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