blob: eb3dd37ccd7c727dcc0b6030f62c183097b956bb [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
87 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
Johannes Weiner241994e2015-02-11 15:26:06 -080095 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
Johannes Weineree814fe2014-08-06 16:06:19 -070098 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108};
109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
Andrew Mortonc255a452012-07-31 16:43:02 -0700151#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800152static bool global_reclaim(struct scan_control *sc)
153{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800154 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155}
Tejun Heo97c93412015-05-22 18:23:36 -0400156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400178 return true;
179#endif
180 return false;
181}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800182#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
Tejun Heo97c93412015-05-22 18:23:36 -0400187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800192#endif
193
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800194static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700195{
Alexandru Moised031a152015-11-05 18:48:08 -0800196 unsigned long nr;
Lisa Du6e543d52013-09-11 14:22:36 -0700197
198 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
Michal Hocko9f6c3992016-01-14 15:19:47 -0800199 zone_page_state(zone, NR_INACTIVE_FILE) +
200 zone_page_state(zone, NR_ISOLATED_FILE);
Lisa Du6e543d52013-09-11 14:22:36 -0700201
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
Michal Hocko9f6c3992016-01-14 15:19:47 -0800204 zone_page_state(zone, NR_INACTIVE_ANON) +
205 zone_page_state(zone, NR_ISOLATED_ANON);
Lisa Du6e543d52013-09-11 14:22:36 -0700206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
Mel Gorman0d5d8232014-08-06 16:07:16 -0700212 return zone_page_state(zone, NR_PAGES_SCANNED) <
213 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700214}
215
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700216static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800217{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700218 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700219 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800220
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800222}
223
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000225 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000227int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000229 size_t size = sizeof(*shrinker->nr_deferred);
230
231 /*
232 * If we only have one possible node in the system anyway, save
233 * ourselves the trouble and disable NUMA aware behavior. This way we
234 * will save memory and some small loop time later.
235 */
236 if (nr_node_ids == 1)
237 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
238
239 if (shrinker->flags & SHRINKER_NUMA_AWARE)
240 size *= nr_node_ids;
241
242 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
243 if (!shrinker->nr_deferred)
244 return -ENOMEM;
245
Rusty Russell8e1f9362007-07-17 04:03:17 -0700246 down_write(&shrinker_rwsem);
247 list_add_tail(&shrinker->list, &shrinker_list);
248 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000249 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700251EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252
253/*
254 * Remove one
255 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700256void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257{
258 down_write(&shrinker_rwsem);
259 list_del(&shrinker->list);
260 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700261 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700263EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
265#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000266
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800267static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
268 struct shrinker *shrinker,
269 unsigned long nr_scanned,
270 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000272 unsigned long freed = 0;
273 unsigned long long delta;
274 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700275 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000276 long nr;
277 long new_nr;
278 int nid = shrinkctl->nid;
279 long batch_size = shrinker->batch ? shrinker->batch
280 : SHRINK_BATCH;
281
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700282 freeable = shrinker->count_objects(shrinker, shrinkctl);
283 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000284 return 0;
285
286 /*
287 * copy the current shrinker scan count into a local variable
288 * and zero it so that other concurrent shrinker invocations
289 * don't also do this scanning work.
290 */
291 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
292
293 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800294 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700295 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800296 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000297 total_scan += delta;
298 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800299 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000300 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700301 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000302 }
303
304 /*
305 * We need to avoid excessive windup on filesystem shrinkers
306 * due to large numbers of GFP_NOFS allocations causing the
307 * shrinkers to return -1 all the time. This results in a large
308 * nr being built up so when a shrink that can do some work
309 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700310 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000311 * memory.
312 *
313 * Hence only allow the shrinker to scan the entire cache when
314 * a large delta change is calculated directly.
315 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700316 if (delta < freeable / 4)
317 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000318
319 /*
320 * Avoid risking looping forever due to too large nr value:
321 * never try to free more than twice the estimate number of
322 * freeable entries.
323 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700324 if (total_scan > freeable * 2)
325 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000326
327 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800328 nr_scanned, nr_eligible,
329 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000330
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800331 /*
332 * Normally, we should not scan less than batch_size objects in one
333 * pass to avoid too frequent shrinker calls, but if the slab has less
334 * than batch_size objects in total and we are really tight on memory,
335 * we will try to reclaim all available objects, otherwise we can end
336 * up failing allocations although there are plenty of reclaimable
337 * objects spread over several slabs with usage less than the
338 * batch_size.
339 *
340 * We detect the "tight on memory" situations by looking at the total
341 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700342 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800343 * scanning at high prio and therefore should try to reclaim as much as
344 * possible.
345 */
346 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700347 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000348 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800349 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000350
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800351 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000352 ret = shrinker->scan_objects(shrinker, shrinkctl);
353 if (ret == SHRINK_STOP)
354 break;
355 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000356
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800357 count_vm_events(SLABS_SCANNED, nr_to_scan);
358 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000359
360 cond_resched();
361 }
362
363 /*
364 * move the unused scan count back into the shrinker in a
365 * manner that handles concurrent updates. If we exhausted the
366 * scan, there is no need to do an update.
367 */
368 if (total_scan > 0)
369 new_nr = atomic_long_add_return(total_scan,
370 &shrinker->nr_deferred[nid]);
371 else
372 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
373
Dave Hansendf9024a2014-06-04 16:08:07 -0700374 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000375 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376}
377
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800378/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800379 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800380 * @gfp_mask: allocation context
381 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800382 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800383 * @nr_scanned: pressure numerator
384 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800386 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800388 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
389 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800391 * @memcg specifies the memory cgroup to target. If it is not NULL,
392 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
393 * objects from the memory cgroup specified. Otherwise all shrinkers
394 * are called, and memcg aware shrinkers are supposed to scan the
395 * global list then.
396 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800397 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
398 * the available objects should be scanned. Page reclaim for example
399 * passes the number of pages scanned and the number of pages on the
400 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
401 * when it encountered mapped pages. The ratio is further biased by
402 * the ->seeks setting of the shrink function, which indicates the
403 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800405 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800407static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
408 struct mem_cgroup *memcg,
409 unsigned long nr_scanned,
410 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
412 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000413 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Johannes Weiner567e9ab2016-01-20 15:02:24 -0800415 if (memcg && !memcg_kmem_online(memcg))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800416 return 0;
417
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800418 if (nr_scanned == 0)
419 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
Minchan Kimf06590b2011-05-24 17:11:11 -0700421 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000422 /*
423 * If we would return 0, our callers would understand that we
424 * have nothing else to shrink and give up trying. By returning
425 * 1 we keep it going and assume we'll be able to shrink next
426 * time.
427 */
428 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700429 goto out;
430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
432 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800433 struct shrink_control sc = {
434 .gfp_mask = gfp_mask,
435 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800436 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800437 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800438
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800439 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
440 continue;
441
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800442 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
443 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000444
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800445 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800447
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700449out:
450 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000451 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800454void drop_slab_node(int nid)
455{
456 unsigned long freed;
457
458 do {
459 struct mem_cgroup *memcg = NULL;
460
461 freed = 0;
462 do {
463 freed += shrink_slab(GFP_KERNEL, nid, memcg,
464 1000, 1000);
465 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
466 } while (freed > 10);
467}
468
469void drop_slab(void)
470{
471 int nid;
472
473 for_each_online_node(nid)
474 drop_slab_node(nid);
475}
476
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477static inline int is_page_cache_freeable(struct page *page)
478{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700479 /*
480 * A freeable page cache page is referenced only by the caller
481 * that isolated the page, the page cache radix tree and
482 * optional buffer heads at page->private.
483 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700484 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
486
Tejun Heo703c2702015-05-22 17:13:44 -0400487static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488{
Christoph Lameter930d9152006-01-08 01:00:47 -0800489 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400491 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400493 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 return 1;
495 return 0;
496}
497
498/*
499 * We detected a synchronous write error writing a page out. Probably
500 * -ENOSPC. We need to propagate that into the address_space for a subsequent
501 * fsync(), msync() or close().
502 *
503 * The tricky part is that after writepage we cannot touch the mapping: nothing
504 * prevents it from being freed up. But we have a ref on the page and once
505 * that page is locked, the mapping is pinned.
506 *
507 * We're allowed to run sleeping lock_page() here because we know the caller has
508 * __GFP_FS.
509 */
510static void handle_write_error(struct address_space *mapping,
511 struct page *page, int error)
512{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100513 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700514 if (page_mapping(page) == mapping)
515 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 unlock_page(page);
517}
518
Christoph Lameter04e62a22006-06-23 02:03:38 -0700519/* possible outcome of pageout() */
520typedef enum {
521 /* failed to write page out, page is locked */
522 PAGE_KEEP,
523 /* move page to the active list, page is locked */
524 PAGE_ACTIVATE,
525 /* page has been sent to the disk successfully, page is unlocked */
526 PAGE_SUCCESS,
527 /* page is clean and locked */
528 PAGE_CLEAN,
529} pageout_t;
530
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531/*
Andrew Morton1742f192006-03-22 00:08:21 -0800532 * pageout is called by shrink_page_list() for each dirty page.
533 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700535static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700536 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
538 /*
539 * If the page is dirty, only perform writeback if that write
540 * will be non-blocking. To prevent this allocation from being
541 * stalled by pagecache activity. But note that there may be
542 * stalls if we need to run get_block(). We could test
543 * PagePrivate for that.
544 *
Al Viro81742022014-04-03 03:17:43 -0400545 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 * this page's queue, we can perform writeback even if that
547 * will block.
548 *
549 * If the page is swapcache, write it back even if that would
550 * block, for some throttling. This happens by accident, because
551 * swap_backing_dev_info is bust: it doesn't reflect the
552 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 */
554 if (!is_page_cache_freeable(page))
555 return PAGE_KEEP;
556 if (!mapping) {
557 /*
558 * Some data journaling orphaned pages can have
559 * page->mapping == NULL while being dirty with clean buffers.
560 */
David Howells266cf652009-04-03 16:42:36 +0100561 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 if (try_to_free_buffers(page)) {
563 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700564 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 return PAGE_CLEAN;
566 }
567 }
568 return PAGE_KEEP;
569 }
570 if (mapping->a_ops->writepage == NULL)
571 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400572 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 return PAGE_KEEP;
574
575 if (clear_page_dirty_for_io(page)) {
576 int res;
577 struct writeback_control wbc = {
578 .sync_mode = WB_SYNC_NONE,
579 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700580 .range_start = 0,
581 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 .for_reclaim = 1,
583 };
584
585 SetPageReclaim(page);
586 res = mapping->a_ops->writepage(page, &wbc);
587 if (res < 0)
588 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800589 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 ClearPageReclaim(page);
591 return PAGE_ACTIVATE;
592 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700593
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 if (!PageWriteback(page)) {
595 /* synchronous write or broken a_ops? */
596 ClearPageReclaim(page);
597 }
yalin wang3aa23852016-01-14 15:18:30 -0800598 trace_mm_vmscan_writepage(page);
Andrew Mortone129b5c2006-09-27 01:50:00 -0700599 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 return PAGE_SUCCESS;
601 }
602
603 return PAGE_CLEAN;
604}
605
Andrew Mortona649fd92006-10-17 00:09:36 -0700606/*
Nick Piggine2867812008-07-25 19:45:30 -0700607 * Same as remove_mapping, but if the page is removed from the mapping, it
608 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700609 */
Johannes Weinera5289102014-04-03 14:47:51 -0700610static int __remove_mapping(struct address_space *mapping, struct page *page,
611 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800612{
Greg Thelenc4843a72015-05-22 17:13:16 -0400613 unsigned long flags;
614 struct mem_cgroup *memcg;
615
Nick Piggin28e4d962006-09-25 23:31:23 -0700616 BUG_ON(!PageLocked(page));
617 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800618
Greg Thelenc4843a72015-05-22 17:13:16 -0400619 memcg = mem_cgroup_begin_page_stat(page);
620 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800621 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700622 * The non racy check for a busy page.
623 *
624 * Must be careful with the order of the tests. When someone has
625 * a ref to the page, it may be possible that they dirty it then
626 * drop the reference. So if PageDirty is tested before page_count
627 * here, then the following race may occur:
628 *
629 * get_user_pages(&page);
630 * [user mapping goes away]
631 * write_to(page);
632 * !PageDirty(page) [good]
633 * SetPageDirty(page);
634 * put_page(page);
635 * !page_count(page) [good, discard it]
636 *
637 * [oops, our write_to data is lost]
638 *
639 * Reversing the order of the tests ensures such a situation cannot
640 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
641 * load is not satisfied before that of page->_count.
642 *
643 * Note that if SetPageDirty is always performed via set_page_dirty,
644 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800645 */
Nick Piggine2867812008-07-25 19:45:30 -0700646 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800647 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700648 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
649 if (unlikely(PageDirty(page))) {
650 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800651 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700652 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800653
654 if (PageSwapCache(page)) {
655 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700656 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800657 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400658 spin_unlock_irqrestore(&mapping->tree_lock, flags);
659 mem_cgroup_end_page_stat(memcg);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700660 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700661 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500662 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700663 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500664
665 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700666 /*
667 * Remember a shadow entry for reclaimed file cache in
668 * order to detect refaults, thus thrashing, later on.
669 *
670 * But don't store shadows in an address space that is
671 * already exiting. This is not just an optizimation,
672 * inode reclaim needs to empty out the radix tree or
673 * the nodes are lost. Don't plant shadows behind its
674 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800675 *
676 * We also don't store shadows for DAX mappings because the
677 * only page cache pages found in these are zero pages
678 * covering holes, and because we don't want to mix DAX
679 * exceptional entries and shadow exceptional entries in the
680 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700681 */
682 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800683 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700684 shadow = workingset_eviction(mapping, page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400685 __delete_from_page_cache(page, shadow, memcg);
686 spin_unlock_irqrestore(&mapping->tree_lock, flags);
687 mem_cgroup_end_page_stat(memcg);
Linus Torvalds6072d132010-12-01 13:35:19 -0500688
689 if (freepage != NULL)
690 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800691 }
692
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800693 return 1;
694
695cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400696 spin_unlock_irqrestore(&mapping->tree_lock, flags);
697 mem_cgroup_end_page_stat(memcg);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800698 return 0;
699}
700
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701/*
Nick Piggine2867812008-07-25 19:45:30 -0700702 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
703 * someone else has a ref on the page, abort and return 0. If it was
704 * successfully detached, return 1. Assumes the caller has a single ref on
705 * this page.
706 */
707int remove_mapping(struct address_space *mapping, struct page *page)
708{
Johannes Weinera5289102014-04-03 14:47:51 -0700709 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700710 /*
711 * Unfreezing the refcount with 1 rather than 2 effectively
712 * drops the pagecache ref for us without requiring another
713 * atomic operation.
714 */
715 page_unfreeze_refs(page, 1);
716 return 1;
717 }
718 return 0;
719}
720
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700721/**
722 * putback_lru_page - put previously isolated page onto appropriate LRU list
723 * @page: page to be put back to appropriate lru list
724 *
725 * Add previously isolated @page to appropriate LRU list.
726 * Page may still be unevictable for other reasons.
727 *
728 * lru_lock must not be held, interrupts must be enabled.
729 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700730void putback_lru_page(struct page *page)
731{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700732 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700733 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700734
Sasha Levin309381fea2014-01-23 15:52:54 -0800735 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700736
737redo:
738 ClearPageUnevictable(page);
739
Hugh Dickins39b5f292012-10-08 16:33:18 -0700740 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700741 /*
742 * For evictable pages, we can use the cache.
743 * In event of a race, worst case is we end up with an
744 * unevictable page on [in]active list.
745 * We know how to handle that.
746 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700747 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700748 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700749 } else {
750 /*
751 * Put unevictable pages directly on zone's unevictable
752 * list.
753 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700754 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700755 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700756 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700757 * When racing with an mlock or AS_UNEVICTABLE clearing
758 * (page is unlocked) make sure that if the other thread
759 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800760 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700761 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700762 * the page back to the evictable list.
763 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700764 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700765 */
766 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700767 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700768
769 /*
770 * page's status can change while we move it among lru. If an evictable
771 * page is on unevictable list, it never be freed. To avoid that,
772 * check after we added it to the list, again.
773 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700774 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700775 if (!isolate_lru_page(page)) {
776 put_page(page);
777 goto redo;
778 }
779 /* This means someone else dropped this page from LRU
780 * So, it will be freed or putback to LRU again. There is
781 * nothing to do here.
782 */
783 }
784
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700785 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700786 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700787 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700788 count_vm_event(UNEVICTABLE_PGCULLED);
789
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700790 put_page(page); /* drop ref from isolate */
791}
792
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800793enum page_references {
794 PAGEREF_RECLAIM,
795 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800796 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800797 PAGEREF_ACTIVATE,
798};
799
800static enum page_references page_check_references(struct page *page,
801 struct scan_control *sc)
802{
Johannes Weiner645747462010-03-05 13:42:22 -0800803 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800804 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800805
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700806 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
807 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800808 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800809
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800810 /*
811 * Mlock lost the isolation race with us. Let try_to_unmap()
812 * move the page to the unevictable list.
813 */
814 if (vm_flags & VM_LOCKED)
815 return PAGEREF_RECLAIM;
816
Johannes Weiner645747462010-03-05 13:42:22 -0800817 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700818 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800819 return PAGEREF_ACTIVATE;
820 /*
821 * All mapped pages start out with page table
822 * references from the instantiating fault, so we need
823 * to look twice if a mapped file page is used more
824 * than once.
825 *
826 * Mark it and spare it for another trip around the
827 * inactive list. Another page table reference will
828 * lead to its activation.
829 *
830 * Note: the mark is set for activated pages as well
831 * so that recently deactivated but used pages are
832 * quickly recovered.
833 */
834 SetPageReferenced(page);
835
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800836 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800837 return PAGEREF_ACTIVATE;
838
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800839 /*
840 * Activate file-backed executable pages after first usage.
841 */
842 if (vm_flags & VM_EXEC)
843 return PAGEREF_ACTIVATE;
844
Johannes Weiner645747462010-03-05 13:42:22 -0800845 return PAGEREF_KEEP;
846 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800847
848 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700849 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800850 return PAGEREF_RECLAIM_CLEAN;
851
852 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800853}
854
Mel Gormane2be15f2013-07-03 15:01:57 -0700855/* Check if a page is dirty or under writeback */
856static void page_check_dirty_writeback(struct page *page,
857 bool *dirty, bool *writeback)
858{
Mel Gormanb4597222013-07-03 15:02:05 -0700859 struct address_space *mapping;
860
Mel Gormane2be15f2013-07-03 15:01:57 -0700861 /*
862 * Anonymous pages are not handled by flushers and must be written
863 * from reclaim context. Do not stall reclaim based on them
864 */
865 if (!page_is_file_cache(page)) {
866 *dirty = false;
867 *writeback = false;
868 return;
869 }
870
871 /* By default assume that the page flags are accurate */
872 *dirty = PageDirty(page);
873 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700874
875 /* Verify dirty/writeback state if the filesystem supports it */
876 if (!page_has_private(page))
877 return;
878
879 mapping = page_mapping(page);
880 if (mapping && mapping->a_ops->is_dirty_writeback)
881 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700882}
883
Nick Piggine2867812008-07-25 19:45:30 -0700884/*
Andrew Morton1742f192006-03-22 00:08:21 -0800885 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 */
Andrew Morton1742f192006-03-22 00:08:21 -0800887static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700888 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700889 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700890 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700891 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700892 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700893 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700894 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700895 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700896 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
898 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700899 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700901 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700902 unsigned long nr_dirty = 0;
903 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800904 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700905 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700906 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907
908 cond_resched();
909
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 while (!list_empty(page_list)) {
911 struct address_space *mapping;
912 struct page *page;
913 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700914 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700915 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800916 bool lazyfree = false;
917 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
919 cond_resched();
920
921 page = lru_to_page(page_list);
922 list_del(&page->lru);
923
Nick Piggin529ae9a2008-08-02 12:01:03 +0200924 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 goto keep;
926
Sasha Levin309381fea2014-01-23 15:52:54 -0800927 VM_BUG_ON_PAGE(PageActive(page), page);
928 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
930 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800931
Hugh Dickins39b5f292012-10-08 16:33:18 -0700932 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700933 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700934
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700935 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800936 goto keep_locked;
937
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 /* Double the slab pressure for mapped and swapcache pages */
939 if (page_mapped(page) || PageSwapCache(page))
940 sc->nr_scanned++;
941
Andy Whitcroftc661b072007-08-22 14:01:26 -0700942 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
943 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
944
Mel Gorman283aba92013-07-03 15:01:51 -0700945 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700946 * The number of dirty pages determines if a zone is marked
947 * reclaim_congested which affects wait_iff_congested. kswapd
948 * will stall and start writing pages if the tail of the LRU
949 * is all dirty unqueued pages.
950 */
951 page_check_dirty_writeback(page, &dirty, &writeback);
952 if (dirty || writeback)
953 nr_dirty++;
954
955 if (dirty && !writeback)
956 nr_unqueued_dirty++;
957
Mel Gormand04e8ac2013-07-03 15:02:03 -0700958 /*
959 * Treat this page as congested if the underlying BDI is or if
960 * pages are cycling through the LRU so quickly that the
961 * pages marked for immediate reclaim are making it to the
962 * end of the LRU a second time.
963 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700964 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800965 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400966 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700967 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700968 nr_congested++;
969
970 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700971 * If a page at the tail of the LRU is under writeback, there
972 * are three cases to consider.
973 *
974 * 1) If reclaim is encountering an excessive number of pages
975 * under writeback and this page is both under writeback and
976 * PageReclaim then it indicates that pages are being queued
977 * for IO but are being recycled through the LRU before the
978 * IO can complete. Waiting on the page itself risks an
979 * indefinite stall if it is impossible to writeback the
980 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700981 * note that the LRU is being scanned too quickly and the
982 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700983 *
Tejun Heo97c93412015-05-22 18:23:36 -0400984 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -0700985 * not marked for immediate reclaim, or the caller does not
986 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
987 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -0400988 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -0700989 *
Michal Hockoecf5fc62015-08-04 14:36:58 -0700990 * Require may_enter_fs because we would wait on fs, which
991 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -0700992 * enter reclaim, and deadlock if it waits on a page for
993 * which it is needed to do the write (loop masks off
994 * __GFP_IO|__GFP_FS for this reason); but more thought
995 * would probably show more reasons.
996 *
Hugh Dickins7fadc822015-09-08 15:03:46 -0700997 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -0700998 * PageReclaim. memcg does not have any dirty pages
999 * throttling so we could easily OOM just because too many
1000 * pages are in writeback and there is nothing else to
1001 * reclaim. Wait for the writeback to complete.
1002 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001003 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001004 /* Case 1 above */
1005 if (current_is_kswapd() &&
1006 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -07001007 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001008 nr_immediate++;
1009 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001010
1011 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001012 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001013 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001014 /*
1015 * This is slightly racy - end_page_writeback()
1016 * might have just cleared PageReclaim, then
1017 * setting PageReclaim here end up interpreted
1018 * as PageReadahead - but that does not matter
1019 * enough to care. What we do want is for this
1020 * page to have PageReclaim set next time memcg
1021 * reclaim reaches the tests above, so it will
1022 * then wait_on_page_writeback() to avoid OOM;
1023 * and it's also appropriate in global reclaim.
1024 */
1025 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001026 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001027 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001028
1029 /* Case 3 above */
1030 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001031 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001032 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001033 /* then go back and try same page again */
1034 list_add_tail(&page->lru, page_list);
1035 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001036 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001037 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038
Minchan Kim02c6de82012-10-08 16:31:55 -07001039 if (!force_reclaim)
1040 references = page_check_references(page, sc);
1041
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001042 switch (references) {
1043 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001045 case PAGEREF_KEEP:
1046 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001047 case PAGEREF_RECLAIM:
1048 case PAGEREF_RECLAIM_CLEAN:
1049 ; /* try to reclaim the page below */
1050 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 /*
1053 * Anonymous process memory has backing store?
1054 * Try to allocate it some swap space here.
1055 */
Nick Pigginb291f002008-10-18 20:26:44 -07001056 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001057 if (!(sc->gfp_mask & __GFP_IO))
1058 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001059 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001061 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001062 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063
Mel Gormane2be15f2013-07-03 15:01:57 -07001064 /* Adding to swap updated mapping */
1065 mapping = page_mapping(page);
1066 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067
1068 /*
1069 * The page is mapped into the page tables of one or more
1070 * processes. Try to unmap it here.
1071 */
1072 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001073 switch (ret = try_to_unmap(page, lazyfree ?
1074 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1075 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 case SWAP_FAIL:
1077 goto activate_locked;
1078 case SWAP_AGAIN:
1079 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001080 case SWAP_MLOCK:
1081 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001082 case SWAP_LZFREE:
1083 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 case SWAP_SUCCESS:
1085 ; /* try to free the page below */
1086 }
1087 }
1088
1089 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001090 /*
1091 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001092 * avoid risk of stack overflow but only writeback
1093 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001094 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001095 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001096 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001097 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001098 /*
1099 * Immediately reclaim when written back.
1100 * Similar in principal to deactivate_page()
1101 * except we already have the page isolated
1102 * and know it's dirty
1103 */
1104 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1105 SetPageReclaim(page);
1106
Mel Gormanee728862011-10-31 17:07:38 -07001107 goto keep_locked;
1108 }
1109
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001110 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001112 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001114 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 goto keep_locked;
1116
Mel Gormand950c942015-09-04 15:47:35 -07001117 /*
1118 * Page is dirty. Flush the TLB if a writable entry
1119 * potentially exists to avoid CPU writes after IO
1120 * starts and then write it out here.
1121 */
1122 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001123 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 case PAGE_KEEP:
1125 goto keep_locked;
1126 case PAGE_ACTIVATE:
1127 goto activate_locked;
1128 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001129 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001130 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001131 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 /*
1135 * A synchronous write - probably a ramdisk. Go
1136 * ahead and try to reclaim the page.
1137 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001138 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 goto keep;
1140 if (PageDirty(page) || PageWriteback(page))
1141 goto keep_locked;
1142 mapping = page_mapping(page);
1143 case PAGE_CLEAN:
1144 ; /* try to free the page below */
1145 }
1146 }
1147
1148 /*
1149 * If the page has buffers, try to free the buffer mappings
1150 * associated with this page. If we succeed we try to free
1151 * the page as well.
1152 *
1153 * We do this even if the page is PageDirty().
1154 * try_to_release_page() does not perform I/O, but it is
1155 * possible for a page to have PageDirty set, but it is actually
1156 * clean (all its buffers are clean). This happens if the
1157 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001158 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 * try_to_release_page() will discover that cleanness and will
1160 * drop the buffers and mark the page clean - it can be freed.
1161 *
1162 * Rarely, pages can have buffers and no ->mapping. These are
1163 * the pages which were not successfully invalidated in
1164 * truncate_complete_page(). We try to drop those buffers here
1165 * and if that worked, and the page is no longer mapped into
1166 * process address space (page_count == 1) it can be freed.
1167 * Otherwise, leave the page on the LRU so it is swappable.
1168 */
David Howells266cf652009-04-03 16:42:36 +01001169 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 if (!try_to_release_page(page, sc->gfp_mask))
1171 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001172 if (!mapping && page_count(page) == 1) {
1173 unlock_page(page);
1174 if (put_page_testzero(page))
1175 goto free_it;
1176 else {
1177 /*
1178 * rare race with speculative reference.
1179 * the speculative reference will free
1180 * this page shortly, so we may
1181 * increment nr_reclaimed here (and
1182 * leave it off the LRU).
1183 */
1184 nr_reclaimed++;
1185 continue;
1186 }
1187 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 }
1189
Minchan Kim854e9ed2016-01-15 16:54:53 -08001190lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001191 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001192 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
Nick Piggina978d6f2008-10-18 20:26:58 -07001194 /*
1195 * At this point, we have no other references and there is
1196 * no way to pick any more up (removed from LRU, removed
1197 * from pagecache). Can use non-atomic bitops now (and
1198 * we obviously don't have to worry about waking up a process
1199 * waiting on the page lock, because there are no references.
1200 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001201 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001202free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001203 if (ret == SWAP_LZFREE)
1204 count_vm_event(PGLAZYFREED);
1205
Andrew Morton05ff5132006-03-22 00:08:20 -08001206 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001207
1208 /*
1209 * Is there need to periodically free_page_list? It would
1210 * appear not as the counts should be low
1211 */
1212 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 continue;
1214
Nick Pigginb291f002008-10-18 20:26:44 -07001215cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001216 if (PageSwapCache(page))
1217 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001218 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001219 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001220 continue;
1221
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001223 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001224 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001225 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001226 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 SetPageActive(page);
1228 pgactivate++;
1229keep_locked:
1230 unlock_page(page);
1231keep:
1232 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001233 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001235
Johannes Weiner747db952014-08-08 14:19:24 -07001236 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001237 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001238 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001241 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001242
Mel Gorman8e950282013-07-03 15:02:02 -07001243 *ret_nr_dirty += nr_dirty;
1244 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001245 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001246 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001247 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001248 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249}
1250
Minchan Kim02c6de82012-10-08 16:31:55 -07001251unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1252 struct list_head *page_list)
1253{
1254 struct scan_control sc = {
1255 .gfp_mask = GFP_KERNEL,
1256 .priority = DEF_PRIORITY,
1257 .may_unmap = 1,
1258 };
Mel Gorman8e950282013-07-03 15:02:02 -07001259 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001260 struct page *page, *next;
1261 LIST_HEAD(clean_pages);
1262
1263 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001264 if (page_is_file_cache(page) && !PageDirty(page) &&
1265 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001266 ClearPageActive(page);
1267 list_move(&page->lru, &clean_pages);
1268 }
1269 }
1270
1271 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001272 TTU_UNMAP|TTU_IGNORE_ACCESS,
1273 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001274 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001275 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001276 return ret;
1277}
1278
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001279/*
1280 * Attempt to remove the specified page from its LRU. Only take this page
1281 * if it is of the appropriate PageActive status. Pages which are being
1282 * freed elsewhere are also ignored.
1283 *
1284 * page: page to consider
1285 * mode: one of the LRU isolation modes defined above
1286 *
1287 * returns 0 on success, -ve errno on failure.
1288 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001289int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001290{
1291 int ret = -EINVAL;
1292
1293 /* Only take pages on the LRU. */
1294 if (!PageLRU(page))
1295 return ret;
1296
Minchan Kime46a2872012-10-08 16:33:48 -07001297 /* Compaction should not handle unevictable pages but CMA can do so */
1298 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001299 return ret;
1300
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001301 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001302
Mel Gormanc8244932012-01-12 17:19:38 -08001303 /*
1304 * To minimise LRU disruption, the caller can indicate that it only
1305 * wants to isolate pages it will be able to operate on without
1306 * blocking - clean pages for the most part.
1307 *
1308 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1309 * is used by reclaim when it is cannot write to backing storage
1310 *
1311 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1312 * that it is possible to migrate without blocking
1313 */
1314 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1315 /* All the caller can do on PageWriteback is block */
1316 if (PageWriteback(page))
1317 return ret;
1318
1319 if (PageDirty(page)) {
1320 struct address_space *mapping;
1321
1322 /* ISOLATE_CLEAN means only clean pages */
1323 if (mode & ISOLATE_CLEAN)
1324 return ret;
1325
1326 /*
1327 * Only pages without mappings or that have a
1328 * ->migratepage callback are possible to migrate
1329 * without blocking
1330 */
1331 mapping = page_mapping(page);
1332 if (mapping && !mapping->a_ops->migratepage)
1333 return ret;
1334 }
1335 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001336
Minchan Kimf80c0672011-10-31 17:06:55 -07001337 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1338 return ret;
1339
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001340 if (likely(get_page_unless_zero(page))) {
1341 /*
1342 * Be careful not to clear PageLRU until after we're
1343 * sure the page is not being freed elsewhere -- the
1344 * page release code relies on it.
1345 */
1346 ClearPageLRU(page);
1347 ret = 0;
1348 }
1349
1350 return ret;
1351}
1352
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001353/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 * zone->lru_lock is heavily contended. Some of the functions that
1355 * shrink the lists perform better by taking out a batch of pages
1356 * and working on them outside the LRU lock.
1357 *
1358 * For pagecache intensive workloads, this function is the hottest
1359 * spot in the kernel (apart from copy_*_user functions).
1360 *
1361 * Appropriate locks must be held before calling this function.
1362 *
1363 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001364 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001366 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001367 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001368 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001369 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 *
1371 * returns how many pages were moved onto *@dst.
1372 */
Andrew Morton69e05942006-03-22 00:08:19 -08001373static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001374 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001375 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001376 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001378 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001379 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001380 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001382 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1383 !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001384 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001385 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001386
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 page = lru_to_page(src);
1388 prefetchw_prev_lru_page(page, src, flags);
1389
Sasha Levin309381fea2014-01-23 15:52:54 -08001390 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001391
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001392 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001393 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001394 nr_pages = hpage_nr_pages(page);
1395 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001396 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001397 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001398 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001399
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001400 case -EBUSY:
1401 /* else it is being freed elsewhere */
1402 list_move(&page->lru, src);
1403 continue;
1404
1405 default:
1406 BUG();
1407 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 }
1409
Hugh Dickinsf6260122012-01-12 17:20:06 -08001410 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001411 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1412 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 return nr_taken;
1414}
1415
Nick Piggin62695a82008-10-18 20:26:09 -07001416/**
1417 * isolate_lru_page - tries to isolate a page from its LRU list
1418 * @page: page to isolate from its LRU list
1419 *
1420 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1421 * vmstat statistic corresponding to whatever LRU list the page was on.
1422 *
1423 * Returns 0 if the page was removed from an LRU list.
1424 * Returns -EBUSY if the page was not on an LRU list.
1425 *
1426 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001427 * the active list, it will have PageActive set. If it was found on
1428 * the unevictable list, it will have the PageUnevictable bit set. That flag
1429 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001430 *
1431 * The vmstat statistic corresponding to the list on which the page was
1432 * found will be decremented.
1433 *
1434 * Restrictions:
1435 * (1) Must be called with an elevated refcount on the page. This is a
1436 * fundamentnal difference from isolate_lru_pages (which is called
1437 * without a stable reference).
1438 * (2) the lru_lock must not be held.
1439 * (3) interrupts must be enabled.
1440 */
1441int isolate_lru_page(struct page *page)
1442{
1443 int ret = -EBUSY;
1444
Sasha Levin309381fea2014-01-23 15:52:54 -08001445 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovbb5b8582016-01-14 15:21:43 -08001446 VM_BUG_ON_PAGE(PageTail(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001447
Nick Piggin62695a82008-10-18 20:26:09 -07001448 if (PageLRU(page)) {
1449 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001450 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001451
1452 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001453 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001454 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001455 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001456 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001457 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001458 del_page_from_lru_list(page, lruvec, lru);
1459 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001460 }
1461 spin_unlock_irq(&zone->lru_lock);
1462 }
1463 return ret;
1464}
1465
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001466/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001467 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1468 * then get resheduled. When there are massive number of tasks doing page
1469 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1470 * the LRU list will go small and be scanned faster than necessary, leading to
1471 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001472 */
1473static int too_many_isolated(struct zone *zone, int file,
1474 struct scan_control *sc)
1475{
1476 unsigned long inactive, isolated;
1477
1478 if (current_is_kswapd())
1479 return 0;
1480
Tejun Heo97c93412015-05-22 18:23:36 -04001481 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001482 return 0;
1483
1484 if (file) {
1485 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1486 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1487 } else {
1488 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1489 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1490 }
1491
Fengguang Wu3cf23842012-12-18 14:23:31 -08001492 /*
1493 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1494 * won't get blocked by normal direct-reclaimers, forming a circular
1495 * deadlock.
1496 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001497 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001498 inactive >>= 3;
1499
Rik van Riel35cd7812009-09-21 17:01:38 -07001500 return isolated > inactive;
1501}
1502
Mel Gorman66635622010-08-09 17:19:30 -07001503static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001504putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001505{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001506 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1507 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001508 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001509
Mel Gorman66635622010-08-09 17:19:30 -07001510 /*
1511 * Put back any unfreeable pages.
1512 */
Mel Gorman66635622010-08-09 17:19:30 -07001513 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001514 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001515 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001516
Sasha Levin309381fea2014-01-23 15:52:54 -08001517 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001518 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001519 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001520 spin_unlock_irq(&zone->lru_lock);
1521 putback_lru_page(page);
1522 spin_lock_irq(&zone->lru_lock);
1523 continue;
1524 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001525
1526 lruvec = mem_cgroup_page_lruvec(page, zone);
1527
Linus Torvalds7a608572011-01-17 14:42:19 -08001528 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001529 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001530 add_page_to_lru_list(page, lruvec, lru);
1531
Mel Gorman66635622010-08-09 17:19:30 -07001532 if (is_active_lru(lru)) {
1533 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001534 int numpages = hpage_nr_pages(page);
1535 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001536 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001537 if (put_page_testzero(page)) {
1538 __ClearPageLRU(page);
1539 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001540 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001541
1542 if (unlikely(PageCompound(page))) {
1543 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001544 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001545 (*get_compound_page_dtor(page))(page);
1546 spin_lock_irq(&zone->lru_lock);
1547 } else
1548 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001549 }
1550 }
Mel Gorman66635622010-08-09 17:19:30 -07001551
Hugh Dickins3f797682012-01-12 17:20:07 -08001552 /*
1553 * To save our caller's stack, now use input list for pages to free.
1554 */
1555 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001556}
1557
1558/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001559 * If a kernel thread (such as nfsd for loop-back mounts) services
1560 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1561 * In that case we should only throttle if the backing device it is
1562 * writing to is congested. In other cases it is safe to throttle.
1563 */
1564static int current_may_throttle(void)
1565{
1566 return !(current->flags & PF_LESS_THROTTLE) ||
1567 current->backing_dev_info == NULL ||
1568 bdi_write_congested(current->backing_dev_info);
1569}
1570
1571/*
Andrew Morton1742f192006-03-22 00:08:21 -08001572 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1573 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 */
Mel Gorman66635622010-08-09 17:19:30 -07001575static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001576shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001577 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578{
1579 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001580 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001581 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001582 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001583 unsigned long nr_dirty = 0;
1584 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001585 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001586 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001587 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001588 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001589 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001590 struct zone *zone = lruvec_zone(lruvec);
1591 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001592
Rik van Riel35cd7812009-09-21 17:01:38 -07001593 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001594 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001595
1596 /* We are about to die and free our memory. Return now. */
1597 if (fatal_signal_pending(current))
1598 return SWAP_CLUSTER_MAX;
1599 }
1600
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001602
1603 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001604 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001605 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001606 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001607
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001610 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1611 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001612
1613 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1614 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1615
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001616 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001617 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001618 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001619 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001620 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001621 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001622 }
Hillf Dantond563c052012-03-21 16:34:02 -07001623 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001624
Hillf Dantond563c052012-03-21 16:34:02 -07001625 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001626 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001627
Minchan Kim02c6de82012-10-08 16:31:55 -07001628 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001629 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1630 &nr_writeback, &nr_immediate,
1631 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001632
Hugh Dickins3f797682012-01-12 17:20:07 -08001633 spin_lock_irq(&zone->lru_lock);
1634
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001635 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001636
Ying Han904249a2012-04-25 16:01:48 -07001637 if (global_reclaim(sc)) {
1638 if (current_is_kswapd())
1639 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1640 nr_reclaimed);
1641 else
1642 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1643 nr_reclaimed);
1644 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001645
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001646 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001647
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001648 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001649
1650 spin_unlock_irq(&zone->lru_lock);
1651
Johannes Weiner747db952014-08-08 14:19:24 -07001652 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001653 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001654
Mel Gorman92df3a72011-10-31 17:07:56 -07001655 /*
1656 * If reclaim is isolating dirty pages under writeback, it implies
1657 * that the long-lived page allocation rate is exceeding the page
1658 * laundering rate. Either the global limits are not being effective
1659 * at throttling processes due to the page distribution throughout
1660 * zones or there is heavy usage of a slow backing device. The
1661 * only option is to throttle from reclaim context which is not ideal
1662 * as there is no guarantee the dirtying process is throttled in the
1663 * same way balance_dirty_pages() manages.
1664 *
Mel Gorman8e950282013-07-03 15:02:02 -07001665 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1666 * of pages under pages flagged for immediate reclaim and stall if any
1667 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001668 */
Mel Gorman918fc712013-07-08 16:00:25 -07001669 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001670 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001671
Mel Gormand43006d2013-07-03 15:01:50 -07001672 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001673 * Legacy memcg will stall in page writeback so avoid forcibly
1674 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001675 */
Tejun Heo97c93412015-05-22 18:23:36 -04001676 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001677 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001678 * Tag a zone as congested if all the dirty pages scanned were
1679 * backed by a congested BDI and wait_iff_congested will stall.
1680 */
1681 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001682 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001683
1684 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001685 * If dirty pages are scanned that are not queued for IO, it
1686 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001687 * the zone ZONE_DIRTY and kswapd will start writing pages from
1688 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001689 */
1690 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001691 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001692
1693 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001694 * If kswapd scans pages marked marked for immediate
1695 * reclaim and under writeback (nr_immediate), it implies
1696 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001697 * they are written so also forcibly stall.
1698 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001699 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001700 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001701 }
Mel Gormand43006d2013-07-03 15:01:50 -07001702
Mel Gorman8e950282013-07-03 15:02:02 -07001703 /*
1704 * Stall direct reclaim for IO completions if underlying BDIs or zone
1705 * is congested. Allow kswapd to continue until it starts encountering
1706 * unqueued dirty pages or cycling through the LRU too quickly.
1707 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001708 if (!sc->hibernation_mode && !current_is_kswapd() &&
1709 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001710 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1711
yalin wangba5e9572016-01-14 15:18:48 -08001712 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1713 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001714 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715}
1716
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001717/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 * This moves pages from the active list to the inactive list.
1719 *
1720 * We move them the other way if the page is referenced by one or more
1721 * processes, from rmap.
1722 *
1723 * If the pages are mostly unmapped, the processing is fast and it is
1724 * appropriate to hold zone->lru_lock across the whole operation. But if
1725 * the pages are mapped, the processing is slow (page_referenced()) so we
1726 * should drop zone->lru_lock around each page. It's impossible to balance
1727 * this, so instead we remove the pages from the LRU while processing them.
1728 * It is safe to rely on PG_active against the non-LRU pages in here because
1729 * nobody will play with that bit on a non-LRU page.
1730 *
1731 * The downside is that we have to touch page->_count against each page.
1732 * But we had to alter page->flags anyway.
1733 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001734
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001735static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001736 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001737 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001738 enum lru_list lru)
1739{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001740 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001741 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001742 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001743 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001744
Wu Fengguang3eb41402009-06-16 15:33:13 -07001745 while (!list_empty(list)) {
1746 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001747 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001748
Sasha Levin309381fea2014-01-23 15:52:54 -08001749 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001750 SetPageLRU(page);
1751
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001752 nr_pages = hpage_nr_pages(page);
1753 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001754 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001755 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001756
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001757 if (put_page_testzero(page)) {
1758 __ClearPageLRU(page);
1759 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001760 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001761
1762 if (unlikely(PageCompound(page))) {
1763 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001764 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001765 (*get_compound_page_dtor(page))(page);
1766 spin_lock_irq(&zone->lru_lock);
1767 } else
1768 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001769 }
1770 }
1771 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1772 if (!is_active_lru(lru))
1773 __count_vm_events(PGDEACTIVATE, pgmoved);
1774}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001775
Hugh Dickinsf6260122012-01-12 17:20:06 -08001776static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001777 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001778 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001779 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001781 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001782 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001783 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001785 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001786 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001788 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001789 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001790 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001791 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001792 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793
1794 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001795
1796 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001797 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001798 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001799 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001800
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001802
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001803 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1804 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001805 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001806 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001807
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001808 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001809
Hugh Dickinsf6260122012-01-12 17:20:06 -08001810 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001811 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001812 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 spin_unlock_irq(&zone->lru_lock);
1814
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 while (!list_empty(&l_hold)) {
1816 cond_resched();
1817 page = lru_to_page(&l_hold);
1818 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001819
Hugh Dickins39b5f292012-10-08 16:33:18 -07001820 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001821 putback_lru_page(page);
1822 continue;
1823 }
1824
Mel Gormancc715d92012-03-21 16:34:00 -07001825 if (unlikely(buffer_heads_over_limit)) {
1826 if (page_has_private(page) && trylock_page(page)) {
1827 if (page_has_private(page))
1828 try_to_release_page(page, 0);
1829 unlock_page(page);
1830 }
1831 }
1832
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001833 if (page_referenced(page, 0, sc->target_mem_cgroup,
1834 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001835 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001836 /*
1837 * Identify referenced, file-backed active pages and
1838 * give them one more trip around the active list. So
1839 * that executable code get better chances to stay in
1840 * memory under moderate memory pressure. Anon pages
1841 * are not likely to be evicted by use-once streaming
1842 * IO, plus JVM can create lots of anon VM_EXEC pages,
1843 * so we ignore them here.
1844 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001845 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001846 list_add(&page->lru, &l_active);
1847 continue;
1848 }
1849 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001850
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001851 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 list_add(&page->lru, &l_inactive);
1853 }
1854
Andrew Mortonb5557492009-01-06 14:40:13 -08001855 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001856 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001857 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001858 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001859 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001860 * Count referenced pages from currently used mappings as rotated,
1861 * even though only some of them are actually re-activated. This
1862 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001863 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001864 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001865 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001866
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001867 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1868 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001869 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001870 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001871
Johannes Weiner747db952014-08-08 14:19:24 -07001872 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001873 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874}
1875
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001876#ifdef CONFIG_SWAP
Yaowei Bai42e2e452015-11-05 18:47:36 -08001877static bool inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001878{
1879 unsigned long active, inactive;
1880
1881 active = zone_page_state(zone, NR_ACTIVE_ANON);
1882 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1883
Yaowei Bai29d06bb2015-11-05 18:47:17 -08001884 return inactive * zone->inactive_ratio < active;
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001885}
1886
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001887/**
1888 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001889 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001890 *
1891 * Returns true if the zone does not have enough inactive anon pages,
1892 * meaning some active anon pages need to be deactivated.
1893 */
Yaowei Bai42e2e452015-11-05 18:47:36 -08001894static bool inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001895{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001896 /*
1897 * If we don't have swap space, anonymous page deactivation
1898 * is pointless.
1899 */
1900 if (!total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08001901 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001902
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001903 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001904 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001905
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001906 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001907}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001908#else
Yaowei Bai42e2e452015-11-05 18:47:36 -08001909static inline bool inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001910{
Yaowei Bai42e2e452015-11-05 18:47:36 -08001911 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001912}
1913#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001914
Rik van Riel56e49d22009-06-16 15:32:28 -07001915/**
1916 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001917 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001918 *
1919 * When the system is doing streaming IO, memory pressure here
1920 * ensures that active file pages get deactivated, until more
1921 * than half of the file pages are on the inactive list.
1922 *
1923 * Once we get to that situation, protect the system's working
1924 * set from being evicted by disabling active file page aging.
1925 *
1926 * This uses a different ratio than the anonymous pages, because
1927 * the page cache uses a use-once replacement algorithm.
1928 */
Yaowei Bai42e2e452015-11-05 18:47:36 -08001929static bool inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001930{
Johannes Weinere3790142013-02-22 16:35:19 -08001931 unsigned long inactive;
1932 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001933
Johannes Weinere3790142013-02-22 16:35:19 -08001934 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1935 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1936
1937 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001938}
1939
Yaowei Bai42e2e452015-11-05 18:47:36 -08001940static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001941{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001942 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001943 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001944 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001945 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001946}
1947
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001948static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001949 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001950{
Rik van Rielb39415b2009-12-14 17:59:48 -08001951 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001952 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001953 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001954 return 0;
1955 }
1956
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001957 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001958}
1959
Johannes Weiner9a265112013-02-22 16:32:17 -08001960enum scan_balance {
1961 SCAN_EQUAL,
1962 SCAN_FRACT,
1963 SCAN_ANON,
1964 SCAN_FILE,
1965};
1966
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001968 * Determine how aggressively the anon and file LRU lists should be
1969 * scanned. The relative value of each set of LRU lists is determined
1970 * by looking at the fraction of the pages scanned we did rotate back
1971 * onto the active list instead of evict.
1972 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001973 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1974 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001975 */
Vladimir Davydov33377672016-01-20 15:02:59 -08001976static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001977 struct scan_control *sc, unsigned long *nr,
1978 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001979{
Vladimir Davydov33377672016-01-20 15:02:59 -08001980 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001981 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001982 u64 fraction[2];
1983 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001984 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001985 unsigned long anon_prio, file_prio;
1986 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001987 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001988 bool force_scan = false;
1989 unsigned long ap, fp;
1990 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001991 bool some_scanned;
1992 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001993
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001994 /*
1995 * If the zone or memcg is small, nr[l] can be 0. This
1996 * results in no scanning on this priority and a potential
1997 * priority drop. Global direct reclaim can go to the next
1998 * zone and tends to have no problems. Global kswapd is for
1999 * zone balancing and it needs to scan a minimum amount. When
2000 * reclaiming for a memcg, a priority drop can cause high
2001 * latencies, so it's better to scan a minimum amount there as
2002 * well.
2003 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002004 if (current_is_kswapd()) {
2005 if (!zone_reclaimable(zone))
2006 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002007 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002008 force_scan = true;
2009 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002010 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002011 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002012
2013 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002014 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002015 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002016 goto out;
2017 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002018
Johannes Weiner10316b32013-02-22 16:32:14 -08002019 /*
2020 * Global reclaim will swap to prevent OOM even with no
2021 * swappiness, but memcg users want to use this knob to
2022 * disable swapping for individual groups completely when
2023 * using the memory controller's swap limit feature would be
2024 * too expensive.
2025 */
Johannes Weiner02695172014-08-06 16:06:17 -07002026 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002027 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002028 goto out;
2029 }
2030
2031 /*
2032 * Do not apply any pressure balancing cleverness when the
2033 * system is close to OOM, scan both anon and file equally
2034 * (unless the swappiness setting disagrees with swapping).
2035 */
Johannes Weiner02695172014-08-06 16:06:17 -07002036 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002037 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002038 goto out;
2039 }
2040
Johannes Weiner11d16c22013-02-22 16:32:15 -08002041 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002042 * Prevent the reclaimer from falling into the cache trap: as
2043 * cache pages start out inactive, every cache fault will tip
2044 * the scan balance towards the file LRU. And as the file LRU
2045 * shrinks, so does the window for rotation from references.
2046 * This means we have a runaway feedback loop where a tiny
2047 * thrashing file LRU becomes infinitely more attractive than
2048 * anon pages. Try to detect this based on file LRU size.
2049 */
2050 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002051 unsigned long zonefile;
2052 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07002053
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002054 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2055 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2056 zone_page_state(zone, NR_INACTIVE_FILE);
2057
2058 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07002059 scan_balance = SCAN_ANON;
2060 goto out;
2061 }
2062 }
2063
2064 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002065 * If there is enough inactive page cache, i.e. if the size of the
2066 * inactive list is greater than that of the active list *and* the
2067 * inactive list actually has some pages to scan on this priority, we
2068 * do not reclaim anything from the anonymous working set right now.
2069 * Without the second condition we could end up never scanning an
2070 * lruvec even if it has plenty of old anonymous pages unless the
2071 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002072 */
Vladimir Davydov316bda02016-01-14 15:19:38 -08002073 if (!inactive_file_is_low(lruvec) &&
2074 get_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002075 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002076 goto out;
2077 }
2078
Johannes Weiner9a265112013-02-22 16:32:17 -08002079 scan_balance = SCAN_FRACT;
2080
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002081 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002082 * With swappiness at 100, anonymous and file have the same priority.
2083 * This scanning priority is essentially the inverse of IO cost.
2084 */
Johannes Weiner02695172014-08-06 16:06:17 -07002085 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002086 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002087
2088 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002089 * OK, so we have swap space and a fair amount of page cache
2090 * pages. We use the recently rotated / recently scanned
2091 * ratios to determine how valuable each cache is.
2092 *
2093 * Because workloads change over time (and to avoid overflow)
2094 * we keep these statistics as a floating average, which ends
2095 * up weighing recent references more than old ones.
2096 *
2097 * anon in [0], file in [1]
2098 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002099
2100 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2101 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2102 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2103 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2104
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002105 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002106 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002107 reclaim_stat->recent_scanned[0] /= 2;
2108 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002109 }
2110
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002111 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002112 reclaim_stat->recent_scanned[1] /= 2;
2113 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002114 }
2115
2116 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002117 * The amount of pressure on anon vs file pages is inversely
2118 * proportional to the fraction of recently scanned pages on
2119 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002120 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002121 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002122 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002123
Satoru Moriyafe350042012-05-29 15:06:47 -07002124 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002125 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002126 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002127
Shaohua Li76a33fc2010-05-24 14:32:36 -07002128 fraction[0] = ap;
2129 fraction[1] = fp;
2130 denominator = ap + fp + 1;
2131out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002132 some_scanned = false;
2133 /* Only use force_scan on second pass. */
2134 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002135 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002136 for_each_evictable_lru(lru) {
2137 int file = is_file_lru(lru);
2138 unsigned long size;
2139 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002140
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002141 size = get_lru_size(lruvec, lru);
2142 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002143
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002144 if (!scan && pass && force_scan)
2145 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002146
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002147 switch (scan_balance) {
2148 case SCAN_EQUAL:
2149 /* Scan lists relative to size */
2150 break;
2151 case SCAN_FRACT:
2152 /*
2153 * Scan types proportional to swappiness and
2154 * their relative recent reclaim efficiency.
2155 */
2156 scan = div64_u64(scan * fraction[file],
2157 denominator);
2158 break;
2159 case SCAN_FILE:
2160 case SCAN_ANON:
2161 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002162 if ((scan_balance == SCAN_FILE) != file) {
2163 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002164 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002165 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002166 break;
2167 default:
2168 /* Look ma, no brain */
2169 BUG();
2170 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002171
2172 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002173 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002174
Johannes Weiner9a265112013-02-22 16:32:17 -08002175 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002176 * Skip the second pass and don't force_scan,
2177 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002178 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002179 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002180 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002181 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002182}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002183
Mel Gorman72b252a2015-09-04 15:47:32 -07002184#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2185static void init_tlb_ubc(void)
2186{
2187 /*
2188 * This deliberately does not clear the cpumask as it's expensive
2189 * and unnecessary. If there happens to be data in there then the
2190 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2191 * then will be cleared.
2192 */
2193 current->tlb_ubc.flush_required = false;
2194}
2195#else
2196static inline void init_tlb_ubc(void)
2197{
2198}
2199#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2200
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002201/*
2202 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2203 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002204static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2205 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002206{
Vladimir Davydov33377672016-01-20 15:02:59 -08002207 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002208 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002209 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002210 unsigned long nr_to_scan;
2211 enum lru_list lru;
2212 unsigned long nr_reclaimed = 0;
2213 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2214 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002215 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002216
Vladimir Davydov33377672016-01-20 15:02:59 -08002217 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002218
Mel Gormane82e0562013-07-03 15:01:44 -07002219 /* Record the original scan target for proportional adjustments later */
2220 memcpy(targets, nr, sizeof(nr));
2221
Mel Gorman1a501902014-06-04 16:10:49 -07002222 /*
2223 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2224 * event that can occur when there is little memory pressure e.g.
2225 * multiple streaming readers/writers. Hence, we do not abort scanning
2226 * when the requested number of pages are reclaimed when scanning at
2227 * DEF_PRIORITY on the assumption that the fact we are direct
2228 * reclaiming implies that kswapd is not keeping up and it is best to
2229 * do a batch of work at once. For memcg reclaim one check is made to
2230 * abort proportional reclaim if either the file or anon lru has already
2231 * dropped to zero at the first pass.
2232 */
2233 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2234 sc->priority == DEF_PRIORITY);
2235
Mel Gorman72b252a2015-09-04 15:47:32 -07002236 init_tlb_ubc();
2237
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002238 blk_start_plug(&plug);
2239 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2240 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002241 unsigned long nr_anon, nr_file, percentage;
2242 unsigned long nr_scanned;
2243
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002244 for_each_evictable_lru(lru) {
2245 if (nr[lru]) {
2246 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2247 nr[lru] -= nr_to_scan;
2248
2249 nr_reclaimed += shrink_list(lru, nr_to_scan,
2250 lruvec, sc);
2251 }
2252 }
Mel Gormane82e0562013-07-03 15:01:44 -07002253
2254 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2255 continue;
2256
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002257 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002258 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002259 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002260 * proportionally what was requested by get_scan_count(). We
2261 * stop reclaiming one LRU and reduce the amount scanning
2262 * proportional to the original scan target.
2263 */
2264 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2265 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2266
Mel Gorman1a501902014-06-04 16:10:49 -07002267 /*
2268 * It's just vindictive to attack the larger once the smaller
2269 * has gone to zero. And given the way we stop scanning the
2270 * smaller below, this makes sure that we only make one nudge
2271 * towards proportionality once we've got nr_to_reclaim.
2272 */
2273 if (!nr_file || !nr_anon)
2274 break;
2275
Mel Gormane82e0562013-07-03 15:01:44 -07002276 if (nr_file > nr_anon) {
2277 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2278 targets[LRU_ACTIVE_ANON] + 1;
2279 lru = LRU_BASE;
2280 percentage = nr_anon * 100 / scan_target;
2281 } else {
2282 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2283 targets[LRU_ACTIVE_FILE] + 1;
2284 lru = LRU_FILE;
2285 percentage = nr_file * 100 / scan_target;
2286 }
2287
2288 /* Stop scanning the smaller of the LRU */
2289 nr[lru] = 0;
2290 nr[lru + LRU_ACTIVE] = 0;
2291
2292 /*
2293 * Recalculate the other LRU scan count based on its original
2294 * scan target and the percentage scanning already complete
2295 */
2296 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2297 nr_scanned = targets[lru] - nr[lru];
2298 nr[lru] = targets[lru] * (100 - percentage) / 100;
2299 nr[lru] -= min(nr[lru], nr_scanned);
2300
2301 lru += LRU_ACTIVE;
2302 nr_scanned = targets[lru] - nr[lru];
2303 nr[lru] = targets[lru] * (100 - percentage) / 100;
2304 nr[lru] -= min(nr[lru], nr_scanned);
2305
2306 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002307 }
2308 blk_finish_plug(&plug);
2309 sc->nr_reclaimed += nr_reclaimed;
2310
2311 /*
2312 * Even if we did not try to evict anon pages at all, we want to
2313 * rebalance the anon lru active/inactive ratio.
2314 */
2315 if (inactive_anon_is_low(lruvec))
2316 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2317 sc, LRU_ACTIVE_ANON);
2318
2319 throttle_vm_writeout(sc->gfp_mask);
2320}
2321
Mel Gorman23b9da52012-05-29 15:06:20 -07002322/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002323static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002324{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002325 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002326 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002327 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002328 return true;
2329
2330 return false;
2331}
2332
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002333/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002334 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2335 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2336 * true if more pages should be reclaimed such that when the page allocator
2337 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2338 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002339 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002340static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002341 unsigned long nr_reclaimed,
2342 unsigned long nr_scanned,
2343 struct scan_control *sc)
2344{
2345 unsigned long pages_for_compaction;
2346 unsigned long inactive_lru_pages;
2347
2348 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002349 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002350 return false;
2351
Mel Gorman28765922011-02-25 14:44:20 -08002352 /* Consider stopping depending on scan and reclaim activity */
2353 if (sc->gfp_mask & __GFP_REPEAT) {
2354 /*
2355 * For __GFP_REPEAT allocations, stop reclaiming if the
2356 * full LRU list has been scanned and we are still failing
2357 * to reclaim pages. This full LRU scan is potentially
2358 * expensive but a __GFP_REPEAT caller really wants to succeed
2359 */
2360 if (!nr_reclaimed && !nr_scanned)
2361 return false;
2362 } else {
2363 /*
2364 * For non-__GFP_REPEAT allocations which can presumably
2365 * fail without consequence, stop if we failed to reclaim
2366 * any pages from the last SWAP_CLUSTER_MAX number of
2367 * pages that were scanned. This will return to the
2368 * caller faster at the risk reclaim/compaction and
2369 * the resulting allocation attempt fails
2370 */
2371 if (!nr_reclaimed)
2372 return false;
2373 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002374
2375 /*
2376 * If we have not reclaimed enough pages for compaction and the
2377 * inactive lists are large enough, continue reclaiming
2378 */
2379 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002380 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002381 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002382 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002383 if (sc->nr_reclaimed < pages_for_compaction &&
2384 inactive_lru_pages > pages_for_compaction)
2385 return true;
2386
2387 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002388 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002389 case COMPACT_PARTIAL:
2390 case COMPACT_CONTINUE:
2391 return false;
2392 default:
2393 return true;
2394 }
2395}
2396
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002397static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2398 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002399{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002400 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002401 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002402 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002403
Johannes Weiner56600482012-01-12 17:17:59 -08002404 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002405 struct mem_cgroup *root = sc->target_mem_cgroup;
2406 struct mem_cgroup_reclaim_cookie reclaim = {
2407 .zone = zone,
2408 .priority = sc->priority,
2409 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002410 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002411 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002412
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002413 nr_reclaimed = sc->nr_reclaimed;
2414 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002415
Andrew Morton694fbc02013-09-24 15:27:37 -07002416 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2417 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002418 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002419 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002420 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002421
Johannes Weiner241994e2015-02-11 15:26:06 -08002422 if (mem_cgroup_low(root, memcg)) {
2423 if (!sc->may_thrash)
2424 continue;
2425 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2426 }
2427
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002428 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002429 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002430
Vladimir Davydov33377672016-01-20 15:02:59 -08002431 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002432 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002433
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002434 if (memcg && is_classzone)
2435 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2436 memcg, sc->nr_scanned - scanned,
2437 lru_pages);
2438
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002439 /* Record the group's reclaim efficiency */
2440 vmpressure(sc->gfp_mask, memcg, false,
2441 sc->nr_scanned - scanned,
2442 sc->nr_reclaimed - reclaimed);
2443
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002444 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002445 * Direct reclaim and kswapd have to scan all memory
2446 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002447 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002448 *
2449 * Limit reclaim, on the other hand, only cares about
2450 * nr_to_reclaim pages to be reclaimed and it will
2451 * retry with decreasing priority if one round over the
2452 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002453 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002454 if (!global_reclaim(sc) &&
2455 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002456 mem_cgroup_iter_break(root, memcg);
2457 break;
2458 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002459 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002460
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002461 /*
2462 * Shrink the slab caches in the same proportion that
2463 * the eligible LRU pages were scanned.
2464 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002465 if (global_reclaim(sc) && is_classzone)
2466 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2467 sc->nr_scanned - nr_scanned,
2468 zone_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002469
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002470 if (reclaim_state) {
2471 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2472 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002473 }
2474
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002475 /* Record the subtree's reclaim efficiency */
2476 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002477 sc->nr_scanned - nr_scanned,
2478 sc->nr_reclaimed - nr_reclaimed);
2479
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002480 if (sc->nr_reclaimed - nr_reclaimed)
2481 reclaimable = true;
2482
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002483 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2484 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002485
2486 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002487}
2488
Vlastimil Babka53853e22014-10-09 15:27:02 -07002489/*
2490 * Returns true if compaction should go ahead for a high-order request, or
2491 * the high-order allocation would succeed without compaction.
2492 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002493static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002494{
2495 unsigned long balance_gap, watermark;
2496 bool watermark_ok;
2497
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002498 /*
2499 * Compaction takes time to run and there are potentially other
2500 * callers using the pages just freed. Continue reclaiming until
2501 * there is a buffer of free pages available to give compaction
2502 * a reasonable chance of completing and allocating the page
2503 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002504 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2505 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002506 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormane2b19192015-11-06 16:28:09 -08002507 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002508
2509 /*
2510 * If compaction is deferred, reclaim up to a point where
2511 * compaction will have a chance of success when re-enabled
2512 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002513 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002514 return watermark_ok;
2515
Vlastimil Babka53853e22014-10-09 15:27:02 -07002516 /*
2517 * If compaction is not ready to start and allocation is not likely
2518 * to succeed without it, then keep reclaiming.
2519 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002520 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002521 return false;
2522
2523 return watermark_ok;
2524}
2525
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526/*
2527 * This is the direct reclaim path, for page-allocating processes. We only
2528 * try to reclaim pages from zones which will satisfy the caller's allocation
2529 * request.
2530 *
Mel Gorman41858962009-06-16 15:32:12 -07002531 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2532 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2534 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002535 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2536 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2537 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 * If a zone is deemed to be full of pinned pages then just give it a light
2540 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002541 *
2542 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002544static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545{
Mel Gormandd1a2392008-04-28 02:12:17 -07002546 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002547 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002548 unsigned long nr_soft_reclaimed;
2549 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002550 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002551 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002552 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002553
Mel Gormancc715d92012-03-21 16:34:00 -07002554 /*
2555 * If the number of buffer_heads in the machine exceeds the maximum
2556 * allowed level, force direct reclaim to scan the highmem zone as
2557 * highmem pages could be pinning lowmem pages storing buffer_heads
2558 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002559 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002560 if (buffer_heads_over_limit)
2561 sc->gfp_mask |= __GFP_HIGHMEM;
2562
Mel Gormand4debc62010-08-09 17:19:29 -07002563 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002564 requested_highidx, sc->nodemask) {
2565 enum zone_type classzone_idx;
2566
Con Kolivasf3fe6512006-01-06 00:11:15 -08002567 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002569
2570 classzone_idx = requested_highidx;
2571 while (!populated_zone(zone->zone_pgdat->node_zones +
2572 classzone_idx))
2573 classzone_idx--;
2574
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002575 /*
2576 * Take care memory controller reclaiming has small influence
2577 * to global LRU.
2578 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002579 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002580 if (!cpuset_zone_allowed(zone,
2581 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002582 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002583
Lisa Du6e543d52013-09-11 14:22:36 -07002584 if (sc->priority != DEF_PRIORITY &&
2585 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002586 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002587
2588 /*
2589 * If we already have plenty of memory free for
2590 * compaction in this zone, don't free any more.
2591 * Even though compaction is invoked for any
2592 * non-zero order, only frequent costly order
2593 * reclamation is disruptive enough to become a
2594 * noticeable problem, like transparent huge
2595 * page allocations.
2596 */
2597 if (IS_ENABLED(CONFIG_COMPACTION) &&
2598 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2599 zonelist_zone_idx(z) <= requested_highidx &&
2600 compaction_ready(zone, sc->order)) {
2601 sc->compaction_ready = true;
2602 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002603 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002604
Andrew Morton0608f432013-09-24 15:27:41 -07002605 /*
2606 * This steals pages from memory cgroups over softlimit
2607 * and returns the number of reclaimed pages and
2608 * scanned pages. This works for global memory pressure
2609 * and balancing, not for a memcg's limit.
2610 */
2611 nr_soft_scanned = 0;
2612 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2613 sc->order, sc->gfp_mask,
2614 &nr_soft_scanned);
2615 sc->nr_reclaimed += nr_soft_reclaimed;
2616 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002617 if (nr_soft_reclaimed)
2618 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002619 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002620 }
Nick Piggin408d8542006-09-25 23:31:27 -07002621
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002622 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002623 reclaimable = true;
2624
2625 if (global_reclaim(sc) &&
2626 !reclaimable && zone_reclaimable(zone))
2627 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 }
Mel Gormane0c23272011-10-31 17:09:33 -07002629
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002630 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002631 * Restore to original mask to avoid the impact on the caller if we
2632 * promoted it to __GFP_HIGHMEM.
2633 */
2634 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002635
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002636 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002638
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639/*
2640 * This is the main entry point to direct page reclaim.
2641 *
2642 * If a full scan of the inactive list fails to free enough memory then we
2643 * are "out of memory" and something needs to be killed.
2644 *
2645 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2646 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002647 * caller can't do much about. We kick the writeback threads and take explicit
2648 * naps in the hope that some of these pages can be written. But if the
2649 * allocating task holds filesystem locks which prevent writeout this might not
2650 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002651 *
2652 * returns: 0, if no pages reclaimed
2653 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 */
Mel Gormandac1d272008-04-28 02:12:12 -07002655static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002656 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657{
Johannes Weiner241994e2015-02-11 15:26:06 -08002658 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002659 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002660 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002661 bool zones_reclaimable;
Johannes Weiner241994e2015-02-11 15:26:06 -08002662retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002663 delayacct_freepages_start();
2664
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002665 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002666 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002668 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002669 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2670 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002671 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002672 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002673
Balbir Singh66e17072008-02-07 00:13:56 -08002674 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002675 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002676 break;
2677
2678 if (sc->compaction_ready)
2679 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
2681 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002682 * If we're getting trouble reclaiming, start doing
2683 * writepage even in laptop mode.
2684 */
2685 if (sc->priority < DEF_PRIORITY - 2)
2686 sc->may_writepage = 1;
2687
2688 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 * Try to write back as many pages as we just scanned. This
2690 * tends to cause slow streaming writers to write data to the
2691 * disk smoothly, at the dirtying rate, which is nice. But
2692 * that's undesirable in laptop mode, where we *want* lumpy
2693 * writeout. So in laptop mode, write out the whole world.
2694 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002695 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2696 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002697 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2698 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002699 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002701 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002702
Keika Kobayashi873b4772008-07-25 01:48:52 -07002703 delayacct_freepages_end();
2704
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002705 if (sc->nr_reclaimed)
2706 return sc->nr_reclaimed;
2707
Mel Gorman0cee34f2012-01-12 17:19:49 -08002708 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002709 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002710 return 1;
2711
Johannes Weiner241994e2015-02-11 15:26:06 -08002712 /* Untapped cgroup reserves? Don't OOM, retry. */
2713 if (!sc->may_thrash) {
2714 sc->priority = initial_priority;
2715 sc->may_thrash = 1;
2716 goto retry;
2717 }
2718
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002719 /* Any of the zones still reclaimable? Don't OOM. */
2720 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002721 return 1;
2722
2723 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724}
2725
Mel Gorman55150612012-07-31 16:44:35 -07002726static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2727{
2728 struct zone *zone;
2729 unsigned long pfmemalloc_reserve = 0;
2730 unsigned long free_pages = 0;
2731 int i;
2732 bool wmark_ok;
2733
2734 for (i = 0; i <= ZONE_NORMAL; i++) {
2735 zone = &pgdat->node_zones[i];
Nishanth Aravamudanf012a842015-06-24 16:56:39 -07002736 if (!populated_zone(zone) ||
2737 zone_reclaimable_pages(zone) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002738 continue;
2739
Mel Gorman55150612012-07-31 16:44:35 -07002740 pfmemalloc_reserve += min_wmark_pages(zone);
2741 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2742 }
2743
Mel Gorman675becc2014-06-04 16:07:35 -07002744 /* If there are no reserves (unexpected config) then do not throttle */
2745 if (!pfmemalloc_reserve)
2746 return true;
2747
Mel Gorman55150612012-07-31 16:44:35 -07002748 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2749
2750 /* kswapd must be awake if processes are being throttled */
2751 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2752 pgdat->classzone_idx = min(pgdat->classzone_idx,
2753 (enum zone_type)ZONE_NORMAL);
2754 wake_up_interruptible(&pgdat->kswapd_wait);
2755 }
2756
2757 return wmark_ok;
2758}
2759
2760/*
2761 * Throttle direct reclaimers if backing storage is backed by the network
2762 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2763 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002764 * when the low watermark is reached.
2765 *
2766 * Returns true if a fatal signal was delivered during throttling. If this
2767 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002768 */
Mel Gorman50694c22012-11-26 16:29:48 -08002769static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002770 nodemask_t *nodemask)
2771{
Mel Gorman675becc2014-06-04 16:07:35 -07002772 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002773 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002774 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002775
2776 /*
2777 * Kernel threads should not be throttled as they may be indirectly
2778 * responsible for cleaning pages necessary for reclaim to make forward
2779 * progress. kjournald for example may enter direct reclaim while
2780 * committing a transaction where throttling it could forcing other
2781 * processes to block on log_wait_commit().
2782 */
2783 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002784 goto out;
2785
2786 /*
2787 * If a fatal signal is pending, this process should not throttle.
2788 * It should return quickly so it can exit and free its memory
2789 */
2790 if (fatal_signal_pending(current))
2791 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002792
Mel Gorman675becc2014-06-04 16:07:35 -07002793 /*
2794 * Check if the pfmemalloc reserves are ok by finding the first node
2795 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2796 * GFP_KERNEL will be required for allocating network buffers when
2797 * swapping over the network so ZONE_HIGHMEM is unusable.
2798 *
2799 * Throttling is based on the first usable node and throttled processes
2800 * wait on a queue until kswapd makes progress and wakes them. There
2801 * is an affinity then between processes waking up and where reclaim
2802 * progress has been made assuming the process wakes on the same node.
2803 * More importantly, processes running on remote nodes will not compete
2804 * for remote pfmemalloc reserves and processes on different nodes
2805 * should make reasonable progress.
2806 */
2807 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002808 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002809 if (zone_idx(zone) > ZONE_NORMAL)
2810 continue;
2811
2812 /* Throttle based on the first usable node */
2813 pgdat = zone->zone_pgdat;
2814 if (pfmemalloc_watermark_ok(pgdat))
2815 goto out;
2816 break;
2817 }
2818
2819 /* If no zone was usable by the allocation flags then do not throttle */
2820 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002821 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002822
Mel Gorman68243e72012-07-31 16:44:39 -07002823 /* Account for the throttling */
2824 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2825
Mel Gorman55150612012-07-31 16:44:35 -07002826 /*
2827 * If the caller cannot enter the filesystem, it's possible that it
2828 * is due to the caller holding an FS lock or performing a journal
2829 * transaction in the case of a filesystem like ext[3|4]. In this case,
2830 * it is not safe to block on pfmemalloc_wait as kswapd could be
2831 * blocked waiting on the same lock. Instead, throttle for up to a
2832 * second before continuing.
2833 */
2834 if (!(gfp_mask & __GFP_FS)) {
2835 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2836 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002837
2838 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002839 }
2840
2841 /* Throttle until kswapd wakes the process */
2842 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2843 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002844
2845check_pending:
2846 if (fatal_signal_pending(current))
2847 return true;
2848
2849out:
2850 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002851}
2852
Mel Gormandac1d272008-04-28 02:12:12 -07002853unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002854 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002855{
Mel Gorman33906bc2010-08-09 17:19:16 -07002856 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002857 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002858 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002859 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2860 .order = order,
2861 .nodemask = nodemask,
2862 .priority = DEF_PRIORITY,
2863 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002864 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002865 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002866 };
2867
Mel Gorman55150612012-07-31 16:44:35 -07002868 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002869 * Do not enter reclaim if fatal signal was delivered while throttled.
2870 * 1 is returned so that the page allocator does not OOM kill at this
2871 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002872 */
Mel Gorman50694c22012-11-26 16:29:48 -08002873 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002874 return 1;
2875
Mel Gorman33906bc2010-08-09 17:19:16 -07002876 trace_mm_vmscan_direct_reclaim_begin(order,
2877 sc.may_writepage,
2878 gfp_mask);
2879
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002880 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002881
2882 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2883
2884 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002885}
2886
Andrew Mortonc255a452012-07-31 16:43:02 -07002887#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002888
Johannes Weiner72835c82012-01-12 17:18:32 -08002889unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002890 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002891 struct zone *zone,
2892 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002893{
2894 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002895 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002896 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002897 .may_writepage = !laptop_mode,
2898 .may_unmap = 1,
2899 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002900 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002901 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002902
Balbir Singh4e416952009-09-23 15:56:39 -07002903 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2904 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002905
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002906 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002907 sc.may_writepage,
2908 sc.gfp_mask);
2909
Balbir Singh4e416952009-09-23 15:56:39 -07002910 /*
2911 * NOTE: Although we can get the priority field, using it
2912 * here is not a good idea, since it limits the pages we can scan.
2913 * if we don't reclaim here, the shrink_zone from balance_pgdat
2914 * will pick up pages from other mem cgroup's as well. We hack
2915 * the priority and make it zero.
2916 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002917 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002918
2919 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2920
Ying Han0ae5e892011-05-26 16:25:25 -07002921 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002922 return sc.nr_reclaimed;
2923}
2924
Johannes Weiner72835c82012-01-12 17:18:32 -08002925unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002926 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002927 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002928 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002929{
Balbir Singh4e416952009-09-23 15:56:39 -07002930 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002931 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002932 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002933 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002934 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002935 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2936 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2937 .target_mem_cgroup = memcg,
2938 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002939 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002940 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002941 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002942 };
Balbir Singh66e17072008-02-07 00:13:56 -08002943
Ying Han889976d2011-05-26 16:25:33 -07002944 /*
2945 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2946 * take care of from where we get pages. So the node where we start the
2947 * scan does not need to be the current node.
2948 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002949 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002950
2951 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002952
2953 trace_mm_vmscan_memcg_reclaim_begin(0,
2954 sc.may_writepage,
2955 sc.gfp_mask);
2956
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002957 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002958
2959 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2960
2961 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002962}
2963#endif
2964
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002965static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002966{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002967 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002968
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002969 if (!total_swap_pages)
2970 return;
2971
2972 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2973 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002974 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002975
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002976 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002977 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002978 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002979
2980 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2981 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002982}
2983
Johannes Weiner60cefed2012-11-29 13:54:23 -08002984static bool zone_balanced(struct zone *zone, int order,
2985 unsigned long balance_gap, int classzone_idx)
2986{
2987 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
Mel Gormane2b19192015-11-06 16:28:09 -08002988 balance_gap, classzone_idx))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002989 return false;
2990
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002991 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2992 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002993 return false;
2994
2995 return true;
2996}
2997
Mel Gorman1741c872011-01-13 15:46:21 -08002998/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002999 * pgdat_balanced() is used when checking if a node is balanced.
3000 *
3001 * For order-0, all zones must be balanced!
3002 *
3003 * For high-order allocations only zones that meet watermarks and are in a
3004 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3005 * total of balanced pages must be at least 25% of the zones allowed by
3006 * classzone_idx for the node to be considered balanced. Forcing all zones to
3007 * be balanced for high orders can cause excessive reclaim when there are
3008 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08003009 * The choice of 25% is due to
3010 * o a 16M DMA zone that is balanced will not balance a zone on any
3011 * reasonable sized machine
3012 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003013 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08003014 * would need to be at least 256M for it to be balance a whole node.
3015 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3016 * to balance a node on its own. These seemed like reasonable ratios.
3017 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003018static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08003019{
Jiang Liub40da042013-02-22 16:33:52 -08003020 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003021 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08003022 int i;
3023
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003024 /* Check the watermark levels */
3025 for (i = 0; i <= classzone_idx; i++) {
3026 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08003027
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003028 if (!populated_zone(zone))
3029 continue;
3030
Jiang Liub40da042013-02-22 16:33:52 -08003031 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003032
3033 /*
3034 * A special case here:
3035 *
3036 * balance_pgdat() skips over all_unreclaimable after
3037 * DEF_PRIORITY. Effectively, it considers them balanced so
3038 * they must be considered balanced here as well!
3039 */
Lisa Du6e543d52013-09-11 14:22:36 -07003040 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08003041 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003042 continue;
3043 }
3044
3045 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08003046 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003047 else if (!order)
3048 return false;
3049 }
3050
3051 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08003052 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003053 else
3054 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08003055}
3056
Mel Gorman55150612012-07-31 16:44:35 -07003057/*
3058 * Prepare kswapd for sleeping. This verifies that there are no processes
3059 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3060 *
3061 * Returns true if kswapd is ready to sleep
3062 */
3063static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08003064 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003065{
Mel Gormanf50de2d2009-12-14 17:58:53 -08003066 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3067 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07003068 return false;
3069
3070 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003071 * The throttled processes are normally woken up in balance_pgdat() as
3072 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3073 * race between when kswapd checks the watermarks and a process gets
3074 * throttled. There is also a potential race if processes get
3075 * throttled, kswapd wakes, a large process exits thereby balancing the
3076 * zones, which causes kswapd to exit balance_pgdat() before reaching
3077 * the wake up checks. If kswapd is going to sleep, no process should
3078 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3079 * the wake up is premature, processes will wake kswapd and get
3080 * throttled again. The difference from wake ups in balance_pgdat() is
3081 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003082 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003083 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3084 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003085
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003086 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003087}
3088
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089/*
Mel Gorman75485362013-07-03 15:01:42 -07003090 * kswapd shrinks the zone by the number of pages required to reach
3091 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003092 *
3093 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003094 * reclaim or if the lack of progress was due to pages under writeback.
3095 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003096 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003097static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07003098 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07003099 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07003100 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07003101{
Mel Gorman7c954f62013-07-03 15:01:54 -07003102 int testorder = sc->order;
3103 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07003104 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07003105
3106 /* Reclaim above the high watermark. */
3107 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07003108
3109 /*
3110 * Kswapd reclaims only single pages with compaction enabled. Trying
3111 * too hard to reclaim until contiguous free pages have become
3112 * available can hurt performance by evicting too much useful data
3113 * from memory. Do not reclaim more than needed for compaction.
3114 */
3115 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08003116 compaction_suitable(zone, sc->order, 0, classzone_idx)
3117 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07003118 testorder = 0;
3119
3120 /*
3121 * We put equal pressure on every zone, unless one zone has way too
3122 * many pages free already. The "too many pages" is defined as the
3123 * high wmark plus a "gap" where the gap is either the low
3124 * watermark or 1% of the zone, whichever is smaller.
3125 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07003126 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3127 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003128
3129 /*
3130 * If there is no low memory pressure or the zone is balanced then no
3131 * reclaim is necessary
3132 */
3133 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3134 if (!lowmem_pressure && zone_balanced(zone, testorder,
3135 balance_gap, classzone_idx))
3136 return true;
3137
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003138 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003139
Mel Gorman2ab44f42013-07-03 15:01:47 -07003140 /* Account for the number of pages attempted to reclaim */
3141 *nr_attempted += sc->nr_to_reclaim;
3142
Johannes Weiner57054652014-10-09 15:28:17 -07003143 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003144
Mel Gorman7c954f62013-07-03 15:01:54 -07003145 /*
3146 * If a zone reaches its high watermark, consider it to be no longer
3147 * congested. It's possible there are dirty pages backed by congested
3148 * BDIs but as pressure is relieved, speculatively avoid congestion
3149 * waits.
3150 */
Lisa Du6e543d52013-09-11 14:22:36 -07003151 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003152 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003153 clear_bit(ZONE_CONGESTED, &zone->flags);
3154 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003155 }
3156
Mel Gormanb8e83b92013-07-03 15:01:45 -07003157 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003158}
3159
3160/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003162 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003164 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 *
3166 * There is special handling here for zones which are full of pinned pages.
3167 * This can happen if the pages are all mlocked, or if they are all used by
3168 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3169 * What we do is to detect the case where all pages in the zone have been
3170 * scanned twice and there has been zero successful reclaim. Mark the zone as
3171 * dead and from now on, only perform a short scan. Basically we're polling
3172 * the zone for when the problem goes away.
3173 *
3174 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003175 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3176 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3177 * lower zones regardless of the number of free pages in the lower zones. This
3178 * interoperates with the page allocator fallback scheme to ensure that aging
3179 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 */
Mel Gorman99504742011-01-13 15:46:20 -08003181static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003182 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003185 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003186 unsigned long nr_soft_reclaimed;
3187 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003188 struct scan_control sc = {
3189 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003190 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003191 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003192 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003193 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003194 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003195 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003196 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003198 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003199 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003200 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003201 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003202
3203 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003205 /*
3206 * Scan in the highmem->dma direction for the highest
3207 * zone which needs scanning
3208 */
3209 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3210 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003212 if (!populated_zone(zone))
3213 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214
Lisa Du6e543d52013-09-11 14:22:36 -07003215 if (sc.priority != DEF_PRIORITY &&
3216 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003217 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218
Rik van Riel556adec2008-10-18 20:26:34 -07003219 /*
3220 * Do some background aging of the anon list, to give
3221 * pages a chance to be referenced before reclaiming.
3222 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003223 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003224
Mel Gormancc715d92012-03-21 16:34:00 -07003225 /*
3226 * If the number of buffer_heads in the machine
3227 * exceeds the maximum allowed level and this node
3228 * has a highmem zone, force kswapd to reclaim from
3229 * it to relieve lowmem pressure.
3230 */
3231 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3232 end_zone = i;
3233 break;
3234 }
3235
Johannes Weiner60cefed2012-11-29 13:54:23 -08003236 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003237 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003238 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003239 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003240 /*
3241 * If balanced, clear the dirty and congested
3242 * flags
3243 */
Johannes Weiner57054652014-10-09 15:28:17 -07003244 clear_bit(ZONE_CONGESTED, &zone->flags);
3245 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003248
Mel Gormanb8e83b92013-07-03 15:01:45 -07003249 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003250 goto out;
3251
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 for (i = 0; i <= end_zone; i++) {
3253 struct zone *zone = pgdat->node_zones + i;
3254
Mel Gorman2ab44f42013-07-03 15:01:47 -07003255 if (!populated_zone(zone))
3256 continue;
3257
Mel Gorman2ab44f42013-07-03 15:01:47 -07003258 /*
3259 * If any zone is currently balanced then kswapd will
3260 * not call compaction as it is expected that the
3261 * necessary pages are already available.
3262 */
3263 if (pgdat_needs_compaction &&
3264 zone_watermark_ok(zone, order,
3265 low_wmark_pages(zone),
3266 *classzone_idx, 0))
3267 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 }
3269
3270 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003271 * If we're getting trouble reclaiming, start doing writepage
3272 * even in laptop mode.
3273 */
3274 if (sc.priority < DEF_PRIORITY - 2)
3275 sc.may_writepage = 1;
3276
3277 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 * Now scan the zone in the dma->highmem direction, stopping
3279 * at the last zone which needs scanning.
3280 *
3281 * We do this because the page allocator works in the opposite
3282 * direction. This prevents the page allocator from allocating
3283 * pages behind kswapd's direction of progress, which would
3284 * cause too much scanning of the lower zones.
3285 */
3286 for (i = 0; i <= end_zone; i++) {
3287 struct zone *zone = pgdat->node_zones + i;
3288
Con Kolivasf3fe6512006-01-06 00:11:15 -08003289 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 continue;
3291
Lisa Du6e543d52013-09-11 14:22:36 -07003292 if (sc.priority != DEF_PRIORITY &&
3293 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 continue;
3295
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003297
Andrew Morton0608f432013-09-24 15:27:41 -07003298 nr_soft_scanned = 0;
3299 /*
3300 * Call soft limit reclaim before calling shrink_zone.
3301 */
3302 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3303 order, sc.gfp_mask,
3304 &nr_soft_scanned);
3305 sc.nr_reclaimed += nr_soft_reclaimed;
3306
Rik van Riel32a43302007-10-16 01:24:50 -07003307 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003308 * There should be no need to raise the scanning
3309 * priority if enough pages are already being scanned
3310 * that that high watermark would be met at 100%
3311 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003312 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003313 if (kswapd_shrink_zone(zone, end_zone,
3314 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003315 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 }
Mel Gorman55150612012-07-31 16:44:35 -07003317
3318 /*
3319 * If the low watermark is met there is no need for processes
3320 * to be throttled on pfmemalloc_wait as they should not be
3321 * able to safely make forward progress. Wake them
3322 */
3323 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3324 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003325 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003326
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003328 * Fragmentation may mean that the system cannot be rebalanced
3329 * for high-order allocations in all zones. If twice the
3330 * allocation size has been reclaimed and the zones are still
3331 * not balanced then recheck the watermarks at order-0 to
3332 * prevent kswapd reclaiming excessively. Assume that a
3333 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003335 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003336 order = sc.order = 0;
3337
Mel Gormanb8e83b92013-07-03 15:01:45 -07003338 /* Check if kswapd should be suspending */
3339 if (try_to_freeze() || kthread_should_stop())
3340 break;
3341
3342 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003343 * Compact if necessary and kswapd is reclaiming at least the
3344 * high watermark number of pages as requsted
3345 */
3346 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3347 compact_pgdat(pgdat, order);
3348
3349 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003350 * Raise priority if scanning rate is too low or there was no
3351 * progress in reclaiming pages
3352 */
3353 if (raise_priority || !sc.nr_reclaimed)
3354 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003355 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003356 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357
Mel Gormanb8e83b92013-07-03 15:01:45 -07003358out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003359 /*
Mel Gorman55150612012-07-31 16:44:35 -07003360 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003361 * makes a decision on the order we were last reclaiming at. However,
3362 * if another caller entered the allocator slow path while kswapd
3363 * was awake, order will remain at the higher level
3364 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003365 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003366 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367}
3368
Mel Gormandc83edd2011-01-13 15:46:26 -08003369static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003370{
3371 long remaining = 0;
3372 DEFINE_WAIT(wait);
3373
3374 if (freezing(current) || kthread_should_stop())
3375 return;
3376
3377 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3378
3379 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003380 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003381 remaining = schedule_timeout(HZ/10);
3382 finish_wait(&pgdat->kswapd_wait, &wait);
3383 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3384 }
3385
3386 /*
3387 * After a short sleep, check if it was a premature sleep. If not, then
3388 * go fully to sleep until explicitly woken up.
3389 */
Mel Gorman55150612012-07-31 16:44:35 -07003390 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003391 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3392
3393 /*
3394 * vmstat counters are not perfectly accurate and the estimated
3395 * value for counters such as NR_FREE_PAGES can deviate from the
3396 * true value by nr_online_cpus * threshold. To avoid the zone
3397 * watermarks being breached while under pressure, we reduce the
3398 * per-cpu vmstat threshold while kswapd is awake and restore
3399 * them before going back to sleep.
3400 */
3401 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003402
Mel Gorman62997022012-10-08 16:32:47 -07003403 /*
3404 * Compaction records what page blocks it recently failed to
3405 * isolate pages from and skips them in the future scanning.
3406 * When kswapd is going to sleep, it is reasonable to assume
3407 * that pages and compaction may succeed so reset the cache.
3408 */
3409 reset_isolation_suitable(pgdat);
3410
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003411 if (!kthread_should_stop())
3412 schedule();
3413
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003414 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3415 } else {
3416 if (remaining)
3417 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3418 else
3419 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3420 }
3421 finish_wait(&pgdat->kswapd_wait, &wait);
3422}
3423
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424/*
3425 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003426 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 *
3428 * This basically trickles out pages so that we have _some_
3429 * free memory available even if there is no other activity
3430 * that frees anything up. This is needed for things like routing
3431 * etc, where we otherwise might have all activity going on in
3432 * asynchronous contexts that cannot page things out.
3433 *
3434 * If there are applications that are active memory-allocators
3435 * (most normal use), this basically shouldn't matter.
3436 */
3437static int kswapd(void *p)
3438{
Mel Gorman215ddd62011-07-08 15:39:40 -07003439 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003440 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003441 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003442 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 pg_data_t *pgdat = (pg_data_t*)p;
3444 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003445
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 struct reclaim_state reclaim_state = {
3447 .reclaimed_slab = 0,
3448 };
Rusty Russella70f7302009-03-13 14:49:46 +10303449 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450
Nick Piggincf40bd12009-01-21 08:12:39 +01003451 lockdep_set_current_reclaim_state(GFP_KERNEL);
3452
Rusty Russell174596a2009-01-01 10:12:29 +10303453 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003454 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455 current->reclaim_state = &reclaim_state;
3456
3457 /*
3458 * Tell the memory management that we're a "memory allocator",
3459 * and that if we need more memory we should get access to it
3460 * regardless (see "__alloc_pages()"). "kswapd" should
3461 * never get caught in the normal page freeing logic.
3462 *
3463 * (Kswapd normally doesn't need memory anyway, but sometimes
3464 * you need a small amount of memory in order to be able to
3465 * page out something else, and this flag essentially protects
3466 * us from recursively trying to free more memory as we're
3467 * trying to free the first piece of memory in the first place).
3468 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003469 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003470 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471
Mel Gorman215ddd62011-07-08 15:39:40 -07003472 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003473 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003474 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003475 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003477 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003478
Mel Gorman215ddd62011-07-08 15:39:40 -07003479 /*
3480 * If the last balance_pgdat was unsuccessful it's unlikely a
3481 * new request of a similar or harder type will succeed soon
3482 * so consider going to sleep on the basis we reclaimed at
3483 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003484 if (balanced_classzone_idx >= new_classzone_idx &&
3485 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003486 new_order = pgdat->kswapd_max_order;
3487 new_classzone_idx = pgdat->classzone_idx;
3488 pgdat->kswapd_max_order = 0;
3489 pgdat->classzone_idx = pgdat->nr_zones - 1;
3490 }
3491
Mel Gorman99504742011-01-13 15:46:20 -08003492 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 /*
3494 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003495 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 */
3497 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003498 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003500 kswapd_try_to_sleep(pgdat, balanced_order,
3501 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003503 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003504 new_order = order;
3505 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003506 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003507 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509
David Rientjes8fe23e02009-12-14 17:58:33 -08003510 ret = try_to_freeze();
3511 if (kthread_should_stop())
3512 break;
3513
3514 /*
3515 * We can speed up thawing tasks if we don't call balance_pgdat
3516 * after returning from the refrigerator
3517 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003518 if (!ret) {
3519 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003520 balanced_classzone_idx = classzone_idx;
3521 balanced_order = balance_pgdat(pgdat, order,
3522 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003523 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003525
Johannes Weiner71abdc12014-06-06 14:35:35 -07003526 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003527 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003528 lockdep_clear_current_reclaim_state();
3529
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 return 0;
3531}
3532
3533/*
3534 * A zone is low on free memory, so wake its kswapd task to service it.
3535 */
Mel Gorman99504742011-01-13 15:46:20 -08003536void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537{
3538 pg_data_t *pgdat;
3539
Con Kolivasf3fe6512006-01-06 00:11:15 -08003540 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 return;
3542
Vladimir Davydov344736f2014-10-20 15:50:30 +04003543 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003545 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003546 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003547 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003548 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3549 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003550 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003552 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003553 return;
3554
3555 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003556 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557}
3558
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003559#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003561 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003562 * freed pages.
3563 *
3564 * Rather than trying to age LRUs the aim is to preserve the overall
3565 * LRU order by reclaiming preferentially
3566 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003568unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003570 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003571 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003572 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003573 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003574 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003575 .may_writepage = 1,
3576 .may_unmap = 1,
3577 .may_swap = 1,
3578 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003580 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003581 struct task_struct *p = current;
3582 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003584 p->flags |= PF_MEMALLOC;
3585 lockdep_set_current_reclaim_state(sc.gfp_mask);
3586 reclaim_state.reclaimed_slab = 0;
3587 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003588
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003589 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003590
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003591 p->reclaim_state = NULL;
3592 lockdep_clear_current_reclaim_state();
3593 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003594
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003595 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003597#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599/* It's optimal to keep kswapds on the same CPUs as their memory, but
3600 not required for correctness. So if the last cpu in a node goes
3601 away, we get changed to run anywhere: as the first one comes back,
3602 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003603static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3604 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003606 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003608 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003609 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003610 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303611 const struct cpumask *mask;
3612
3613 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003614
Rusty Russell3e597942009-01-01 10:12:24 +10303615 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003617 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618 }
3619 }
3620 return NOTIFY_OK;
3621}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Yasunori Goto3218ae12006-06-27 02:53:33 -07003623/*
3624 * This kswapd start function will be called by init and node-hot-add.
3625 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3626 */
3627int kswapd_run(int nid)
3628{
3629 pg_data_t *pgdat = NODE_DATA(nid);
3630 int ret = 0;
3631
3632 if (pgdat->kswapd)
3633 return 0;
3634
3635 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3636 if (IS_ERR(pgdat->kswapd)) {
3637 /* failure at boot is fatal */
3638 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003639 pr_err("Failed to start kswapd on node %d\n", nid);
3640 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003641 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003642 }
3643 return ret;
3644}
3645
David Rientjes8fe23e02009-12-14 17:58:33 -08003646/*
Jiang Liud8adde12012-07-11 14:01:52 -07003647 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003648 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003649 */
3650void kswapd_stop(int nid)
3651{
3652 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3653
Jiang Liud8adde12012-07-11 14:01:52 -07003654 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003655 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003656 NODE_DATA(nid)->kswapd = NULL;
3657 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003658}
3659
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660static int __init kswapd_init(void)
3661{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003662 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003663
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003665 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003666 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 hotcpu_notifier(cpu_callback, 0);
3668 return 0;
3669}
3670
3671module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003672
3673#ifdef CONFIG_NUMA
3674/*
3675 * Zone reclaim mode
3676 *
3677 * If non-zero call zone_reclaim when the number of free pages falls below
3678 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003679 */
3680int zone_reclaim_mode __read_mostly;
3681
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003682#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003683#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003684#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003685#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003686
Christoph Lameter9eeff232006-01-18 17:42:31 -08003687/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003688 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3689 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3690 * a zone.
3691 */
3692#define ZONE_RECLAIM_PRIORITY 4
3693
Christoph Lameter9eeff232006-01-18 17:42:31 -08003694/*
Christoph Lameter96146342006-07-03 00:24:13 -07003695 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3696 * occur.
3697 */
3698int sysctl_min_unmapped_ratio = 1;
3699
3700/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003701 * If the number of slab pages in a zone grows beyond this percentage then
3702 * slab reclaim needs to occur.
3703 */
3704int sysctl_min_slab_ratio = 5;
3705
Mel Gorman90afa5d2009-06-16 15:33:20 -07003706static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3707{
3708 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3709 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3710 zone_page_state(zone, NR_ACTIVE_FILE);
3711
3712 /*
3713 * It's possible for there to be more file mapped pages than
3714 * accounted for by the pages on the file LRU lists because
3715 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3716 */
3717 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3718}
3719
3720/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Alexandru Moised031a152015-11-05 18:48:08 -08003721static unsigned long zone_pagecache_reclaimable(struct zone *zone)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003722{
Alexandru Moised031a152015-11-05 18:48:08 -08003723 unsigned long nr_pagecache_reclaimable;
3724 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003725
3726 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003727 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003728 * potentially reclaimable. Otherwise, we have to worry about
3729 * pages like swapcache and zone_unmapped_file_pages() provides
3730 * a better estimate
3731 */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003732 if (zone_reclaim_mode & RECLAIM_UNMAP)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003733 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3734 else
3735 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3736
3737 /* If we can't clean pages, remove dirty pages from consideration */
3738 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3739 delta += zone_page_state(zone, NR_FILE_DIRTY);
3740
3741 /* Watch for any possible underflows due to delta */
3742 if (unlikely(delta > nr_pagecache_reclaimable))
3743 delta = nr_pagecache_reclaimable;
3744
3745 return nr_pagecache_reclaimable - delta;
3746}
3747
Christoph Lameter0ff38492006-09-25 23:31:52 -07003748/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003749 * Try to free up some pages from this zone through reclaim.
3750 */
Andrew Morton179e9632006-03-22 00:08:18 -08003751static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003752{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003753 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003754 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003755 struct task_struct *p = current;
3756 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003757 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003758 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003759 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003760 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003761 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003762 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003763 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003764 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003765 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003766
Christoph Lameter9eeff232006-01-18 17:42:31 -08003767 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003768 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003769 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003770 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003771 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003772 */
3773 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003774 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003775 reclaim_state.reclaimed_slab = 0;
3776 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003777
Mel Gorman90afa5d2009-06-16 15:33:20 -07003778 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003779 /*
3780 * Free memory by calling shrink zone with increasing
3781 * priorities until we have enough memory freed.
3782 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003783 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003784 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003785 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003786 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003787
Christoph Lameter9eeff232006-01-18 17:42:31 -08003788 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003789 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003790 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003791 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003792}
Andrew Morton179e9632006-03-22 00:08:18 -08003793
3794int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3795{
Andrew Morton179e9632006-03-22 00:08:18 -08003796 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003797 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003798
3799 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003800 * Zone reclaim reclaims unmapped file backed pages and
3801 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003802 *
Christoph Lameter96146342006-07-03 00:24:13 -07003803 * A small portion of unmapped file backed pages is needed for
3804 * file I/O otherwise pages read by file I/O will be immediately
3805 * thrown out if the zone is overallocated. So we do not reclaim
3806 * if less than a specified percentage of the zone is used by
3807 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003808 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003809 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3810 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003811 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003812
Lisa Du6e543d52013-09-11 14:22:36 -07003813 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003814 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003815
Andrew Morton179e9632006-03-22 00:08:18 -08003816 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003817 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003818 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003819 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003820 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003821
3822 /*
3823 * Only run zone reclaim on the local zone or on zones that do not
3824 * have associated processors. This will favor the local processor
3825 * over remote processors and spread off node memory allocations
3826 * as wide as possible.
3827 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003828 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003829 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003830 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003831
Johannes Weiner57054652014-10-09 15:28:17 -07003832 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003833 return ZONE_RECLAIM_NOSCAN;
3834
David Rientjesd773ed62007-10-16 23:26:01 -07003835 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003836 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003837
Mel Gorman24cf725182009-06-16 15:33:23 -07003838 if (!ret)
3839 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3840
David Rientjesd773ed62007-10-16 23:26:01 -07003841 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003842}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003843#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003844
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003845/*
3846 * page_evictable - test whether a page is evictable
3847 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003848 *
3849 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003850 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003851 *
3852 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003853 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003854 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003855 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003856 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003857int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003858{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003859 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003860}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003861
Hugh Dickins850465792012-01-20 14:34:19 -08003862#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003863/**
Hugh Dickins24513262012-01-20 14:34:21 -08003864 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3865 * @pages: array of pages to check
3866 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003867 *
Hugh Dickins24513262012-01-20 14:34:21 -08003868 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003869 *
3870 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003871 */
Hugh Dickins24513262012-01-20 14:34:21 -08003872void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003873{
Johannes Weiner925b7672012-01-12 17:18:15 -08003874 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003875 struct zone *zone = NULL;
3876 int pgscanned = 0;
3877 int pgrescued = 0;
3878 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003879
Hugh Dickins24513262012-01-20 14:34:21 -08003880 for (i = 0; i < nr_pages; i++) {
3881 struct page *page = pages[i];
3882 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003883
Hugh Dickins24513262012-01-20 14:34:21 -08003884 pgscanned++;
3885 pagezone = page_zone(page);
3886 if (pagezone != zone) {
3887 if (zone)
3888 spin_unlock_irq(&zone->lru_lock);
3889 zone = pagezone;
3890 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003891 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003892 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003893
Hugh Dickins24513262012-01-20 14:34:21 -08003894 if (!PageLRU(page) || !PageUnevictable(page))
3895 continue;
3896
Hugh Dickins39b5f292012-10-08 16:33:18 -07003897 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003898 enum lru_list lru = page_lru_base_type(page);
3899
Sasha Levin309381fea2014-01-23 15:52:54 -08003900 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003901 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003902 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3903 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003904 pgrescued++;
3905 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003906 }
Hugh Dickins24513262012-01-20 14:34:21 -08003907
3908 if (zone) {
3909 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3910 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3911 spin_unlock_irq(&zone->lru_lock);
3912 }
Hugh Dickins850465792012-01-20 14:34:19 -08003913}
3914#endif /* CONFIG_SHMEM */