blob: 2d978b28a410b25df1acde351630dee387efbbe5 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049
50#include <asm/tlbflush.h>
51#include <asm/div64.h>
52
53#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070054#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Nick Piggin0f8053a2006-03-22 00:08:33 -080056#include "internal.h"
57
Mel Gorman33906bc2010-08-09 17:19:16 -070058#define CREATE_TRACE_POINTS
59#include <trace/events/vmscan.h>
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080062 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
64
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040066 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Johannes Weineree814fe2014-08-06 16:06:19 -070068 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070069 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080070
Johannes Weineree814fe2014-08-06 16:06:19 -070071 /*
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
73 * are scanned.
74 */
75 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070076
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070077 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080078 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
80 */
81 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080082
Johannes Weineree814fe2014-08-06 16:06:19 -070083 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
86 unsigned int may_writepage:1;
87
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
90
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
93
Johannes Weiner241994ed2015-02-11 15:26:06 -080094 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
96
Johannes Weineree814fe2014-08-06 16:06:19 -070097 unsigned int hibernation_mode:1;
98
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
101
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
104
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107};
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
110
111#ifdef ARCH_HAS_PREFETCH
112#define prefetch_prev_lru_page(_page, _base, _field) \
113 do { \
114 if ((_page)->lru.prev != _base) { \
115 struct page *prev; \
116 \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetch(&prev->_field); \
119 } \
120 } while (0)
121#else
122#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
123#endif
124
125#ifdef ARCH_HAS_PREFETCHW
126#define prefetchw_prev_lru_page(_page, _base, _field) \
127 do { \
128 if ((_page)->lru.prev != _base) { \
129 struct page *prev; \
130 \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetchw(&prev->_field); \
133 } \
134 } while (0)
135#else
136#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137#endif
138
139/*
140 * From 0 .. 100. Higher means more swappy.
141 */
142int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700143/*
144 * The total number of pages which are beyond the high watermark within all
145 * zones.
146 */
147unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
149static LIST_HEAD(shrinker_list);
150static DECLARE_RWSEM(shrinker_rwsem);
151
Andrew Mortonc255a452012-07-31 16:43:02 -0700152#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800153static bool global_reclaim(struct scan_control *sc)
154{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800155 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800156}
Tejun Heo97c93412015-05-22 18:23:36 -0400157
158/**
159 * sane_reclaim - is the usual dirty throttling mechanism operational?
160 * @sc: scan_control in question
161 *
162 * The normal page dirty throttling mechanism in balance_dirty_pages() is
163 * completely broken with the legacy memcg and direct stalling in
164 * shrink_page_list() is used for throttling instead, which lacks all the
165 * niceties such as fairness, adaptive pausing, bandwidth proportional
166 * allocation and configurability.
167 *
168 * This function tests whether the vmscan currently in progress can assume
169 * that the normal dirty throttling mechanism is operational.
170 */
171static bool sane_reclaim(struct scan_control *sc)
172{
173 struct mem_cgroup *memcg = sc->target_mem_cgroup;
174
175 if (!memcg)
176 return true;
177#ifdef CONFIG_CGROUP_WRITEBACK
Michal Hocko33398cf2015-09-08 15:01:02 -0700178 if (memcg->css.cgroup)
Tejun Heo97c93412015-05-22 18:23:36 -0400179 return true;
180#endif
181 return false;
182}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800183#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800184static bool global_reclaim(struct scan_control *sc)
185{
186 return true;
187}
Tejun Heo97c93412015-05-22 18:23:36 -0400188
189static bool sane_reclaim(struct scan_control *sc)
190{
191 return true;
192}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800193#endif
194
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800195static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700196{
197 int nr;
198
199 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
200 zone_page_state(zone, NR_INACTIVE_FILE);
201
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
204 zone_page_state(zone, NR_INACTIVE_ANON);
205
206 return nr;
207}
208
209bool zone_reclaimable(struct zone *zone)
210{
Mel Gorman0d5d8232014-08-06 16:07:16 -0700211 return zone_page_state(zone, NR_PAGES_SCANNED) <
212 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700213}
214
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700215static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800216{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700217 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700218 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800219
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700220 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800221}
222
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000224 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000226int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000228 size_t size = sizeof(*shrinker->nr_deferred);
229
230 /*
231 * If we only have one possible node in the system anyway, save
232 * ourselves the trouble and disable NUMA aware behavior. This way we
233 * will save memory and some small loop time later.
234 */
235 if (nr_node_ids == 1)
236 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
237
238 if (shrinker->flags & SHRINKER_NUMA_AWARE)
239 size *= nr_node_ids;
240
241 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
242 if (!shrinker->nr_deferred)
243 return -ENOMEM;
244
Rusty Russell8e1f9362007-07-17 04:03:17 -0700245 down_write(&shrinker_rwsem);
246 list_add_tail(&shrinker->list, &shrinker_list);
247 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000248 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700250EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
252/*
253 * Remove one
254 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700255void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256{
257 down_write(&shrinker_rwsem);
258 list_del(&shrinker->list);
259 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700260 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700262EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
264#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000265
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800266static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
267 struct shrinker *shrinker,
268 unsigned long nr_scanned,
269 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000271 unsigned long freed = 0;
272 unsigned long long delta;
273 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700274 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000275 long nr;
276 long new_nr;
277 int nid = shrinkctl->nid;
278 long batch_size = shrinker->batch ? shrinker->batch
279 : SHRINK_BATCH;
280
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700281 freeable = shrinker->count_objects(shrinker, shrinkctl);
282 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000283 return 0;
284
285 /*
286 * copy the current shrinker scan count into a local variable
287 * and zero it so that other concurrent shrinker invocations
288 * don't also do this scanning work.
289 */
290 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
291
292 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800293 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700294 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800295 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000296 total_scan += delta;
297 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800298 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000299 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700300 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000301 }
302
303 /*
304 * We need to avoid excessive windup on filesystem shrinkers
305 * due to large numbers of GFP_NOFS allocations causing the
306 * shrinkers to return -1 all the time. This results in a large
307 * nr being built up so when a shrink that can do some work
308 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700309 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000310 * memory.
311 *
312 * Hence only allow the shrinker to scan the entire cache when
313 * a large delta change is calculated directly.
314 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700315 if (delta < freeable / 4)
316 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000317
318 /*
319 * Avoid risking looping forever due to too large nr value:
320 * never try to free more than twice the estimate number of
321 * freeable entries.
322 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700323 if (total_scan > freeable * 2)
324 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000325
326 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800327 nr_scanned, nr_eligible,
328 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000329
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800330 /*
331 * Normally, we should not scan less than batch_size objects in one
332 * pass to avoid too frequent shrinker calls, but if the slab has less
333 * than batch_size objects in total and we are really tight on memory,
334 * we will try to reclaim all available objects, otherwise we can end
335 * up failing allocations although there are plenty of reclaimable
336 * objects spread over several slabs with usage less than the
337 * batch_size.
338 *
339 * We detect the "tight on memory" situations by looking at the total
340 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700341 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800342 * scanning at high prio and therefore should try to reclaim as much as
343 * possible.
344 */
345 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700346 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000347 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800348 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000349
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800350 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000351 ret = shrinker->scan_objects(shrinker, shrinkctl);
352 if (ret == SHRINK_STOP)
353 break;
354 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000355
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800356 count_vm_events(SLABS_SCANNED, nr_to_scan);
357 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000358
359 cond_resched();
360 }
361
362 /*
363 * move the unused scan count back into the shrinker in a
364 * manner that handles concurrent updates. If we exhausted the
365 * scan, there is no need to do an update.
366 */
367 if (total_scan > 0)
368 new_nr = atomic_long_add_return(total_scan,
369 &shrinker->nr_deferred[nid]);
370 else
371 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
372
Dave Hansendf9024a2014-06-04 16:08:07 -0700373 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000374 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800377/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800378 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800379 * @gfp_mask: allocation context
380 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800381 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800382 * @nr_scanned: pressure numerator
383 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800385 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800387 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
388 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800390 * @memcg specifies the memory cgroup to target. If it is not NULL,
391 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
392 * objects from the memory cgroup specified. Otherwise all shrinkers
393 * are called, and memcg aware shrinkers are supposed to scan the
394 * global list then.
395 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800396 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
397 * the available objects should be scanned. Page reclaim for example
398 * passes the number of pages scanned and the number of pages on the
399 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
400 * when it encountered mapped pages. The ratio is further biased by
401 * the ->seeks setting of the shrink function, which indicates the
402 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800404 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800406static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
407 struct mem_cgroup *memcg,
408 unsigned long nr_scanned,
409 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410{
411 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000412 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800414 if (memcg && !memcg_kmem_is_active(memcg))
415 return 0;
416
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800417 if (nr_scanned == 0)
418 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Minchan Kimf06590b2011-05-24 17:11:11 -0700420 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000421 /*
422 * If we would return 0, our callers would understand that we
423 * have nothing else to shrink and give up trying. By returning
424 * 1 we keep it going and assume we'll be able to shrink next
425 * time.
426 */
427 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700428 goto out;
429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
431 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800432 struct shrink_control sc = {
433 .gfp_mask = gfp_mask,
434 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800435 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800436 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800437
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800438 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
439 continue;
440
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800441 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
442 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000443
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800444 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800446
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700448out:
449 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000450 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451}
452
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800453void drop_slab_node(int nid)
454{
455 unsigned long freed;
456
457 do {
458 struct mem_cgroup *memcg = NULL;
459
460 freed = 0;
461 do {
462 freed += shrink_slab(GFP_KERNEL, nid, memcg,
463 1000, 1000);
464 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
465 } while (freed > 10);
466}
467
468void drop_slab(void)
469{
470 int nid;
471
472 for_each_online_node(nid)
473 drop_slab_node(nid);
474}
475
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476static inline int is_page_cache_freeable(struct page *page)
477{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700478 /*
479 * A freeable page cache page is referenced only by the caller
480 * that isolated the page, the page cache radix tree and
481 * optional buffer heads at page->private.
482 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700483 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484}
485
Tejun Heo703c2702015-05-22 17:13:44 -0400486static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
Christoph Lameter930d9152006-01-08 01:00:47 -0800488 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400490 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400492 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 return 1;
494 return 0;
495}
496
497/*
498 * We detected a synchronous write error writing a page out. Probably
499 * -ENOSPC. We need to propagate that into the address_space for a subsequent
500 * fsync(), msync() or close().
501 *
502 * The tricky part is that after writepage we cannot touch the mapping: nothing
503 * prevents it from being freed up. But we have a ref on the page and once
504 * that page is locked, the mapping is pinned.
505 *
506 * We're allowed to run sleeping lock_page() here because we know the caller has
507 * __GFP_FS.
508 */
509static void handle_write_error(struct address_space *mapping,
510 struct page *page, int error)
511{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100512 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700513 if (page_mapping(page) == mapping)
514 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 unlock_page(page);
516}
517
Christoph Lameter04e62a22006-06-23 02:03:38 -0700518/* possible outcome of pageout() */
519typedef enum {
520 /* failed to write page out, page is locked */
521 PAGE_KEEP,
522 /* move page to the active list, page is locked */
523 PAGE_ACTIVATE,
524 /* page has been sent to the disk successfully, page is unlocked */
525 PAGE_SUCCESS,
526 /* page is clean and locked */
527 PAGE_CLEAN,
528} pageout_t;
529
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530/*
Andrew Morton1742f192006-03-22 00:08:21 -0800531 * pageout is called by shrink_page_list() for each dirty page.
532 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700534static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700535 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
537 /*
538 * If the page is dirty, only perform writeback if that write
539 * will be non-blocking. To prevent this allocation from being
540 * stalled by pagecache activity. But note that there may be
541 * stalls if we need to run get_block(). We could test
542 * PagePrivate for that.
543 *
Al Viro81742022014-04-03 03:17:43 -0400544 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 * this page's queue, we can perform writeback even if that
546 * will block.
547 *
548 * If the page is swapcache, write it back even if that would
549 * block, for some throttling. This happens by accident, because
550 * swap_backing_dev_info is bust: it doesn't reflect the
551 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 */
553 if (!is_page_cache_freeable(page))
554 return PAGE_KEEP;
555 if (!mapping) {
556 /*
557 * Some data journaling orphaned pages can have
558 * page->mapping == NULL while being dirty with clean buffers.
559 */
David Howells266cf652009-04-03 16:42:36 +0100560 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 if (try_to_free_buffers(page)) {
562 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700563 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 return PAGE_CLEAN;
565 }
566 }
567 return PAGE_KEEP;
568 }
569 if (mapping->a_ops->writepage == NULL)
570 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400571 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 return PAGE_KEEP;
573
574 if (clear_page_dirty_for_io(page)) {
575 int res;
576 struct writeback_control wbc = {
577 .sync_mode = WB_SYNC_NONE,
578 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700579 .range_start = 0,
580 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 .for_reclaim = 1,
582 };
583
584 SetPageReclaim(page);
585 res = mapping->a_ops->writepage(page, &wbc);
586 if (res < 0)
587 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800588 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 ClearPageReclaim(page);
590 return PAGE_ACTIVATE;
591 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700592
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 if (!PageWriteback(page)) {
594 /* synchronous write or broken a_ops? */
595 ClearPageReclaim(page);
596 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700597 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700598 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 return PAGE_SUCCESS;
600 }
601
602 return PAGE_CLEAN;
603}
604
Andrew Mortona649fd92006-10-17 00:09:36 -0700605/*
Nick Piggine2867812008-07-25 19:45:30 -0700606 * Same as remove_mapping, but if the page is removed from the mapping, it
607 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700608 */
Johannes Weinera5289102014-04-03 14:47:51 -0700609static int __remove_mapping(struct address_space *mapping, struct page *page,
610 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800611{
Greg Thelenc4843a72015-05-22 17:13:16 -0400612 unsigned long flags;
613 struct mem_cgroup *memcg;
614
Nick Piggin28e4d962006-09-25 23:31:23 -0700615 BUG_ON(!PageLocked(page));
616 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800617
Greg Thelenc4843a72015-05-22 17:13:16 -0400618 memcg = mem_cgroup_begin_page_stat(page);
619 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800620 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700621 * The non racy check for a busy page.
622 *
623 * Must be careful with the order of the tests. When someone has
624 * a ref to the page, it may be possible that they dirty it then
625 * drop the reference. So if PageDirty is tested before page_count
626 * here, then the following race may occur:
627 *
628 * get_user_pages(&page);
629 * [user mapping goes away]
630 * write_to(page);
631 * !PageDirty(page) [good]
632 * SetPageDirty(page);
633 * put_page(page);
634 * !page_count(page) [good, discard it]
635 *
636 * [oops, our write_to data is lost]
637 *
638 * Reversing the order of the tests ensures such a situation cannot
639 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
640 * load is not satisfied before that of page->_count.
641 *
642 * Note that if SetPageDirty is always performed via set_page_dirty,
643 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800644 */
Nick Piggine2867812008-07-25 19:45:30 -0700645 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800646 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700647 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
648 if (unlikely(PageDirty(page))) {
649 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800650 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700651 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800652
653 if (PageSwapCache(page)) {
654 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700655 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800656 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400657 spin_unlock_irqrestore(&mapping->tree_lock, flags);
658 mem_cgroup_end_page_stat(memcg);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700659 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700660 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500661 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700662 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500663
664 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700665 /*
666 * Remember a shadow entry for reclaimed file cache in
667 * order to detect refaults, thus thrashing, later on.
668 *
669 * But don't store shadows in an address space that is
670 * already exiting. This is not just an optizimation,
671 * inode reclaim needs to empty out the radix tree or
672 * the nodes are lost. Don't plant shadows behind its
673 * back.
674 */
675 if (reclaimed && page_is_file_cache(page) &&
676 !mapping_exiting(mapping))
677 shadow = workingset_eviction(mapping, page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400678 __delete_from_page_cache(page, shadow, memcg);
679 spin_unlock_irqrestore(&mapping->tree_lock, flags);
680 mem_cgroup_end_page_stat(memcg);
Linus Torvalds6072d132010-12-01 13:35:19 -0500681
682 if (freepage != NULL)
683 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800684 }
685
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800686 return 1;
687
688cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 mem_cgroup_end_page_stat(memcg);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800691 return 0;
692}
693
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694/*
Nick Piggine2867812008-07-25 19:45:30 -0700695 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
696 * someone else has a ref on the page, abort and return 0. If it was
697 * successfully detached, return 1. Assumes the caller has a single ref on
698 * this page.
699 */
700int remove_mapping(struct address_space *mapping, struct page *page)
701{
Johannes Weinera5289102014-04-03 14:47:51 -0700702 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700703 /*
704 * Unfreezing the refcount with 1 rather than 2 effectively
705 * drops the pagecache ref for us without requiring another
706 * atomic operation.
707 */
708 page_unfreeze_refs(page, 1);
709 return 1;
710 }
711 return 0;
712}
713
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700714/**
715 * putback_lru_page - put previously isolated page onto appropriate LRU list
716 * @page: page to be put back to appropriate lru list
717 *
718 * Add previously isolated @page to appropriate LRU list.
719 * Page may still be unevictable for other reasons.
720 *
721 * lru_lock must not be held, interrupts must be enabled.
722 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700723void putback_lru_page(struct page *page)
724{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700725 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700726 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700727
Sasha Levin309381fea2014-01-23 15:52:54 -0800728 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700729
730redo:
731 ClearPageUnevictable(page);
732
Hugh Dickins39b5f292012-10-08 16:33:18 -0700733 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700734 /*
735 * For evictable pages, we can use the cache.
736 * In event of a race, worst case is we end up with an
737 * unevictable page on [in]active list.
738 * We know how to handle that.
739 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700740 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700741 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700742 } else {
743 /*
744 * Put unevictable pages directly on zone's unevictable
745 * list.
746 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700747 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700748 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700749 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700750 * When racing with an mlock or AS_UNEVICTABLE clearing
751 * (page is unlocked) make sure that if the other thread
752 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800753 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700754 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700755 * the page back to the evictable list.
756 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700757 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700758 */
759 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700760 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700761
762 /*
763 * page's status can change while we move it among lru. If an evictable
764 * page is on unevictable list, it never be freed. To avoid that,
765 * check after we added it to the list, again.
766 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700767 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700768 if (!isolate_lru_page(page)) {
769 put_page(page);
770 goto redo;
771 }
772 /* This means someone else dropped this page from LRU
773 * So, it will be freed or putback to LRU again. There is
774 * nothing to do here.
775 */
776 }
777
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700778 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700779 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700780 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700781 count_vm_event(UNEVICTABLE_PGCULLED);
782
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700783 put_page(page); /* drop ref from isolate */
784}
785
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800786enum page_references {
787 PAGEREF_RECLAIM,
788 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800789 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800790 PAGEREF_ACTIVATE,
791};
792
793static enum page_references page_check_references(struct page *page,
794 struct scan_control *sc)
795{
Johannes Weiner645747462010-03-05 13:42:22 -0800796 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800797 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800798
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700799 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
800 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800801 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800802
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800803 /*
804 * Mlock lost the isolation race with us. Let try_to_unmap()
805 * move the page to the unevictable list.
806 */
807 if (vm_flags & VM_LOCKED)
808 return PAGEREF_RECLAIM;
809
Johannes Weiner645747462010-03-05 13:42:22 -0800810 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700811 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800812 return PAGEREF_ACTIVATE;
813 /*
814 * All mapped pages start out with page table
815 * references from the instantiating fault, so we need
816 * to look twice if a mapped file page is used more
817 * than once.
818 *
819 * Mark it and spare it for another trip around the
820 * inactive list. Another page table reference will
821 * lead to its activation.
822 *
823 * Note: the mark is set for activated pages as well
824 * so that recently deactivated but used pages are
825 * quickly recovered.
826 */
827 SetPageReferenced(page);
828
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800829 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800830 return PAGEREF_ACTIVATE;
831
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800832 /*
833 * Activate file-backed executable pages after first usage.
834 */
835 if (vm_flags & VM_EXEC)
836 return PAGEREF_ACTIVATE;
837
Johannes Weiner645747462010-03-05 13:42:22 -0800838 return PAGEREF_KEEP;
839 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800840
841 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700842 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800843 return PAGEREF_RECLAIM_CLEAN;
844
845 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800846}
847
Mel Gormane2be15f2013-07-03 15:01:57 -0700848/* Check if a page is dirty or under writeback */
849static void page_check_dirty_writeback(struct page *page,
850 bool *dirty, bool *writeback)
851{
Mel Gormanb4597222013-07-03 15:02:05 -0700852 struct address_space *mapping;
853
Mel Gormane2be15f2013-07-03 15:01:57 -0700854 /*
855 * Anonymous pages are not handled by flushers and must be written
856 * from reclaim context. Do not stall reclaim based on them
857 */
858 if (!page_is_file_cache(page)) {
859 *dirty = false;
860 *writeback = false;
861 return;
862 }
863
864 /* By default assume that the page flags are accurate */
865 *dirty = PageDirty(page);
866 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700867
868 /* Verify dirty/writeback state if the filesystem supports it */
869 if (!page_has_private(page))
870 return;
871
872 mapping = page_mapping(page);
873 if (mapping && mapping->a_ops->is_dirty_writeback)
874 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700875}
876
Nick Piggine2867812008-07-25 19:45:30 -0700877/*
Andrew Morton1742f192006-03-22 00:08:21 -0800878 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 */
Andrew Morton1742f192006-03-22 00:08:21 -0800880static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700881 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700882 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700883 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700884 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700885 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700886 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700887 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700888 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700889 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890{
891 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700892 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700894 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700895 unsigned long nr_dirty = 0;
896 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800897 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700898 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700899 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
901 cond_resched();
902
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 while (!list_empty(page_list)) {
904 struct address_space *mapping;
905 struct page *page;
906 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700907 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700908 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909
910 cond_resched();
911
912 page = lru_to_page(page_list);
913 list_del(&page->lru);
914
Nick Piggin529ae9a2008-08-02 12:01:03 +0200915 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 goto keep;
917
Sasha Levin309381fea2014-01-23 15:52:54 -0800918 VM_BUG_ON_PAGE(PageActive(page), page);
919 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
921 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800922
Hugh Dickins39b5f292012-10-08 16:33:18 -0700923 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700924 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700925
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700926 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800927 goto keep_locked;
928
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 /* Double the slab pressure for mapped and swapcache pages */
930 if (page_mapped(page) || PageSwapCache(page))
931 sc->nr_scanned++;
932
Andy Whitcroftc661b072007-08-22 14:01:26 -0700933 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
934 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
935
Mel Gorman283aba92013-07-03 15:01:51 -0700936 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700937 * The number of dirty pages determines if a zone is marked
938 * reclaim_congested which affects wait_iff_congested. kswapd
939 * will stall and start writing pages if the tail of the LRU
940 * is all dirty unqueued pages.
941 */
942 page_check_dirty_writeback(page, &dirty, &writeback);
943 if (dirty || writeback)
944 nr_dirty++;
945
946 if (dirty && !writeback)
947 nr_unqueued_dirty++;
948
Mel Gormand04e8ac2013-07-03 15:02:03 -0700949 /*
950 * Treat this page as congested if the underlying BDI is or if
951 * pages are cycling through the LRU so quickly that the
952 * pages marked for immediate reclaim are making it to the
953 * end of the LRU a second time.
954 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700955 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800956 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400957 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700958 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700959 nr_congested++;
960
961 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700962 * If a page at the tail of the LRU is under writeback, there
963 * are three cases to consider.
964 *
965 * 1) If reclaim is encountering an excessive number of pages
966 * under writeback and this page is both under writeback and
967 * PageReclaim then it indicates that pages are being queued
968 * for IO but are being recycled through the LRU before the
969 * IO can complete. Waiting on the page itself risks an
970 * indefinite stall if it is impossible to writeback the
971 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700972 * note that the LRU is being scanned too quickly and the
973 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700974 *
Tejun Heo97c93412015-05-22 18:23:36 -0400975 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -0700976 * not marked for immediate reclaim, or the caller does not
977 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
978 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -0400979 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -0700980 *
Michal Hockoecf5fc62015-08-04 14:36:58 -0700981 * Require may_enter_fs because we would wait on fs, which
982 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -0700983 * enter reclaim, and deadlock if it waits on a page for
984 * which it is needed to do the write (loop masks off
985 * __GFP_IO|__GFP_FS for this reason); but more thought
986 * would probably show more reasons.
987 *
Hugh Dickins7fadc822015-09-08 15:03:46 -0700988 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -0700989 * PageReclaim. memcg does not have any dirty pages
990 * throttling so we could easily OOM just because too many
991 * pages are in writeback and there is nothing else to
992 * reclaim. Wait for the writeback to complete.
993 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700994 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700995 /* Case 1 above */
996 if (current_is_kswapd() &&
997 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700998 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700999 nr_immediate++;
1000 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001001
1002 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001003 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001004 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001005 /*
1006 * This is slightly racy - end_page_writeback()
1007 * might have just cleared PageReclaim, then
1008 * setting PageReclaim here end up interpreted
1009 * as PageReadahead - but that does not matter
1010 * enough to care. What we do want is for this
1011 * page to have PageReclaim set next time memcg
1012 * reclaim reaches the tests above, so it will
1013 * then wait_on_page_writeback() to avoid OOM;
1014 * and it's also appropriate in global reclaim.
1015 */
1016 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001017 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001018 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001019
1020 /* Case 3 above */
1021 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001022 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001023 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001024 /* then go back and try same page again */
1025 list_add_tail(&page->lru, page_list);
1026 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001027 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001028 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029
Minchan Kim02c6de82012-10-08 16:31:55 -07001030 if (!force_reclaim)
1031 references = page_check_references(page, sc);
1032
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001033 switch (references) {
1034 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001036 case PAGEREF_KEEP:
1037 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001038 case PAGEREF_RECLAIM:
1039 case PAGEREF_RECLAIM_CLEAN:
1040 ; /* try to reclaim the page below */
1041 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 /*
1044 * Anonymous process memory has backing store?
1045 * Try to allocate it some swap space here.
1046 */
Nick Pigginb291f002008-10-18 20:26:44 -07001047 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001048 if (!(sc->gfp_mask & __GFP_IO))
1049 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001050 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001052 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Mel Gormane2be15f2013-07-03 15:01:57 -07001054 /* Adding to swap updated mapping */
1055 mapping = page_mapping(page);
1056 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
1058 /*
1059 * The page is mapped into the page tables of one or more
1060 * processes. Try to unmap it here.
1061 */
1062 if (page_mapped(page) && mapping) {
Mel Gorman72b252a2015-09-04 15:47:32 -07001063 switch (try_to_unmap(page,
1064 ttu_flags|TTU_BATCH_FLUSH)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 case SWAP_FAIL:
1066 goto activate_locked;
1067 case SWAP_AGAIN:
1068 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001069 case SWAP_MLOCK:
1070 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 case SWAP_SUCCESS:
1072 ; /* try to free the page below */
1073 }
1074 }
1075
1076 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001077 /*
1078 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001079 * avoid risk of stack overflow but only writeback
1080 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001081 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001082 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001083 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001084 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001085 /*
1086 * Immediately reclaim when written back.
1087 * Similar in principal to deactivate_page()
1088 * except we already have the page isolated
1089 * and know it's dirty
1090 */
1091 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1092 SetPageReclaim(page);
1093
Mel Gormanee728862011-10-31 17:07:38 -07001094 goto keep_locked;
1095 }
1096
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001097 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001099 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001101 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 goto keep_locked;
1103
Mel Gormand950c942015-09-04 15:47:35 -07001104 /*
1105 * Page is dirty. Flush the TLB if a writable entry
1106 * potentially exists to avoid CPU writes after IO
1107 * starts and then write it out here.
1108 */
1109 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001110 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 case PAGE_KEEP:
1112 goto keep_locked;
1113 case PAGE_ACTIVATE:
1114 goto activate_locked;
1115 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001116 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001117 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001118 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 /*
1122 * A synchronous write - probably a ramdisk. Go
1123 * ahead and try to reclaim the page.
1124 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001125 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 goto keep;
1127 if (PageDirty(page) || PageWriteback(page))
1128 goto keep_locked;
1129 mapping = page_mapping(page);
1130 case PAGE_CLEAN:
1131 ; /* try to free the page below */
1132 }
1133 }
1134
1135 /*
1136 * If the page has buffers, try to free the buffer mappings
1137 * associated with this page. If we succeed we try to free
1138 * the page as well.
1139 *
1140 * We do this even if the page is PageDirty().
1141 * try_to_release_page() does not perform I/O, but it is
1142 * possible for a page to have PageDirty set, but it is actually
1143 * clean (all its buffers are clean). This happens if the
1144 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001145 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 * try_to_release_page() will discover that cleanness and will
1147 * drop the buffers and mark the page clean - it can be freed.
1148 *
1149 * Rarely, pages can have buffers and no ->mapping. These are
1150 * the pages which were not successfully invalidated in
1151 * truncate_complete_page(). We try to drop those buffers here
1152 * and if that worked, and the page is no longer mapped into
1153 * process address space (page_count == 1) it can be freed.
1154 * Otherwise, leave the page on the LRU so it is swappable.
1155 */
David Howells266cf652009-04-03 16:42:36 +01001156 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 if (!try_to_release_page(page, sc->gfp_mask))
1158 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001159 if (!mapping && page_count(page) == 1) {
1160 unlock_page(page);
1161 if (put_page_testzero(page))
1162 goto free_it;
1163 else {
1164 /*
1165 * rare race with speculative reference.
1166 * the speculative reference will free
1167 * this page shortly, so we may
1168 * increment nr_reclaimed here (and
1169 * leave it off the LRU).
1170 */
1171 nr_reclaimed++;
1172 continue;
1173 }
1174 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 }
1176
Johannes Weinera5289102014-04-03 14:47:51 -07001177 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001178 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
Nick Piggina978d6f2008-10-18 20:26:58 -07001180 /*
1181 * At this point, we have no other references and there is
1182 * no way to pick any more up (removed from LRU, removed
1183 * from pagecache). Can use non-atomic bitops now (and
1184 * we obviously don't have to worry about waking up a process
1185 * waiting on the page lock, because there are no references.
1186 */
1187 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001188free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001189 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001190
1191 /*
1192 * Is there need to periodically free_page_list? It would
1193 * appear not as the counts should be low
1194 */
1195 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 continue;
1197
Nick Pigginb291f002008-10-18 20:26:44 -07001198cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001199 if (PageSwapCache(page))
1200 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001201 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001202 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001203 continue;
1204
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001206 /* Not a candidate for swapping, so reclaim swap space. */
1207 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001208 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001209 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 SetPageActive(page);
1211 pgactivate++;
1212keep_locked:
1213 unlock_page(page);
1214keep:
1215 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001216 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001218
Johannes Weiner747db952014-08-08 14:19:24 -07001219 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001220 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001221 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001222
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001224 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001225
Mel Gorman8e950282013-07-03 15:02:02 -07001226 *ret_nr_dirty += nr_dirty;
1227 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001228 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001229 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001230 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001231 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232}
1233
Minchan Kim02c6de82012-10-08 16:31:55 -07001234unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1235 struct list_head *page_list)
1236{
1237 struct scan_control sc = {
1238 .gfp_mask = GFP_KERNEL,
1239 .priority = DEF_PRIORITY,
1240 .may_unmap = 1,
1241 };
Mel Gorman8e950282013-07-03 15:02:02 -07001242 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001243 struct page *page, *next;
1244 LIST_HEAD(clean_pages);
1245
1246 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001247 if (page_is_file_cache(page) && !PageDirty(page) &&
1248 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001249 ClearPageActive(page);
1250 list_move(&page->lru, &clean_pages);
1251 }
1252 }
1253
1254 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001255 TTU_UNMAP|TTU_IGNORE_ACCESS,
1256 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001257 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001258 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001259 return ret;
1260}
1261
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001262/*
1263 * Attempt to remove the specified page from its LRU. Only take this page
1264 * if it is of the appropriate PageActive status. Pages which are being
1265 * freed elsewhere are also ignored.
1266 *
1267 * page: page to consider
1268 * mode: one of the LRU isolation modes defined above
1269 *
1270 * returns 0 on success, -ve errno on failure.
1271 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001272int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001273{
1274 int ret = -EINVAL;
1275
1276 /* Only take pages on the LRU. */
1277 if (!PageLRU(page))
1278 return ret;
1279
Minchan Kime46a2872012-10-08 16:33:48 -07001280 /* Compaction should not handle unevictable pages but CMA can do so */
1281 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001282 return ret;
1283
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001284 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001285
Mel Gormanc8244932012-01-12 17:19:38 -08001286 /*
1287 * To minimise LRU disruption, the caller can indicate that it only
1288 * wants to isolate pages it will be able to operate on without
1289 * blocking - clean pages for the most part.
1290 *
1291 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1292 * is used by reclaim when it is cannot write to backing storage
1293 *
1294 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1295 * that it is possible to migrate without blocking
1296 */
1297 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1298 /* All the caller can do on PageWriteback is block */
1299 if (PageWriteback(page))
1300 return ret;
1301
1302 if (PageDirty(page)) {
1303 struct address_space *mapping;
1304
1305 /* ISOLATE_CLEAN means only clean pages */
1306 if (mode & ISOLATE_CLEAN)
1307 return ret;
1308
1309 /*
1310 * Only pages without mappings or that have a
1311 * ->migratepage callback are possible to migrate
1312 * without blocking
1313 */
1314 mapping = page_mapping(page);
1315 if (mapping && !mapping->a_ops->migratepage)
1316 return ret;
1317 }
1318 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001319
Minchan Kimf80c0672011-10-31 17:06:55 -07001320 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1321 return ret;
1322
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001323 if (likely(get_page_unless_zero(page))) {
1324 /*
1325 * Be careful not to clear PageLRU until after we're
1326 * sure the page is not being freed elsewhere -- the
1327 * page release code relies on it.
1328 */
1329 ClearPageLRU(page);
1330 ret = 0;
1331 }
1332
1333 return ret;
1334}
1335
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001336/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 * zone->lru_lock is heavily contended. Some of the functions that
1338 * shrink the lists perform better by taking out a batch of pages
1339 * and working on them outside the LRU lock.
1340 *
1341 * For pagecache intensive workloads, this function is the hottest
1342 * spot in the kernel (apart from copy_*_user functions).
1343 *
1344 * Appropriate locks must be held before calling this function.
1345 *
1346 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001347 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001349 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001350 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001351 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001352 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 *
1354 * returns how many pages were moved onto *@dst.
1355 */
Andrew Morton69e05942006-03-22 00:08:19 -08001356static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001357 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001358 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001359 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001361 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001362 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001363 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001365 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1366 !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001367 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001368 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 page = lru_to_page(src);
1371 prefetchw_prev_lru_page(page, src, flags);
1372
Sasha Levin309381fea2014-01-23 15:52:54 -08001373 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001374
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001375 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001376 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001377 nr_pages = hpage_nr_pages(page);
1378 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001379 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001380 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001381 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001382
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001383 case -EBUSY:
1384 /* else it is being freed elsewhere */
1385 list_move(&page->lru, src);
1386 continue;
1387
1388 default:
1389 BUG();
1390 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 }
1392
Hugh Dickinsf6260122012-01-12 17:20:06 -08001393 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001394 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1395 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 return nr_taken;
1397}
1398
Nick Piggin62695a82008-10-18 20:26:09 -07001399/**
1400 * isolate_lru_page - tries to isolate a page from its LRU list
1401 * @page: page to isolate from its LRU list
1402 *
1403 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1404 * vmstat statistic corresponding to whatever LRU list the page was on.
1405 *
1406 * Returns 0 if the page was removed from an LRU list.
1407 * Returns -EBUSY if the page was not on an LRU list.
1408 *
1409 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001410 * the active list, it will have PageActive set. If it was found on
1411 * the unevictable list, it will have the PageUnevictable bit set. That flag
1412 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001413 *
1414 * The vmstat statistic corresponding to the list on which the page was
1415 * found will be decremented.
1416 *
1417 * Restrictions:
1418 * (1) Must be called with an elevated refcount on the page. This is a
1419 * fundamentnal difference from isolate_lru_pages (which is called
1420 * without a stable reference).
1421 * (2) the lru_lock must not be held.
1422 * (3) interrupts must be enabled.
1423 */
1424int isolate_lru_page(struct page *page)
1425{
1426 int ret = -EBUSY;
1427
Sasha Levin309381fea2014-01-23 15:52:54 -08001428 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001429
Nick Piggin62695a82008-10-18 20:26:09 -07001430 if (PageLRU(page)) {
1431 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001432 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001433
1434 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001435 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001436 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001437 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001438 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001439 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001440 del_page_from_lru_list(page, lruvec, lru);
1441 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001442 }
1443 spin_unlock_irq(&zone->lru_lock);
1444 }
1445 return ret;
1446}
1447
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001448/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001449 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1450 * then get resheduled. When there are massive number of tasks doing page
1451 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1452 * the LRU list will go small and be scanned faster than necessary, leading to
1453 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001454 */
1455static int too_many_isolated(struct zone *zone, int file,
1456 struct scan_control *sc)
1457{
1458 unsigned long inactive, isolated;
1459
1460 if (current_is_kswapd())
1461 return 0;
1462
Tejun Heo97c93412015-05-22 18:23:36 -04001463 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001464 return 0;
1465
1466 if (file) {
1467 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1468 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1469 } else {
1470 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1471 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1472 }
1473
Fengguang Wu3cf23842012-12-18 14:23:31 -08001474 /*
1475 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1476 * won't get blocked by normal direct-reclaimers, forming a circular
1477 * deadlock.
1478 */
1479 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1480 inactive >>= 3;
1481
Rik van Riel35cd7812009-09-21 17:01:38 -07001482 return isolated > inactive;
1483}
1484
Mel Gorman66635622010-08-09 17:19:30 -07001485static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001486putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001487{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001488 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1489 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001490 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001491
Mel Gorman66635622010-08-09 17:19:30 -07001492 /*
1493 * Put back any unfreeable pages.
1494 */
Mel Gorman66635622010-08-09 17:19:30 -07001495 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001496 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001497 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001498
Sasha Levin309381fea2014-01-23 15:52:54 -08001499 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001500 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001501 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001502 spin_unlock_irq(&zone->lru_lock);
1503 putback_lru_page(page);
1504 spin_lock_irq(&zone->lru_lock);
1505 continue;
1506 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001507
1508 lruvec = mem_cgroup_page_lruvec(page, zone);
1509
Linus Torvalds7a608572011-01-17 14:42:19 -08001510 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001511 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001512 add_page_to_lru_list(page, lruvec, lru);
1513
Mel Gorman66635622010-08-09 17:19:30 -07001514 if (is_active_lru(lru)) {
1515 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001516 int numpages = hpage_nr_pages(page);
1517 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001518 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001519 if (put_page_testzero(page)) {
1520 __ClearPageLRU(page);
1521 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001522 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001523
1524 if (unlikely(PageCompound(page))) {
1525 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001526 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001527 (*get_compound_page_dtor(page))(page);
1528 spin_lock_irq(&zone->lru_lock);
1529 } else
1530 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001531 }
1532 }
Mel Gorman66635622010-08-09 17:19:30 -07001533
Hugh Dickins3f797682012-01-12 17:20:07 -08001534 /*
1535 * To save our caller's stack, now use input list for pages to free.
1536 */
1537 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001538}
1539
1540/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001541 * If a kernel thread (such as nfsd for loop-back mounts) services
1542 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1543 * In that case we should only throttle if the backing device it is
1544 * writing to is congested. In other cases it is safe to throttle.
1545 */
1546static int current_may_throttle(void)
1547{
1548 return !(current->flags & PF_LESS_THROTTLE) ||
1549 current->backing_dev_info == NULL ||
1550 bdi_write_congested(current->backing_dev_info);
1551}
1552
1553/*
Andrew Morton1742f192006-03-22 00:08:21 -08001554 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1555 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 */
Mel Gorman66635622010-08-09 17:19:30 -07001557static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001558shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001559 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560{
1561 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001562 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001563 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001564 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001565 unsigned long nr_dirty = 0;
1566 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001567 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001568 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001569 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001570 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001571 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001572 struct zone *zone = lruvec_zone(lruvec);
1573 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001574
Rik van Riel35cd7812009-09-21 17:01:38 -07001575 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001576 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001577
1578 /* We are about to die and free our memory. Return now. */
1579 if (fatal_signal_pending(current))
1580 return SWAP_CLUSTER_MAX;
1581 }
1582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001584
1585 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001586 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001587 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001588 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001589
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001592 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1593 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001594
1595 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1596 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1597
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001598 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001599 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001600 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001601 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001602 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001603 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001604 }
Hillf Dantond563c052012-03-21 16:34:02 -07001605 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001606
Hillf Dantond563c052012-03-21 16:34:02 -07001607 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001608 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001609
Minchan Kim02c6de82012-10-08 16:31:55 -07001610 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001611 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1612 &nr_writeback, &nr_immediate,
1613 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001614
Hugh Dickins3f797682012-01-12 17:20:07 -08001615 spin_lock_irq(&zone->lru_lock);
1616
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001617 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001618
Ying Han904249a2012-04-25 16:01:48 -07001619 if (global_reclaim(sc)) {
1620 if (current_is_kswapd())
1621 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1622 nr_reclaimed);
1623 else
1624 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1625 nr_reclaimed);
1626 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001627
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001628 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001629
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001630 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001631
1632 spin_unlock_irq(&zone->lru_lock);
1633
Johannes Weiner747db952014-08-08 14:19:24 -07001634 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001635 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001636
Mel Gorman92df3a72011-10-31 17:07:56 -07001637 /*
1638 * If reclaim is isolating dirty pages under writeback, it implies
1639 * that the long-lived page allocation rate is exceeding the page
1640 * laundering rate. Either the global limits are not being effective
1641 * at throttling processes due to the page distribution throughout
1642 * zones or there is heavy usage of a slow backing device. The
1643 * only option is to throttle from reclaim context which is not ideal
1644 * as there is no guarantee the dirtying process is throttled in the
1645 * same way balance_dirty_pages() manages.
1646 *
Mel Gorman8e950282013-07-03 15:02:02 -07001647 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1648 * of pages under pages flagged for immediate reclaim and stall if any
1649 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001650 */
Mel Gorman918fc712013-07-08 16:00:25 -07001651 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001652 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001653
Mel Gormand43006d2013-07-03 15:01:50 -07001654 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001655 * Legacy memcg will stall in page writeback so avoid forcibly
1656 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001657 */
Tejun Heo97c93412015-05-22 18:23:36 -04001658 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001659 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001660 * Tag a zone as congested if all the dirty pages scanned were
1661 * backed by a congested BDI and wait_iff_congested will stall.
1662 */
1663 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001664 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001665
1666 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001667 * If dirty pages are scanned that are not queued for IO, it
1668 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001669 * the zone ZONE_DIRTY and kswapd will start writing pages from
1670 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001671 */
1672 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001673 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001674
1675 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001676 * If kswapd scans pages marked marked for immediate
1677 * reclaim and under writeback (nr_immediate), it implies
1678 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001679 * they are written so also forcibly stall.
1680 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001681 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001682 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001683 }
Mel Gormand43006d2013-07-03 15:01:50 -07001684
Mel Gorman8e950282013-07-03 15:02:02 -07001685 /*
1686 * Stall direct reclaim for IO completions if underlying BDIs or zone
1687 * is congested. Allow kswapd to continue until it starts encountering
1688 * unqueued dirty pages or cycling through the LRU too quickly.
1689 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001690 if (!sc->hibernation_mode && !current_is_kswapd() &&
1691 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001692 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1693
Mel Gormane11da5b2010-10-26 14:21:40 -07001694 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1695 zone_idx(zone),
1696 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001697 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001698 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001699 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700}
1701
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001702/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 * This moves pages from the active list to the inactive list.
1704 *
1705 * We move them the other way if the page is referenced by one or more
1706 * processes, from rmap.
1707 *
1708 * If the pages are mostly unmapped, the processing is fast and it is
1709 * appropriate to hold zone->lru_lock across the whole operation. But if
1710 * the pages are mapped, the processing is slow (page_referenced()) so we
1711 * should drop zone->lru_lock around each page. It's impossible to balance
1712 * this, so instead we remove the pages from the LRU while processing them.
1713 * It is safe to rely on PG_active against the non-LRU pages in here because
1714 * nobody will play with that bit on a non-LRU page.
1715 *
1716 * The downside is that we have to touch page->_count against each page.
1717 * But we had to alter page->flags anyway.
1718 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001719
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001720static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001721 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001722 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001723 enum lru_list lru)
1724{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001725 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001726 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001727 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001728 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001729
Wu Fengguang3eb41402009-06-16 15:33:13 -07001730 while (!list_empty(list)) {
1731 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001732 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001733
Sasha Levin309381fea2014-01-23 15:52:54 -08001734 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001735 SetPageLRU(page);
1736
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001737 nr_pages = hpage_nr_pages(page);
1738 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001739 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001740 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001741
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001742 if (put_page_testzero(page)) {
1743 __ClearPageLRU(page);
1744 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001745 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001746
1747 if (unlikely(PageCompound(page))) {
1748 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001749 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001750 (*get_compound_page_dtor(page))(page);
1751 spin_lock_irq(&zone->lru_lock);
1752 } else
1753 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001754 }
1755 }
1756 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1757 if (!is_active_lru(lru))
1758 __count_vm_events(PGDEACTIVATE, pgmoved);
1759}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001760
Hugh Dickinsf6260122012-01-12 17:20:06 -08001761static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001762 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001763 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001764 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001766 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001767 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001768 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001770 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001771 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001773 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001774 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001775 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001776 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001777 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
1779 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001780
1781 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001782 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001783 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001784 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001785
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001787
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001788 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1789 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001790 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001791 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001792
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001793 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001794
Hugh Dickinsf6260122012-01-12 17:20:06 -08001795 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001796 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001797 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 spin_unlock_irq(&zone->lru_lock);
1799
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 while (!list_empty(&l_hold)) {
1801 cond_resched();
1802 page = lru_to_page(&l_hold);
1803 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001804
Hugh Dickins39b5f292012-10-08 16:33:18 -07001805 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001806 putback_lru_page(page);
1807 continue;
1808 }
1809
Mel Gormancc715d92012-03-21 16:34:00 -07001810 if (unlikely(buffer_heads_over_limit)) {
1811 if (page_has_private(page) && trylock_page(page)) {
1812 if (page_has_private(page))
1813 try_to_release_page(page, 0);
1814 unlock_page(page);
1815 }
1816 }
1817
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001818 if (page_referenced(page, 0, sc->target_mem_cgroup,
1819 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001820 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001821 /*
1822 * Identify referenced, file-backed active pages and
1823 * give them one more trip around the active list. So
1824 * that executable code get better chances to stay in
1825 * memory under moderate memory pressure. Anon pages
1826 * are not likely to be evicted by use-once streaming
1827 * IO, plus JVM can create lots of anon VM_EXEC pages,
1828 * so we ignore them here.
1829 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001830 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001831 list_add(&page->lru, &l_active);
1832 continue;
1833 }
1834 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001835
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001836 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 list_add(&page->lru, &l_inactive);
1838 }
1839
Andrew Mortonb5557492009-01-06 14:40:13 -08001840 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001841 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001842 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001843 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001844 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001845 * Count referenced pages from currently used mappings as rotated,
1846 * even though only some of them are actually re-activated. This
1847 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001848 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001849 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001850 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001851
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001852 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1853 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001854 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001855 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001856
Johannes Weiner747db952014-08-08 14:19:24 -07001857 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001858 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859}
1860
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001861#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001862static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001863{
1864 unsigned long active, inactive;
1865
1866 active = zone_page_state(zone, NR_ACTIVE_ANON);
1867 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1868
1869 if (inactive * zone->inactive_ratio < active)
1870 return 1;
1871
1872 return 0;
1873}
1874
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001875/**
1876 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001877 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001878 *
1879 * Returns true if the zone does not have enough inactive anon pages,
1880 * meaning some active anon pages need to be deactivated.
1881 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001882static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001883{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001884 /*
1885 * If we don't have swap space, anonymous page deactivation
1886 * is pointless.
1887 */
1888 if (!total_swap_pages)
1889 return 0;
1890
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001891 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001892 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001893
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001894 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001895}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001896#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001897static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001898{
1899 return 0;
1900}
1901#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001902
Rik van Riel56e49d22009-06-16 15:32:28 -07001903/**
1904 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001905 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001906 *
1907 * When the system is doing streaming IO, memory pressure here
1908 * ensures that active file pages get deactivated, until more
1909 * than half of the file pages are on the inactive list.
1910 *
1911 * Once we get to that situation, protect the system's working
1912 * set from being evicted by disabling active file page aging.
1913 *
1914 * This uses a different ratio than the anonymous pages, because
1915 * the page cache uses a use-once replacement algorithm.
1916 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001917static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001918{
Johannes Weinere3790142013-02-22 16:35:19 -08001919 unsigned long inactive;
1920 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001921
Johannes Weinere3790142013-02-22 16:35:19 -08001922 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1923 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1924
1925 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001926}
1927
Hugh Dickins75b00af2012-05-29 15:07:09 -07001928static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001929{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001930 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001931 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001932 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001933 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001934}
1935
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001936static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001937 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001938{
Rik van Rielb39415b2009-12-14 17:59:48 -08001939 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001940 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001941 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001942 return 0;
1943 }
1944
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001945 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001946}
1947
Johannes Weiner9a265112013-02-22 16:32:17 -08001948enum scan_balance {
1949 SCAN_EQUAL,
1950 SCAN_FRACT,
1951 SCAN_ANON,
1952 SCAN_FILE,
1953};
1954
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001956 * Determine how aggressively the anon and file LRU lists should be
1957 * scanned. The relative value of each set of LRU lists is determined
1958 * by looking at the fraction of the pages scanned we did rotate back
1959 * onto the active list instead of evict.
1960 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001961 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1962 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001963 */
Johannes Weiner02695172014-08-06 16:06:17 -07001964static void get_scan_count(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001965 struct scan_control *sc, unsigned long *nr,
1966 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001967{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001968 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001969 u64 fraction[2];
1970 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001971 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001972 unsigned long anon_prio, file_prio;
1973 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001974 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001975 bool force_scan = false;
1976 unsigned long ap, fp;
1977 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001978 bool some_scanned;
1979 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001980
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001981 /*
1982 * If the zone or memcg is small, nr[l] can be 0. This
1983 * results in no scanning on this priority and a potential
1984 * priority drop. Global direct reclaim can go to the next
1985 * zone and tends to have no problems. Global kswapd is for
1986 * zone balancing and it needs to scan a minimum amount. When
1987 * reclaiming for a memcg, a priority drop can cause high
1988 * latencies, so it's better to scan a minimum amount there as
1989 * well.
1990 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08001991 if (current_is_kswapd()) {
1992 if (!zone_reclaimable(zone))
1993 force_scan = true;
1994 if (!mem_cgroup_lruvec_online(lruvec))
1995 force_scan = true;
1996 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001997 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001998 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001999
2000 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08002001 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002002 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002003 goto out;
2004 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002005
Johannes Weiner10316b32013-02-22 16:32:14 -08002006 /*
2007 * Global reclaim will swap to prevent OOM even with no
2008 * swappiness, but memcg users want to use this knob to
2009 * disable swapping for individual groups completely when
2010 * using the memory controller's swap limit feature would be
2011 * too expensive.
2012 */
Johannes Weiner02695172014-08-06 16:06:17 -07002013 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002014 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002015 goto out;
2016 }
2017
2018 /*
2019 * Do not apply any pressure balancing cleverness when the
2020 * system is close to OOM, scan both anon and file equally
2021 * (unless the swappiness setting disagrees with swapping).
2022 */
Johannes Weiner02695172014-08-06 16:06:17 -07002023 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002024 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002025 goto out;
2026 }
2027
Johannes Weiner11d16c22013-02-22 16:32:15 -08002028 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002029 * Prevent the reclaimer from falling into the cache trap: as
2030 * cache pages start out inactive, every cache fault will tip
2031 * the scan balance towards the file LRU. And as the file LRU
2032 * shrinks, so does the window for rotation from references.
2033 * This means we have a runaway feedback loop where a tiny
2034 * thrashing file LRU becomes infinitely more attractive than
2035 * anon pages. Try to detect this based on file LRU size.
2036 */
2037 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002038 unsigned long zonefile;
2039 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07002040
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002041 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2042 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2043 zone_page_state(zone, NR_INACTIVE_FILE);
2044
2045 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07002046 scan_balance = SCAN_ANON;
2047 goto out;
2048 }
2049 }
2050
2051 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002052 * There is enough inactive page cache, do not reclaim
2053 * anything from the anonymous working set right now.
2054 */
2055 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002056 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002057 goto out;
2058 }
2059
Johannes Weiner9a265112013-02-22 16:32:17 -08002060 scan_balance = SCAN_FRACT;
2061
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002062 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002063 * With swappiness at 100, anonymous and file have the same priority.
2064 * This scanning priority is essentially the inverse of IO cost.
2065 */
Johannes Weiner02695172014-08-06 16:06:17 -07002066 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002067 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002068
2069 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002070 * OK, so we have swap space and a fair amount of page cache
2071 * pages. We use the recently rotated / recently scanned
2072 * ratios to determine how valuable each cache is.
2073 *
2074 * Because workloads change over time (and to avoid overflow)
2075 * we keep these statistics as a floating average, which ends
2076 * up weighing recent references more than old ones.
2077 *
2078 * anon in [0], file in [1]
2079 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002080
2081 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2082 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2083 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2084 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2085
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002086 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002087 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002088 reclaim_stat->recent_scanned[0] /= 2;
2089 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002090 }
2091
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002092 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002093 reclaim_stat->recent_scanned[1] /= 2;
2094 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002095 }
2096
2097 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002098 * The amount of pressure on anon vs file pages is inversely
2099 * proportional to the fraction of recently scanned pages on
2100 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002101 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002102 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002103 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002104
Satoru Moriyafe350042012-05-29 15:06:47 -07002105 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002106 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002107 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002108
Shaohua Li76a33fc2010-05-24 14:32:36 -07002109 fraction[0] = ap;
2110 fraction[1] = fp;
2111 denominator = ap + fp + 1;
2112out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002113 some_scanned = false;
2114 /* Only use force_scan on second pass. */
2115 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002116 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002117 for_each_evictable_lru(lru) {
2118 int file = is_file_lru(lru);
2119 unsigned long size;
2120 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002121
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002122 size = get_lru_size(lruvec, lru);
2123 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002124
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002125 if (!scan && pass && force_scan)
2126 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002127
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002128 switch (scan_balance) {
2129 case SCAN_EQUAL:
2130 /* Scan lists relative to size */
2131 break;
2132 case SCAN_FRACT:
2133 /*
2134 * Scan types proportional to swappiness and
2135 * their relative recent reclaim efficiency.
2136 */
2137 scan = div64_u64(scan * fraction[file],
2138 denominator);
2139 break;
2140 case SCAN_FILE:
2141 case SCAN_ANON:
2142 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002143 if ((scan_balance == SCAN_FILE) != file) {
2144 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002145 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002146 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002147 break;
2148 default:
2149 /* Look ma, no brain */
2150 BUG();
2151 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002152
2153 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002154 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002155
Johannes Weiner9a265112013-02-22 16:32:17 -08002156 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002157 * Skip the second pass and don't force_scan,
2158 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002159 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002160 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002161 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002162 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002163}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002164
Mel Gorman72b252a2015-09-04 15:47:32 -07002165#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2166static void init_tlb_ubc(void)
2167{
2168 /*
2169 * This deliberately does not clear the cpumask as it's expensive
2170 * and unnecessary. If there happens to be data in there then the
2171 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2172 * then will be cleared.
2173 */
2174 current->tlb_ubc.flush_required = false;
2175}
2176#else
2177static inline void init_tlb_ubc(void)
2178{
2179}
2180#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2181
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002182/*
2183 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2184 */
Johannes Weiner02695172014-08-06 16:06:17 -07002185static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002186 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002187{
2188 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002189 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002190 unsigned long nr_to_scan;
2191 enum lru_list lru;
2192 unsigned long nr_reclaimed = 0;
2193 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2194 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002195 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002196
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002197 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002198
Mel Gormane82e0562013-07-03 15:01:44 -07002199 /* Record the original scan target for proportional adjustments later */
2200 memcpy(targets, nr, sizeof(nr));
2201
Mel Gorman1a501902014-06-04 16:10:49 -07002202 /*
2203 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2204 * event that can occur when there is little memory pressure e.g.
2205 * multiple streaming readers/writers. Hence, we do not abort scanning
2206 * when the requested number of pages are reclaimed when scanning at
2207 * DEF_PRIORITY on the assumption that the fact we are direct
2208 * reclaiming implies that kswapd is not keeping up and it is best to
2209 * do a batch of work at once. For memcg reclaim one check is made to
2210 * abort proportional reclaim if either the file or anon lru has already
2211 * dropped to zero at the first pass.
2212 */
2213 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2214 sc->priority == DEF_PRIORITY);
2215
Mel Gorman72b252a2015-09-04 15:47:32 -07002216 init_tlb_ubc();
2217
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002218 blk_start_plug(&plug);
2219 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2220 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002221 unsigned long nr_anon, nr_file, percentage;
2222 unsigned long nr_scanned;
2223
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002224 for_each_evictable_lru(lru) {
2225 if (nr[lru]) {
2226 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2227 nr[lru] -= nr_to_scan;
2228
2229 nr_reclaimed += shrink_list(lru, nr_to_scan,
2230 lruvec, sc);
2231 }
2232 }
Mel Gormane82e0562013-07-03 15:01:44 -07002233
2234 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2235 continue;
2236
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002237 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002238 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002239 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002240 * proportionally what was requested by get_scan_count(). We
2241 * stop reclaiming one LRU and reduce the amount scanning
2242 * proportional to the original scan target.
2243 */
2244 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2245 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2246
Mel Gorman1a501902014-06-04 16:10:49 -07002247 /*
2248 * It's just vindictive to attack the larger once the smaller
2249 * has gone to zero. And given the way we stop scanning the
2250 * smaller below, this makes sure that we only make one nudge
2251 * towards proportionality once we've got nr_to_reclaim.
2252 */
2253 if (!nr_file || !nr_anon)
2254 break;
2255
Mel Gormane82e0562013-07-03 15:01:44 -07002256 if (nr_file > nr_anon) {
2257 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2258 targets[LRU_ACTIVE_ANON] + 1;
2259 lru = LRU_BASE;
2260 percentage = nr_anon * 100 / scan_target;
2261 } else {
2262 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2263 targets[LRU_ACTIVE_FILE] + 1;
2264 lru = LRU_FILE;
2265 percentage = nr_file * 100 / scan_target;
2266 }
2267
2268 /* Stop scanning the smaller of the LRU */
2269 nr[lru] = 0;
2270 nr[lru + LRU_ACTIVE] = 0;
2271
2272 /*
2273 * Recalculate the other LRU scan count based on its original
2274 * scan target and the percentage scanning already complete
2275 */
2276 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2277 nr_scanned = targets[lru] - nr[lru];
2278 nr[lru] = targets[lru] * (100 - percentage) / 100;
2279 nr[lru] -= min(nr[lru], nr_scanned);
2280
2281 lru += LRU_ACTIVE;
2282 nr_scanned = targets[lru] - nr[lru];
2283 nr[lru] = targets[lru] * (100 - percentage) / 100;
2284 nr[lru] -= min(nr[lru], nr_scanned);
2285
2286 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002287 }
2288 blk_finish_plug(&plug);
2289 sc->nr_reclaimed += nr_reclaimed;
2290
2291 /*
2292 * Even if we did not try to evict anon pages at all, we want to
2293 * rebalance the anon lru active/inactive ratio.
2294 */
2295 if (inactive_anon_is_low(lruvec))
2296 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2297 sc, LRU_ACTIVE_ANON);
2298
2299 throttle_vm_writeout(sc->gfp_mask);
2300}
2301
Mel Gorman23b9da52012-05-29 15:06:20 -07002302/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002303static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002304{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002305 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002306 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002307 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002308 return true;
2309
2310 return false;
2311}
2312
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002313/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002314 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2315 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2316 * true if more pages should be reclaimed such that when the page allocator
2317 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2318 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002319 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002320static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002321 unsigned long nr_reclaimed,
2322 unsigned long nr_scanned,
2323 struct scan_control *sc)
2324{
2325 unsigned long pages_for_compaction;
2326 unsigned long inactive_lru_pages;
2327
2328 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002329 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002330 return false;
2331
Mel Gorman28765922011-02-25 14:44:20 -08002332 /* Consider stopping depending on scan and reclaim activity */
2333 if (sc->gfp_mask & __GFP_REPEAT) {
2334 /*
2335 * For __GFP_REPEAT allocations, stop reclaiming if the
2336 * full LRU list has been scanned and we are still failing
2337 * to reclaim pages. This full LRU scan is potentially
2338 * expensive but a __GFP_REPEAT caller really wants to succeed
2339 */
2340 if (!nr_reclaimed && !nr_scanned)
2341 return false;
2342 } else {
2343 /*
2344 * For non-__GFP_REPEAT allocations which can presumably
2345 * fail without consequence, stop if we failed to reclaim
2346 * any pages from the last SWAP_CLUSTER_MAX number of
2347 * pages that were scanned. This will return to the
2348 * caller faster at the risk reclaim/compaction and
2349 * the resulting allocation attempt fails
2350 */
2351 if (!nr_reclaimed)
2352 return false;
2353 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002354
2355 /*
2356 * If we have not reclaimed enough pages for compaction and the
2357 * inactive lists are large enough, continue reclaiming
2358 */
2359 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002360 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002361 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002362 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002363 if (sc->nr_reclaimed < pages_for_compaction &&
2364 inactive_lru_pages > pages_for_compaction)
2365 return true;
2366
2367 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002368 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002369 case COMPACT_PARTIAL:
2370 case COMPACT_CONTINUE:
2371 return false;
2372 default:
2373 return true;
2374 }
2375}
2376
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002377static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2378 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002379{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002380 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002381 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002382 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002383
Johannes Weiner56600482012-01-12 17:17:59 -08002384 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002385 struct mem_cgroup *root = sc->target_mem_cgroup;
2386 struct mem_cgroup_reclaim_cookie reclaim = {
2387 .zone = zone,
2388 .priority = sc->priority,
2389 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002390 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002391 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002392
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002393 nr_reclaimed = sc->nr_reclaimed;
2394 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002395
Andrew Morton694fbc02013-09-24 15:27:37 -07002396 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2397 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002398 unsigned long lru_pages;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002399 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002400 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002401 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002402
Johannes Weiner241994ed2015-02-11 15:26:06 -08002403 if (mem_cgroup_low(root, memcg)) {
2404 if (!sc->may_thrash)
2405 continue;
2406 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2407 }
2408
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002409 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002410 swappiness = mem_cgroup_swappiness(memcg);
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002411 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002412
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002413 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2414 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002415
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002416 if (memcg && is_classzone)
2417 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2418 memcg, sc->nr_scanned - scanned,
2419 lru_pages);
2420
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002421 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002422 * Direct reclaim and kswapd have to scan all memory
2423 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002424 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002425 *
2426 * Limit reclaim, on the other hand, only cares about
2427 * nr_to_reclaim pages to be reclaimed and it will
2428 * retry with decreasing priority if one round over the
2429 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002430 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002431 if (!global_reclaim(sc) &&
2432 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002433 mem_cgroup_iter_break(root, memcg);
2434 break;
2435 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002436 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002437
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002438 /*
2439 * Shrink the slab caches in the same proportion that
2440 * the eligible LRU pages were scanned.
2441 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002442 if (global_reclaim(sc) && is_classzone)
2443 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2444 sc->nr_scanned - nr_scanned,
2445 zone_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002446
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002447 if (reclaim_state) {
2448 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2449 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002450 }
2451
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002452 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2453 sc->nr_scanned - nr_scanned,
2454 sc->nr_reclaimed - nr_reclaimed);
2455
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002456 if (sc->nr_reclaimed - nr_reclaimed)
2457 reclaimable = true;
2458
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002459 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2460 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002461
2462 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002463}
2464
Vlastimil Babka53853e22014-10-09 15:27:02 -07002465/*
2466 * Returns true if compaction should go ahead for a high-order request, or
2467 * the high-order allocation would succeed without compaction.
2468 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002469static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002470{
2471 unsigned long balance_gap, watermark;
2472 bool watermark_ok;
2473
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002474 /*
2475 * Compaction takes time to run and there are potentially other
2476 * callers using the pages just freed. Continue reclaiming until
2477 * there is a buffer of free pages available to give compaction
2478 * a reasonable chance of completing and allocating the page
2479 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002480 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2481 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002482 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002483 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2484
2485 /*
2486 * If compaction is deferred, reclaim up to a point where
2487 * compaction will have a chance of success when re-enabled
2488 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002489 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002490 return watermark_ok;
2491
Vlastimil Babka53853e22014-10-09 15:27:02 -07002492 /*
2493 * If compaction is not ready to start and allocation is not likely
2494 * to succeed without it, then keep reclaiming.
2495 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002496 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002497 return false;
2498
2499 return watermark_ok;
2500}
2501
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502/*
2503 * This is the direct reclaim path, for page-allocating processes. We only
2504 * try to reclaim pages from zones which will satisfy the caller's allocation
2505 * request.
2506 *
Mel Gorman41858962009-06-16 15:32:12 -07002507 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2508 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2510 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002511 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2512 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2513 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 * If a zone is deemed to be full of pinned pages then just give it a light
2516 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002517 *
2518 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002520static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
Mel Gormandd1a2392008-04-28 02:12:17 -07002522 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002523 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002524 unsigned long nr_soft_reclaimed;
2525 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002526 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002527 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002528 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002529
Mel Gormancc715d92012-03-21 16:34:00 -07002530 /*
2531 * If the number of buffer_heads in the machine exceeds the maximum
2532 * allowed level, force direct reclaim to scan the highmem zone as
2533 * highmem pages could be pinning lowmem pages storing buffer_heads
2534 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002535 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002536 if (buffer_heads_over_limit)
2537 sc->gfp_mask |= __GFP_HIGHMEM;
2538
Mel Gormand4debc62010-08-09 17:19:29 -07002539 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002540 requested_highidx, sc->nodemask) {
2541 enum zone_type classzone_idx;
2542
Con Kolivasf3fe6512006-01-06 00:11:15 -08002543 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002545
2546 classzone_idx = requested_highidx;
2547 while (!populated_zone(zone->zone_pgdat->node_zones +
2548 classzone_idx))
2549 classzone_idx--;
2550
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002551 /*
2552 * Take care memory controller reclaiming has small influence
2553 * to global LRU.
2554 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002555 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002556 if (!cpuset_zone_allowed(zone,
2557 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002558 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002559
Lisa Du6e543d52013-09-11 14:22:36 -07002560 if (sc->priority != DEF_PRIORITY &&
2561 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002562 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002563
2564 /*
2565 * If we already have plenty of memory free for
2566 * compaction in this zone, don't free any more.
2567 * Even though compaction is invoked for any
2568 * non-zero order, only frequent costly order
2569 * reclamation is disruptive enough to become a
2570 * noticeable problem, like transparent huge
2571 * page allocations.
2572 */
2573 if (IS_ENABLED(CONFIG_COMPACTION) &&
2574 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2575 zonelist_zone_idx(z) <= requested_highidx &&
2576 compaction_ready(zone, sc->order)) {
2577 sc->compaction_ready = true;
2578 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002579 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002580
Andrew Morton0608f432013-09-24 15:27:41 -07002581 /*
2582 * This steals pages from memory cgroups over softlimit
2583 * and returns the number of reclaimed pages and
2584 * scanned pages. This works for global memory pressure
2585 * and balancing, not for a memcg's limit.
2586 */
2587 nr_soft_scanned = 0;
2588 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2589 sc->order, sc->gfp_mask,
2590 &nr_soft_scanned);
2591 sc->nr_reclaimed += nr_soft_reclaimed;
2592 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002593 if (nr_soft_reclaimed)
2594 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002595 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002596 }
Nick Piggin408d8542006-09-25 23:31:27 -07002597
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002598 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002599 reclaimable = true;
2600
2601 if (global_reclaim(sc) &&
2602 !reclaimable && zone_reclaimable(zone))
2603 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 }
Mel Gormane0c23272011-10-31 17:09:33 -07002605
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002606 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002607 * Restore to original mask to avoid the impact on the caller if we
2608 * promoted it to __GFP_HIGHMEM.
2609 */
2610 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002611
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002612 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002614
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615/*
2616 * This is the main entry point to direct page reclaim.
2617 *
2618 * If a full scan of the inactive list fails to free enough memory then we
2619 * are "out of memory" and something needs to be killed.
2620 *
2621 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2622 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002623 * caller can't do much about. We kick the writeback threads and take explicit
2624 * naps in the hope that some of these pages can be written. But if the
2625 * allocating task holds filesystem locks which prevent writeout this might not
2626 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002627 *
2628 * returns: 0, if no pages reclaimed
2629 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 */
Mel Gormandac1d272008-04-28 02:12:12 -07002631static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002632 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002634 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002635 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002636 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002637 bool zones_reclaimable;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002638retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002639 delayacct_freepages_start();
2640
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002641 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002642 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002644 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002645 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2646 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002647 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002648 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002649
Balbir Singh66e17072008-02-07 00:13:56 -08002650 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002651 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002652 break;
2653
2654 if (sc->compaction_ready)
2655 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656
2657 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002658 * If we're getting trouble reclaiming, start doing
2659 * writepage even in laptop mode.
2660 */
2661 if (sc->priority < DEF_PRIORITY - 2)
2662 sc->may_writepage = 1;
2663
2664 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 * Try to write back as many pages as we just scanned. This
2666 * tends to cause slow streaming writers to write data to the
2667 * disk smoothly, at the dirtying rate, which is nice. But
2668 * that's undesirable in laptop mode, where we *want* lumpy
2669 * writeout. So in laptop mode, write out the whole world.
2670 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002671 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2672 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002673 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2674 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002675 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002677 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002678
Keika Kobayashi873b4772008-07-25 01:48:52 -07002679 delayacct_freepages_end();
2680
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002681 if (sc->nr_reclaimed)
2682 return sc->nr_reclaimed;
2683
Mel Gorman0cee34f2012-01-12 17:19:49 -08002684 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002685 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002686 return 1;
2687
Johannes Weiner241994ed2015-02-11 15:26:06 -08002688 /* Untapped cgroup reserves? Don't OOM, retry. */
2689 if (!sc->may_thrash) {
2690 sc->priority = initial_priority;
2691 sc->may_thrash = 1;
2692 goto retry;
2693 }
2694
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002695 /* Any of the zones still reclaimable? Don't OOM. */
2696 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002697 return 1;
2698
2699 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700}
2701
Mel Gorman55150612012-07-31 16:44:35 -07002702static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2703{
2704 struct zone *zone;
2705 unsigned long pfmemalloc_reserve = 0;
2706 unsigned long free_pages = 0;
2707 int i;
2708 bool wmark_ok;
2709
2710 for (i = 0; i <= ZONE_NORMAL; i++) {
2711 zone = &pgdat->node_zones[i];
Nishanth Aravamudanf012a842015-06-24 16:56:39 -07002712 if (!populated_zone(zone) ||
2713 zone_reclaimable_pages(zone) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002714 continue;
2715
Mel Gorman55150612012-07-31 16:44:35 -07002716 pfmemalloc_reserve += min_wmark_pages(zone);
2717 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2718 }
2719
Mel Gorman675becc2014-06-04 16:07:35 -07002720 /* If there are no reserves (unexpected config) then do not throttle */
2721 if (!pfmemalloc_reserve)
2722 return true;
2723
Mel Gorman55150612012-07-31 16:44:35 -07002724 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2725
2726 /* kswapd must be awake if processes are being throttled */
2727 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2728 pgdat->classzone_idx = min(pgdat->classzone_idx,
2729 (enum zone_type)ZONE_NORMAL);
2730 wake_up_interruptible(&pgdat->kswapd_wait);
2731 }
2732
2733 return wmark_ok;
2734}
2735
2736/*
2737 * Throttle direct reclaimers if backing storage is backed by the network
2738 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2739 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002740 * when the low watermark is reached.
2741 *
2742 * Returns true if a fatal signal was delivered during throttling. If this
2743 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002744 */
Mel Gorman50694c22012-11-26 16:29:48 -08002745static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002746 nodemask_t *nodemask)
2747{
Mel Gorman675becc2014-06-04 16:07:35 -07002748 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002749 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002750 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002751
2752 /*
2753 * Kernel threads should not be throttled as they may be indirectly
2754 * responsible for cleaning pages necessary for reclaim to make forward
2755 * progress. kjournald for example may enter direct reclaim while
2756 * committing a transaction where throttling it could forcing other
2757 * processes to block on log_wait_commit().
2758 */
2759 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002760 goto out;
2761
2762 /*
2763 * If a fatal signal is pending, this process should not throttle.
2764 * It should return quickly so it can exit and free its memory
2765 */
2766 if (fatal_signal_pending(current))
2767 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002768
Mel Gorman675becc2014-06-04 16:07:35 -07002769 /*
2770 * Check if the pfmemalloc reserves are ok by finding the first node
2771 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2772 * GFP_KERNEL will be required for allocating network buffers when
2773 * swapping over the network so ZONE_HIGHMEM is unusable.
2774 *
2775 * Throttling is based on the first usable node and throttled processes
2776 * wait on a queue until kswapd makes progress and wakes them. There
2777 * is an affinity then between processes waking up and where reclaim
2778 * progress has been made assuming the process wakes on the same node.
2779 * More importantly, processes running on remote nodes will not compete
2780 * for remote pfmemalloc reserves and processes on different nodes
2781 * should make reasonable progress.
2782 */
2783 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002784 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002785 if (zone_idx(zone) > ZONE_NORMAL)
2786 continue;
2787
2788 /* Throttle based on the first usable node */
2789 pgdat = zone->zone_pgdat;
2790 if (pfmemalloc_watermark_ok(pgdat))
2791 goto out;
2792 break;
2793 }
2794
2795 /* If no zone was usable by the allocation flags then do not throttle */
2796 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002797 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002798
Mel Gorman68243e72012-07-31 16:44:39 -07002799 /* Account for the throttling */
2800 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2801
Mel Gorman55150612012-07-31 16:44:35 -07002802 /*
2803 * If the caller cannot enter the filesystem, it's possible that it
2804 * is due to the caller holding an FS lock or performing a journal
2805 * transaction in the case of a filesystem like ext[3|4]. In this case,
2806 * it is not safe to block on pfmemalloc_wait as kswapd could be
2807 * blocked waiting on the same lock. Instead, throttle for up to a
2808 * second before continuing.
2809 */
2810 if (!(gfp_mask & __GFP_FS)) {
2811 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2812 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002813
2814 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002815 }
2816
2817 /* Throttle until kswapd wakes the process */
2818 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2819 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002820
2821check_pending:
2822 if (fatal_signal_pending(current))
2823 return true;
2824
2825out:
2826 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002827}
2828
Mel Gormandac1d272008-04-28 02:12:12 -07002829unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002830 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002831{
Mel Gorman33906bc2010-08-09 17:19:16 -07002832 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002833 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002834 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002835 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2836 .order = order,
2837 .nodemask = nodemask,
2838 .priority = DEF_PRIORITY,
2839 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002840 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002841 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002842 };
2843
Mel Gorman55150612012-07-31 16:44:35 -07002844 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002845 * Do not enter reclaim if fatal signal was delivered while throttled.
2846 * 1 is returned so that the page allocator does not OOM kill at this
2847 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002848 */
Mel Gorman50694c22012-11-26 16:29:48 -08002849 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002850 return 1;
2851
Mel Gorman33906bc2010-08-09 17:19:16 -07002852 trace_mm_vmscan_direct_reclaim_begin(order,
2853 sc.may_writepage,
2854 gfp_mask);
2855
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002856 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002857
2858 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2859
2860 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002861}
2862
Andrew Mortonc255a452012-07-31 16:43:02 -07002863#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002864
Johannes Weiner72835c82012-01-12 17:18:32 -08002865unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002866 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002867 struct zone *zone,
2868 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002869{
2870 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002871 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002872 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002873 .may_writepage = !laptop_mode,
2874 .may_unmap = 1,
2875 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002876 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002877 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002878 int swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002879 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002880
Balbir Singh4e416952009-09-23 15:56:39 -07002881 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2882 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002883
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002884 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002885 sc.may_writepage,
2886 sc.gfp_mask);
2887
Balbir Singh4e416952009-09-23 15:56:39 -07002888 /*
2889 * NOTE: Although we can get the priority field, using it
2890 * here is not a good idea, since it limits the pages we can scan.
2891 * if we don't reclaim here, the shrink_zone from balance_pgdat
2892 * will pick up pages from other mem cgroup's as well. We hack
2893 * the priority and make it zero.
2894 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002895 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002896
2897 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2898
Ying Han0ae5e892011-05-26 16:25:25 -07002899 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002900 return sc.nr_reclaimed;
2901}
2902
Johannes Weiner72835c82012-01-12 17:18:32 -08002903unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002904 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002905 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002906 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002907{
Balbir Singh4e416952009-09-23 15:56:39 -07002908 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002909 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002910 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002911 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002912 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002913 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2914 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2915 .target_mem_cgroup = memcg,
2916 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002917 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002918 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002919 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002920 };
Balbir Singh66e17072008-02-07 00:13:56 -08002921
Ying Han889976d2011-05-26 16:25:33 -07002922 /*
2923 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2924 * take care of from where we get pages. So the node where we start the
2925 * scan does not need to be the current node.
2926 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002927 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002928
2929 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002930
2931 trace_mm_vmscan_memcg_reclaim_begin(0,
2932 sc.may_writepage,
2933 sc.gfp_mask);
2934
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002935 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002936
2937 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2938
2939 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002940}
2941#endif
2942
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002943static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002944{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002945 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002946
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002947 if (!total_swap_pages)
2948 return;
2949
2950 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2951 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002952 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002953
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002954 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002955 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002956 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002957
2958 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2959 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002960}
2961
Johannes Weiner60cefed2012-11-29 13:54:23 -08002962static bool zone_balanced(struct zone *zone, int order,
2963 unsigned long balance_gap, int classzone_idx)
2964{
2965 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2966 balance_gap, classzone_idx, 0))
2967 return false;
2968
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002969 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2970 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002971 return false;
2972
2973 return true;
2974}
2975
Mel Gorman1741c872011-01-13 15:46:21 -08002976/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002977 * pgdat_balanced() is used when checking if a node is balanced.
2978 *
2979 * For order-0, all zones must be balanced!
2980 *
2981 * For high-order allocations only zones that meet watermarks and are in a
2982 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2983 * total of balanced pages must be at least 25% of the zones allowed by
2984 * classzone_idx for the node to be considered balanced. Forcing all zones to
2985 * be balanced for high orders can cause excessive reclaim when there are
2986 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002987 * The choice of 25% is due to
2988 * o a 16M DMA zone that is balanced will not balance a zone on any
2989 * reasonable sized machine
2990 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002991 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002992 * would need to be at least 256M for it to be balance a whole node.
2993 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2994 * to balance a node on its own. These seemed like reasonable ratios.
2995 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002996static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002997{
Jiang Liub40da042013-02-22 16:33:52 -08002998 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002999 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08003000 int i;
3001
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003002 /* Check the watermark levels */
3003 for (i = 0; i <= classzone_idx; i++) {
3004 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08003005
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003006 if (!populated_zone(zone))
3007 continue;
3008
Jiang Liub40da042013-02-22 16:33:52 -08003009 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003010
3011 /*
3012 * A special case here:
3013 *
3014 * balance_pgdat() skips over all_unreclaimable after
3015 * DEF_PRIORITY. Effectively, it considers them balanced so
3016 * they must be considered balanced here as well!
3017 */
Lisa Du6e543d52013-09-11 14:22:36 -07003018 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08003019 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003020 continue;
3021 }
3022
3023 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08003024 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003025 else if (!order)
3026 return false;
3027 }
3028
3029 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08003030 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003031 else
3032 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08003033}
3034
Mel Gorman55150612012-07-31 16:44:35 -07003035/*
3036 * Prepare kswapd for sleeping. This verifies that there are no processes
3037 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3038 *
3039 * Returns true if kswapd is ready to sleep
3040 */
3041static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08003042 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003043{
Mel Gormanf50de2d2009-12-14 17:58:53 -08003044 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3045 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07003046 return false;
3047
3048 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003049 * The throttled processes are normally woken up in balance_pgdat() as
3050 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3051 * race between when kswapd checks the watermarks and a process gets
3052 * throttled. There is also a potential race if processes get
3053 * throttled, kswapd wakes, a large process exits thereby balancing the
3054 * zones, which causes kswapd to exit balance_pgdat() before reaching
3055 * the wake up checks. If kswapd is going to sleep, no process should
3056 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3057 * the wake up is premature, processes will wake kswapd and get
3058 * throttled again. The difference from wake ups in balance_pgdat() is
3059 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003060 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003061 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3062 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003063
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003064 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003065}
3066
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067/*
Mel Gorman75485362013-07-03 15:01:42 -07003068 * kswapd shrinks the zone by the number of pages required to reach
3069 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003070 *
3071 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003072 * reclaim or if the lack of progress was due to pages under writeback.
3073 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003074 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003075static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07003076 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07003077 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07003078 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07003079{
Mel Gorman7c954f62013-07-03 15:01:54 -07003080 int testorder = sc->order;
3081 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07003082 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07003083
3084 /* Reclaim above the high watermark. */
3085 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07003086
3087 /*
3088 * Kswapd reclaims only single pages with compaction enabled. Trying
3089 * too hard to reclaim until contiguous free pages have become
3090 * available can hurt performance by evicting too much useful data
3091 * from memory. Do not reclaim more than needed for compaction.
3092 */
3093 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08003094 compaction_suitable(zone, sc->order, 0, classzone_idx)
3095 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07003096 testorder = 0;
3097
3098 /*
3099 * We put equal pressure on every zone, unless one zone has way too
3100 * many pages free already. The "too many pages" is defined as the
3101 * high wmark plus a "gap" where the gap is either the low
3102 * watermark or 1% of the zone, whichever is smaller.
3103 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07003104 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3105 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003106
3107 /*
3108 * If there is no low memory pressure or the zone is balanced then no
3109 * reclaim is necessary
3110 */
3111 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3112 if (!lowmem_pressure && zone_balanced(zone, testorder,
3113 balance_gap, classzone_idx))
3114 return true;
3115
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003116 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003117
Mel Gorman2ab44f42013-07-03 15:01:47 -07003118 /* Account for the number of pages attempted to reclaim */
3119 *nr_attempted += sc->nr_to_reclaim;
3120
Johannes Weiner57054652014-10-09 15:28:17 -07003121 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003122
Mel Gorman7c954f62013-07-03 15:01:54 -07003123 /*
3124 * If a zone reaches its high watermark, consider it to be no longer
3125 * congested. It's possible there are dirty pages backed by congested
3126 * BDIs but as pressure is relieved, speculatively avoid congestion
3127 * waits.
3128 */
Lisa Du6e543d52013-09-11 14:22:36 -07003129 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003130 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003131 clear_bit(ZONE_CONGESTED, &zone->flags);
3132 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003133 }
3134
Mel Gormanb8e83b92013-07-03 15:01:45 -07003135 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003136}
3137
3138/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003140 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003142 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 *
3144 * There is special handling here for zones which are full of pinned pages.
3145 * This can happen if the pages are all mlocked, or if they are all used by
3146 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3147 * What we do is to detect the case where all pages in the zone have been
3148 * scanned twice and there has been zero successful reclaim. Mark the zone as
3149 * dead and from now on, only perform a short scan. Basically we're polling
3150 * the zone for when the problem goes away.
3151 *
3152 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003153 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3154 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3155 * lower zones regardless of the number of free pages in the lower zones. This
3156 * interoperates with the page allocator fallback scheme to ensure that aging
3157 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 */
Mel Gorman99504742011-01-13 15:46:20 -08003159static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003160 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003163 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003164 unsigned long nr_soft_reclaimed;
3165 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003166 struct scan_control sc = {
3167 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003168 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003169 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003170 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003171 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003172 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003173 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003174 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003176 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003177 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003178 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003179 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003180
3181 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003183 /*
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3186 */
3187 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003190 if (!populated_zone(zone))
3191 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192
Lisa Du6e543d52013-09-11 14:22:36 -07003193 if (sc.priority != DEF_PRIORITY &&
3194 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003195 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196
Rik van Riel556adec2008-10-18 20:26:34 -07003197 /*
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3200 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003201 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003202
Mel Gormancc715d92012-03-21 16:34:00 -07003203 /*
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3208 */
3209 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210 end_zone = i;
3211 break;
3212 }
3213
Johannes Weiner60cefed2012-11-29 13:54:23 -08003214 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003215 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003216 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003217 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003218 /*
3219 * If balanced, clear the dirty and congested
3220 * flags
3221 */
Johannes Weiner57054652014-10-09 15:28:17 -07003222 clear_bit(ZONE_CONGESTED, &zone->flags);
3223 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003226
Mel Gormanb8e83b92013-07-03 15:01:45 -07003227 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003228 goto out;
3229
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 for (i = 0; i <= end_zone; i++) {
3231 struct zone *zone = pgdat->node_zones + i;
3232
Mel Gorman2ab44f42013-07-03 15:01:47 -07003233 if (!populated_zone(zone))
3234 continue;
3235
Mel Gorman2ab44f42013-07-03 15:01:47 -07003236 /*
3237 * If any zone is currently balanced then kswapd will
3238 * not call compaction as it is expected that the
3239 * necessary pages are already available.
3240 */
3241 if (pgdat_needs_compaction &&
3242 zone_watermark_ok(zone, order,
3243 low_wmark_pages(zone),
3244 *classzone_idx, 0))
3245 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 }
3247
3248 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003249 * If we're getting trouble reclaiming, start doing writepage
3250 * even in laptop mode.
3251 */
3252 if (sc.priority < DEF_PRIORITY - 2)
3253 sc.may_writepage = 1;
3254
3255 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 * Now scan the zone in the dma->highmem direction, stopping
3257 * at the last zone which needs scanning.
3258 *
3259 * We do this because the page allocator works in the opposite
3260 * direction. This prevents the page allocator from allocating
3261 * pages behind kswapd's direction of progress, which would
3262 * cause too much scanning of the lower zones.
3263 */
3264 for (i = 0; i <= end_zone; i++) {
3265 struct zone *zone = pgdat->node_zones + i;
3266
Con Kolivasf3fe6512006-01-06 00:11:15 -08003267 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 continue;
3269
Lisa Du6e543d52013-09-11 14:22:36 -07003270 if (sc.priority != DEF_PRIORITY &&
3271 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272 continue;
3273
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003275
Andrew Morton0608f432013-09-24 15:27:41 -07003276 nr_soft_scanned = 0;
3277 /*
3278 * Call soft limit reclaim before calling shrink_zone.
3279 */
3280 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3281 order, sc.gfp_mask,
3282 &nr_soft_scanned);
3283 sc.nr_reclaimed += nr_soft_reclaimed;
3284
Rik van Riel32a43302007-10-16 01:24:50 -07003285 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003286 * There should be no need to raise the scanning
3287 * priority if enough pages are already being scanned
3288 * that that high watermark would be met at 100%
3289 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003290 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003291 if (kswapd_shrink_zone(zone, end_zone,
3292 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003293 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 }
Mel Gorman55150612012-07-31 16:44:35 -07003295
3296 /*
3297 * If the low watermark is met there is no need for processes
3298 * to be throttled on pfmemalloc_wait as they should not be
3299 * able to safely make forward progress. Wake them
3300 */
3301 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3302 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003303 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003304
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003306 * Fragmentation may mean that the system cannot be rebalanced
3307 * for high-order allocations in all zones. If twice the
3308 * allocation size has been reclaimed and the zones are still
3309 * not balanced then recheck the watermarks at order-0 to
3310 * prevent kswapd reclaiming excessively. Assume that a
3311 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003313 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003314 order = sc.order = 0;
3315
Mel Gormanb8e83b92013-07-03 15:01:45 -07003316 /* Check if kswapd should be suspending */
3317 if (try_to_freeze() || kthread_should_stop())
3318 break;
3319
3320 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003321 * Compact if necessary and kswapd is reclaiming at least the
3322 * high watermark number of pages as requsted
3323 */
3324 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3325 compact_pgdat(pgdat, order);
3326
3327 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003328 * Raise priority if scanning rate is too low or there was no
3329 * progress in reclaiming pages
3330 */
3331 if (raise_priority || !sc.nr_reclaimed)
3332 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003333 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003334 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335
Mel Gormanb8e83b92013-07-03 15:01:45 -07003336out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003337 /*
Mel Gorman55150612012-07-31 16:44:35 -07003338 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003339 * makes a decision on the order we were last reclaiming at. However,
3340 * if another caller entered the allocator slow path while kswapd
3341 * was awake, order will remain at the higher level
3342 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003343 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003344 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345}
3346
Mel Gormandc83edd2011-01-13 15:46:26 -08003347static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003348{
3349 long remaining = 0;
3350 DEFINE_WAIT(wait);
3351
3352 if (freezing(current) || kthread_should_stop())
3353 return;
3354
3355 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3356
3357 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003358 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003359 remaining = schedule_timeout(HZ/10);
3360 finish_wait(&pgdat->kswapd_wait, &wait);
3361 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3362 }
3363
3364 /*
3365 * After a short sleep, check if it was a premature sleep. If not, then
3366 * go fully to sleep until explicitly woken up.
3367 */
Mel Gorman55150612012-07-31 16:44:35 -07003368 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003369 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3370
3371 /*
3372 * vmstat counters are not perfectly accurate and the estimated
3373 * value for counters such as NR_FREE_PAGES can deviate from the
3374 * true value by nr_online_cpus * threshold. To avoid the zone
3375 * watermarks being breached while under pressure, we reduce the
3376 * per-cpu vmstat threshold while kswapd is awake and restore
3377 * them before going back to sleep.
3378 */
3379 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003380
Mel Gorman62997022012-10-08 16:32:47 -07003381 /*
3382 * Compaction records what page blocks it recently failed to
3383 * isolate pages from and skips them in the future scanning.
3384 * When kswapd is going to sleep, it is reasonable to assume
3385 * that pages and compaction may succeed so reset the cache.
3386 */
3387 reset_isolation_suitable(pgdat);
3388
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003389 if (!kthread_should_stop())
3390 schedule();
3391
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003392 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3393 } else {
3394 if (remaining)
3395 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3396 else
3397 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3398 }
3399 finish_wait(&pgdat->kswapd_wait, &wait);
3400}
3401
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402/*
3403 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003404 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 *
3406 * This basically trickles out pages so that we have _some_
3407 * free memory available even if there is no other activity
3408 * that frees anything up. This is needed for things like routing
3409 * etc, where we otherwise might have all activity going on in
3410 * asynchronous contexts that cannot page things out.
3411 *
3412 * If there are applications that are active memory-allocators
3413 * (most normal use), this basically shouldn't matter.
3414 */
3415static int kswapd(void *p)
3416{
Mel Gorman215ddd62011-07-08 15:39:40 -07003417 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003418 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003419 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003420 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 pg_data_t *pgdat = (pg_data_t*)p;
3422 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003423
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 struct reclaim_state reclaim_state = {
3425 .reclaimed_slab = 0,
3426 };
Rusty Russella70f7302009-03-13 14:49:46 +10303427 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Nick Piggincf40bd12009-01-21 08:12:39 +01003429 lockdep_set_current_reclaim_state(GFP_KERNEL);
3430
Rusty Russell174596a2009-01-01 10:12:29 +10303431 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003432 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 current->reclaim_state = &reclaim_state;
3434
3435 /*
3436 * Tell the memory management that we're a "memory allocator",
3437 * and that if we need more memory we should get access to it
3438 * regardless (see "__alloc_pages()"). "kswapd" should
3439 * never get caught in the normal page freeing logic.
3440 *
3441 * (Kswapd normally doesn't need memory anyway, but sometimes
3442 * you need a small amount of memory in order to be able to
3443 * page out something else, and this flag essentially protects
3444 * us from recursively trying to free more memory as we're
3445 * trying to free the first piece of memory in the first place).
3446 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003447 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003448 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449
Mel Gorman215ddd62011-07-08 15:39:40 -07003450 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003451 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003452 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003453 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003455 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003456
Mel Gorman215ddd62011-07-08 15:39:40 -07003457 /*
3458 * If the last balance_pgdat was unsuccessful it's unlikely a
3459 * new request of a similar or harder type will succeed soon
3460 * so consider going to sleep on the basis we reclaimed at
3461 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003462 if (balanced_classzone_idx >= new_classzone_idx &&
3463 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003464 new_order = pgdat->kswapd_max_order;
3465 new_classzone_idx = pgdat->classzone_idx;
3466 pgdat->kswapd_max_order = 0;
3467 pgdat->classzone_idx = pgdat->nr_zones - 1;
3468 }
3469
Mel Gorman99504742011-01-13 15:46:20 -08003470 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 /*
3472 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003473 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 */
3475 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003476 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003478 kswapd_try_to_sleep(pgdat, balanced_order,
3479 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003481 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003482 new_order = order;
3483 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003484 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003485 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487
David Rientjes8fe23e02009-12-14 17:58:33 -08003488 ret = try_to_freeze();
3489 if (kthread_should_stop())
3490 break;
3491
3492 /*
3493 * We can speed up thawing tasks if we don't call balance_pgdat
3494 * after returning from the refrigerator
3495 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003496 if (!ret) {
3497 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003498 balanced_classzone_idx = classzone_idx;
3499 balanced_order = balance_pgdat(pgdat, order,
3500 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003501 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003503
Johannes Weiner71abdc12014-06-06 14:35:35 -07003504 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003505 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003506 lockdep_clear_current_reclaim_state();
3507
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 return 0;
3509}
3510
3511/*
3512 * A zone is low on free memory, so wake its kswapd task to service it.
3513 */
Mel Gorman99504742011-01-13 15:46:20 -08003514void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515{
3516 pg_data_t *pgdat;
3517
Con Kolivasf3fe6512006-01-06 00:11:15 -08003518 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 return;
3520
Vladimir Davydov344736f2014-10-20 15:50:30 +04003521 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003523 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003524 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003525 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003526 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3527 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003528 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003530 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003531 return;
3532
3533 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003534 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535}
3536
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003537#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003539 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003540 * freed pages.
3541 *
3542 * Rather than trying to age LRUs the aim is to preserve the overall
3543 * LRU order by reclaiming preferentially
3544 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003546unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003548 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003549 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003550 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003551 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003552 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003553 .may_writepage = 1,
3554 .may_unmap = 1,
3555 .may_swap = 1,
3556 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003558 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003559 struct task_struct *p = current;
3560 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003562 p->flags |= PF_MEMALLOC;
3563 lockdep_set_current_reclaim_state(sc.gfp_mask);
3564 reclaim_state.reclaimed_slab = 0;
3565 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003566
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003567 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003568
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003569 p->reclaim_state = NULL;
3570 lockdep_clear_current_reclaim_state();
3571 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003572
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003573 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003575#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577/* It's optimal to keep kswapds on the same CPUs as their memory, but
3578 not required for correctness. So if the last cpu in a node goes
3579 away, we get changed to run anywhere: as the first one comes back,
3580 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003581static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3582 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003584 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003586 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003587 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003588 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303589 const struct cpumask *mask;
3590
3591 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003592
Rusty Russell3e597942009-01-01 10:12:24 +10303593 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003595 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596 }
3597 }
3598 return NOTIFY_OK;
3599}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
Yasunori Goto3218ae12006-06-27 02:53:33 -07003601/*
3602 * This kswapd start function will be called by init and node-hot-add.
3603 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3604 */
3605int kswapd_run(int nid)
3606{
3607 pg_data_t *pgdat = NODE_DATA(nid);
3608 int ret = 0;
3609
3610 if (pgdat->kswapd)
3611 return 0;
3612
3613 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3614 if (IS_ERR(pgdat->kswapd)) {
3615 /* failure at boot is fatal */
3616 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003617 pr_err("Failed to start kswapd on node %d\n", nid);
3618 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003619 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003620 }
3621 return ret;
3622}
3623
David Rientjes8fe23e02009-12-14 17:58:33 -08003624/*
Jiang Liud8adde12012-07-11 14:01:52 -07003625 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003626 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003627 */
3628void kswapd_stop(int nid)
3629{
3630 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3631
Jiang Liud8adde12012-07-11 14:01:52 -07003632 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003633 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003634 NODE_DATA(nid)->kswapd = NULL;
3635 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003636}
3637
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638static int __init kswapd_init(void)
3639{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003640 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003641
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003643 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003644 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 hotcpu_notifier(cpu_callback, 0);
3646 return 0;
3647}
3648
3649module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003650
3651#ifdef CONFIG_NUMA
3652/*
3653 * Zone reclaim mode
3654 *
3655 * If non-zero call zone_reclaim when the number of free pages falls below
3656 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003657 */
3658int zone_reclaim_mode __read_mostly;
3659
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003660#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003661#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003662#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003663#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003664
Christoph Lameter9eeff232006-01-18 17:42:31 -08003665/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003666 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3667 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3668 * a zone.
3669 */
3670#define ZONE_RECLAIM_PRIORITY 4
3671
Christoph Lameter9eeff232006-01-18 17:42:31 -08003672/*
Christoph Lameter96146342006-07-03 00:24:13 -07003673 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3674 * occur.
3675 */
3676int sysctl_min_unmapped_ratio = 1;
3677
3678/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003679 * If the number of slab pages in a zone grows beyond this percentage then
3680 * slab reclaim needs to occur.
3681 */
3682int sysctl_min_slab_ratio = 5;
3683
Mel Gorman90afa5d2009-06-16 15:33:20 -07003684static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3685{
3686 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3687 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3688 zone_page_state(zone, NR_ACTIVE_FILE);
3689
3690 /*
3691 * It's possible for there to be more file mapped pages than
3692 * accounted for by the pages on the file LRU lists because
3693 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3694 */
3695 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3696}
3697
3698/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3699static long zone_pagecache_reclaimable(struct zone *zone)
3700{
3701 long nr_pagecache_reclaimable;
3702 long delta = 0;
3703
3704 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003705 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003706 * potentially reclaimable. Otherwise, we have to worry about
3707 * pages like swapcache and zone_unmapped_file_pages() provides
3708 * a better estimate
3709 */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003710 if (zone_reclaim_mode & RECLAIM_UNMAP)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003711 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3712 else
3713 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3714
3715 /* If we can't clean pages, remove dirty pages from consideration */
3716 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3717 delta += zone_page_state(zone, NR_FILE_DIRTY);
3718
3719 /* Watch for any possible underflows due to delta */
3720 if (unlikely(delta > nr_pagecache_reclaimable))
3721 delta = nr_pagecache_reclaimable;
3722
3723 return nr_pagecache_reclaimable - delta;
3724}
3725
Christoph Lameter0ff38492006-09-25 23:31:52 -07003726/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003727 * Try to free up some pages from this zone through reclaim.
3728 */
Andrew Morton179e9632006-03-22 00:08:18 -08003729static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003730{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003731 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003732 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003733 struct task_struct *p = current;
3734 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003735 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003736 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003737 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003738 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003739 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003740 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003741 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003742 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003743 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003744
Christoph Lameter9eeff232006-01-18 17:42:31 -08003745 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003746 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003747 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003748 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003749 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003750 */
3751 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003752 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003753 reclaim_state.reclaimed_slab = 0;
3754 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003755
Mel Gorman90afa5d2009-06-16 15:33:20 -07003756 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003757 /*
3758 * Free memory by calling shrink zone with increasing
3759 * priorities until we have enough memory freed.
3760 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003761 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003762 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003763 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003764 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003765
Christoph Lameter9eeff232006-01-18 17:42:31 -08003766 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003767 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003768 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003769 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003770}
Andrew Morton179e9632006-03-22 00:08:18 -08003771
3772int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3773{
Andrew Morton179e9632006-03-22 00:08:18 -08003774 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003775 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003776
3777 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003778 * Zone reclaim reclaims unmapped file backed pages and
3779 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003780 *
Christoph Lameter96146342006-07-03 00:24:13 -07003781 * A small portion of unmapped file backed pages is needed for
3782 * file I/O otherwise pages read by file I/O will be immediately
3783 * thrown out if the zone is overallocated. So we do not reclaim
3784 * if less than a specified percentage of the zone is used by
3785 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003786 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003787 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3788 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003789 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003790
Lisa Du6e543d52013-09-11 14:22:36 -07003791 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003792 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003793
Andrew Morton179e9632006-03-22 00:08:18 -08003794 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003795 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003796 */
David Rientjesd773ed62007-10-16 23:26:01 -07003797 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003798 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003799
3800 /*
3801 * Only run zone reclaim on the local zone or on zones that do not
3802 * have associated processors. This will favor the local processor
3803 * over remote processors and spread off node memory allocations
3804 * as wide as possible.
3805 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003806 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003807 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003808 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003809
Johannes Weiner57054652014-10-09 15:28:17 -07003810 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003811 return ZONE_RECLAIM_NOSCAN;
3812
David Rientjesd773ed62007-10-16 23:26:01 -07003813 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003814 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003815
Mel Gorman24cf725182009-06-16 15:33:23 -07003816 if (!ret)
3817 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3818
David Rientjesd773ed62007-10-16 23:26:01 -07003819 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003820}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003821#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003822
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003823/*
3824 * page_evictable - test whether a page is evictable
3825 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003826 *
3827 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003828 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003829 *
3830 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003831 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003832 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003833 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003834 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003835int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003836{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003837 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003838}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003839
Hugh Dickins850465792012-01-20 14:34:19 -08003840#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003841/**
Hugh Dickins24513262012-01-20 14:34:21 -08003842 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3843 * @pages: array of pages to check
3844 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003845 *
Hugh Dickins24513262012-01-20 14:34:21 -08003846 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003847 *
3848 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003849 */
Hugh Dickins24513262012-01-20 14:34:21 -08003850void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003851{
Johannes Weiner925b7672012-01-12 17:18:15 -08003852 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003853 struct zone *zone = NULL;
3854 int pgscanned = 0;
3855 int pgrescued = 0;
3856 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003857
Hugh Dickins24513262012-01-20 14:34:21 -08003858 for (i = 0; i < nr_pages; i++) {
3859 struct page *page = pages[i];
3860 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003861
Hugh Dickins24513262012-01-20 14:34:21 -08003862 pgscanned++;
3863 pagezone = page_zone(page);
3864 if (pagezone != zone) {
3865 if (zone)
3866 spin_unlock_irq(&zone->lru_lock);
3867 zone = pagezone;
3868 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003869 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003870 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003871
Hugh Dickins24513262012-01-20 14:34:21 -08003872 if (!PageLRU(page) || !PageUnevictable(page))
3873 continue;
3874
Hugh Dickins39b5f292012-10-08 16:33:18 -07003875 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003876 enum lru_list lru = page_lru_base_type(page);
3877
Sasha Levin309381fea2014-01-23 15:52:54 -08003878 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003879 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003880 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3881 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003882 pgrescued++;
3883 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003884 }
Hugh Dickins24513262012-01-20 14:34:21 -08003885
3886 if (zone) {
3887 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3888 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3889 spin_unlock_irq(&zone->lru_lock);
3890 }
Hugh Dickins850465792012-01-20 14:34:19 -08003891}
3892#endif /* CONFIG_SHMEM */