recently deactivated but used pages are
807                  * quickly recovered.
808                  */
809                 SetPageReferenced(page);
810
811                 if (referenced_page || referenced_ptes > 1)
812                         return PAGEREF_ACTIVATE;
813
814                 /*
815                  * Activate file-backed executable pages after first usage.
816                  */
817                 if (vm_flags & VM_EXEC)
818                         return PAGEREF_ACTIVATE;
819
820                 return PAGEREF_KEEP;
821         }
822
823         /* Reclaim if clean, defer dirty pages to writeback */
824         if (referenced_page && !PageSwapBacked(page))
825                 return PAGEREF_RECLAIM_CLEAN;
826
827         return PAGEREF_RECLAIM;
828 }
829
830 /* Check if a page is dirty or under writeback */
831 static void page_check_dirty_writeback(struct page *page,
832                                        bool *dirty, bool *writeback)
833 {
834         struct address_space *mapping;
835
836         /*
837          * Anonymous pages are not handled by flushers and must be written
838          * from reclaim context. Do not stall reclaim based on them
839          */
840         if (!page_is_file_cache(page) ||
841             (PageAnon(page) && !PageSwapBacked(page))) {
842                 *dirty = false;
843                 *writeback = false;
844                 return;
845         }
846
847         /* By default assume that the page flags are accurate */
848         *dirty = PageDirty(page);
849         *writeback = PageWriteback(page);
850
851         /* Verify dirty/writeback state if the filesystem supports it */
852         if (!page_has_private(page))
853                 return;
854
855         mapping = page_mapping(page);
856         if (mapping && mapping->a_ops->is_dirty_writeback)
857                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
858 }
859
860 struct reclaim_stat {
861         unsigned nr_dirty;
862         unsigned nr_unqueued_dirty;
863         unsigned nr_congested;
864         unsigned nr_writeback;
865         unsigned nr_immediate;
866         unsigned nr_activate;
867         unsigned nr_ref_keep;
868         unsigned nr_unmap_fail;
869 };
870
871 /*
872  * shrink_page_list() returns the number of reclaimed pages
873  */
874 static unsigned long shrink_page_list(struct list_head *page_list,
875                                       struct pglist_data *pgdat,
876                                       struct scan_control *sc,
877                                       enum ttu_flags ttu_flags,
878                                       struct reclaim_stat *stat,
879                                       bool force_reclaim)
880 {
881         LIST_HEAD(ret_pages);
882         LIST_HEAD(free_pages);
883         int pgactivate = 0;
884         unsigned nr_unqueued_dirty = 0;
885         unsigned nr_dirty = 0;
886         unsigned nr_congested = 0;
887         unsigned nr_reclaimed = 0;
888         unsigned nr_writeback = 0;
889         unsigned nr_immediate = 0;
890         unsigned nr_ref_keep = 0;
891         unsigned nr_unmap_fail = 0;
892
893         cond_resched();
894
895         while (!list_empty(page_list)) {
896                 struct address_space *mapping;
897                 struct page *page;
898                 int may_enter_fs;
899                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
900                 bool dirty, writeback;
901
902                 cond_resched();
903
904                 page = lru_to_page(page_list);
905                 list_del(&page->lru);
906
907                 if (!trylock_page(page))
908                         goto keep;
909
910                 VM_BUG_ON_PAGE(PageActive(page), page);
911
912                 sc->nr_scanned++;
913
914                 if (unlikely(!page_evictable(page)))
915                         goto activate_locked;
916
917                 if (!sc->may_unmap && page_mapped(page))
918                         goto keep_locked;
919
920                 /* Double the slab pressure for mapped and swapcache pages */
921                 if ((page_mapped(page) || PageSwapCache(page)) &&
922                     !(PageAnon(page) && !PageSwapBacked(page)))
923                         sc->nr_scanned++;
924
925                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
926                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
927
928                 /*
929                  * The number of dirty pages determines if a zone is marked
930                  * reclaim_congested which affects wait_iff_congested. kswapd
931                  * will stall and start writing pages if the tail of the LRU
932                  * is all dirty unqueued pages.
933                  */
934                 page_check_dirty_writeback(page, &dirty, &writeback);
935                 if (dirty || writeback)
936                         nr_dirty++;
937
938                 if (dirty && !writeback)
939                         nr_unqueued_dirty++;
940
941                 /*
942                  * Treat this page as congested if the underlying BDI is or if
943                  * pages are cycling through the LRU so quickly that the
944                  * pages marked for immediate reclaim are making it to the
945                  * end of the LRU a second time.
946                  */
947                 mapping = page_mapping(page);
948                 if (((dirty || writeback) && mapping &&
949                      inode_write_congested(mapping->host)) ||
950                     (writeback && PageReclaim(page)))
951                         nr_congested++;
952
953                 /*
954                  * If a page at the tail of the LRU is under writeback, there
955                  * are three cases to consider.
956                  *
957                  * 1) If reclaim is encountering an excessive number of pages
958                  *    under writeback and this page is both under writeback and
959                  *    PageReclaim then it indicates that pages are being queued
960                  *    for IO but are being recycled through the LRU before the
961                  *    IO can complete. Waiting on the page itself risks an
962                  *    indefinite stall if it is impossible to writeback the
963                  *    page due to IO error or disconnected storage so instead
964                  *    note that the LRU is being scanned too quickly and the
965                  *    caller can stall after page list has been processed.
966                  *
967                  * 2) Global or new memcg reclaim encounters a page that is
968                  *    not marked for immediate reclaim, or the caller does not
969                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
970                  *    not to fs). In this case mark the page for immediate
971                  *    reclaim and continue scanning.
972                  *
973                  *    Require may_enter_fs because we would wait on fs, which
974                  *    may not have submitted IO yet. And the loop driver might
975                  *    enter reclaim, and deadlock if it waits on a page for
976                  *    which it is needed to do the write (loop masks off
977                  *    __GFP_IO|__GFP_FS for this reason); but more thought
978                  *    would probably show more reasons.
979                  *
980                  * 3) Legacy memcg encounters a page that is already marked
981                  *    PageReclaim. memcg does not have any dirty pages
982                  *    throttling so we could easily OOM just because too many
983                  *    pages are in writeback and there is nothing else to
984                  *    reclaim. Wait for the writeback to complete.
985                  *
986                  * In cases 1) and 2) we activate the pages to get them out of
987                  * the way while we continue scanning for clean pages on the
988                  * inactive list and refilling from the active list. The
989                  * observation here is that waiting for disk writes is more
990                  * expensive than potentially causing reloads down the line.
991                  * Since they're marked for immediate reclaim, they won't put
992                  * memory pressure on the cache working set any longer than it
993                  * takes to write them to disk.
994                  */
995                 if (PageWriteback(page)) {
996                         /* Case 1 above */
997                         if (current_is_kswapd() &&
998                             PageReclaim(page) &&
999                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1000                                 nr_immediate++;
1001                                 goto activate_locked;
1002
1003                         /* Case 2 above */
1004                         } else if (sane_reclaim(sc) ||
1005                             !PageReclaim(page) || !may_enter_fs) {
1006                                 /*
1007                                  * This is slightly racy - end_page_writeback()
1008                                  * might have just cleared PageReclaim, then
1009                                  * setting PageReclaim here end up interpreted
1010                                  * as PageReadahead - but that does not matter
1011                                  * enough to care.  What we do want is for this
1012                                  * page to have PageReclaim set next time memcg
1013                                  * reclaim reaches the tests above, so it will
1014                                  * then wait_on_page_writeback() to avoid OOM;
1015                                  * and it's also appropriate in global reclaim.
1016                                  */
1017                                 SetPageReclaim(page);
1018                                 nr_writeback++;
1019                                 goto activate_locked;
1020
1021                         /* Case 3 above */
1022                         } else {
1023                                 unlock_page(page);
1024                                 wait_on_page_writeback(page);
1025                                 /* then go back and try same page again */
1026                                 list_add_tail(&page->lru, page_list);
1027                                 continue;
1028                         }
1029                 }
1030
1031                 if (!force_reclaim)
1032                         references = page_check_references(page, sc);
1033
1034                 switch (references) {
1035                 case PAGEREF_ACTIVATE:
1036                         goto activate_locked;
1037                 case PAGEREF_KEEP:
1038                         nr_ref_keep++;
1039                         goto keep_locked;
1040                 case PAGEREF_RECLAIM:
1041                 case PAGEREF_RECLAIM_CLEAN:
1042                         ; /* try to reclaim the page below */
1043                 }
1044
1045                 /*
1046                  * Anonymous process memory has backing store?
1047                  * Try to allocate it some swap space here.
1048                  * Lazyfree page could be freed directly
1049                  */
1050                 if (PageAnon(page) && PageSwapBacked(page)) {
1051                         if (!PageSwapCache(page)) {
1052                                 if (!(sc->gfp_mask & __GFP_IO))
1053                                         goto keep_locked;
1054                                 if (PageTransHuge(page)) {
1055                                         /* cannot split THP, skip it */
1056                                         if (!can_split_huge_page(page, NULL))
1057                                                 goto activate_locked;
1058                                         /*
1059                                          * Split pages without a PMD map right
1060                                          * away. Chances are some or all of the
1061                                          * tail pages can be freed without IO.
1062                                          */
1063                                         if (!compound_mapcount(page) &&
1064                                             split_huge_page_to_list(page,
1065                                                                     page_list))
1066                                                 goto activate_locked;
1067                                 }
1068                                 if (!add_to_swap(page)) {
1069                                         if (!PageTransHuge(page))
1070                                                 goto activate_locked;
1071                                         /* Fallback to swap normal pages */
1072                                         if (split_huge_page_to_list(page,
1073                                                                     page_list))
1074                                                 goto activate_locked;
1075 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1076                                         count_vm_event(THP_SWPOUT_FALLBACK);
1077 #endif
1078                                         if (!add_to_swap(page))
1079                                                 goto activate_locked;
1080                                 }
1081
1082                                 may_enter_fs = 1;
1083
1084                                 /* Adding to swap updated mapping */
1085                                 mapping = page_mapping(page);
1086                         }
1087                 } else if (unlikely(PageTransHuge(page))) {
1088                         /* Split file THP */
1089                         if (split_huge_page_to_list(page, page_list))
1090                                 goto keep_locked;
1091                 }
1092
1093                 /*
1094                  * The page is mapped into the page tables of one or more
1095                  * processes. Try to unmap it here.
1096                  */
1097                 if (page_mapped(page)) {
1098                         enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1099
1100                         if (unlikely(PageTransHuge(page)))
1101                                 flags |= TTU_SPLIT_HUGE_PMD;
1102                         if (!try_to_unmap(page, flags)) {
1103                                 nr_unmap_fail++;
1104                                 goto activate_locked;
1105                         }
1106                 }
1107
1108                 if (PageDirty(page)) {
1109                         /*
1110                          * Only kswapd can writeback filesystem pages
1111                          * to avoid risk of stack overflow. But avoid
1112                          * injecting inefficient single-page IO into
1113                          * flusher writeback as much as possible: only
1114                          * write pages when we've encountered many
1115                          * dirty pages, and when we've already scanned
1116                          * the rest of the LRU for clean pages and see
1117                          * the same dirty pages again (PageReclaim).
1118                          */
1119                         if (page_is_file_cache(page) &&
1120                             (!current_is_kswapd() || !PageReclaim(page) ||
1121                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1122                                 /*
1123                                  * Immediately reclaim when written back.
1124                                  * Similar in principal to deactivate_page()
1125                                  * except we already have the page isolated
1126                                  * and know it's dirty
1127                                  */
1128                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1129                                 SetPageReclaim(page);
1130
1131                                 goto activate_locked;
1132                         }
1133
1134                         if (references == PAGEREF_RECLAIM_CLEAN)
1135                                 goto keep_locked;
1136                         if (!may_enter_fs)
1137                                 goto keep_locked;
1138                         if (!sc->may_writepage)
1139                                 goto keep_locked;
1140
1141                         /*
1142                          * Page is dirty. Flush the TLB if a writable entry
1143                          * potentially exists to avoid CPU writes after IO
1144                          * starts and then write it out here.
1145                          */
1146                         try_to_unmap_flush_dirty();
1147                         switch (pageout(page, mapping, sc)) {
1148                         case PAGE_KEEP:
1149                                 goto keep_locked;
1150                         case PAGE_ACTIVATE:
1151                                 goto activate_locked;
1152                         case PAGE_SUCCESS:
1153                                 if (PageWriteback(page))
1154                                         goto keep;
1155                                 if (PageDirty(page))
1156                                         goto keep;
1157
1158                                 /*
1159                                  * A synchronous write - probably a ramdisk.  Go
1160                                  * ahead and try to reclaim the page.
1161                                  */
1162                                 if (!trylock_page(page))
1163                                         goto keep;
1164                                 if (PageDirty(page) || PageWriteback(page))
1165                                         goto keep_locked;
1166                                 mapping = page_mapping(page);
1167                         case PAGE_CLEAN:
1168                                 ; /* try to free the page below */
1169                         }
1170                 }
1171
1172                 /*
1173                  * If the page has buffers, try to free the buffer mappings
1174                  * associated with this page. If we succeed we try to free
1175                  * the page as well.
1176                  *
1177                  * We do this even if the page is PageDirty().
1178                  * try_to_release_page() does not perform I/O, but it is
1179                  * possible for a page to have PageDirty set, but it is actually
1180                  * clean (all its buffers are clean).  This happens if the
1181                  * buffers were written out directly, with submit_bh(). ext3
1182                  * will do this, as well as the blockdev mapping.
1183                  * try_to_release_page() will discover that cleanness and will
1184                  * drop the buffers and mark the page clean - it can be freed.
1185                  *
1186                  * Rarely, pages can have buffers and no ->mapping.  These are
1187                  * the pages which were not successfully invalidated in
1188                  * truncate_complete_page().  We try to drop those buffers here
1189                  * and if that worked, and the page is no longer mapped into
1190                  * process address space (page_count == 1) it can be freed.
1191                  * Otherwise, leave the page on the LRU so it is swappable.
1192                  */
1193                 if (page_has_private(page)) {
1194                         if (!try_to_release_page(page, sc->gfp_mask))
1195                                 goto activate_locked;
1196                         if (!mapping && page_count(page) == 1) {
1197                                 unlock_page(page);
1198                                 if (put_page_testzero(page))
1199                                         goto free_it;
1200                                 else {
1201                                         /*
1202                                          * rare race with speculative reference.
1203                                          * the speculative reference will free
1204                                          * this page shortly, so we may
1205                                          * increment nr_reclaimed here (and
1206                                          * leave it off the LRU).
1207                                          */
1208                                         nr_reclaimed++;
1209                                         continue;
1210                                 }
1211                         }
1212                 }
1213
1214                 if (PageAnon(page) && !PageSwapBacked(page)) {
1215                         /* follow __remove_mapping for reference */
1216                         if (!page_ref_freeze(page, 1))
1217                                 goto keep_locked;
1218                         if (PageDirty(page)) {
1219                                 page_ref_unfreeze(page, 1);
1220                                 goto keep_locked;
1221                         }
1222
1223                         count_vm_event(PGLAZYFREED);
1224                         count_memcg_page_event(page, PGLAZYFREED);
1225                 } else if (!mapping || !__remove_mapping(mapping, page, true))
1226                         goto keep_locked;
1227                 /*
1228                  * At this point, we have no other references and there is
1229                  * no way to pick any more up (removed from LRU, removed
1230                  * from pagecache). Can use non-atomic bitops now (and
1231                  * we obviously don't have to worry about waking up a process
1232                  * waiting on the page lock, because there are no references.
1233                  */
1234                 __ClearPageLocked(page);
1235 free_it:
1236                 nr_reclaimed++;
1237
1238                 /*
1239                  * Is there need to periodically free_page_list? It would
1240                  * appear not as the counts should be low
1241                  */
1242                 if (unlikely(PageTransHuge(page))) {
1243                         mem_cgroup_uncharge(page);
1244                         (*get_compound_page_dtor(page))(page);
1245                 } else
1246                         list_add(&page->lru, &free_pages);
1247                 continue;
1248
1249 activate_locked:
1250                 /* Not a candidate for swapping, so reclaim swap space. */
1251                 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1252                                                 PageMlocked(page)))
1253                         try_to_free_swap(page);
1254                 VM_BUG_ON_PAGE(PageActive(page), page);
1255                 if (!PageMlocked(page)) {
1256                         SetPageActive(page);
1257                         pgactivate++;
1258                         count_memcg_page_event(page, PGACTIVATE);
1259                 }
1260 keep_locked:
1261                 unlock_page(page);
1262 keep:
1263                 list_add(&page->lru, &ret_pages);
1264                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1265         }
1266
1267         mem_cgroup_uncharge_list(&free_pages);
1268         try_to_unmap_flush();
1269         free_unref_page_list(&free_pages);
1270
1271         list_splice(&ret_pages, page_list);
1272         count_vm_events(PGACTIVATE, pgactivate);
1273
1274         if (stat) {
1275                 stat->nr_dirty = nr_dirty;
1276                 stat->nr_congested = nr_congested;
1277                 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1278                 stat->nr_writeback = nr_writeback;
1279                 stat->nr_immediate = nr_immediate;
1280                 stat->nr_activate = pgactivate;
1281                 stat->nr_ref_keep = nr_ref_keep;
1282                 stat->nr_unmap_fail = nr_unmap_fail;
1283         }
1284         return nr_reclaimed;
1285 }
1286
1287 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1288                                             struct list_head *page_list)
1289 {
1290         struct scan_control sc = {
1291                 .gfp_mask = GFP_KERNEL,
1292                 .priority = DEF_PRIORITY,
1293                 .may_unmap = 1,
1294         };
1295         unsigned long ret;
1296         struct page *page, *next;
1297         LIST_HEAD(clean_pages);
1298
1299         list_for_each_entry_safe(page, next, page_list, lru) {
1300                 if (page_is_file_cache(page) && !PageDirty(page) &&
1301                     !__PageMovable(page)) {
1302                         ClearPageActive(page);
1303                         list_move(&page->lru, &clean_pages);
1304                 }
1305         }
1306
1307         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1308                         TTU_IGNORE_ACCESS, NULL, true);
1309         list_splice(&clean_pages, page_list);
1310         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1311         return ret;
1312 }
1313
1314 /*
1315  * Attempt to remove the specified page from its LRU.  Only take this page
1316  * if it is of the appropriate PageActive status.  Pages which are being
1317  * freed elsewhere are also ignored.
1318  *
1319  * page:        page to consider
1320  * mode:        one of the LRU isolation modes defined above
1321  *
1322  * returns 0 on success, -ve errno on failure.
1323  */
1324 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1325 {
1326         int ret = -EINVAL;
1327
1328         /* Only take pages on the LRU. */
1329         if (!PageLRU(page))
1330                 return ret;
1331
1332         /* Compaction should not handle unevictable pages but CMA can do so */
1333         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1334                 return ret;
1335
1336         ret = -EBUSY;
1337
1338         /*
1339          * To minimise LRU disruption, the caller can indicate that it only
1340          * wants to isolate pages it will be able to operate on without
1341          * blocking - clean pages for the most part.
1342          *
1343          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1344          * that it is possible to migrate without blocking
1345          */
1346         if (mode & ISOLATE_ASYNC_MIGRATE) {
1347                 /* All the caller can do on PageWriteback is block */
1348                 if (PageWriteback(page))
1349                         return ret;
1350
1351                 if (PageDirty(page)) {
1352                         struct address_space *mapping;
1353                         bool migrate_dirty;
1354
1355                         /*
1356                          * Only pages without mappings or that have a
1357                          * ->migratepage callback are possible to migrate
1358                          * without blocking. However, we can be racing with
1359                          * truncation so it's necessary to lock the page
1360                          * to stabilise the mapping as truncation holds
1361                          * the page lock until after the page is removed
1362                          * from the page cache.
1363                          */
1364                         if (!trylock_page(page))
1365                                 return ret;
1366
1367                         mapping = page_mapping(page);
1368                         migrate_dirty = mapping && mapping->a_ops->migratepage;
1369                         unlock_page(page);
1370                         if (!migrate_dirty)
1371                                 return ret;
1372                 }
1373         }
1374
1375         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1376                 return ret;
1377
1378         if (likely(get_page_unless_zero(page))) {
1379                 /*
1380                  * Be careful not to clear PageLRU until after we're
1381                  * sure the page is not being freed elsewhere -- the
1382                  * page release code relies on it.
1383                  */
1384                 ClearPageLRU(page);
1385                 ret = 0;
1386         }
1387
1388         return ret;
1389 }
1390
1391
1392 /*
1393  * Update LRU sizes after isolating pages. The LRU size updates must
1394  * be complete before mem_cgroup_update_lru_size due to a santity check.
1395  */
1396 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1397                         enum lru_list lru, unsigned long *nr_zone_taken)
1398 {
1399         int zid;
1400
1401         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1402                 if (!nr_zone_taken[zid])
1403                         continue;
1404
1405                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1406 #ifdef CONFIG_MEMCG
1407                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1408 #endif
1409         }
1410
1411 }
1412
1413 /*
1414  * zone_lru_lock is heavily contended.  Some of the functions that
1415  * shrink the lists perform better by taking out a batch of pages
1416  * and working on them outside the LRU lock.
1417  *
1418  * For pagecache intensive workloads, this function is the hottest
1419  * spot in the kernel (apart from copy_*_user functions).
1420  *
1421  * Appropriate locks must be held before calling this function.
1422  *
1423  * @nr_to_scan: The number of eligible pages to look through on the list.
1424  * @lruvec:     The LRU vector to pull pages from.
1425  * @dst:        The temp list to put pages on to.
1426  * @nr_scanned: The number of pages that were scanned.
1427  * @sc:         The scan_control struct for this reclaim session
1428  * @mode:       One of the LRU isolation modes
1429  * @lru:        LRU list id for isolating
1430  *
1431  * returns how many pages were moved onto *@dst.
1432  */
1433 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1434                 struct lruvec *lruvec, struct list_head *dst,
1435                 unsigned long *nr_scanned, struct scan_control *sc,
1436                 isolate_mode_t mode, enum lru_list lru)
1437 {
1438         struct list_head *src = &lruvec->lists[lru];
1439         unsigned long nr_taken = 0;
1440         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1441         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1442         unsigned long skipped = 0;
1443         unsigned long scan, total_scan, nr_pages;
1444         LIST_HEAD(pages_skipped);
1445
1446         scan = 0;
1447         for (total_scan = 0;
1448              scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1449              total_scan++) {
1450                 struct page *page;
1451
1452                 page = lru_to_page(src);
1453                 prefetchw_prev_lru_page(page, src, flags);
1454
1455                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1456
1457                 if (page_zonenum(page) > sc->reclaim_idx) {
1458                         list_move(&page->lru, &pages_skipped);
1459                         nr_skipped[page_zonenum(page)]++;
1460                         continue;
1461                 }
1462
1463                 /*
1464                  * Do not count skipped pages because that makes the function
1465                  * return with no isolated pages if the LRU mostly contains
1466                  * ineligible pages.  This causes the VM to not reclaim any
1467                  * pages, triggering a premature OOM.
1468                  */
1469                 scan++;
1470                 switch (__isolate_lru_page(page, mode)) {
1471                 case 0:
1472                         nr_pages = hpage_nr_pages(page);
1473                         nr_taken += nr_pages;
1474                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1475                         list_move(&page->lru, dst);
1476                         break;
1477
1478                 case -EBUSY:
1479                         /* else it is being freed elsewhere */
1480                         list_move(&page->lru, src);
1481                         continue;
1482
1483                 default:
1484                         BUG();
1485                 }
1486         }
1487
1488         /*
1489          * Splice any skipped pages to the start of the LRU list. Note that
1490          * this disrupts the LRU order when reclaiming for lower zones but
1491          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1492          * scanning would soon rescan the same pages to skip and put the
1493          * system at risk of premature OOM.
1494          */
1495         if (!list_empty(&pages_skipped)) {
1496                 int zid;
1497
1498                 list_splice(&pages_skipped, src);
1499                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1500                         if (!nr_skipped[zid])
1501                                 continue;
1502
1503                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1504                         skipped += nr_skipped[zid];
1505                 }
1506         }
1507         *nr_scanned = total_scan;
1508         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1509                                     total_scan, skipped, nr_taken, mode, lru);
1510         update_lru_sizes(lruvec, lru, nr_zone_taken);
1511         return nr_taken;
1512 }
1513
1514 /**
1515  * isolate_lru_page - tries to isolate a page from its LRU list
1516  * @page: page to isolate from its LRU list
1517  *
1518  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1519  * vmstat statistic corresponding to whatever LRU list the page was on.
1520  *
1521  * Returns 0 if the page was removed from an LRU list.
1522  * Returns -EBUSY if the page was not on an LRU list.
1523  *
1524  * The returned page will have PageLRU() cleared.  If it was found on
1525  * the active list, it will have PageActive set.  If it was found on
1526  * the unevictable list, it will have the PageUnevictable bit set. That flag
1527  * may need to be cleared by the caller before letting the page go.
1528  *
1529  * The vmstat statistic corresponding to the list on which the page was
1530  * found will be decremented.
1531  *
1532  * Restrictions:
1533  *
1534  * (1) Must be called with an elevated refcount on the page. This is a
1535  *     fundamentnal difference from isolate_lru_pages (which is called
1536  *     without a stable reference).
1537  * (2) the lru_lock must not be held.
1538  * (3) interrupts must be enabled.
1539  */
1540 int isolate_lru_page(struct page *page)
1541 {
1542         int ret = -EBUSY;
1543
1544         VM_BUG_ON_PAGE(!page_count(page), page);
1545         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1546
1547         if (PageLRU(page)) {
1548                 struct zone *zone = page_zone(page);
1549                 struct lruvec *lruvec;
1550
1551                 spin_lock_irq(zone_lru_lock(zone));
1552                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1553                 if (PageLRU(page)) {
1554                         int lru = page_lru(page);
1555                         get_page(page);
1556                         ClearPageLRU(page);
1557                         del_page_from_lru_list(page, lruvec, lru);
1558                         ret = 0;
1559                 }
1560                 spin_unlock_irq(zone_lru_lock(zone));
1561         }
1562         return ret;
1563 }
1564
1565 /*
1566  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1567  * then get resheduled. When there are massive number of tasks doing page
1568  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1569  * the LRU list will go small and be scanned faster than necessary, leading to
1570  * unnecessary swapping, thrashing and OOM.
1571  */
1572 static int too_many_isolated(struct pglist_data *pgdat, int file,
1573                 struct scan_control *sc)
1574 {
1575         unsigned long inactive, isolated;
1576
1577         if (current_is_kswapd())
1578                 return 0;
1579
1580         if (!sane_reclaim(sc))
1581                 return 0;
1582
1583         if (file) {
1584                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1585                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1586         } else {
1587                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1588                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1589         }
1590
1591         /*
1592          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1593          * won't get blocked by normal direct-reclaimers, forming a circular
1594          * deadlock.
1595          */
1596         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1597                 inactive >>= 3;
1598
1599         return isolated > inactive;
1600 }
1601
1602 static noinline_for_stack void
1603 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1604 {
1605         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1606         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1607         LIST_HEAD(pages_to_free);
1608
1609         /*
1610          * Put back any unfreeable pages.
1611          */
1612         while (!list_empty(page_list)) {
1613                 struct page *page = lru_to_page(page_list);
1614                 int lru;
1615
1616                 VM_BUG_ON_PAGE(PageLRU(page), page);
1617                 list_del(&page->lru);
1618                 if (unlikely(!page_evictable(page))) {
1619                         spin_unlock_irq(&pgdat->lru_lock);
1620                         putback_lru_page(page);
1621                         spin_lock_irq(&pgdat->lru_lock);
1622                         continue;
1623                 }
1624
1625                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1626
1627                 SetPageLRU(page);
1628                 lru = page_lru(page);
1629                 add_page_to_lru_list(page, lruvec, lru);
1630
1631                 if (is_active_lru(lru)) {
1632                         int file = is_file_lru(lru);
1633                         int numpages = hpage_nr_pages(page);
1634                         reclaim_stat->recent_rotated[file] += numpages;
1635                 }
1636                 if (put_page_testzero(page)) {
1637                         __ClearPageLRU(page);
1638                         __ClearPageActive(page);
1639                         del_page_from_lru_list(page, lruvec, lru);
1640
1641                         if (unlikely(PageCompound(page))) {
1642                                 spin_unlock_irq(&pgdat->lru_lock);
1643                                 mem_cgroup_uncharge(page);
1644                                 (*get_compound_page_dtor(page))(page);
1645                                 spin_lock_irq(&pgdat->lru_lock);
1646                         } else
1647                                 list_add(&page->lru, &pages_to_free);
1648                 }
1649         }
1650
1651         /*
1652          * To save our caller's stack, now use input list for pages to free.
1653          */
1654         list_splice(&pages_to_free, page_list);
1655 }
1656
1657 /*
1658  * If a kernel thread (such as nfsd for loop-back mounts) services
1659  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1660  * In that case we should only throttle if the backing device it is
1661  * writing to is congested.  In other cases it is safe to throttle.
1662  */
1663 static int current_may_throttle(void)
1664 {
1665         return !(current->flags & PF_LESS_THROTTLE) ||
1666                 current->backing_dev_info == NULL ||
1667                 bdi_write_congested(current->backing_dev_info);
1668 }
1669
1670 /*
1671  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1672  * of reclaimed pages
1673  */
1674 static noinline_for_stack unsigned long
1675 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1676                      struct scan_control *sc, enum lru_list lru)
1677 {
1678         LIST_HEAD(page_list);
1679         unsigned long nr_scanned;
1680         unsigned long nr_reclaimed = 0;
1681         unsigned long nr_taken;
1682         struct reclaim_stat stat = {};
1683         isolate_mode_t isolate_mode = 0;
1684         int file = is_file_lru(lru);
1685         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1686         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1687         bool stalled = false;
1688
1689         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1690                 if (stalled)
1691                         return 0;
1692
1693                 /* wait a bit for the reclaimer. */
1694                 msleep(100);
1695                 stalled = true;
1696
1697                 /* We are about to die and free our memory. Return now. */
1698                 if (fatal_signal_pending(current))
1699                         return SWAP_CLUSTER_MAX;
1700         }
1701
1702         lru_add_drain();
1703
1704         if (!sc->may_unmap)
1705                 isolate_mode |= ISOLATE_UNMAPPED;
1706
1707         spin_lock_irq(&pgdat->lru_lock);
1708
1709         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1710                                      &nr_scanned, sc, isolate_mode, lru);
1711
1712         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1713         reclaim_stat->recent_scanned[file] += nr_taken;
1714
1715         if (current_is_kswapd()) {
1716                 if (global_reclaim(sc))
1717                         __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1718                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1719                                    nr_scanned);
1720         } else {
1721                 if (global_reclaim(sc))
1722                         __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1723                 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1724                                    nr_scanned);
1725         }
1726         spin_unlock_irq(&pgdat->lru_lock);
1727
1728         if (nr_taken == 0)
1729                 return 0;
1730
1731         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1732                                 &stat, false);
1733
1734         spin_lock_irq(&pgdat->lru_lock);
1735
1736         if (current_is_kswapd()) {
1737                 if (global_reclaim(sc))
1738                         __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1739                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1740                                    nr_reclaimed);
1741         } else {
1742                 if (global_reclaim(sc))
1743                         __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1744                 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1745                                    nr_reclaimed);
1746         }
1747
1748         putback_inactive_pages(lruvec, &page_list);
1749
1750         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1751
1752         spin_unlock_irq(&pgdat->lru_lock);
1753
1754         mem_cgroup_uncharge_list(&page_list);
1755         free_unref_page_list(&page_list);
1756
1757         /*
1758          * If reclaim is isolating dirty pages under writeback, it implies
1759          * that the long-lived page allocation rate is exceeding the page
1760          * laundering rate. Either the global limits are not being effective
1761          * at throttling processes due to the page distribution throughout
1762          * zones or there is heavy usage of a slow backing device. The
1763          * only option is to throttle from reclaim context which is not ideal
1764          * as there is no guarantee the dirtying process is throttled in the
1765          * same way balance_dirty_pages() manages.
1766          *
1767          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1768          * of pages under pages flagged for immediate reclaim and stall if any
1769          * are encountered in the nr_immediate check below.
1770          */
1771         if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1772                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1773
1774         /*
1775          * If dirty pages are scanned that are not queued for IO, it
1776          * implies that flushers are not doing their job. This can
1777          * happen when memory pressure pushes dirty pages to the end of
1778          * the LRU before the dirty limits are breached and the dirty
1779          * data has expired. It can also happen when the proportion of
1780          * dirty pages grows not through writes but through memory
1781          * pressure reclaiming all the clean cache. And in some cases,
1782          * the flushers simply cannot keep up with the allocation
1783          * rate. Nudge the flusher threads in case they are asleep.
1784          */
1785         if (stat.nr_unqueued_dirty == nr_taken)
1786                 wakeup_flusher_threads(WB_REASON_VMSCAN);
1787
1788         /*
1789          * Legacy memcg will stall in page writeback so avoid forcibly
1790          * stalling here.
1791          */
1792         if (sane_reclaim(sc)) {
1793                 /*
1794                  * Tag a zone as congested if all the dirty pages scanned were
1795                  * backed by a congested BDI and wait_iff_congested will stall.
1796                  */
1797                 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1798                         set_bit(PGDAT_CONGESTED, &pgdat->flags);
1799
1800                 /* Allow kswapd to start writing pages during reclaim. */
1801                 if (stat.nr_unqueued_dirty == nr_taken)
1802                         set_bit(PGDAT_DIRTY, &pgdat->flags);
1803
1804                 /*
1805                  * If kswapd scans pages marked marked for immediate
1806                  * reclaim and under writeback (nr_immediate), it implies
1807                  * that pages are cycling through the LRU faster than
1808                  * they are written so also forcibly stall.
1809                  */
1810                 if (stat.nr_immediate && current_may_throttle())
1811                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1812         }
1813
1814         /*
1815          * Stall direct reclaim for IO completions if underlying BDIs or zone
1816          * is congested. Allow kswapd to continue until it starts encountering
1817          * unqueued dirty pages or cycling through the LRU too quickly.
1818          */
1819         if (!sc->hibernation_mode && !current_is_kswapd() &&
1820             current_may_throttle())
1821                 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1822
1823         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1824                         nr_scanned, nr_reclaimed,
1825                         stat.nr_dirty,  stat.nr_writeback,
1826                         stat.nr_congested, stat.nr_immediate,
1827                         stat.nr_activate, stat.nr_ref_keep,
1828                         stat.nr_unmap_fail,
1829                         sc->priority, file);
1830         return nr_reclaimed;
1831 }
1832
1833 /*
1834  * This moves pages from the active list to the inactive list.
1835  *
1836  * We move them the other way if the page is referenced by one or more
1837  * processes, from rmap.
1838  *
1839  * If the pages are mostly unmapped, the processing is fast and it is
1840  * appropriate to hold zone_lru_lock across the whole operation.  But if
1841  * the pages are mapped, the processing is slow (page_referenced()) so we
1842  * should drop zone_lru_lock around each page.  It's impossible to balance
1843  * this, so instead we remove the pages from the LRU while processing them.
1844  * It is safe to rely on PG_active against the non-LRU pages in here because
1845  * nobody will play with that bit on a non-LRU page.
1846  *
1847  * The downside is that we have to touch page->_refcount against each page.
1848  * But we had to alter page->flags anyway.
1849  *
1850  * Returns the number of pages moved to the given lru.
1851  */
1852
1853 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1854                                      struct list_head *list,
1855                                      struct list_head *pages_to_free,
1856                                      enum lru_list lru)
1857 {
1858         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1859         struct page *page;
1860         int nr_pages;
1861         int nr_moved = 0;
1862
1863         while (!list_empty(list)) {
1864                 page = lru_to_page(list);
1865                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1866
1867                 VM_BUG_ON_PAGE(PageLRU(page), page);
1868                 SetPageLRU(page);
1869
1870                 nr_pages = hpage_nr_pages(page);
1871                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1872                 list_move(&page->lru, &lruvec->lists[lru]);
1873
1874                 if (put_page_testzero(page)) {
1875                         __ClearPageLRU(page);
1876                         __ClearPageActive(page);
1877                         del_page_from_lru_list(page, lruvec, lru);
1878
1879                         if (unlikely(PageCompound(page))) {
1880                                 spin_unlock_irq(&pgdat->lru_lock);
1881                                 mem_cgroup_uncharge(page);
1882                                 (*get_compound_page_dtor(page))(page);
1883                                 spin_lock_irq(&pgdat->lru_lock);
1884                         } else
1885                                 list_add(&page->lru, pages_to_free);
1886                 } else {
1887                         nr_moved += nr_pages;
1888                 }
1889         }
1890
1891         if (!is_active_lru(lru)) {
1892                 __count_vm_events(PGDEACTIVATE, nr_moved);
1893                 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1894                                    nr_moved);
1895         }
1896
1897         return nr_moved;
1898 }
1899
1900 static void shrink_active_list(unsigned long nr_to_scan,
1901                                struct lruvec *lruvec,
1902                                struct scan_control *sc,
1903                                enum lru_list lru)
1904 {
1905         unsigned long nr_taken;
1906         unsigned long nr_scanned;
1907         unsigned long vm_flags;
1908         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1909         LIST_HEAD(l_active);
1910         LIST_HEAD(l_inactive);
1911         struct page *page;
1912         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1913         unsigned nr_deactivate, nr_activate;
1914         unsigned nr_rotated = 0;
1915         isolate_mode_t isolate_mode = 0;
1916         int file = is_file_lru(lru);
1917         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1918
1919         lru_add_drain();
1920
1921         if (!sc->may_unmap)
1922                 isolate_mode |= ISOLATE_UNMAPPED;
1923
1924         spin_lock_irq(&pgdat->lru_lock);
1925
1926         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1927                                      &nr_scanned, sc, isolate_mode, lru);
1928
1929         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1930         reclaim_stat->recent_scanned[file] += nr_taken;
1931
1932         __count_vm_events(PGREFILL, nr_scanned);
1933         count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
1934
1935         spin_unlock_irq(&pgdat->lru_lock);
1936
1937         while (!list_empty(&l_hold)) {
1938                 cond_resched();
1939                 page = lru_to_page(&l_hold);
1940                 list_del(&page->lru);
1941
1942                 if (unlikely(!page_evictable(page))) {
1943                         putback_lru_page(page);
1944                         continue;
1945                 }
1946
1947                 if (unlikely(buffer_heads_over_limit)) {
1948                         if (page_has_private(page) && trylock_page(page)) {
1949                                 if (page_has_private(page))
1950                                         try_to_release_page(page, 0);
1951                                 unlock_page(page);
1952                         }
1953                 }
1954
1955                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1956                                     &vm_flags)) {
1957                         nr_rotated += hpage_nr_pages(page);
1958                         /*
1959                          * Identify referenced, file-backed active pages and
1960                          * give them one more trip around the active list. So
1961                          * that executable code get better chances to stay in
1962                          * memory under moderate memory pressure.  Anon pages
1963                          * are not likely to be evicted by use-once streaming
1964                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1965                          * so we ignore them here.
1966                          */
1967                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1968                                 list_add(&page->lru, &l_active);
1969                                 continue;
1970                         }
1971                 }
1972
1973                 ClearPageActive(page);  /* we are de-activating */
1974                 list_add(&page->lru, &l_inactive);
1975         }
1976
1977         /*
1978          * Move pages back to the lru list.
1979          */
1980         spin_lock_irq(&pgdat->lru_lock);
1981         /*
1982          * Count referenced pages from currently used mappings as rotated,
1983          * even though only some of them are actually re-activated.  This
1984          * helps balance scan pressure between file and anonymous pages in
1985          * get_scan_count.
1986          */
1987         reclaim_stat->recent_rotated[file] += nr_rotated;
1988
1989         nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1990         nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1991         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1992         spin_unlock_irq(&pgdat->lru_lock);
1993
1994         mem_cgroup_uncharge_list(&l_hold);
1995         free_unref_page_list(&l_hold);
1996         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
1997                         nr_deactivate, nr_rotated, sc->priority, file);
1998 }
1999
2000 /*
2001  * The inactive anon list should be small enough that the VM never has
2002  * to do too much work.
2003  *
2004  * The inactive file list should be small enough to leave most memory
2005  * to the established workingset on the scan-resistant active list,
2006  * but large enough to avoid thrashing the aggregate readahead window.
2007  *
2008  * Both inactive lists should also be large enough that each inactive
2009  * page has a chance to be referenced again before it is reclaimed.
2010  *
2011  * If that fails and refaulting is observed, the inactive list grows.
2012  *
2013  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2014  * on this LRU, maintained by the pageout code. An inactive_ratio
2015  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2016  *
2017  * total     target    max
2018  * memory    ratio     inactive
2019  * -------------------------------------
2020  *   10MB       1         5MB
2021  *  100MB       1        50MB
2022  *    1GB       3       250MB
2023  *   10GB      10       0.9GB
2024  *  100GB      31         3GB
2025  *    1TB     101        10GB
2026  *   10TB     320        32GB
2027  */
2028 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2029                                  struct mem_cgroup *memcg,
2030                                  struct scan_control *sc, bool actual_reclaim)
2031 {
2032         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2033         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2034         enum lru_list inactive_lru = file * LRU_FILE;
2035         unsigned long inactive, active;
2036         unsigned long inactive_ratio;
2037         unsigned long refaults;
2038         unsigned long gb;
2039
2040         /*
2041          * If we don't have swap space, anonymous page deactivation
2042          * is pointless.
2043          */
2044         if (!file && !total_swap_pages)
2045                 return false;
2046
2047         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2048         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2049
2050         if (memcg)
2051                 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2052         else
2053                 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2054
2055         /*
2056          * When refaults are being observed, it means a new workingset
2057          * is being established. Disable active list protection to get
2058          * rid of the stale workingset quickly.
2059          */
2060         if (file && actual_reclaim && lruvec->refaults != refaults) {
2061                 inactive_ratio = 0;
2062         } else {
2063                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2064                 if (gb)
2065                         inactive_ratio = int_sqrt(10 * gb);
2066                 else
2067                         inactive_ratio = 1;
2068         }
2069
2070         if (actual_reclaim)
2071                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2072                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2073                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2074                         inactive_ratio, file);
2075
2076         return inactive * inactive_ratio < active;
2077 }
2078
2079 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2080                                  struct lruvec *lruvec, struct mem_cgroup *memcg,
2081                                  struct scan_control *sc)
2082 {
2083         if (is_active_lru(lru)) {
2084                 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2085                                          memcg, sc, true))
2086                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2087                 return 0;
2088         }
2089
2090         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2091 }
2092
2093 enum scan_balance {
2094         SCAN_EQUAL,
2095         SCAN_FRACT,
2096         SCAN_ANON,
2097         SCAN_FILE,
2098 };
2099
2100 /*
2101  * Determine how aggressively the anon and file LRU lists should be
2102  * scanned.  The relative value of each set of LRU lists is determined
2103  * by looking at the fraction of the pages scanned we did rotate back
2104  * onto the active list instead of evict.
2105  *
2106  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2107  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2108  */
2109 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2110                            struct scan_control *sc, unsigned long *nr,
2111                            unsigned long *lru_pages)
2112 {
2113         int swappiness = mem_cgroup_swappiness(memcg);
2114         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2115         u64 fraction[2];
2116         u64 denominator = 0;    /* gcc */
2117         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2118         unsigned long anon_prio, file_prio;
2119         enum scan_balance scan_balance;
2120         unsigned long anon, file;
2121         unsigned long ap, fp;
2122         enum lru_list lru;
2123
2124         /* If we have no swap space, do not bother scanning anon pages. */
2125         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2126                 scan_balance = SCAN_FILE;
2127                 goto out;
2128         }
2129
2130         /*
2131          * Global reclaim will swap to prevent OOM even with no
2132          * swappiness, but memcg users want to use this knob to
2133          * disable swapping for individual groups completely when
2134          * using the memory controller's swap limit feature would be
2135          * too expensive.
2136          */
2137         if (!global_reclaim(sc) && !swappiness) {
2138                 scan_balance = SCAN_FILE;
2139                 goto out;
2140         }
2141
2142         /*
2143          * Do not apply any pressure balancing cleverness when the
2144          * system is close to OOM, scan both anon and file equally
2145          * (unless the swappiness setting disagrees with swapping).
2146          */
2147         if (!sc->priority && swappiness) {
2148                 scan_balance = SCAN_EQUAL;
2149                 goto out;
2150         }
2151
2152         /*
2153          * Prevent the reclaimer from falling into the cache trap: as
2154          * cache pages start out inactive, every cache fault will tip
2155          * the scan balance towards the file LRU.  And as the file LRU
2156          * shrinks, so does the window for rotation from references.
2157          * This means we have a runaway feedback loop where a tiny
2158          * thrashing file LRU becomes infinitely more attractive than
2159          * anon pages.  Try to detect this based on file LRU size.
2160          */
2161         if (global_reclaim(sc)) {
2162                 unsigned long pgdatfile;
2163                 unsigned long pgdatfree;
2164                 int z;
2165                 unsigned long total_high_wmark = 0;
2166
2167                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2168                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2169                            node_page_state(pgdat, NR_INACTIVE_FILE);
2170
2171                 for (z = 0; z < MAX_NR_ZONES; z++) {
2172                         struct zone *zone = &pgdat->node_zones[z];
2173                         if (!managed_zone(zone))
2174                                 continue;
2175
2176                         total_high_wmark += high_wmark_pages(zone);
2177                 }
2178
2179                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2180                         /*
2181                          * Force SCAN_ANON if there are enough inactive
2182                          * anonymous pages on the LRU in eligible zones.
2183                          * Otherwise, the small LRU gets thrashed.
2184                          */
2185                         if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2186                             lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2187                                         >> sc->priority) {
2188                                 scan_balance = SCAN_ANON;
2189                                 goto out;
2190                         }
2191                 }
2192         }
2193
2194         /*
2195          * If there is enough inactive page cache, i.e. if the size of the
2196          * inactive list is greater than that of the active list *and* the
2197          * inactive list actually has some pages to scan on this priority, we
2198          * do not reclaim anything from the anonymous working set right now.
2199          * Without the second condition we could end up never scanning an
2200          * lruvec even if it has plenty of old anonymous pages unless the
2201          * system is under heavy pressure.
2202          */
2203         if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2204             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2205                 scan_balance = SCAN_FILE;
2206                 goto out;
2207         }
2208
2209         scan_balance = SCAN_FRACT;
2210
2211         /*
2212          * With swappiness at 100, anonymous and file have the same priority.
2213          * This scanning priority is essentially the inverse of IO cost.
2214          */
2215         anon_prio = swappiness;
2216         file_prio = 200 - anon_prio;
2217
2218         /*
2219          * OK, so we have swap space and a fair amount of page cache
2220          * pages.  We use the recently rotated / recently scanned
2221          * ratios to determine how valuable each cache is.
2222          *
2223          * Because workloads change over time (and to avoid overflow)
2224          * we keep these statistics as a floating average, which ends
2225          * up weighing recent references more than old ones.
2226          *
2227          * anon in [0], file in [1]
2228          */
2229
2230         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2231                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2232         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2233                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2234
2235         spin_lock_irq(&pgdat->lru_lock);
2236         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2237                 reclaim_stat->recent_scanned[0] /= 2;
2238                 reclaim_stat->recent_rotated[0] /= 2;
2239         }
2240
2241         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2242                 reclaim_stat->recent_scanned[1] /= 2;
2243                 reclaim_stat->recent_rotated[1] /= 2;
2244         }
2245
2246         /*
2247          * The amount of pressure on anon vs file pages is inversely
2248          * proportional to the fraction of recently scanned pages on
2249          * each list that were recently referenced and in active use.
2250          */
2251         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2252         ap /= reclaim_stat->recent_rotated[0] + 1;
2253
2254         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2255         fp /= reclaim_stat->recent_rotated[1] + 1;
2256         spin_unlock_irq(&pgdat->lru_lock);
2257
2258         fraction[0] = ap;
2259         fraction[1] = fp;
2260         denominator = ap + fp + 1;
2261 out:
2262         *lru_pages = 0;
2263         for_each_evictable_lru(lru) {
2264                 int file = is_file_lru(lru);
2265                 unsigned long size;
2266                 unsigned long scan;
2267
2268                 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2269                 scan = size >> sc->priority;
2270                 /*
2271                  * If the cgroup's already been deleted, make sure to
2272                  * scrape out the remaining cache.
2273                  */
2274                 if (!scan && !mem_cgroup_online(memcg))
2275                         scan = min(size, SWAP_CLUSTER_MAX);
2276
2277                 switch (scan_balance) {
2278                 case SCAN_EQUAL:
2279                         /* Scan lists relative to size */
2280                         break;
2281                 case SCAN_FRACT:
2282                         /*
2283                          * Scan types proportional to swappiness and
2284                          * their relative recent reclaim efficiency.
2285                          */
2286                         scan = div64_u64(scan * fraction[file],
2287                                          denominator);
2288                         break;
2289                 case SCAN_FILE:
2290                 case SCAN_ANON:
2291                         /* Scan one type exclusively */
2292                         if ((scan_balance == SCAN_FILE) != file) {
2293                                 size = 0;
2294                                 scan = 0;
2295                         }
2296                         break;
2297                 default:
2298                         /* Look ma, no brain */
2299                         BUG();
2300                 }
2301
2302                 *lru_pages += size;
2303                 nr[lru] = scan;
2304         }
2305 }
2306
2307 /*
2308  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2309  */
2310 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2311                               struct scan_control *sc, unsigned long *lru_pages)
2312 {
2313         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2314         unsigned long nr[NR_LRU_LISTS];
2315         unsigned long targets[NR_LRU_LISTS];
2316         unsigned long nr_to_scan;
2317         enum lru_list lru;
2318         unsigned long nr_reclaimed = 0;
2319         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2320         struct blk_plug plug;
2321         bool scan_adjusted;
2322
2323         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2324
2325         /* Record the original scan target for proportional adjustments later */
2326         memcpy(targets, nr, sizeof(nr));
2327
2328         /*
2329          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2330          * event that can occur when there is little memory pressure e.g.
2331          * multiple streaming readers/writers. Hence, we do not abort scanning
2332          * when the requested number of pages are reclaimed when scanning at
2333          * DEF_PRIORITY on the assumption that the fact we are direct
2334          * reclaiming implies that kswapd is not keeping up and it is best to
2335          * do a batch of work at once. For memcg reclaim one check is made to
2336          * abort proportional reclaim if either the file or anon lru has already
2337          * dropped to zero at the first pass.
2338          */
2339         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2340                          sc->priority == DEF_PRIORITY);
2341
2342         blk_start_plug(&plug);
2343         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2344                                         nr[LRU_INACTIVE_FILE]) {
2345                 unsigned long nr_anon, nr_file, percentage;
2346                 unsigned long nr_scanned;
2347
2348                 for_each_evictable_lru(lru) {
2349                         if (nr[lru]) {
2350                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2351                                 nr[lru] -= nr_to_scan;
2352
2353                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2354                                                             lruvec, memcg, sc);
2355                         }
2356                 }
2357
2358                 cond_resched();
2359
2360                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2361                         continue;
2362
2363                 /*
2364                  * For kswapd and memcg, reclaim at least the number of pages
2365                  * requested. Ensure that the anon and file LRUs are scanned
2366                  * proportionally what was requested by get_scan_count(). We
2367                  * stop reclaiming one LRU and reduce the amount scanning
2368                  * proportional to the original scan target.
2369                  */
2370                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2371                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2372
2373                 /*
2374                  * It's just vindictive to attack the larger once the smaller
2375                  * has gone to zero.  And given the way we stop scanning the
2376                  * smaller below, this makes sure that we only make one nudge
2377                  * towards proportionality once we've got nr_to_reclaim.
2378                  */
2379                 if (!nr_file || !nr_anon)
2380                         break;
2381
2382                 if (nr_file > nr_anon) {
2383                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2384                                                 targets[LRU_ACTIVE_ANON] + 1;
2385                         lru = LRU_BASE;
2386                         percentage = nr_anon * 100 / scan_target;
2387                 } else {
2388                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2389                                                 targets[LRU_ACTIVE_FILE] + 1;
2390                         lru = LRU_FILE;
2391                         percentage = nr_file * 100 / scan_target;
2392                 }
2393
2394                 /* Stop scanning the smaller of the LRU */
2395                 nr[lru] = 0;
2396                 nr[lru + LRU_ACTIVE] = 0;
2397
2398                 /*
2399                  * Recalculate the other LRU scan count based on its original
2400                  * scan target and the percentage scanning already complete
2401                  */
2402                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2403                 nr_scanned = targets[lru] - nr[lru];
2404                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2405                 nr[lru] -= min(nr[lru], nr_scanned);
2406
2407                 lru += LRU_ACTIVE;
2408                 nr_scanned = targets[lru] - nr[lru];
2409                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2410                 nr[lru] -= min(nr[lru], nr_scanned);
2411
2412                 scan_adjusted = true;
2413         }
2414         blk_finish_plug(&plug);
2415         sc->nr_reclaimed += nr_reclaimed;
2416
2417         /*
2418          * Even if we did not try to evict anon pages at all, we want to
2419          * rebalance the anon lru active/inactive ratio.
2420          */
2421         if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2422                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2423                                    sc, LRU_ACTIVE_ANON);
2424 }
2425
2426 /* Use reclaim/compaction for costly allocs or under memory pressure */
2427 static bool in_reclaim_compaction(struct scan_control *sc)
2428 {
2429         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2430                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2431                          sc->priority < DEF_PRIORITY - 2))
2432                 return true;
2433
2434         return false;
2435 }
2436
2437 /*
2438  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2439  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2440  * true if more pages should be reclaimed such that when the page allocator
2441  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2442  * It will give up earlier than that if there is difficulty reclaiming pages.
2443  */
2444 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2445                                         unsigned long nr_reclaimed,
2446                                         unsigned long nr_scanned,
2447                                         struct scan_control *sc)
2448 {
2449         unsigned long pages_for_compaction;
2450         unsigned long inactive_lru_pages;
2451         int z;
2452
2453         /* If not in reclaim/compaction mode, stop */
2454         if (!in_reclaim_compaction(sc))
2455                 return false;
2456
2457         /* Consider stopping depending on scan and reclaim activity */
2458         if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2459                 /*
2460                  * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2461                  * full LRU list has been scanned and we are still failing
2462                  * to reclaim pages. This full LRU scan is potentially
2463                  * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2464                  */
2465                 if (!nr_reclaimed && !nr_scanned)
2466                         return false;
2467         } else {
2468                 /*
2469                  * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2470                  * fail without consequence, stop if we failed to reclaim
2471                  * any pages from the last SWAP_CLUSTER_MAX number of
2472                  * pages that were scanned. This will return to the
2473                  * caller faster at the risk reclaim/compaction and
2474                  * the resulting allocation attempt fails
2475                  */
2476                 if (!nr_reclaimed)
2477                         return false;
2478         }
2479
2480         /*
2481          * If we have not reclaimed enough pages for compaction and the
2482          * inactive lists are large enough, continue reclaiming
2483          */
2484         pages_for_compaction = compact_gap(sc->order);
2485         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2486         if (get_nr_swap_pages() > 0)
2487                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2488         if (sc->nr_reclaimed < pages_for_compaction &&
2489                         inactive_lru_pages > pages_for_compaction)
2490                 return true;
2491
2492         /* If compaction would go ahead or the allocation would succeed, stop */
2493         for (z = 0; z <= sc->reclaim_idx; z++) {
2494                 struct zone *zone = &pgdat->node_zones[z];
2495                 if (!managed_zone(zone))
2496                         continue;
2497
2498                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2499                 case COMPACT_SUCCESS:
2500                 case COMPACT_CONTINUE:
2501                         return false;
2502                 default:
2503                         /* check next zone */
2504                         ;
2505                 }
2506         }
2507         return true;
2508 }
2509
2510 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2511 {
2512         struct reclaim_state *reclaim_state = current->reclaim_state;
2513         unsigned long nr_reclaimed, nr_scanned;
2514         bool reclaimable = false;
2515
2516         do {
2517                 struct mem_cgroup *root = sc->target_mem_cgroup;
2518                 struct mem_cgroup_reclaim_cookie reclaim = {
2519                         .pgdat = pgdat,
2520                         .priority = sc->priority,
2521                 };
2522                 unsigned long node_lru_pages = 0;
2523                 struct mem_cgroup *memcg;
2524
2525                 nr_reclaimed = sc->nr_reclaimed;
2526                 nr_scanned = sc->nr_scanned;
2527
2528                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2529                 do {
2530                         unsigned long lru_pages;
2531                         unsigned long reclaimed;
2532                         unsigned long scanned;
2533
2534                         if (mem_cgroup_low(root, memcg)) {
2535                                 if (!sc->memcg_low_reclaim) {
2536                                         sc->memcg_low_skipped = 1;
2537                                         continue;
2538                                 }
2539                                 mem_cgroup_event(memcg, MEMCG_LOW);
2540                         }
2541
2542                         reclaimed = sc->nr_reclaimed;
2543                         scanned = sc->nr_scanned;
2544                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2545                         node_lru_pages += lru_pages;
2546
2547                         if (memcg)
2548                                 shrink_slab(sc->gfp_mask, pgdat->node_id,
2549                                             memcg, sc->priority);
2550
2551                         /* Record the group's reclaim efficiency */
2552                         vmpressure(sc->gfp_mask, memcg, false,
2553                                    sc->nr_scanned - scanned,
2554                                    sc->nr_reclaimed - reclaimed);
2555
2556                         /*
2557                          * Direct reclaim and kswapd have to scan all memory
2558                          * cgroups to fulfill the overall scan target for the
2559                          * node.
2560                          *
2561                          * Limit reclaim, on the other hand, only cares about
2562                          * nr_to_reclaim pages to be reclaimed and it will
2563                          * retry with decreasing priority if one round over the
2564                          * whole hierarchy is not sufficient.
2565                          */
2566                         if (!global_reclaim(sc) &&
2567                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2568                                 mem_cgroup_iter_break(root, memcg);
2569                                 break;
2570                         }
2571                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2572
2573                 if (global_reclaim(sc))
2574                         shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2575                                     sc->priority);
2576
2577                 if (reclaim_state) {
2578                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2579                         reclaim_state->reclaimed_slab = 0;
2580                 }
2581
2582                 /* Record the subtree's reclaim efficiency */
2583                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2584                            sc->nr_scanned - nr_scanned,
2585                            sc->nr_reclaimed - nr_reclaimed);
2586
2587                 if (sc->nr_reclaimed - nr_reclaimed)
2588                         reclaimable = true;
2589
2590         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2591                                          sc->nr_scanned - nr_scanned, sc));
2592
2593         /*
2594          * Kswapd gives up on balancing particular nodes after too
2595          * many failures to reclaim anything from them and goes to
2596          * sleep. On reclaim progress, reset the failure counter. A
2597          * successful direct reclaim run will revive a dormant kswapd.
2598          */
2599         if (reclaimable)
2600                 pgdat->kswapd_failures = 0;
2601
2602         return reclaimable;
2603 }
2604
2605 /*
2606  * Returns true if compaction should go ahead for a costly-order request, or
2607  * the allocation would already succeed without compaction. Return false if we
2608  * should reclaim first.
2609  */
2610 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2611 {
2612         unsigned long watermark;
2613         enum compact_result suitable;
2614
2615         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2616         if (suitable == COMPACT_SUCCESS)
2617                 /* Allocation should succeed already. Don't reclaim. */
2618                 return true;
2619         if (suitable == COMPACT_SKIPPED)
2620                 /* Compaction cannot yet proceed. Do reclaim. */
2621                 return false;
2622
2623         /*
2624          * Compaction is already possible, but it takes time to run and there
2625          * are potentially other callers using the pages just freed. So proceed
2626          * with reclaim to make a buffer of free pages available to give
2627          * compaction a reasonable chance of completing and allocating the page.
2628          * Note that we won't actually reclaim the whole buffer in one attempt
2629          * as the target watermark in should_continue_reclaim() is lower. But if
2630          * we are already above the high+gap watermark, don't reclaim at all.
2631          */
2632         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2633
2634         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2635 }
2636
2637 /*
2638  * This is the direct reclaim path, for page-allocating processes.  We only
2639  * try to reclaim pages from zones which will satisfy the caller's allocation
2640  * request.
2641  *
2642  * If a zone is deemed to be full of pinned pages then just give it a light
2643  * scan then give up on it.
2644  */
2645 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2646 {
2647         struct zoneref *z;
2648         struct zone *zone;
2649         unsigned long nr_soft_reclaimed;
2650         unsigned long nr_soft_scanned;
2651         gfp_t orig_mask;
2652         pg_data_t *last_pgdat = NULL;
2653
2654         /*
2655          * If the number of buffer_heads in the machine exceeds the maximum
2656          * allowed level, force direct reclaim to scan the highmem zone as
2657          * highmem pages could be pinning lowmem pages storing buffer_heads
2658          */
2659         orig_mask = sc->gfp_mask;
2660         if (buffer_heads_over_limit) {
2661                 sc->gfp_mask |= __GFP_HIGHMEM;
2662                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2663         }
2664
2665         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2666                                         sc->reclaim_idx, sc->nodemask) {
2667                 /*
2668                  * Take care memory controller reclaiming has small influence
2669                  * to global LRU.
2670                  */
2671                 if (global_reclaim(sc)) {
2672                         if (!cpuset_zone_allowed(zone,
2673                                                  GFP_KERNEL | __GFP_HARDWALL))
2674                                 continue;
2675
2676                         /*
2677                          * If we already have plenty of memory free for
2678                          * compaction in this zone, don't free any more.
2679                          * Even though compaction is invoked for any
2680                          * non-zero order, only frequent costly order
2681                          * reclamation is disruptive enough to become a
2682                          * noticeable problem, like transparent huge
2683                          * page allocations.
2684                          */
2685                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2686                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2687                             compaction_ready(zone, sc)) {
2688                                 sc->compaction_ready = true;
2689                                 continue;
2690                         }
2691
2692                         /*
2693                          * Shrink each node in the zonelist once. If the
2694                          * zonelist is ordered by zone (not the default) then a
2695                          * node may be shrunk multiple times but in that case
2696                          * the user prefers lower zones being preserved.
2697                          */
2698                         if (zone->zone_pgdat == last_pgdat)
2699                                 continue;
2700
2701                         /*
2702                          * This steals pages from memory cgroups over softlimit
2703                          * and returns the number of reclaimed pages and
2704                          * scanned pages. This works for global memory pressure
2705                          * and balancing, not for a memcg's limit.
2706                          */
2707                         nr_soft_scanned = 0;
2708                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2709                                                 sc->order, sc->gfp_mask,
2710                                                 &nr_soft_scanned);
2711                         sc->nr_reclaimed += nr_soft_reclaimed;
2712                         sc->nr_scanned += nr_soft_scanned;
2713                         /* need some check for avoid more shrink_zone() */
2714                 }
2715
2716                 /* See comment about same check for global reclaim above */
2717                 if (zone->zone_pgdat == last_pgdat)
2718                         continue;
2719                 last_pgdat = zone->zone_pgdat;
2720                 shrink_node(zone->zone_pgdat, sc);
2721         }
2722
2723         /*
2724          * Restore to original mask to avoid the impact on the caller if we
2725          * promoted it to __GFP_HIGHMEM.
2726          */
2727         sc->gfp_mask = orig_mask;
2728 }
2729
2730 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2731 {
2732         struct mem_cgroup *memcg;
2733
2734         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2735         do {
2736                 unsigned long refaults;
2737                 struct lruvec *lruvec;
2738
2739                 if (memcg)
2740                         refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2741                 else
2742                         refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2743
2744                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2745                 lruvec->refaults = refaults;
2746         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2747 }
2748
2749 /*
2750  * This is the main entry point to direct page reclaim.
2751  *
2752  * If a full scan of the inactive list fails to free enough memory then we
2753  * are "out of memory" and something needs to be killed.
2754  *
2755  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2756  * high - the zone may be full of dirty or under-writeback pages, which this
2757  * caller can't do much about.  We kick the writeback threads and take explicit
2758  * naps in the hope that some of these pages can be written.  But if the
2759  * allocating task holds filesystem locks which prevent writeout this might not
2760  * work, and the allocation attempt will fail.
2761  *
2762  * returns:     0, if no pages reclaimed
2763  *              else, the number of pages reclaimed
2764  */
2765 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2766                                           struct scan_control *sc)
2767 {
2768         int initial_priority = sc->priority;
2769         pg_data_t *last_pgdat;
2770         struct zoneref *z;
2771         struct zone *zone;
2772 retry:
2773         delayacct_freepages_start();
2774
2775         if (global_reclaim(sc))
2776                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2777
2778         do {
2779                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2780                                 sc->priority);
2781                 sc->nr_scanned = 0;
2782                 shrink_zones(zonelist, sc);
2783
2784                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2785                         break;
2786
2787                 if (sc->compaction_ready)
2788                         break;
2789
2790                 /*
2791                  * If we're getting trouble reclaiming, start doing
2792                  * writepage even in laptop mode.
2793                  */
2794                 if (sc->priority < DEF_PRIORITY - 2)
2795                         sc->may_writepage = 1;
2796         } while (--sc->priority >= 0);
2797
2798         last_pgdat = NULL;
2799         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2800                                         sc->nodemask) {
2801                 if (zone->zone_pgdat == last_pgdat)
2802                         continue;
2803                 last_pgdat = zone->zone_pgdat;
2804                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2805         }
2806
2807         delayacct_freepages_end();
2808
2809         if (sc->nr_reclaimed)
2810                 return sc->nr_reclaimed;
2811
2812         /* Aborted reclaim to try compaction? don't OOM, then */
2813         if (sc->compaction_ready)
2814                 return 1;
2815
2816         /* Untapped cgroup reserves?  Don't OOM, retry. */
2817         if (sc->memcg_low_skipped) {
2818                 sc->priority = initial_priority;
2819                 sc->memcg_low_reclaim = 1;
2820                 sc->memcg_low_skipped = 0;
2821                 goto retry;
2822         }
2823
2824         return 0;
2825 }
2826
2827 static bool allow_direct_reclaim(pg_data_t *pgdat)
2828 {
2829         struct zone *zone;
2830         unsigned long pfmemalloc_reserve = 0;
2831         unsigned long free_pages = 0;
2832         int i;
2833         bool wmark_ok;
2834
2835         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2836                 return true;
2837
2838         for (i = 0; i <= ZONE_NORMAL; i++) {
2839                 zone = &pgdat->node_zones[i];
2840                 if (!managed_zone(zone))
2841                         continue;
2842
2843                 if (!zone_reclaimable_pages(zone))
2844                         continue;
2845
2846                 pfmemalloc_reserve += min_wmark_pages(zone);
2847                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2848         }
2849
2850         /* If there are no reserves (unexpected config) then do not throttle */
2851         if (!pfmemalloc_reserve)
2852                 return true;
2853
2854         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2855
2856         /* kswapd must be awake if processes are being throttled */
2857         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2858                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2859                                                 (enum zone_type)ZONE_NORMAL);
2860                 wake_up_interruptible(&pgdat->kswapd_wait);
2861         }
2862
2863         return wmark_ok;
2864 }
2865
2866 /*
2867  * Throttle direct reclaimers if backing storage is backed by the network
2868  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2869  * depleted. kswapd will continue to make progress and wake the processes
2870  * when the low watermark is reached.
2871  *
2872  * Returns true if a fatal signal was delivered during throttling. If this
2873  * happens, the page allocator should not consider triggering the OOM killer.
2874  */
2875 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2876                                         nodemask_t *nodemask)
2877 {
2878         struct zoneref *z;
2879         struct zone *zone;
2880         pg_data_t *pgdat = NULL;
2881
2882         /*
2883          * Kernel threads should not be throttled as they may be indirectly
2884          * responsible for cleaning pages necessary for reclaim to make forward
2885          * progress. kjournald for example may enter direct reclaim while
2886          * committing a transaction where throttling it could forcing other
2887          * processes to block on log_wait_commit().
2888          */
2889         if (current->flags & PF_KTHREAD)
2890                 goto out;
2891
2892         /*
2893          * If a fatal signal is pending, this process should not throttle.
2894          * It should return quickly so it can exit and free its memory
2895          */
2896         if (fatal_signal_pending(current))
2897                 goto out;
2898
2899         /*
2900          * Check if the pfmemalloc reserves are ok by finding the first node
2901          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2902          * GFP_KERNEL will be required for allocating network buffers when
2903          * swapping over the network so ZONE_HIGHMEM is unusable.
2904          *
2905          * Throttling is based on the first usable node and throttled processes
2906          * wait on a queue until kswapd makes progress and wakes them. There
2907          * is an affinity then between processes waking up and where reclaim
2908          * progress has been made assuming the process wakes on the same node.
2909          * More importantly, processes running on remote nodes will not compete
2910          * for remote pfmemalloc reserves and processes on different nodes
2911          * should make reasonable progress.
2912          */
2913         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2914                                         gfp_zone(gfp_mask), nodemask) {
2915                 if (zone_idx(zone) > ZONE_NORMAL)
2916                         continue;
2917
2918                 /* Throttle based on the first usable node */
2919                 pgdat = zone->zone_pgdat;
2920                 if (allow_direct_reclaim(pgdat))
2921                         goto out;
2922                 break;
2923         }
2924
2925         /* If no zone was usable by the allocation flags then do not throttle */
2926         if (!pgdat)
2927                 goto out;
2928
2929         /* Account for the throttling */
2930         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2931
2932         /*
2933          * If the caller cannot enter the filesystem, it's possible that it
2934          * is due to the caller holding an FS lock or performing a journal
2935          * transaction in the case of a filesystem like ext[3|4]. In this case,
2936          * it is not safe to block on pfmemalloc_wait as kswapd could be
2937          * blocked waiting on the same lock. Instead, throttle for up to a
2938          * second before continuing.
2939          */
2940         if (!(gfp_mask & __GFP_FS)) {
2941                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2942                         allow_direct_reclaim(pgdat), HZ);
2943
2944                 goto check_pending;
2945         }
2946
2947         /* Throttle until kswapd wakes the process */
2948         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2949                 allow_direct_reclaim(pgdat));
2950
2951 check_pending:
2952         if (fatal_signal_pending(current))
2953                 return true;
2954
2955 out:
2956         return false;
2957 }
2958
2959 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2960                                 gfp_t gfp_mask, nodemask_t *nodemask)
2961 {
2962         unsigned long nr_reclaimed;
2963         struct scan_control sc = {
2964                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2965                 .gfp_mask = current_gfp_context(gfp_mask),
2966                 .reclaim_idx = gfp_zone(gfp_mask),
2967                 .order = order,
2968                 .nodemask = nodemask,
2969                 .priority = DEF_PRIORITY,
2970                 .may_writepage = !laptop_mode,
2971                 .may_unmap = 1,
2972                 .may_swap = 1,
2973         };
2974
2975         /*
2976          * Do not enter reclaim if fatal signal was delivered while throttled.
2977          * 1 is returned so that the page allocator does not OOM kill at this
2978          * point.
2979          */
2980         if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
2981                 return 1;
2982
2983         trace_mm_vmscan_direct_reclaim_begin(order,
2984                                 sc.may_writepage,
2985                                 sc.gfp_mask,
2986                                 sc.reclaim_idx);
2987
2988         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2989
2990         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2991
2992         return nr_reclaimed;
2993 }
2994
2995 #ifdef CONFIG_MEMCG
2996
2997 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2998                                                 gfp_t gfp_mask, bool noswap,
2999                                                 pg_data_t *pgdat,
3000                                                 unsigned long *nr_scanned)
3001 {
3002         struct scan_control sc = {
3003                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3004                 .target_mem_cgroup = memcg,
3005                 .may_writepage = !laptop_mode,
3006                 .may_unmap = 1,
3007                 .reclaim_idx = MAX_NR_ZONES - 1,
3008                 .may_swap = !noswap,
3009         };
3010         unsigned long lru_pages;
3011
3012         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3013                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3014
3015         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3016                                                       sc.may_writepage,
3017                                                       sc.gfp_mask,
3018                                                       sc.reclaim_idx);
3019
3020         /*
3021          * NOTE: Although we can get the priority field, using it
3022          * here is not a good idea, since it limits the pages we can scan.
3023          * if we don't reclaim here, the shrink_node from balance_pgdat
3024          * will pick up pages from other mem cgroup's as well. We hack
3025          * the priority and make it zero.
3026          */
3027         shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3028
3029         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3030
3031         *nr_scanned = sc.nr_scanned;
3032         return sc.nr_reclaimed;
3033 }
3034
3035 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3036                                            unsigned long nr_pages,
3037                                            gfp_t gfp_mask,
3038                                            bool may_swap)
3039 {
3040         struct zonelist *zonelist;
3041         unsigned long nr_reclaimed;
3042         int nid;
3043         unsigned int noreclaim_flag;
3044         struct scan_control sc = {
3045                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3046                 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3047                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3048                 .reclaim_idx = MAX_NR_ZONES - 1,
3049                 .target_mem_cgroup = memcg,
3050                 .priority = DEF_PRIORITY,
3051                 .may_writepage = !laptop_mode,
3052                 .may_unmap = 1,
3053                 .may_swap = may_swap,
3054         };
3055
3056         /*
3057          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3058          * take care of from where we get pages. So the node where we start the
3059          * scan does not need to be the current node.
3060          */
3061         nid = mem_cgroup_select_victim_node(memcg);
3062
3063         zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3064
3065         trace_mm_vmscan_memcg_reclaim_begin(0,
3066                                             sc.may_writepage,
3067                                             sc.gfp_mask,
3068                                             sc.reclaim_idx);
3069
3070         noreclaim_flag = memalloc_noreclaim_save();
3071         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3072         memalloc_noreclaim_restore(noreclaim_flag);
3073
3074         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3075
3076         return nr_reclaimed;
3077 }
3078 #endif
3079
3080 static void age_active_anon(struct pglist_data *pgdat,
3081                                 struct scan_control *sc)
3082 {
3083         struct mem_cgroup *memcg;
3084
3085         if (!total_swap_pages)
3086                 return;
3087
3088         memcg = mem_cgroup_iter(NULL, NULL, NULL);
3089         do {
3090                 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3091
3092                 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3093                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3094                                            sc, LRU_ACTIVE_ANON);
3095
3096                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3097         } while (memcg);
3098 }
3099
3100 /*
3101  * Returns true if there is an eligible zone balanced for the request order
3102  * and classzone_idx
3103  */
3104 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3105 {
3106         int i;
3107         unsigned long mark = -1;
3108         struct zone *zone;
3109
3110         for (i = 0; i <= classzone_idx; i++) {
3111                 zone = pgdat->node_zones + i;
3112
3113                 if (!managed_zone(zone))
3114                         continue;
3115
3116                 mark = high_wmark_pages(zone);
3117                 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3118                         return true;
3119         }
3120
3121         /*
3122          * If a node has no populated zone within classzone_idx, it does not
3123          * need balancing by definition. This can happen if a zone-restricted
3124          * allocation tries to wake a remote kswapd.
3125          */
3126         if (mark == -1)
3127                 return true;
3128
3129         return false;
3130 }
3131
3132 /* Clear pgdat state for congested, dirty or under writeback. */
3133 static void clear_pgdat_congested(pg_data_t *pgdat)
3134 {
3135         clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3136         clear_bit(PGDAT_DIRTY, &pgdat->flags);
3137         clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3138 }
3139
3140 /*
3141  * Prepare kswapd for sleeping. This verifies that there are no processes
3142  * waiting in throttle_direct_reclaim() and that watermarks have been met.
3143  *
3144  * Returns true if kswapd is ready to sleep
3145  */
3146 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3147 {
3148         /*
3149          * The throttled processes are normally woken up in balance_pgdat() as
3150          * soon as allow_direct_reclaim() is true. But there is a potential
3151          * race between when kswapd checks the watermarks and a process gets
3152          * throttled. There is also a potential race if processes get
3153          * throttled, kswapd wakes, a large process exits thereby balancing the
3154          * zones, which causes kswapd to exit balance_pgdat() before reaching
3155          * the wake up checks. If kswapd is going to sleep, no process should
3156          * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3157          * the wake up is premature, processes will wake kswapd and get
3158          * throttled again. The difference from wake ups in balance_pgdat() is
3159          * that here we are under prepare_to_wait().
3160          */
3161         if (waitqueue_active(&pgdat->pfmemalloc_wait))
3162                 wake_up_all(&pgdat->pfmemalloc_wait);
3163
3164         /* Hopeless node, leave it to direct reclaim */
3165         if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3166                 return true;
3167
3168         if (pgdat_balanced(pgdat, order, classzone_idx)) {
3169                 clear_pgdat_congested(pgdat);
3170                 return true;
3171         }
3172
3173         return false;
3174 }
3175
3176 /*
3177  * kswapd shrinks a node of pages that are at or below the highest usable
3178  * zone that is currently unbalanced.
3179  *
3180  * Returns true if kswapd scanned at least the requested number of pages to
3181  * reclaim or if the lack of progress was due to pages under writeback.
3182  * This is used to determine if the scanning priority needs to be raised.
3183  */
3184 static bool kswapd_shrink_node(pg_data_t *pgdat,
3185                                struct scan_control *sc)
3186 {
3187         struct zone *zone;
3188         int z;
3189
3190         /* Reclaim a number of pages proportional to the number of zones */
3191         sc->nr_to_reclaim = 0;
3192         for (z = 0; z <= sc->reclaim_idx; z++) {
3193                 zone = pgdat->node_zones + z;
3194                 if (!managed_zone(zone))
3195                         continue;
3196
3197                 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3198         }
3199
3200         /*
3201          * Historically care was taken to put equal pressure on all zones but
3202          * now pressure is applied based on node LRU order.
3203          */
3204         shrink_node(pgdat, sc);
3205
3206         /*
3207          * Fragmentation may mean that the system cannot be rebalanced for
3208          * high-order allocations. If twice the allocation size has been
3209          * reclaimed then recheck watermarks only at order-0 to prevent
3210          * excessive reclaim. Assume that a process requested a high-order
3211          * can direct reclaim/compact.
3212          */
3213         if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3214                 sc->order = 0;
3215
3216         return sc->nr_scanned >= sc->nr_to_reclaim;
3217 }
3218
3219 /*
3220  * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3221  * that are eligible for use by the caller until at least one zone is
3222  * balanced.
3223  *
3224  * Returns the order kswapd finished reclaiming at.
3225  *
3226  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3227  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3228  * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3229  * or lower is eligible for reclaim until at least one usable zone is
3230  * balanced.
3231  */
3232 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3233 {
3234         int i;
3235         unsigned long nr_soft_reclaimed;
3236         unsigned long nr_soft_scanned;
3237         struct zone *zone;
3238         struct scan_control sc = {
3239                 .gfp_mask = GFP_KERNEL,
3240                 .order = order,
3241                 .priority = DEF_PRIORITY,
3242                 .may_writepage = !laptop_mode,
3243                 .may_unmap = 1,
3244                 .may_swap = 1,
3245         };
3246         count_vm_event(PAGEOUTRUN);
3247
3248         do {
3249                 unsigned long nr_reclaimed = sc.nr_reclaimed;
3250                 bool raise_priority = true;
3251
3252                 sc.reclaim_idx = classzone_idx;
3253
3254                 /*
3255                  * If the number of buffer_heads exceeds the maximum allowed
3256                  * then consider reclaiming from all zones. This has a dual
3257                  * purpose -- on 64-bit systems it is expected that
3258                  * buffer_heads are stripped during active rotation. On 32-bit
3259                  * systems, highmem pages can pin lowmem memory and shrinking
3260                  * buffers can relieve lowmem pressure. Reclaim may still not
3261                  * go ahead if all eligible zones for the original allocation
3262                  * request are balanced to avoid excessive reclaim from kswapd.
3263                  */
3264                 if (buffer_heads_over_limit) {
3265                         for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3266                                 zone = pgdat->node_zones + i;
3267                                 if (!managed_zone(zone))
3268                                         continue;
3269
3270                                 sc.reclaim_idx = i;
3271                                 break;
3272                         }
3273                 }
3274
3275                 /*
3276                  * Only reclaim if there are no eligible zones. Note that
3277                  * sc.reclaim_idx is not used as buffer_heads_over_limit may
3278                  * have adjusted it.
3279                  */
3280                 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3281                         goto out;
3282
3283                 /*
3284                  * Do some background aging of the anon list, to give
3285                  * pages a chance to be referenced before reclaiming. All
3286                  * pages are rotated regardless of classzone as this is
3287                  * about consistent aging.
3288                  */
3289                 age_active_anon(pgdat, &sc);
3290
3291                 /*
3292                  * If we're getting trouble reclaiming, start doing writepage
3293                  * even in laptop mode.
3294                  */
3295                 if (sc.priority < DEF_PRIORITY - 2)
3296                         sc.may_writepage = 1;
3297
3298                 /* Call soft limit reclaim before calling shrink_node. */
3299                 sc.nr_scanned = 0;
3300                 nr_soft_scanned = 0;
3301                 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3302                                                 sc.gfp_mask, &nr_soft_scanned);
3303                 sc.nr_reclaimed += nr_soft_reclaimed;
3304
3305                 /*
3306                  * There should be no need to raise the scanning priority if
3307                  * enough pages are already being scanned that that high
3308                  * watermark would be met at 100% efficiency.
3309                  */
3310                 if (kswapd_shrink_node(pgdat, &sc))
3311                         raise_priority = false;
3312
3313                 /*
3314                  * If the low watermark is met there is no need for processes
3315                  * to be throttled on pfmemalloc_wait as they should not be
3316                  * able to safely make forward progress. Wake them
3317                  */
3318                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3319                                 allow_direct_reclaim(pgdat))
3320                         wake_up_all(&pgdat->pfmemalloc_wait);
3321
3322                 /* Check if kswapd should be suspending */
3323                 if (try_to_freeze() || kthread_should_stop())
3324                         break;
3325
3326                 /*