blob: b1e12a1ea9cfcc6ece1690305fc2d44bfdcf5b37 [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 Weiner241994e2015-02-11 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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000246 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000248int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000250 size_t size = sizeof(*shrinker->nr_deferred);
251
Glauber Costa1d3d4432013-08-28 10:18:04 +1000252 if (shrinker->flags & SHRINKER_NUMA_AWARE)
253 size *= nr_node_ids;
254
255 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
256 if (!shrinker->nr_deferred)
257 return -ENOMEM;
258
Rusty Russell8e1f9362007-07-17 04:03:17 -0700259 down_write(&shrinker_rwsem);
260 list_add_tail(&shrinker->list, &shrinker_list);
261 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000262 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700264EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
266/*
267 * Remove one
268 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700269void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270{
271 down_write(&shrinker_rwsem);
272 list_del(&shrinker->list);
273 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700274 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700276EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
278#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800280static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
281 struct shrinker *shrinker,
282 unsigned long nr_scanned,
283 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000285 unsigned long freed = 0;
286 unsigned long long delta;
287 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700288 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000289 long nr;
290 long new_nr;
291 int nid = shrinkctl->nid;
292 long batch_size = shrinker->batch ? shrinker->batch
293 : SHRINK_BATCH;
294
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700295 freeable = shrinker->count_objects(shrinker, shrinkctl);
296 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000297 return 0;
298
299 /*
300 * copy the current shrinker scan count into a local variable
301 * and zero it so that other concurrent shrinker invocations
302 * don't also do this scanning work.
303 */
304 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
305
306 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800307 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700308 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800309 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000310 total_scan += delta;
311 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800312 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000313 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700314 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000315 }
316
317 /*
318 * We need to avoid excessive windup on filesystem shrinkers
319 * due to large numbers of GFP_NOFS allocations causing the
320 * shrinkers to return -1 all the time. This results in a large
321 * nr being built up so when a shrink that can do some work
322 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700323 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000324 * memory.
325 *
326 * Hence only allow the shrinker to scan the entire cache when
327 * a large delta change is calculated directly.
328 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700329 if (delta < freeable / 4)
330 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000331
332 /*
333 * Avoid risking looping forever due to too large nr value:
334 * never try to free more than twice the estimate number of
335 * freeable entries.
336 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700337 if (total_scan > freeable * 2)
338 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000339
340 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800341 nr_scanned, nr_eligible,
342 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000343
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800344 /*
345 * Normally, we should not scan less than batch_size objects in one
346 * pass to avoid too frequent shrinker calls, but if the slab has less
347 * than batch_size objects in total and we are really tight on memory,
348 * we will try to reclaim all available objects, otherwise we can end
349 * up failing allocations although there are plenty of reclaimable
350 * objects spread over several slabs with usage less than the
351 * batch_size.
352 *
353 * We detect the "tight on memory" situations by looking at the total
354 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700355 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800356 * scanning at high prio and therefore should try to reclaim as much as
357 * possible.
358 */
359 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700360 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000361 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800362 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000363
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800364 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000365 ret = shrinker->scan_objects(shrinker, shrinkctl);
366 if (ret == SHRINK_STOP)
367 break;
368 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000369
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800370 count_vm_events(SLABS_SCANNED, nr_to_scan);
371 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000372
373 cond_resched();
374 }
375
376 /*
377 * move the unused scan count back into the shrinker in a
378 * manner that handles concurrent updates. If we exhausted the
379 * scan, there is no need to do an update.
380 */
381 if (total_scan > 0)
382 new_nr = atomic_long_add_return(total_scan,
383 &shrinker->nr_deferred[nid]);
384 else
385 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
386
Dave Hansendf9024a2014-06-04 16:08:07 -0700387 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000388 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389}
390
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800391/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800392 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800393 * @gfp_mask: allocation context
394 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800395 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800396 * @nr_scanned: pressure numerator
397 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800399 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800401 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
402 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800404 * @memcg specifies the memory cgroup to target. If it is not NULL,
405 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700406 * objects from the memory cgroup specified. Otherwise, only unaware
407 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800408 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800409 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
410 * the available objects should be scanned. Page reclaim for example
411 * passes the number of pages scanned and the number of pages on the
412 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
413 * when it encountered mapped pages. The ratio is further biased by
414 * the ->seeks setting of the shrink function, which indicates the
415 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800417 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800419static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
420 struct mem_cgroup *memcg,
421 unsigned long nr_scanned,
422 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423{
424 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000425 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700427 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800428 return 0;
429
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800430 if (nr_scanned == 0)
431 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
Minchan Kimf06590b2011-05-24 17:11:11 -0700433 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000434 /*
435 * If we would return 0, our callers would understand that we
436 * have nothing else to shrink and give up trying. By returning
437 * 1 we keep it going and assume we'll be able to shrink next
438 * time.
439 */
440 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700441 goto out;
442 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800445 struct shrink_control sc = {
446 .gfp_mask = gfp_mask,
447 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800448 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800449 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800450
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700451 /*
452 * If kernel memory accounting is disabled, we ignore
453 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
454 * passing NULL for memcg.
455 */
456 if (memcg_kmem_enabled() &&
457 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800458 continue;
459
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800460 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
461 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000462
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800463 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800465
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700467out:
468 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000469 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470}
471
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800472void drop_slab_node(int nid)
473{
474 unsigned long freed;
475
476 do {
477 struct mem_cgroup *memcg = NULL;
478
479 freed = 0;
480 do {
481 freed += shrink_slab(GFP_KERNEL, nid, memcg,
482 1000, 1000);
483 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
484 } while (freed > 10);
485}
486
487void drop_slab(void)
488{
489 int nid;
490
491 for_each_online_node(nid)
492 drop_slab_node(nid);
493}
494
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495static inline int is_page_cache_freeable(struct page *page)
496{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700497 /*
498 * A freeable page cache page is referenced only by the caller
499 * that isolated the page, the page cache radix tree and
500 * optional buffer heads at page->private.
501 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700502 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503}
504
Tejun Heo703c2702015-05-22 17:13:44 -0400505static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506{
Christoph Lameter930d9152006-01-08 01:00:47 -0800507 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400509 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400511 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 return 1;
513 return 0;
514}
515
516/*
517 * We detected a synchronous write error writing a page out. Probably
518 * -ENOSPC. We need to propagate that into the address_space for a subsequent
519 * fsync(), msync() or close().
520 *
521 * The tricky part is that after writepage we cannot touch the mapping: nothing
522 * prevents it from being freed up. But we have a ref on the page and once
523 * that page is locked, the mapping is pinned.
524 *
525 * We're allowed to run sleeping lock_page() here because we know the caller has
526 * __GFP_FS.
527 */
528static void handle_write_error(struct address_space *mapping,
529 struct page *page, int error)
530{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100531 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700532 if (page_mapping(page) == mapping)
533 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 unlock_page(page);
535}
536
Christoph Lameter04e62a22006-06-23 02:03:38 -0700537/* possible outcome of pageout() */
538typedef enum {
539 /* failed to write page out, page is locked */
540 PAGE_KEEP,
541 /* move page to the active list, page is locked */
542 PAGE_ACTIVATE,
543 /* page has been sent to the disk successfully, page is unlocked */
544 PAGE_SUCCESS,
545 /* page is clean and locked */
546 PAGE_CLEAN,
547} pageout_t;
548
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549/*
Andrew Morton1742f192006-03-22 00:08:21 -0800550 * pageout is called by shrink_page_list() for each dirty page.
551 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700553static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700554 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555{
556 /*
557 * If the page is dirty, only perform writeback if that write
558 * will be non-blocking. To prevent this allocation from being
559 * stalled by pagecache activity. But note that there may be
560 * stalls if we need to run get_block(). We could test
561 * PagePrivate for that.
562 *
Al Viro81742022014-04-03 03:17:43 -0400563 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 * this page's queue, we can perform writeback even if that
565 * will block.
566 *
567 * If the page is swapcache, write it back even if that would
568 * block, for some throttling. This happens by accident, because
569 * swap_backing_dev_info is bust: it doesn't reflect the
570 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 */
572 if (!is_page_cache_freeable(page))
573 return PAGE_KEEP;
574 if (!mapping) {
575 /*
576 * Some data journaling orphaned pages can have
577 * page->mapping == NULL while being dirty with clean buffers.
578 */
David Howells266cf652009-04-03 16:42:36 +0100579 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 if (try_to_free_buffers(page)) {
581 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700582 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 return PAGE_CLEAN;
584 }
585 }
586 return PAGE_KEEP;
587 }
588 if (mapping->a_ops->writepage == NULL)
589 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400590 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 return PAGE_KEEP;
592
593 if (clear_page_dirty_for_io(page)) {
594 int res;
595 struct writeback_control wbc = {
596 .sync_mode = WB_SYNC_NONE,
597 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700598 .range_start = 0,
599 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 .for_reclaim = 1,
601 };
602
603 SetPageReclaim(page);
604 res = mapping->a_ops->writepage(page, &wbc);
605 if (res < 0)
606 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800607 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 ClearPageReclaim(page);
609 return PAGE_ACTIVATE;
610 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700611
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 if (!PageWriteback(page)) {
613 /* synchronous write or broken a_ops? */
614 ClearPageReclaim(page);
615 }
yalin wang3aa23852016-01-14 15:18:30 -0800616 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700617 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 return PAGE_SUCCESS;
619 }
620
621 return PAGE_CLEAN;
622}
623
Andrew Mortona649fd92006-10-17 00:09:36 -0700624/*
Nick Piggine2867812008-07-25 19:45:30 -0700625 * Same as remove_mapping, but if the page is removed from the mapping, it
626 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700627 */
Johannes Weinera5289102014-04-03 14:47:51 -0700628static int __remove_mapping(struct address_space *mapping, struct page *page,
629 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800630{
Greg Thelenc4843a72015-05-22 17:13:16 -0400631 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400632
Nick Piggin28e4d962006-09-25 23:31:23 -0700633 BUG_ON(!PageLocked(page));
634 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800635
Greg Thelenc4843a72015-05-22 17:13:16 -0400636 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800637 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700638 * The non racy check for a busy page.
639 *
640 * Must be careful with the order of the tests. When someone has
641 * a ref to the page, it may be possible that they dirty it then
642 * drop the reference. So if PageDirty is tested before page_count
643 * here, then the following race may occur:
644 *
645 * get_user_pages(&page);
646 * [user mapping goes away]
647 * write_to(page);
648 * !PageDirty(page) [good]
649 * SetPageDirty(page);
650 * put_page(page);
651 * !page_count(page) [good, discard it]
652 *
653 * [oops, our write_to data is lost]
654 *
655 * Reversing the order of the tests ensures such a situation cannot
656 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700657 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700658 *
659 * Note that if SetPageDirty is always performed via set_page_dirty,
660 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800661 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700662 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800663 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700664 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
665 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700666 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800667 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700668 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800669
670 if (PageSwapCache(page)) {
671 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700672 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800673 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400674 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700675 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700676 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500677 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700678 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500679
680 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700681 /*
682 * Remember a shadow entry for reclaimed file cache in
683 * order to detect refaults, thus thrashing, later on.
684 *
685 * But don't store shadows in an address space that is
686 * already exiting. This is not just an optizimation,
687 * inode reclaim needs to empty out the radix tree or
688 * the nodes are lost. Don't plant shadows behind its
689 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800690 *
691 * We also don't store shadows for DAX mappings because the
692 * only page cache pages found in these are zero pages
693 * covering holes, and because we don't want to mix DAX
694 * exceptional entries and shadow exceptional entries in the
695 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700696 */
697 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800698 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700699 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700700 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400701 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500702
703 if (freepage != NULL)
704 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800705 }
706
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800707 return 1;
708
709cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800711 return 0;
712}
713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714/*
Nick Piggine2867812008-07-25 19:45:30 -0700715 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
716 * someone else has a ref on the page, abort and return 0. If it was
717 * successfully detached, return 1. Assumes the caller has a single ref on
718 * this page.
719 */
720int remove_mapping(struct address_space *mapping, struct page *page)
721{
Johannes Weinera5289102014-04-03 14:47:51 -0700722 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700723 /*
724 * Unfreezing the refcount with 1 rather than 2 effectively
725 * drops the pagecache ref for us without requiring another
726 * atomic operation.
727 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700728 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700729 return 1;
730 }
731 return 0;
732}
733
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700734/**
735 * putback_lru_page - put previously isolated page onto appropriate LRU list
736 * @page: page to be put back to appropriate lru list
737 *
738 * Add previously isolated @page to appropriate LRU list.
739 * Page may still be unevictable for other reasons.
740 *
741 * lru_lock must not be held, interrupts must be enabled.
742 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700743void putback_lru_page(struct page *page)
744{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700745 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700746 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700747
Sasha Levin309381fea2014-01-23 15:52:54 -0800748 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700749
750redo:
751 ClearPageUnevictable(page);
752
Hugh Dickins39b5f292012-10-08 16:33:18 -0700753 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700754 /*
755 * For evictable pages, we can use the cache.
756 * In event of a race, worst case is we end up with an
757 * unevictable page on [in]active list.
758 * We know how to handle that.
759 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700760 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700761 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700762 } else {
763 /*
764 * Put unevictable pages directly on zone's unevictable
765 * list.
766 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700767 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700768 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700769 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700770 * When racing with an mlock or AS_UNEVICTABLE clearing
771 * (page is unlocked) make sure that if the other thread
772 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800773 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700774 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700775 * the page back to the evictable list.
776 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700777 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700778 */
779 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700780 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700781
782 /*
783 * page's status can change while we move it among lru. If an evictable
784 * page is on unevictable list, it never be freed. To avoid that,
785 * check after we added it to the list, again.
786 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700787 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700788 if (!isolate_lru_page(page)) {
789 put_page(page);
790 goto redo;
791 }
792 /* This means someone else dropped this page from LRU
793 * So, it will be freed or putback to LRU again. There is
794 * nothing to do here.
795 */
796 }
797
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700798 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700799 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700800 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700801 count_vm_event(UNEVICTABLE_PGCULLED);
802
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700803 put_page(page); /* drop ref from isolate */
804}
805
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800806enum page_references {
807 PAGEREF_RECLAIM,
808 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800809 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800810 PAGEREF_ACTIVATE,
811};
812
813static enum page_references page_check_references(struct page *page,
814 struct scan_control *sc)
815{
Johannes Weiner645747462010-03-05 13:42:22 -0800816 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800817 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800818
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700819 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
820 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800821 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800822
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800823 /*
824 * Mlock lost the isolation race with us. Let try_to_unmap()
825 * move the page to the unevictable list.
826 */
827 if (vm_flags & VM_LOCKED)
828 return PAGEREF_RECLAIM;
829
Johannes Weiner645747462010-03-05 13:42:22 -0800830 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700831 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800832 return PAGEREF_ACTIVATE;
833 /*
834 * All mapped pages start out with page table
835 * references from the instantiating fault, so we need
836 * to look twice if a mapped file page is used more
837 * than once.
838 *
839 * Mark it and spare it for another trip around the
840 * inactive list. Another page table reference will
841 * lead to its activation.
842 *
843 * Note: the mark is set for activated pages as well
844 * so that recently deactivated but used pages are
845 * quickly recovered.
846 */
847 SetPageReferenced(page);
848
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800849 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800850 return PAGEREF_ACTIVATE;
851
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800852 /*
853 * Activate file-backed executable pages after first usage.
854 */
855 if (vm_flags & VM_EXEC)
856 return PAGEREF_ACTIVATE;
857
Johannes Weiner645747462010-03-05 13:42:22 -0800858 return PAGEREF_KEEP;
859 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800860
861 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700862 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800863 return PAGEREF_RECLAIM_CLEAN;
864
865 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800866}
867
Mel Gormane2be15f2013-07-03 15:01:57 -0700868/* Check if a page is dirty or under writeback */
869static void page_check_dirty_writeback(struct page *page,
870 bool *dirty, bool *writeback)
871{
Mel Gormanb4597222013-07-03 15:02:05 -0700872 struct address_space *mapping;
873
Mel Gormane2be15f2013-07-03 15:01:57 -0700874 /*
875 * Anonymous pages are not handled by flushers and must be written
876 * from reclaim context. Do not stall reclaim based on them
877 */
878 if (!page_is_file_cache(page)) {
879 *dirty = false;
880 *writeback = false;
881 return;
882 }
883
884 /* By default assume that the page flags are accurate */
885 *dirty = PageDirty(page);
886 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700887
888 /* Verify dirty/writeback state if the filesystem supports it */
889 if (!page_has_private(page))
890 return;
891
892 mapping = page_mapping(page);
893 if (mapping && mapping->a_ops->is_dirty_writeback)
894 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700895}
896
Nick Piggine2867812008-07-25 19:45:30 -0700897/*
Andrew Morton1742f192006-03-22 00:08:21 -0800898 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 */
Andrew Morton1742f192006-03-22 00:08:21 -0800900static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700901 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700902 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700903 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700904 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700905 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700906 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700907 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700908 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700909 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910{
911 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700912 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700914 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700915 unsigned long nr_dirty = 0;
916 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800917 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700918 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700919 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
921 cond_resched();
922
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 while (!list_empty(page_list)) {
924 struct address_space *mapping;
925 struct page *page;
926 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700927 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700928 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800929 bool lazyfree = false;
930 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931
932 cond_resched();
933
934 page = lru_to_page(page_list);
935 list_del(&page->lru);
936
Nick Piggin529ae9a2008-08-02 12:01:03 +0200937 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 goto keep;
939
Sasha Levin309381fea2014-01-23 15:52:54 -0800940 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941
942 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800943
Hugh Dickins39b5f292012-10-08 16:33:18 -0700944 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700945 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700946
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700947 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800948 goto keep_locked;
949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 /* Double the slab pressure for mapped and swapcache pages */
951 if (page_mapped(page) || PageSwapCache(page))
952 sc->nr_scanned++;
953
Andy Whitcroftc661b072007-08-22 14:01:26 -0700954 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
955 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
956
Mel Gorman283aba92013-07-03 15:01:51 -0700957 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700958 * The number of dirty pages determines if a zone is marked
959 * reclaim_congested which affects wait_iff_congested. kswapd
960 * will stall and start writing pages if the tail of the LRU
961 * is all dirty unqueued pages.
962 */
963 page_check_dirty_writeback(page, &dirty, &writeback);
964 if (dirty || writeback)
965 nr_dirty++;
966
967 if (dirty && !writeback)
968 nr_unqueued_dirty++;
969
Mel Gormand04e8ac2013-07-03 15:02:03 -0700970 /*
971 * Treat this page as congested if the underlying BDI is or if
972 * pages are cycling through the LRU so quickly that the
973 * pages marked for immediate reclaim are making it to the
974 * end of the LRU a second time.
975 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700976 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800977 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400978 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700979 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700980 nr_congested++;
981
982 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700983 * If a page at the tail of the LRU is under writeback, there
984 * are three cases to consider.
985 *
986 * 1) If reclaim is encountering an excessive number of pages
987 * under writeback and this page is both under writeback and
988 * PageReclaim then it indicates that pages are being queued
989 * for IO but are being recycled through the LRU before the
990 * IO can complete. Waiting on the page itself risks an
991 * indefinite stall if it is impossible to writeback the
992 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700993 * note that the LRU is being scanned too quickly and the
994 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700995 *
Tejun Heo97c93412015-05-22 18:23:36 -0400996 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -0700997 * not marked for immediate reclaim, or the caller does not
998 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
999 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001000 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001001 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001002 * Require may_enter_fs because we would wait on fs, which
1003 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001004 * enter reclaim, and deadlock if it waits on a page for
1005 * which it is needed to do the write (loop masks off
1006 * __GFP_IO|__GFP_FS for this reason); but more thought
1007 * would probably show more reasons.
1008 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001009 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001010 * PageReclaim. memcg does not have any dirty pages
1011 * throttling so we could easily OOM just because too many
1012 * pages are in writeback and there is nothing else to
1013 * reclaim. Wait for the writeback to complete.
1014 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001015 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001016 /* Case 1 above */
1017 if (current_is_kswapd() &&
1018 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001019 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001020 nr_immediate++;
1021 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001022
1023 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001024 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001025 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001026 /*
1027 * This is slightly racy - end_page_writeback()
1028 * might have just cleared PageReclaim, then
1029 * setting PageReclaim here end up interpreted
1030 * as PageReadahead - but that does not matter
1031 * enough to care. What we do want is for this
1032 * page to have PageReclaim set next time memcg
1033 * reclaim reaches the tests above, so it will
1034 * then wait_on_page_writeback() to avoid OOM;
1035 * and it's also appropriate in global reclaim.
1036 */
1037 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001038 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001039 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001040
1041 /* Case 3 above */
1042 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001043 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001044 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001045 /* then go back and try same page again */
1046 list_add_tail(&page->lru, page_list);
1047 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001048 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001049 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050
Minchan Kim02c6de82012-10-08 16:31:55 -07001051 if (!force_reclaim)
1052 references = page_check_references(page, sc);
1053
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001054 switch (references) {
1055 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001057 case PAGEREF_KEEP:
1058 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001059 case PAGEREF_RECLAIM:
1060 case PAGEREF_RECLAIM_CLEAN:
1061 ; /* try to reclaim the page below */
1062 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 /*
1065 * Anonymous process memory has backing store?
1066 * Try to allocate it some swap space here.
1067 */
Nick Pigginb291f002008-10-18 20:26:44 -07001068 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001069 if (!(sc->gfp_mask & __GFP_IO))
1070 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001071 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001073 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001074 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075
Mel Gormane2be15f2013-07-03 15:01:57 -07001076 /* Adding to swap updated mapping */
1077 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001078 } else if (unlikely(PageTransHuge(page))) {
1079 /* Split file THP */
1080 if (split_huge_page_to_list(page, page_list))
1081 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001082 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001084 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1085
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 /*
1087 * The page is mapped into the page tables of one or more
1088 * processes. Try to unmap it here.
1089 */
1090 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001091 switch (ret = try_to_unmap(page, lazyfree ?
1092 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1093 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 case SWAP_FAIL:
1095 goto activate_locked;
1096 case SWAP_AGAIN:
1097 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001098 case SWAP_MLOCK:
1099 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001100 case SWAP_LZFREE:
1101 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 case SWAP_SUCCESS:
1103 ; /* try to free the page below */
1104 }
1105 }
1106
1107 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001108 /*
1109 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001110 * avoid risk of stack overflow but only writeback
1111 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001112 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001113 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001114 (!current_is_kswapd() ||
Mel Gorman599d0c92016-07-28 15:45:31 -07001115 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001116 /*
1117 * Immediately reclaim when written back.
1118 * Similar in principal to deactivate_page()
1119 * except we already have the page isolated
1120 * and know it's dirty
1121 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001122 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001123 SetPageReclaim(page);
1124
Mel Gormanee728862011-10-31 17:07:38 -07001125 goto keep_locked;
1126 }
1127
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001128 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001130 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001132 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 goto keep_locked;
1134
Mel Gormand950c942015-09-04 15:47:35 -07001135 /*
1136 * Page is dirty. Flush the TLB if a writable entry
1137 * potentially exists to avoid CPU writes after IO
1138 * starts and then write it out here.
1139 */
1140 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001141 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 case PAGE_KEEP:
1143 goto keep_locked;
1144 case PAGE_ACTIVATE:
1145 goto activate_locked;
1146 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001147 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001148 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001149 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001151
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 /*
1153 * A synchronous write - probably a ramdisk. Go
1154 * ahead and try to reclaim the page.
1155 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001156 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 goto keep;
1158 if (PageDirty(page) || PageWriteback(page))
1159 goto keep_locked;
1160 mapping = page_mapping(page);
1161 case PAGE_CLEAN:
1162 ; /* try to free the page below */
1163 }
1164 }
1165
1166 /*
1167 * If the page has buffers, try to free the buffer mappings
1168 * associated with this page. If we succeed we try to free
1169 * the page as well.
1170 *
1171 * We do this even if the page is PageDirty().
1172 * try_to_release_page() does not perform I/O, but it is
1173 * possible for a page to have PageDirty set, but it is actually
1174 * clean (all its buffers are clean). This happens if the
1175 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001176 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 * try_to_release_page() will discover that cleanness and will
1178 * drop the buffers and mark the page clean - it can be freed.
1179 *
1180 * Rarely, pages can have buffers and no ->mapping. These are
1181 * the pages which were not successfully invalidated in
1182 * truncate_complete_page(). We try to drop those buffers here
1183 * and if that worked, and the page is no longer mapped into
1184 * process address space (page_count == 1) it can be freed.
1185 * Otherwise, leave the page on the LRU so it is swappable.
1186 */
David Howells266cf652009-04-03 16:42:36 +01001187 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 if (!try_to_release_page(page, sc->gfp_mask))
1189 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001190 if (!mapping && page_count(page) == 1) {
1191 unlock_page(page);
1192 if (put_page_testzero(page))
1193 goto free_it;
1194 else {
1195 /*
1196 * rare race with speculative reference.
1197 * the speculative reference will free
1198 * this page shortly, so we may
1199 * increment nr_reclaimed here (and
1200 * leave it off the LRU).
1201 */
1202 nr_reclaimed++;
1203 continue;
1204 }
1205 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 }
1207
Minchan Kim854e9ed2016-01-15 16:54:53 -08001208lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001209 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001210 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211
Nick Piggina978d6f2008-10-18 20:26:58 -07001212 /*
1213 * At this point, we have no other references and there is
1214 * no way to pick any more up (removed from LRU, removed
1215 * from pagecache). Can use non-atomic bitops now (and
1216 * we obviously don't have to worry about waking up a process
1217 * waiting on the page lock, because there are no references.
1218 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001219 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001220free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001221 if (ret == SWAP_LZFREE)
1222 count_vm_event(PGLAZYFREED);
1223
Andrew Morton05ff5132006-03-22 00:08:20 -08001224 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001225
1226 /*
1227 * Is there need to periodically free_page_list? It would
1228 * appear not as the counts should be low
1229 */
1230 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 continue;
1232
Nick Pigginb291f002008-10-18 20:26:44 -07001233cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001234 if (PageSwapCache(page))
1235 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001236 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001237 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001238 continue;
1239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001241 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001242 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001243 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001244 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 SetPageActive(page);
1246 pgactivate++;
1247keep_locked:
1248 unlock_page(page);
1249keep:
1250 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001251 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001253
Johannes Weiner747db952014-08-08 14:19:24 -07001254 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001255 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001256 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001257
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001259 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001260
Mel Gorman8e950282013-07-03 15:02:02 -07001261 *ret_nr_dirty += nr_dirty;
1262 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001263 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001264 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001265 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001266 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267}
1268
Minchan Kim02c6de82012-10-08 16:31:55 -07001269unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1270 struct list_head *page_list)
1271{
1272 struct scan_control sc = {
1273 .gfp_mask = GFP_KERNEL,
1274 .priority = DEF_PRIORITY,
1275 .may_unmap = 1,
1276 };
Mel Gorman8e950282013-07-03 15:02:02 -07001277 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001278 struct page *page, *next;
1279 LIST_HEAD(clean_pages);
1280
1281 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001282 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001283 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001284 ClearPageActive(page);
1285 list_move(&page->lru, &clean_pages);
1286 }
1287 }
1288
Mel Gorman599d0c92016-07-28 15:45:31 -07001289 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001290 TTU_UNMAP|TTU_IGNORE_ACCESS,
1291 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001292 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001293 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001294 return ret;
1295}
1296
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001297/*
1298 * Attempt to remove the specified page from its LRU. Only take this page
1299 * if it is of the appropriate PageActive status. Pages which are being
1300 * freed elsewhere are also ignored.
1301 *
1302 * page: page to consider
1303 * mode: one of the LRU isolation modes defined above
1304 *
1305 * returns 0 on success, -ve errno on failure.
1306 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001307int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001308{
1309 int ret = -EINVAL;
1310
1311 /* Only take pages on the LRU. */
1312 if (!PageLRU(page))
1313 return ret;
1314
Minchan Kime46a2872012-10-08 16:33:48 -07001315 /* Compaction should not handle unevictable pages but CMA can do so */
1316 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001317 return ret;
1318
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001319 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001320
Mel Gormanc8244932012-01-12 17:19:38 -08001321 /*
1322 * To minimise LRU disruption, the caller can indicate that it only
1323 * wants to isolate pages it will be able to operate on without
1324 * blocking - clean pages for the most part.
1325 *
1326 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1327 * is used by reclaim when it is cannot write to backing storage
1328 *
1329 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1330 * that it is possible to migrate without blocking
1331 */
1332 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1333 /* All the caller can do on PageWriteback is block */
1334 if (PageWriteback(page))
1335 return ret;
1336
1337 if (PageDirty(page)) {
1338 struct address_space *mapping;
1339
1340 /* ISOLATE_CLEAN means only clean pages */
1341 if (mode & ISOLATE_CLEAN)
1342 return ret;
1343
1344 /*
1345 * Only pages without mappings or that have a
1346 * ->migratepage callback are possible to migrate
1347 * without blocking
1348 */
1349 mapping = page_mapping(page);
1350 if (mapping && !mapping->a_ops->migratepage)
1351 return ret;
1352 }
1353 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001354
Minchan Kimf80c0672011-10-31 17:06:55 -07001355 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1356 return ret;
1357
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001358 if (likely(get_page_unless_zero(page))) {
1359 /*
1360 * Be careful not to clear PageLRU until after we're
1361 * sure the page is not being freed elsewhere -- the
1362 * page release code relies on it.
1363 */
1364 ClearPageLRU(page);
1365 ret = 0;
1366 }
1367
1368 return ret;
1369}
1370
Mel Gorman7ee36a12016-07-28 15:47:17 -07001371
1372/*
1373 * Update LRU sizes after isolating pages. The LRU size updates must
1374 * be complete before mem_cgroup_update_lru_size due to a santity check.
1375 */
1376static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1377 enum lru_list lru, unsigned long *nr_zone_taken,
1378 unsigned long nr_taken)
1379{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001380 int zid;
1381
Mel Gorman7ee36a12016-07-28 15:47:17 -07001382 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1383 if (!nr_zone_taken[zid])
1384 continue;
1385
1386 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1387 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001388
1389#ifdef CONFIG_MEMCG
1390 mem_cgroup_update_lru_size(lruvec, lru, -nr_taken);
1391#endif
1392}
1393
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001394/*
Mel Gormana52633d2016-07-28 15:45:28 -07001395 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 * shrink the lists perform better by taking out a batch of pages
1397 * and working on them outside the LRU lock.
1398 *
1399 * For pagecache intensive workloads, this function is the hottest
1400 * spot in the kernel (apart from copy_*_user functions).
1401 *
1402 * Appropriate locks must be held before calling this function.
1403 *
1404 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001405 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001407 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001408 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001409 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001410 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 *
1412 * returns how many pages were moved onto *@dst.
1413 */
Andrew Morton69e05942006-03-22 00:08:19 -08001414static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001415 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001416 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001417 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001419 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001420 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001421 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001422 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Mel Gorman599d0c92016-07-28 15:45:31 -07001423 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001424 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001426 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Mel Gormand7f05522016-07-28 15:47:37 -07001427 !list_empty(src);) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001428 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001429
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 page = lru_to_page(src);
1431 prefetchw_prev_lru_page(page, src, flags);
1432
Sasha Levin309381fea2014-01-23 15:52:54 -08001433 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001434
Mel Gormanb2e18752016-07-28 15:45:37 -07001435 if (page_zonenum(page) > sc->reclaim_idx) {
1436 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001437 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001438 continue;
1439 }
1440
Mel Gormand7f05522016-07-28 15:47:37 -07001441 /*
1442 * Account for scanned and skipped separetly to avoid the pgdat
1443 * being prematurely marked unreclaimable by pgdat_reclaimable.
1444 */
1445 scan++;
1446
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001447 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001448 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001449 nr_pages = hpage_nr_pages(page);
1450 nr_taken += nr_pages;
1451 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001452 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001453 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001454
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001455 case -EBUSY:
1456 /* else it is being freed elsewhere */
1457 list_move(&page->lru, src);
1458 continue;
1459
1460 default:
1461 BUG();
1462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 }
1464
Mel Gormanb2e18752016-07-28 15:45:37 -07001465 /*
1466 * Splice any skipped pages to the start of the LRU list. Note that
1467 * this disrupts the LRU order when reclaiming for lower zones but
1468 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1469 * scanning would soon rescan the same pages to skip and put the
1470 * system at risk of premature OOM.
1471 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001472 if (!list_empty(&pages_skipped)) {
1473 int zid;
Mel Gormand7f05522016-07-28 15:47:37 -07001474 unsigned long total_skipped = 0;
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001475
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001476 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1477 if (!nr_skipped[zid])
1478 continue;
1479
1480 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Mel Gormand7f05522016-07-28 15:47:37 -07001481 total_skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001482 }
Mel Gormand7f05522016-07-28 15:47:37 -07001483
1484 /*
1485 * Account skipped pages as a partial scan as the pgdat may be
1486 * close to unreclaimable. If the LRU list is empty, account
1487 * skipped pages as a full scan.
1488 */
1489 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1490
1491 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001492 }
Hugh Dickinsf6260122012-01-12 17:20:06 -08001493 *nr_scanned = scan;
Mel Gormane5146b12016-07-28 15:46:47 -07001494 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
Hugh Dickins75b00af2012-05-29 15:07:09 -07001495 nr_taken, mode, is_file_lru(lru));
Mel Gorman7ee36a12016-07-28 15:47:17 -07001496 update_lru_sizes(lruvec, lru, nr_zone_taken, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 return nr_taken;
1498}
1499
Nick Piggin62695a82008-10-18 20:26:09 -07001500/**
1501 * isolate_lru_page - tries to isolate a page from its LRU list
1502 * @page: page to isolate from its LRU list
1503 *
1504 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1505 * vmstat statistic corresponding to whatever LRU list the page was on.
1506 *
1507 * Returns 0 if the page was removed from an LRU list.
1508 * Returns -EBUSY if the page was not on an LRU list.
1509 *
1510 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001511 * the active list, it will have PageActive set. If it was found on
1512 * the unevictable list, it will have the PageUnevictable bit set. That flag
1513 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001514 *
1515 * The vmstat statistic corresponding to the list on which the page was
1516 * found will be decremented.
1517 *
1518 * Restrictions:
1519 * (1) Must be called with an elevated refcount on the page. This is a
1520 * fundamentnal difference from isolate_lru_pages (which is called
1521 * without a stable reference).
1522 * (2) the lru_lock must not be held.
1523 * (3) interrupts must be enabled.
1524 */
1525int isolate_lru_page(struct page *page)
1526{
1527 int ret = -EBUSY;
1528
Sasha Levin309381fea2014-01-23 15:52:54 -08001529 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001530 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001531
Nick Piggin62695a82008-10-18 20:26:09 -07001532 if (PageLRU(page)) {
1533 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001534 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001535
Mel Gormana52633d2016-07-28 15:45:28 -07001536 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001537 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001538 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001539 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001540 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001541 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001542 del_page_from_lru_list(page, lruvec, lru);
1543 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001544 }
Mel Gormana52633d2016-07-28 15:45:28 -07001545 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001546 }
1547 return ret;
1548}
1549
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001550/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001551 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1552 * then get resheduled. When there are massive number of tasks doing page
1553 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1554 * the LRU list will go small and be scanned faster than necessary, leading to
1555 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001556 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001557static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001558 struct scan_control *sc)
1559{
1560 unsigned long inactive, isolated;
1561
1562 if (current_is_kswapd())
1563 return 0;
1564
Tejun Heo97c93412015-05-22 18:23:36 -04001565 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001566 return 0;
1567
1568 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001569 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1570 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001571 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001572 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1573 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001574 }
1575
Fengguang Wu3cf23842012-12-18 14:23:31 -08001576 /*
1577 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1578 * won't get blocked by normal direct-reclaimers, forming a circular
1579 * deadlock.
1580 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001581 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001582 inactive >>= 3;
1583
Rik van Riel35cd7812009-09-21 17:01:38 -07001584 return isolated > inactive;
1585}
1586
Mel Gorman66635622010-08-09 17:19:30 -07001587static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001588putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001589{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001590 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001591 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001592 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001593
Mel Gorman66635622010-08-09 17:19:30 -07001594 /*
1595 * Put back any unfreeable pages.
1596 */
Mel Gorman66635622010-08-09 17:19:30 -07001597 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001598 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001599 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001600
Sasha Levin309381fea2014-01-23 15:52:54 -08001601 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001602 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001603 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001604 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001605 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001606 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001607 continue;
1608 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001609
Mel Gorman599d0c92016-07-28 15:45:31 -07001610 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001611
Linus Torvalds7a608572011-01-17 14:42:19 -08001612 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001613 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001614 add_page_to_lru_list(page, lruvec, lru);
1615
Mel Gorman66635622010-08-09 17:19:30 -07001616 if (is_active_lru(lru)) {
1617 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001618 int numpages = hpage_nr_pages(page);
1619 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001620 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001621 if (put_page_testzero(page)) {
1622 __ClearPageLRU(page);
1623 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001624 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001625
1626 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001627 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001628 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001629 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001630 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001631 } else
1632 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001633 }
1634 }
Mel Gorman66635622010-08-09 17:19:30 -07001635
Hugh Dickins3f797682012-01-12 17:20:07 -08001636 /*
1637 * To save our caller's stack, now use input list for pages to free.
1638 */
1639 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001640}
1641
1642/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001643 * If a kernel thread (such as nfsd for loop-back mounts) services
1644 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1645 * In that case we should only throttle if the backing device it is
1646 * writing to is congested. In other cases it is safe to throttle.
1647 */
1648static int current_may_throttle(void)
1649{
1650 return !(current->flags & PF_LESS_THROTTLE) ||
1651 current->backing_dev_info == NULL ||
1652 bdi_write_congested(current->backing_dev_info);
1653}
1654
Minchan Kim91dcade2016-07-28 15:47:40 -07001655static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1656 struct scan_control *sc, enum lru_list lru)
1657{
1658 int zid;
1659 struct zone *zone;
1660 int file = is_file_lru(lru);
1661 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1662
1663 if (!global_reclaim(sc))
1664 return true;
1665
1666 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1667 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001668 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001669 continue;
1670
1671 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1672 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1673 return true;
1674 }
1675
1676 return false;
1677}
1678
NeilBrown399ba0b2014-06-04 16:07:42 -07001679/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001680 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001681 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 */
Mel Gorman66635622010-08-09 17:19:30 -07001683static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001684shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001685 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686{
1687 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001688 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001689 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001690 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001691 unsigned long nr_dirty = 0;
1692 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001693 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001694 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001695 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001696 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001697 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001698 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001699 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001700
Minchan Kim91dcade2016-07-28 15:47:40 -07001701 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1702 return 0;
1703
Mel Gorman599d0c92016-07-28 15:45:31 -07001704 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001705 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001706
1707 /* We are about to die and free our memory. Return now. */
1708 if (fatal_signal_pending(current))
1709 return SWAP_CLUSTER_MAX;
1710 }
1711
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001713
1714 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001715 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001716 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001717 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001718
Mel Gorman599d0c92016-07-28 15:45:31 -07001719 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001721 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1722 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001723
Mel Gorman599d0c92016-07-28 15:45:31 -07001724 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001725 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001726
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001727 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001728 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001729 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001730 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001731 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001732 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001733 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001734 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001735
Hillf Dantond563c052012-03-21 16:34:02 -07001736 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001737 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001738
Mel Gorman599d0c92016-07-28 15:45:31 -07001739 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001740 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1741 &nr_writeback, &nr_immediate,
1742 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001743
Mel Gorman599d0c92016-07-28 15:45:31 -07001744 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001745
Ying Han904249a2012-04-25 16:01:48 -07001746 if (global_reclaim(sc)) {
1747 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001748 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001749 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001750 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001751 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001752
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001753 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001754
Mel Gorman599d0c92016-07-28 15:45:31 -07001755 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001756
Mel Gorman599d0c92016-07-28 15:45:31 -07001757 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001758
Johannes Weiner747db952014-08-08 14:19:24 -07001759 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001760 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001761
Mel Gorman92df3a72011-10-31 17:07:56 -07001762 /*
1763 * If reclaim is isolating dirty pages under writeback, it implies
1764 * that the long-lived page allocation rate is exceeding the page
1765 * laundering rate. Either the global limits are not being effective
1766 * at throttling processes due to the page distribution throughout
1767 * zones or there is heavy usage of a slow backing device. The
1768 * only option is to throttle from reclaim context which is not ideal
1769 * as there is no guarantee the dirtying process is throttled in the
1770 * same way balance_dirty_pages() manages.
1771 *
Mel Gorman8e950282013-07-03 15:02:02 -07001772 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1773 * of pages under pages flagged for immediate reclaim and stall if any
1774 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001775 */
Mel Gorman918fc712013-07-08 16:00:25 -07001776 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001777 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001778
Mel Gormand43006d2013-07-03 15:01:50 -07001779 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001780 * Legacy memcg will stall in page writeback so avoid forcibly
1781 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001782 */
Tejun Heo97c93412015-05-22 18:23:36 -04001783 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001784 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001785 * Tag a zone as congested if all the dirty pages scanned were
1786 * backed by a congested BDI and wait_iff_congested will stall.
1787 */
1788 if (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001789 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001790
1791 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001792 * If dirty pages are scanned that are not queued for IO, it
1793 * implies that flushers are not keeping up. In this case, flag
Mel Gorman599d0c92016-07-28 15:45:31 -07001794 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
Johannes Weiner57054652014-10-09 15:28:17 -07001795 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001796 */
1797 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001798 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001799
1800 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001801 * If kswapd scans pages marked marked for immediate
1802 * reclaim and under writeback (nr_immediate), it implies
1803 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001804 * they are written so also forcibly stall.
1805 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001806 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001807 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001808 }
Mel Gormand43006d2013-07-03 15:01:50 -07001809
Mel Gorman8e950282013-07-03 15:02:02 -07001810 /*
1811 * Stall direct reclaim for IO completions if underlying BDIs or zone
1812 * is congested. Allow kswapd to continue until it starts encountering
1813 * unqueued dirty pages or cycling through the LRU too quickly.
1814 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001815 if (!sc->hibernation_mode && !current_is_kswapd() &&
1816 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001817 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001818
Mel Gorman599d0c92016-07-28 15:45:31 -07001819 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1820 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001821 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001822 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823}
1824
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001825/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 * This moves pages from the active list to the inactive list.
1827 *
1828 * We move them the other way if the page is referenced by one or more
1829 * processes, from rmap.
1830 *
1831 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001832 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001834 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 * this, so instead we remove the pages from the LRU while processing them.
1836 * It is safe to rely on PG_active against the non-LRU pages in here because
1837 * nobody will play with that bit on a non-LRU page.
1838 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001839 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 * But we had to alter page->flags anyway.
1841 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001842
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001843static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001844 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001845 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001846 enum lru_list lru)
1847{
Mel Gorman599d0c92016-07-28 15:45:31 -07001848 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001849 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001850 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001851 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001852
Wu Fengguang3eb41402009-06-16 15:33:13 -07001853 while (!list_empty(list)) {
1854 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001855 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001856
Sasha Levin309381fea2014-01-23 15:52:54 -08001857 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001858 SetPageLRU(page);
1859
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001860 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001861 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001862 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001863 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001864
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001865 if (put_page_testzero(page)) {
1866 __ClearPageLRU(page);
1867 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001868 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001869
1870 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001871 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001872 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001873 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001874 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001875 } else
1876 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001877 }
1878 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001879
Wu Fengguang3eb41402009-06-16 15:33:13 -07001880 if (!is_active_lru(lru))
1881 __count_vm_events(PGDEACTIVATE, pgmoved);
1882}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001883
Hugh Dickinsf6260122012-01-12 17:20:06 -08001884static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001885 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001886 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001887 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001889 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001890 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001891 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001893 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001894 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001896 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001897 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001898 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001899 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001900 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901
1902 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001903
1904 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001905 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001906 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001907 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001908
Mel Gorman599d0c92016-07-28 15:45:31 -07001909 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001910
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001911 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1912 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001913
Mel Gorman599d0c92016-07-28 15:45:31 -07001914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001915 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001916
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001917 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001918 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1919 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001920
Mel Gorman599d0c92016-07-28 15:45:31 -07001921 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 while (!list_empty(&l_hold)) {
1924 cond_resched();
1925 page = lru_to_page(&l_hold);
1926 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001927
Hugh Dickins39b5f292012-10-08 16:33:18 -07001928 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001929 putback_lru_page(page);
1930 continue;
1931 }
1932
Mel Gormancc715d92012-03-21 16:34:00 -07001933 if (unlikely(buffer_heads_over_limit)) {
1934 if (page_has_private(page) && trylock_page(page)) {
1935 if (page_has_private(page))
1936 try_to_release_page(page, 0);
1937 unlock_page(page);
1938 }
1939 }
1940
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001941 if (page_referenced(page, 0, sc->target_mem_cgroup,
1942 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001943 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001944 /*
1945 * Identify referenced, file-backed active pages and
1946 * give them one more trip around the active list. So
1947 * that executable code get better chances to stay in
1948 * memory under moderate memory pressure. Anon pages
1949 * are not likely to be evicted by use-once streaming
1950 * IO, plus JVM can create lots of anon VM_EXEC pages,
1951 * so we ignore them here.
1952 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001953 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001954 list_add(&page->lru, &l_active);
1955 continue;
1956 }
1957 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001958
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001959 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 list_add(&page->lru, &l_inactive);
1961 }
1962
Andrew Mortonb5557492009-01-06 14:40:13 -08001963 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001964 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001965 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001966 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001967 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001968 * Count referenced pages from currently used mappings as rotated,
1969 * even though only some of them are actually re-activated. This
1970 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001971 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001972 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001973 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001974
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001975 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1976 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07001977 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1978 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001979
Johannes Weiner747db952014-08-08 14:19:24 -07001980 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001981 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982}
1983
Rik van Riel59dc76b2016-05-20 16:56:31 -07001984/*
1985 * The inactive anon list should be small enough that the VM never has
1986 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001987 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07001988 * The inactive file list should be small enough to leave most memory
1989 * to the established workingset on the scan-resistant active list,
1990 * but large enough to avoid thrashing the aggregate readahead window.
1991 *
1992 * Both inactive lists should also be large enough that each inactive
1993 * page has a chance to be referenced again before it is reclaimed.
1994 *
1995 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1996 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1997 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1998 *
1999 * total target max
2000 * memory ratio inactive
2001 * -------------------------------------
2002 * 10MB 1 5MB
2003 * 100MB 1 50MB
2004 * 1GB 3 250MB
2005 * 10GB 10 0.9GB
2006 * 100GB 31 3GB
2007 * 1TB 101 10GB
2008 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002009 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002010static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2011 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002012{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002013 unsigned long inactive_ratio;
2014 unsigned long inactive;
2015 unsigned long active;
2016 unsigned long gb;
Mel Gormanf8d1a312016-07-28 15:47:34 -07002017 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2018 int zid;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002019
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002020 /*
2021 * If we don't have swap space, anonymous page deactivation
2022 * is pointless.
2023 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002024 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002025 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002026
Rik van Riel59dc76b2016-05-20 16:56:31 -07002027 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2028 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002029
Mel Gormanf8d1a312016-07-28 15:47:34 -07002030 /*
2031 * For zone-constrained allocations, it is necessary to check if
2032 * deactivations are required for lowmem to be reclaimed. This
2033 * calculates the inactive/active pages available in eligible zones.
2034 */
2035 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2036 struct zone *zone = &pgdat->node_zones[zid];
2037 unsigned long inactive_zone, active_zone;
2038
Mel Gorman6aa303d2016-09-01 16:14:55 -07002039 if (!managed_zone(zone))
Mel Gormanf8d1a312016-07-28 15:47:34 -07002040 continue;
2041
2042 inactive_zone = zone_page_state(zone,
2043 NR_ZONE_LRU_BASE + (file * LRU_FILE));
2044 active_zone = zone_page_state(zone,
2045 NR_ZONE_LRU_BASE + (file * LRU_FILE) + LRU_ACTIVE);
2046
2047 inactive -= min(inactive, inactive_zone);
2048 active -= min(active, active_zone);
2049 }
2050
Rik van Riel59dc76b2016-05-20 16:56:31 -07002051 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2052 if (gb)
2053 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002054 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002055 inactive_ratio = 1;
2056
2057 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002058}
2059
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002060static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002061 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002062{
Rik van Rielb39415b2009-12-14 17:59:48 -08002063 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002064 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002065 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002066 return 0;
2067 }
2068
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002069 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002070}
2071
Johannes Weiner9a265112013-02-22 16:32:17 -08002072enum scan_balance {
2073 SCAN_EQUAL,
2074 SCAN_FRACT,
2075 SCAN_ANON,
2076 SCAN_FILE,
2077};
2078
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002080 * Determine how aggressively the anon and file LRU lists should be
2081 * scanned. The relative value of each set of LRU lists is determined
2082 * by looking at the fraction of the pages scanned we did rotate back
2083 * onto the active list instead of evict.
2084 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002085 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2086 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002087 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002088static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002089 struct scan_control *sc, unsigned long *nr,
2090 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002091{
Vladimir Davydov33377672016-01-20 15:02:59 -08002092 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002093 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002094 u64 fraction[2];
2095 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002096 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002097 unsigned long anon_prio, file_prio;
2098 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002099 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002100 bool force_scan = false;
2101 unsigned long ap, fp;
2102 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002103 bool some_scanned;
2104 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07002105
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002106 /*
2107 * If the zone or memcg is small, nr[l] can be 0. This
2108 * results in no scanning on this priority and a potential
2109 * priority drop. Global direct reclaim can go to the next
2110 * zone and tends to have no problems. Global kswapd is for
2111 * zone balancing and it needs to scan a minimum amount. When
2112 * reclaiming for a memcg, a priority drop can cause high
2113 * latencies, so it's better to scan a minimum amount there as
2114 * well.
2115 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002116 if (current_is_kswapd()) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002117 if (!pgdat_reclaimable(pgdat))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002118 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002119 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002120 force_scan = true;
2121 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002122 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002123 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002124
2125 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002126 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002127 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002128 goto out;
2129 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002130
Johannes Weiner10316b32013-02-22 16:32:14 -08002131 /*
2132 * Global reclaim will swap to prevent OOM even with no
2133 * swappiness, but memcg users want to use this knob to
2134 * disable swapping for individual groups completely when
2135 * using the memory controller's swap limit feature would be
2136 * too expensive.
2137 */
Johannes Weiner02695172014-08-06 16:06:17 -07002138 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002139 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002140 goto out;
2141 }
2142
2143 /*
2144 * Do not apply any pressure balancing cleverness when the
2145 * system is close to OOM, scan both anon and file equally
2146 * (unless the swappiness setting disagrees with swapping).
2147 */
Johannes Weiner02695172014-08-06 16:06:17 -07002148 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002149 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002150 goto out;
2151 }
2152
Johannes Weiner11d16c22013-02-22 16:32:15 -08002153 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002154 * Prevent the reclaimer from falling into the cache trap: as
2155 * cache pages start out inactive, every cache fault will tip
2156 * the scan balance towards the file LRU. And as the file LRU
2157 * shrinks, so does the window for rotation from references.
2158 * This means we have a runaway feedback loop where a tiny
2159 * thrashing file LRU becomes infinitely more attractive than
2160 * anon pages. Try to detect this based on file LRU size.
2161 */
2162 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002163 unsigned long pgdatfile;
2164 unsigned long pgdatfree;
2165 int z;
2166 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002167
Mel Gorman599d0c92016-07-28 15:45:31 -07002168 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2169 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2170 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002171
Mel Gorman599d0c92016-07-28 15:45:31 -07002172 for (z = 0; z < MAX_NR_ZONES; z++) {
2173 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002174 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002175 continue;
2176
2177 total_high_wmark += high_wmark_pages(zone);
2178 }
2179
2180 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002181 scan_balance = SCAN_ANON;
2182 goto out;
2183 }
2184 }
2185
2186 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002187 * If there is enough inactive page cache, i.e. if the size of the
2188 * inactive list is greater than that of the active list *and* the
2189 * inactive list actually has some pages to scan on this priority, we
2190 * do not reclaim anything from the anonymous working set right now.
2191 * Without the second condition we could end up never scanning an
2192 * lruvec even if it has plenty of old anonymous pages unless the
2193 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002194 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002195 if (!inactive_list_is_low(lruvec, true, sc) &&
Johannes Weiner23047a92016-03-15 14:57:16 -07002196 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002197 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002198 goto out;
2199 }
2200
Johannes Weiner9a265112013-02-22 16:32:17 -08002201 scan_balance = SCAN_FRACT;
2202
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002203 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002204 * With swappiness at 100, anonymous and file have the same priority.
2205 * This scanning priority is essentially the inverse of IO cost.
2206 */
Johannes Weiner02695172014-08-06 16:06:17 -07002207 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002208 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002209
2210 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002211 * OK, so we have swap space and a fair amount of page cache
2212 * pages. We use the recently rotated / recently scanned
2213 * ratios to determine how valuable each cache is.
2214 *
2215 * Because workloads change over time (and to avoid overflow)
2216 * we keep these statistics as a floating average, which ends
2217 * up weighing recent references more than old ones.
2218 *
2219 * anon in [0], file in [1]
2220 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002221
Johannes Weiner23047a92016-03-15 14:57:16 -07002222 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2223 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2224 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2225 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002226
Mel Gorman599d0c92016-07-28 15:45:31 -07002227 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002228 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002229 reclaim_stat->recent_scanned[0] /= 2;
2230 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002231 }
2232
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002233 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002234 reclaim_stat->recent_scanned[1] /= 2;
2235 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002236 }
2237
2238 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002239 * The amount of pressure on anon vs file pages is inversely
2240 * proportional to the fraction of recently scanned pages on
2241 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002242 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002243 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002244 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002245
Satoru Moriyafe350042012-05-29 15:06:47 -07002246 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002247 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002248 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002249
Shaohua Li76a33fc2010-05-24 14:32:36 -07002250 fraction[0] = ap;
2251 fraction[1] = fp;
2252 denominator = ap + fp + 1;
2253out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002254 some_scanned = false;
2255 /* Only use force_scan on second pass. */
2256 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002257 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002258 for_each_evictable_lru(lru) {
2259 int file = is_file_lru(lru);
2260 unsigned long size;
2261 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002262
Johannes Weiner23047a92016-03-15 14:57:16 -07002263 size = lruvec_lru_size(lruvec, lru);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002264 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002265
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002266 if (!scan && pass && force_scan)
2267 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002268
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002269 switch (scan_balance) {
2270 case SCAN_EQUAL:
2271 /* Scan lists relative to size */
2272 break;
2273 case SCAN_FRACT:
2274 /*
2275 * Scan types proportional to swappiness and
2276 * their relative recent reclaim efficiency.
2277 */
2278 scan = div64_u64(scan * fraction[file],
2279 denominator);
2280 break;
2281 case SCAN_FILE:
2282 case SCAN_ANON:
2283 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002284 if ((scan_balance == SCAN_FILE) != file) {
2285 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002286 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002287 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002288 break;
2289 default:
2290 /* Look ma, no brain */
2291 BUG();
2292 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002293
2294 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002295 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002296
Johannes Weiner9a265112013-02-22 16:32:17 -08002297 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002298 * Skip the second pass and don't force_scan,
2299 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002300 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002301 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002302 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002303 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002304}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002305
Mel Gorman72b252a2015-09-04 15:47:32 -07002306#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2307static void init_tlb_ubc(void)
2308{
2309 /*
2310 * This deliberately does not clear the cpumask as it's expensive
2311 * and unnecessary. If there happens to be data in there then the
2312 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2313 * then will be cleared.
2314 */
2315 current->tlb_ubc.flush_required = false;
2316}
2317#else
2318static inline void init_tlb_ubc(void)
2319{
2320}
2321#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2322
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002323/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002324 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002325 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002326static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002327 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002328{
Mel Gormanef8f2322016-07-28 15:46:05 -07002329 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002330 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002331 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002332 unsigned long nr_to_scan;
2333 enum lru_list lru;
2334 unsigned long nr_reclaimed = 0;
2335 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2336 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002337 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002338
Vladimir Davydov33377672016-01-20 15:02:59 -08002339 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002340
Mel Gormane82e0562013-07-03 15:01:44 -07002341 /* Record the original scan target for proportional adjustments later */
2342 memcpy(targets, nr, sizeof(nr));
2343
Mel Gorman1a501902014-06-04 16:10:49 -07002344 /*
2345 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2346 * event that can occur when there is little memory pressure e.g.
2347 * multiple streaming readers/writers. Hence, we do not abort scanning
2348 * when the requested number of pages are reclaimed when scanning at
2349 * DEF_PRIORITY on the assumption that the fact we are direct
2350 * reclaiming implies that kswapd is not keeping up and it is best to
2351 * do a batch of work at once. For memcg reclaim one check is made to
2352 * abort proportional reclaim if either the file or anon lru has already
2353 * dropped to zero at the first pass.
2354 */
2355 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2356 sc->priority == DEF_PRIORITY);
2357
Mel Gorman72b252a2015-09-04 15:47:32 -07002358 init_tlb_ubc();
2359
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002360 blk_start_plug(&plug);
2361 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2362 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002363 unsigned long nr_anon, nr_file, percentage;
2364 unsigned long nr_scanned;
2365
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002366 for_each_evictable_lru(lru) {
2367 if (nr[lru]) {
2368 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2369 nr[lru] -= nr_to_scan;
2370
2371 nr_reclaimed += shrink_list(lru, nr_to_scan,
2372 lruvec, sc);
2373 }
2374 }
Mel Gormane82e0562013-07-03 15:01:44 -07002375
2376 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2377 continue;
2378
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002379 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002380 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002381 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002382 * proportionally what was requested by get_scan_count(). We
2383 * stop reclaiming one LRU and reduce the amount scanning
2384 * proportional to the original scan target.
2385 */
2386 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2387 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2388
Mel Gorman1a501902014-06-04 16:10:49 -07002389 /*
2390 * It's just vindictive to attack the larger once the smaller
2391 * has gone to zero. And given the way we stop scanning the
2392 * smaller below, this makes sure that we only make one nudge
2393 * towards proportionality once we've got nr_to_reclaim.
2394 */
2395 if (!nr_file || !nr_anon)
2396 break;
2397
Mel Gormane82e0562013-07-03 15:01:44 -07002398 if (nr_file > nr_anon) {
2399 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2400 targets[LRU_ACTIVE_ANON] + 1;
2401 lru = LRU_BASE;
2402 percentage = nr_anon * 100 / scan_target;
2403 } else {
2404 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2405 targets[LRU_ACTIVE_FILE] + 1;
2406 lru = LRU_FILE;
2407 percentage = nr_file * 100 / scan_target;
2408 }
2409
2410 /* Stop scanning the smaller of the LRU */
2411 nr[lru] = 0;
2412 nr[lru + LRU_ACTIVE] = 0;
2413
2414 /*
2415 * Recalculate the other LRU scan count based on its original
2416 * scan target and the percentage scanning already complete
2417 */
2418 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2419 nr_scanned = targets[lru] - nr[lru];
2420 nr[lru] = targets[lru] * (100 - percentage) / 100;
2421 nr[lru] -= min(nr[lru], nr_scanned);
2422
2423 lru += LRU_ACTIVE;
2424 nr_scanned = targets[lru] - nr[lru];
2425 nr[lru] = targets[lru] * (100 - percentage) / 100;
2426 nr[lru] -= min(nr[lru], nr_scanned);
2427
2428 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002429 }
2430 blk_finish_plug(&plug);
2431 sc->nr_reclaimed += nr_reclaimed;
2432
2433 /*
2434 * Even if we did not try to evict anon pages at all, we want to
2435 * rebalance the anon lru active/inactive ratio.
2436 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002437 if (inactive_list_is_low(lruvec, false, sc))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002438 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2439 sc, LRU_ACTIVE_ANON);
2440
2441 throttle_vm_writeout(sc->gfp_mask);
2442}
2443
Mel Gorman23b9da52012-05-29 15:06:20 -07002444/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002445static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002446{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002447 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002448 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002449 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002450 return true;
2451
2452 return false;
2453}
2454
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002455/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002456 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2457 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2458 * true if more pages should be reclaimed such that when the page allocator
2459 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2460 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002461 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002462static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002463 unsigned long nr_reclaimed,
2464 unsigned long nr_scanned,
2465 struct scan_control *sc)
2466{
2467 unsigned long pages_for_compaction;
2468 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002469 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002470
2471 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002472 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002473 return false;
2474
Mel Gorman28765922011-02-25 14:44:20 -08002475 /* Consider stopping depending on scan and reclaim activity */
2476 if (sc->gfp_mask & __GFP_REPEAT) {
2477 /*
2478 * For __GFP_REPEAT allocations, stop reclaiming if the
2479 * full LRU list has been scanned and we are still failing
2480 * to reclaim pages. This full LRU scan is potentially
2481 * expensive but a __GFP_REPEAT caller really wants to succeed
2482 */
2483 if (!nr_reclaimed && !nr_scanned)
2484 return false;
2485 } else {
2486 /*
2487 * For non-__GFP_REPEAT allocations which can presumably
2488 * fail without consequence, stop if we failed to reclaim
2489 * any pages from the last SWAP_CLUSTER_MAX number of
2490 * pages that were scanned. This will return to the
2491 * caller faster at the risk reclaim/compaction and
2492 * the resulting allocation attempt fails
2493 */
2494 if (!nr_reclaimed)
2495 return false;
2496 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002497
2498 /*
2499 * If we have not reclaimed enough pages for compaction and the
2500 * inactive lists are large enough, continue reclaiming
2501 */
2502 pages_for_compaction = (2UL << sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002503 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002504 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002505 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002506 if (sc->nr_reclaimed < pages_for_compaction &&
2507 inactive_lru_pages > pages_for_compaction)
2508 return true;
2509
2510 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002511 for (z = 0; z <= sc->reclaim_idx; z++) {
2512 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002513 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002514 continue;
2515
2516 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2517 case COMPACT_PARTIAL:
2518 case COMPACT_CONTINUE:
2519 return false;
2520 default:
2521 /* check next zone */
2522 ;
2523 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002524 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002525 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002526}
2527
Mel Gorman970a39a2016-07-28 15:46:35 -07002528static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002529{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002530 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002531 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002532 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002533
Johannes Weiner56600482012-01-12 17:17:59 -08002534 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002535 struct mem_cgroup *root = sc->target_mem_cgroup;
2536 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002537 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002538 .priority = sc->priority,
2539 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002540 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002541 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002542
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002543 nr_reclaimed = sc->nr_reclaimed;
2544 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002545
Andrew Morton694fbc02013-09-24 15:27:37 -07002546 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2547 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002548 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002549 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002550 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002551
Johannes Weiner241994e2015-02-11 15:26:06 -08002552 if (mem_cgroup_low(root, memcg)) {
2553 if (!sc->may_thrash)
2554 continue;
2555 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2556 }
2557
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002558 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002559 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002560
Mel Gormana9dd0a82016-07-28 15:46:02 -07002561 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2562 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002563
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002564 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002565 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002566 memcg, sc->nr_scanned - scanned,
2567 lru_pages);
2568
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002569 /* Record the group's reclaim efficiency */
2570 vmpressure(sc->gfp_mask, memcg, false,
2571 sc->nr_scanned - scanned,
2572 sc->nr_reclaimed - reclaimed);
2573
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002574 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002575 * Direct reclaim and kswapd have to scan all memory
2576 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002577 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002578 *
2579 * Limit reclaim, on the other hand, only cares about
2580 * nr_to_reclaim pages to be reclaimed and it will
2581 * retry with decreasing priority if one round over the
2582 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002583 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002584 if (!global_reclaim(sc) &&
2585 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002586 mem_cgroup_iter_break(root, memcg);
2587 break;
2588 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002589 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002590
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002591 /*
2592 * Shrink the slab caches in the same proportion that
2593 * the eligible LRU pages were scanned.
2594 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002595 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002596 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002597 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002598 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002599
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002600 if (reclaim_state) {
2601 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2602 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002603 }
2604
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002605 /* Record the subtree's reclaim efficiency */
2606 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002607 sc->nr_scanned - nr_scanned,
2608 sc->nr_reclaimed - nr_reclaimed);
2609
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002610 if (sc->nr_reclaimed - nr_reclaimed)
2611 reclaimable = true;
2612
Mel Gormana9dd0a82016-07-28 15:46:02 -07002613 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002614 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002615
2616 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002617}
2618
Vlastimil Babka53853e22014-10-09 15:27:02 -07002619/*
2620 * Returns true if compaction should go ahead for a high-order request, or
2621 * the high-order allocation would succeed without compaction.
2622 */
Mel Gorman4f588332016-07-28 15:46:38 -07002623static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002624{
Mel Gorman31483b62016-07-28 15:45:46 -07002625 unsigned long watermark;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002626 bool watermark_ok;
2627
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002628 /*
2629 * Compaction takes time to run and there are potentially other
2630 * callers using the pages just freed. Continue reclaiming until
2631 * there is a buffer of free pages available to give compaction
2632 * a reasonable chance of completing and allocating the page
2633 */
Mel Gorman4f588332016-07-28 15:46:38 -07002634 watermark = high_wmark_pages(zone) + (2UL << sc->order);
2635 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002636
2637 /*
2638 * If compaction is deferred, reclaim up to a point where
2639 * compaction will have a chance of success when re-enabled
2640 */
Mel Gorman4f588332016-07-28 15:46:38 -07002641 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002642 return watermark_ok;
2643
Vlastimil Babka53853e22014-10-09 15:27:02 -07002644 /*
2645 * If compaction is not ready to start and allocation is not likely
2646 * to succeed without it, then keep reclaiming.
2647 */
Mel Gorman4f588332016-07-28 15:46:38 -07002648 if (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002649 return false;
2650
2651 return watermark_ok;
2652}
2653
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654/*
2655 * This is the direct reclaim path, for page-allocating processes. We only
2656 * try to reclaim pages from zones which will satisfy the caller's allocation
2657 * request.
2658 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 * If a zone is deemed to be full of pinned pages then just give it a light
2660 * scan then give up on it.
2661 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002662static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663{
Mel Gormandd1a2392008-04-28 02:12:17 -07002664 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002665 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002666 unsigned long nr_soft_reclaimed;
2667 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002668 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002669 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002670
Mel Gormancc715d92012-03-21 16:34:00 -07002671 /*
2672 * If the number of buffer_heads in the machine exceeds the maximum
2673 * allowed level, force direct reclaim to scan the highmem zone as
2674 * highmem pages could be pinning lowmem pages storing buffer_heads
2675 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002676 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002677 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002678 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002679 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002680 }
Mel Gormancc715d92012-03-21 16:34:00 -07002681
Mel Gormand4debc62010-08-09 17:19:29 -07002682 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002683 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002684 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002685 * Take care memory controller reclaiming has small influence
2686 * to global LRU.
2687 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002688 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002689 if (!cpuset_zone_allowed(zone,
2690 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002691 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002692
Lisa Du6e543d52013-09-11 14:22:36 -07002693 if (sc->priority != DEF_PRIORITY &&
Mel Gorman599d0c92016-07-28 15:45:31 -07002694 !pgdat_reclaimable(zone->zone_pgdat))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002695 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002696
2697 /*
2698 * If we already have plenty of memory free for
2699 * compaction in this zone, don't free any more.
2700 * Even though compaction is invoked for any
2701 * non-zero order, only frequent costly order
2702 * reclamation is disruptive enough to become a
2703 * noticeable problem, like transparent huge
2704 * page allocations.
2705 */
2706 if (IS_ENABLED(CONFIG_COMPACTION) &&
2707 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002708 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002709 sc->compaction_ready = true;
2710 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002711 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002712
Andrew Morton0608f432013-09-24 15:27:41 -07002713 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002714 * Shrink each node in the zonelist once. If the
2715 * zonelist is ordered by zone (not the default) then a
2716 * node may be shrunk multiple times but in that case
2717 * the user prefers lower zones being preserved.
2718 */
2719 if (zone->zone_pgdat == last_pgdat)
2720 continue;
2721
2722 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002723 * This steals pages from memory cgroups over softlimit
2724 * and returns the number of reclaimed pages and
2725 * scanned pages. This works for global memory pressure
2726 * and balancing, not for a memcg's limit.
2727 */
2728 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002729 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002730 sc->order, sc->gfp_mask,
2731 &nr_soft_scanned);
2732 sc->nr_reclaimed += nr_soft_reclaimed;
2733 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002734 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002735 }
Nick Piggin408d8542006-09-25 23:31:27 -07002736
Mel Gorman79dafcd2016-07-28 15:45:53 -07002737 /* See comment about same check for global reclaim above */
2738 if (zone->zone_pgdat == last_pgdat)
2739 continue;
2740 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002741 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 }
Mel Gormane0c23272011-10-31 17:09:33 -07002743
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002744 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002745 * Restore to original mask to avoid the impact on the caller if we
2746 * promoted it to __GFP_HIGHMEM.
2747 */
2748 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002750
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751/*
2752 * This is the main entry point to direct page reclaim.
2753 *
2754 * If a full scan of the inactive list fails to free enough memory then we
2755 * are "out of memory" and something needs to be killed.
2756 *
2757 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2758 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002759 * caller can't do much about. We kick the writeback threads and take explicit
2760 * naps in the hope that some of these pages can be written. But if the
2761 * allocating task holds filesystem locks which prevent writeout this might not
2762 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002763 *
2764 * returns: 0, if no pages reclaimed
2765 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 */
Mel Gormandac1d272008-04-28 02:12:12 -07002767static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002768 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769{
Johannes Weiner241994e2015-02-11 15:26:06 -08002770 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002771 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002772 unsigned long writeback_threshold;
Johannes Weiner241994e2015-02-11 15:26:06 -08002773retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002774 delayacct_freepages_start();
2775
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002776 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002777 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002779 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002780 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2781 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002782 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002783 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002784
Balbir Singh66e17072008-02-07 00:13:56 -08002785 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002786 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002787 break;
2788
2789 if (sc->compaction_ready)
2790 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
2792 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002793 * If we're getting trouble reclaiming, start doing
2794 * writepage even in laptop mode.
2795 */
2796 if (sc->priority < DEF_PRIORITY - 2)
2797 sc->may_writepage = 1;
2798
2799 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 * Try to write back as many pages as we just scanned. This
2801 * tends to cause slow streaming writers to write data to the
2802 * disk smoothly, at the dirtying rate, which is nice. But
2803 * that's undesirable in laptop mode, where we *want* lumpy
2804 * writeout. So in laptop mode, write out the whole world.
2805 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002806 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2807 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002808 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2809 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002810 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002812 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002813
Keika Kobayashi873b4772008-07-25 01:48:52 -07002814 delayacct_freepages_end();
2815
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002816 if (sc->nr_reclaimed)
2817 return sc->nr_reclaimed;
2818
Mel Gorman0cee34f2012-01-12 17:19:49 -08002819 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002820 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002821 return 1;
2822
Johannes Weiner241994e2015-02-11 15:26:06 -08002823 /* Untapped cgroup reserves? Don't OOM, retry. */
2824 if (!sc->may_thrash) {
2825 sc->priority = initial_priority;
2826 sc->may_thrash = 1;
2827 goto retry;
2828 }
2829
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002830 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831}
2832
Mel Gorman55150612012-07-31 16:44:35 -07002833static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2834{
2835 struct zone *zone;
2836 unsigned long pfmemalloc_reserve = 0;
2837 unsigned long free_pages = 0;
2838 int i;
2839 bool wmark_ok;
2840
2841 for (i = 0; i <= ZONE_NORMAL; i++) {
2842 zone = &pgdat->node_zones[i];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002843 if (!managed_zone(zone) ||
Mel Gorman599d0c92016-07-28 15:45:31 -07002844 pgdat_reclaimable_pages(pgdat) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002845 continue;
2846
Mel Gorman55150612012-07-31 16:44:35 -07002847 pfmemalloc_reserve += min_wmark_pages(zone);
2848 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2849 }
2850
Mel Gorman675becc2014-06-04 16:07:35 -07002851 /* If there are no reserves (unexpected config) then do not throttle */
2852 if (!pfmemalloc_reserve)
2853 return true;
2854
Mel Gorman55150612012-07-31 16:44:35 -07002855 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2856
2857 /* kswapd must be awake if processes are being throttled */
2858 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002859 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002860 (enum zone_type)ZONE_NORMAL);
2861 wake_up_interruptible(&pgdat->kswapd_wait);
2862 }
2863
2864 return wmark_ok;
2865}
2866
2867/*
2868 * Throttle direct reclaimers if backing storage is backed by the network
2869 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2870 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002871 * when the low watermark is reached.
2872 *
2873 * Returns true if a fatal signal was delivered during throttling. If this
2874 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002875 */
Mel Gorman50694c22012-11-26 16:29:48 -08002876static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002877 nodemask_t *nodemask)
2878{
Mel Gorman675becc2014-06-04 16:07:35 -07002879 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002880 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002881 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002882
2883 /*
2884 * Kernel threads should not be throttled as they may be indirectly
2885 * responsible for cleaning pages necessary for reclaim to make forward
2886 * progress. kjournald for example may enter direct reclaim while
2887 * committing a transaction where throttling it could forcing other
2888 * processes to block on log_wait_commit().
2889 */
2890 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002891 goto out;
2892
2893 /*
2894 * If a fatal signal is pending, this process should not throttle.
2895 * It should return quickly so it can exit and free its memory
2896 */
2897 if (fatal_signal_pending(current))
2898 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002899
Mel Gorman675becc2014-06-04 16:07:35 -07002900 /*
2901 * Check if the pfmemalloc reserves are ok by finding the first node
2902 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2903 * GFP_KERNEL will be required for allocating network buffers when
2904 * swapping over the network so ZONE_HIGHMEM is unusable.
2905 *
2906 * Throttling is based on the first usable node and throttled processes
2907 * wait on a queue until kswapd makes progress and wakes them. There
2908 * is an affinity then between processes waking up and where reclaim
2909 * progress has been made assuming the process wakes on the same node.
2910 * More importantly, processes running on remote nodes will not compete
2911 * for remote pfmemalloc reserves and processes on different nodes
2912 * should make reasonable progress.
2913 */
2914 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002915 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002916 if (zone_idx(zone) > ZONE_NORMAL)
2917 continue;
2918
2919 /* Throttle based on the first usable node */
2920 pgdat = zone->zone_pgdat;
2921 if (pfmemalloc_watermark_ok(pgdat))
2922 goto out;
2923 break;
2924 }
2925
2926 /* If no zone was usable by the allocation flags then do not throttle */
2927 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002928 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002929
Mel Gorman68243e72012-07-31 16:44:39 -07002930 /* Account for the throttling */
2931 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2932
Mel Gorman55150612012-07-31 16:44:35 -07002933 /*
2934 * If the caller cannot enter the filesystem, it's possible that it
2935 * is due to the caller holding an FS lock or performing a journal
2936 * transaction in the case of a filesystem like ext[3|4]. In this case,
2937 * it is not safe to block on pfmemalloc_wait as kswapd could be
2938 * blocked waiting on the same lock. Instead, throttle for up to a
2939 * second before continuing.
2940 */
2941 if (!(gfp_mask & __GFP_FS)) {
2942 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2943 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002944
2945 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002946 }
2947
2948 /* Throttle until kswapd wakes the process */
2949 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2950 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002951
2952check_pending:
2953 if (fatal_signal_pending(current))
2954 return true;
2955
2956out:
2957 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002958}
2959
Mel Gormandac1d272008-04-28 02:12:12 -07002960unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002961 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002962{
Mel Gorman33906bc2010-08-09 17:19:16 -07002963 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002964 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002965 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002966 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07002967 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07002968 .order = order,
2969 .nodemask = nodemask,
2970 .priority = DEF_PRIORITY,
2971 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002972 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002973 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002974 };
2975
Mel Gorman55150612012-07-31 16:44:35 -07002976 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002977 * Do not enter reclaim if fatal signal was delivered while throttled.
2978 * 1 is returned so that the page allocator does not OOM kill at this
2979 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002980 */
Mel Gorman50694c22012-11-26 16:29:48 -08002981 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002982 return 1;
2983
Mel Gorman33906bc2010-08-09 17:19:16 -07002984 trace_mm_vmscan_direct_reclaim_begin(order,
2985 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07002986 gfp_mask,
2987 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002988
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002989 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002990
2991 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2992
2993 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002994}
2995
Andrew Mortonc255a452012-07-31 16:43:02 -07002996#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002997
Mel Gormana9dd0a82016-07-28 15:46:02 -07002998unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002999 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003000 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003001 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003002{
3003 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003004 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003005 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003006 .may_writepage = !laptop_mode,
3007 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003008 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003009 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003010 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003011 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003012
Balbir Singh4e416952009-09-23 15:56:39 -07003013 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3014 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003015
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003016 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003017 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003018 sc.gfp_mask,
3019 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003020
Balbir Singh4e416952009-09-23 15:56:39 -07003021 /*
3022 * NOTE: Although we can get the priority field, using it
3023 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003024 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003025 * will pick up pages from other mem cgroup's as well. We hack
3026 * the priority and make it zero.
3027 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003028 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003029
3030 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3031
Ying Han0ae5e892011-05-26 16:25:25 -07003032 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003033 return sc.nr_reclaimed;
3034}
3035
Johannes Weiner72835c82012-01-12 17:18:32 -08003036unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003037 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003038 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003039 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003040{
Balbir Singh4e416952009-09-23 15:56:39 -07003041 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003042 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003043 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003044 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003045 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003046 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3047 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003048 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003049 .target_mem_cgroup = memcg,
3050 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003051 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003052 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003053 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003054 };
Balbir Singh66e17072008-02-07 00:13:56 -08003055
Ying Han889976d2011-05-26 16:25:33 -07003056 /*
3057 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3058 * take care of from where we get pages. So the node where we start the
3059 * scan does not need to be the current node.
3060 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003061 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003062
3063 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003064
3065 trace_mm_vmscan_memcg_reclaim_begin(0,
3066 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003067 sc.gfp_mask,
3068 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003069
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003070 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003071
3072 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3073
3074 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003075}
3076#endif
3077
Mel Gorman1d82de62016-07-28 15:45:43 -07003078static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003079 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003080{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003081 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003082
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003083 if (!total_swap_pages)
3084 return;
3085
3086 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3087 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003088 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003089
Mel Gormanf8d1a312016-07-28 15:47:34 -07003090 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003091 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003092 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003093
3094 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3095 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003096}
3097
Mel Gorman31483b62016-07-28 15:45:46 -07003098static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003099{
Mel Gorman31483b62016-07-28 15:45:46 -07003100 unsigned long mark = high_wmark_pages(zone);
Johannes Weiner60cefed2012-11-29 13:54:23 -08003101
Mel Gorman6256c6b42016-07-28 15:45:56 -07003102 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3103 return false;
3104
3105 /*
3106 * If any eligible zone is balanced then the node is not considered
3107 * to be congested or dirty
3108 */
3109 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3110 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3111
3112 return true;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003113}
3114
Mel Gorman1741c872011-01-13 15:46:21 -08003115/*
Mel Gorman55150612012-07-31 16:44:35 -07003116 * Prepare kswapd for sleeping. This verifies that there are no processes
3117 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3118 *
3119 * Returns true if kswapd is ready to sleep
3120 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003121static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003122{
Mel Gorman1d82de62016-07-28 15:45:43 -07003123 int i;
3124
Mel Gorman55150612012-07-31 16:44:35 -07003125 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003126 * The throttled processes are normally woken up in balance_pgdat() as
3127 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3128 * race between when kswapd checks the watermarks and a process gets
3129 * throttled. There is also a potential race if processes get
3130 * throttled, kswapd wakes, a large process exits thereby balancing the
3131 * zones, which causes kswapd to exit balance_pgdat() before reaching
3132 * the wake up checks. If kswapd is going to sleep, no process should
3133 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3134 * the wake up is premature, processes will wake kswapd and get
3135 * throttled again. The difference from wake ups in balance_pgdat() is
3136 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003137 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003138 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3139 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003140
Mel Gorman1d82de62016-07-28 15:45:43 -07003141 for (i = 0; i <= classzone_idx; i++) {
3142 struct zone *zone = pgdat->node_zones + i;
3143
Mel Gorman6aa303d2016-09-01 16:14:55 -07003144 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003145 continue;
3146
Mel Gorman38087d92016-07-28 15:45:49 -07003147 if (!zone_balanced(zone, order, classzone_idx))
3148 return false;
Mel Gorman1d82de62016-07-28 15:45:43 -07003149 }
3150
Mel Gorman38087d92016-07-28 15:45:49 -07003151 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003152}
3153
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003155 * kswapd shrinks a node of pages that are at or below the highest usable
3156 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003157 *
3158 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003159 * reclaim or if the lack of progress was due to pages under writeback.
3160 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003161 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003162static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003163 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003164{
Mel Gorman1d82de62016-07-28 15:45:43 -07003165 struct zone *zone;
3166 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003167
Mel Gorman1d82de62016-07-28 15:45:43 -07003168 /* Reclaim a number of pages proportional to the number of zones */
3169 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003170 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003171 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003172 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003173 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003174
Mel Gorman1d82de62016-07-28 15:45:43 -07003175 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003176 }
3177
Mel Gorman1d82de62016-07-28 15:45:43 -07003178 /*
3179 * Historically care was taken to put equal pressure on all zones but
3180 * now pressure is applied based on node LRU order.
3181 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003182 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003183
3184 /*
3185 * Fragmentation may mean that the system cannot be rebalanced for
3186 * high-order allocations. If twice the allocation size has been
3187 * reclaimed then recheck watermarks only at order-0 to prevent
3188 * excessive reclaim. Assume that a process requested a high-order
3189 * can direct reclaim/compact.
3190 */
3191 if (sc->order && sc->nr_reclaimed >= 2UL << sc->order)
3192 sc->order = 0;
3193
Mel Gormanb8e83b92013-07-03 15:01:45 -07003194 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003195}
3196
3197/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003198 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3199 * that are eligible for use by the caller until at least one zone is
3200 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003202 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203 *
3204 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003205 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003206 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3207 * or lower is eligible for reclaim until at least one usable zone is
3208 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003210static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003213 unsigned long nr_soft_reclaimed;
3214 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003215 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003216 struct scan_control sc = {
3217 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003218 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003219 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003220 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003221 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003222 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003223 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003224 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003226 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003227 bool raise_priority = true;
3228
3229 sc.nr_reclaimed = 0;
Mel Gorman84c7a772016-07-28 15:46:44 -07003230 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231
Mel Gorman86c79f62016-07-28 15:45:59 -07003232 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003233 * If the number of buffer_heads exceeds the maximum allowed
3234 * then consider reclaiming from all zones. This has a dual
3235 * purpose -- on 64-bit systems it is expected that
3236 * buffer_heads are stripped during active rotation. On 32-bit
3237 * systems, highmem pages can pin lowmem memory and shrinking
3238 * buffers can relieve lowmem pressure. Reclaim may still not
3239 * go ahead if all eligible zones for the original allocation
3240 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003241 */
3242 if (buffer_heads_over_limit) {
3243 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3244 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003245 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003246 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247
Mel Gorman970a39a2016-07-28 15:46:35 -07003248 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003249 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003252
Mel Gorman86c79f62016-07-28 15:45:59 -07003253 /*
3254 * Only reclaim if there are no eligible zones. Check from
3255 * high to low zone as allocations prefer higher zones.
3256 * Scanning from low to high zone would allow congestion to be
3257 * cleared during a very small window when a small low
3258 * zone was balanced even under extreme pressure when the
Mel Gorman84c7a772016-07-28 15:46:44 -07003259 * overall node may be congested. Note that sc.reclaim_idx
3260 * is not used as buffer_heads_over_limit may have adjusted
3261 * it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003262 */
Mel Gorman84c7a772016-07-28 15:46:44 -07003263 for (i = classzone_idx; i >= 0; i--) {
Mel Gorman86c79f62016-07-28 15:45:59 -07003264 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003265 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003266 continue;
3267
Mel Gorman84c7a772016-07-28 15:46:44 -07003268 if (zone_balanced(zone, sc.order, classzone_idx))
Mel Gorman86c79f62016-07-28 15:45:59 -07003269 goto out;
3270 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003271
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003273 * Do some background aging of the anon list, to give
3274 * pages a chance to be referenced before reclaiming. All
3275 * pages are rotated regardless of classzone as this is
3276 * about consistent aging.
3277 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003278 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003279
3280 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003281 * If we're getting trouble reclaiming, start doing writepage
3282 * even in laptop mode.
3283 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003284 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003285 sc.may_writepage = 1;
3286
Mel Gorman1d82de62016-07-28 15:45:43 -07003287 /* Call soft limit reclaim before calling shrink_node. */
3288 sc.nr_scanned = 0;
3289 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003290 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003291 sc.gfp_mask, &nr_soft_scanned);
3292 sc.nr_reclaimed += nr_soft_reclaimed;
3293
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003294 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003295 * There should be no need to raise the scanning priority if
3296 * enough pages are already being scanned that that high
3297 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003299 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003300 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003301
3302 /*
3303 * If the low watermark is met there is no need for processes
3304 * to be throttled on pfmemalloc_wait as they should not be
3305 * able to safely make forward progress. Wake them
3306 */
3307 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3308 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003309 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003310
Mel Gormanb8e83b92013-07-03 15:01:45 -07003311 /* Check if kswapd should be suspending */
3312 if (try_to_freeze() || kthread_should_stop())
3313 break;
3314
3315 /*
3316 * Raise priority if scanning rate is too low or there was no
3317 * progress in reclaiming pages
3318 */
3319 if (raise_priority || !sc.nr_reclaimed)
3320 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003321 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Mel Gormanb8e83b92013-07-03 15:01:45 -07003323out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003324 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003325 * Return the order kswapd stopped reclaiming at as
3326 * prepare_kswapd_sleep() takes it into account. If another caller
3327 * entered the allocator slow path while kswapd was awake, order will
3328 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003329 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003330 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331}
3332
Mel Gorman38087d92016-07-28 15:45:49 -07003333static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3334 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003335{
3336 long remaining = 0;
3337 DEFINE_WAIT(wait);
3338
3339 if (freezing(current) || kthread_should_stop())
3340 return;
3341
3342 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3343
3344 /* Try to sleep for a short interval */
Mel Gormand9f21d42016-07-28 15:46:41 -07003345 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003346 /*
3347 * Compaction records what page blocks it recently failed to
3348 * isolate pages from and skips them in the future scanning.
3349 * When kswapd is going to sleep, it is reasonable to assume
3350 * that pages and compaction may succeed so reset the cache.
3351 */
3352 reset_isolation_suitable(pgdat);
3353
3354 /*
3355 * We have freed the memory, now we should compact it to make
3356 * allocation of the requested order possible.
3357 */
Mel Gorman38087d92016-07-28 15:45:49 -07003358 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003359
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003360 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003361
3362 /*
3363 * If woken prematurely then reset kswapd_classzone_idx and
3364 * order. The values will either be from a wakeup request or
3365 * the previous request that slept prematurely.
3366 */
3367 if (remaining) {
3368 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3369 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3370 }
3371
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003372 finish_wait(&pgdat->kswapd_wait, &wait);
3373 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3374 }
3375
3376 /*
3377 * After a short sleep, check if it was a premature sleep. If not, then
3378 * go fully to sleep until explicitly woken up.
3379 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003380 if (!remaining &&
3381 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003382 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3383
3384 /*
3385 * vmstat counters are not perfectly accurate and the estimated
3386 * value for counters such as NR_FREE_PAGES can deviate from the
3387 * true value by nr_online_cpus * threshold. To avoid the zone
3388 * watermarks being breached while under pressure, we reduce the
3389 * per-cpu vmstat threshold while kswapd is awake and restore
3390 * them before going back to sleep.
3391 */
3392 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003393
3394 if (!kthread_should_stop())
3395 schedule();
3396
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003397 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3398 } else {
3399 if (remaining)
3400 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3401 else
3402 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3403 }
3404 finish_wait(&pgdat->kswapd_wait, &wait);
3405}
3406
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407/*
3408 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003409 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 *
3411 * This basically trickles out pages so that we have _some_
3412 * free memory available even if there is no other activity
3413 * that frees anything up. This is needed for things like routing
3414 * etc, where we otherwise might have all activity going on in
3415 * asynchronous contexts that cannot page things out.
3416 *
3417 * If there are applications that are active memory-allocators
3418 * (most normal use), this basically shouldn't matter.
3419 */
3420static int kswapd(void *p)
3421{
Mel Gorman38087d92016-07-28 15:45:49 -07003422 unsigned int alloc_order, reclaim_order, classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423 pg_data_t *pgdat = (pg_data_t*)p;
3424 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003425
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 struct reclaim_state reclaim_state = {
3427 .reclaimed_slab = 0,
3428 };
Rusty Russella70f7302009-03-13 14:49:46 +10303429 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430
Nick Piggincf40bd12009-01-21 08:12:39 +01003431 lockdep_set_current_reclaim_state(GFP_KERNEL);
3432
Rusty Russell174596a2009-01-01 10:12:29 +10303433 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003434 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435 current->reclaim_state = &reclaim_state;
3436
3437 /*
3438 * Tell the memory management that we're a "memory allocator",
3439 * and that if we need more memory we should get access to it
3440 * regardless (see "__alloc_pages()"). "kswapd" should
3441 * never get caught in the normal page freeing logic.
3442 *
3443 * (Kswapd normally doesn't need memory anyway, but sometimes
3444 * you need a small amount of memory in order to be able to
3445 * page out something else, and this flag essentially protects
3446 * us from recursively trying to free more memory as we're
3447 * trying to free the first piece of memory in the first place).
3448 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003449 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003450 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451
Mel Gorman38087d92016-07-28 15:45:49 -07003452 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3453 pgdat->kswapd_classzone_idx = classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003455 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003456
Mel Gorman38087d92016-07-28 15:45:49 -07003457kswapd_try_sleep:
3458 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3459 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003460
Mel Gorman38087d92016-07-28 15:45:49 -07003461 /* Read the new order and classzone_idx */
3462 alloc_order = reclaim_order = pgdat->kswapd_order;
3463 classzone_idx = pgdat->kswapd_classzone_idx;
3464 pgdat->kswapd_order = 0;
3465 pgdat->kswapd_classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466
David Rientjes8fe23e02009-12-14 17:58:33 -08003467 ret = try_to_freeze();
3468 if (kthread_should_stop())
3469 break;
3470
3471 /*
3472 * We can speed up thawing tasks if we don't call balance_pgdat
3473 * after returning from the refrigerator
3474 */
Mel Gorman38087d92016-07-28 15:45:49 -07003475 if (ret)
3476 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003477
Mel Gorman38087d92016-07-28 15:45:49 -07003478 /*
3479 * Reclaim begins at the requested order but if a high-order
3480 * reclaim fails then kswapd falls back to reclaiming for
3481 * order-0. If that happens, kswapd will consider sleeping
3482 * for the order it finished reclaiming at (reclaim_order)
3483 * but kcompactd is woken to compact for the original
3484 * request (alloc_order).
3485 */
Mel Gormane5146b12016-07-28 15:46:47 -07003486 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3487 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003488 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3489 if (reclaim_order < alloc_order)
3490 goto kswapd_try_sleep;
3491
3492 alloc_order = reclaim_order = pgdat->kswapd_order;
3493 classzone_idx = pgdat->kswapd_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003495
Johannes Weiner71abdc12014-06-06 14:35:35 -07003496 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003497 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003498 lockdep_clear_current_reclaim_state();
3499
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 return 0;
3501}
3502
3503/*
3504 * A zone is low on free memory, so wake its kswapd task to service it.
3505 */
Mel Gorman99504742011-01-13 15:46:20 -08003506void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507{
3508 pg_data_t *pgdat;
Mel Gormane1a55632016-07-28 15:46:26 -07003509 int z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510
Mel Gorman6aa303d2016-09-01 16:14:55 -07003511 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 return;
3513
Vladimir Davydov344736f2014-10-20 15:50:30 +04003514 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003516 pgdat = zone->zone_pgdat;
Mel Gorman38087d92016-07-28 15:45:49 -07003517 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3518 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003519 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003521
3522 /* Only wake kswapd if all zones are unbalanced */
3523 for (z = 0; z <= classzone_idx; z++) {
3524 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003525 if (!managed_zone(zone))
Mel Gormane1a55632016-07-28 15:46:26 -07003526 continue;
3527
3528 if (zone_balanced(zone, order, classzone_idx))
3529 return;
3530 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08003531
3532 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003533 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534}
3535
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003536#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003538 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003539 * freed pages.
3540 *
3541 * Rather than trying to age LRUs the aim is to preserve the overall
3542 * LRU order by reclaiming preferentially
3543 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003545unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003547 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003548 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003549 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003550 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003551 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003552 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003553 .may_writepage = 1,
3554 .may_unmap = 1,
3555 .may_swap = 1,
3556 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003558 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003559 struct task_struct *p = current;
3560 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003562 p->flags |= PF_MEMALLOC;
3563 lockdep_set_current_reclaim_state(sc.gfp_mask);
3564 reclaim_state.reclaimed_slab = 0;
3565 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003566
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003567 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003568
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003569 p->reclaim_state = NULL;
3570 lockdep_clear_current_reclaim_state();
3571 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003572
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003573 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003575#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577/* It's optimal to keep kswapds on the same CPUs as their memory, but
3578 not required for correctness. So if the last cpu in a node goes
3579 away, we get changed to run anywhere: as the first one comes back,
3580 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003581static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3582 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003584 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003586 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003587 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003588 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303589 const struct cpumask *mask;
3590
3591 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003592
Rusty Russell3e597942009-01-01 10:12:24 +10303593 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003595 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596 }
3597 }
3598 return NOTIFY_OK;
3599}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
Yasunori Goto3218ae12006-06-27 02:53:33 -07003601/*
3602 * This kswapd start function will be called by init and node-hot-add.
3603 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3604 */
3605int kswapd_run(int nid)
3606{
3607 pg_data_t *pgdat = NODE_DATA(nid);
3608 int ret = 0;
3609
3610 if (pgdat->kswapd)
3611 return 0;
3612
3613 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3614 if (IS_ERR(pgdat->kswapd)) {
3615 /* failure at boot is fatal */
3616 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003617 pr_err("Failed to start kswapd on node %d\n", nid);
3618 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003619 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003620 }
3621 return ret;
3622}
3623
David Rientjes8fe23e02009-12-14 17:58:33 -08003624/*
Jiang Liud8adde12012-07-11 14:01:52 -07003625 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003626 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003627 */
3628void kswapd_stop(int nid)
3629{
3630 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3631
Jiang Liud8adde12012-07-11 14:01:52 -07003632 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003633 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003634 NODE_DATA(nid)->kswapd = NULL;
3635 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003636}
3637
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638static int __init kswapd_init(void)
3639{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003640 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003641
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003643 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003644 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 hotcpu_notifier(cpu_callback, 0);
3646 return 0;
3647}
3648
3649module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003650
3651#ifdef CONFIG_NUMA
3652/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003653 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003654 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003655 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003656 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003657 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003658int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003659
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003660#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003661#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003662#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003663#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003664
Christoph Lameter9eeff232006-01-18 17:42:31 -08003665/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003666 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003667 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3668 * a zone.
3669 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003670#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003671
Christoph Lameter9eeff232006-01-18 17:42:31 -08003672/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003673 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003674 * occur.
3675 */
3676int sysctl_min_unmapped_ratio = 1;
3677
3678/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003679 * If the number of slab pages in a zone grows beyond this percentage then
3680 * slab reclaim needs to occur.
3681 */
3682int sysctl_min_slab_ratio = 5;
3683
Mel Gorman11fb9982016-07-28 15:46:20 -07003684static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003685{
Mel Gorman11fb9982016-07-28 15:46:20 -07003686 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3687 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3688 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003689
3690 /*
3691 * It's possible for there to be more file mapped pages than
3692 * accounted for by the pages on the file LRU lists because
3693 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3694 */
3695 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3696}
3697
3698/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003699static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003700{
Alexandru Moised031a152015-11-05 18:48:08 -08003701 unsigned long nr_pagecache_reclaimable;
3702 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003703
3704 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003705 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003706 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003707 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003708 * a better estimate
3709 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003710 if (node_reclaim_mode & RECLAIM_UNMAP)
3711 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003712 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003713 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003714
3715 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003716 if (!(node_reclaim_mode & RECLAIM_WRITE))
3717 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003718
3719 /* Watch for any possible underflows due to delta */
3720 if (unlikely(delta > nr_pagecache_reclaimable))
3721 delta = nr_pagecache_reclaimable;
3722
3723 return nr_pagecache_reclaimable - delta;
3724}
3725
Christoph Lameter0ff38492006-09-25 23:31:52 -07003726/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003727 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003728 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003729static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003730{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003731 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003732 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003733 struct task_struct *p = current;
3734 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003735 int classzone_idx = gfp_zone(gfp_mask);
Andrew Morton179e9632006-03-22 00:08:18 -08003736 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003737 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003738 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003739 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003740 .priority = NODE_RECLAIM_PRIORITY,
3741 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3742 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003743 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003744 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003745 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003746
Christoph Lameter9eeff232006-01-18 17:42:31 -08003747 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003748 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003749 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003750 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003751 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003752 */
3753 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003754 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003755 reclaim_state.reclaimed_slab = 0;
3756 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003757
Mel Gormana5f5f912016-07-28 15:46:32 -07003758 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003759 /*
3760 * Free memory by calling shrink zone with increasing
3761 * priorities until we have enough memory freed.
3762 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003763 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003764 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003765 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003766 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003767
Christoph Lameter9eeff232006-01-18 17:42:31 -08003768 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003769 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003770 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003771 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003772}
Andrew Morton179e9632006-03-22 00:08:18 -08003773
Mel Gormana5f5f912016-07-28 15:46:32 -07003774int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003775{
David Rientjesd773ed62007-10-16 23:26:01 -07003776 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003777
3778 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003779 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003780 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003781 *
Christoph Lameter96146342006-07-03 00:24:13 -07003782 * A small portion of unmapped file backed pages is needed for
3783 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003784 * thrown out if the node is overallocated. So we do not reclaim
3785 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003786 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003787 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003788 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3789 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3790 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003791
Mel Gormana5f5f912016-07-28 15:46:32 -07003792 if (!pgdat_reclaimable(pgdat))
3793 return NODE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003794
Andrew Morton179e9632006-03-22 00:08:18 -08003795 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003796 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003797 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003798 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003799 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003800
3801 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003802 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003803 * have associated processors. This will favor the local processor
3804 * over remote processors and spread off node memory allocations
3805 * as wide as possible.
3806 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003807 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3808 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003809
Mel Gormana5f5f912016-07-28 15:46:32 -07003810 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3811 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003812
Mel Gormana5f5f912016-07-28 15:46:32 -07003813 ret = __node_reclaim(pgdat, gfp_mask, order);
3814 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003815
Mel Gorman24cf725182009-06-16 15:33:23 -07003816 if (!ret)
3817 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3818
David Rientjesd773ed62007-10-16 23:26:01 -07003819 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003820}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003821#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003822
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003823/*
3824 * page_evictable - test whether a page is evictable
3825 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003826 *
3827 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003828 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003829 *
3830 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003831 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003832 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003833 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003834 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003835int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003836{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003837 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003838}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003839
Hugh Dickins850465792012-01-20 14:34:19 -08003840#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003841/**
Hugh Dickins24513262012-01-20 14:34:21 -08003842 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3843 * @pages: array of pages to check
3844 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003845 *
Hugh Dickins24513262012-01-20 14:34:21 -08003846 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003847 *
3848 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003849 */
Hugh Dickins24513262012-01-20 14:34:21 -08003850void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003851{
Johannes Weiner925b7672012-01-12 17:18:15 -08003852 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003853 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003854 int pgscanned = 0;
3855 int pgrescued = 0;
3856 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003857
Hugh Dickins24513262012-01-20 14:34:21 -08003858 for (i = 0; i < nr_pages; i++) {
3859 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003860 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003861
Hugh Dickins24513262012-01-20 14:34:21 -08003862 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003863 if (pagepgdat != pgdat) {
3864 if (pgdat)
3865 spin_unlock_irq(&pgdat->lru_lock);
3866 pgdat = pagepgdat;
3867 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003868 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003869 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003870
Hugh Dickins24513262012-01-20 14:34:21 -08003871 if (!PageLRU(page) || !PageUnevictable(page))
3872 continue;
3873
Hugh Dickins39b5f292012-10-08 16:33:18 -07003874 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003875 enum lru_list lru = page_lru_base_type(page);
3876
Sasha Levin309381fea2014-01-23 15:52:54 -08003877 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003878 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003879 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3880 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003881 pgrescued++;
3882 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003883 }
Hugh Dickins24513262012-01-20 14:34:21 -08003884
Mel Gorman785b99f2016-07-28 15:47:23 -07003885 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003886 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3887 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003888 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003889 }
Hugh Dickins850465792012-01-20 14:34:19 -08003890}
3891#endif /* CONFIG_SHMEM */