blob: 1778828f1bb92e6ef6755c249cd51feb6e98b88b [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 Kim13e7b912015-01-09 18:36:55 +0530966 enum page_references references = PAGEREF_RECLAIM;
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);
Minchan Kim07e6c8d2013-05-09 16:21:25 +0900980 if (pgdat)
981 VM_BUG_ON_PAGE(page_pgdat(page) != pgdat, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
983 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800984
Hugh Dickins39b5f292012-10-08 16:33:18 -0700985 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700986 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700987
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700988 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800989 goto keep_locked;
990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 /* Double the slab pressure for mapped and swapcache pages */
992 if (page_mapped(page) || PageSwapCache(page))
993 sc->nr_scanned++;
994
Andy Whitcroftc661b072007-08-22 14:01:26 -0700995 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
996 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
997
Mel Gorman283aba92013-07-03 15:01:51 -0700998 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700999 * The number of dirty pages determines if a zone is marked
1000 * reclaim_congested which affects wait_iff_congested. kswapd
1001 * will stall and start writing pages if the tail of the LRU
1002 * is all dirty unqueued pages.
1003 */
1004 page_check_dirty_writeback(page, &dirty, &writeback);
1005 if (dirty || writeback)
1006 nr_dirty++;
1007
1008 if (dirty && !writeback)
1009 nr_unqueued_dirty++;
1010
Mel Gormand04e8ac2013-07-03 15:02:03 -07001011 /*
1012 * Treat this page as congested if the underlying BDI is or if
1013 * pages are cycling through the LRU so quickly that the
1014 * pages marked for immediate reclaim are making it to the
1015 * end of the LRU a second time.
1016 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001017 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001018 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001019 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001020 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001021 nr_congested++;
1022
1023 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001024 * If a page at the tail of the LRU is under writeback, there
1025 * are three cases to consider.
1026 *
1027 * 1) If reclaim is encountering an excessive number of pages
1028 * under writeback and this page is both under writeback and
1029 * PageReclaim then it indicates that pages are being queued
1030 * for IO but are being recycled through the LRU before the
1031 * IO can complete. Waiting on the page itself risks an
1032 * indefinite stall if it is impossible to writeback the
1033 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001034 * note that the LRU is being scanned too quickly and the
1035 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001036 *
Tejun Heo97c93412015-05-22 18:23:36 -04001037 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001038 * not marked for immediate reclaim, or the caller does not
1039 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1040 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001041 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001042 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001043 * Require may_enter_fs because we would wait on fs, which
1044 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001045 * enter reclaim, and deadlock if it waits on a page for
1046 * which it is needed to do the write (loop masks off
1047 * __GFP_IO|__GFP_FS for this reason); but more thought
1048 * would probably show more reasons.
1049 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001050 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001051 * PageReclaim. memcg does not have any dirty pages
1052 * throttling so we could easily OOM just because too many
1053 * pages are in writeback and there is nothing else to
1054 * reclaim. Wait for the writeback to complete.
1055 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001056 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001057 /* Case 1 above */
1058 if (current_is_kswapd() &&
1059 PageReclaim(page) &&
Minchan Kim07e6c8d2013-05-09 16:21:25 +09001060 (pgdat && test_bit(PGDAT_WRITEBACK, &pgdat->flags))) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001061 nr_immediate++;
1062 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001063
1064 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001065 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001066 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001067 /*
1068 * This is slightly racy - end_page_writeback()
1069 * might have just cleared PageReclaim, then
1070 * setting PageReclaim here end up interpreted
1071 * as PageReadahead - but that does not matter
1072 * enough to care. What we do want is for this
1073 * page to have PageReclaim set next time memcg
1074 * reclaim reaches the tests above, so it will
1075 * then wait_on_page_writeback() to avoid OOM;
1076 * and it's also appropriate in global reclaim.
1077 */
1078 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001079 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001080 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001081
1082 /* Case 3 above */
1083 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001084 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001085 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001086 /* then go back and try same page again */
1087 list_add_tail(&page->lru, page_list);
1088 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001089 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001090 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091
Minchan Kim02c6de82012-10-08 16:31:55 -07001092 if (!force_reclaim)
1093 references = page_check_references(page, sc);
1094
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001095 switch (references) {
1096 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001098 case PAGEREF_KEEP:
1099 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001100 case PAGEREF_RECLAIM:
1101 case PAGEREF_RECLAIM_CLEAN:
1102 ; /* try to reclaim the page below */
1103 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 /*
1106 * Anonymous process memory has backing store?
1107 * Try to allocate it some swap space here.
1108 */
Nick Pigginb291f002008-10-18 20:26:44 -07001109 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001110 if (!(sc->gfp_mask & __GFP_IO))
1111 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001112 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001114 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001115 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
Mel Gormane2be15f2013-07-03 15:01:57 -07001117 /* Adding to swap updated mapping */
1118 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001119 } else if (unlikely(PageTransHuge(page))) {
1120 /* Split file THP */
1121 if (split_huge_page_to_list(page, page_list))
1122 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001123 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001125 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1126
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 /*
1128 * The page is mapped into the page tables of one or more
1129 * processes. Try to unmap it here.
1130 */
1131 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001132 switch (ret = try_to_unmap(page, lazyfree ?
1133 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1134 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 case SWAP_FAIL:
1136 goto activate_locked;
1137 case SWAP_AGAIN:
1138 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001139 case SWAP_MLOCK:
1140 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001141 case SWAP_LZFREE:
1142 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 case SWAP_SUCCESS:
1144 ; /* try to free the page below */
1145 }
1146 }
1147
1148 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001149 /*
1150 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001151 * avoid risk of stack overflow but only writeback
1152 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001153 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001154 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001155 (!current_is_kswapd() ||
Minchan Kim07e6c8d2013-05-09 16:21:25 +09001156 (pgdat &&
1157 !test_bit(PGDAT_DIRTY, &pgdat->flags)))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001158 /*
1159 * Immediately reclaim when written back.
1160 * Similar in principal to deactivate_page()
1161 * except we already have the page isolated
1162 * and know it's dirty
1163 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001164 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001165 SetPageReclaim(page);
1166
Mel Gormanee728862011-10-31 17:07:38 -07001167 goto keep_locked;
1168 }
1169
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001170 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001172 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001174 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 goto keep_locked;
1176
Mel Gormand950c942015-09-04 15:47:35 -07001177 /*
1178 * Page is dirty. Flush the TLB if a writable entry
1179 * potentially exists to avoid CPU writes after IO
1180 * starts and then write it out here.
1181 */
1182 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001183 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 case PAGE_KEEP:
1185 goto keep_locked;
1186 case PAGE_ACTIVATE:
1187 goto activate_locked;
1188 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001189 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001190 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001191 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001193
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 /*
1195 * A synchronous write - probably a ramdisk. Go
1196 * ahead and try to reclaim the page.
1197 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001198 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 goto keep;
1200 if (PageDirty(page) || PageWriteback(page))
1201 goto keep_locked;
1202 mapping = page_mapping(page);
1203 case PAGE_CLEAN:
1204 ; /* try to free the page below */
1205 }
1206 }
1207
1208 /*
1209 * If the page has buffers, try to free the buffer mappings
1210 * associated with this page. If we succeed we try to free
1211 * the page as well.
1212 *
1213 * We do this even if the page is PageDirty().
1214 * try_to_release_page() does not perform I/O, but it is
1215 * possible for a page to have PageDirty set, but it is actually
1216 * clean (all its buffers are clean). This happens if the
1217 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001218 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 * try_to_release_page() will discover that cleanness and will
1220 * drop the buffers and mark the page clean - it can be freed.
1221 *
1222 * Rarely, pages can have buffers and no ->mapping. These are
1223 * the pages which were not successfully invalidated in
1224 * truncate_complete_page(). We try to drop those buffers here
1225 * and if that worked, and the page is no longer mapped into
1226 * process address space (page_count == 1) it can be freed.
1227 * Otherwise, leave the page on the LRU so it is swappable.
1228 */
David Howells266cf652009-04-03 16:42:36 +01001229 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 if (!try_to_release_page(page, sc->gfp_mask))
1231 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001232 if (!mapping && page_count(page) == 1) {
1233 unlock_page(page);
1234 if (put_page_testzero(page))
1235 goto free_it;
1236 else {
1237 /*
1238 * rare race with speculative reference.
1239 * the speculative reference will free
1240 * this page shortly, so we may
1241 * increment nr_reclaimed here (and
1242 * leave it off the LRU).
1243 */
1244 nr_reclaimed++;
1245 continue;
1246 }
1247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 }
1249
Minchan Kim854e9ed2016-01-15 16:54:53 -08001250lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001251 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001252 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Nick Piggina978d6f2008-10-18 20:26:58 -07001254 /*
1255 * At this point, we have no other references and there is
1256 * no way to pick any more up (removed from LRU, removed
1257 * from pagecache). Can use non-atomic bitops now (and
1258 * we obviously don't have to worry about waking up a process
1259 * waiting on the page lock, because there are no references.
1260 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001261 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001262free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001263 if (ret == SWAP_LZFREE)
1264 count_vm_event(PGLAZYFREED);
1265
Andrew Morton05ff5132006-03-22 00:08:20 -08001266 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001267
1268 /*
1269 * Is there need to periodically free_page_list? It would
1270 * appear not as the counts should be low
1271 */
1272 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 continue;
1274
Nick Pigginb291f002008-10-18 20:26:44 -07001275cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001276 if (PageSwapCache(page))
1277 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001278 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001279 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001280 continue;
1281
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001283 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001284 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001285 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001286 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 SetPageActive(page);
1288 pgactivate++;
1289keep_locked:
1290 unlock_page(page);
1291keep:
1292 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001293 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001295
Johannes Weiner747db952014-08-08 14:19:24 -07001296 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001297 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001298 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001299
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001301 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001302
Mel Gorman8e950282013-07-03 15:02:02 -07001303 *ret_nr_dirty += nr_dirty;
1304 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001305 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001306 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001307 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001308 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309}
1310
Minchan Kim02c6de82012-10-08 16:31:55 -07001311unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1312 struct list_head *page_list)
1313{
1314 struct scan_control sc = {
1315 .gfp_mask = GFP_KERNEL,
1316 .priority = DEF_PRIORITY,
1317 .may_unmap = 1,
Minchan Kim13e7b912015-01-09 18:36:55 +05301318 /* Doesn't allow to write out dirty page */
1319 .may_writepage = 0,
Minchan Kim02c6de82012-10-08 16:31:55 -07001320 };
Mel Gorman8e950282013-07-03 15:02:02 -07001321 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001322 struct page *page, *next;
1323 LIST_HEAD(clean_pages);
1324
1325 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001326 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001327 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001328 ClearPageActive(page);
1329 list_move(&page->lru, &clean_pages);
1330 }
1331 }
1332
Mel Gorman599d0c92016-07-28 15:45:31 -07001333 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001334 TTU_UNMAP|TTU_IGNORE_ACCESS,
1335 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001336 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001337 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001338 return ret;
1339}
1340
Minchan Kimff926892013-05-09 16:21:24 +09001341#ifdef CONFIG_PROCESS_RECLAIM
1342static unsigned long shrink_page(struct page *page,
1343 struct zone *zone,
1344 struct scan_control *sc,
1345 enum ttu_flags ttu_flags,
1346 unsigned long *dummy1,
1347 unsigned long *dummy2,
1348 unsigned long *dummy3,
1349 unsigned long *dummy4,
1350 unsigned long *dummy5,
1351 bool force_reclaim,
1352 struct list_head *ret_pages)
1353{
1354 int reclaimed;
1355 LIST_HEAD(page_list);
1356 list_add(&page->lru, &page_list);
1357
1358 reclaimed = shrink_page_list(&page_list, zone->zone_pgdat, sc,
1359 ttu_flags, dummy1, dummy2, dummy3, dummy4,
1360 dummy5, force_reclaim);
1361 if (!reclaimed)
1362 list_splice(&page_list, ret_pages);
1363
1364 return reclaimed;
1365}
1366
1367unsigned long reclaim_pages_from_list(struct list_head *page_list)
1368{
1369 struct scan_control sc = {
1370 .gfp_mask = GFP_KERNEL,
1371 .priority = DEF_PRIORITY,
1372 .may_writepage = 1,
1373 .may_unmap = 1,
1374 .may_swap = 1,
1375 };
1376
1377 LIST_HEAD(ret_pages);
1378 struct page *page;
1379 unsigned long dummy1, dummy2, dummy3, dummy4, dummy5;
1380 unsigned long nr_reclaimed = 0;
1381
1382 while (!list_empty(page_list)) {
1383 page = lru_to_page(page_list);
1384 list_del(&page->lru);
1385
1386 ClearPageActive(page);
1387 nr_reclaimed += shrink_page(page, page_zone(page), &sc,
1388 TTU_UNMAP|TTU_IGNORE_ACCESS,
1389 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true,
1390 &ret_pages);
1391 }
1392
1393 while (!list_empty(&ret_pages)) {
1394 page = lru_to_page(&ret_pages);
1395 list_del(&page->lru);
1396 dec_node_page_state(page, NR_ISOLATED_ANON +
1397 page_is_file_cache(page));
1398 putback_lru_page(page);
1399 }
1400
1401 return nr_reclaimed;
1402}
1403#endif
1404
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001405/*
1406 * Attempt to remove the specified page from its LRU. Only take this page
1407 * if it is of the appropriate PageActive status. Pages which are being
1408 * freed elsewhere are also ignored.
1409 *
1410 * page: page to consider
1411 * mode: one of the LRU isolation modes defined above
1412 *
1413 * returns 0 on success, -ve errno on failure.
1414 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001415int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001416{
1417 int ret = -EINVAL;
1418
1419 /* Only take pages on the LRU. */
1420 if (!PageLRU(page))
1421 return ret;
1422
Minchan Kime46a2872012-10-08 16:33:48 -07001423 /* Compaction should not handle unevictable pages but CMA can do so */
1424 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001425 return ret;
1426
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001427 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001428
Mel Gormanc8244932012-01-12 17:19:38 -08001429 /*
1430 * To minimise LRU disruption, the caller can indicate that it only
1431 * wants to isolate pages it will be able to operate on without
1432 * blocking - clean pages for the most part.
1433 *
1434 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1435 * is used by reclaim when it is cannot write to backing storage
1436 *
1437 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1438 * that it is possible to migrate without blocking
1439 */
1440 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1441 /* All the caller can do on PageWriteback is block */
1442 if (PageWriteback(page))
1443 return ret;
1444
1445 if (PageDirty(page)) {
1446 struct address_space *mapping;
1447
1448 /* ISOLATE_CLEAN means only clean pages */
1449 if (mode & ISOLATE_CLEAN)
1450 return ret;
1451
1452 /*
1453 * Only pages without mappings or that have a
1454 * ->migratepage callback are possible to migrate
1455 * without blocking
1456 */
1457 mapping = page_mapping(page);
1458 if (mapping && !mapping->a_ops->migratepage)
1459 return ret;
1460 }
1461 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001462
Minchan Kimf80c0672011-10-31 17:06:55 -07001463 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1464 return ret;
1465
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001466 if (likely(get_page_unless_zero(page))) {
1467 /*
1468 * Be careful not to clear PageLRU until after we're
1469 * sure the page is not being freed elsewhere -- the
1470 * page release code relies on it.
1471 */
1472 ClearPageLRU(page);
1473 ret = 0;
1474 }
1475
1476 return ret;
1477}
1478
Mel Gorman7ee36a12016-07-28 15:47:17 -07001479
1480/*
1481 * Update LRU sizes after isolating pages. The LRU size updates must
1482 * be complete before mem_cgroup_update_lru_size due to a santity check.
1483 */
1484static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hocko07fc9572017-01-10 16:58:04 -08001485 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001486{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001487 int zid;
1488
Mel Gorman7ee36a12016-07-28 15:47:17 -07001489 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1490 if (!nr_zone_taken[zid])
1491 continue;
1492
1493 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hocko07fc9572017-01-10 16:58:04 -08001494#ifdef CONFIG_MEMCG
1495 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1496#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001497 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001498
Mel Gorman7ee36a12016-07-28 15:47:17 -07001499}
1500
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001501/*
Mel Gormana52633d2016-07-28 15:45:28 -07001502 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * shrink the lists perform better by taking out a batch of pages
1504 * and working on them outside the LRU lock.
1505 *
1506 * For pagecache intensive workloads, this function is the hottest
1507 * spot in the kernel (apart from copy_*_user functions).
1508 *
1509 * Appropriate locks must be held before calling this function.
1510 *
1511 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001512 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001514 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001515 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001516 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001517 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 *
1519 * returns how many pages were moved onto *@dst.
1520 */
Andrew Morton69e05942006-03-22 00:08:19 -08001521static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001522 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001523 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001524 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001526 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001527 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001528 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001529 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Mel Gorman599d0c92016-07-28 15:45:31 -07001530 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001531 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001533 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Mel Gormand7f05522016-07-28 15:47:37 -07001534 !list_empty(src);) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001535 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001536
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 page = lru_to_page(src);
1538 prefetchw_prev_lru_page(page, src, flags);
1539
Sasha Levin309381fea2014-01-23 15:52:54 -08001540 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001541
Mel Gormanb2e18752016-07-28 15:45:37 -07001542 if (page_zonenum(page) > sc->reclaim_idx) {
1543 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001544 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001545 continue;
1546 }
1547
Mel Gormand7f05522016-07-28 15:47:37 -07001548 /*
1549 * Account for scanned and skipped separetly to avoid the pgdat
1550 * being prematurely marked unreclaimable by pgdat_reclaimable.
1551 */
1552 scan++;
1553
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001554 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001555 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001556 nr_pages = hpage_nr_pages(page);
1557 nr_taken += nr_pages;
1558 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001559 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001560 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001561
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001562 case -EBUSY:
1563 /* else it is being freed elsewhere */
1564 list_move(&page->lru, src);
1565 continue;
1566
1567 default:
1568 BUG();
1569 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 }
1571
Mel Gormanb2e18752016-07-28 15:45:37 -07001572 /*
1573 * Splice any skipped pages to the start of the LRU list. Note that
1574 * this disrupts the LRU order when reclaiming for lower zones but
1575 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1576 * scanning would soon rescan the same pages to skip and put the
1577 * system at risk of premature OOM.
1578 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001579 if (!list_empty(&pages_skipped)) {
1580 int zid;
Mel Gormand7f05522016-07-28 15:47:37 -07001581 unsigned long total_skipped = 0;
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001582
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001583 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1584 if (!nr_skipped[zid])
1585 continue;
1586
1587 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Mel Gormand7f05522016-07-28 15:47:37 -07001588 total_skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001589 }
Mel Gormand7f05522016-07-28 15:47:37 -07001590
1591 /*
1592 * Account skipped pages as a partial scan as the pgdat may be
1593 * close to unreclaimable. If the LRU list is empty, account
1594 * skipped pages as a full scan.
1595 */
1596 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1597
1598 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001599 }
Hugh Dickinsf6260122012-01-12 17:20:06 -08001600 *nr_scanned = scan;
Mel Gormane5146b12016-07-28 15:46:47 -07001601 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
Hugh Dickins75b00af2012-05-29 15:07:09 -07001602 nr_taken, mode, is_file_lru(lru));
Michal Hocko07fc9572017-01-10 16:58:04 -08001603 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 return nr_taken;
1605}
1606
Nick Piggin62695a82008-10-18 20:26:09 -07001607/**
1608 * isolate_lru_page - tries to isolate a page from its LRU list
1609 * @page: page to isolate from its LRU list
1610 *
1611 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1612 * vmstat statistic corresponding to whatever LRU list the page was on.
1613 *
1614 * Returns 0 if the page was removed from an LRU list.
1615 * Returns -EBUSY if the page was not on an LRU list.
1616 *
1617 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001618 * the active list, it will have PageActive set. If it was found on
1619 * the unevictable list, it will have the PageUnevictable bit set. That flag
1620 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001621 *
1622 * The vmstat statistic corresponding to the list on which the page was
1623 * found will be decremented.
1624 *
1625 * Restrictions:
1626 * (1) Must be called with an elevated refcount on the page. This is a
1627 * fundamentnal difference from isolate_lru_pages (which is called
1628 * without a stable reference).
1629 * (2) the lru_lock must not be held.
1630 * (3) interrupts must be enabled.
1631 */
1632int isolate_lru_page(struct page *page)
1633{
1634 int ret = -EBUSY;
1635
Sasha Levin309381fea2014-01-23 15:52:54 -08001636 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001637 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001638
Nick Piggin62695a82008-10-18 20:26:09 -07001639 if (PageLRU(page)) {
1640 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001641 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001642
Mel Gormana52633d2016-07-28 15:45:28 -07001643 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001644 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001645 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001646 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001647 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001648 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001649 del_page_from_lru_list(page, lruvec, lru);
1650 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001651 }
Mel Gormana52633d2016-07-28 15:45:28 -07001652 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001653 }
1654 return ret;
1655}
1656
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001657/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001658 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1659 * then get resheduled. When there are massive number of tasks doing page
1660 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1661 * the LRU list will go small and be scanned faster than necessary, leading to
1662 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001663 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001664static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001665 struct scan_control *sc)
1666{
1667 unsigned long inactive, isolated;
1668
1669 if (current_is_kswapd())
1670 return 0;
1671
Tejun Heo97c93412015-05-22 18:23:36 -04001672 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001673 return 0;
1674
1675 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001676 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1677 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001678 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001679 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1680 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001681 }
1682
Fengguang Wu3cf23842012-12-18 14:23:31 -08001683 /*
1684 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1685 * won't get blocked by normal direct-reclaimers, forming a circular
1686 * deadlock.
1687 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001688 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001689 inactive >>= 3;
1690
Rik van Riel35cd7812009-09-21 17:01:38 -07001691 return isolated > inactive;
1692}
1693
Mel Gorman66635622010-08-09 17:19:30 -07001694static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001695putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001696{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001697 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001698 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001699 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001700
Mel Gorman66635622010-08-09 17:19:30 -07001701 /*
1702 * Put back any unfreeable pages.
1703 */
Mel Gorman66635622010-08-09 17:19:30 -07001704 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001705 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001706 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001707
Sasha Levin309381fea2014-01-23 15:52:54 -08001708 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001709 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001710 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001711 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001712 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001713 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001714 continue;
1715 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001716
Mel Gorman599d0c92016-07-28 15:45:31 -07001717 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001718
Linus Torvalds7a608572011-01-17 14:42:19 -08001719 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001720 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001721 add_page_to_lru_list(page, lruvec, lru);
1722
Mel Gorman66635622010-08-09 17:19:30 -07001723 if (is_active_lru(lru)) {
1724 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001725 int numpages = hpage_nr_pages(page);
1726 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001727 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001728 if (put_page_testzero(page)) {
1729 __ClearPageLRU(page);
1730 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001731 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001732
1733 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001734 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001735 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001736 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001737 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001738 } else
1739 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001740 }
1741 }
Mel Gorman66635622010-08-09 17:19:30 -07001742
Hugh Dickins3f797682012-01-12 17:20:07 -08001743 /*
1744 * To save our caller's stack, now use input list for pages to free.
1745 */
1746 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001747}
1748
1749/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001750 * If a kernel thread (such as nfsd for loop-back mounts) services
1751 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1752 * In that case we should only throttle if the backing device it is
1753 * writing to is congested. In other cases it is safe to throttle.
1754 */
1755static int current_may_throttle(void)
1756{
1757 return !(current->flags & PF_LESS_THROTTLE) ||
1758 current->backing_dev_info == NULL ||
1759 bdi_write_congested(current->backing_dev_info);
1760}
1761
Minchan Kim91dcade2016-07-28 15:47:40 -07001762static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1763 struct scan_control *sc, enum lru_list lru)
1764{
1765 int zid;
1766 struct zone *zone;
1767 int file = is_file_lru(lru);
1768 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1769
1770 if (!global_reclaim(sc))
1771 return true;
1772
1773 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1774 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001775 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001776 continue;
1777
1778 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1779 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1780 return true;
1781 }
1782
1783 return false;
1784}
1785
NeilBrown399ba0b2014-06-04 16:07:42 -07001786/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001787 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001788 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 */
Mel Gorman66635622010-08-09 17:19:30 -07001790static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001791shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001792 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793{
1794 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001795 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001796 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001797 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001798 unsigned long nr_dirty = 0;
1799 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001800 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001801 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001802 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001803 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001804 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001805 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001806 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001807
Minchan Kim91dcade2016-07-28 15:47:40 -07001808 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1809 return 0;
1810
Mel Gorman599d0c92016-07-28 15:45:31 -07001811 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001812 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001813
1814 /* We are about to die and free our memory. Return now. */
1815 if (fatal_signal_pending(current))
1816 return SWAP_CLUSTER_MAX;
1817 }
1818
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001820
1821 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001822 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001823 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001824 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001825
Mel Gorman599d0c92016-07-28 15:45:31 -07001826 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001828 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1829 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001830
Mel Gorman599d0c92016-07-28 15:45:31 -07001831 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001832 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001833
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001834 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001835 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001836 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001837 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001838 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001839 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001840 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001841 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001842
Hillf Dantond563c052012-03-21 16:34:02 -07001843 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001844 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001845
Mel Gorman599d0c92016-07-28 15:45:31 -07001846 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001847 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1848 &nr_writeback, &nr_immediate,
1849 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001850
Mel Gorman599d0c92016-07-28 15:45:31 -07001851 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001852
Ying Han904249a2012-04-25 16:01:48 -07001853 if (global_reclaim(sc)) {
1854 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001855 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001856 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001857 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001858 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001859
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001860 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001861
Mel Gorman599d0c92016-07-28 15:45:31 -07001862 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001863
Mel Gorman599d0c92016-07-28 15:45:31 -07001864 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001865
Johannes Weiner747db952014-08-08 14:19:24 -07001866 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001867 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001868
Mel Gorman92df3a72011-10-31 17:07:56 -07001869 /*
1870 * If reclaim is isolating dirty pages under writeback, it implies
1871 * that the long-lived page allocation rate is exceeding the page
1872 * laundering rate. Either the global limits are not being effective
1873 * at throttling processes due to the page distribution throughout
1874 * zones or there is heavy usage of a slow backing device. The
1875 * only option is to throttle from reclaim context which is not ideal
1876 * as there is no guarantee the dirtying process is throttled in the
1877 * same way balance_dirty_pages() manages.
1878 *
Mel Gorman8e950282013-07-03 15:02:02 -07001879 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1880 * of pages under pages flagged for immediate reclaim and stall if any
1881 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001882 */
Mel Gorman918fc712013-07-08 16:00:25 -07001883 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001884 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001885
Mel Gormand43006d2013-07-03 15:01:50 -07001886 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001887 * Legacy memcg will stall in page writeback so avoid forcibly
1888 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001889 */
Tejun Heo97c93412015-05-22 18:23:36 -04001890 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001891 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001892 * Tag a zone as congested if all the dirty pages scanned were
1893 * backed by a congested BDI and wait_iff_congested will stall.
1894 */
1895 if (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001896 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001897
1898 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001899 * If dirty pages are scanned that are not queued for IO, it
1900 * implies that flushers are not keeping up. In this case, flag
Mel Gorman599d0c92016-07-28 15:45:31 -07001901 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
Johannes Weiner57054652014-10-09 15:28:17 -07001902 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001903 */
1904 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001905 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001906
1907 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001908 * If kswapd scans pages marked marked for immediate
1909 * reclaim and under writeback (nr_immediate), it implies
1910 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001911 * they are written so also forcibly stall.
1912 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001913 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001914 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001915 }
Mel Gormand43006d2013-07-03 15:01:50 -07001916
Mel Gorman8e950282013-07-03 15:02:02 -07001917 /*
1918 * Stall direct reclaim for IO completions if underlying BDIs or zone
1919 * is congested. Allow kswapd to continue until it starts encountering
1920 * unqueued dirty pages or cycling through the LRU too quickly.
1921 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001922 if (!sc->hibernation_mode && !current_is_kswapd() &&
1923 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001924 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001925
Mel Gorman599d0c92016-07-28 15:45:31 -07001926 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1927 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001928 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001929 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930}
1931
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001932/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 * This moves pages from the active list to the inactive list.
1934 *
1935 * We move them the other way if the page is referenced by one or more
1936 * processes, from rmap.
1937 *
1938 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001939 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001941 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 * this, so instead we remove the pages from the LRU while processing them.
1943 * It is safe to rely on PG_active against the non-LRU pages in here because
1944 * nobody will play with that bit on a non-LRU page.
1945 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001946 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 * But we had to alter page->flags anyway.
1948 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001949
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001950static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001951 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001952 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001953 enum lru_list lru)
1954{
Mel Gorman599d0c92016-07-28 15:45:31 -07001955 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001956 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001957 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001958 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001959
Wu Fengguang3eb41402009-06-16 15:33:13 -07001960 while (!list_empty(list)) {
1961 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001962 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001963
Sasha Levin309381fea2014-01-23 15:52:54 -08001964 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001965 SetPageLRU(page);
1966
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001967 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001968 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001969 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001970 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001971
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001972 if (put_page_testzero(page)) {
1973 __ClearPageLRU(page);
1974 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001975 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001976
1977 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001978 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001979 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001980 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001981 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001982 } else
1983 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001984 }
1985 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001986
Wu Fengguang3eb41402009-06-16 15:33:13 -07001987 if (!is_active_lru(lru))
1988 __count_vm_events(PGDEACTIVATE, pgmoved);
1989}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001990
Hugh Dickinsf6260122012-01-12 17:20:06 -08001991static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001992 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001993 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001994 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001996 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001997 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001998 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002000 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002001 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002003 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002004 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07002005 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002006 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002007 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
2009 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002010
2011 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07002012 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07002013 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07002014 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07002015
Mel Gorman599d0c92016-07-28 15:45:31 -07002016 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002017
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002018 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2019 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002020
Mel Gorman599d0c92016-07-28 15:45:31 -07002021 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002022 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002023
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002024 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07002025 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
2026 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002027
Mel Gorman599d0c92016-07-28 15:45:31 -07002028 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 while (!list_empty(&l_hold)) {
2031 cond_resched();
2032 page = lru_to_page(&l_hold);
2033 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002034
Hugh Dickins39b5f292012-10-08 16:33:18 -07002035 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002036 putback_lru_page(page);
2037 continue;
2038 }
2039
Mel Gormancc715d92012-03-21 16:34:00 -07002040 if (unlikely(buffer_heads_over_limit)) {
2041 if (page_has_private(page) && trylock_page(page)) {
2042 if (page_has_private(page))
2043 try_to_release_page(page, 0);
2044 unlock_page(page);
2045 }
2046 }
2047
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002048 if (page_referenced(page, 0, sc->target_mem_cgroup,
2049 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08002050 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002051 /*
2052 * Identify referenced, file-backed active pages and
2053 * give them one more trip around the active list. So
2054 * that executable code get better chances to stay in
2055 * memory under moderate memory pressure. Anon pages
2056 * are not likely to be evicted by use-once streaming
2057 * IO, plus JVM can create lots of anon VM_EXEC pages,
2058 * so we ignore them here.
2059 */
Wu Fengguang41e20982009-10-26 16:49:53 -07002060 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002061 list_add(&page->lru, &l_active);
2062 continue;
2063 }
2064 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002065
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002066 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 list_add(&page->lru, &l_inactive);
2068 }
2069
Andrew Mortonb5557492009-01-06 14:40:13 -08002070 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002071 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002072 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002073 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002074 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002075 * Count referenced pages from currently used mappings as rotated,
2076 * even though only some of them are actually re-activated. This
2077 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002078 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002079 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002080 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002081
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002082 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2083 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07002084 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2085 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002086
Johannes Weiner747db952014-08-08 14:19:24 -07002087 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07002088 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089}
2090
Rik van Riel59dc76b2016-05-20 16:56:31 -07002091/*
2092 * The inactive anon list should be small enough that the VM never has
2093 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002094 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002095 * The inactive file list should be small enough to leave most memory
2096 * to the established workingset on the scan-resistant active list,
2097 * but large enough to avoid thrashing the aggregate readahead window.
2098 *
2099 * Both inactive lists should also be large enough that each inactive
2100 * page has a chance to be referenced again before it is reclaimed.
2101 *
2102 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2103 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2104 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2105 *
2106 * total target max
2107 * memory ratio inactive
2108 * -------------------------------------
2109 * 10MB 1 5MB
2110 * 100MB 1 50MB
2111 * 1GB 3 250MB
2112 * 10GB 10 0.9GB
2113 * 100GB 31 3GB
2114 * 1TB 101 10GB
2115 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002116 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002117static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2118 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002119{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002120 unsigned long inactive_ratio;
Michal Hocko71053132017-02-22 15:45:58 -08002121 unsigned long inactive, active;
2122 enum lru_list inactive_lru = file * LRU_FILE;
2123 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002124 unsigned long gb;
2125
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002126 /*
2127 * If we don't have swap space, anonymous page deactivation
2128 * is pointless.
2129 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002130 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002131 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002132
Michal Hocko71053132017-02-22 15:45:58 -08002133 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2134 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002135
Rik van Riel59dc76b2016-05-20 16:56:31 -07002136 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2137 if (gb)
2138 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002139 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002140 inactive_ratio = 1;
2141
2142 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002143}
2144
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002145static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002146 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002147{
Rik van Rielb39415b2009-12-14 17:59:48 -08002148 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002149 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002150 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002151 return 0;
2152 }
2153
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002154 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002155}
2156
Johannes Weiner9a265112013-02-22 16:32:17 -08002157enum scan_balance {
2158 SCAN_EQUAL,
2159 SCAN_FRACT,
2160 SCAN_ANON,
2161 SCAN_FILE,
2162};
2163
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002165 * Determine how aggressively the anon and file LRU lists should be
2166 * scanned. The relative value of each set of LRU lists is determined
2167 * by looking at the fraction of the pages scanned we did rotate back
2168 * onto the active list instead of evict.
2169 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002170 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2171 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002172 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002173static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002174 struct scan_control *sc, unsigned long *nr,
2175 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002176{
Vladimir Davydov33377672016-01-20 15:02:59 -08002177 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002178 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002179 u64 fraction[2];
2180 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002181 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002182 unsigned long anon_prio, file_prio;
2183 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002184 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002185 bool force_scan = false;
2186 unsigned long ap, fp;
2187 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002188 bool some_scanned;
2189 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07002190
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002191 /*
2192 * If the zone or memcg is small, nr[l] can be 0. This
2193 * results in no scanning on this priority and a potential
2194 * priority drop. Global direct reclaim can go to the next
2195 * zone and tends to have no problems. Global kswapd is for
2196 * zone balancing and it needs to scan a minimum amount. When
2197 * reclaiming for a memcg, a priority drop can cause high
2198 * latencies, so it's better to scan a minimum amount there as
2199 * well.
2200 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002201 if (current_is_kswapd()) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002202 if (!pgdat_reclaimable(pgdat))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002203 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002204 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002205 force_scan = true;
2206 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002207 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002208 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002209
2210 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002211 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002212 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002213 goto out;
2214 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002215
Johannes Weiner10316b32013-02-22 16:32:14 -08002216 /*
2217 * Global reclaim will swap to prevent OOM even with no
2218 * swappiness, but memcg users want to use this knob to
2219 * disable swapping for individual groups completely when
2220 * using the memory controller's swap limit feature would be
2221 * too expensive.
2222 */
Johannes Weiner02695172014-08-06 16:06:17 -07002223 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002224 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002225 goto out;
2226 }
2227
2228 /*
2229 * Do not apply any pressure balancing cleverness when the
2230 * system is close to OOM, scan both anon and file equally
2231 * (unless the swappiness setting disagrees with swapping).
2232 */
Johannes Weiner02695172014-08-06 16:06:17 -07002233 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002234 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002235 goto out;
2236 }
2237
Johannes Weiner11d16c22013-02-22 16:32:15 -08002238 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002239 * Prevent the reclaimer from falling into the cache trap: as
2240 * cache pages start out inactive, every cache fault will tip
2241 * the scan balance towards the file LRU. And as the file LRU
2242 * shrinks, so does the window for rotation from references.
2243 * This means we have a runaway feedback loop where a tiny
2244 * thrashing file LRU becomes infinitely more attractive than
2245 * anon pages. Try to detect this based on file LRU size.
2246 */
2247 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002248 unsigned long pgdatfile;
2249 unsigned long pgdatfree;
2250 int z;
2251 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002252
Mel Gorman599d0c92016-07-28 15:45:31 -07002253 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2254 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2255 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002256
Mel Gorman599d0c92016-07-28 15:45:31 -07002257 for (z = 0; z < MAX_NR_ZONES; z++) {
2258 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002259 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002260 continue;
2261
2262 total_high_wmark += high_wmark_pages(zone);
2263 }
2264
2265 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002266 scan_balance = SCAN_ANON;
2267 goto out;
2268 }
2269 }
2270
2271 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002272 * If there is enough inactive page cache, i.e. if the size of the
2273 * inactive list is greater than that of the active list *and* the
2274 * inactive list actually has some pages to scan on this priority, we
2275 * do not reclaim anything from the anonymous working set right now.
2276 * Without the second condition we could end up never scanning an
2277 * lruvec even if it has plenty of old anonymous pages unless the
2278 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002279 */
Patrick Daly72312122017-03-14 18:30:41 -07002280 if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) &&
2281 !inactive_list_is_low(lruvec, true, sc) &&
Michal Hocko521e92b2017-02-22 15:46:01 -08002282 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002283 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002284 goto out;
2285 }
2286
Johannes Weiner9a265112013-02-22 16:32:17 -08002287 scan_balance = SCAN_FRACT;
2288
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002289 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002290 * With swappiness at 100, anonymous and file have the same priority.
2291 * This scanning priority is essentially the inverse of IO cost.
2292 */
Johannes Weiner02695172014-08-06 16:06:17 -07002293 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002294 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002295
2296 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002297 * OK, so we have swap space and a fair amount of page cache
2298 * pages. We use the recently rotated / recently scanned
2299 * ratios to determine how valuable each cache is.
2300 *
2301 * Because workloads change over time (and to avoid overflow)
2302 * we keep these statistics as a floating average, which ends
2303 * up weighing recent references more than old ones.
2304 *
2305 * anon in [0], file in [1]
2306 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002307
Michal Hocko71053132017-02-22 15:45:58 -08002308 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2309 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2310 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2311 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002312
Mel Gorman599d0c92016-07-28 15:45:31 -07002313 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002314 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002315 reclaim_stat->recent_scanned[0] /= 2;
2316 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002317 }
2318
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002319 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002320 reclaim_stat->recent_scanned[1] /= 2;
2321 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002322 }
2323
2324 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002325 * The amount of pressure on anon vs file pages is inversely
2326 * proportional to the fraction of recently scanned pages on
2327 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002328 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002329 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002330 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002331
Satoru Moriyafe350042012-05-29 15:06:47 -07002332 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002333 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002334 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002335
Shaohua Li76a33fc2010-05-24 14:32:36 -07002336 fraction[0] = ap;
2337 fraction[1] = fp;
2338 denominator = ap + fp + 1;
2339out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002340 some_scanned = false;
2341 /* Only use force_scan on second pass. */
2342 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002343 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002344 for_each_evictable_lru(lru) {
2345 int file = is_file_lru(lru);
2346 unsigned long size;
2347 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002348
Michal Hocko521e92b2017-02-22 15:46:01 -08002349 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002350 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002351
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002352 if (!scan && pass && force_scan)
2353 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002354
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002355 switch (scan_balance) {
2356 case SCAN_EQUAL:
2357 /* Scan lists relative to size */
2358 break;
2359 case SCAN_FRACT:
2360 /*
2361 * Scan types proportional to swappiness and
2362 * their relative recent reclaim efficiency.
2363 */
2364 scan = div64_u64(scan * fraction[file],
2365 denominator);
2366 break;
2367 case SCAN_FILE:
2368 case SCAN_ANON:
2369 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002370 if ((scan_balance == SCAN_FILE) != file) {
2371 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002372 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002373 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002374 break;
2375 default:
2376 /* Look ma, no brain */
2377 BUG();
2378 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002379
2380 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002381 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002382
Johannes Weiner9a265112013-02-22 16:32:17 -08002383 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002384 * Skip the second pass and don't force_scan,
2385 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002386 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002387 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002388 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002389 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002390}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002391
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002392/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002393 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002394 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002395static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002396 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002397{
Mel Gormanef8f2322016-07-28 15:46:05 -07002398 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002399 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002400 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002401 unsigned long nr_to_scan;
2402 enum lru_list lru;
2403 unsigned long nr_reclaimed = 0;
2404 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2405 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002406 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002407
Vladimir Davydov33377672016-01-20 15:02:59 -08002408 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002409
Mel Gormane82e0562013-07-03 15:01:44 -07002410 /* Record the original scan target for proportional adjustments later */
2411 memcpy(targets, nr, sizeof(nr));
2412
Mel Gorman1a501902014-06-04 16:10:49 -07002413 /*
2414 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2415 * event that can occur when there is little memory pressure e.g.
2416 * multiple streaming readers/writers. Hence, we do not abort scanning
2417 * when the requested number of pages are reclaimed when scanning at
2418 * DEF_PRIORITY on the assumption that the fact we are direct
2419 * reclaiming implies that kswapd is not keeping up and it is best to
2420 * do a batch of work at once. For memcg reclaim one check is made to
2421 * abort proportional reclaim if either the file or anon lru has already
2422 * dropped to zero at the first pass.
2423 */
2424 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2425 sc->priority == DEF_PRIORITY);
2426
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002427 blk_start_plug(&plug);
2428 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2429 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002430 unsigned long nr_anon, nr_file, percentage;
2431 unsigned long nr_scanned;
2432
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002433 for_each_evictable_lru(lru) {
2434 if (nr[lru]) {
2435 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2436 nr[lru] -= nr_to_scan;
2437
2438 nr_reclaimed += shrink_list(lru, nr_to_scan,
2439 lruvec, sc);
2440 }
2441 }
Mel Gormane82e0562013-07-03 15:01:44 -07002442
Michal Hockobd041732016-12-02 17:26:48 -08002443 cond_resched();
2444
Mel Gormane82e0562013-07-03 15:01:44 -07002445 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2446 continue;
2447
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002448 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002449 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002450 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002451 * proportionally what was requested by get_scan_count(). We
2452 * stop reclaiming one LRU and reduce the amount scanning
2453 * proportional to the original scan target.
2454 */
2455 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2456 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2457
Mel Gorman1a501902014-06-04 16:10:49 -07002458 /*
2459 * It's just vindictive to attack the larger once the smaller
2460 * has gone to zero. And given the way we stop scanning the
2461 * smaller below, this makes sure that we only make one nudge
2462 * towards proportionality once we've got nr_to_reclaim.
2463 */
2464 if (!nr_file || !nr_anon)
2465 break;
2466
Mel Gormane82e0562013-07-03 15:01:44 -07002467 if (nr_file > nr_anon) {
2468 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2469 targets[LRU_ACTIVE_ANON] + 1;
2470 lru = LRU_BASE;
2471 percentage = nr_anon * 100 / scan_target;
2472 } else {
2473 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2474 targets[LRU_ACTIVE_FILE] + 1;
2475 lru = LRU_FILE;
2476 percentage = nr_file * 100 / scan_target;
2477 }
2478
2479 /* Stop scanning the smaller of the LRU */
2480 nr[lru] = 0;
2481 nr[lru + LRU_ACTIVE] = 0;
2482
2483 /*
2484 * Recalculate the other LRU scan count based on its original
2485 * scan target and the percentage scanning already complete
2486 */
2487 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2488 nr_scanned = targets[lru] - nr[lru];
2489 nr[lru] = targets[lru] * (100 - percentage) / 100;
2490 nr[lru] -= min(nr[lru], nr_scanned);
2491
2492 lru += LRU_ACTIVE;
2493 nr_scanned = targets[lru] - nr[lru];
2494 nr[lru] = targets[lru] * (100 - percentage) / 100;
2495 nr[lru] -= min(nr[lru], nr_scanned);
2496
2497 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002498 }
2499 blk_finish_plug(&plug);
2500 sc->nr_reclaimed += nr_reclaimed;
2501
2502 /*
2503 * Even if we did not try to evict anon pages at all, we want to
2504 * rebalance the anon lru active/inactive ratio.
2505 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002506 if (inactive_list_is_low(lruvec, false, sc))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002507 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2508 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002509}
2510
Mel Gorman23b9da52012-05-29 15:06:20 -07002511/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002512static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002513{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002514 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002515 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002516 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002517 return true;
2518
2519 return false;
2520}
2521
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002522/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002523 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2524 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2525 * true if more pages should be reclaimed such that when the page allocator
2526 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2527 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002528 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002529static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002530 unsigned long nr_reclaimed,
2531 unsigned long nr_scanned,
2532 struct scan_control *sc)
2533{
2534 unsigned long pages_for_compaction;
2535 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002536 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002537
2538 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002539 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002540 return false;
2541
Mel Gorman28765922011-02-25 14:44:20 -08002542 /* Consider stopping depending on scan and reclaim activity */
2543 if (sc->gfp_mask & __GFP_REPEAT) {
2544 /*
2545 * For __GFP_REPEAT allocations, stop reclaiming if the
2546 * full LRU list has been scanned and we are still failing
2547 * to reclaim pages. This full LRU scan is potentially
2548 * expensive but a __GFP_REPEAT caller really wants to succeed
2549 */
2550 if (!nr_reclaimed && !nr_scanned)
2551 return false;
2552 } else {
2553 /*
2554 * For non-__GFP_REPEAT allocations which can presumably
2555 * fail without consequence, stop if we failed to reclaim
2556 * any pages from the last SWAP_CLUSTER_MAX number of
2557 * pages that were scanned. This will return to the
2558 * caller faster at the risk reclaim/compaction and
2559 * the resulting allocation attempt fails
2560 */
2561 if (!nr_reclaimed)
2562 return false;
2563 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002564
2565 /*
2566 * If we have not reclaimed enough pages for compaction and the
2567 * inactive lists are large enough, continue reclaiming
2568 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002569 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002570 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002571 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002572 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002573 if (sc->nr_reclaimed < pages_for_compaction &&
2574 inactive_lru_pages > pages_for_compaction)
2575 return true;
2576
2577 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002578 for (z = 0; z <= sc->reclaim_idx; z++) {
2579 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002580 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002581 continue;
2582
2583 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002584 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002585 case COMPACT_CONTINUE:
2586 return false;
2587 default:
2588 /* check next zone */
2589 ;
2590 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002591 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002592 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002593}
2594
Mel Gorman970a39a2016-07-28 15:46:35 -07002595static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002596{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002597 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002598 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002599 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002600
Johannes Weiner56600482012-01-12 17:17:59 -08002601 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002602 struct mem_cgroup *root = sc->target_mem_cgroup;
2603 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002604 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002605 .priority = sc->priority,
2606 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002607 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002608 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002609
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002610 nr_reclaimed = sc->nr_reclaimed;
2611 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002612
Andrew Morton694fbc02013-09-24 15:27:37 -07002613 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2614 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002615 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002616 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002617 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002618
Johannes Weiner241994e2015-02-11 15:26:06 -08002619 if (mem_cgroup_low(root, memcg)) {
2620 if (!sc->may_thrash)
2621 continue;
2622 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2623 }
2624
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002625 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002626 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002627
Mel Gormana9dd0a82016-07-28 15:46:02 -07002628 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2629 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002630
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002631 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002632 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002633 memcg, sc->nr_scanned - scanned,
2634 lru_pages);
2635
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002636 /* Record the group's reclaim efficiency */
2637 vmpressure(sc->gfp_mask, memcg, false,
2638 sc->nr_scanned - scanned,
2639 sc->nr_reclaimed - reclaimed);
2640
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002641 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002642 * Direct reclaim and kswapd have to scan all memory
2643 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002644 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002645 *
2646 * Limit reclaim, on the other hand, only cares about
2647 * nr_to_reclaim pages to be reclaimed and it will
2648 * retry with decreasing priority if one round over the
2649 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002650 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002651 if (!global_reclaim(sc) &&
2652 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002653 mem_cgroup_iter_break(root, memcg);
2654 break;
2655 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002656 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002657
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002658 /*
2659 * Shrink the slab caches in the same proportion that
2660 * the eligible LRU pages were scanned.
2661 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002662 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002663 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002664 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002665 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002666
Vinayak Menon57e6f442017-02-19 20:00:31 +11002667 /*
2668 * Record the subtree's reclaim efficiency. The reclaimed
2669 * pages from slab is excluded here because the corresponding
2670 * scanned pages is not accounted. Moreover, freeing a page
2671 * by slab shrinking depends on each slab's object population,
2672 * making the cost model (i.e. scan:free) different from that
2673 * of LRU.
2674 */
2675 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2676 sc->nr_scanned - nr_scanned,
2677 sc->nr_reclaimed - nr_reclaimed);
2678
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002679 if (reclaim_state) {
2680 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2681 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002682 }
2683
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002684 if (sc->nr_reclaimed - nr_reclaimed)
2685 reclaimable = true;
2686
Mel Gormana9dd0a82016-07-28 15:46:02 -07002687 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002688 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002689
2690 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002691}
2692
Vlastimil Babka53853e22014-10-09 15:27:02 -07002693/*
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002694 * Returns true if compaction should go ahead for a costly-order request, or
2695 * the allocation would already succeed without compaction. Return false if we
2696 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002697 */
Mel Gorman4f588332016-07-28 15:46:38 -07002698static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002699{
Mel Gorman31483b62016-07-28 15:45:46 -07002700 unsigned long watermark;
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002701 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002702
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002703 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2704 if (suitable == COMPACT_SUCCESS)
2705 /* Allocation should succeed already. Don't reclaim. */
2706 return true;
2707 if (suitable == COMPACT_SKIPPED)
2708 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002709 return false;
2710
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002711 /*
2712 * Compaction is already possible, but it takes time to run and there
2713 * are potentially other callers using the pages just freed. So proceed
2714 * with reclaim to make a buffer of free pages available to give
2715 * compaction a reasonable chance of completing and allocating the page.
2716 * Note that we won't actually reclaim the whole buffer in one attempt
2717 * as the target watermark in should_continue_reclaim() is lower. But if
2718 * we are already above the high+gap watermark, don't reclaim at all.
2719 */
2720 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2721
2722 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002723}
2724
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725/*
2726 * This is the direct reclaim path, for page-allocating processes. We only
2727 * try to reclaim pages from zones which will satisfy the caller's allocation
2728 * request.
2729 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 * If a zone is deemed to be full of pinned pages then just give it a light
2731 * scan then give up on it.
2732 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002733static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734{
Mel Gormandd1a2392008-04-28 02:12:17 -07002735 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002736 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002737 unsigned long nr_soft_reclaimed;
2738 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002739 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002740 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002741
Mel Gormancc715d92012-03-21 16:34:00 -07002742 /*
2743 * If the number of buffer_heads in the machine exceeds the maximum
2744 * allowed level, force direct reclaim to scan the highmem zone as
2745 * highmem pages could be pinning lowmem pages storing buffer_heads
2746 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002747 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002748 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002749 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002750 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002751 }
Mel Gormancc715d92012-03-21 16:34:00 -07002752
Mel Gormand4debc62010-08-09 17:19:29 -07002753 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002754 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002755 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002756 * Take care memory controller reclaiming has small influence
2757 * to global LRU.
2758 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002759 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002760 if (!cpuset_zone_allowed(zone,
2761 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002762 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002763
Lisa Du6e543d52013-09-11 14:22:36 -07002764 if (sc->priority != DEF_PRIORITY &&
Mel Gorman599d0c92016-07-28 15:45:31 -07002765 !pgdat_reclaimable(zone->zone_pgdat))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002766 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002767
2768 /*
2769 * If we already have plenty of memory free for
2770 * compaction in this zone, don't free any more.
2771 * Even though compaction is invoked for any
2772 * non-zero order, only frequent costly order
2773 * reclamation is disruptive enough to become a
2774 * noticeable problem, like transparent huge
2775 * page allocations.
2776 */
2777 if (IS_ENABLED(CONFIG_COMPACTION) &&
2778 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002779 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002780 sc->compaction_ready = true;
2781 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002782 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002783
Andrew Morton0608f432013-09-24 15:27:41 -07002784 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002785 * Shrink each node in the zonelist once. If the
2786 * zonelist is ordered by zone (not the default) then a
2787 * node may be shrunk multiple times but in that case
2788 * the user prefers lower zones being preserved.
2789 */
2790 if (zone->zone_pgdat == last_pgdat)
2791 continue;
2792
2793 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002794 * This steals pages from memory cgroups over softlimit
2795 * and returns the number of reclaimed pages and
2796 * scanned pages. This works for global memory pressure
2797 * and balancing, not for a memcg's limit.
2798 */
2799 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002800 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002801 sc->order, sc->gfp_mask,
2802 &nr_soft_scanned);
2803 sc->nr_reclaimed += nr_soft_reclaimed;
2804 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002805 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002806 }
Nick Piggin408d8542006-09-25 23:31:27 -07002807
Mel Gorman79dafcd2016-07-28 15:45:53 -07002808 /* See comment about same check for global reclaim above */
2809 if (zone->zone_pgdat == last_pgdat)
2810 continue;
2811 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002812 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 }
Mel Gormane0c23272011-10-31 17:09:33 -07002814
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002815 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002816 * Restore to original mask to avoid the impact on the caller if we
2817 * promoted it to __GFP_HIGHMEM.
2818 */
2819 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002821
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822/*
2823 * This is the main entry point to direct page reclaim.
2824 *
2825 * If a full scan of the inactive list fails to free enough memory then we
2826 * are "out of memory" and something needs to be killed.
2827 *
2828 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2829 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002830 * caller can't do much about. We kick the writeback threads and take explicit
2831 * naps in the hope that some of these pages can be written. But if the
2832 * allocating task holds filesystem locks which prevent writeout this might not
2833 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002834 *
2835 * returns: 0, if no pages reclaimed
2836 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 */
Mel Gormandac1d272008-04-28 02:12:12 -07002838static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002839 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840{
Johannes Weiner241994e2015-02-11 15:26:06 -08002841 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002842 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002843 unsigned long writeback_threshold;
Johannes Weiner241994e2015-02-11 15:26:06 -08002844retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002845 delayacct_freepages_start();
2846
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002847 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002848 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002850 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002851 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2852 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002853 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002854 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002855
Balbir Singh66e17072008-02-07 00:13:56 -08002856 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002857 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002858 break;
2859
2860 if (sc->compaction_ready)
2861 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862
2863 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002864 * If we're getting trouble reclaiming, start doing
2865 * writepage even in laptop mode.
2866 */
2867 if (sc->priority < DEF_PRIORITY - 2)
2868 sc->may_writepage = 1;
2869
2870 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 * Try to write back as many pages as we just scanned. This
2872 * tends to cause slow streaming writers to write data to the
2873 * disk smoothly, at the dirtying rate, which is nice. But
2874 * that's undesirable in laptop mode, where we *want* lumpy
2875 * writeout. So in laptop mode, write out the whole world.
2876 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002877 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2878 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002879 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2880 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002881 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002883 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002884
Keika Kobayashi873b4772008-07-25 01:48:52 -07002885 delayacct_freepages_end();
2886
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002887 if (sc->nr_reclaimed)
2888 return sc->nr_reclaimed;
2889
Mel Gorman0cee34f2012-01-12 17:19:49 -08002890 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002891 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002892 return 1;
2893
Johannes Weiner241994e2015-02-11 15:26:06 -08002894 /* Untapped cgroup reserves? Don't OOM, retry. */
2895 if (!sc->may_thrash) {
2896 sc->priority = initial_priority;
2897 sc->may_thrash = 1;
2898 goto retry;
2899 }
2900
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002901 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902}
2903
Mel Gorman55150612012-07-31 16:44:35 -07002904static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2905{
2906 struct zone *zone;
2907 unsigned long pfmemalloc_reserve = 0;
2908 unsigned long free_pages = 0;
2909 int i;
2910 bool wmark_ok;
2911
2912 for (i = 0; i <= ZONE_NORMAL; i++) {
2913 zone = &pgdat->node_zones[i];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002914 if (!managed_zone(zone) ||
Mel Gorman599d0c92016-07-28 15:45:31 -07002915 pgdat_reclaimable_pages(pgdat) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002916 continue;
2917
Mel Gorman55150612012-07-31 16:44:35 -07002918 pfmemalloc_reserve += min_wmark_pages(zone);
2919 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2920 }
2921
Mel Gorman675becc2014-06-04 16:07:35 -07002922 /* If there are no reserves (unexpected config) then do not throttle */
2923 if (!pfmemalloc_reserve)
2924 return true;
2925
Mel Gorman55150612012-07-31 16:44:35 -07002926 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2927
2928 /* kswapd must be awake if processes are being throttled */
2929 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002930 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002931 (enum zone_type)ZONE_NORMAL);
2932 wake_up_interruptible(&pgdat->kswapd_wait);
2933 }
2934
2935 return wmark_ok;
2936}
2937
2938/*
2939 * Throttle direct reclaimers if backing storage is backed by the network
2940 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2941 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002942 * when the low watermark is reached.
2943 *
2944 * Returns true if a fatal signal was delivered during throttling. If this
2945 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002946 */
Mel Gorman50694c22012-11-26 16:29:48 -08002947static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002948 nodemask_t *nodemask)
2949{
Mel Gorman675becc2014-06-04 16:07:35 -07002950 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002951 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002952 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002953
2954 /*
2955 * Kernel threads should not be throttled as they may be indirectly
2956 * responsible for cleaning pages necessary for reclaim to make forward
2957 * progress. kjournald for example may enter direct reclaim while
2958 * committing a transaction where throttling it could forcing other
2959 * processes to block on log_wait_commit().
2960 */
2961 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002962 goto out;
2963
2964 /*
2965 * If a fatal signal is pending, this process should not throttle.
2966 * It should return quickly so it can exit and free its memory
2967 */
2968 if (fatal_signal_pending(current))
2969 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002970
Mel Gorman675becc2014-06-04 16:07:35 -07002971 /*
2972 * Check if the pfmemalloc reserves are ok by finding the first node
2973 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2974 * GFP_KERNEL will be required for allocating network buffers when
2975 * swapping over the network so ZONE_HIGHMEM is unusable.
2976 *
2977 * Throttling is based on the first usable node and throttled processes
2978 * wait on a queue until kswapd makes progress and wakes them. There
2979 * is an affinity then between processes waking up and where reclaim
2980 * progress has been made assuming the process wakes on the same node.
2981 * More importantly, processes running on remote nodes will not compete
2982 * for remote pfmemalloc reserves and processes on different nodes
2983 * should make reasonable progress.
2984 */
2985 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002986 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002987 if (zone_idx(zone) > ZONE_NORMAL)
2988 continue;
2989
2990 /* Throttle based on the first usable node */
2991 pgdat = zone->zone_pgdat;
2992 if (pfmemalloc_watermark_ok(pgdat))
2993 goto out;
2994 break;
2995 }
2996
2997 /* If no zone was usable by the allocation flags then do not throttle */
2998 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002999 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003000
Mel Gorman68243e72012-07-31 16:44:39 -07003001 /* Account for the throttling */
3002 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3003
Mel Gorman55150612012-07-31 16:44:35 -07003004 /*
3005 * If the caller cannot enter the filesystem, it's possible that it
3006 * is due to the caller holding an FS lock or performing a journal
3007 * transaction in the case of a filesystem like ext[3|4]. In this case,
3008 * it is not safe to block on pfmemalloc_wait as kswapd could be
3009 * blocked waiting on the same lock. Instead, throttle for up to a
3010 * second before continuing.
3011 */
3012 if (!(gfp_mask & __GFP_FS)) {
3013 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3014 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003015
3016 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003017 }
3018
3019 /* Throttle until kswapd wakes the process */
3020 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3021 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003022
3023check_pending:
3024 if (fatal_signal_pending(current))
3025 return true;
3026
3027out:
3028 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003029}
3030
Mel Gormandac1d272008-04-28 02:12:12 -07003031unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003032 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003033{
Mel Gorman33906bc2010-08-09 17:19:16 -07003034 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003035 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003036 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003037 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07003038 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003039 .order = order,
3040 .nodemask = nodemask,
3041 .priority = DEF_PRIORITY,
3042 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003043 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003044 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003045 };
3046
Mel Gorman55150612012-07-31 16:44:35 -07003047 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003048 * Do not enter reclaim if fatal signal was delivered while throttled.
3049 * 1 is returned so that the page allocator does not OOM kill at this
3050 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003051 */
Mel Gorman50694c22012-11-26 16:29:48 -08003052 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003053 return 1;
3054
Mel Gorman33906bc2010-08-09 17:19:16 -07003055 trace_mm_vmscan_direct_reclaim_begin(order,
3056 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003057 gfp_mask,
3058 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003059
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003060 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003061
3062 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3063
3064 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003065}
3066
Andrew Mortonc255a452012-07-31 16:43:02 -07003067#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003068
Mel Gormana9dd0a82016-07-28 15:46:02 -07003069unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003070 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003071 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003072 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003073{
3074 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003075 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003076 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003077 .may_writepage = !laptop_mode,
3078 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003079 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003080 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003081 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003082 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003083
Balbir Singh4e416952009-09-23 15:56:39 -07003084 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3085 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003086
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003087 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003088 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003089 sc.gfp_mask,
3090 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003091
Balbir Singh4e416952009-09-23 15:56:39 -07003092 /*
3093 * NOTE: Although we can get the priority field, using it
3094 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003095 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003096 * will pick up pages from other mem cgroup's as well. We hack
3097 * the priority and make it zero.
3098 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003099 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003100
3101 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3102
Ying Han0ae5e892011-05-26 16:25:25 -07003103 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003104 return sc.nr_reclaimed;
3105}
3106
Johannes Weiner72835c82012-01-12 17:18:32 -08003107unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003108 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003109 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003110 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003111{
Balbir Singh4e416952009-09-23 15:56:39 -07003112 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003113 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003114 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003115 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003116 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003117 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3118 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003119 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003120 .target_mem_cgroup = memcg,
3121 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003122 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003123 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003124 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003125 };
Balbir Singh66e17072008-02-07 00:13:56 -08003126
Ying Han889976d2011-05-26 16:25:33 -07003127 /*
3128 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3129 * take care of from where we get pages. So the node where we start the
3130 * scan does not need to be the current node.
3131 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003132 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003133
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003134 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003135
3136 trace_mm_vmscan_memcg_reclaim_begin(0,
3137 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003138 sc.gfp_mask,
3139 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003140
Johannes Weiner89a28482016-10-27 17:46:56 -07003141 current->flags |= PF_MEMALLOC;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003142 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weiner89a28482016-10-27 17:46:56 -07003143 current->flags &= ~PF_MEMALLOC;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003144
3145 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3146
3147 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003148}
3149#endif
3150
Mel Gorman1d82de62016-07-28 15:45:43 -07003151static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003152 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003153{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003154 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003155
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003156 if (!total_swap_pages)
3157 return;
3158
3159 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3160 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003161 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003162
Mel Gormanf8d1a312016-07-28 15:47:34 -07003163 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003164 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003165 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003166
3167 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3168 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003169}
3170
Mel Gorman31483b62016-07-28 15:45:46 -07003171static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003172{
Mel Gorman31483b62016-07-28 15:45:46 -07003173 unsigned long mark = high_wmark_pages(zone);
Johannes Weiner60cefed2012-11-29 13:54:23 -08003174
Mel Gorman6256c6b42016-07-28 15:45:56 -07003175 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3176 return false;
3177
3178 /*
3179 * If any eligible zone is balanced then the node is not considered
3180 * to be congested or dirty
3181 */
3182 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3183 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3184
3185 return true;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003186}
3187
Mel Gorman1741c872011-01-13 15:46:21 -08003188/*
Mel Gorman55150612012-07-31 16:44:35 -07003189 * Prepare kswapd for sleeping. This verifies that there are no processes
3190 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3191 *
3192 * Returns true if kswapd is ready to sleep
3193 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003194static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003195{
Mel Gorman1d82de62016-07-28 15:45:43 -07003196 int i;
3197
Mel Gorman55150612012-07-31 16:44:35 -07003198 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003199 * The throttled processes are normally woken up in balance_pgdat() as
3200 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3201 * race between when kswapd checks the watermarks and a process gets
3202 * throttled. There is also a potential race if processes get
3203 * throttled, kswapd wakes, a large process exits thereby balancing the
3204 * zones, which causes kswapd to exit balance_pgdat() before reaching
3205 * the wake up checks. If kswapd is going to sleep, no process should
3206 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3207 * the wake up is premature, processes will wake kswapd and get
3208 * throttled again. The difference from wake ups in balance_pgdat() is
3209 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003210 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003211 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3212 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003213
Mel Gorman1d82de62016-07-28 15:45:43 -07003214 for (i = 0; i <= classzone_idx; i++) {
3215 struct zone *zone = pgdat->node_zones + i;
3216
Mel Gorman6aa303d2016-09-01 16:14:55 -07003217 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003218 continue;
3219
Mel Gorman38087d92016-07-28 15:45:49 -07003220 if (!zone_balanced(zone, order, classzone_idx))
3221 return false;
Mel Gorman1d82de62016-07-28 15:45:43 -07003222 }
3223
Mel Gorman38087d92016-07-28 15:45:49 -07003224 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003225}
3226
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003228 * kswapd shrinks a node of pages that are at or below the highest usable
3229 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003230 *
3231 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003232 * reclaim or if the lack of progress was due to pages under writeback.
3233 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003234 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003235static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003236 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003237{
Mel Gorman1d82de62016-07-28 15:45:43 -07003238 struct zone *zone;
3239 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003240
Mel Gorman1d82de62016-07-28 15:45:43 -07003241 /* Reclaim a number of pages proportional to the number of zones */
3242 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003243 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003244 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003245 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003246 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003247
Mel Gorman1d82de62016-07-28 15:45:43 -07003248 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003249 }
3250
Mel Gorman1d82de62016-07-28 15:45:43 -07003251 /*
3252 * Historically care was taken to put equal pressure on all zones but
3253 * now pressure is applied based on node LRU order.
3254 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003255 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003256
3257 /*
3258 * Fragmentation may mean that the system cannot be rebalanced for
3259 * high-order allocations. If twice the allocation size has been
3260 * reclaimed then recheck watermarks only at order-0 to prevent
3261 * excessive reclaim. Assume that a process requested a high-order
3262 * can direct reclaim/compact.
3263 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003264 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003265 sc->order = 0;
3266
Mel Gormanb8e83b92013-07-03 15:01:45 -07003267 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003268}
3269
3270/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003271 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3272 * that are eligible for use by the caller until at least one zone is
3273 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003275 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 *
3277 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003278 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003279 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3280 * or lower is eligible for reclaim until at least one usable zone is
3281 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003283static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003286 unsigned long nr_soft_reclaimed;
3287 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003288 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003289 struct scan_control sc = {
3290 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003291 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003292 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003293 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003294 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003295 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003296 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003297 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003299 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003300 bool raise_priority = true;
3301
3302 sc.nr_reclaimed = 0;
Mel Gorman84c7a772016-07-28 15:46:44 -07003303 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304
Mel Gorman86c79f62016-07-28 15:45:59 -07003305 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003306 * If the number of buffer_heads exceeds the maximum allowed
3307 * then consider reclaiming from all zones. This has a dual
3308 * purpose -- on 64-bit systems it is expected that
3309 * buffer_heads are stripped during active rotation. On 32-bit
3310 * systems, highmem pages can pin lowmem memory and shrinking
3311 * buffers can relieve lowmem pressure. Reclaim may still not
3312 * go ahead if all eligible zones for the original allocation
3313 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003314 */
3315 if (buffer_heads_over_limit) {
3316 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3317 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003318 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003319 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320
Mel Gorman970a39a2016-07-28 15:46:35 -07003321 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003322 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003325
Mel Gorman86c79f62016-07-28 15:45:59 -07003326 /*
3327 * Only reclaim if there are no eligible zones. Check from
3328 * high to low zone as allocations prefer higher zones.
3329 * Scanning from low to high zone would allow congestion to be
3330 * cleared during a very small window when a small low
3331 * zone was balanced even under extreme pressure when the
Mel Gorman84c7a772016-07-28 15:46:44 -07003332 * overall node may be congested. Note that sc.reclaim_idx
3333 * is not used as buffer_heads_over_limit may have adjusted
3334 * it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003335 */
Mel Gorman84c7a772016-07-28 15:46:44 -07003336 for (i = classzone_idx; i >= 0; i--) {
Mel Gorman86c79f62016-07-28 15:45:59 -07003337 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003338 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003339 continue;
3340
Mel Gorman84c7a772016-07-28 15:46:44 -07003341 if (zone_balanced(zone, sc.order, classzone_idx))
Mel Gorman86c79f62016-07-28 15:45:59 -07003342 goto out;
3343 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003344
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003346 * Do some background aging of the anon list, to give
3347 * pages a chance to be referenced before reclaiming. All
3348 * pages are rotated regardless of classzone as this is
3349 * about consistent aging.
3350 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003351 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003352
3353 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003354 * If we're getting trouble reclaiming, start doing writepage
3355 * even in laptop mode.
3356 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003357 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003358 sc.may_writepage = 1;
3359
Mel Gorman1d82de62016-07-28 15:45:43 -07003360 /* Call soft limit reclaim before calling shrink_node. */
3361 sc.nr_scanned = 0;
3362 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003363 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003364 sc.gfp_mask, &nr_soft_scanned);
3365 sc.nr_reclaimed += nr_soft_reclaimed;
3366
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003367 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003368 * There should be no need to raise the scanning priority if
3369 * enough pages are already being scanned that that high
3370 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003372 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003373 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003374
3375 /*
3376 * If the low watermark is met there is no need for processes
3377 * to be throttled on pfmemalloc_wait as they should not be
3378 * able to safely make forward progress. Wake them
3379 */
3380 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3381 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003382 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003383
Mel Gormanb8e83b92013-07-03 15:01:45 -07003384 /* Check if kswapd should be suspending */
3385 if (try_to_freeze() || kthread_should_stop())
3386 break;
3387
3388 /*
3389 * Raise priority if scanning rate is too low or there was no
3390 * progress in reclaiming pages
3391 */
3392 if (raise_priority || !sc.nr_reclaimed)
3393 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003394 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395
Mel Gormanb8e83b92013-07-03 15:01:45 -07003396out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003397 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003398 * Return the order kswapd stopped reclaiming at as
3399 * prepare_kswapd_sleep() takes it into account. If another caller
3400 * entered the allocator slow path while kswapd was awake, order will
3401 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003402 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003403 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404}
3405
Mel Gorman38087d92016-07-28 15:45:49 -07003406static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3407 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003408{
3409 long remaining = 0;
3410 DEFINE_WAIT(wait);
3411
3412 if (freezing(current) || kthread_should_stop())
3413 return;
3414
3415 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3416
3417 /* Try to sleep for a short interval */
Mel Gormand9f21d42016-07-28 15:46:41 -07003418 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003419 /*
3420 * Compaction records what page blocks it recently failed to
3421 * isolate pages from and skips them in the future scanning.
3422 * When kswapd is going to sleep, it is reasonable to assume
3423 * that pages and compaction may succeed so reset the cache.
3424 */
3425 reset_isolation_suitable(pgdat);
3426
3427 /*
3428 * We have freed the memory, now we should compact it to make
3429 * allocation of the requested order possible.
3430 */
Mel Gorman38087d92016-07-28 15:45:49 -07003431 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003432
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003433 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003434
3435 /*
3436 * If woken prematurely then reset kswapd_classzone_idx and
3437 * order. The values will either be from a wakeup request or
3438 * the previous request that slept prematurely.
3439 */
3440 if (remaining) {
3441 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3442 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3443 }
3444
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003445 finish_wait(&pgdat->kswapd_wait, &wait);
3446 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3447 }
3448
3449 /*
3450 * After a short sleep, check if it was a premature sleep. If not, then
3451 * go fully to sleep until explicitly woken up.
3452 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003453 if (!remaining &&
3454 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003455 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3456
3457 /*
3458 * vmstat counters are not perfectly accurate and the estimated
3459 * value for counters such as NR_FREE_PAGES can deviate from the
3460 * true value by nr_online_cpus * threshold. To avoid the zone
3461 * watermarks being breached while under pressure, we reduce the
3462 * per-cpu vmstat threshold while kswapd is awake and restore
3463 * them before going back to sleep.
3464 */
3465 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003466
3467 if (!kthread_should_stop())
3468 schedule();
3469
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003470 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3471 } else {
3472 if (remaining)
3473 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3474 else
3475 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3476 }
3477 finish_wait(&pgdat->kswapd_wait, &wait);
3478}
3479
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480/*
3481 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003482 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 *
3484 * This basically trickles out pages so that we have _some_
3485 * free memory available even if there is no other activity
3486 * that frees anything up. This is needed for things like routing
3487 * etc, where we otherwise might have all activity going on in
3488 * asynchronous contexts that cannot page things out.
3489 *
3490 * If there are applications that are active memory-allocators
3491 * (most normal use), this basically shouldn't matter.
3492 */
3493static int kswapd(void *p)
3494{
Mel Gorman38087d92016-07-28 15:45:49 -07003495 unsigned int alloc_order, reclaim_order, classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 pg_data_t *pgdat = (pg_data_t*)p;
3497 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003498
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 struct reclaim_state reclaim_state = {
3500 .reclaimed_slab = 0,
3501 };
Rusty Russella70f7302009-03-13 14:49:46 +10303502 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
Nick Piggincf40bd12009-01-21 08:12:39 +01003504 lockdep_set_current_reclaim_state(GFP_KERNEL);
3505
Rusty Russell174596a2009-01-01 10:12:29 +10303506 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003507 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 current->reclaim_state = &reclaim_state;
3509
3510 /*
3511 * Tell the memory management that we're a "memory allocator",
3512 * and that if we need more memory we should get access to it
3513 * regardless (see "__alloc_pages()"). "kswapd" should
3514 * never get caught in the normal page freeing logic.
3515 *
3516 * (Kswapd normally doesn't need memory anyway, but sometimes
3517 * you need a small amount of memory in order to be able to
3518 * page out something else, and this flag essentially protects
3519 * us from recursively trying to free more memory as we're
3520 * trying to free the first piece of memory in the first place).
3521 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003522 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003523 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
Mel Gorman38087d92016-07-28 15:45:49 -07003525 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3526 pgdat->kswapd_classzone_idx = classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003528 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003529
Mel Gorman38087d92016-07-28 15:45:49 -07003530kswapd_try_sleep:
3531 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3532 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003533
Mel Gorman38087d92016-07-28 15:45:49 -07003534 /* Read the new order and classzone_idx */
3535 alloc_order = reclaim_order = pgdat->kswapd_order;
3536 classzone_idx = pgdat->kswapd_classzone_idx;
3537 pgdat->kswapd_order = 0;
3538 pgdat->kswapd_classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539
David Rientjes8fe23e02009-12-14 17:58:33 -08003540 ret = try_to_freeze();
3541 if (kthread_should_stop())
3542 break;
3543
3544 /*
3545 * We can speed up thawing tasks if we don't call balance_pgdat
3546 * after returning from the refrigerator
3547 */
Mel Gorman38087d92016-07-28 15:45:49 -07003548 if (ret)
3549 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003550
Mel Gorman38087d92016-07-28 15:45:49 -07003551 /*
3552 * Reclaim begins at the requested order but if a high-order
3553 * reclaim fails then kswapd falls back to reclaiming for
3554 * order-0. If that happens, kswapd will consider sleeping
3555 * for the order it finished reclaiming at (reclaim_order)
3556 * but kcompactd is woken to compact for the original
3557 * request (alloc_order).
3558 */
Mel Gormane5146b12016-07-28 15:46:47 -07003559 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3560 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003561 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3562 if (reclaim_order < alloc_order)
3563 goto kswapd_try_sleep;
3564
3565 alloc_order = reclaim_order = pgdat->kswapd_order;
3566 classzone_idx = pgdat->kswapd_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003568
Johannes Weiner71abdc12014-06-06 14:35:35 -07003569 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003570 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003571 lockdep_clear_current_reclaim_state();
3572
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 return 0;
3574}
3575
3576/*
3577 * A zone is low on free memory, so wake its kswapd task to service it.
3578 */
Mel Gorman99504742011-01-13 15:46:20 -08003579void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580{
3581 pg_data_t *pgdat;
Mel Gormane1a55632016-07-28 15:46:26 -07003582 int z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Mel Gorman6aa303d2016-09-01 16:14:55 -07003584 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 return;
3586
Vladimir Davydov344736f2014-10-20 15:50:30 +04003587 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003589 pgdat = zone->zone_pgdat;
Mel Gorman38087d92016-07-28 15:45:49 -07003590 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3591 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003592 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003594
3595 /* Only wake kswapd if all zones are unbalanced */
3596 for (z = 0; z <= classzone_idx; z++) {
3597 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003598 if (!managed_zone(zone))
Mel Gormane1a55632016-07-28 15:46:26 -07003599 continue;
3600
3601 if (zone_balanced(zone, order, classzone_idx))
3602 return;
3603 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08003604
3605 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003606 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607}
3608
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003609#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003611 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003612 * freed pages.
3613 *
3614 * Rather than trying to age LRUs the aim is to preserve the overall
3615 * LRU order by reclaiming preferentially
3616 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003618unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003620 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003621 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003622 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003623 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003624 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003625 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003626 .may_writepage = 1,
3627 .may_unmap = 1,
3628 .may_swap = 1,
3629 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003631 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003632 struct task_struct *p = current;
3633 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003635 p->flags |= PF_MEMALLOC;
3636 lockdep_set_current_reclaim_state(sc.gfp_mask);
3637 reclaim_state.reclaimed_slab = 0;
3638 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003639
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003640 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003641
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003642 p->reclaim_state = NULL;
3643 lockdep_clear_current_reclaim_state();
3644 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003645
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003646 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003648#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650/* It's optimal to keep kswapds on the same CPUs as their memory, but
3651 not required for correctness. So if the last cpu in a node goes
3652 away, we get changed to run anywhere: as the first one comes back,
3653 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003654static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3655 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003657 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003659 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003660 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003661 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303662 const struct cpumask *mask;
3663
3664 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003665
Rusty Russell3e597942009-01-01 10:12:24 +10303666 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003668 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 }
3670 }
3671 return NOTIFY_OK;
3672}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673
Yasunori Goto3218ae12006-06-27 02:53:33 -07003674/*
3675 * This kswapd start function will be called by init and node-hot-add.
3676 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3677 */
3678int kswapd_run(int nid)
3679{
3680 pg_data_t *pgdat = NODE_DATA(nid);
3681 int ret = 0;
3682
3683 if (pgdat->kswapd)
3684 return 0;
3685
3686 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3687 if (IS_ERR(pgdat->kswapd)) {
3688 /* failure at boot is fatal */
3689 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003690 pr_err("Failed to start kswapd on node %d\n", nid);
3691 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003692 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003693 }
3694 return ret;
3695}
3696
David Rientjes8fe23e02009-12-14 17:58:33 -08003697/*
Jiang Liud8adde12012-07-11 14:01:52 -07003698 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003699 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003700 */
3701void kswapd_stop(int nid)
3702{
3703 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3704
Jiang Liud8adde12012-07-11 14:01:52 -07003705 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003706 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003707 NODE_DATA(nid)->kswapd = NULL;
3708 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003709}
3710
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711static int __init kswapd_init(void)
3712{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003713 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003714
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003716 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003717 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 hotcpu_notifier(cpu_callback, 0);
3719 return 0;
3720}
3721
3722module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003723
3724#ifdef CONFIG_NUMA
3725/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003726 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003727 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003728 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003729 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003730 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003731int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003732
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003733#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003734#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003735#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003736#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003737
Christoph Lameter9eeff232006-01-18 17:42:31 -08003738/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003739 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003740 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3741 * a zone.
3742 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003743#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003744
Christoph Lameter9eeff232006-01-18 17:42:31 -08003745/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003746 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003747 * occur.
3748 */
3749int sysctl_min_unmapped_ratio = 1;
3750
3751/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003752 * If the number of slab pages in a zone grows beyond this percentage then
3753 * slab reclaim needs to occur.
3754 */
3755int sysctl_min_slab_ratio = 5;
3756
Mel Gorman11fb9982016-07-28 15:46:20 -07003757static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003758{
Mel Gorman11fb9982016-07-28 15:46:20 -07003759 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3760 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3761 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003762
3763 /*
3764 * It's possible for there to be more file mapped pages than
3765 * accounted for by the pages on the file LRU lists because
3766 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3767 */
3768 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3769}
3770
3771/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003772static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003773{
Alexandru Moised031a152015-11-05 18:48:08 -08003774 unsigned long nr_pagecache_reclaimable;
3775 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003776
3777 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003778 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003779 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003780 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003781 * a better estimate
3782 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003783 if (node_reclaim_mode & RECLAIM_UNMAP)
3784 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003785 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003786 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003787
3788 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003789 if (!(node_reclaim_mode & RECLAIM_WRITE))
3790 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003791
3792 /* Watch for any possible underflows due to delta */
3793 if (unlikely(delta > nr_pagecache_reclaimable))
3794 delta = nr_pagecache_reclaimable;
3795
3796 return nr_pagecache_reclaimable - delta;
3797}
3798
Christoph Lameter0ff38492006-09-25 23:31:52 -07003799/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003800 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003801 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003802static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003803{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003804 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003805 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003806 struct task_struct *p = current;
3807 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003808 int classzone_idx = gfp_zone(gfp_mask);
Andrew Morton179e9632006-03-22 00:08:18 -08003809 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003810 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003811 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003812 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003813 .priority = NODE_RECLAIM_PRIORITY,
3814 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3815 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003816 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003817 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003818 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003819
Christoph Lameter9eeff232006-01-18 17:42:31 -08003820 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003821 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003822 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003823 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003824 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003825 */
3826 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003827 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003828 reclaim_state.reclaimed_slab = 0;
3829 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003830
Mel Gormana5f5f912016-07-28 15:46:32 -07003831 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003832 /*
3833 * Free memory by calling shrink zone with increasing
3834 * priorities until we have enough memory freed.
3835 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003836 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003837 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003838 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003839 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003840
Christoph Lameter9eeff232006-01-18 17:42:31 -08003841 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003842 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003843 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003844 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003845}
Andrew Morton179e9632006-03-22 00:08:18 -08003846
Mel Gormana5f5f912016-07-28 15:46:32 -07003847int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003848{
David Rientjesd773ed62007-10-16 23:26:01 -07003849 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003850
3851 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003852 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003853 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003854 *
Christoph Lameter96146342006-07-03 00:24:13 -07003855 * A small portion of unmapped file backed pages is needed for
3856 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003857 * thrown out if the node is overallocated. So we do not reclaim
3858 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003859 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003860 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003861 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3862 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3863 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003864
Mel Gormana5f5f912016-07-28 15:46:32 -07003865 if (!pgdat_reclaimable(pgdat))
3866 return NODE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003867
Andrew Morton179e9632006-03-22 00:08:18 -08003868 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003869 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003870 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003871 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003872 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003873
3874 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003875 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003876 * have associated processors. This will favor the local processor
3877 * over remote processors and spread off node memory allocations
3878 * as wide as possible.
3879 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003880 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3881 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003882
Mel Gormana5f5f912016-07-28 15:46:32 -07003883 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3884 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003885
Mel Gormana5f5f912016-07-28 15:46:32 -07003886 ret = __node_reclaim(pgdat, gfp_mask, order);
3887 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003888
Mel Gorman24cf725182009-06-16 15:33:23 -07003889 if (!ret)
3890 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3891
David Rientjesd773ed62007-10-16 23:26:01 -07003892 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003893}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003894#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003895
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003896/*
3897 * page_evictable - test whether a page is evictable
3898 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003899 *
3900 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003901 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003902 *
3903 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003904 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003905 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003906 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003907 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003908int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003909{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003910 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003911}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003912
Hugh Dickins850465792012-01-20 14:34:19 -08003913#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003914/**
Hugh Dickins24513262012-01-20 14:34:21 -08003915 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3916 * @pages: array of pages to check
3917 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003918 *
Hugh Dickins24513262012-01-20 14:34:21 -08003919 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003920 *
3921 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003922 */
Hugh Dickins24513262012-01-20 14:34:21 -08003923void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003924{
Johannes Weiner925b7672012-01-12 17:18:15 -08003925 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003926 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003927 int pgscanned = 0;
3928 int pgrescued = 0;
3929 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003930
Hugh Dickins24513262012-01-20 14:34:21 -08003931 for (i = 0; i < nr_pages; i++) {
3932 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003933 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003934
Hugh Dickins24513262012-01-20 14:34:21 -08003935 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003936 if (pagepgdat != pgdat) {
3937 if (pgdat)
3938 spin_unlock_irq(&pgdat->lru_lock);
3939 pgdat = pagepgdat;
3940 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003941 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003942 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003943
Hugh Dickins24513262012-01-20 14:34:21 -08003944 if (!PageLRU(page) || !PageUnevictable(page))
3945 continue;
3946
Hugh Dickins39b5f292012-10-08 16:33:18 -07003947 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003948 enum lru_list lru = page_lru_base_type(page);
3949
Sasha Levin309381fea2014-01-23 15:52:54 -08003950 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003951 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003952 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3953 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003954 pgrescued++;
3955 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003956 }
Hugh Dickins24513262012-01-20 14:34:21 -08003957
Mel Gorman785b99f2016-07-28 15:47:23 -07003958 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003959 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3960 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003961 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003962 }
Hugh Dickins850465792012-01-20 14:34:19 -08003963}
3964#endif /* CONFIG_SHMEM */