blob: 532a2a750952daffed4e4242d36a449eb9ba8837 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080049#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070055#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
Nick Piggin0f8053a2006-03-22 00:08:33 -080057#include "internal.h"
58
Mel Gorman33906bc2010-08-09 17:19:16 -070059#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
Linus Torvalds1da177e2005-04-16 15:20:36 -070062struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080063 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040067 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
Johannes Weineree814fe2014-08-06 16:06:19 -070069 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070070 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080071
Johannes Weineree814fe2014-08-06 16:06:19 -070072 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070077
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070078 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080079 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080083
Johannes Weineree814fe2014-08-06 16:06:19 -070084 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
Mel Gormanb2e18752016-07-28 15:45:37 -070087 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
89
Johannes Weineree814fe2014-08-06 16:06:19 -070090 unsigned int may_writepage:1;
91
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
94
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
97
Johannes Weiner241994ed2015-02-11 15:26:06 -080098 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
100
Johannes Weineree814fe2014-08-06 16:06:19 -0700101 unsigned int hibernation_mode:1;
102
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
105
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
108
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111};
112
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113#ifdef ARCH_HAS_PREFETCH
114#define prefetch_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127#ifdef ARCH_HAS_PREFETCHW
128#define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
132 \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
135 } \
136 } while (0)
137#else
138#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139#endif
140
141/*
142 * From 0 .. 100. Higher means more swappy.
143 */
144int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700145/*
146 * The total number of pages which are beyond the high watermark within all
147 * zones.
148 */
149unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150
151static LIST_HEAD(shrinker_list);
152static DECLARE_RWSEM(shrinker_rwsem);
153
Andrew Mortonc255a452012-07-31 16:43:02 -0700154#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155static bool global_reclaim(struct scan_control *sc)
156{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800157 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800158}
Tejun Heo97c93412015-05-22 18:23:36 -0400159
160/**
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
163 *
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
169 *
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
172 */
173static bool sane_reclaim(struct scan_control *sc)
174{
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
176
177 if (!memcg)
178 return true;
179#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400181 return true;
182#endif
183 return false;
184}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800185#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800186static bool global_reclaim(struct scan_control *sc)
187{
188 return true;
189}
Tejun Heo97c93412015-05-22 18:23:36 -0400190
191static bool sane_reclaim(struct scan_control *sc)
192{
193 return true;
194}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800195#endif
196
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700197/*
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
201 */
202unsigned long zone_reclaimable_pages(struct zone *zone)
203{
204 unsigned long nr;
205
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
211
212 return nr;
213}
214
Mel Gorman599d0c92016-07-28 15:45:31 -0700215unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
Lisa Du6e543d52013-09-11 14:22:36 -0700216{
Mel Gorman599d0c92016-07-28 15:45:31 -0700217 unsigned long nr;
218
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
222
223 if (get_nr_swap_pages() > 0)
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
227
228 return nr;
229}
230
231bool pgdat_reclaimable(struct pglist_data *pgdat)
232{
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700235}
236
Johannes Weiner23047a92016-03-15 14:57:16 -0700237unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800238{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700239 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700240 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800241
Mel Gorman599d0c92016-07-28 15:45:31 -0700242 return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800243}
244
Michal Hockob4536f0c82017-01-10 16:58:04 -0800245unsigned long lruvec_zone_lru_size(struct lruvec *lruvec, enum lru_list lru,
246 int zone_idx)
247{
248 if (!mem_cgroup_disabled())
249 return mem_cgroup_get_zone_lru_size(lruvec, lru, zone_idx);
250
251 return zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zone_idx],
252 NR_ZONE_LRU_BASE + lru);
253}
254
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000256 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000258int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000260 size_t size = sizeof(*shrinker->nr_deferred);
261
Glauber Costa1d3d4432013-08-28 10:18:04 +1000262 if (shrinker->flags & SHRINKER_NUMA_AWARE)
263 size *= nr_node_ids;
264
265 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
266 if (!shrinker->nr_deferred)
267 return -ENOMEM;
268
Rusty Russell8e1f9362007-07-17 04:03:17 -0700269 down_write(&shrinker_rwsem);
270 list_add_tail(&shrinker->list, &shrinker_list);
271 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000272 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700274EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
276/*
277 * Remove one
278 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700279void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280{
281 down_write(&shrinker_rwsem);
282 list_del(&shrinker->list);
283 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700284 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700286EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287
288#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000289
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800290static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
291 struct shrinker *shrinker,
292 unsigned long nr_scanned,
293 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000295 unsigned long freed = 0;
296 unsigned long long delta;
297 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700298 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000299 long nr;
300 long new_nr;
301 int nid = shrinkctl->nid;
302 long batch_size = shrinker->batch ? shrinker->batch
303 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800304 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000305
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700306 freeable = shrinker->count_objects(shrinker, shrinkctl);
307 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000308 return 0;
309
310 /*
311 * copy the current shrinker scan count into a local variable
312 * and zero it so that other concurrent shrinker invocations
313 * don't also do this scanning work.
314 */
315 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
316
317 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800318 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700319 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800320 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000321 total_scan += delta;
322 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800323 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000324 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700325 total_scan = freeable;
Shaohua Li5f33a082016-12-12 16:41:50 -0800326 next_deferred = nr;
327 } else
328 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000329
330 /*
331 * We need to avoid excessive windup on filesystem shrinkers
332 * due to large numbers of GFP_NOFS allocations causing the
333 * shrinkers to return -1 all the time. This results in a large
334 * nr being built up so when a shrink that can do some work
335 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700336 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000337 * memory.
338 *
339 * Hence only allow the shrinker to scan the entire cache when
340 * a large delta change is calculated directly.
341 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700342 if (delta < freeable / 4)
343 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000344
345 /*
346 * Avoid risking looping forever due to too large nr value:
347 * never try to free more than twice the estimate number of
348 * freeable entries.
349 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700350 if (total_scan > freeable * 2)
351 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000352
353 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800354 nr_scanned, nr_eligible,
355 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000356
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800357 /*
358 * Normally, we should not scan less than batch_size objects in one
359 * pass to avoid too frequent shrinker calls, but if the slab has less
360 * than batch_size objects in total and we are really tight on memory,
361 * we will try to reclaim all available objects, otherwise we can end
362 * up failing allocations although there are plenty of reclaimable
363 * objects spread over several slabs with usage less than the
364 * batch_size.
365 *
366 * We detect the "tight on memory" situations by looking at the total
367 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700368 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800369 * scanning at high prio and therefore should try to reclaim as much as
370 * possible.
371 */
372 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700373 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000374 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800375 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000376
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800377 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000378 ret = shrinker->scan_objects(shrinker, shrinkctl);
379 if (ret == SHRINK_STOP)
380 break;
381 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000382
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800383 count_vm_events(SLABS_SCANNED, nr_to_scan);
384 total_scan -= nr_to_scan;
Shaohua Li5f33a082016-12-12 16:41:50 -0800385 scanned += nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000386
387 cond_resched();
388 }
389
Shaohua Li5f33a082016-12-12 16:41:50 -0800390 if (next_deferred >= scanned)
391 next_deferred -= scanned;
392 else
393 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000394 /*
395 * move the unused scan count back into the shrinker in a
396 * manner that handles concurrent updates. If we exhausted the
397 * scan, there is no need to do an update.
398 */
Shaohua Li5f33a082016-12-12 16:41:50 -0800399 if (next_deferred > 0)
400 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000401 &shrinker->nr_deferred[nid]);
402 else
403 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
404
Dave Hansendf9024a2014-06-04 16:08:07 -0700405 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000406 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407}
408
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800409/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800410 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800411 * @gfp_mask: allocation context
412 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800413 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800414 * @nr_scanned: pressure numerator
415 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800417 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800419 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
420 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800422 * @memcg specifies the memory cgroup to target. If it is not NULL,
423 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700424 * objects from the memory cgroup specified. Otherwise, only unaware
425 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800426 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800427 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
428 * the available objects should be scanned. Page reclaim for example
429 * passes the number of pages scanned and the number of pages on the
430 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
431 * when it encountered mapped pages. The ratio is further biased by
432 * the ->seeks setting of the shrink function, which indicates the
433 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800435 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800437static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
438 struct mem_cgroup *memcg,
439 unsigned long nr_scanned,
440 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
442 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000443 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700445 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800446 return 0;
447
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800448 if (nr_scanned == 0)
449 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
Minchan Kimf06590b2011-05-24 17:11:11 -0700451 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000452 /*
453 * If we would return 0, our callers would understand that we
454 * have nothing else to shrink and give up trying. By returning
455 * 1 we keep it going and assume we'll be able to shrink next
456 * time.
457 */
458 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700459 goto out;
460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
462 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800463 struct shrink_control sc = {
464 .gfp_mask = gfp_mask,
465 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800466 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800467 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800468
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700469 /*
470 * If kernel memory accounting is disabled, we ignore
471 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
472 * passing NULL for memcg.
473 */
474 if (memcg_kmem_enabled() &&
475 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800476 continue;
477
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800478 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
479 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000480
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800481 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800483
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700485out:
486 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000487 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488}
489
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800490void drop_slab_node(int nid)
491{
492 unsigned long freed;
493
494 do {
495 struct mem_cgroup *memcg = NULL;
496
497 freed = 0;
498 do {
499 freed += shrink_slab(GFP_KERNEL, nid, memcg,
500 1000, 1000);
501 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
502 } while (freed > 10);
503}
504
505void drop_slab(void)
506{
507 int nid;
508
509 for_each_online_node(nid)
510 drop_slab_node(nid);
511}
512
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513static inline int is_page_cache_freeable(struct page *page)
514{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700515 /*
516 * A freeable page cache page is referenced only by the caller
517 * that isolated the page, the page cache radix tree and
518 * optional buffer heads at page->private.
519 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700520 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Tejun Heo703c2702015-05-22 17:13:44 -0400523static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Christoph Lameter930d9152006-01-08 01:00:47 -0800525 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400527 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400529 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 return 1;
531 return 0;
532}
533
534/*
535 * We detected a synchronous write error writing a page out. Probably
536 * -ENOSPC. We need to propagate that into the address_space for a subsequent
537 * fsync(), msync() or close().
538 *
539 * The tricky part is that after writepage we cannot touch the mapping: nothing
540 * prevents it from being freed up. But we have a ref on the page and once
541 * that page is locked, the mapping is pinned.
542 *
543 * We're allowed to run sleeping lock_page() here because we know the caller has
544 * __GFP_FS.
545 */
546static void handle_write_error(struct address_space *mapping,
547 struct page *page, int error)
548{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100549 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700550 if (page_mapping(page) == mapping)
551 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 unlock_page(page);
553}
554
Christoph Lameter04e62a22006-06-23 02:03:38 -0700555/* possible outcome of pageout() */
556typedef enum {
557 /* failed to write page out, page is locked */
558 PAGE_KEEP,
559 /* move page to the active list, page is locked */
560 PAGE_ACTIVATE,
561 /* page has been sent to the disk successfully, page is unlocked */
562 PAGE_SUCCESS,
563 /* page is clean and locked */
564 PAGE_CLEAN,
565} pageout_t;
566
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567/*
Andrew Morton1742f192006-03-22 00:08:21 -0800568 * pageout is called by shrink_page_list() for each dirty page.
569 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700571static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700572 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573{
574 /*
575 * If the page is dirty, only perform writeback if that write
576 * will be non-blocking. To prevent this allocation from being
577 * stalled by pagecache activity. But note that there may be
578 * stalls if we need to run get_block(). We could test
579 * PagePrivate for that.
580 *
Al Viro81742022014-04-03 03:17:43 -0400581 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 * this page's queue, we can perform writeback even if that
583 * will block.
584 *
585 * If the page is swapcache, write it back even if that would
586 * block, for some throttling. This happens by accident, because
587 * swap_backing_dev_info is bust: it doesn't reflect the
588 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 */
590 if (!is_page_cache_freeable(page))
591 return PAGE_KEEP;
592 if (!mapping) {
593 /*
594 * Some data journaling orphaned pages can have
595 * page->mapping == NULL while being dirty with clean buffers.
596 */
David Howells266cf652009-04-03 16:42:36 +0100597 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 if (try_to_free_buffers(page)) {
599 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700600 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 return PAGE_CLEAN;
602 }
603 }
604 return PAGE_KEEP;
605 }
606 if (mapping->a_ops->writepage == NULL)
607 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400608 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 return PAGE_KEEP;
610
611 if (clear_page_dirty_for_io(page)) {
612 int res;
613 struct writeback_control wbc = {
614 .sync_mode = WB_SYNC_NONE,
615 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700616 .range_start = 0,
617 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 .for_reclaim = 1,
619 };
620
621 SetPageReclaim(page);
622 res = mapping->a_ops->writepage(page, &wbc);
623 if (res < 0)
624 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800625 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 ClearPageReclaim(page);
627 return PAGE_ACTIVATE;
628 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700629
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 if (!PageWriteback(page)) {
631 /* synchronous write or broken a_ops? */
632 ClearPageReclaim(page);
633 }
yalin wang3aa23852016-01-14 15:18:30 -0800634 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700635 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 return PAGE_SUCCESS;
637 }
638
639 return PAGE_CLEAN;
640}
641
Andrew Mortona649fd92006-10-17 00:09:36 -0700642/*
Nick Piggine2867812008-07-25 19:45:30 -0700643 * Same as remove_mapping, but if the page is removed from the mapping, it
644 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700645 */
Johannes Weinera5289102014-04-03 14:47:51 -0700646static int __remove_mapping(struct address_space *mapping, struct page *page,
647 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800648{
Greg Thelenc4843a72015-05-22 17:13:16 -0400649 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400650
Nick Piggin28e4d962006-09-25 23:31:23 -0700651 BUG_ON(!PageLocked(page));
652 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800653
Greg Thelenc4843a72015-05-22 17:13:16 -0400654 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800655 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700656 * The non racy check for a busy page.
657 *
658 * Must be careful with the order of the tests. When someone has
659 * a ref to the page, it may be possible that they dirty it then
660 * drop the reference. So if PageDirty is tested before page_count
661 * here, then the following race may occur:
662 *
663 * get_user_pages(&page);
664 * [user mapping goes away]
665 * write_to(page);
666 * !PageDirty(page) [good]
667 * SetPageDirty(page);
668 * put_page(page);
669 * !page_count(page) [good, discard it]
670 *
671 * [oops, our write_to data is lost]
672 *
673 * Reversing the order of the tests ensures such a situation cannot
674 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700675 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700676 *
677 * Note that if SetPageDirty is always performed via set_page_dirty,
678 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800679 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700680 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800681 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700682 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
683 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700684 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800685 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700686 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800687
688 if (PageSwapCache(page)) {
689 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700690 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800691 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400692 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700693 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700694 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500695 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700696 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500697
698 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700699 /*
700 * Remember a shadow entry for reclaimed file cache in
701 * order to detect refaults, thus thrashing, later on.
702 *
703 * But don't store shadows in an address space that is
704 * already exiting. This is not just an optizimation,
705 * inode reclaim needs to empty out the radix tree or
706 * the nodes are lost. Don't plant shadows behind its
707 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800708 *
709 * We also don't store shadows for DAX mappings because the
710 * only page cache pages found in these are zero pages
711 * covering holes, and because we don't want to mix DAX
712 * exceptional entries and shadow exceptional entries in the
713 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700714 */
715 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800716 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700717 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700718 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400719 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500720
721 if (freepage != NULL)
722 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800723 }
724
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800725 return 1;
726
727cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400728 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800729 return 0;
730}
731
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732/*
Nick Piggine2867812008-07-25 19:45:30 -0700733 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
734 * someone else has a ref on the page, abort and return 0. If it was
735 * successfully detached, return 1. Assumes the caller has a single ref on
736 * this page.
737 */
738int remove_mapping(struct address_space *mapping, struct page *page)
739{
Johannes Weinera5289102014-04-03 14:47:51 -0700740 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700741 /*
742 * Unfreezing the refcount with 1 rather than 2 effectively
743 * drops the pagecache ref for us without requiring another
744 * atomic operation.
745 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700746 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700747 return 1;
748 }
749 return 0;
750}
751
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700752/**
753 * putback_lru_page - put previously isolated page onto appropriate LRU list
754 * @page: page to be put back to appropriate lru list
755 *
756 * Add previously isolated @page to appropriate LRU list.
757 * Page may still be unevictable for other reasons.
758 *
759 * lru_lock must not be held, interrupts must be enabled.
760 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700761void putback_lru_page(struct page *page)
762{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700763 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700764 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700765
Sasha Levin309381fea2014-01-23 15:52:54 -0800766 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700767
768redo:
769 ClearPageUnevictable(page);
770
Hugh Dickins39b5f292012-10-08 16:33:18 -0700771 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700772 /*
773 * For evictable pages, we can use the cache.
774 * In event of a race, worst case is we end up with an
775 * unevictable page on [in]active list.
776 * We know how to handle that.
777 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700778 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700779 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700780 } else {
781 /*
782 * Put unevictable pages directly on zone's unevictable
783 * list.
784 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700785 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700786 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700787 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700788 * When racing with an mlock or AS_UNEVICTABLE clearing
789 * (page is unlocked) make sure that if the other thread
790 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800791 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700792 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700793 * the page back to the evictable list.
794 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700795 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700796 */
797 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700798 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700799
800 /*
801 * page's status can change while we move it among lru. If an evictable
802 * page is on unevictable list, it never be freed. To avoid that,
803 * check after we added it to the list, again.
804 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700805 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700806 if (!isolate_lru_page(page)) {
807 put_page(page);
808 goto redo;
809 }
810 /* This means someone else dropped this page from LRU
811 * So, it will be freed or putback to LRU again. There is
812 * nothing to do here.
813 */
814 }
815
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700816 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700817 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700818 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700819 count_vm_event(UNEVICTABLE_PGCULLED);
820
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700821 put_page(page); /* drop ref from isolate */
822}
823
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800824enum page_references {
825 PAGEREF_RECLAIM,
826 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800827 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800828 PAGEREF_ACTIVATE,
829};
830
831static enum page_references page_check_references(struct page *page,
832 struct scan_control *sc)
833{
Johannes Weiner645747462010-03-05 13:42:22 -0800834 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800835 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800836
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700837 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
838 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800839 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800840
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800841 /*
842 * Mlock lost the isolation race with us. Let try_to_unmap()
843 * move the page to the unevictable list.
844 */
845 if (vm_flags & VM_LOCKED)
846 return PAGEREF_RECLAIM;
847
Johannes Weiner645747462010-03-05 13:42:22 -0800848 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700849 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800850 return PAGEREF_ACTIVATE;
851 /*
852 * All mapped pages start out with page table
853 * references from the instantiating fault, so we need
854 * to look twice if a mapped file page is used more
855 * than once.
856 *
857 * Mark it and spare it for another trip around the
858 * inactive list. Another page table reference will
859 * lead to its activation.
860 *
861 * Note: the mark is set for activated pages as well
862 * so that recently deactivated but used pages are
863 * quickly recovered.
864 */
865 SetPageReferenced(page);
866
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800867 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800868 return PAGEREF_ACTIVATE;
869
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800870 /*
871 * Activate file-backed executable pages after first usage.
872 */
873 if (vm_flags & VM_EXEC)
874 return PAGEREF_ACTIVATE;
875
Johannes Weiner645747462010-03-05 13:42:22 -0800876 return PAGEREF_KEEP;
877 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800878
879 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700880 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800881 return PAGEREF_RECLAIM_CLEAN;
882
883 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800884}
885
Mel Gormane2be15f2013-07-03 15:01:57 -0700886/* Check if a page is dirty or under writeback */
887static void page_check_dirty_writeback(struct page *page,
888 bool *dirty, bool *writeback)
889{
Mel Gormanb4597222013-07-03 15:02:05 -0700890 struct address_space *mapping;
891
Mel Gormane2be15f2013-07-03 15:01:57 -0700892 /*
893 * Anonymous pages are not handled by flushers and must be written
894 * from reclaim context. Do not stall reclaim based on them
895 */
896 if (!page_is_file_cache(page)) {
897 *dirty = false;
898 *writeback = false;
899 return;
900 }
901
902 /* By default assume that the page flags are accurate */
903 *dirty = PageDirty(page);
904 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700905
906 /* Verify dirty/writeback state if the filesystem supports it */
907 if (!page_has_private(page))
908 return;
909
910 mapping = page_mapping(page);
911 if (mapping && mapping->a_ops->is_dirty_writeback)
912 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700913}
914
Nick Piggine2867812008-07-25 19:45:30 -0700915/*
Andrew Morton1742f192006-03-22 00:08:21 -0800916 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 */
Andrew Morton1742f192006-03-22 00:08:21 -0800918static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700919 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700920 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700921 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700922 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700923 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700924 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700925 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700926 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700927 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928{
929 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700930 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700932 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700933 unsigned long nr_dirty = 0;
934 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800935 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700936 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700937 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
939 cond_resched();
940
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 while (!list_empty(page_list)) {
942 struct address_space *mapping;
943 struct page *page;
944 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700945 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700946 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800947 bool lazyfree = false;
948 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
950 cond_resched();
951
952 page = lru_to_page(page_list);
953 list_del(&page->lru);
954
Nick Piggin529ae9a2008-08-02 12:01:03 +0200955 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 goto keep;
957
Sasha Levin309381fea2014-01-23 15:52:54 -0800958 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959
960 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800961
Hugh Dickins39b5f292012-10-08 16:33:18 -0700962 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700963 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700964
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700965 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800966 goto keep_locked;
967
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 /* Double the slab pressure for mapped and swapcache pages */
969 if (page_mapped(page) || PageSwapCache(page))
970 sc->nr_scanned++;
971
Andy Whitcroftc661b072007-08-22 14:01:26 -0700972 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
973 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
974
Mel Gorman283aba92013-07-03 15:01:51 -0700975 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700976 * The number of dirty pages determines if a zone is marked
977 * reclaim_congested which affects wait_iff_congested. kswapd
978 * will stall and start writing pages if the tail of the LRU
979 * is all dirty unqueued pages.
980 */
981 page_check_dirty_writeback(page, &dirty, &writeback);
982 if (dirty || writeback)
983 nr_dirty++;
984
985 if (dirty && !writeback)
986 nr_unqueued_dirty++;
987
Mel Gormand04e8ac2013-07-03 15:02:03 -0700988 /*
989 * Treat this page as congested if the underlying BDI is or if
990 * pages are cycling through the LRU so quickly that the
991 * pages marked for immediate reclaim are making it to the
992 * end of the LRU a second time.
993 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700994 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800995 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400996 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700997 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700998 nr_congested++;
999
1000 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001001 * If a page at the tail of the LRU is under writeback, there
1002 * are three cases to consider.
1003 *
1004 * 1) If reclaim is encountering an excessive number of pages
1005 * under writeback and this page is both under writeback and
1006 * PageReclaim then it indicates that pages are being queued
1007 * for IO but are being recycled through the LRU before the
1008 * IO can complete. Waiting on the page itself risks an
1009 * indefinite stall if it is impossible to writeback the
1010 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001011 * note that the LRU is being scanned too quickly and the
1012 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001013 *
Tejun Heo97c93412015-05-22 18:23:36 -04001014 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001015 * not marked for immediate reclaim, or the caller does not
1016 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1017 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001018 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001019 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001020 * Require may_enter_fs because we would wait on fs, which
1021 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001022 * enter reclaim, and deadlock if it waits on a page for
1023 * which it is needed to do the write (loop masks off
1024 * __GFP_IO|__GFP_FS for this reason); but more thought
1025 * would probably show more reasons.
1026 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001027 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001028 * PageReclaim. memcg does not have any dirty pages
1029 * throttling so we could easily OOM just because too many
1030 * pages are in writeback and there is nothing else to
1031 * reclaim. Wait for the writeback to complete.
1032 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001033 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001034 /* Case 1 above */
1035 if (current_is_kswapd() &&
1036 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001037 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001038 nr_immediate++;
1039 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001040
1041 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001042 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001043 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001044 /*
1045 * This is slightly racy - end_page_writeback()
1046 * might have just cleared PageReclaim, then
1047 * setting PageReclaim here end up interpreted
1048 * as PageReadahead - but that does not matter
1049 * enough to care. What we do want is for this
1050 * page to have PageReclaim set next time memcg
1051 * reclaim reaches the tests above, so it will
1052 * then wait_on_page_writeback() to avoid OOM;
1053 * and it's also appropriate in global reclaim.
1054 */
1055 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001056 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001057 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001058
1059 /* Case 3 above */
1060 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001061 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001062 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001063 /* then go back and try same page again */
1064 list_add_tail(&page->lru, page_list);
1065 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001066 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001067 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
Minchan Kim02c6de82012-10-08 16:31:55 -07001069 if (!force_reclaim)
1070 references = page_check_references(page, sc);
1071
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001072 switch (references) {
1073 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001075 case PAGEREF_KEEP:
1076 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001077 case PAGEREF_RECLAIM:
1078 case PAGEREF_RECLAIM_CLEAN:
1079 ; /* try to reclaim the page below */
1080 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 /*
1083 * Anonymous process memory has backing store?
1084 * Try to allocate it some swap space here.
1085 */
Nick Pigginb291f002008-10-18 20:26:44 -07001086 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001087 if (!(sc->gfp_mask & __GFP_IO))
1088 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001089 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001091 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001092 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093
Mel Gormane2be15f2013-07-03 15:01:57 -07001094 /* Adding to swap updated mapping */
1095 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001096 } else if (unlikely(PageTransHuge(page))) {
1097 /* Split file THP */
1098 if (split_huge_page_to_list(page, page_list))
1099 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001100 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001102 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1103
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 /*
1105 * The page is mapped into the page tables of one or more
1106 * processes. Try to unmap it here.
1107 */
1108 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001109 switch (ret = try_to_unmap(page, lazyfree ?
1110 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1111 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 case SWAP_FAIL:
1113 goto activate_locked;
1114 case SWAP_AGAIN:
1115 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001116 case SWAP_MLOCK:
1117 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001118 case SWAP_LZFREE:
1119 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 case SWAP_SUCCESS:
1121 ; /* try to free the page below */
1122 }
1123 }
1124
1125 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001126 /*
1127 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001128 * avoid risk of stack overflow but only writeback
1129 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001130 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001131 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001132 (!current_is_kswapd() ||
Mel Gorman599d0c92016-07-28 15:45:31 -07001133 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001134 /*
1135 * Immediately reclaim when written back.
1136 * Similar in principal to deactivate_page()
1137 * except we already have the page isolated
1138 * and know it's dirty
1139 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001140 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001141 SetPageReclaim(page);
1142
Mel Gormanee728862011-10-31 17:07:38 -07001143 goto keep_locked;
1144 }
1145
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001146 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001148 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001150 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 goto keep_locked;
1152
Mel Gormand950c942015-09-04 15:47:35 -07001153 /*
1154 * Page is dirty. Flush the TLB if a writable entry
1155 * potentially exists to avoid CPU writes after IO
1156 * starts and then write it out here.
1157 */
1158 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001159 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 case PAGE_KEEP:
1161 goto keep_locked;
1162 case PAGE_ACTIVATE:
1163 goto activate_locked;
1164 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001165 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001166 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001167 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001169
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 /*
1171 * A synchronous write - probably a ramdisk. Go
1172 * ahead and try to reclaim the page.
1173 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001174 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 goto keep;
1176 if (PageDirty(page) || PageWriteback(page))
1177 goto keep_locked;
1178 mapping = page_mapping(page);
1179 case PAGE_CLEAN:
1180 ; /* try to free the page below */
1181 }
1182 }
1183
1184 /*
1185 * If the page has buffers, try to free the buffer mappings
1186 * associated with this page. If we succeed we try to free
1187 * the page as well.
1188 *
1189 * We do this even if the page is PageDirty().
1190 * try_to_release_page() does not perform I/O, but it is
1191 * possible for a page to have PageDirty set, but it is actually
1192 * clean (all its buffers are clean). This happens if the
1193 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001194 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 * try_to_release_page() will discover that cleanness and will
1196 * drop the buffers and mark the page clean - it can be freed.
1197 *
1198 * Rarely, pages can have buffers and no ->mapping. These are
1199 * the pages which were not successfully invalidated in
1200 * truncate_complete_page(). We try to drop those buffers here
1201 * and if that worked, and the page is no longer mapped into
1202 * process address space (page_count == 1) it can be freed.
1203 * Otherwise, leave the page on the LRU so it is swappable.
1204 */
David Howells266cf652009-04-03 16:42:36 +01001205 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 if (!try_to_release_page(page, sc->gfp_mask))
1207 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001208 if (!mapping && page_count(page) == 1) {
1209 unlock_page(page);
1210 if (put_page_testzero(page))
1211 goto free_it;
1212 else {
1213 /*
1214 * rare race with speculative reference.
1215 * the speculative reference will free
1216 * this page shortly, so we may
1217 * increment nr_reclaimed here (and
1218 * leave it off the LRU).
1219 */
1220 nr_reclaimed++;
1221 continue;
1222 }
1223 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 }
1225
Minchan Kim854e9ed2016-01-15 16:54:53 -08001226lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001227 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001228 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
Nick Piggina978d6f2008-10-18 20:26:58 -07001230 /*
1231 * At this point, we have no other references and there is
1232 * no way to pick any more up (removed from LRU, removed
1233 * from pagecache). Can use non-atomic bitops now (and
1234 * we obviously don't have to worry about waking up a process
1235 * waiting on the page lock, because there are no references.
1236 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001237 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001238free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001239 if (ret == SWAP_LZFREE)
1240 count_vm_event(PGLAZYFREED);
1241
Andrew Morton05ff5132006-03-22 00:08:20 -08001242 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001243
1244 /*
1245 * Is there need to periodically free_page_list? It would
1246 * appear not as the counts should be low
1247 */
1248 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 continue;
1250
Nick Pigginb291f002008-10-18 20:26:44 -07001251cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001252 if (PageSwapCache(page))
1253 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001254 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001255 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001256 continue;
1257
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001259 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001260 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001261 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001262 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 SetPageActive(page);
1264 pgactivate++;
1265keep_locked:
1266 unlock_page(page);
1267keep:
1268 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001269 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001271
Johannes Weiner747db952014-08-08 14:19:24 -07001272 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001273 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001274 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001275
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001277 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001278
Mel Gorman8e950282013-07-03 15:02:02 -07001279 *ret_nr_dirty += nr_dirty;
1280 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001281 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001282 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001283 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001284 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Minchan Kim02c6de82012-10-08 16:31:55 -07001287unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1288 struct list_head *page_list)
1289{
1290 struct scan_control sc = {
1291 .gfp_mask = GFP_KERNEL,
1292 .priority = DEF_PRIORITY,
1293 .may_unmap = 1,
1294 };
Mel Gorman8e950282013-07-03 15:02:02 -07001295 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001296 struct page *page, *next;
1297 LIST_HEAD(clean_pages);
1298
1299 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001300 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001301 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001302 ClearPageActive(page);
1303 list_move(&page->lru, &clean_pages);
1304 }
1305 }
1306
Mel Gorman599d0c92016-07-28 15:45:31 -07001307 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001308 TTU_UNMAP|TTU_IGNORE_ACCESS,
1309 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001310 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001311 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001312 return ret;
1313}
1314
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001315/*
1316 * Attempt to remove the specified page from its LRU. Only take this page
1317 * if it is of the appropriate PageActive status. Pages which are being
1318 * freed elsewhere are also ignored.
1319 *
1320 * page: page to consider
1321 * mode: one of the LRU isolation modes defined above
1322 *
1323 * returns 0 on success, -ve errno on failure.
1324 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001325int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001326{
1327 int ret = -EINVAL;
1328
1329 /* Only take pages on the LRU. */
1330 if (!PageLRU(page))
1331 return ret;
1332
Minchan Kime46a2872012-10-08 16:33:48 -07001333 /* Compaction should not handle unevictable pages but CMA can do so */
1334 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001335 return ret;
1336
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001337 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001338
Mel Gormanc8244932012-01-12 17:19:38 -08001339 /*
1340 * To minimise LRU disruption, the caller can indicate that it only
1341 * wants to isolate pages it will be able to operate on without
1342 * blocking - clean pages for the most part.
1343 *
1344 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1345 * is used by reclaim when it is cannot write to backing storage
1346 *
1347 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1348 * that it is possible to migrate without blocking
1349 */
1350 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1351 /* All the caller can do on PageWriteback is block */
1352 if (PageWriteback(page))
1353 return ret;
1354
1355 if (PageDirty(page)) {
1356 struct address_space *mapping;
1357
1358 /* ISOLATE_CLEAN means only clean pages */
1359 if (mode & ISOLATE_CLEAN)
1360 return ret;
1361
1362 /*
1363 * Only pages without mappings or that have a
1364 * ->migratepage callback are possible to migrate
1365 * without blocking
1366 */
1367 mapping = page_mapping(page);
1368 if (mapping && !mapping->a_ops->migratepage)
1369 return ret;
1370 }
1371 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001372
Minchan Kimf80c0672011-10-31 17:06:55 -07001373 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1374 return ret;
1375
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001376 if (likely(get_page_unless_zero(page))) {
1377 /*
1378 * Be careful not to clear PageLRU until after we're
1379 * sure the page is not being freed elsewhere -- the
1380 * page release code relies on it.
1381 */
1382 ClearPageLRU(page);
1383 ret = 0;
1384 }
1385
1386 return ret;
1387}
1388
Mel Gorman7ee36a12016-07-28 15:47:17 -07001389
1390/*
1391 * Update LRU sizes after isolating pages. The LRU size updates must
1392 * be complete before mem_cgroup_update_lru_size due to a santity check.
1393 */
1394static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001395 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001396{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001397 int zid;
1398
Mel Gorman7ee36a12016-07-28 15:47:17 -07001399 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1400 if (!nr_zone_taken[zid])
1401 continue;
1402
1403 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001404#ifdef CONFIG_MEMCG
1405 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1406#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001407 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001408
Mel Gorman7ee36a12016-07-28 15:47:17 -07001409}
1410
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001411/*
Mel Gormana52633d2016-07-28 15:45:28 -07001412 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 * shrink the lists perform better by taking out a batch of pages
1414 * and working on them outside the LRU lock.
1415 *
1416 * For pagecache intensive workloads, this function is the hottest
1417 * spot in the kernel (apart from copy_*_user functions).
1418 *
1419 * Appropriate locks must be held before calling this function.
1420 *
1421 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001422 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001424 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001425 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001426 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001427 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 *
1429 * returns how many pages were moved onto *@dst.
1430 */
Andrew Morton69e05942006-03-22 00:08:19 -08001431static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001432 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001433 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001434 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001436 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001437 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001438 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001439 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Mel Gorman599d0c92016-07-28 15:45:31 -07001440 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001441 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001443 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Mel Gormand7f05522016-07-28 15:47:37 -07001444 !list_empty(src);) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001445 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001446
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 page = lru_to_page(src);
1448 prefetchw_prev_lru_page(page, src, flags);
1449
Sasha Levin309381fea2014-01-23 15:52:54 -08001450 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001451
Mel Gormanb2e18752016-07-28 15:45:37 -07001452 if (page_zonenum(page) > sc->reclaim_idx) {
1453 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001454 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001455 continue;
1456 }
1457
Mel Gormand7f05522016-07-28 15:47:37 -07001458 /*
1459 * Account for scanned and skipped separetly to avoid the pgdat
1460 * being prematurely marked unreclaimable by pgdat_reclaimable.
1461 */
1462 scan++;
1463
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001464 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001465 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001466 nr_pages = hpage_nr_pages(page);
1467 nr_taken += nr_pages;
1468 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001469 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001470 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001471
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001472 case -EBUSY:
1473 /* else it is being freed elsewhere */
1474 list_move(&page->lru, src);
1475 continue;
1476
1477 default:
1478 BUG();
1479 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 }
1481
Mel Gormanb2e18752016-07-28 15:45:37 -07001482 /*
1483 * Splice any skipped pages to the start of the LRU list. Note that
1484 * this disrupts the LRU order when reclaiming for lower zones but
1485 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1486 * scanning would soon rescan the same pages to skip and put the
1487 * system at risk of premature OOM.
1488 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001489 if (!list_empty(&pages_skipped)) {
1490 int zid;
Mel Gormand7f05522016-07-28 15:47:37 -07001491 unsigned long total_skipped = 0;
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001492
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001493 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1494 if (!nr_skipped[zid])
1495 continue;
1496
1497 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Mel Gormand7f05522016-07-28 15:47:37 -07001498 total_skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001499 }
Mel Gormand7f05522016-07-28 15:47:37 -07001500
1501 /*
1502 * Account skipped pages as a partial scan as the pgdat may be
1503 * close to unreclaimable. If the LRU list is empty, account
1504 * skipped pages as a full scan.
1505 */
1506 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1507
1508 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001509 }
Hugh Dickinsf6260122012-01-12 17:20:06 -08001510 *nr_scanned = scan;
Mel Gormane5146b12016-07-28 15:46:47 -07001511 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
Hugh Dickins75b00af2012-05-29 15:07:09 -07001512 nr_taken, mode, is_file_lru(lru));
Michal Hockob4536f0c82017-01-10 16:58:04 -08001513 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 return nr_taken;
1515}
1516
Nick Piggin62695a82008-10-18 20:26:09 -07001517/**
1518 * isolate_lru_page - tries to isolate a page from its LRU list
1519 * @page: page to isolate from its LRU list
1520 *
1521 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1522 * vmstat statistic corresponding to whatever LRU list the page was on.
1523 *
1524 * Returns 0 if the page was removed from an LRU list.
1525 * Returns -EBUSY if the page was not on an LRU list.
1526 *
1527 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001528 * the active list, it will have PageActive set. If it was found on
1529 * the unevictable list, it will have the PageUnevictable bit set. That flag
1530 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001531 *
1532 * The vmstat statistic corresponding to the list on which the page was
1533 * found will be decremented.
1534 *
1535 * Restrictions:
1536 * (1) Must be called with an elevated refcount on the page. This is a
1537 * fundamentnal difference from isolate_lru_pages (which is called
1538 * without a stable reference).
1539 * (2) the lru_lock must not be held.
1540 * (3) interrupts must be enabled.
1541 */
1542int isolate_lru_page(struct page *page)
1543{
1544 int ret = -EBUSY;
1545
Sasha Levin309381fea2014-01-23 15:52:54 -08001546 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001547 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001548
Nick Piggin62695a82008-10-18 20:26:09 -07001549 if (PageLRU(page)) {
1550 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001551 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001552
Mel Gormana52633d2016-07-28 15:45:28 -07001553 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001554 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001555 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001556 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001557 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001558 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001559 del_page_from_lru_list(page, lruvec, lru);
1560 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001561 }
Mel Gormana52633d2016-07-28 15:45:28 -07001562 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001563 }
1564 return ret;
1565}
1566
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001567/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001568 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1569 * then get resheduled. When there are massive number of tasks doing page
1570 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1571 * the LRU list will go small and be scanned faster than necessary, leading to
1572 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001573 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001574static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001575 struct scan_control *sc)
1576{
1577 unsigned long inactive, isolated;
1578
1579 if (current_is_kswapd())
1580 return 0;
1581
Tejun Heo97c93412015-05-22 18:23:36 -04001582 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001583 return 0;
1584
1585 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001586 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1587 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001588 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001589 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1590 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001591 }
1592
Fengguang Wu3cf23842012-12-18 14:23:31 -08001593 /*
1594 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1595 * won't get blocked by normal direct-reclaimers, forming a circular
1596 * deadlock.
1597 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001598 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001599 inactive >>= 3;
1600
Rik van Riel35cd7812009-09-21 17:01:38 -07001601 return isolated > inactive;
1602}
1603
Mel Gorman66635622010-08-09 17:19:30 -07001604static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001605putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001606{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001607 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001608 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001609 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001610
Mel Gorman66635622010-08-09 17:19:30 -07001611 /*
1612 * Put back any unfreeable pages.
1613 */
Mel Gorman66635622010-08-09 17:19:30 -07001614 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001615 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001616 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001617
Sasha Levin309381fea2014-01-23 15:52:54 -08001618 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001619 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001620 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001621 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001622 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001623 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001624 continue;
1625 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001626
Mel Gorman599d0c92016-07-28 15:45:31 -07001627 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001628
Linus Torvalds7a608572011-01-17 14:42:19 -08001629 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001630 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001631 add_page_to_lru_list(page, lruvec, lru);
1632
Mel Gorman66635622010-08-09 17:19:30 -07001633 if (is_active_lru(lru)) {
1634 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001635 int numpages = hpage_nr_pages(page);
1636 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001637 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001638 if (put_page_testzero(page)) {
1639 __ClearPageLRU(page);
1640 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001641 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001642
1643 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001644 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001645 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001646 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001647 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001648 } else
1649 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001650 }
1651 }
Mel Gorman66635622010-08-09 17:19:30 -07001652
Hugh Dickins3f797682012-01-12 17:20:07 -08001653 /*
1654 * To save our caller's stack, now use input list for pages to free.
1655 */
1656 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001657}
1658
1659/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001660 * If a kernel thread (such as nfsd for loop-back mounts) services
1661 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1662 * In that case we should only throttle if the backing device it is
1663 * writing to is congested. In other cases it is safe to throttle.
1664 */
1665static int current_may_throttle(void)
1666{
1667 return !(current->flags & PF_LESS_THROTTLE) ||
1668 current->backing_dev_info == NULL ||
1669 bdi_write_congested(current->backing_dev_info);
1670}
1671
Minchan Kim91dcade2016-07-28 15:47:40 -07001672static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1673 struct scan_control *sc, enum lru_list lru)
1674{
1675 int zid;
1676 struct zone *zone;
1677 int file = is_file_lru(lru);
1678 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1679
1680 if (!global_reclaim(sc))
1681 return true;
1682
1683 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1684 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001685 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001686 continue;
1687
1688 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1689 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1690 return true;
1691 }
1692
1693 return false;
1694}
1695
NeilBrown399ba0b2014-06-04 16:07:42 -07001696/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001697 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001698 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 */
Mel Gorman66635622010-08-09 17:19:30 -07001700static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001701shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001702 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703{
1704 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001705 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001706 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001707 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001708 unsigned long nr_dirty = 0;
1709 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001710 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001711 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001712 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001713 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001714 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001715 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001716 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001717
Minchan Kim91dcade2016-07-28 15:47:40 -07001718 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1719 return 0;
1720
Mel Gorman599d0c92016-07-28 15:45:31 -07001721 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001722 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001723
1724 /* We are about to die and free our memory. Return now. */
1725 if (fatal_signal_pending(current))
1726 return SWAP_CLUSTER_MAX;
1727 }
1728
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001730
1731 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001732 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001733 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001734 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001735
Mel Gorman599d0c92016-07-28 15:45:31 -07001736 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001738 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1739 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001740
Mel Gorman599d0c92016-07-28 15:45:31 -07001741 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001742 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001743
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001744 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001745 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001746 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001747 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001748 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001749 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001750 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001751 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001752
Hillf Dantond563c052012-03-21 16:34:02 -07001753 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001754 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001755
Mel Gorman599d0c92016-07-28 15:45:31 -07001756 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001757 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1758 &nr_writeback, &nr_immediate,
1759 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001760
Mel Gorman599d0c92016-07-28 15:45:31 -07001761 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001762
Ying Han904249a2012-04-25 16:01:48 -07001763 if (global_reclaim(sc)) {
1764 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001765 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001766 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001767 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001768 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001769
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001770 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001771
Mel Gorman599d0c92016-07-28 15:45:31 -07001772 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001773
Mel Gorman599d0c92016-07-28 15:45:31 -07001774 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001775
Johannes Weiner747db952014-08-08 14:19:24 -07001776 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001777 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001778
Mel Gorman92df3a72011-10-31 17:07:56 -07001779 /*
1780 * If reclaim is isolating dirty pages under writeback, it implies
1781 * that the long-lived page allocation rate is exceeding the page
1782 * laundering rate. Either the global limits are not being effective
1783 * at throttling processes due to the page distribution throughout
1784 * zones or there is heavy usage of a slow backing device. The
1785 * only option is to throttle from reclaim context which is not ideal
1786 * as there is no guarantee the dirtying process is throttled in the
1787 * same way balance_dirty_pages() manages.
1788 *
Mel Gorman8e950282013-07-03 15:02:02 -07001789 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1790 * of pages under pages flagged for immediate reclaim and stall if any
1791 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001792 */
Mel Gorman918fc712013-07-08 16:00:25 -07001793 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001794 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001795
Mel Gormand43006d2013-07-03 15:01:50 -07001796 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001797 * Legacy memcg will stall in page writeback so avoid forcibly
1798 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001799 */
Tejun Heo97c93412015-05-22 18:23:36 -04001800 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001801 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001802 * 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 (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001806 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001807
1808 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001809 * If dirty pages are scanned that are not queued for IO, it
1810 * implies that flushers are not keeping up. In this case, flag
Mel Gorman599d0c92016-07-28 15:45:31 -07001811 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
Johannes Weiner57054652014-10-09 15:28:17 -07001812 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001813 */
1814 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001815 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001816
1817 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001818 * If kswapd scans pages marked marked for immediate
1819 * reclaim and under writeback (nr_immediate), it implies
1820 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001821 * they are written so also forcibly stall.
1822 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001823 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001824 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001825 }
Mel Gormand43006d2013-07-03 15:01:50 -07001826
Mel Gorman8e950282013-07-03 15:02:02 -07001827 /*
1828 * Stall direct reclaim for IO completions if underlying BDIs or zone
1829 * is congested. Allow kswapd to continue until it starts encountering
1830 * unqueued dirty pages or cycling through the LRU too quickly.
1831 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001832 if (!sc->hibernation_mode && !current_is_kswapd() &&
1833 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001834 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001835
Mel Gorman599d0c92016-07-28 15:45:31 -07001836 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1837 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001838 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001839 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840}
1841
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001842/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 * This moves pages from the active list to the inactive list.
1844 *
1845 * We move them the other way if the page is referenced by one or more
1846 * processes, from rmap.
1847 *
1848 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001849 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001851 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 * this, so instead we remove the pages from the LRU while processing them.
1853 * It is safe to rely on PG_active against the non-LRU pages in here because
1854 * nobody will play with that bit on a non-LRU page.
1855 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001856 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 * But we had to alter page->flags anyway.
1858 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001859
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001860static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001861 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001862 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001863 enum lru_list lru)
1864{
Mel Gorman599d0c92016-07-28 15:45:31 -07001865 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001866 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001867 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001868 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001869
Wu Fengguang3eb41402009-06-16 15:33:13 -07001870 while (!list_empty(list)) {
1871 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001872 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001873
Sasha Levin309381fea2014-01-23 15:52:54 -08001874 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001875 SetPageLRU(page);
1876
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001877 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001878 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001879 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001880 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001881
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001882 if (put_page_testzero(page)) {
1883 __ClearPageLRU(page);
1884 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001885 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001886
1887 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001888 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001889 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001890 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001891 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001892 } else
1893 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001894 }
1895 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001896
Wu Fengguang3eb41402009-06-16 15:33:13 -07001897 if (!is_active_lru(lru))
1898 __count_vm_events(PGDEACTIVATE, pgmoved);
1899}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001900
Hugh Dickinsf6260122012-01-12 17:20:06 -08001901static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001902 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001903 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001904 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001906 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001907 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001908 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001910 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001911 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001913 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001914 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001915 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001916 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001917 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
1919 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001920
1921 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001922 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001923 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001924 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001925
Mel Gorman599d0c92016-07-28 15:45:31 -07001926 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001927
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001928 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1929 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001930
Mel Gorman599d0c92016-07-28 15:45:31 -07001931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001932 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001933
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001934 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001935 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1936 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001937
Mel Gorman599d0c92016-07-28 15:45:31 -07001938 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 while (!list_empty(&l_hold)) {
1941 cond_resched();
1942 page = lru_to_page(&l_hold);
1943 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001944
Hugh Dickins39b5f292012-10-08 16:33:18 -07001945 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001946 putback_lru_page(page);
1947 continue;
1948 }
1949
Mel Gormancc715d92012-03-21 16:34:00 -07001950 if (unlikely(buffer_heads_over_limit)) {
1951 if (page_has_private(page) && trylock_page(page)) {
1952 if (page_has_private(page))
1953 try_to_release_page(page, 0);
1954 unlock_page(page);
1955 }
1956 }
1957
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001958 if (page_referenced(page, 0, sc->target_mem_cgroup,
1959 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001960 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001961 /*
1962 * Identify referenced, file-backed active pages and
1963 * give them one more trip around the active list. So
1964 * that executable code get better chances to stay in
1965 * memory under moderate memory pressure. Anon pages
1966 * are not likely to be evicted by use-once streaming
1967 * IO, plus JVM can create lots of anon VM_EXEC pages,
1968 * so we ignore them here.
1969 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001970 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001971 list_add(&page->lru, &l_active);
1972 continue;
1973 }
1974 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001975
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001976 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 list_add(&page->lru, &l_inactive);
1978 }
1979
Andrew Mortonb5557492009-01-06 14:40:13 -08001980 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001981 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001982 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001983 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001984 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001985 * Count referenced pages from currently used mappings as rotated,
1986 * even though only some of them are actually re-activated. This
1987 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001988 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001989 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001990 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001991
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001992 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1993 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07001994 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1995 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001996
Johannes Weiner747db952014-08-08 14:19:24 -07001997 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001998 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
2000
Rik van Riel59dc76b2016-05-20 16:56:31 -07002001/*
2002 * The inactive anon list should be small enough that the VM never has
2003 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002004 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002005 * The inactive file list should be small enough to leave most memory
2006 * to the established workingset on the scan-resistant active list,
2007 * but large enough to avoid thrashing the aggregate readahead window.
2008 *
2009 * Both inactive lists should also be large enough that each inactive
2010 * page has a chance to be referenced again before it is reclaimed.
2011 *
2012 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2013 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2014 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2015 *
2016 * total target max
2017 * memory ratio inactive
2018 * -------------------------------------
2019 * 10MB 1 5MB
2020 * 100MB 1 50MB
2021 * 1GB 3 250MB
2022 * 10GB 10 0.9GB
2023 * 100GB 31 3GB
2024 * 1TB 101 10GB
2025 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002026 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002027static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2028 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002029{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002030 unsigned long inactive_ratio;
2031 unsigned long inactive;
2032 unsigned long active;
2033 unsigned long gb;
Mel Gormanf8d1a312016-07-28 15:47:34 -07002034 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2035 int zid;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002036
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002037 /*
2038 * If we don't have swap space, anonymous page deactivation
2039 * is pointless.
2040 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002041 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002042 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002043
Rik van Riel59dc76b2016-05-20 16:56:31 -07002044 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2045 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002046
Mel Gormanf8d1a312016-07-28 15:47:34 -07002047 /*
2048 * For zone-constrained allocations, it is necessary to check if
2049 * deactivations are required for lowmem to be reclaimed. This
2050 * calculates the inactive/active pages available in eligible zones.
2051 */
2052 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2053 struct zone *zone = &pgdat->node_zones[zid];
2054 unsigned long inactive_zone, active_zone;
2055
Mel Gorman6aa303d2016-09-01 16:14:55 -07002056 if (!managed_zone(zone))
Mel Gormanf8d1a312016-07-28 15:47:34 -07002057 continue;
2058
Michal Hockob4536f0c82017-01-10 16:58:04 -08002059 inactive_zone = lruvec_zone_lru_size(lruvec, file * LRU_FILE, zid);
2060 active_zone = lruvec_zone_lru_size(lruvec, (file * LRU_FILE) + LRU_ACTIVE, zid);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002061
2062 inactive -= min(inactive, inactive_zone);
2063 active -= min(active, active_zone);
2064 }
2065
Rik van Riel59dc76b2016-05-20 16:56:31 -07002066 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2067 if (gb)
2068 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002069 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002070 inactive_ratio = 1;
2071
2072 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002073}
2074
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002075static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002076 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002077{
Rik van Rielb39415b2009-12-14 17:59:48 -08002078 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002079 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002080 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002081 return 0;
2082 }
2083
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002084 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002085}
2086
Johannes Weiner9a265112013-02-22 16:32:17 -08002087enum scan_balance {
2088 SCAN_EQUAL,
2089 SCAN_FRACT,
2090 SCAN_ANON,
2091 SCAN_FILE,
2092};
2093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002095 * Determine how aggressively the anon and file LRU lists should be
2096 * scanned. The relative value of each set of LRU lists is determined
2097 * by looking at the fraction of the pages scanned we did rotate back
2098 * onto the active list instead of evict.
2099 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002100 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2101 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002102 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002103static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002104 struct scan_control *sc, unsigned long *nr,
2105 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002106{
Vladimir Davydov33377672016-01-20 15:02:59 -08002107 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002108 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002109 u64 fraction[2];
2110 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002111 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002112 unsigned long anon_prio, file_prio;
2113 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002114 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002115 bool force_scan = false;
2116 unsigned long ap, fp;
2117 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002118 bool some_scanned;
2119 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07002120
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002121 /*
2122 * If the zone or memcg is small, nr[l] can be 0. This
2123 * results in no scanning on this priority and a potential
2124 * priority drop. Global direct reclaim can go to the next
2125 * zone and tends to have no problems. Global kswapd is for
2126 * zone balancing and it needs to scan a minimum amount. When
2127 * reclaiming for a memcg, a priority drop can cause high
2128 * latencies, so it's better to scan a minimum amount there as
2129 * well.
2130 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002131 if (current_is_kswapd()) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002132 if (!pgdat_reclaimable(pgdat))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002133 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002134 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002135 force_scan = true;
2136 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002137 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002138 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002139
2140 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002141 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002142 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002143 goto out;
2144 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002145
Johannes Weiner10316b32013-02-22 16:32:14 -08002146 /*
2147 * Global reclaim will swap to prevent OOM even with no
2148 * swappiness, but memcg users want to use this knob to
2149 * disable swapping for individual groups completely when
2150 * using the memory controller's swap limit feature would be
2151 * too expensive.
2152 */
Johannes Weiner02695172014-08-06 16:06:17 -07002153 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002154 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002155 goto out;
2156 }
2157
2158 /*
2159 * Do not apply any pressure balancing cleverness when the
2160 * system is close to OOM, scan both anon and file equally
2161 * (unless the swappiness setting disagrees with swapping).
2162 */
Johannes Weiner02695172014-08-06 16:06:17 -07002163 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002164 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002165 goto out;
2166 }
2167
Johannes Weiner11d16c22013-02-22 16:32:15 -08002168 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002169 * Prevent the reclaimer from falling into the cache trap: as
2170 * cache pages start out inactive, every cache fault will tip
2171 * the scan balance towards the file LRU. And as the file LRU
2172 * shrinks, so does the window for rotation from references.
2173 * This means we have a runaway feedback loop where a tiny
2174 * thrashing file LRU becomes infinitely more attractive than
2175 * anon pages. Try to detect this based on file LRU size.
2176 */
2177 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002178 unsigned long pgdatfile;
2179 unsigned long pgdatfree;
2180 int z;
2181 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002182
Mel Gorman599d0c92016-07-28 15:45:31 -07002183 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2184 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2185 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002186
Mel Gorman599d0c92016-07-28 15:45:31 -07002187 for (z = 0; z < MAX_NR_ZONES; z++) {
2188 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002189 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002190 continue;
2191
2192 total_high_wmark += high_wmark_pages(zone);
2193 }
2194
2195 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002196 scan_balance = SCAN_ANON;
2197 goto out;
2198 }
2199 }
2200
2201 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002202 * If there is enough inactive page cache, i.e. if the size of the
2203 * inactive list is greater than that of the active list *and* the
2204 * inactive list actually has some pages to scan on this priority, we
2205 * do not reclaim anything from the anonymous working set right now.
2206 * Without the second condition we could end up never scanning an
2207 * lruvec even if it has plenty of old anonymous pages unless the
2208 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002209 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002210 if (!inactive_list_is_low(lruvec, true, sc) &&
Johannes Weiner23047a92016-03-15 14:57:16 -07002211 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002212 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002213 goto out;
2214 }
2215
Johannes Weiner9a265112013-02-22 16:32:17 -08002216 scan_balance = SCAN_FRACT;
2217
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002218 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002219 * With swappiness at 100, anonymous and file have the same priority.
2220 * This scanning priority is essentially the inverse of IO cost.
2221 */
Johannes Weiner02695172014-08-06 16:06:17 -07002222 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002223 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002224
2225 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002226 * OK, so we have swap space and a fair amount of page cache
2227 * pages. We use the recently rotated / recently scanned
2228 * ratios to determine how valuable each cache is.
2229 *
2230 * Because workloads change over time (and to avoid overflow)
2231 * we keep these statistics as a floating average, which ends
2232 * up weighing recent references more than old ones.
2233 *
2234 * anon in [0], file in [1]
2235 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002236
Johannes Weiner23047a92016-03-15 14:57:16 -07002237 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2238 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2239 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2240 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002241
Mel Gorman599d0c92016-07-28 15:45:31 -07002242 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002243 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002244 reclaim_stat->recent_scanned[0] /= 2;
2245 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002246 }
2247
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002248 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002249 reclaim_stat->recent_scanned[1] /= 2;
2250 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002251 }
2252
2253 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002254 * The amount of pressure on anon vs file pages is inversely
2255 * proportional to the fraction of recently scanned pages on
2256 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002257 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002258 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002259 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002260
Satoru Moriyafe350042012-05-29 15:06:47 -07002261 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002262 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002263 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002264
Shaohua Li76a33fc2010-05-24 14:32:36 -07002265 fraction[0] = ap;
2266 fraction[1] = fp;
2267 denominator = ap + fp + 1;
2268out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002269 some_scanned = false;
2270 /* Only use force_scan on second pass. */
2271 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002272 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002273 for_each_evictable_lru(lru) {
2274 int file = is_file_lru(lru);
2275 unsigned long size;
2276 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002277
Johannes Weiner23047a92016-03-15 14:57:16 -07002278 size = lruvec_lru_size(lruvec, lru);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002279 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002280
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002281 if (!scan && pass && force_scan)
2282 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002283
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002284 switch (scan_balance) {
2285 case SCAN_EQUAL:
2286 /* Scan lists relative to size */
2287 break;
2288 case SCAN_FRACT:
2289 /*
2290 * Scan types proportional to swappiness and
2291 * their relative recent reclaim efficiency.
2292 */
2293 scan = div64_u64(scan * fraction[file],
2294 denominator);
2295 break;
2296 case SCAN_FILE:
2297 case SCAN_ANON:
2298 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002299 if ((scan_balance == SCAN_FILE) != file) {
2300 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002301 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002302 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002303 break;
2304 default:
2305 /* Look ma, no brain */
2306 BUG();
2307 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002308
2309 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002310 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002311
Johannes Weiner9a265112013-02-22 16:32:17 -08002312 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002313 * Skip the second pass and don't force_scan,
2314 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002315 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002316 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002317 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002318 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002319}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002320
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002321/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002322 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002323 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002324static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002325 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002326{
Mel Gormanef8f2322016-07-28 15:46:05 -07002327 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002328 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002329 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002330 unsigned long nr_to_scan;
2331 enum lru_list lru;
2332 unsigned long nr_reclaimed = 0;
2333 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2334 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002335 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002336
Vladimir Davydov33377672016-01-20 15:02:59 -08002337 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002338
Mel Gormane82e0562013-07-03 15:01:44 -07002339 /* Record the original scan target for proportional adjustments later */
2340 memcpy(targets, nr, sizeof(nr));
2341
Mel Gorman1a501902014-06-04 16:10:49 -07002342 /*
2343 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2344 * event that can occur when there is little memory pressure e.g.
2345 * multiple streaming readers/writers. Hence, we do not abort scanning
2346 * when the requested number of pages are reclaimed when scanning at
2347 * DEF_PRIORITY on the assumption that the fact we are direct
2348 * reclaiming implies that kswapd is not keeping up and it is best to
2349 * do a batch of work at once. For memcg reclaim one check is made to
2350 * abort proportional reclaim if either the file or anon lru has already
2351 * dropped to zero at the first pass.
2352 */
2353 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2354 sc->priority == DEF_PRIORITY);
2355
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002356 blk_start_plug(&plug);
2357 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2358 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002359 unsigned long nr_anon, nr_file, percentage;
2360 unsigned long nr_scanned;
2361
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002362 for_each_evictable_lru(lru) {
2363 if (nr[lru]) {
2364 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2365 nr[lru] -= nr_to_scan;
2366
2367 nr_reclaimed += shrink_list(lru, nr_to_scan,
2368 lruvec, sc);
2369 }
2370 }
Mel Gormane82e0562013-07-03 15:01:44 -07002371
Michal Hockobd041732016-12-02 17:26:48 -08002372 cond_resched();
2373
Mel Gormane82e0562013-07-03 15:01:44 -07002374 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2375 continue;
2376
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002377 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002378 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002379 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002380 * proportionally what was requested by get_scan_count(). We
2381 * stop reclaiming one LRU and reduce the amount scanning
2382 * proportional to the original scan target.
2383 */
2384 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2385 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2386
Mel Gorman1a501902014-06-04 16:10:49 -07002387 /*
2388 * It's just vindictive to attack the larger once the smaller
2389 * has gone to zero. And given the way we stop scanning the
2390 * smaller below, this makes sure that we only make one nudge
2391 * towards proportionality once we've got nr_to_reclaim.
2392 */
2393 if (!nr_file || !nr_anon)
2394 break;
2395
Mel Gormane82e0562013-07-03 15:01:44 -07002396 if (nr_file > nr_anon) {
2397 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2398 targets[LRU_ACTIVE_ANON] + 1;
2399 lru = LRU_BASE;
2400 percentage = nr_anon * 100 / scan_target;
2401 } else {
2402 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2403 targets[LRU_ACTIVE_FILE] + 1;
2404 lru = LRU_FILE;
2405 percentage = nr_file * 100 / scan_target;
2406 }
2407
2408 /* Stop scanning the smaller of the LRU */
2409 nr[lru] = 0;
2410 nr[lru + LRU_ACTIVE] = 0;
2411
2412 /*
2413 * Recalculate the other LRU scan count based on its original
2414 * scan target and the percentage scanning already complete
2415 */
2416 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2417 nr_scanned = targets[lru] - nr[lru];
2418 nr[lru] = targets[lru] * (100 - percentage) / 100;
2419 nr[lru] -= min(nr[lru], nr_scanned);
2420
2421 lru += LRU_ACTIVE;
2422 nr_scanned = targets[lru] - nr[lru];
2423 nr[lru] = targets[lru] * (100 - percentage) / 100;
2424 nr[lru] -= min(nr[lru], nr_scanned);
2425
2426 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002427 }
2428 blk_finish_plug(&plug);
2429 sc->nr_reclaimed += nr_reclaimed;
2430
2431 /*
2432 * Even if we did not try to evict anon pages at all, we want to
2433 * rebalance the anon lru active/inactive ratio.
2434 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002435 if (inactive_list_is_low(lruvec, false, sc))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002436 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2437 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002438}
2439
Mel Gorman23b9da52012-05-29 15:06:20 -07002440/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002441static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002442{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002443 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002444 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002445 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002446 return true;
2447
2448 return false;
2449}
2450
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002451/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002452 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2453 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2454 * true if more pages should be reclaimed such that when the page allocator
2455 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2456 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002457 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002458static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002459 unsigned long nr_reclaimed,
2460 unsigned long nr_scanned,
2461 struct scan_control *sc)
2462{
2463 unsigned long pages_for_compaction;
2464 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002465 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002466
2467 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002468 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002469 return false;
2470
Mel Gorman28765922011-02-25 14:44:20 -08002471 /* Consider stopping depending on scan and reclaim activity */
2472 if (sc->gfp_mask & __GFP_REPEAT) {
2473 /*
2474 * For __GFP_REPEAT allocations, stop reclaiming if the
2475 * full LRU list has been scanned and we are still failing
2476 * to reclaim pages. This full LRU scan is potentially
2477 * expensive but a __GFP_REPEAT caller really wants to succeed
2478 */
2479 if (!nr_reclaimed && !nr_scanned)
2480 return false;
2481 } else {
2482 /*
2483 * For non-__GFP_REPEAT allocations which can presumably
2484 * fail without consequence, stop if we failed to reclaim
2485 * any pages from the last SWAP_CLUSTER_MAX number of
2486 * pages that were scanned. This will return to the
2487 * caller faster at the risk reclaim/compaction and
2488 * the resulting allocation attempt fails
2489 */
2490 if (!nr_reclaimed)
2491 return false;
2492 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002493
2494 /*
2495 * If we have not reclaimed enough pages for compaction and the
2496 * inactive lists are large enough, continue reclaiming
2497 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002498 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002499 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002500 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002501 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002502 if (sc->nr_reclaimed < pages_for_compaction &&
2503 inactive_lru_pages > pages_for_compaction)
2504 return true;
2505
2506 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002507 for (z = 0; z <= sc->reclaim_idx; z++) {
2508 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002509 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002510 continue;
2511
2512 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002513 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002514 case COMPACT_CONTINUE:
2515 return false;
2516 default:
2517 /* check next zone */
2518 ;
2519 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002520 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002521 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002522}
2523
Mel Gorman970a39a2016-07-28 15:46:35 -07002524static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002525{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002526 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002527 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002528 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002529
Johannes Weiner56600482012-01-12 17:17:59 -08002530 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002531 struct mem_cgroup *root = sc->target_mem_cgroup;
2532 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002533 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002534 .priority = sc->priority,
2535 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002536 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002537 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002538
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002539 nr_reclaimed = sc->nr_reclaimed;
2540 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002541
Andrew Morton694fbc02013-09-24 15:27:37 -07002542 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2543 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002544 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002545 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002546 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002547
Johannes Weiner241994ed2015-02-11 15:26:06 -08002548 if (mem_cgroup_low(root, memcg)) {
2549 if (!sc->may_thrash)
2550 continue;
2551 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2552 }
2553
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002554 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002555 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002556
Mel Gormana9dd0a82016-07-28 15:46:02 -07002557 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2558 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002559
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002560 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002561 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002562 memcg, sc->nr_scanned - scanned,
2563 lru_pages);
2564
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002565 /* Record the group's reclaim efficiency */
2566 vmpressure(sc->gfp_mask, memcg, false,
2567 sc->nr_scanned - scanned,
2568 sc->nr_reclaimed - reclaimed);
2569
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002570 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002571 * Direct reclaim and kswapd have to scan all memory
2572 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002573 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002574 *
2575 * Limit reclaim, on the other hand, only cares about
2576 * nr_to_reclaim pages to be reclaimed and it will
2577 * retry with decreasing priority if one round over the
2578 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002579 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002580 if (!global_reclaim(sc) &&
2581 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002582 mem_cgroup_iter_break(root, memcg);
2583 break;
2584 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002585 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002586
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002587 /*
2588 * Shrink the slab caches in the same proportion that
2589 * the eligible LRU pages were scanned.
2590 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002591 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002592 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002593 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002594 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002595
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002596 if (reclaim_state) {
2597 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2598 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002599 }
2600
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002601 /* Record the subtree's reclaim efficiency */
2602 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002603 sc->nr_scanned - nr_scanned,
2604 sc->nr_reclaimed - nr_reclaimed);
2605
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002606 if (sc->nr_reclaimed - nr_reclaimed)
2607 reclaimable = true;
2608
Mel Gormana9dd0a82016-07-28 15:46:02 -07002609 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002610 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002611
2612 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002613}
2614
Vlastimil Babka53853e22014-10-09 15:27:02 -07002615/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002616 * Returns true if compaction should go ahead for a costly-order request, or
2617 * the allocation would already succeed without compaction. Return false if we
2618 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002619 */
Mel Gorman4f588332016-07-28 15:46:38 -07002620static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002621{
Mel Gorman31483b62016-07-28 15:45:46 -07002622 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002623 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002624
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002625 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2626 if (suitable == COMPACT_SUCCESS)
2627 /* Allocation should succeed already. Don't reclaim. */
2628 return true;
2629 if (suitable == COMPACT_SKIPPED)
2630 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002631 return false;
2632
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002633 /*
2634 * Compaction is already possible, but it takes time to run and there
2635 * are potentially other callers using the pages just freed. So proceed
2636 * with reclaim to make a buffer of free pages available to give
2637 * compaction a reasonable chance of completing and allocating the page.
2638 * Note that we won't actually reclaim the whole buffer in one attempt
2639 * as the target watermark in should_continue_reclaim() is lower. But if
2640 * we are already above the high+gap watermark, don't reclaim at all.
2641 */
2642 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2643
2644 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002645}
2646
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647/*
2648 * This is the direct reclaim path, for page-allocating processes. We only
2649 * try to reclaim pages from zones which will satisfy the caller's allocation
2650 * request.
2651 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 * If a zone is deemed to be full of pinned pages then just give it a light
2653 * scan then give up on it.
2654 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002655static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656{
Mel Gormandd1a2392008-04-28 02:12:17 -07002657 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002658 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002659 unsigned long nr_soft_reclaimed;
2660 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002661 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002662 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002663
Mel Gormancc715d92012-03-21 16:34:00 -07002664 /*
2665 * If the number of buffer_heads in the machine exceeds the maximum
2666 * allowed level, force direct reclaim to scan the highmem zone as
2667 * highmem pages could be pinning lowmem pages storing buffer_heads
2668 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002669 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002670 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002671 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002672 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002673 }
Mel Gormancc715d92012-03-21 16:34:00 -07002674
Mel Gormand4debc62010-08-09 17:19:29 -07002675 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002676 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002677 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002678 * Take care memory controller reclaiming has small influence
2679 * to global LRU.
2680 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002681 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002682 if (!cpuset_zone_allowed(zone,
2683 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002684 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002685
Lisa Du6e543d52013-09-11 14:22:36 -07002686 if (sc->priority != DEF_PRIORITY &&
Mel Gorman599d0c92016-07-28 15:45:31 -07002687 !pgdat_reclaimable(zone->zone_pgdat))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002688 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002689
2690 /*
2691 * If we already have plenty of memory free for
2692 * compaction in this zone, don't free any more.
2693 * Even though compaction is invoked for any
2694 * non-zero order, only frequent costly order
2695 * reclamation is disruptive enough to become a
2696 * noticeable problem, like transparent huge
2697 * page allocations.
2698 */
2699 if (IS_ENABLED(CONFIG_COMPACTION) &&
2700 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002701 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002702 sc->compaction_ready = true;
2703 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002704 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002705
Andrew Morton0608f432013-09-24 15:27:41 -07002706 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002707 * Shrink each node in the zonelist once. If the
2708 * zonelist is ordered by zone (not the default) then a
2709 * node may be shrunk multiple times but in that case
2710 * the user prefers lower zones being preserved.
2711 */
2712 if (zone->zone_pgdat == last_pgdat)
2713 continue;
2714
2715 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002716 * This steals pages from memory cgroups over softlimit
2717 * and returns the number of reclaimed pages and
2718 * scanned pages. This works for global memory pressure
2719 * and balancing, not for a memcg's limit.
2720 */
2721 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002722 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002723 sc->order, sc->gfp_mask,
2724 &nr_soft_scanned);
2725 sc->nr_reclaimed += nr_soft_reclaimed;
2726 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002727 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002728 }
Nick Piggin408d8542006-09-25 23:31:27 -07002729
Mel Gorman79dafcd2016-07-28 15:45:53 -07002730 /* See comment about same check for global reclaim above */
2731 if (zone->zone_pgdat == last_pgdat)
2732 continue;
2733 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002734 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 }
Mel Gormane0c23272011-10-31 17:09:33 -07002736
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002737 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002738 * Restore to original mask to avoid the impact on the caller if we
2739 * promoted it to __GFP_HIGHMEM.
2740 */
2741 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002743
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744/*
2745 * This is the main entry point to direct page reclaim.
2746 *
2747 * If a full scan of the inactive list fails to free enough memory then we
2748 * are "out of memory" and something needs to be killed.
2749 *
2750 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2751 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002752 * caller can't do much about. We kick the writeback threads and take explicit
2753 * naps in the hope that some of these pages can be written. But if the
2754 * allocating task holds filesystem locks which prevent writeout this might not
2755 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002756 *
2757 * returns: 0, if no pages reclaimed
2758 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 */
Mel Gormandac1d272008-04-28 02:12:12 -07002760static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002761 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002763 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002764 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002765 unsigned long writeback_threshold;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002766retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002767 delayacct_freepages_start();
2768
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002769 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002770 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002772 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002773 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2774 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002775 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002776 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002777
Balbir Singh66e17072008-02-07 00:13:56 -08002778 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002779 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002780 break;
2781
2782 if (sc->compaction_ready)
2783 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
2785 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002786 * If we're getting trouble reclaiming, start doing
2787 * writepage even in laptop mode.
2788 */
2789 if (sc->priority < DEF_PRIORITY - 2)
2790 sc->may_writepage = 1;
2791
2792 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 * Try to write back as many pages as we just scanned. This
2794 * tends to cause slow streaming writers to write data to the
2795 * disk smoothly, at the dirtying rate, which is nice. But
2796 * that's undesirable in laptop mode, where we *want* lumpy
2797 * writeout. So in laptop mode, write out the whole world.
2798 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002799 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2800 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002801 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2802 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002803 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002805 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002806
Keika Kobayashi873b4772008-07-25 01:48:52 -07002807 delayacct_freepages_end();
2808
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002809 if (sc->nr_reclaimed)
2810 return sc->nr_reclaimed;
2811
Mel Gorman0cee34f2012-01-12 17:19:49 -08002812 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002813 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002814 return 1;
2815
Johannes Weiner241994ed2015-02-11 15:26:06 -08002816 /* Untapped cgroup reserves? Don't OOM, retry. */
2817 if (!sc->may_thrash) {
2818 sc->priority = initial_priority;
2819 sc->may_thrash = 1;
2820 goto retry;
2821 }
2822
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002823 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824}
2825
Mel Gorman55150612012-07-31 16:44:35 -07002826static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2827{
2828 struct zone *zone;
2829 unsigned long pfmemalloc_reserve = 0;
2830 unsigned long free_pages = 0;
2831 int i;
2832 bool wmark_ok;
2833
2834 for (i = 0; i <= ZONE_NORMAL; i++) {
2835 zone = &pgdat->node_zones[i];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002836 if (!managed_zone(zone) ||
Mel Gorman599d0c92016-07-28 15:45:31 -07002837 pgdat_reclaimable_pages(pgdat) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002838 continue;
2839
Mel Gorman55150612012-07-31 16:44:35 -07002840 pfmemalloc_reserve += min_wmark_pages(zone);
2841 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2842 }
2843
Mel Gorman675becc2014-06-04 16:07:35 -07002844 /* If there are no reserves (unexpected config) then do not throttle */
2845 if (!pfmemalloc_reserve)
2846 return true;
2847
Mel Gorman55150612012-07-31 16:44:35 -07002848 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2849
2850 /* kswapd must be awake if processes are being throttled */
2851 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002852 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002853 (enum zone_type)ZONE_NORMAL);
2854 wake_up_interruptible(&pgdat->kswapd_wait);
2855 }
2856
2857 return wmark_ok;
2858}
2859
2860/*
2861 * Throttle direct reclaimers if backing storage is backed by the network
2862 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2863 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002864 * when the low watermark is reached.
2865 *
2866 * Returns true if a fatal signal was delivered during throttling. If this
2867 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002868 */
Mel Gorman50694c22012-11-26 16:29:48 -08002869static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002870 nodemask_t *nodemask)
2871{
Mel Gorman675becc2014-06-04 16:07:35 -07002872 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002873 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002874 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002875
2876 /*
2877 * Kernel threads should not be throttled as they may be indirectly
2878 * responsible for cleaning pages necessary for reclaim to make forward
2879 * progress. kjournald for example may enter direct reclaim while
2880 * committing a transaction where throttling it could forcing other
2881 * processes to block on log_wait_commit().
2882 */
2883 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002884 goto out;
2885
2886 /*
2887 * If a fatal signal is pending, this process should not throttle.
2888 * It should return quickly so it can exit and free its memory
2889 */
2890 if (fatal_signal_pending(current))
2891 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002892
Mel Gorman675becc2014-06-04 16:07:35 -07002893 /*
2894 * Check if the pfmemalloc reserves are ok by finding the first node
2895 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2896 * GFP_KERNEL will be required for allocating network buffers when
2897 * swapping over the network so ZONE_HIGHMEM is unusable.
2898 *
2899 * Throttling is based on the first usable node and throttled processes
2900 * wait on a queue until kswapd makes progress and wakes them. There
2901 * is an affinity then between processes waking up and where reclaim
2902 * progress has been made assuming the process wakes on the same node.
2903 * More importantly, processes running on remote nodes will not compete
2904 * for remote pfmemalloc reserves and processes on different nodes
2905 * should make reasonable progress.
2906 */
2907 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002908 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002909 if (zone_idx(zone) > ZONE_NORMAL)
2910 continue;
2911
2912 /* Throttle based on the first usable node */
2913 pgdat = zone->zone_pgdat;
2914 if (pfmemalloc_watermark_ok(pgdat))
2915 goto out;
2916 break;
2917 }
2918
2919 /* If no zone was usable by the allocation flags then do not throttle */
2920 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002921 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002922
Mel Gorman68243e72012-07-31 16:44:39 -07002923 /* Account for the throttling */
2924 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2925
Mel Gorman55150612012-07-31 16:44:35 -07002926 /*
2927 * If the caller cannot enter the filesystem, it's possible that it
2928 * is due to the caller holding an FS lock or performing a journal
2929 * transaction in the case of a filesystem like ext[3|4]. In this case,
2930 * it is not safe to block on pfmemalloc_wait as kswapd could be
2931 * blocked waiting on the same lock. Instead, throttle for up to a
2932 * second before continuing.
2933 */
2934 if (!(gfp_mask & __GFP_FS)) {
2935 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2936 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002937
2938 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002939 }
2940
2941 /* Throttle until kswapd wakes the process */
2942 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2943 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002944
2945check_pending:
2946 if (fatal_signal_pending(current))
2947 return true;
2948
2949out:
2950 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002951}
2952
Mel Gormandac1d272008-04-28 02:12:12 -07002953unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002954 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002955{
Mel Gorman33906bc2010-08-09 17:19:16 -07002956 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002957 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002958 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002959 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07002960 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07002961 .order = order,
2962 .nodemask = nodemask,
2963 .priority = DEF_PRIORITY,
2964 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002965 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002966 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002967 };
2968
Mel Gorman55150612012-07-31 16:44:35 -07002969 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002970 * Do not enter reclaim if fatal signal was delivered while throttled.
2971 * 1 is returned so that the page allocator does not OOM kill at this
2972 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002973 */
Mel Gorman50694c22012-11-26 16:29:48 -08002974 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002975 return 1;
2976
Mel Gorman33906bc2010-08-09 17:19:16 -07002977 trace_mm_vmscan_direct_reclaim_begin(order,
2978 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07002979 gfp_mask,
2980 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002981
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002982 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002983
2984 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2985
2986 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002987}
2988
Andrew Mortonc255a452012-07-31 16:43:02 -07002989#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002990
Mel Gormana9dd0a82016-07-28 15:46:02 -07002991unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002992 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07002993 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07002994 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002995{
2996 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002997 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002998 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002999 .may_writepage = !laptop_mode,
3000 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003001 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003002 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003003 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003004 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003005
Balbir Singh4e416952009-09-23 15:56:39 -07003006 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3007 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003008
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003009 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003010 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003011 sc.gfp_mask,
3012 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003013
Balbir Singh4e416952009-09-23 15:56:39 -07003014 /*
3015 * NOTE: Although we can get the priority field, using it
3016 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003017 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003018 * will pick up pages from other mem cgroup's as well. We hack
3019 * the priority and make it zero.
3020 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003021 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003022
3023 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3024
Ying Han0ae5e892011-05-26 16:25:25 -07003025 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003026 return sc.nr_reclaimed;
3027}
3028
Johannes Weiner72835c82012-01-12 17:18:32 -08003029unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003030 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003031 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003032 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003033{
Balbir Singh4e416952009-09-23 15:56:39 -07003034 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003035 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003036 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003037 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003038 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003039 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3040 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003041 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003042 .target_mem_cgroup = memcg,
3043 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003044 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003045 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003046 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003047 };
Balbir Singh66e17072008-02-07 00:13:56 -08003048
Ying Han889976d2011-05-26 16:25:33 -07003049 /*
3050 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3051 * take care of from where we get pages. So the node where we start the
3052 * scan does not need to be the current node.
3053 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003054 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003055
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003056 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003057
3058 trace_mm_vmscan_memcg_reclaim_begin(0,
3059 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003060 sc.gfp_mask,
3061 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003062
Johannes Weiner89a28482016-10-27 17:46:56 -07003063 current->flags |= PF_MEMALLOC;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003064 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weiner89a28482016-10-27 17:46:56 -07003065 current->flags &= ~PF_MEMALLOC;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003066
3067 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3068
3069 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003070}
3071#endif
3072
Mel Gorman1d82de62016-07-28 15:45:43 -07003073static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003074 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003075{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003076 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003077
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003078 if (!total_swap_pages)
3079 return;
3080
3081 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3082 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003083 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003084
Mel Gormanf8d1a312016-07-28 15:47:34 -07003085 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003086 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003087 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003088
3089 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3090 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003091}
3092
Mel Gorman31483b62016-07-28 15:45:46 -07003093static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003094{
Mel Gorman31483b62016-07-28 15:45:46 -07003095 unsigned long mark = high_wmark_pages(zone);
Johannes Weiner60cefed2012-11-29 13:54:23 -08003096
Mel Gorman6256c6b2016-07-28 15:45:56 -07003097 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3098 return false;
3099
3100 /*
3101 * If any eligible zone is balanced then the node is not considered
3102 * to be congested or dirty
3103 */
3104 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3105 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3106
3107 return true;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003108}
3109
Mel Gorman1741c872011-01-13 15:46:21 -08003110/*
Mel Gorman55150612012-07-31 16:44:35 -07003111 * Prepare kswapd for sleeping. This verifies that there are no processes
3112 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3113 *
3114 * Returns true if kswapd is ready to sleep
3115 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003116static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003117{
Mel Gorman1d82de62016-07-28 15:45:43 -07003118 int i;
3119
Mel Gorman55150612012-07-31 16:44:35 -07003120 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003121 * The throttled processes are normally woken up in balance_pgdat() as
3122 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3123 * race between when kswapd checks the watermarks and a process gets
3124 * throttled. There is also a potential race if processes get
3125 * throttled, kswapd wakes, a large process exits thereby balancing the
3126 * zones, which causes kswapd to exit balance_pgdat() before reaching
3127 * the wake up checks. If kswapd is going to sleep, no process should
3128 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3129 * the wake up is premature, processes will wake kswapd and get
3130 * throttled again. The difference from wake ups in balance_pgdat() is
3131 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003132 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003133 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3134 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003135
Mel Gorman1d82de62016-07-28 15:45:43 -07003136 for (i = 0; i <= classzone_idx; i++) {
3137 struct zone *zone = pgdat->node_zones + i;
3138
Mel Gorman6aa303d2016-09-01 16:14:55 -07003139 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003140 continue;
3141
Mel Gorman38087d92016-07-28 15:45:49 -07003142 if (!zone_balanced(zone, order, classzone_idx))
3143 return false;
Mel Gorman1d82de62016-07-28 15:45:43 -07003144 }
3145
Mel Gorman38087d92016-07-28 15:45:49 -07003146 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003147}
3148
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003150 * kswapd shrinks a node of pages that are at or below the highest usable
3151 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003152 *
3153 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003154 * reclaim or if the lack of progress was due to pages under writeback.
3155 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003156 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003157static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003158 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003159{
Mel Gorman1d82de62016-07-28 15:45:43 -07003160 struct zone *zone;
3161 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003162
Mel Gorman1d82de62016-07-28 15:45:43 -07003163 /* Reclaim a number of pages proportional to the number of zones */
3164 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003165 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003166 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003167 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003168 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003169
Mel Gorman1d82de62016-07-28 15:45:43 -07003170 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003171 }
3172
Mel Gorman1d82de62016-07-28 15:45:43 -07003173 /*
3174 * Historically care was taken to put equal pressure on all zones but
3175 * now pressure is applied based on node LRU order.
3176 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003177 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003178
3179 /*
3180 * Fragmentation may mean that the system cannot be rebalanced for
3181 * high-order allocations. If twice the allocation size has been
3182 * reclaimed then recheck watermarks only at order-0 to prevent
3183 * excessive reclaim. Assume that a process requested a high-order
3184 * can direct reclaim/compact.
3185 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003186 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003187 sc->order = 0;
3188
Mel Gormanb8e83b92013-07-03 15:01:45 -07003189 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003190}
3191
3192/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003193 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3194 * that are eligible for use by the caller until at least one zone is
3195 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003197 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 *
3199 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003200 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003201 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3202 * or lower is eligible for reclaim until at least one usable zone is
3203 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003205static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003208 unsigned long nr_soft_reclaimed;
3209 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003210 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003211 struct scan_control sc = {
3212 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003213 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003214 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003215 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003216 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003217 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003218 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003219 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003221 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003222 bool raise_priority = true;
3223
3224 sc.nr_reclaimed = 0;
Mel Gorman84c7a772016-07-28 15:46:44 -07003225 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226
Mel Gorman86c79f62016-07-28 15:45:59 -07003227 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003228 * If the number of buffer_heads exceeds the maximum allowed
3229 * then consider reclaiming from all zones. This has a dual
3230 * purpose -- on 64-bit systems it is expected that
3231 * buffer_heads are stripped during active rotation. On 32-bit
3232 * systems, highmem pages can pin lowmem memory and shrinking
3233 * buffers can relieve lowmem pressure. Reclaim may still not
3234 * go ahead if all eligible zones for the original allocation
3235 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003236 */
3237 if (buffer_heads_over_limit) {
3238 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3239 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003240 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003241 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242
Mel Gorman970a39a2016-07-28 15:46:35 -07003243 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003244 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003247
Mel Gorman86c79f62016-07-28 15:45:59 -07003248 /*
3249 * Only reclaim if there are no eligible zones. Check from
3250 * high to low zone as allocations prefer higher zones.
3251 * Scanning from low to high zone would allow congestion to be
3252 * cleared during a very small window when a small low
3253 * zone was balanced even under extreme pressure when the
Mel Gorman84c7a772016-07-28 15:46:44 -07003254 * overall node may be congested. Note that sc.reclaim_idx
3255 * is not used as buffer_heads_over_limit may have adjusted
3256 * it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003257 */
Mel Gorman84c7a772016-07-28 15:46:44 -07003258 for (i = classzone_idx; i >= 0; i--) {
Mel Gorman86c79f62016-07-28 15:45:59 -07003259 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003260 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003261 continue;
3262
Mel Gorman84c7a772016-07-28 15:46:44 -07003263 if (zone_balanced(zone, sc.order, classzone_idx))
Mel Gorman86c79f62016-07-28 15:45:59 -07003264 goto out;
3265 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003266
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003268 * Do some background aging of the anon list, to give
3269 * pages a chance to be referenced before reclaiming. All
3270 * pages are rotated regardless of classzone as this is
3271 * about consistent aging.
3272 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003273 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003274
3275 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003276 * If we're getting trouble reclaiming, start doing writepage
3277 * even in laptop mode.
3278 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003279 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003280 sc.may_writepage = 1;
3281
Mel Gorman1d82de62016-07-28 15:45:43 -07003282 /* Call soft limit reclaim before calling shrink_node. */
3283 sc.nr_scanned = 0;
3284 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003285 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003286 sc.gfp_mask, &nr_soft_scanned);
3287 sc.nr_reclaimed += nr_soft_reclaimed;
3288
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003289 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003290 * There should be no need to raise the scanning priority if
3291 * enough pages are already being scanned that that high
3292 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003294 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003295 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003296
3297 /*
3298 * If the low watermark is met there is no need for processes
3299 * to be throttled on pfmemalloc_wait as they should not be
3300 * able to safely make forward progress. Wake them
3301 */
3302 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3303 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003304 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003305
Mel Gormanb8e83b92013-07-03 15:01:45 -07003306 /* Check if kswapd should be suspending */
3307 if (try_to_freeze() || kthread_should_stop())
3308 break;
3309
3310 /*
3311 * Raise priority if scanning rate is too low or there was no
3312 * progress in reclaiming pages
3313 */
3314 if (raise_priority || !sc.nr_reclaimed)
3315 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003316 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317
Mel Gormanb8e83b92013-07-03 15:01:45 -07003318out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003319 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003320 * Return the order kswapd stopped reclaiming at as
3321 * prepare_kswapd_sleep() takes it into account. If another caller
3322 * entered the allocator slow path while kswapd was awake, order will
3323 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003324 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003325 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326}
3327
Mel Gorman38087d92016-07-28 15:45:49 -07003328static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3329 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003330{
3331 long remaining = 0;
3332 DEFINE_WAIT(wait);
3333
3334 if (freezing(current) || kthread_should_stop())
3335 return;
3336
3337 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3338
3339 /* Try to sleep for a short interval */
Mel Gormand9f21d42016-07-28 15:46:41 -07003340 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003341 /*
3342 * Compaction records what page blocks it recently failed to
3343 * isolate pages from and skips them in the future scanning.
3344 * When kswapd is going to sleep, it is reasonable to assume
3345 * that pages and compaction may succeed so reset the cache.
3346 */
3347 reset_isolation_suitable(pgdat);
3348
3349 /*
3350 * We have freed the memory, now we should compact it to make
3351 * allocation of the requested order possible.
3352 */
Mel Gorman38087d92016-07-28 15:45:49 -07003353 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003354
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003355 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003356
3357 /*
3358 * If woken prematurely then reset kswapd_classzone_idx and
3359 * order. The values will either be from a wakeup request or
3360 * the previous request that slept prematurely.
3361 */
3362 if (remaining) {
3363 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3364 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3365 }
3366
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003367 finish_wait(&pgdat->kswapd_wait, &wait);
3368 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3369 }
3370
3371 /*
3372 * After a short sleep, check if it was a premature sleep. If not, then
3373 * go fully to sleep until explicitly woken up.
3374 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003375 if (!remaining &&
3376 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003377 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3378
3379 /*
3380 * vmstat counters are not perfectly accurate and the estimated
3381 * value for counters such as NR_FREE_PAGES can deviate from the
3382 * true value by nr_online_cpus * threshold. To avoid the zone
3383 * watermarks being breached while under pressure, we reduce the
3384 * per-cpu vmstat threshold while kswapd is awake and restore
3385 * them before going back to sleep.
3386 */
3387 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003388
3389 if (!kthread_should_stop())
3390 schedule();
3391
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003392 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3393 } else {
3394 if (remaining)
3395 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3396 else
3397 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3398 }
3399 finish_wait(&pgdat->kswapd_wait, &wait);
3400}
3401
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402/*
3403 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003404 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 *
3406 * This basically trickles out pages so that we have _some_
3407 * free memory available even if there is no other activity
3408 * that frees anything up. This is needed for things like routing
3409 * etc, where we otherwise might have all activity going on in
3410 * asynchronous contexts that cannot page things out.
3411 *
3412 * If there are applications that are active memory-allocators
3413 * (most normal use), this basically shouldn't matter.
3414 */
3415static int kswapd(void *p)
3416{
Mel Gorman38087d92016-07-28 15:45:49 -07003417 unsigned int alloc_order, reclaim_order, classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 pg_data_t *pgdat = (pg_data_t*)p;
3419 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003420
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 struct reclaim_state reclaim_state = {
3422 .reclaimed_slab = 0,
3423 };
Rusty Russella70f7302009-03-13 14:49:46 +10303424 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
Nick Piggincf40bd12009-01-21 08:12:39 +01003426 lockdep_set_current_reclaim_state(GFP_KERNEL);
3427
Rusty Russell174596a2009-01-01 10:12:29 +10303428 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003429 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 current->reclaim_state = &reclaim_state;
3431
3432 /*
3433 * Tell the memory management that we're a "memory allocator",
3434 * and that if we need more memory we should get access to it
3435 * regardless (see "__alloc_pages()"). "kswapd" should
3436 * never get caught in the normal page freeing logic.
3437 *
3438 * (Kswapd normally doesn't need memory anyway, but sometimes
3439 * you need a small amount of memory in order to be able to
3440 * page out something else, and this flag essentially protects
3441 * us from recursively trying to free more memory as we're
3442 * trying to free the first piece of memory in the first place).
3443 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003444 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003445 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446
Mel Gorman38087d92016-07-28 15:45:49 -07003447 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3448 pgdat->kswapd_classzone_idx = classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003450 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003451
Mel Gorman38087d92016-07-28 15:45:49 -07003452kswapd_try_sleep:
3453 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3454 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003455
Mel Gorman38087d92016-07-28 15:45:49 -07003456 /* Read the new order and classzone_idx */
3457 alloc_order = reclaim_order = pgdat->kswapd_order;
3458 classzone_idx = pgdat->kswapd_classzone_idx;
3459 pgdat->kswapd_order = 0;
3460 pgdat->kswapd_classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461
David Rientjes8fe23e02009-12-14 17:58:33 -08003462 ret = try_to_freeze();
3463 if (kthread_should_stop())
3464 break;
3465
3466 /*
3467 * We can speed up thawing tasks if we don't call balance_pgdat
3468 * after returning from the refrigerator
3469 */
Mel Gorman38087d92016-07-28 15:45:49 -07003470 if (ret)
3471 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003472
Mel Gorman38087d92016-07-28 15:45:49 -07003473 /*
3474 * Reclaim begins at the requested order but if a high-order
3475 * reclaim fails then kswapd falls back to reclaiming for
3476 * order-0. If that happens, kswapd will consider sleeping
3477 * for the order it finished reclaiming at (reclaim_order)
3478 * but kcompactd is woken to compact for the original
3479 * request (alloc_order).
3480 */
Mel Gormane5146b12016-07-28 15:46:47 -07003481 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3482 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003483 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3484 if (reclaim_order < alloc_order)
3485 goto kswapd_try_sleep;
3486
3487 alloc_order = reclaim_order = pgdat->kswapd_order;
3488 classzone_idx = pgdat->kswapd_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003490
Johannes Weiner71abdc12014-06-06 14:35:35 -07003491 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003492 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003493 lockdep_clear_current_reclaim_state();
3494
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 return 0;
3496}
3497
3498/*
3499 * A zone is low on free memory, so wake its kswapd task to service it.
3500 */
Mel Gorman99504742011-01-13 15:46:20 -08003501void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502{
3503 pg_data_t *pgdat;
Mel Gormane1a55632016-07-28 15:46:26 -07003504 int z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505
Mel Gorman6aa303d2016-09-01 16:14:55 -07003506 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 return;
3508
Vladimir Davydov344736f2014-10-20 15:50:30 +04003509 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003511 pgdat = zone->zone_pgdat;
Mel Gorman38087d92016-07-28 15:45:49 -07003512 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3513 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003514 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003516
3517 /* Only wake kswapd if all zones are unbalanced */
3518 for (z = 0; z <= classzone_idx; z++) {
3519 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003520 if (!managed_zone(zone))
Mel Gormane1a55632016-07-28 15:46:26 -07003521 continue;
3522
3523 if (zone_balanced(zone, order, classzone_idx))
3524 return;
3525 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08003526
3527 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003528 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529}
3530
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003531#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003533 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003534 * freed pages.
3535 *
3536 * Rather than trying to age LRUs the aim is to preserve the overall
3537 * LRU order by reclaiming preferentially
3538 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003540unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003542 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003543 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003544 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003545 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003546 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003547 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003548 .may_writepage = 1,
3549 .may_unmap = 1,
3550 .may_swap = 1,
3551 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003553 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003554 struct task_struct *p = current;
3555 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003557 p->flags |= PF_MEMALLOC;
3558 lockdep_set_current_reclaim_state(sc.gfp_mask);
3559 reclaim_state.reclaimed_slab = 0;
3560 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003561
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003562 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003563
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003564 p->reclaim_state = NULL;
3565 lockdep_clear_current_reclaim_state();
3566 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003567
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003568 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003570#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572/* It's optimal to keep kswapds on the same CPUs as their memory, but
3573 not required for correctness. So if the last cpu in a node goes
3574 away, we get changed to run anywhere: as the first one comes back,
3575 restore their cpu bindings. */
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003576static int kswapd_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003578 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003580 for_each_node_state(nid, N_MEMORY) {
3581 pg_data_t *pgdat = NODE_DATA(nid);
3582 const struct cpumask *mask;
Rusty Russella70f7302009-03-13 14:49:46 +10303583
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003584 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003585
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003586 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3587 /* One of our CPUs online: restore mask */
3588 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 }
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003590 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592
Yasunori Goto3218ae12006-06-27 02:53:33 -07003593/*
3594 * This kswapd start function will be called by init and node-hot-add.
3595 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3596 */
3597int kswapd_run(int nid)
3598{
3599 pg_data_t *pgdat = NODE_DATA(nid);
3600 int ret = 0;
3601
3602 if (pgdat->kswapd)
3603 return 0;
3604
3605 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3606 if (IS_ERR(pgdat->kswapd)) {
3607 /* failure at boot is fatal */
3608 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003609 pr_err("Failed to start kswapd on node %d\n", nid);
3610 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003611 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003612 }
3613 return ret;
3614}
3615
David Rientjes8fe23e02009-12-14 17:58:33 -08003616/*
Jiang Liud8adde12012-07-11 14:01:52 -07003617 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003618 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003619 */
3620void kswapd_stop(int nid)
3621{
3622 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3623
Jiang Liud8adde12012-07-11 14:01:52 -07003624 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003625 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003626 NODE_DATA(nid)->kswapd = NULL;
3627 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003628}
3629
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630static int __init kswapd_init(void)
3631{
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003632 int nid, ret;
Andrew Morton69e05942006-03-22 00:08:19 -08003633
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003635 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003636 kswapd_run(nid);
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003637 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3638 "mm/vmscan:online", kswapd_cpu_online,
3639 NULL);
3640 WARN_ON(ret < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641 return 0;
3642}
3643
3644module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003645
3646#ifdef CONFIG_NUMA
3647/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003648 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003649 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003650 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003651 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003652 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003653int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003654
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003655#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003656#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003657#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003658#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003659
Christoph Lameter9eeff232006-01-18 17:42:31 -08003660/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003661 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003662 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3663 * a zone.
3664 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003665#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003666
Christoph Lameter9eeff232006-01-18 17:42:31 -08003667/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003668 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003669 * occur.
3670 */
3671int sysctl_min_unmapped_ratio = 1;
3672
3673/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003674 * If the number of slab pages in a zone grows beyond this percentage then
3675 * slab reclaim needs to occur.
3676 */
3677int sysctl_min_slab_ratio = 5;
3678
Mel Gorman11fb9982016-07-28 15:46:20 -07003679static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003680{
Mel Gorman11fb9982016-07-28 15:46:20 -07003681 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3682 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3683 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003684
3685 /*
3686 * It's possible for there to be more file mapped pages than
3687 * accounted for by the pages on the file LRU lists because
3688 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3689 */
3690 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3691}
3692
3693/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003694static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003695{
Alexandru Moised031a152015-11-05 18:48:08 -08003696 unsigned long nr_pagecache_reclaimable;
3697 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003698
3699 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003700 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003701 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003702 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003703 * a better estimate
3704 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003705 if (node_reclaim_mode & RECLAIM_UNMAP)
3706 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003707 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003708 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003709
3710 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003711 if (!(node_reclaim_mode & RECLAIM_WRITE))
3712 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003713
3714 /* Watch for any possible underflows due to delta */
3715 if (unlikely(delta > nr_pagecache_reclaimable))
3716 delta = nr_pagecache_reclaimable;
3717
3718 return nr_pagecache_reclaimable - delta;
3719}
3720
Christoph Lameter0ff38492006-09-25 23:31:52 -07003721/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003722 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003723 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003724static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003725{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003726 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003727 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003728 struct task_struct *p = current;
3729 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003730 int classzone_idx = gfp_zone(gfp_mask);
Andrew Morton179e9632006-03-22 00:08:18 -08003731 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003732 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003733 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003734 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003735 .priority = NODE_RECLAIM_PRIORITY,
3736 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3737 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003738 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003739 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003740 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003741
Christoph Lameter9eeff232006-01-18 17:42:31 -08003742 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003743 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003744 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003745 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003746 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003747 */
3748 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003749 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003750 reclaim_state.reclaimed_slab = 0;
3751 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003752
Mel Gormana5f5f912016-07-28 15:46:32 -07003753 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003754 /*
3755 * Free memory by calling shrink zone with increasing
3756 * priorities until we have enough memory freed.
3757 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003758 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003759 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003760 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003761 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003762
Christoph Lameter9eeff232006-01-18 17:42:31 -08003763 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003764 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003765 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003766 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003767}
Andrew Morton179e9632006-03-22 00:08:18 -08003768
Mel Gormana5f5f912016-07-28 15:46:32 -07003769int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003770{
David Rientjesd773ed62007-10-16 23:26:01 -07003771 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003772
3773 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003774 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003775 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003776 *
Christoph Lameter96146342006-07-03 00:24:13 -07003777 * A small portion of unmapped file backed pages is needed for
3778 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003779 * thrown out if the node is overallocated. So we do not reclaim
3780 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003781 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003782 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003783 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3784 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3785 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003786
Mel Gormana5f5f912016-07-28 15:46:32 -07003787 if (!pgdat_reclaimable(pgdat))
3788 return NODE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003789
Andrew Morton179e9632006-03-22 00:08:18 -08003790 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003791 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003792 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003793 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003794 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003795
3796 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003797 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003798 * have associated processors. This will favor the local processor
3799 * over remote processors and spread off node memory allocations
3800 * as wide as possible.
3801 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003802 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3803 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003804
Mel Gormana5f5f912016-07-28 15:46:32 -07003805 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3806 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003807
Mel Gormana5f5f912016-07-28 15:46:32 -07003808 ret = __node_reclaim(pgdat, gfp_mask, order);
3809 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003810
Mel Gorman24cf725182009-06-16 15:33:23 -07003811 if (!ret)
3812 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3813
David Rientjesd773ed62007-10-16 23:26:01 -07003814 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003815}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003816#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003817
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003818/*
3819 * page_evictable - test whether a page is evictable
3820 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003821 *
3822 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003823 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003824 *
3825 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003826 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003827 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003828 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003829 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003830int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003831{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003832 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003833}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003834
Hugh Dickins850465792012-01-20 14:34:19 -08003835#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003836/**
Hugh Dickins24513262012-01-20 14:34:21 -08003837 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3838 * @pages: array of pages to check
3839 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003840 *
Hugh Dickins24513262012-01-20 14:34:21 -08003841 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003842 *
3843 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003844 */
Hugh Dickins24513262012-01-20 14:34:21 -08003845void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003846{
Johannes Weiner925b7672012-01-12 17:18:15 -08003847 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003848 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003849 int pgscanned = 0;
3850 int pgrescued = 0;
3851 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003852
Hugh Dickins24513262012-01-20 14:34:21 -08003853 for (i = 0; i < nr_pages; i++) {
3854 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003855 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003856
Hugh Dickins24513262012-01-20 14:34:21 -08003857 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003858 if (pagepgdat != pgdat) {
3859 if (pgdat)
3860 spin_unlock_irq(&pgdat->lru_lock);
3861 pgdat = pagepgdat;
3862 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003863 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003864 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003865
Hugh Dickins24513262012-01-20 14:34:21 -08003866 if (!PageLRU(page) || !PageUnevictable(page))
3867 continue;
3868
Hugh Dickins39b5f292012-10-08 16:33:18 -07003869 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003870 enum lru_list lru = page_lru_base_type(page);
3871
Sasha Levin309381fea2014-01-23 15:52:54 -08003872 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003873 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003874 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3875 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003876 pgrescued++;
3877 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003878 }
Hugh Dickins24513262012-01-20 14:34:21 -08003879
Mel Gorman785b99f2016-07-28 15:47:23 -07003880 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003881 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3882 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003883 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003884 }
Hugh Dickins850465792012-01-20 14:34:19 -08003885}
3886#endif /* CONFIG_SHMEM */