blob: b4d398bb92acb2b1e4a78cf985a3584d77f08b67 [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
Michal Hocko71053132017-02-22 15:45:58 -0800237/**
238 * lruvec_lru_size - Returns the number of pages on the given LRU list.
239 * @lruvec: lru vector
240 * @lru: lru to use
241 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
242 */
243unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800244{
Michal Hocko71053132017-02-22 15:45:58 -0800245 unsigned long lru_size;
246 int zid;
247
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700248 if (!mem_cgroup_disabled())
Michal Hocko71053132017-02-22 15:45:58 -0800249 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
250 else
251 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800252
Michal Hocko71053132017-02-22 15:45:58 -0800253 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
254 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
255 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800256
Michal Hocko71053132017-02-22 15:45:58 -0800257 if (!managed_zone(zone))
258 continue;
Michal Hocko07fc9572017-01-10 16:58:04 -0800259
Michal Hocko71053132017-02-22 15:45:58 -0800260 if (!mem_cgroup_disabled())
261 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
262 else
263 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
264 NR_ZONE_LRU_BASE + lru);
265 lru_size -= min(size, lru_size);
266 }
267
268 return lru_size;
269
Michal Hocko07fc9572017-01-10 16:58:04 -0800270}
271
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000273 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000275int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000277 size_t size = sizeof(*shrinker->nr_deferred);
278
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279 if (shrinker->flags & SHRINKER_NUMA_AWARE)
280 size *= nr_node_ids;
281
282 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
283 if (!shrinker->nr_deferred)
284 return -ENOMEM;
285
Rusty Russell8e1f9362007-07-17 04:03:17 -0700286 down_write(&shrinker_rwsem);
287 list_add_tail(&shrinker->list, &shrinker_list);
288 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000289 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700291EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
293/*
294 * Remove one
295 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700296void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 down_write(&shrinker_rwsem);
299 list_del(&shrinker->list);
300 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700301 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700303EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
305#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000306
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800307static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
308 struct shrinker *shrinker,
309 unsigned long nr_scanned,
310 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000312 unsigned long freed = 0;
313 unsigned long long delta;
314 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700315 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000316 long nr;
317 long new_nr;
318 int nid = shrinkctl->nid;
319 long batch_size = shrinker->batch ? shrinker->batch
320 : SHRINK_BATCH;
Shaohua Li04597be2016-12-12 16:41:50 -0800321 long scanned = 0, next_deferred;
Liam Marka708a022014-11-25 14:43:42 -0800322 long min_cache_size = batch_size;
323
324 if (current_is_kswapd())
325 min_cache_size = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000326
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700327 freeable = shrinker->count_objects(shrinker, shrinkctl);
328 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000329 return 0;
330
331 /*
332 * copy the current shrinker scan count into a local variable
333 * and zero it so that other concurrent shrinker invocations
334 * don't also do this scanning work.
335 */
336 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
337
338 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800339 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700340 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800341 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000342 total_scan += delta;
343 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800344 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000345 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700346 total_scan = freeable;
Shaohua Li04597be2016-12-12 16:41:50 -0800347 next_deferred = nr;
348 } else
349 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000350
351 /*
352 * We need to avoid excessive windup on filesystem shrinkers
353 * due to large numbers of GFP_NOFS allocations causing the
354 * shrinkers to return -1 all the time. This results in a large
355 * nr being built up so when a shrink that can do some work
356 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700357 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000358 * memory.
359 *
360 * Hence only allow the shrinker to scan the entire cache when
361 * a large delta change is calculated directly.
362 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700363 if (delta < freeable / 4)
364 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000365
366 /*
367 * Avoid risking looping forever due to too large nr value:
368 * never try to free more than twice the estimate number of
369 * freeable entries.
370 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700371 if (total_scan > freeable * 2)
372 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000373
374 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800375 nr_scanned, nr_eligible,
376 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000377
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800378 /*
379 * Normally, we should not scan less than batch_size objects in one
380 * pass to avoid too frequent shrinker calls, but if the slab has less
381 * than batch_size objects in total and we are really tight on memory,
382 * we will try to reclaim all available objects, otherwise we can end
383 * up failing allocations although there are plenty of reclaimable
384 * objects spread over several slabs with usage less than the
385 * batch_size.
386 *
387 * We detect the "tight on memory" situations by looking at the total
388 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700389 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800390 * scanning at high prio and therefore should try to reclaim as much as
391 * possible.
392 */
Liam Marka708a022014-11-25 14:43:42 -0800393 while (total_scan > min_cache_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700394 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000395 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800396 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000397
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800398 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000399 ret = shrinker->scan_objects(shrinker, shrinkctl);
400 if (ret == SHRINK_STOP)
401 break;
402 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000403
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800404 count_vm_events(SLABS_SCANNED, nr_to_scan);
405 total_scan -= nr_to_scan;
Shaohua Li04597be2016-12-12 16:41:50 -0800406 scanned += nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000407
408 cond_resched();
409 }
410
Shaohua Li04597be2016-12-12 16:41:50 -0800411 if (next_deferred >= scanned)
412 next_deferred -= scanned;
413 else
414 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000415 /*
416 * move the unused scan count back into the shrinker in a
417 * manner that handles concurrent updates. If we exhausted the
418 * scan, there is no need to do an update.
419 */
Shaohua Li04597be2016-12-12 16:41:50 -0800420 if (next_deferred > 0)
421 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000422 &shrinker->nr_deferred[nid]);
423 else
424 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
425
Dave Hansendf9024a2014-06-04 16:08:07 -0700426 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000427 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800430/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800431 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800432 * @gfp_mask: allocation context
433 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800434 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800435 * @nr_scanned: pressure numerator
436 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800438 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800440 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
441 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800443 * @memcg specifies the memory cgroup to target. If it is not NULL,
444 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700445 * objects from the memory cgroup specified. Otherwise, only unaware
446 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800447 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800448 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
449 * the available objects should be scanned. Page reclaim for example
450 * passes the number of pages scanned and the number of pages on the
451 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
452 * when it encountered mapped pages. The ratio is further biased by
453 * the ->seeks setting of the shrink function, which indicates the
454 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800456 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800458static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
459 struct mem_cgroup *memcg,
460 unsigned long nr_scanned,
461 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462{
463 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000464 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700466 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800467 return 0;
468
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800469 if (nr_scanned == 0)
470 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Minchan Kimf06590b2011-05-24 17:11:11 -0700472 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000473 /*
474 * If we would return 0, our callers would understand that we
475 * have nothing else to shrink and give up trying. By returning
476 * 1 we keep it going and assume we'll be able to shrink next
477 * time.
478 */
479 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700480 goto out;
481 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
483 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800484 struct shrink_control sc = {
485 .gfp_mask = gfp_mask,
486 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800487 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800488 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800489
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700490 /*
491 * If kernel memory accounting is disabled, we ignore
492 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
493 * passing NULL for memcg.
494 */
495 if (memcg_kmem_enabled() &&
496 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800497 continue;
498
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800499 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
500 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000501
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800502 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800504
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700506out:
507 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000508 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509}
510
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800511void drop_slab_node(int nid)
512{
513 unsigned long freed;
514
515 do {
516 struct mem_cgroup *memcg = NULL;
517
518 freed = 0;
519 do {
520 freed += shrink_slab(GFP_KERNEL, nid, memcg,
521 1000, 1000);
522 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
523 } while (freed > 10);
524}
525
526void drop_slab(void)
527{
528 int nid;
529
530 for_each_online_node(nid)
531 drop_slab_node(nid);
532}
533
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534static inline int is_page_cache_freeable(struct page *page)
535{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700536 /*
537 * A freeable page cache page is referenced only by the caller
538 * that isolated the page, the page cache radix tree and
539 * optional buffer heads at page->private.
540 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700541 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
Tejun Heo703c2702015-05-22 17:13:44 -0400544static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
Christoph Lameter930d9152006-01-08 01:00:47 -0800546 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400548 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400550 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 return 1;
552 return 0;
553}
554
555/*
556 * We detected a synchronous write error writing a page out. Probably
557 * -ENOSPC. We need to propagate that into the address_space for a subsequent
558 * fsync(), msync() or close().
559 *
560 * The tricky part is that after writepage we cannot touch the mapping: nothing
561 * prevents it from being freed up. But we have a ref on the page and once
562 * that page is locked, the mapping is pinned.
563 *
564 * We're allowed to run sleeping lock_page() here because we know the caller has
565 * __GFP_FS.
566 */
567static void handle_write_error(struct address_space *mapping,
568 struct page *page, int error)
569{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100570 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700571 if (page_mapping(page) == mapping)
572 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 unlock_page(page);
574}
575
Christoph Lameter04e62a22006-06-23 02:03:38 -0700576/* possible outcome of pageout() */
577typedef enum {
578 /* failed to write page out, page is locked */
579 PAGE_KEEP,
580 /* move page to the active list, page is locked */
581 PAGE_ACTIVATE,
582 /* page has been sent to the disk successfully, page is unlocked */
583 PAGE_SUCCESS,
584 /* page is clean and locked */
585 PAGE_CLEAN,
586} pageout_t;
587
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588/*
Andrew Morton1742f192006-03-22 00:08:21 -0800589 * pageout is called by shrink_page_list() for each dirty page.
590 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700592static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700593 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594{
595 /*
596 * If the page is dirty, only perform writeback if that write
597 * will be non-blocking. To prevent this allocation from being
598 * stalled by pagecache activity. But note that there may be
599 * stalls if we need to run get_block(). We could test
600 * PagePrivate for that.
601 *
Al Viro81742022014-04-03 03:17:43 -0400602 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 * this page's queue, we can perform writeback even if that
604 * will block.
605 *
606 * If the page is swapcache, write it back even if that would
607 * block, for some throttling. This happens by accident, because
608 * swap_backing_dev_info is bust: it doesn't reflect the
609 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 */
611 if (!is_page_cache_freeable(page))
612 return PAGE_KEEP;
613 if (!mapping) {
614 /*
615 * Some data journaling orphaned pages can have
616 * page->mapping == NULL while being dirty with clean buffers.
617 */
David Howells266cf652009-04-03 16:42:36 +0100618 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 if (try_to_free_buffers(page)) {
620 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700621 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 return PAGE_CLEAN;
623 }
624 }
625 return PAGE_KEEP;
626 }
627 if (mapping->a_ops->writepage == NULL)
628 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400629 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 return PAGE_KEEP;
631
632 if (clear_page_dirty_for_io(page)) {
633 int res;
634 struct writeback_control wbc = {
635 .sync_mode = WB_SYNC_NONE,
636 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700637 .range_start = 0,
638 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 .for_reclaim = 1,
640 };
641
642 SetPageReclaim(page);
643 res = mapping->a_ops->writepage(page, &wbc);
644 if (res < 0)
645 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800646 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 ClearPageReclaim(page);
648 return PAGE_ACTIVATE;
649 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700650
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 if (!PageWriteback(page)) {
652 /* synchronous write or broken a_ops? */
653 ClearPageReclaim(page);
654 }
yalin wang3aa23852016-01-14 15:18:30 -0800655 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700656 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 return PAGE_SUCCESS;
658 }
659
660 return PAGE_CLEAN;
661}
662
Andrew Mortona649fd92006-10-17 00:09:36 -0700663/*
Nick Piggine2867812008-07-25 19:45:30 -0700664 * Same as remove_mapping, but if the page is removed from the mapping, it
665 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700666 */
Johannes Weinera5289102014-04-03 14:47:51 -0700667static int __remove_mapping(struct address_space *mapping, struct page *page,
668 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800669{
Greg Thelenc4843a72015-05-22 17:13:16 -0400670 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400671
Nick Piggin28e4d962006-09-25 23:31:23 -0700672 BUG_ON(!PageLocked(page));
673 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800674
Greg Thelenc4843a72015-05-22 17:13:16 -0400675 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800676 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700677 * The non racy check for a busy page.
678 *
679 * Must be careful with the order of the tests. When someone has
680 * a ref to the page, it may be possible that they dirty it then
681 * drop the reference. So if PageDirty is tested before page_count
682 * here, then the following race may occur:
683 *
684 * get_user_pages(&page);
685 * [user mapping goes away]
686 * write_to(page);
687 * !PageDirty(page) [good]
688 * SetPageDirty(page);
689 * put_page(page);
690 * !page_count(page) [good, discard it]
691 *
692 * [oops, our write_to data is lost]
693 *
694 * Reversing the order of the tests ensures such a situation cannot
695 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700696 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700697 *
698 * Note that if SetPageDirty is always performed via set_page_dirty,
699 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800700 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700701 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800702 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700703 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
704 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700705 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800706 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700707 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800708
709 if (PageSwapCache(page)) {
710 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700711 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800712 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400713 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700714 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700715 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500716 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700717 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500718
719 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700720 /*
721 * Remember a shadow entry for reclaimed file cache in
722 * order to detect refaults, thus thrashing, later on.
723 *
724 * But don't store shadows in an address space that is
725 * already exiting. This is not just an optizimation,
726 * inode reclaim needs to empty out the radix tree or
727 * the nodes are lost. Don't plant shadows behind its
728 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800729 *
730 * We also don't store shadows for DAX mappings because the
731 * only page cache pages found in these are zero pages
732 * covering holes, and because we don't want to mix DAX
733 * exceptional entries and shadow exceptional entries in the
734 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700735 */
736 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800737 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700738 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700739 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400740 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500741
742 if (freepage != NULL)
743 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800744 }
745
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800746 return 1;
747
748cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400749 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800750 return 0;
751}
752
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753/*
Nick Piggine2867812008-07-25 19:45:30 -0700754 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
755 * someone else has a ref on the page, abort and return 0. If it was
756 * successfully detached, return 1. Assumes the caller has a single ref on
757 * this page.
758 */
759int remove_mapping(struct address_space *mapping, struct page *page)
760{
Johannes Weinera5289102014-04-03 14:47:51 -0700761 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700762 /*
763 * Unfreezing the refcount with 1 rather than 2 effectively
764 * drops the pagecache ref for us without requiring another
765 * atomic operation.
766 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700767 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700768 return 1;
769 }
770 return 0;
771}
772
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700773/**
774 * putback_lru_page - put previously isolated page onto appropriate LRU list
775 * @page: page to be put back to appropriate lru list
776 *
777 * Add previously isolated @page to appropriate LRU list.
778 * Page may still be unevictable for other reasons.
779 *
780 * lru_lock must not be held, interrupts must be enabled.
781 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700782void putback_lru_page(struct page *page)
783{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700784 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700785 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700786
Sasha Levin309381fea2014-01-23 15:52:54 -0800787 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700788
789redo:
790 ClearPageUnevictable(page);
791
Hugh Dickins39b5f292012-10-08 16:33:18 -0700792 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700793 /*
794 * For evictable pages, we can use the cache.
795 * In event of a race, worst case is we end up with an
796 * unevictable page on [in]active list.
797 * We know how to handle that.
798 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700799 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700800 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700801 } else {
802 /*
803 * Put unevictable pages directly on zone's unevictable
804 * list.
805 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700806 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700807 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700808 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700809 * When racing with an mlock or AS_UNEVICTABLE clearing
810 * (page is unlocked) make sure that if the other thread
811 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800812 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700813 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700814 * the page back to the evictable list.
815 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700816 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700817 */
818 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700819 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700820
821 /*
822 * page's status can change while we move it among lru. If an evictable
823 * page is on unevictable list, it never be freed. To avoid that,
824 * check after we added it to the list, again.
825 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700826 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700827 if (!isolate_lru_page(page)) {
828 put_page(page);
829 goto redo;
830 }
831 /* This means someone else dropped this page from LRU
832 * So, it will be freed or putback to LRU again. There is
833 * nothing to do here.
834 */
835 }
836
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700837 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700838 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700839 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700840 count_vm_event(UNEVICTABLE_PGCULLED);
841
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700842 put_page(page); /* drop ref from isolate */
843}
844
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800845enum page_references {
846 PAGEREF_RECLAIM,
847 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800848 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800849 PAGEREF_ACTIVATE,
850};
851
852static enum page_references page_check_references(struct page *page,
853 struct scan_control *sc)
854{
Johannes Weiner645747462010-03-05 13:42:22 -0800855 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800856 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800857
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700858 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
859 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800860 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800861
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800862 /*
863 * Mlock lost the isolation race with us. Let try_to_unmap()
864 * move the page to the unevictable list.
865 */
866 if (vm_flags & VM_LOCKED)
867 return PAGEREF_RECLAIM;
868
Johannes Weiner645747462010-03-05 13:42:22 -0800869 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700870 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800871 return PAGEREF_ACTIVATE;
872 /*
873 * All mapped pages start out with page table
874 * references from the instantiating fault, so we need
875 * to look twice if a mapped file page is used more
876 * than once.
877 *
878 * Mark it and spare it for another trip around the
879 * inactive list. Another page table reference will
880 * lead to its activation.
881 *
882 * Note: the mark is set for activated pages as well
883 * so that recently deactivated but used pages are
884 * quickly recovered.
885 */
886 SetPageReferenced(page);
887
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800888 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800889 return PAGEREF_ACTIVATE;
890
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800891 /*
892 * Activate file-backed executable pages after first usage.
893 */
894 if (vm_flags & VM_EXEC)
895 return PAGEREF_ACTIVATE;
896
Johannes Weiner645747462010-03-05 13:42:22 -0800897 return PAGEREF_KEEP;
898 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800899
900 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700901 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800902 return PAGEREF_RECLAIM_CLEAN;
903
904 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800905}
906
Mel Gormane2be15f2013-07-03 15:01:57 -0700907/* Check if a page is dirty or under writeback */
908static void page_check_dirty_writeback(struct page *page,
909 bool *dirty, bool *writeback)
910{
Mel Gormanb4597222013-07-03 15:02:05 -0700911 struct address_space *mapping;
912
Mel Gormane2be15f2013-07-03 15:01:57 -0700913 /*
914 * Anonymous pages are not handled by flushers and must be written
915 * from reclaim context. Do not stall reclaim based on them
916 */
917 if (!page_is_file_cache(page)) {
918 *dirty = false;
919 *writeback = false;
920 return;
921 }
922
923 /* By default assume that the page flags are accurate */
924 *dirty = PageDirty(page);
925 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700926
927 /* Verify dirty/writeback state if the filesystem supports it */
928 if (!page_has_private(page))
929 return;
930
931 mapping = page_mapping(page);
932 if (mapping && mapping->a_ops->is_dirty_writeback)
933 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700934}
935
Nick Piggine2867812008-07-25 19:45:30 -0700936/*
Andrew Morton1742f192006-03-22 00:08:21 -0800937 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 */
Andrew Morton1742f192006-03-22 00:08:21 -0800939static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700940 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700941 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700942 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700943 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700944 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700945 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700946 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700947 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700948 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949{
950 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700951 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700953 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700954 unsigned long nr_dirty = 0;
955 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800956 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700957 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700958 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959
960 cond_resched();
961
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 while (!list_empty(page_list)) {
963 struct address_space *mapping;
964 struct page *page;
965 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700966 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700967 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800968 bool lazyfree = false;
969 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
971 cond_resched();
972
973 page = lru_to_page(page_list);
974 list_del(&page->lru);
975
Nick Piggin529ae9a2008-08-02 12:01:03 +0200976 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 goto keep;
978
Sasha Levin309381fea2014-01-23 15:52:54 -0800979 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
981 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800982
Hugh Dickins39b5f292012-10-08 16:33:18 -0700983 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700984 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700985
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700986 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800987 goto keep_locked;
988
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 /* Double the slab pressure for mapped and swapcache pages */
990 if (page_mapped(page) || PageSwapCache(page))
991 sc->nr_scanned++;
992
Andy Whitcroftc661b072007-08-22 14:01:26 -0700993 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
994 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
995
Mel Gorman283aba92013-07-03 15:01:51 -0700996 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700997 * The number of dirty pages determines if a zone is marked
998 * reclaim_congested which affects wait_iff_congested. kswapd
999 * will stall and start writing pages if the tail of the LRU
1000 * is all dirty unqueued pages.
1001 */
1002 page_check_dirty_writeback(page, &dirty, &writeback);
1003 if (dirty || writeback)
1004 nr_dirty++;
1005
1006 if (dirty && !writeback)
1007 nr_unqueued_dirty++;
1008
Mel Gormand04e8ac2013-07-03 15:02:03 -07001009 /*
1010 * Treat this page as congested if the underlying BDI is or if
1011 * pages are cycling through the LRU so quickly that the
1012 * pages marked for immediate reclaim are making it to the
1013 * end of the LRU a second time.
1014 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001015 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001016 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001017 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001018 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001019 nr_congested++;
1020
1021 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001022 * If a page at the tail of the LRU is under writeback, there
1023 * are three cases to consider.
1024 *
1025 * 1) If reclaim is encountering an excessive number of pages
1026 * under writeback and this page is both under writeback and
1027 * PageReclaim then it indicates that pages are being queued
1028 * for IO but are being recycled through the LRU before the
1029 * IO can complete. Waiting on the page itself risks an
1030 * indefinite stall if it is impossible to writeback the
1031 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001032 * note that the LRU is being scanned too quickly and the
1033 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001034 *
Tejun Heo97c93412015-05-22 18:23:36 -04001035 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001036 * not marked for immediate reclaim, or the caller does not
1037 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1038 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001039 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001040 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001041 * Require may_enter_fs because we would wait on fs, which
1042 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001043 * enter reclaim, and deadlock if it waits on a page for
1044 * which it is needed to do the write (loop masks off
1045 * __GFP_IO|__GFP_FS for this reason); but more thought
1046 * would probably show more reasons.
1047 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001048 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001049 * PageReclaim. memcg does not have any dirty pages
1050 * throttling so we could easily OOM just because too many
1051 * pages are in writeback and there is nothing else to
1052 * reclaim. Wait for the writeback to complete.
1053 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001054 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001055 /* Case 1 above */
1056 if (current_is_kswapd() &&
1057 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001058 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001059 nr_immediate++;
1060 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001061
1062 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001063 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001064 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001065 /*
1066 * This is slightly racy - end_page_writeback()
1067 * might have just cleared PageReclaim, then
1068 * setting PageReclaim here end up interpreted
1069 * as PageReadahead - but that does not matter
1070 * enough to care. What we do want is for this
1071 * page to have PageReclaim set next time memcg
1072 * reclaim reaches the tests above, so it will
1073 * then wait_on_page_writeback() to avoid OOM;
1074 * and it's also appropriate in global reclaim.
1075 */
1076 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001077 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001078 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001079
1080 /* Case 3 above */
1081 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001082 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001083 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001084 /* then go back and try same page again */
1085 list_add_tail(&page->lru, page_list);
1086 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001087 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001088 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089
Minchan Kim02c6de82012-10-08 16:31:55 -07001090 if (!force_reclaim)
1091 references = page_check_references(page, sc);
1092
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001093 switch (references) {
1094 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001096 case PAGEREF_KEEP:
1097 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001098 case PAGEREF_RECLAIM:
1099 case PAGEREF_RECLAIM_CLEAN:
1100 ; /* try to reclaim the page below */
1101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 /*
1104 * Anonymous process memory has backing store?
1105 * Try to allocate it some swap space here.
1106 */
Nick Pigginb291f002008-10-18 20:26:44 -07001107 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001108 if (!(sc->gfp_mask & __GFP_IO))
1109 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001110 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001112 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001113 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
Mel Gormane2be15f2013-07-03 15:01:57 -07001115 /* Adding to swap updated mapping */
1116 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001117 } else if (unlikely(PageTransHuge(page))) {
1118 /* Split file THP */
1119 if (split_huge_page_to_list(page, page_list))
1120 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001123 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1124
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 /*
1126 * The page is mapped into the page tables of one or more
1127 * processes. Try to unmap it here.
1128 */
1129 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001130 switch (ret = try_to_unmap(page, lazyfree ?
1131 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1132 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 case SWAP_FAIL:
1134 goto activate_locked;
1135 case SWAP_AGAIN:
1136 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001137 case SWAP_MLOCK:
1138 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001139 case SWAP_LZFREE:
1140 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 case SWAP_SUCCESS:
1142 ; /* try to free the page below */
1143 }
1144 }
1145
1146 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001147 /*
1148 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001149 * avoid risk of stack overflow but only writeback
1150 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001151 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001152 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001153 (!current_is_kswapd() ||
Mel Gorman599d0c92016-07-28 15:45:31 -07001154 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001155 /*
1156 * Immediately reclaim when written back.
1157 * Similar in principal to deactivate_page()
1158 * except we already have the page isolated
1159 * and know it's dirty
1160 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001161 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001162 SetPageReclaim(page);
1163
Mel Gormanee728862011-10-31 17:07:38 -07001164 goto keep_locked;
1165 }
1166
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001167 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001169 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001171 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 goto keep_locked;
1173
Mel Gormand950c942015-09-04 15:47:35 -07001174 /*
1175 * Page is dirty. Flush the TLB if a writable entry
1176 * potentially exists to avoid CPU writes after IO
1177 * starts and then write it out here.
1178 */
1179 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001180 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 case PAGE_KEEP:
1182 goto keep_locked;
1183 case PAGE_ACTIVATE:
1184 goto activate_locked;
1185 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001186 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001187 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001188 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001190
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 /*
1192 * A synchronous write - probably a ramdisk. Go
1193 * ahead and try to reclaim the page.
1194 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001195 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 goto keep;
1197 if (PageDirty(page) || PageWriteback(page))
1198 goto keep_locked;
1199 mapping = page_mapping(page);
1200 case PAGE_CLEAN:
1201 ; /* try to free the page below */
1202 }
1203 }
1204
1205 /*
1206 * If the page has buffers, try to free the buffer mappings
1207 * associated with this page. If we succeed we try to free
1208 * the page as well.
1209 *
1210 * We do this even if the page is PageDirty().
1211 * try_to_release_page() does not perform I/O, but it is
1212 * possible for a page to have PageDirty set, but it is actually
1213 * clean (all its buffers are clean). This happens if the
1214 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001215 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 * try_to_release_page() will discover that cleanness and will
1217 * drop the buffers and mark the page clean - it can be freed.
1218 *
1219 * Rarely, pages can have buffers and no ->mapping. These are
1220 * the pages which were not successfully invalidated in
1221 * truncate_complete_page(). We try to drop those buffers here
1222 * and if that worked, and the page is no longer mapped into
1223 * process address space (page_count == 1) it can be freed.
1224 * Otherwise, leave the page on the LRU so it is swappable.
1225 */
David Howells266cf652009-04-03 16:42:36 +01001226 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 if (!try_to_release_page(page, sc->gfp_mask))
1228 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001229 if (!mapping && page_count(page) == 1) {
1230 unlock_page(page);
1231 if (put_page_testzero(page))
1232 goto free_it;
1233 else {
1234 /*
1235 * rare race with speculative reference.
1236 * the speculative reference will free
1237 * this page shortly, so we may
1238 * increment nr_reclaimed here (and
1239 * leave it off the LRU).
1240 */
1241 nr_reclaimed++;
1242 continue;
1243 }
1244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 }
1246
Minchan Kim854e9ed2016-01-15 16:54:53 -08001247lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001248 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001249 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250
Nick Piggina978d6f2008-10-18 20:26:58 -07001251 /*
1252 * At this point, we have no other references and there is
1253 * no way to pick any more up (removed from LRU, removed
1254 * from pagecache). Can use non-atomic bitops now (and
1255 * we obviously don't have to worry about waking up a process
1256 * waiting on the page lock, because there are no references.
1257 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001258 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001259free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001260 if (ret == SWAP_LZFREE)
1261 count_vm_event(PGLAZYFREED);
1262
Andrew Morton05ff5132006-03-22 00:08:20 -08001263 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001264
1265 /*
1266 * Is there need to periodically free_page_list? It would
1267 * appear not as the counts should be low
1268 */
1269 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 continue;
1271
Nick Pigginb291f002008-10-18 20:26:44 -07001272cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001273 if (PageSwapCache(page))
1274 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001275 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001276 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001277 continue;
1278
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001280 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001281 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001282 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001283 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 SetPageActive(page);
1285 pgactivate++;
1286keep_locked:
1287 unlock_page(page);
1288keep:
1289 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001290 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001292
Johannes Weiner747db952014-08-08 14:19:24 -07001293 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001294 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001295 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001296
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001298 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001299
Mel Gorman8e950282013-07-03 15:02:02 -07001300 *ret_nr_dirty += nr_dirty;
1301 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001302 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001303 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001304 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001305 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306}
1307
Minchan Kim02c6de82012-10-08 16:31:55 -07001308unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1309 struct list_head *page_list)
1310{
1311 struct scan_control sc = {
1312 .gfp_mask = GFP_KERNEL,
1313 .priority = DEF_PRIORITY,
1314 .may_unmap = 1,
1315 };
Mel Gorman8e950282013-07-03 15:02:02 -07001316 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001317 struct page *page, *next;
1318 LIST_HEAD(clean_pages);
1319
1320 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001321 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001322 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001323 ClearPageActive(page);
1324 list_move(&page->lru, &clean_pages);
1325 }
1326 }
1327
Mel Gorman599d0c92016-07-28 15:45:31 -07001328 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001329 TTU_UNMAP|TTU_IGNORE_ACCESS,
1330 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001331 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001332 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001333 return ret;
1334}
1335
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001336/*
1337 * Attempt to remove the specified page from its LRU. Only take this page
1338 * if it is of the appropriate PageActive status. Pages which are being
1339 * freed elsewhere are also ignored.
1340 *
1341 * page: page to consider
1342 * mode: one of the LRU isolation modes defined above
1343 *
1344 * returns 0 on success, -ve errno on failure.
1345 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001346int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001347{
1348 int ret = -EINVAL;
1349
1350 /* Only take pages on the LRU. */
1351 if (!PageLRU(page))
1352 return ret;
1353
Minchan Kime46a2872012-10-08 16:33:48 -07001354 /* Compaction should not handle unevictable pages but CMA can do so */
1355 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001356 return ret;
1357
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001358 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001359
Mel Gormanc8244932012-01-12 17:19:38 -08001360 /*
1361 * To minimise LRU disruption, the caller can indicate that it only
1362 * wants to isolate pages it will be able to operate on without
1363 * blocking - clean pages for the most part.
1364 *
1365 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1366 * is used by reclaim when it is cannot write to backing storage
1367 *
1368 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1369 * that it is possible to migrate without blocking
1370 */
1371 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1372 /* All the caller can do on PageWriteback is block */
1373 if (PageWriteback(page))
1374 return ret;
1375
1376 if (PageDirty(page)) {
1377 struct address_space *mapping;
1378
1379 /* ISOLATE_CLEAN means only clean pages */
1380 if (mode & ISOLATE_CLEAN)
1381 return ret;
1382
1383 /*
1384 * Only pages without mappings or that have a
1385 * ->migratepage callback are possible to migrate
1386 * without blocking
1387 */
1388 mapping = page_mapping(page);
1389 if (mapping && !mapping->a_ops->migratepage)
1390 return ret;
1391 }
1392 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001393
Minchan Kimf80c0672011-10-31 17:06:55 -07001394 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1395 return ret;
1396
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001397 if (likely(get_page_unless_zero(page))) {
1398 /*
1399 * Be careful not to clear PageLRU until after we're
1400 * sure the page is not being freed elsewhere -- the
1401 * page release code relies on it.
1402 */
1403 ClearPageLRU(page);
1404 ret = 0;
1405 }
1406
1407 return ret;
1408}
1409
Mel Gorman7ee36a12016-07-28 15:47:17 -07001410
1411/*
1412 * Update LRU sizes after isolating pages. The LRU size updates must
1413 * be complete before mem_cgroup_update_lru_size due to a santity check.
1414 */
1415static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hocko07fc9572017-01-10 16:58:04 -08001416 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001417{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001418 int zid;
1419
Mel Gorman7ee36a12016-07-28 15:47:17 -07001420 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1421 if (!nr_zone_taken[zid])
1422 continue;
1423
1424 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hocko07fc9572017-01-10 16:58:04 -08001425#ifdef CONFIG_MEMCG
1426 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1427#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001428 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001429
Mel Gorman7ee36a12016-07-28 15:47:17 -07001430}
1431
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001432/*
Mel Gormana52633d2016-07-28 15:45:28 -07001433 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 * shrink the lists perform better by taking out a batch of pages
1435 * and working on them outside the LRU lock.
1436 *
1437 * For pagecache intensive workloads, this function is the hottest
1438 * spot in the kernel (apart from copy_*_user functions).
1439 *
1440 * Appropriate locks must be held before calling this function.
1441 *
1442 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001443 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001445 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001446 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001447 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001448 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 *
1450 * returns how many pages were moved onto *@dst.
1451 */
Andrew Morton69e05942006-03-22 00:08:19 -08001452static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001453 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001454 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001455 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001457 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001458 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001459 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001460 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Mel Gorman599d0c92016-07-28 15:45:31 -07001461 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001462 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001464 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Mel Gormand7f05522016-07-28 15:47:37 -07001465 !list_empty(src);) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001466 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 page = lru_to_page(src);
1469 prefetchw_prev_lru_page(page, src, flags);
1470
Sasha Levin309381fea2014-01-23 15:52:54 -08001471 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001472
Mel Gormanb2e18752016-07-28 15:45:37 -07001473 if (page_zonenum(page) > sc->reclaim_idx) {
1474 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001475 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001476 continue;
1477 }
1478
Mel Gormand7f05522016-07-28 15:47:37 -07001479 /*
1480 * Account for scanned and skipped separetly to avoid the pgdat
1481 * being prematurely marked unreclaimable by pgdat_reclaimable.
1482 */
1483 scan++;
1484
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001485 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001486 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001487 nr_pages = hpage_nr_pages(page);
1488 nr_taken += nr_pages;
1489 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001490 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001491 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001492
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001493 case -EBUSY:
1494 /* else it is being freed elsewhere */
1495 list_move(&page->lru, src);
1496 continue;
1497
1498 default:
1499 BUG();
1500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 }
1502
Mel Gormanb2e18752016-07-28 15:45:37 -07001503 /*
1504 * Splice any skipped pages to the start of the LRU list. Note that
1505 * this disrupts the LRU order when reclaiming for lower zones but
1506 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1507 * scanning would soon rescan the same pages to skip and put the
1508 * system at risk of premature OOM.
1509 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001510 if (!list_empty(&pages_skipped)) {
1511 int zid;
Mel Gormand7f05522016-07-28 15:47:37 -07001512 unsigned long total_skipped = 0;
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001513
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001514 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1515 if (!nr_skipped[zid])
1516 continue;
1517
1518 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Mel Gormand7f05522016-07-28 15:47:37 -07001519 total_skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001520 }
Mel Gormand7f05522016-07-28 15:47:37 -07001521
1522 /*
1523 * Account skipped pages as a partial scan as the pgdat may be
1524 * close to unreclaimable. If the LRU list is empty, account
1525 * skipped pages as a full scan.
1526 */
1527 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1528
1529 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001530 }
Hugh Dickinsf6260122012-01-12 17:20:06 -08001531 *nr_scanned = scan;
Mel Gormane5146b12016-07-28 15:46:47 -07001532 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
Hugh Dickins75b00af2012-05-29 15:07:09 -07001533 nr_taken, mode, is_file_lru(lru));
Michal Hocko07fc9572017-01-10 16:58:04 -08001534 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 return nr_taken;
1536}
1537
Nick Piggin62695a82008-10-18 20:26:09 -07001538/**
1539 * isolate_lru_page - tries to isolate a page from its LRU list
1540 * @page: page to isolate from its LRU list
1541 *
1542 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1543 * vmstat statistic corresponding to whatever LRU list the page was on.
1544 *
1545 * Returns 0 if the page was removed from an LRU list.
1546 * Returns -EBUSY if the page was not on an LRU list.
1547 *
1548 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001549 * the active list, it will have PageActive set. If it was found on
1550 * the unevictable list, it will have the PageUnevictable bit set. That flag
1551 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001552 *
1553 * The vmstat statistic corresponding to the list on which the page was
1554 * found will be decremented.
1555 *
1556 * Restrictions:
1557 * (1) Must be called with an elevated refcount on the page. This is a
1558 * fundamentnal difference from isolate_lru_pages (which is called
1559 * without a stable reference).
1560 * (2) the lru_lock must not be held.
1561 * (3) interrupts must be enabled.
1562 */
1563int isolate_lru_page(struct page *page)
1564{
1565 int ret = -EBUSY;
1566
Sasha Levin309381fea2014-01-23 15:52:54 -08001567 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001568 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001569
Nick Piggin62695a82008-10-18 20:26:09 -07001570 if (PageLRU(page)) {
1571 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001572 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001573
Mel Gormana52633d2016-07-28 15:45:28 -07001574 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001575 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001576 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001577 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001578 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001579 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001580 del_page_from_lru_list(page, lruvec, lru);
1581 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001582 }
Mel Gormana52633d2016-07-28 15:45:28 -07001583 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001584 }
1585 return ret;
1586}
1587
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001588/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001589 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1590 * then get resheduled. When there are massive number of tasks doing page
1591 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1592 * the LRU list will go small and be scanned faster than necessary, leading to
1593 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001594 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001595static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001596 struct scan_control *sc)
1597{
1598 unsigned long inactive, isolated;
1599
1600 if (current_is_kswapd())
1601 return 0;
1602
Tejun Heo97c93412015-05-22 18:23:36 -04001603 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001604 return 0;
1605
1606 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001607 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1608 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001609 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001610 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1611 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001612 }
1613
Fengguang Wu3cf23842012-12-18 14:23:31 -08001614 /*
1615 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1616 * won't get blocked by normal direct-reclaimers, forming a circular
1617 * deadlock.
1618 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001619 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001620 inactive >>= 3;
1621
Rik van Riel35cd7812009-09-21 17:01:38 -07001622 return isolated > inactive;
1623}
1624
Mel Gorman66635622010-08-09 17:19:30 -07001625static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001626putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001627{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001628 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001629 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001630 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001631
Mel Gorman66635622010-08-09 17:19:30 -07001632 /*
1633 * Put back any unfreeable pages.
1634 */
Mel Gorman66635622010-08-09 17:19:30 -07001635 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001636 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001637 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001638
Sasha Levin309381fea2014-01-23 15:52:54 -08001639 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001640 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001641 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001642 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001643 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001644 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001645 continue;
1646 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001647
Mel Gorman599d0c92016-07-28 15:45:31 -07001648 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001649
Linus Torvalds7a608572011-01-17 14:42:19 -08001650 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001651 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001652 add_page_to_lru_list(page, lruvec, lru);
1653
Mel Gorman66635622010-08-09 17:19:30 -07001654 if (is_active_lru(lru)) {
1655 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001656 int numpages = hpage_nr_pages(page);
1657 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001658 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001659 if (put_page_testzero(page)) {
1660 __ClearPageLRU(page);
1661 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001662 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001663
1664 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001665 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001666 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001667 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001668 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001669 } else
1670 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001671 }
1672 }
Mel Gorman66635622010-08-09 17:19:30 -07001673
Hugh Dickins3f797682012-01-12 17:20:07 -08001674 /*
1675 * To save our caller's stack, now use input list for pages to free.
1676 */
1677 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001678}
1679
1680/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001681 * If a kernel thread (such as nfsd for loop-back mounts) services
1682 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1683 * In that case we should only throttle if the backing device it is
1684 * writing to is congested. In other cases it is safe to throttle.
1685 */
1686static int current_may_throttle(void)
1687{
1688 return !(current->flags & PF_LESS_THROTTLE) ||
1689 current->backing_dev_info == NULL ||
1690 bdi_write_congested(current->backing_dev_info);
1691}
1692
Minchan Kim91dcade2016-07-28 15:47:40 -07001693static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1694 struct scan_control *sc, enum lru_list lru)
1695{
1696 int zid;
1697 struct zone *zone;
1698 int file = is_file_lru(lru);
1699 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1700
1701 if (!global_reclaim(sc))
1702 return true;
1703
1704 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1705 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001706 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001707 continue;
1708
1709 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1710 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1711 return true;
1712 }
1713
1714 return false;
1715}
1716
NeilBrown399ba0b2014-06-04 16:07:42 -07001717/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001718 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001719 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 */
Mel Gorman66635622010-08-09 17:19:30 -07001721static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001722shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001723 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724{
1725 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001726 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001727 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001728 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001729 unsigned long nr_dirty = 0;
1730 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001731 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001732 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001733 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001734 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001735 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001736 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001737 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001738
Minchan Kim91dcade2016-07-28 15:47:40 -07001739 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1740 return 0;
1741
Mel Gorman599d0c92016-07-28 15:45:31 -07001742 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001743 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001744
1745 /* We are about to die and free our memory. Return now. */
1746 if (fatal_signal_pending(current))
1747 return SWAP_CLUSTER_MAX;
1748 }
1749
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001751
1752 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001753 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001754 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001755 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001756
Mel Gorman599d0c92016-07-28 15:45:31 -07001757 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001759 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1760 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001761
Mel Gorman599d0c92016-07-28 15:45:31 -07001762 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001763 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001764
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001765 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001766 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001767 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001768 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001769 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001770 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001771 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001772 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001773
Hillf Dantond563c052012-03-21 16:34:02 -07001774 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001775 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001776
Mel Gorman599d0c92016-07-28 15:45:31 -07001777 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001778 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1779 &nr_writeback, &nr_immediate,
1780 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001781
Mel Gorman599d0c92016-07-28 15:45:31 -07001782 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001783
Ying Han904249a2012-04-25 16:01:48 -07001784 if (global_reclaim(sc)) {
1785 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001786 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001787 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001788 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001789 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001790
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001791 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001792
Mel Gorman599d0c92016-07-28 15:45:31 -07001793 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001794
Mel Gorman599d0c92016-07-28 15:45:31 -07001795 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001796
Johannes Weiner747db952014-08-08 14:19:24 -07001797 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001798 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001799
Mel Gorman92df3a72011-10-31 17:07:56 -07001800 /*
1801 * If reclaim is isolating dirty pages under writeback, it implies
1802 * that the long-lived page allocation rate is exceeding the page
1803 * laundering rate. Either the global limits are not being effective
1804 * at throttling processes due to the page distribution throughout
1805 * zones or there is heavy usage of a slow backing device. The
1806 * only option is to throttle from reclaim context which is not ideal
1807 * as there is no guarantee the dirtying process is throttled in the
1808 * same way balance_dirty_pages() manages.
1809 *
Mel Gorman8e950282013-07-03 15:02:02 -07001810 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1811 * of pages under pages flagged for immediate reclaim and stall if any
1812 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001813 */
Mel Gorman918fc712013-07-08 16:00:25 -07001814 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001815 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001816
Mel Gormand43006d2013-07-03 15:01:50 -07001817 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001818 * Legacy memcg will stall in page writeback so avoid forcibly
1819 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001820 */
Tejun Heo97c93412015-05-22 18:23:36 -04001821 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001822 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001823 * Tag a zone as congested if all the dirty pages scanned were
1824 * backed by a congested BDI and wait_iff_congested will stall.
1825 */
1826 if (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001827 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001828
1829 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001830 * If dirty pages are scanned that are not queued for IO, it
1831 * implies that flushers are not keeping up. In this case, flag
Mel Gorman599d0c92016-07-28 15:45:31 -07001832 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
Johannes Weiner57054652014-10-09 15:28:17 -07001833 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001834 */
1835 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001836 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001837
1838 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001839 * If kswapd scans pages marked marked for immediate
1840 * reclaim and under writeback (nr_immediate), it implies
1841 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001842 * they are written so also forcibly stall.
1843 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001844 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001845 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001846 }
Mel Gormand43006d2013-07-03 15:01:50 -07001847
Mel Gorman8e950282013-07-03 15:02:02 -07001848 /*
1849 * Stall direct reclaim for IO completions if underlying BDIs or zone
1850 * is congested. Allow kswapd to continue until it starts encountering
1851 * unqueued dirty pages or cycling through the LRU too quickly.
1852 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001853 if (!sc->hibernation_mode && !current_is_kswapd() &&
1854 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001855 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001856
Mel Gorman599d0c92016-07-28 15:45:31 -07001857 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1858 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001859 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001860 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861}
1862
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001863/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 * This moves pages from the active list to the inactive list.
1865 *
1866 * We move them the other way if the page is referenced by one or more
1867 * processes, from rmap.
1868 *
1869 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001870 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001872 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 * this, so instead we remove the pages from the LRU while processing them.
1874 * It is safe to rely on PG_active against the non-LRU pages in here because
1875 * nobody will play with that bit on a non-LRU page.
1876 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001877 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 * But we had to alter page->flags anyway.
1879 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001880
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001881static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001882 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001883 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001884 enum lru_list lru)
1885{
Mel Gorman599d0c92016-07-28 15:45:31 -07001886 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001887 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001888 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001889 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001890
Wu Fengguang3eb41402009-06-16 15:33:13 -07001891 while (!list_empty(list)) {
1892 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001893 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001894
Sasha Levin309381fea2014-01-23 15:52:54 -08001895 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001896 SetPageLRU(page);
1897
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001898 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001899 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001900 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001901 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001902
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001903 if (put_page_testzero(page)) {
1904 __ClearPageLRU(page);
1905 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001906 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001907
1908 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001909 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001910 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001911 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001912 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001913 } else
1914 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001915 }
1916 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001917
Wu Fengguang3eb41402009-06-16 15:33:13 -07001918 if (!is_active_lru(lru))
1919 __count_vm_events(PGDEACTIVATE, pgmoved);
1920}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001921
Hugh Dickinsf6260122012-01-12 17:20:06 -08001922static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001923 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001924 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001925 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001927 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001928 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001929 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001931 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001932 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001934 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001935 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001936 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001937 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001938 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
1940 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001941
1942 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001943 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001944 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001945 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001946
Mel Gorman599d0c92016-07-28 15:45:31 -07001947 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001948
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001949 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1950 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001951
Mel Gorman599d0c92016-07-28 15:45:31 -07001952 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001953 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001954
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001955 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001956 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1957 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001958
Mel Gorman599d0c92016-07-28 15:45:31 -07001959 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 while (!list_empty(&l_hold)) {
1962 cond_resched();
1963 page = lru_to_page(&l_hold);
1964 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001965
Hugh Dickins39b5f292012-10-08 16:33:18 -07001966 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001967 putback_lru_page(page);
1968 continue;
1969 }
1970
Mel Gormancc715d92012-03-21 16:34:00 -07001971 if (unlikely(buffer_heads_over_limit)) {
1972 if (page_has_private(page) && trylock_page(page)) {
1973 if (page_has_private(page))
1974 try_to_release_page(page, 0);
1975 unlock_page(page);
1976 }
1977 }
1978
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001979 if (page_referenced(page, 0, sc->target_mem_cgroup,
1980 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001981 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001982 /*
1983 * Identify referenced, file-backed active pages and
1984 * give them one more trip around the active list. So
1985 * that executable code get better chances to stay in
1986 * memory under moderate memory pressure. Anon pages
1987 * are not likely to be evicted by use-once streaming
1988 * IO, plus JVM can create lots of anon VM_EXEC pages,
1989 * so we ignore them here.
1990 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001991 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001992 list_add(&page->lru, &l_active);
1993 continue;
1994 }
1995 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001996
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001997 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 list_add(&page->lru, &l_inactive);
1999 }
2000
Andrew Mortonb5557492009-01-06 14:40:13 -08002001 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002002 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002003 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002004 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002005 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002006 * Count referenced pages from currently used mappings as rotated,
2007 * even though only some of them are actually re-activated. This
2008 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002009 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002010 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002011 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002012
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002013 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2014 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07002015 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2016 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002017
Johannes Weiner747db952014-08-08 14:19:24 -07002018 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07002019 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020}
2021
Rik van Riel59dc76b2016-05-20 16:56:31 -07002022/*
2023 * The inactive anon list should be small enough that the VM never has
2024 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002025 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002026 * The inactive file list should be small enough to leave most memory
2027 * to the established workingset on the scan-resistant active list,
2028 * but large enough to avoid thrashing the aggregate readahead window.
2029 *
2030 * Both inactive lists should also be large enough that each inactive
2031 * page has a chance to be referenced again before it is reclaimed.
2032 *
2033 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2034 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2035 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2036 *
2037 * total target max
2038 * memory ratio inactive
2039 * -------------------------------------
2040 * 10MB 1 5MB
2041 * 100MB 1 50MB
2042 * 1GB 3 250MB
2043 * 10GB 10 0.9GB
2044 * 100GB 31 3GB
2045 * 1TB 101 10GB
2046 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002047 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002048static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2049 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002050{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002051 unsigned long inactive_ratio;
Michal Hocko71053132017-02-22 15:45:58 -08002052 unsigned long inactive, active;
2053 enum lru_list inactive_lru = file * LRU_FILE;
2054 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002055 unsigned long gb;
2056
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002057 /*
2058 * If we don't have swap space, anonymous page deactivation
2059 * is pointless.
2060 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002061 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002062 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002063
Michal Hocko71053132017-02-22 15:45:58 -08002064 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2065 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002066
Rik van Riel59dc76b2016-05-20 16:56:31 -07002067 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2068 if (gb)
2069 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002070 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002071 inactive_ratio = 1;
2072
2073 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002074}
2075
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002076static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002077 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002078{
Rik van Rielb39415b2009-12-14 17:59:48 -08002079 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002080 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002081 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002082 return 0;
2083 }
2084
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002085 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002086}
2087
Johannes Weiner9a265112013-02-22 16:32:17 -08002088enum scan_balance {
2089 SCAN_EQUAL,
2090 SCAN_FRACT,
2091 SCAN_ANON,
2092 SCAN_FILE,
2093};
2094
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002096 * Determine how aggressively the anon and file LRU lists should be
2097 * scanned. The relative value of each set of LRU lists is determined
2098 * by looking at the fraction of the pages scanned we did rotate back
2099 * onto the active list instead of evict.
2100 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002101 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2102 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002103 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002104static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002105 struct scan_control *sc, unsigned long *nr,
2106 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002107{
Vladimir Davydov33377672016-01-20 15:02:59 -08002108 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002109 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002110 u64 fraction[2];
2111 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002112 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002113 unsigned long anon_prio, file_prio;
2114 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002115 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002116 bool force_scan = false;
2117 unsigned long ap, fp;
2118 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002119 bool some_scanned;
2120 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07002121
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002122 /*
2123 * If the zone or memcg is small, nr[l] can be 0. This
2124 * results in no scanning on this priority and a potential
2125 * priority drop. Global direct reclaim can go to the next
2126 * zone and tends to have no problems. Global kswapd is for
2127 * zone balancing and it needs to scan a minimum amount. When
2128 * reclaiming for a memcg, a priority drop can cause high
2129 * latencies, so it's better to scan a minimum amount there as
2130 * well.
2131 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002132 if (current_is_kswapd()) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002133 if (!pgdat_reclaimable(pgdat))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002134 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002135 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002136 force_scan = true;
2137 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002138 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002139 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002140
2141 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002142 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002143 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002144 goto out;
2145 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002146
Johannes Weiner10316b32013-02-22 16:32:14 -08002147 /*
2148 * Global reclaim will swap to prevent OOM even with no
2149 * swappiness, but memcg users want to use this knob to
2150 * disable swapping for individual groups completely when
2151 * using the memory controller's swap limit feature would be
2152 * too expensive.
2153 */
Johannes Weiner02695172014-08-06 16:06:17 -07002154 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002155 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002156 goto out;
2157 }
2158
2159 /*
2160 * Do not apply any pressure balancing cleverness when the
2161 * system is close to OOM, scan both anon and file equally
2162 * (unless the swappiness setting disagrees with swapping).
2163 */
Johannes Weiner02695172014-08-06 16:06:17 -07002164 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002165 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002166 goto out;
2167 }
2168
Johannes Weiner11d16c22013-02-22 16:32:15 -08002169 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002170 * Prevent the reclaimer from falling into the cache trap: as
2171 * cache pages start out inactive, every cache fault will tip
2172 * the scan balance towards the file LRU. And as the file LRU
2173 * shrinks, so does the window for rotation from references.
2174 * This means we have a runaway feedback loop where a tiny
2175 * thrashing file LRU becomes infinitely more attractive than
2176 * anon pages. Try to detect this based on file LRU size.
2177 */
2178 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002179 unsigned long pgdatfile;
2180 unsigned long pgdatfree;
2181 int z;
2182 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002183
Mel Gorman599d0c92016-07-28 15:45:31 -07002184 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2185 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2186 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002187
Mel Gorman599d0c92016-07-28 15:45:31 -07002188 for (z = 0; z < MAX_NR_ZONES; z++) {
2189 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002190 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002191 continue;
2192
2193 total_high_wmark += high_wmark_pages(zone);
2194 }
2195
2196 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002197 scan_balance = SCAN_ANON;
2198 goto out;
2199 }
2200 }
2201
2202 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002203 * If there is enough inactive page cache, i.e. if the size of the
2204 * inactive list is greater than that of the active list *and* the
2205 * inactive list actually has some pages to scan on this priority, we
2206 * do not reclaim anything from the anonymous working set right now.
2207 * Without the second condition we could end up never scanning an
2208 * lruvec even if it has plenty of old anonymous pages unless the
2209 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002210 */
Patrick Daly72312122017-03-14 18:30:41 -07002211 if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) &&
2212 !inactive_list_is_low(lruvec, true, sc) &&
Michal Hocko521e92b2017-02-22 15:46:01 -08002213 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002214 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002215 goto out;
2216 }
2217
Johannes Weiner9a265112013-02-22 16:32:17 -08002218 scan_balance = SCAN_FRACT;
2219
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002220 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002221 * With swappiness at 100, anonymous and file have the same priority.
2222 * This scanning priority is essentially the inverse of IO cost.
2223 */
Johannes Weiner02695172014-08-06 16:06:17 -07002224 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002225 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002226
2227 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002228 * OK, so we have swap space and a fair amount of page cache
2229 * pages. We use the recently rotated / recently scanned
2230 * ratios to determine how valuable each cache is.
2231 *
2232 * Because workloads change over time (and to avoid overflow)
2233 * we keep these statistics as a floating average, which ends
2234 * up weighing recent references more than old ones.
2235 *
2236 * anon in [0], file in [1]
2237 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002238
Michal Hocko71053132017-02-22 15:45:58 -08002239 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2240 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2241 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2242 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002243
Mel Gorman599d0c92016-07-28 15:45:31 -07002244 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002245 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002246 reclaim_stat->recent_scanned[0] /= 2;
2247 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002248 }
2249
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002250 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002251 reclaim_stat->recent_scanned[1] /= 2;
2252 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002253 }
2254
2255 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002256 * The amount of pressure on anon vs file pages is inversely
2257 * proportional to the fraction of recently scanned pages on
2258 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002259 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002260 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002261 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002262
Satoru Moriyafe350042012-05-29 15:06:47 -07002263 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002264 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002265 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002266
Shaohua Li76a33fc2010-05-24 14:32:36 -07002267 fraction[0] = ap;
2268 fraction[1] = fp;
2269 denominator = ap + fp + 1;
2270out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002271 some_scanned = false;
2272 /* Only use force_scan on second pass. */
2273 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002274 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002275 for_each_evictable_lru(lru) {
2276 int file = is_file_lru(lru);
2277 unsigned long size;
2278 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002279
Michal Hocko521e92b2017-02-22 15:46:01 -08002280 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002281 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002282
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002283 if (!scan && pass && force_scan)
2284 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002285
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002286 switch (scan_balance) {
2287 case SCAN_EQUAL:
2288 /* Scan lists relative to size */
2289 break;
2290 case SCAN_FRACT:
2291 /*
2292 * Scan types proportional to swappiness and
2293 * their relative recent reclaim efficiency.
2294 */
2295 scan = div64_u64(scan * fraction[file],
2296 denominator);
2297 break;
2298 case SCAN_FILE:
2299 case SCAN_ANON:
2300 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002301 if ((scan_balance == SCAN_FILE) != file) {
2302 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002303 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002304 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002305 break;
2306 default:
2307 /* Look ma, no brain */
2308 BUG();
2309 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002310
2311 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002312 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002313
Johannes Weiner9a265112013-02-22 16:32:17 -08002314 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002315 * Skip the second pass and don't force_scan,
2316 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002317 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002318 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002319 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002320 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002321}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002322
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
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002358 blk_start_plug(&plug);
2359 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2360 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002361 unsigned long nr_anon, nr_file, percentage;
2362 unsigned long nr_scanned;
2363
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002364 for_each_evictable_lru(lru) {
2365 if (nr[lru]) {
2366 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2367 nr[lru] -= nr_to_scan;
2368
2369 nr_reclaimed += shrink_list(lru, nr_to_scan,
2370 lruvec, sc);
2371 }
2372 }
Mel Gormane82e0562013-07-03 15:01:44 -07002373
Michal Hockobd041732016-12-02 17:26:48 -08002374 cond_resched();
2375
Mel Gormane82e0562013-07-03 15:01:44 -07002376 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);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002440}
2441
Mel Gorman23b9da52012-05-29 15:06:20 -07002442/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002443static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002444{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002445 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002446 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002447 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002448 return true;
2449
2450 return false;
2451}
2452
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002453/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002454 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2455 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2456 * true if more pages should be reclaimed such that when the page allocator
2457 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2458 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002459 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002460static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002461 unsigned long nr_reclaimed,
2462 unsigned long nr_scanned,
2463 struct scan_control *sc)
2464{
2465 unsigned long pages_for_compaction;
2466 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002467 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002468
2469 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002470 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002471 return false;
2472
Mel Gorman28765922011-02-25 14:44:20 -08002473 /* Consider stopping depending on scan and reclaim activity */
2474 if (sc->gfp_mask & __GFP_REPEAT) {
2475 /*
2476 * For __GFP_REPEAT allocations, stop reclaiming if the
2477 * full LRU list has been scanned and we are still failing
2478 * to reclaim pages. This full LRU scan is potentially
2479 * expensive but a __GFP_REPEAT caller really wants to succeed
2480 */
2481 if (!nr_reclaimed && !nr_scanned)
2482 return false;
2483 } else {
2484 /*
2485 * For non-__GFP_REPEAT allocations which can presumably
2486 * fail without consequence, stop if we failed to reclaim
2487 * any pages from the last SWAP_CLUSTER_MAX number of
2488 * pages that were scanned. This will return to the
2489 * caller faster at the risk reclaim/compaction and
2490 * the resulting allocation attempt fails
2491 */
2492 if (!nr_reclaimed)
2493 return false;
2494 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002495
2496 /*
2497 * If we have not reclaimed enough pages for compaction and the
2498 * inactive lists are large enough, continue reclaiming
2499 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002500 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002501 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002502 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002503 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002504 if (sc->nr_reclaimed < pages_for_compaction &&
2505 inactive_lru_pages > pages_for_compaction)
2506 return true;
2507
2508 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002509 for (z = 0; z <= sc->reclaim_idx; z++) {
2510 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002511 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002512 continue;
2513
2514 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002515 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002516 case COMPACT_CONTINUE:
2517 return false;
2518 default:
2519 /* check next zone */
2520 ;
2521 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002522 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002523 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002524}
2525
Mel Gorman970a39a2016-07-28 15:46:35 -07002526static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002527{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002528 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002529 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002530 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002531
Johannes Weiner56600482012-01-12 17:17:59 -08002532 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002533 struct mem_cgroup *root = sc->target_mem_cgroup;
2534 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002535 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002536 .priority = sc->priority,
2537 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002538 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002539 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002540
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002541 nr_reclaimed = sc->nr_reclaimed;
2542 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002543
Andrew Morton694fbc02013-09-24 15:27:37 -07002544 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2545 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002546 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002547 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002548 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002549
Johannes Weiner241994e2015-02-11 15:26:06 -08002550 if (mem_cgroup_low(root, memcg)) {
2551 if (!sc->may_thrash)
2552 continue;
2553 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2554 }
2555
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002556 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002557 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002558
Mel Gormana9dd0a82016-07-28 15:46:02 -07002559 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2560 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002561
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002562 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002563 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002564 memcg, sc->nr_scanned - scanned,
2565 lru_pages);
2566
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002567 /* Record the group's reclaim efficiency */
2568 vmpressure(sc->gfp_mask, memcg, false,
2569 sc->nr_scanned - scanned,
2570 sc->nr_reclaimed - reclaimed);
2571
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002572 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002573 * Direct reclaim and kswapd have to scan all memory
2574 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002575 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002576 *
2577 * Limit reclaim, on the other hand, only cares about
2578 * nr_to_reclaim pages to be reclaimed and it will
2579 * retry with decreasing priority if one round over the
2580 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002581 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002582 if (!global_reclaim(sc) &&
2583 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002584 mem_cgroup_iter_break(root, memcg);
2585 break;
2586 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002587 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002588
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002589 /*
2590 * Shrink the slab caches in the same proportion that
2591 * the eligible LRU pages were scanned.
2592 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002593 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002594 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002595 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002596 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002597
Vinayak Menon57e6f442017-02-19 20:00:31 +11002598 /*
2599 * Record the subtree's reclaim efficiency. The reclaimed
2600 * pages from slab is excluded here because the corresponding
2601 * scanned pages is not accounted. Moreover, freeing a page
2602 * by slab shrinking depends on each slab's object population,
2603 * making the cost model (i.e. scan:free) different from that
2604 * of LRU.
2605 */
2606 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2607 sc->nr_scanned - nr_scanned,
2608 sc->nr_reclaimed - nr_reclaimed);
2609
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002610 if (reclaim_state) {
2611 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2612 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002613 }
2614
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002615 if (sc->nr_reclaimed - nr_reclaimed)
2616 reclaimable = true;
2617
Mel Gormana9dd0a82016-07-28 15:46:02 -07002618 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002619 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002620
2621 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002622}
2623
Vlastimil Babka53853e22014-10-09 15:27:02 -07002624/*
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002625 * Returns true if compaction should go ahead for a costly-order request, or
2626 * the allocation would already succeed without compaction. Return false if we
2627 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002628 */
Mel Gorman4f588332016-07-28 15:46:38 -07002629static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002630{
Mel Gorman31483b62016-07-28 15:45:46 -07002631 unsigned long watermark;
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002632 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002633
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002634 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2635 if (suitable == COMPACT_SUCCESS)
2636 /* Allocation should succeed already. Don't reclaim. */
2637 return true;
2638 if (suitable == COMPACT_SKIPPED)
2639 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002640 return false;
2641
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002642 /*
2643 * Compaction is already possible, but it takes time to run and there
2644 * are potentially other callers using the pages just freed. So proceed
2645 * with reclaim to make a buffer of free pages available to give
2646 * compaction a reasonable chance of completing and allocating the page.
2647 * Note that we won't actually reclaim the whole buffer in one attempt
2648 * as the target watermark in should_continue_reclaim() is lower. But if
2649 * we are already above the high+gap watermark, don't reclaim at all.
2650 */
2651 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2652
2653 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002654}
2655
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656/*
2657 * This is the direct reclaim path, for page-allocating processes. We only
2658 * try to reclaim pages from zones which will satisfy the caller's allocation
2659 * request.
2660 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 * If a zone is deemed to be full of pinned pages then just give it a light
2662 * scan then give up on it.
2663 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002664static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665{
Mel Gormandd1a2392008-04-28 02:12:17 -07002666 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002667 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002668 unsigned long nr_soft_reclaimed;
2669 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002670 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002671 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002672
Mel Gormancc715d92012-03-21 16:34:00 -07002673 /*
2674 * If the number of buffer_heads in the machine exceeds the maximum
2675 * allowed level, force direct reclaim to scan the highmem zone as
2676 * highmem pages could be pinning lowmem pages storing buffer_heads
2677 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002678 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002679 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002680 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002681 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002682 }
Mel Gormancc715d92012-03-21 16:34:00 -07002683
Mel Gormand4debc62010-08-09 17:19:29 -07002684 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002685 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002686 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002687 * Take care memory controller reclaiming has small influence
2688 * to global LRU.
2689 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002690 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002691 if (!cpuset_zone_allowed(zone,
2692 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002693 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002694
Lisa Du6e543d52013-09-11 14:22:36 -07002695 if (sc->priority != DEF_PRIORITY &&
Mel Gorman599d0c92016-07-28 15:45:31 -07002696 !pgdat_reclaimable(zone->zone_pgdat))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002697 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002698
2699 /*
2700 * If we already have plenty of memory free for
2701 * compaction in this zone, don't free any more.
2702 * Even though compaction is invoked for any
2703 * non-zero order, only frequent costly order
2704 * reclamation is disruptive enough to become a
2705 * noticeable problem, like transparent huge
2706 * page allocations.
2707 */
2708 if (IS_ENABLED(CONFIG_COMPACTION) &&
2709 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002710 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002711 sc->compaction_ready = true;
2712 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002713 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002714
Andrew Morton0608f432013-09-24 15:27:41 -07002715 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002716 * Shrink each node in the zonelist once. If the
2717 * zonelist is ordered by zone (not the default) then a
2718 * node may be shrunk multiple times but in that case
2719 * the user prefers lower zones being preserved.
2720 */
2721 if (zone->zone_pgdat == last_pgdat)
2722 continue;
2723
2724 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002725 * This steals pages from memory cgroups over softlimit
2726 * and returns the number of reclaimed pages and
2727 * scanned pages. This works for global memory pressure
2728 * and balancing, not for a memcg's limit.
2729 */
2730 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002731 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002732 sc->order, sc->gfp_mask,
2733 &nr_soft_scanned);
2734 sc->nr_reclaimed += nr_soft_reclaimed;
2735 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002736 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002737 }
Nick Piggin408d8542006-09-25 23:31:27 -07002738
Mel Gorman79dafcd2016-07-28 15:45:53 -07002739 /* See comment about same check for global reclaim above */
2740 if (zone->zone_pgdat == last_pgdat)
2741 continue;
2742 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002743 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 }
Mel Gormane0c23272011-10-31 17:09:33 -07002745
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002746 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002747 * Restore to original mask to avoid the impact on the caller if we
2748 * promoted it to __GFP_HIGHMEM.
2749 */
2750 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002752
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753/*
2754 * This is the main entry point to direct page reclaim.
2755 *
2756 * If a full scan of the inactive list fails to free enough memory then we
2757 * are "out of memory" and something needs to be killed.
2758 *
2759 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2760 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002761 * caller can't do much about. We kick the writeback threads and take explicit
2762 * naps in the hope that some of these pages can be written. But if the
2763 * allocating task holds filesystem locks which prevent writeout this might not
2764 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002765 *
2766 * returns: 0, if no pages reclaimed
2767 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 */
Mel Gormandac1d272008-04-28 02:12:12 -07002769static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002770 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771{
Johannes Weiner241994e2015-02-11 15:26:06 -08002772 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002773 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002774 unsigned long writeback_threshold;
Johannes Weiner241994e2015-02-11 15:26:06 -08002775retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002776 delayacct_freepages_start();
2777
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002778 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002779 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002781 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002782 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2783 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002784 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002785 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002786
Balbir Singh66e17072008-02-07 00:13:56 -08002787 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002788 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002789 break;
2790
2791 if (sc->compaction_ready)
2792 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
2794 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002795 * If we're getting trouble reclaiming, start doing
2796 * writepage even in laptop mode.
2797 */
2798 if (sc->priority < DEF_PRIORITY - 2)
2799 sc->may_writepage = 1;
2800
2801 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 * Try to write back as many pages as we just scanned. This
2803 * tends to cause slow streaming writers to write data to the
2804 * disk smoothly, at the dirtying rate, which is nice. But
2805 * that's undesirable in laptop mode, where we *want* lumpy
2806 * writeout. So in laptop mode, write out the whole world.
2807 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002808 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2809 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002810 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2811 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002812 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002814 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002815
Keika Kobayashi873b4772008-07-25 01:48:52 -07002816 delayacct_freepages_end();
2817
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002818 if (sc->nr_reclaimed)
2819 return sc->nr_reclaimed;
2820
Mel Gorman0cee34f2012-01-12 17:19:49 -08002821 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002822 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002823 return 1;
2824
Johannes Weiner241994e2015-02-11 15:26:06 -08002825 /* Untapped cgroup reserves? Don't OOM, retry. */
2826 if (!sc->may_thrash) {
2827 sc->priority = initial_priority;
2828 sc->may_thrash = 1;
2829 goto retry;
2830 }
2831
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002832 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833}
2834
Mel Gorman55150612012-07-31 16:44:35 -07002835static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2836{
2837 struct zone *zone;
2838 unsigned long pfmemalloc_reserve = 0;
2839 unsigned long free_pages = 0;
2840 int i;
2841 bool wmark_ok;
2842
2843 for (i = 0; i <= ZONE_NORMAL; i++) {
2844 zone = &pgdat->node_zones[i];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002845 if (!managed_zone(zone) ||
Mel Gorman599d0c92016-07-28 15:45:31 -07002846 pgdat_reclaimable_pages(pgdat) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002847 continue;
2848
Mel Gorman55150612012-07-31 16:44:35 -07002849 pfmemalloc_reserve += min_wmark_pages(zone);
2850 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2851 }
2852
Mel Gorman675becc2014-06-04 16:07:35 -07002853 /* If there are no reserves (unexpected config) then do not throttle */
2854 if (!pfmemalloc_reserve)
2855 return true;
2856
Mel Gorman55150612012-07-31 16:44:35 -07002857 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2858
2859 /* kswapd must be awake if processes are being throttled */
2860 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002861 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002862 (enum zone_type)ZONE_NORMAL);
2863 wake_up_interruptible(&pgdat->kswapd_wait);
2864 }
2865
2866 return wmark_ok;
2867}
2868
2869/*
2870 * Throttle direct reclaimers if backing storage is backed by the network
2871 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2872 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002873 * when the low watermark is reached.
2874 *
2875 * Returns true if a fatal signal was delivered during throttling. If this
2876 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002877 */
Mel Gorman50694c22012-11-26 16:29:48 -08002878static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002879 nodemask_t *nodemask)
2880{
Mel Gorman675becc2014-06-04 16:07:35 -07002881 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002882 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002883 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002884
2885 /*
2886 * Kernel threads should not be throttled as they may be indirectly
2887 * responsible for cleaning pages necessary for reclaim to make forward
2888 * progress. kjournald for example may enter direct reclaim while
2889 * committing a transaction where throttling it could forcing other
2890 * processes to block on log_wait_commit().
2891 */
2892 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002893 goto out;
2894
2895 /*
2896 * If a fatal signal is pending, this process should not throttle.
2897 * It should return quickly so it can exit and free its memory
2898 */
2899 if (fatal_signal_pending(current))
2900 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002901
Mel Gorman675becc2014-06-04 16:07:35 -07002902 /*
2903 * Check if the pfmemalloc reserves are ok by finding the first node
2904 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2905 * GFP_KERNEL will be required for allocating network buffers when
2906 * swapping over the network so ZONE_HIGHMEM is unusable.
2907 *
2908 * Throttling is based on the first usable node and throttled processes
2909 * wait on a queue until kswapd makes progress and wakes them. There
2910 * is an affinity then between processes waking up and where reclaim
2911 * progress has been made assuming the process wakes on the same node.
2912 * More importantly, processes running on remote nodes will not compete
2913 * for remote pfmemalloc reserves and processes on different nodes
2914 * should make reasonable progress.
2915 */
2916 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002917 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002918 if (zone_idx(zone) > ZONE_NORMAL)
2919 continue;
2920
2921 /* Throttle based on the first usable node */
2922 pgdat = zone->zone_pgdat;
2923 if (pfmemalloc_watermark_ok(pgdat))
2924 goto out;
2925 break;
2926 }
2927
2928 /* If no zone was usable by the allocation flags then do not throttle */
2929 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002930 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002931
Mel Gorman68243e72012-07-31 16:44:39 -07002932 /* Account for the throttling */
2933 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2934
Mel Gorman55150612012-07-31 16:44:35 -07002935 /*
2936 * If the caller cannot enter the filesystem, it's possible that it
2937 * is due to the caller holding an FS lock or performing a journal
2938 * transaction in the case of a filesystem like ext[3|4]. In this case,
2939 * it is not safe to block on pfmemalloc_wait as kswapd could be
2940 * blocked waiting on the same lock. Instead, throttle for up to a
2941 * second before continuing.
2942 */
2943 if (!(gfp_mask & __GFP_FS)) {
2944 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2945 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002946
2947 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002948 }
2949
2950 /* Throttle until kswapd wakes the process */
2951 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2952 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002953
2954check_pending:
2955 if (fatal_signal_pending(current))
2956 return true;
2957
2958out:
2959 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002960}
2961
Mel Gormandac1d272008-04-28 02:12:12 -07002962unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002963 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002964{
Mel Gorman33906bc2010-08-09 17:19:16 -07002965 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002966 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002967 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002968 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07002969 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07002970 .order = order,
2971 .nodemask = nodemask,
2972 .priority = DEF_PRIORITY,
2973 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002974 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002975 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002976 };
2977
Mel Gorman55150612012-07-31 16:44:35 -07002978 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002979 * Do not enter reclaim if fatal signal was delivered while throttled.
2980 * 1 is returned so that the page allocator does not OOM kill at this
2981 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002982 */
Mel Gorman50694c22012-11-26 16:29:48 -08002983 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002984 return 1;
2985
Mel Gorman33906bc2010-08-09 17:19:16 -07002986 trace_mm_vmscan_direct_reclaim_begin(order,
2987 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07002988 gfp_mask,
2989 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002990
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002991 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002992
2993 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2994
2995 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002996}
2997
Andrew Mortonc255a452012-07-31 16:43:02 -07002998#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002999
Mel Gormana9dd0a82016-07-28 15:46:02 -07003000unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003001 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003002 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003003 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003004{
3005 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003006 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003007 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003008 .may_writepage = !laptop_mode,
3009 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003010 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003011 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003012 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003013 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003014
Balbir Singh4e416952009-09-23 15:56:39 -07003015 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3016 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003017
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003018 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003019 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003020 sc.gfp_mask,
3021 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003022
Balbir Singh4e416952009-09-23 15:56:39 -07003023 /*
3024 * NOTE: Although we can get the priority field, using it
3025 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003026 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003027 * will pick up pages from other mem cgroup's as well. We hack
3028 * the priority and make it zero.
3029 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003030 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003031
3032 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3033
Ying Han0ae5e892011-05-26 16:25:25 -07003034 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003035 return sc.nr_reclaimed;
3036}
3037
Johannes Weiner72835c82012-01-12 17:18:32 -08003038unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003039 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003040 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003041 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003042{
Balbir Singh4e416952009-09-23 15:56:39 -07003043 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003044 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003045 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003046 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003047 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003048 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3049 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003050 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003051 .target_mem_cgroup = memcg,
3052 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003053 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003054 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003055 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003056 };
Balbir Singh66e17072008-02-07 00:13:56 -08003057
Ying Han889976d2011-05-26 16:25:33 -07003058 /*
3059 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3060 * take care of from where we get pages. So the node where we start the
3061 * scan does not need to be the current node.
3062 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003063 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003064
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003065 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003066
3067 trace_mm_vmscan_memcg_reclaim_begin(0,
3068 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003069 sc.gfp_mask,
3070 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003071
Johannes Weiner89a28482016-10-27 17:46:56 -07003072 current->flags |= PF_MEMALLOC;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003073 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weiner89a28482016-10-27 17:46:56 -07003074 current->flags &= ~PF_MEMALLOC;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003075
3076 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3077
3078 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003079}
3080#endif
3081
Mel Gorman1d82de62016-07-28 15:45:43 -07003082static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003083 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003084{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003085 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003086
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003087 if (!total_swap_pages)
3088 return;
3089
3090 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3091 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003092 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003093
Mel Gormanf8d1a312016-07-28 15:47:34 -07003094 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003095 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003096 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003097
3098 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3099 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003100}
3101
Mel Gorman31483b62016-07-28 15:45:46 -07003102static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003103{
Mel Gorman31483b62016-07-28 15:45:46 -07003104 unsigned long mark = high_wmark_pages(zone);
Johannes Weiner60cefed2012-11-29 13:54:23 -08003105
Mel Gorman6256c6b42016-07-28 15:45:56 -07003106 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3107 return false;
3108
3109 /*
3110 * If any eligible zone is balanced then the node is not considered
3111 * to be congested or dirty
3112 */
3113 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3114 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3115
3116 return true;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003117}
3118
Mel Gorman1741c872011-01-13 15:46:21 -08003119/*
Mel Gorman55150612012-07-31 16:44:35 -07003120 * Prepare kswapd for sleeping. This verifies that there are no processes
3121 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3122 *
3123 * Returns true if kswapd is ready to sleep
3124 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003125static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003126{
Mel Gorman1d82de62016-07-28 15:45:43 -07003127 int i;
3128
Mel Gorman55150612012-07-31 16:44:35 -07003129 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003130 * The throttled processes are normally woken up in balance_pgdat() as
3131 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3132 * race between when kswapd checks the watermarks and a process gets
3133 * throttled. There is also a potential race if processes get
3134 * throttled, kswapd wakes, a large process exits thereby balancing the
3135 * zones, which causes kswapd to exit balance_pgdat() before reaching
3136 * the wake up checks. If kswapd is going to sleep, no process should
3137 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3138 * the wake up is premature, processes will wake kswapd and get
3139 * throttled again. The difference from wake ups in balance_pgdat() is
3140 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003141 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003142 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3143 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003144
Mel Gorman1d82de62016-07-28 15:45:43 -07003145 for (i = 0; i <= classzone_idx; i++) {
3146 struct zone *zone = pgdat->node_zones + i;
3147
Mel Gorman6aa303d2016-09-01 16:14:55 -07003148 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003149 continue;
3150
Mel Gorman38087d92016-07-28 15:45:49 -07003151 if (!zone_balanced(zone, order, classzone_idx))
3152 return false;
Mel Gorman1d82de62016-07-28 15:45:43 -07003153 }
3154
Mel Gorman38087d92016-07-28 15:45:49 -07003155 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003156}
3157
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003159 * kswapd shrinks a node of pages that are at or below the highest usable
3160 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003161 *
3162 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003163 * reclaim or if the lack of progress was due to pages under writeback.
3164 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003165 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003166static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003167 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003168{
Mel Gorman1d82de62016-07-28 15:45:43 -07003169 struct zone *zone;
3170 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003171
Mel Gorman1d82de62016-07-28 15:45:43 -07003172 /* Reclaim a number of pages proportional to the number of zones */
3173 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003174 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003175 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003176 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003177 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003178
Mel Gorman1d82de62016-07-28 15:45:43 -07003179 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003180 }
3181
Mel Gorman1d82de62016-07-28 15:45:43 -07003182 /*
3183 * Historically care was taken to put equal pressure on all zones but
3184 * now pressure is applied based on node LRU order.
3185 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003186 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003187
3188 /*
3189 * Fragmentation may mean that the system cannot be rebalanced for
3190 * high-order allocations. If twice the allocation size has been
3191 * reclaimed then recheck watermarks only at order-0 to prevent
3192 * excessive reclaim. Assume that a process requested a high-order
3193 * can direct reclaim/compact.
3194 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003195 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003196 sc->order = 0;
3197
Mel Gormanb8e83b92013-07-03 15:01:45 -07003198 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003199}
3200
3201/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003202 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3203 * that are eligible for use by the caller until at least one zone is
3204 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003206 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 *
3208 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003209 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003210 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3211 * or lower is eligible for reclaim until at least one usable zone is
3212 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003214static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003217 unsigned long nr_soft_reclaimed;
3218 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003219 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003220 struct scan_control sc = {
3221 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003222 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003223 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003224 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003225 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003226 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003227 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003228 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003230 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003231 bool raise_priority = true;
3232
3233 sc.nr_reclaimed = 0;
Mel Gorman84c7a772016-07-28 15:46:44 -07003234 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235
Mel Gorman86c79f62016-07-28 15:45:59 -07003236 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003237 * If the number of buffer_heads exceeds the maximum allowed
3238 * then consider reclaiming from all zones. This has a dual
3239 * purpose -- on 64-bit systems it is expected that
3240 * buffer_heads are stripped during active rotation. On 32-bit
3241 * systems, highmem pages can pin lowmem memory and shrinking
3242 * buffers can relieve lowmem pressure. Reclaim may still not
3243 * go ahead if all eligible zones for the original allocation
3244 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003245 */
3246 if (buffer_heads_over_limit) {
3247 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3248 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003249 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003250 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251
Mel Gorman970a39a2016-07-28 15:46:35 -07003252 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003253 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003256
Mel Gorman86c79f62016-07-28 15:45:59 -07003257 /*
3258 * Only reclaim if there are no eligible zones. Check from
3259 * high to low zone as allocations prefer higher zones.
3260 * Scanning from low to high zone would allow congestion to be
3261 * cleared during a very small window when a small low
3262 * zone was balanced even under extreme pressure when the
Mel Gorman84c7a772016-07-28 15:46:44 -07003263 * overall node may be congested. Note that sc.reclaim_idx
3264 * is not used as buffer_heads_over_limit may have adjusted
3265 * it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003266 */
Mel Gorman84c7a772016-07-28 15:46:44 -07003267 for (i = classzone_idx; i >= 0; i--) {
Mel Gorman86c79f62016-07-28 15:45:59 -07003268 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003269 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003270 continue;
3271
Mel Gorman84c7a772016-07-28 15:46:44 -07003272 if (zone_balanced(zone, sc.order, classzone_idx))
Mel Gorman86c79f62016-07-28 15:45:59 -07003273 goto out;
3274 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003275
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003277 * Do some background aging of the anon list, to give
3278 * pages a chance to be referenced before reclaiming. All
3279 * pages are rotated regardless of classzone as this is
3280 * about consistent aging.
3281 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003282 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003283
3284 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003285 * If we're getting trouble reclaiming, start doing writepage
3286 * even in laptop mode.
3287 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003288 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003289 sc.may_writepage = 1;
3290
Mel Gorman1d82de62016-07-28 15:45:43 -07003291 /* Call soft limit reclaim before calling shrink_node. */
3292 sc.nr_scanned = 0;
3293 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003294 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003295 sc.gfp_mask, &nr_soft_scanned);
3296 sc.nr_reclaimed += nr_soft_reclaimed;
3297
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003298 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003299 * There should be no need to raise the scanning priority if
3300 * enough pages are already being scanned that that high
3301 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003303 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003304 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003305
3306 /*
3307 * If the low watermark is met there is no need for processes
3308 * to be throttled on pfmemalloc_wait as they should not be
3309 * able to safely make forward progress. Wake them
3310 */
3311 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3312 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003313 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003314
Mel Gormanb8e83b92013-07-03 15:01:45 -07003315 /* Check if kswapd should be suspending */
3316 if (try_to_freeze() || kthread_should_stop())
3317 break;
3318
3319 /*
3320 * Raise priority if scanning rate is too low or there was no
3321 * progress in reclaiming pages
3322 */
3323 if (raise_priority || !sc.nr_reclaimed)
3324 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003325 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326
Mel Gormanb8e83b92013-07-03 15:01:45 -07003327out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003328 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003329 * Return the order kswapd stopped reclaiming at as
3330 * prepare_kswapd_sleep() takes it into account. If another caller
3331 * entered the allocator slow path while kswapd was awake, order will
3332 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003333 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003334 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335}
3336
Mel Gorman38087d92016-07-28 15:45:49 -07003337static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3338 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003339{
3340 long remaining = 0;
3341 DEFINE_WAIT(wait);
3342
3343 if (freezing(current) || kthread_should_stop())
3344 return;
3345
3346 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3347
3348 /* Try to sleep for a short interval */
Mel Gormand9f21d42016-07-28 15:46:41 -07003349 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003350 /*
3351 * Compaction records what page blocks it recently failed to
3352 * isolate pages from and skips them in the future scanning.
3353 * When kswapd is going to sleep, it is reasonable to assume
3354 * that pages and compaction may succeed so reset the cache.
3355 */
3356 reset_isolation_suitable(pgdat);
3357
3358 /*
3359 * We have freed the memory, now we should compact it to make
3360 * allocation of the requested order possible.
3361 */
Mel Gorman38087d92016-07-28 15:45:49 -07003362 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003363
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003364 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003365
3366 /*
3367 * If woken prematurely then reset kswapd_classzone_idx and
3368 * order. The values will either be from a wakeup request or
3369 * the previous request that slept prematurely.
3370 */
3371 if (remaining) {
3372 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3373 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3374 }
3375
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003376 finish_wait(&pgdat->kswapd_wait, &wait);
3377 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3378 }
3379
3380 /*
3381 * After a short sleep, check if it was a premature sleep. If not, then
3382 * go fully to sleep until explicitly woken up.
3383 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003384 if (!remaining &&
3385 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003386 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3387
3388 /*
3389 * vmstat counters are not perfectly accurate and the estimated
3390 * value for counters such as NR_FREE_PAGES can deviate from the
3391 * true value by nr_online_cpus * threshold. To avoid the zone
3392 * watermarks being breached while under pressure, we reduce the
3393 * per-cpu vmstat threshold while kswapd is awake and restore
3394 * them before going back to sleep.
3395 */
3396 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003397
3398 if (!kthread_should_stop())
3399 schedule();
3400
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003401 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3402 } else {
3403 if (remaining)
3404 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3405 else
3406 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3407 }
3408 finish_wait(&pgdat->kswapd_wait, &wait);
3409}
3410
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411/*
3412 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003413 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 *
3415 * This basically trickles out pages so that we have _some_
3416 * free memory available even if there is no other activity
3417 * that frees anything up. This is needed for things like routing
3418 * etc, where we otherwise might have all activity going on in
3419 * asynchronous contexts that cannot page things out.
3420 *
3421 * If there are applications that are active memory-allocators
3422 * (most normal use), this basically shouldn't matter.
3423 */
3424static int kswapd(void *p)
3425{
Mel Gorman38087d92016-07-28 15:45:49 -07003426 unsigned int alloc_order, reclaim_order, classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 pg_data_t *pgdat = (pg_data_t*)p;
3428 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003429
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 struct reclaim_state reclaim_state = {
3431 .reclaimed_slab = 0,
3432 };
Rusty Russella70f7302009-03-13 14:49:46 +10303433 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434
Nick Piggincf40bd12009-01-21 08:12:39 +01003435 lockdep_set_current_reclaim_state(GFP_KERNEL);
3436
Rusty Russell174596a2009-01-01 10:12:29 +10303437 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003438 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 current->reclaim_state = &reclaim_state;
3440
3441 /*
3442 * Tell the memory management that we're a "memory allocator",
3443 * and that if we need more memory we should get access to it
3444 * regardless (see "__alloc_pages()"). "kswapd" should
3445 * never get caught in the normal page freeing logic.
3446 *
3447 * (Kswapd normally doesn't need memory anyway, but sometimes
3448 * you need a small amount of memory in order to be able to
3449 * page out something else, and this flag essentially protects
3450 * us from recursively trying to free more memory as we're
3451 * trying to free the first piece of memory in the first place).
3452 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003453 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003454 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455
Mel Gorman38087d92016-07-28 15:45:49 -07003456 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3457 pgdat->kswapd_classzone_idx = classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003459 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003460
Mel Gorman38087d92016-07-28 15:45:49 -07003461kswapd_try_sleep:
3462 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3463 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003464
Mel Gorman38087d92016-07-28 15:45:49 -07003465 /* Read the new order and classzone_idx */
3466 alloc_order = reclaim_order = pgdat->kswapd_order;
3467 classzone_idx = pgdat->kswapd_classzone_idx;
3468 pgdat->kswapd_order = 0;
3469 pgdat->kswapd_classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
David Rientjes8fe23e02009-12-14 17:58:33 -08003471 ret = try_to_freeze();
3472 if (kthread_should_stop())
3473 break;
3474
3475 /*
3476 * We can speed up thawing tasks if we don't call balance_pgdat
3477 * after returning from the refrigerator
3478 */
Mel Gorman38087d92016-07-28 15:45:49 -07003479 if (ret)
3480 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003481
Mel Gorman38087d92016-07-28 15:45:49 -07003482 /*
3483 * Reclaim begins at the requested order but if a high-order
3484 * reclaim fails then kswapd falls back to reclaiming for
3485 * order-0. If that happens, kswapd will consider sleeping
3486 * for the order it finished reclaiming at (reclaim_order)
3487 * but kcompactd is woken to compact for the original
3488 * request (alloc_order).
3489 */
Mel Gormane5146b12016-07-28 15:46:47 -07003490 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3491 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003492 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3493 if (reclaim_order < alloc_order)
3494 goto kswapd_try_sleep;
3495
3496 alloc_order = reclaim_order = pgdat->kswapd_order;
3497 classzone_idx = pgdat->kswapd_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003499
Johannes Weiner71abdc12014-06-06 14:35:35 -07003500 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003501 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003502 lockdep_clear_current_reclaim_state();
3503
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 return 0;
3505}
3506
3507/*
3508 * A zone is low on free memory, so wake its kswapd task to service it.
3509 */
Mel Gorman99504742011-01-13 15:46:20 -08003510void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511{
3512 pg_data_t *pgdat;
Mel Gormane1a55632016-07-28 15:46:26 -07003513 int z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514
Mel Gorman6aa303d2016-09-01 16:14:55 -07003515 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516 return;
3517
Vladimir Davydov344736f2014-10-20 15:50:30 +04003518 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003520 pgdat = zone->zone_pgdat;
Mel Gorman38087d92016-07-28 15:45:49 -07003521 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3522 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003523 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003525
3526 /* Only wake kswapd if all zones are unbalanced */
3527 for (z = 0; z <= classzone_idx; z++) {
3528 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003529 if (!managed_zone(zone))
Mel Gormane1a55632016-07-28 15:46:26 -07003530 continue;
3531
3532 if (zone_balanced(zone, order, classzone_idx))
3533 return;
3534 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08003535
3536 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003537 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538}
3539
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003540#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003542 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003543 * freed pages.
3544 *
3545 * Rather than trying to age LRUs the aim is to preserve the overall
3546 * LRU order by reclaiming preferentially
3547 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003549unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003551 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003552 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003553 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003554 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003555 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003556 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003557 .may_writepage = 1,
3558 .may_unmap = 1,
3559 .may_swap = 1,
3560 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003562 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003563 struct task_struct *p = current;
3564 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003566 p->flags |= PF_MEMALLOC;
3567 lockdep_set_current_reclaim_state(sc.gfp_mask);
3568 reclaim_state.reclaimed_slab = 0;
3569 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003570
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003571 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003572
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003573 p->reclaim_state = NULL;
3574 lockdep_clear_current_reclaim_state();
3575 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003576
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003577 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003579#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581/* It's optimal to keep kswapds on the same CPUs as their memory, but
3582 not required for correctness. So if the last cpu in a node goes
3583 away, we get changed to run anywhere: as the first one comes back,
3584 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003585static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3586 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003588 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003590 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003591 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003592 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303593 const struct cpumask *mask;
3594
3595 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003596
Rusty Russell3e597942009-01-01 10:12:24 +10303597 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003599 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 }
3601 }
3602 return NOTIFY_OK;
3603}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604
Yasunori Goto3218ae12006-06-27 02:53:33 -07003605/*
3606 * This kswapd start function will be called by init and node-hot-add.
3607 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3608 */
3609int kswapd_run(int nid)
3610{
3611 pg_data_t *pgdat = NODE_DATA(nid);
3612 int ret = 0;
3613
3614 if (pgdat->kswapd)
3615 return 0;
3616
3617 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3618 if (IS_ERR(pgdat->kswapd)) {
3619 /* failure at boot is fatal */
3620 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003621 pr_err("Failed to start kswapd on node %d\n", nid);
3622 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003623 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003624 }
3625 return ret;
3626}
3627
David Rientjes8fe23e02009-12-14 17:58:33 -08003628/*
Jiang Liud8adde12012-07-11 14:01:52 -07003629 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003630 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003631 */
3632void kswapd_stop(int nid)
3633{
3634 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3635
Jiang Liud8adde12012-07-11 14:01:52 -07003636 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003637 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003638 NODE_DATA(nid)->kswapd = NULL;
3639 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003640}
3641
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642static int __init kswapd_init(void)
3643{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003644 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003645
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003647 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003648 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649 hotcpu_notifier(cpu_callback, 0);
3650 return 0;
3651}
3652
3653module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003654
3655#ifdef CONFIG_NUMA
3656/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003657 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003658 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003659 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003660 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003661 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003662int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003663
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003664#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003665#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003666#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003667#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003668
Christoph Lameter9eeff232006-01-18 17:42:31 -08003669/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003670 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003671 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3672 * a zone.
3673 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003674#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003675
Christoph Lameter9eeff232006-01-18 17:42:31 -08003676/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003677 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003678 * occur.
3679 */
3680int sysctl_min_unmapped_ratio = 1;
3681
3682/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003683 * If the number of slab pages in a zone grows beyond this percentage then
3684 * slab reclaim needs to occur.
3685 */
3686int sysctl_min_slab_ratio = 5;
3687
Mel Gorman11fb9982016-07-28 15:46:20 -07003688static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003689{
Mel Gorman11fb9982016-07-28 15:46:20 -07003690 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3691 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3692 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003693
3694 /*
3695 * It's possible for there to be more file mapped pages than
3696 * accounted for by the pages on the file LRU lists because
3697 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3698 */
3699 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3700}
3701
3702/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003703static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003704{
Alexandru Moised031a152015-11-05 18:48:08 -08003705 unsigned long nr_pagecache_reclaimable;
3706 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003707
3708 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003709 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003710 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003711 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003712 * a better estimate
3713 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003714 if (node_reclaim_mode & RECLAIM_UNMAP)
3715 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003716 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003717 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003718
3719 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003720 if (!(node_reclaim_mode & RECLAIM_WRITE))
3721 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003722
3723 /* Watch for any possible underflows due to delta */
3724 if (unlikely(delta > nr_pagecache_reclaimable))
3725 delta = nr_pagecache_reclaimable;
3726
3727 return nr_pagecache_reclaimable - delta;
3728}
3729
Christoph Lameter0ff38492006-09-25 23:31:52 -07003730/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003731 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003732 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003733static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003734{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003735 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003736 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003737 struct task_struct *p = current;
3738 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003739 int classzone_idx = gfp_zone(gfp_mask);
Andrew Morton179e9632006-03-22 00:08:18 -08003740 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003741 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003742 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003743 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003744 .priority = NODE_RECLAIM_PRIORITY,
3745 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3746 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003747 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003748 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003749 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003750
Christoph Lameter9eeff232006-01-18 17:42:31 -08003751 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003752 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003753 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003754 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003755 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003756 */
3757 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003758 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003759 reclaim_state.reclaimed_slab = 0;
3760 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003761
Mel Gormana5f5f912016-07-28 15:46:32 -07003762 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003763 /*
3764 * Free memory by calling shrink zone with increasing
3765 * priorities until we have enough memory freed.
3766 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003767 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003768 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003769 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003770 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003771
Christoph Lameter9eeff232006-01-18 17:42:31 -08003772 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003773 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003774 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003775 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003776}
Andrew Morton179e9632006-03-22 00:08:18 -08003777
Mel Gormana5f5f912016-07-28 15:46:32 -07003778int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003779{
David Rientjesd773ed62007-10-16 23:26:01 -07003780 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003781
3782 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003783 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003784 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003785 *
Christoph Lameter96146342006-07-03 00:24:13 -07003786 * A small portion of unmapped file backed pages is needed for
3787 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003788 * thrown out if the node is overallocated. So we do not reclaim
3789 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003790 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003791 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003792 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3793 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3794 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003795
Mel Gormana5f5f912016-07-28 15:46:32 -07003796 if (!pgdat_reclaimable(pgdat))
3797 return NODE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003798
Andrew Morton179e9632006-03-22 00:08:18 -08003799 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003800 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003801 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003802 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003803 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003804
3805 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003806 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003807 * have associated processors. This will favor the local processor
3808 * over remote processors and spread off node memory allocations
3809 * as wide as possible.
3810 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003811 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3812 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003813
Mel Gormana5f5f912016-07-28 15:46:32 -07003814 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3815 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003816
Mel Gormana5f5f912016-07-28 15:46:32 -07003817 ret = __node_reclaim(pgdat, gfp_mask, order);
3818 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003819
Mel Gorman24cf725182009-06-16 15:33:23 -07003820 if (!ret)
3821 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3822
David Rientjesd773ed62007-10-16 23:26:01 -07003823 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003824}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003825#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003826
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003827/*
3828 * page_evictable - test whether a page is evictable
3829 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003830 *
3831 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003832 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003833 *
3834 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003835 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003836 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003837 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003838 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003839int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003840{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003841 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003842}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003843
Hugh Dickins850465792012-01-20 14:34:19 -08003844#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003845/**
Hugh Dickins24513262012-01-20 14:34:21 -08003846 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3847 * @pages: array of pages to check
3848 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003849 *
Hugh Dickins24513262012-01-20 14:34:21 -08003850 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003851 *
3852 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003853 */
Hugh Dickins24513262012-01-20 14:34:21 -08003854void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003855{
Johannes Weiner925b7672012-01-12 17:18:15 -08003856 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003857 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003858 int pgscanned = 0;
3859 int pgrescued = 0;
3860 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003861
Hugh Dickins24513262012-01-20 14:34:21 -08003862 for (i = 0; i < nr_pages; i++) {
3863 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003864 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003865
Hugh Dickins24513262012-01-20 14:34:21 -08003866 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003867 if (pagepgdat != pgdat) {
3868 if (pgdat)
3869 spin_unlock_irq(&pgdat->lru_lock);
3870 pgdat = pagepgdat;
3871 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003872 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003873 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003874
Hugh Dickins24513262012-01-20 14:34:21 -08003875 if (!PageLRU(page) || !PageUnevictable(page))
3876 continue;
3877
Hugh Dickins39b5f292012-10-08 16:33:18 -07003878 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003879 enum lru_list lru = page_lru_base_type(page);
3880
Sasha Levin309381fea2014-01-23 15:52:54 -08003881 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003882 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003883 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3884 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003885 pgrescued++;
3886 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003887 }
Hugh Dickins24513262012-01-20 14:34:21 -08003888
Mel Gorman785b99f2016-07-28 15:47:23 -07003889 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003890 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3891 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003892 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003893 }
Hugh Dickins850465792012-01-20 14:34:19 -08003894}
3895#endif /* CONFIG_SHMEM */