blob: f46339870147fef6604d15f0a540046b4d12e220 [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 Weiner241994e2015-02-11 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}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800157#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800158static bool global_reclaim(struct scan_control *sc)
159{
160 return true;
161}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800162#endif
163
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800164static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700165{
166 int nr;
167
168 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
169 zone_page_state(zone, NR_INACTIVE_FILE);
170
171 if (get_nr_swap_pages() > 0)
172 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
173 zone_page_state(zone, NR_INACTIVE_ANON);
174
175 return nr;
176}
177
178bool zone_reclaimable(struct zone *zone)
179{
Mel Gorman0d5d8232014-08-06 16:07:16 -0700180 return zone_page_state(zone, NR_PAGES_SCANNED) <
181 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700182}
183
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700184static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800185{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700186 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700187 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800188
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700189 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800190}
191
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000193 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000195int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000197 size_t size = sizeof(*shrinker->nr_deferred);
198
199 /*
200 * If we only have one possible node in the system anyway, save
201 * ourselves the trouble and disable NUMA aware behavior. This way we
202 * will save memory and some small loop time later.
203 */
204 if (nr_node_ids == 1)
205 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
206
207 if (shrinker->flags & SHRINKER_NUMA_AWARE)
208 size *= nr_node_ids;
209
210 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
211 if (!shrinker->nr_deferred)
212 return -ENOMEM;
213
Rusty Russell8e1f9362007-07-17 04:03:17 -0700214 down_write(&shrinker_rwsem);
215 list_add_tail(&shrinker->list, &shrinker_list);
216 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000217 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700219EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220
221/*
222 * Remove one
223 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700224void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225{
226 down_write(&shrinker_rwsem);
227 list_del(&shrinker->list);
228 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700229 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700231EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
233#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000234
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800235static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
236 struct shrinker *shrinker,
237 unsigned long nr_scanned,
238 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000240 unsigned long freed = 0;
241 unsigned long long delta;
242 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700243 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000244 long nr;
245 long new_nr;
246 int nid = shrinkctl->nid;
247 long batch_size = shrinker->batch ? shrinker->batch
248 : SHRINK_BATCH;
249
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700250 freeable = shrinker->count_objects(shrinker, shrinkctl);
251 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000252 return 0;
253
254 /*
255 * copy the current shrinker scan count into a local variable
256 * and zero it so that other concurrent shrinker invocations
257 * don't also do this scanning work.
258 */
259 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
260
261 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800262 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700263 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800264 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000265 total_scan += delta;
266 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800267 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000268 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700269 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000270 }
271
272 /*
273 * We need to avoid excessive windup on filesystem shrinkers
274 * due to large numbers of GFP_NOFS allocations causing the
275 * shrinkers to return -1 all the time. This results in a large
276 * nr being built up so when a shrink that can do some work
277 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700278 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279 * memory.
280 *
281 * Hence only allow the shrinker to scan the entire cache when
282 * a large delta change is calculated directly.
283 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700284 if (delta < freeable / 4)
285 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000286
287 /*
288 * Avoid risking looping forever due to too large nr value:
289 * never try to free more than twice the estimate number of
290 * freeable entries.
291 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700292 if (total_scan > freeable * 2)
293 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000294
295 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800296 nr_scanned, nr_eligible,
297 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000298
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800299 /*
300 * Normally, we should not scan less than batch_size objects in one
301 * pass to avoid too frequent shrinker calls, but if the slab has less
302 * than batch_size objects in total and we are really tight on memory,
303 * we will try to reclaim all available objects, otherwise we can end
304 * up failing allocations although there are plenty of reclaimable
305 * objects spread over several slabs with usage less than the
306 * batch_size.
307 *
308 * We detect the "tight on memory" situations by looking at the total
309 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700310 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800311 * scanning at high prio and therefore should try to reclaim as much as
312 * possible.
313 */
314 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700315 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000316 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800317 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000318
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800319 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000320 ret = shrinker->scan_objects(shrinker, shrinkctl);
321 if (ret == SHRINK_STOP)
322 break;
323 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000324
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800325 count_vm_events(SLABS_SCANNED, nr_to_scan);
326 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000327
328 cond_resched();
329 }
330
331 /*
332 * move the unused scan count back into the shrinker in a
333 * manner that handles concurrent updates. If we exhausted the
334 * scan, there is no need to do an update.
335 */
336 if (total_scan > 0)
337 new_nr = atomic_long_add_return(total_scan,
338 &shrinker->nr_deferred[nid]);
339 else
340 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
341
Dave Hansendf9024a2014-06-04 16:08:07 -0700342 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000343 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344}
345
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800346/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800347 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800348 * @gfp_mask: allocation context
349 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800350 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800351 * @nr_scanned: pressure numerator
352 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800354 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800356 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
357 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800359 * @memcg specifies the memory cgroup to target. If it is not NULL,
360 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
361 * objects from the memory cgroup specified. Otherwise all shrinkers
362 * are called, and memcg aware shrinkers are supposed to scan the
363 * global list then.
364 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800365 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
366 * the available objects should be scanned. Page reclaim for example
367 * passes the number of pages scanned and the number of pages on the
368 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
369 * when it encountered mapped pages. The ratio is further biased by
370 * the ->seeks setting of the shrink function, which indicates the
371 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800373 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800375static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
376 struct mem_cgroup *memcg,
377 unsigned long nr_scanned,
378 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379{
380 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000381 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800383 if (memcg && !memcg_kmem_is_active(memcg))
384 return 0;
385
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800386 if (nr_scanned == 0)
387 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
Minchan Kimf06590b2011-05-24 17:11:11 -0700389 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000390 /*
391 * If we would return 0, our callers would understand that we
392 * have nothing else to shrink and give up trying. By returning
393 * 1 we keep it going and assume we'll be able to shrink next
394 * time.
395 */
396 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700397 goto out;
398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800401 struct shrink_control sc = {
402 .gfp_mask = gfp_mask,
403 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800404 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800405 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800406
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800407 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
408 continue;
409
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800410 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
411 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000412
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800413 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800415
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700417out:
418 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000419 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420}
421
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800422void drop_slab_node(int nid)
423{
424 unsigned long freed;
425
426 do {
427 struct mem_cgroup *memcg = NULL;
428
429 freed = 0;
430 do {
431 freed += shrink_slab(GFP_KERNEL, nid, memcg,
432 1000, 1000);
433 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
434 } while (freed > 10);
435}
436
437void drop_slab(void)
438{
439 int nid;
440
441 for_each_online_node(nid)
442 drop_slab_node(nid);
443}
444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445static inline int is_page_cache_freeable(struct page *page)
446{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700447 /*
448 * A freeable page cache page is referenced only by the caller
449 * that isolated the page, the page cache radix tree and
450 * optional buffer heads at page->private.
451 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700452 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453}
454
Tejun Heo703c2702015-05-22 17:13:44 -0400455static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Christoph Lameter930d9152006-01-08 01:00:47 -0800457 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400459 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400461 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 return 1;
463 return 0;
464}
465
466/*
467 * We detected a synchronous write error writing a page out. Probably
468 * -ENOSPC. We need to propagate that into the address_space for a subsequent
469 * fsync(), msync() or close().
470 *
471 * The tricky part is that after writepage we cannot touch the mapping: nothing
472 * prevents it from being freed up. But we have a ref on the page and once
473 * that page is locked, the mapping is pinned.
474 *
475 * We're allowed to run sleeping lock_page() here because we know the caller has
476 * __GFP_FS.
477 */
478static void handle_write_error(struct address_space *mapping,
479 struct page *page, int error)
480{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100481 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700482 if (page_mapping(page) == mapping)
483 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 unlock_page(page);
485}
486
Christoph Lameter04e62a22006-06-23 02:03:38 -0700487/* possible outcome of pageout() */
488typedef enum {
489 /* failed to write page out, page is locked */
490 PAGE_KEEP,
491 /* move page to the active list, page is locked */
492 PAGE_ACTIVATE,
493 /* page has been sent to the disk successfully, page is unlocked */
494 PAGE_SUCCESS,
495 /* page is clean and locked */
496 PAGE_CLEAN,
497} pageout_t;
498
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499/*
Andrew Morton1742f192006-03-22 00:08:21 -0800500 * pageout is called by shrink_page_list() for each dirty page.
501 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700503static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700504 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505{
506 /*
507 * If the page is dirty, only perform writeback if that write
508 * will be non-blocking. To prevent this allocation from being
509 * stalled by pagecache activity. But note that there may be
510 * stalls if we need to run get_block(). We could test
511 * PagePrivate for that.
512 *
Al Viro81742022014-04-03 03:17:43 -0400513 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * this page's queue, we can perform writeback even if that
515 * will block.
516 *
517 * If the page is swapcache, write it back even if that would
518 * block, for some throttling. This happens by accident, because
519 * swap_backing_dev_info is bust: it doesn't reflect the
520 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 */
522 if (!is_page_cache_freeable(page))
523 return PAGE_KEEP;
524 if (!mapping) {
525 /*
526 * Some data journaling orphaned pages can have
527 * page->mapping == NULL while being dirty with clean buffers.
528 */
David Howells266cf652009-04-03 16:42:36 +0100529 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 if (try_to_free_buffers(page)) {
531 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700532 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 return PAGE_CLEAN;
534 }
535 }
536 return PAGE_KEEP;
537 }
538 if (mapping->a_ops->writepage == NULL)
539 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400540 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 return PAGE_KEEP;
542
543 if (clear_page_dirty_for_io(page)) {
544 int res;
545 struct writeback_control wbc = {
546 .sync_mode = WB_SYNC_NONE,
547 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700548 .range_start = 0,
549 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 .for_reclaim = 1,
551 };
552
553 SetPageReclaim(page);
554 res = mapping->a_ops->writepage(page, &wbc);
555 if (res < 0)
556 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800557 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 ClearPageReclaim(page);
559 return PAGE_ACTIVATE;
560 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700561
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 if (!PageWriteback(page)) {
563 /* synchronous write or broken a_ops? */
564 ClearPageReclaim(page);
565 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700566 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700567 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 return PAGE_SUCCESS;
569 }
570
571 return PAGE_CLEAN;
572}
573
Andrew Mortona649fd92006-10-17 00:09:36 -0700574/*
Nick Piggine2867812008-07-25 19:45:30 -0700575 * Same as remove_mapping, but if the page is removed from the mapping, it
576 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700577 */
Johannes Weinera5289102014-04-03 14:47:51 -0700578static int __remove_mapping(struct address_space *mapping, struct page *page,
579 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800580{
Greg Thelenc4843a72015-05-22 17:13:16 -0400581 unsigned long flags;
582 struct mem_cgroup *memcg;
583
Nick Piggin28e4d962006-09-25 23:31:23 -0700584 BUG_ON(!PageLocked(page));
585 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800586
Greg Thelenc4843a72015-05-22 17:13:16 -0400587 memcg = mem_cgroup_begin_page_stat(page);
588 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800589 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700590 * The non racy check for a busy page.
591 *
592 * Must be careful with the order of the tests. When someone has
593 * a ref to the page, it may be possible that they dirty it then
594 * drop the reference. So if PageDirty is tested before page_count
595 * here, then the following race may occur:
596 *
597 * get_user_pages(&page);
598 * [user mapping goes away]
599 * write_to(page);
600 * !PageDirty(page) [good]
601 * SetPageDirty(page);
602 * put_page(page);
603 * !page_count(page) [good, discard it]
604 *
605 * [oops, our write_to data is lost]
606 *
607 * Reversing the order of the tests ensures such a situation cannot
608 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
609 * load is not satisfied before that of page->_count.
610 *
611 * Note that if SetPageDirty is always performed via set_page_dirty,
612 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800613 */
Nick Piggine2867812008-07-25 19:45:30 -0700614 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800615 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700616 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
617 if (unlikely(PageDirty(page))) {
618 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800619 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700620 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800621
622 if (PageSwapCache(page)) {
623 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700624 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800625 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400626 spin_unlock_irqrestore(&mapping->tree_lock, flags);
627 mem_cgroup_end_page_stat(memcg);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700628 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700629 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500630 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700631 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500632
633 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700634 /*
635 * Remember a shadow entry for reclaimed file cache in
636 * order to detect refaults, thus thrashing, later on.
637 *
638 * But don't store shadows in an address space that is
639 * already exiting. This is not just an optizimation,
640 * inode reclaim needs to empty out the radix tree or
641 * the nodes are lost. Don't plant shadows behind its
642 * back.
643 */
644 if (reclaimed && page_is_file_cache(page) &&
645 !mapping_exiting(mapping))
646 shadow = workingset_eviction(mapping, page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400647 __delete_from_page_cache(page, shadow, memcg);
648 spin_unlock_irqrestore(&mapping->tree_lock, flags);
649 mem_cgroup_end_page_stat(memcg);
Linus Torvalds6072d132010-12-01 13:35:19 -0500650
651 if (freepage != NULL)
652 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800653 }
654
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800655 return 1;
656
657cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400658 spin_unlock_irqrestore(&mapping->tree_lock, flags);
659 mem_cgroup_end_page_stat(memcg);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800660 return 0;
661}
662
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663/*
Nick Piggine2867812008-07-25 19:45:30 -0700664 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
665 * someone else has a ref on the page, abort and return 0. If it was
666 * successfully detached, return 1. Assumes the caller has a single ref on
667 * this page.
668 */
669int remove_mapping(struct address_space *mapping, struct page *page)
670{
Johannes Weinera5289102014-04-03 14:47:51 -0700671 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700672 /*
673 * Unfreezing the refcount with 1 rather than 2 effectively
674 * drops the pagecache ref for us without requiring another
675 * atomic operation.
676 */
677 page_unfreeze_refs(page, 1);
678 return 1;
679 }
680 return 0;
681}
682
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700683/**
684 * putback_lru_page - put previously isolated page onto appropriate LRU list
685 * @page: page to be put back to appropriate lru list
686 *
687 * Add previously isolated @page to appropriate LRU list.
688 * Page may still be unevictable for other reasons.
689 *
690 * lru_lock must not be held, interrupts must be enabled.
691 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700692void putback_lru_page(struct page *page)
693{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700694 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700695 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700696
Sasha Levin309381fea2014-01-23 15:52:54 -0800697 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700698
699redo:
700 ClearPageUnevictable(page);
701
Hugh Dickins39b5f292012-10-08 16:33:18 -0700702 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700703 /*
704 * For evictable pages, we can use the cache.
705 * In event of a race, worst case is we end up with an
706 * unevictable page on [in]active list.
707 * We know how to handle that.
708 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700709 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700710 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700711 } else {
712 /*
713 * Put unevictable pages directly on zone's unevictable
714 * list.
715 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700716 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700717 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700718 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700719 * When racing with an mlock or AS_UNEVICTABLE clearing
720 * (page is unlocked) make sure that if the other thread
721 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800722 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700723 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700724 * the page back to the evictable list.
725 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700726 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700727 */
728 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700729 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700730
731 /*
732 * page's status can change while we move it among lru. If an evictable
733 * page is on unevictable list, it never be freed. To avoid that,
734 * check after we added it to the list, again.
735 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700736 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700737 if (!isolate_lru_page(page)) {
738 put_page(page);
739 goto redo;
740 }
741 /* This means someone else dropped this page from LRU
742 * So, it will be freed or putback to LRU again. There is
743 * nothing to do here.
744 */
745 }
746
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700747 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700748 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700749 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700750 count_vm_event(UNEVICTABLE_PGCULLED);
751
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700752 put_page(page); /* drop ref from isolate */
753}
754
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800755enum page_references {
756 PAGEREF_RECLAIM,
757 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800758 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800759 PAGEREF_ACTIVATE,
760};
761
762static enum page_references page_check_references(struct page *page,
763 struct scan_control *sc)
764{
Johannes Weiner645747462010-03-05 13:42:22 -0800765 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800766 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800767
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700768 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
769 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800770 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800771
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800772 /*
773 * Mlock lost the isolation race with us. Let try_to_unmap()
774 * move the page to the unevictable list.
775 */
776 if (vm_flags & VM_LOCKED)
777 return PAGEREF_RECLAIM;
778
Johannes Weiner645747462010-03-05 13:42:22 -0800779 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700780 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800781 return PAGEREF_ACTIVATE;
782 /*
783 * All mapped pages start out with page table
784 * references from the instantiating fault, so we need
785 * to look twice if a mapped file page is used more
786 * than once.
787 *
788 * Mark it and spare it for another trip around the
789 * inactive list. Another page table reference will
790 * lead to its activation.
791 *
792 * Note: the mark is set for activated pages as well
793 * so that recently deactivated but used pages are
794 * quickly recovered.
795 */
796 SetPageReferenced(page);
797
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800798 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800799 return PAGEREF_ACTIVATE;
800
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800801 /*
802 * Activate file-backed executable pages after first usage.
803 */
804 if (vm_flags & VM_EXEC)
805 return PAGEREF_ACTIVATE;
806
Johannes Weiner645747462010-03-05 13:42:22 -0800807 return PAGEREF_KEEP;
808 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800809
810 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700811 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800812 return PAGEREF_RECLAIM_CLEAN;
813
814 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800815}
816
Mel Gormane2be15f2013-07-03 15:01:57 -0700817/* Check if a page is dirty or under writeback */
818static void page_check_dirty_writeback(struct page *page,
819 bool *dirty, bool *writeback)
820{
Mel Gormanb4597222013-07-03 15:02:05 -0700821 struct address_space *mapping;
822
Mel Gormane2be15f2013-07-03 15:01:57 -0700823 /*
824 * Anonymous pages are not handled by flushers and must be written
825 * from reclaim context. Do not stall reclaim based on them
826 */
827 if (!page_is_file_cache(page)) {
828 *dirty = false;
829 *writeback = false;
830 return;
831 }
832
833 /* By default assume that the page flags are accurate */
834 *dirty = PageDirty(page);
835 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700836
837 /* Verify dirty/writeback state if the filesystem supports it */
838 if (!page_has_private(page))
839 return;
840
841 mapping = page_mapping(page);
842 if (mapping && mapping->a_ops->is_dirty_writeback)
843 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700844}
845
Nick Piggine2867812008-07-25 19:45:30 -0700846/*
Andrew Morton1742f192006-03-22 00:08:21 -0800847 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 */
Andrew Morton1742f192006-03-22 00:08:21 -0800849static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700850 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700851 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700852 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700853 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700854 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700855 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700856 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700857 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700858 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859{
860 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700861 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700863 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700864 unsigned long nr_dirty = 0;
865 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800866 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700867 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700868 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869
870 cond_resched();
871
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 while (!list_empty(page_list)) {
873 struct address_space *mapping;
874 struct page *page;
875 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700876 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700877 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878
879 cond_resched();
880
881 page = lru_to_page(page_list);
882 list_del(&page->lru);
883
Nick Piggin529ae9a2008-08-02 12:01:03 +0200884 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 goto keep;
886
Sasha Levin309381fea2014-01-23 15:52:54 -0800887 VM_BUG_ON_PAGE(PageActive(page), page);
888 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
890 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800891
Hugh Dickins39b5f292012-10-08 16:33:18 -0700892 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700893 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700894
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700895 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800896 goto keep_locked;
897
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 /* Double the slab pressure for mapped and swapcache pages */
899 if (page_mapped(page) || PageSwapCache(page))
900 sc->nr_scanned++;
901
Andy Whitcroftc661b072007-08-22 14:01:26 -0700902 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
903 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
904
Mel Gorman283aba92013-07-03 15:01:51 -0700905 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700906 * The number of dirty pages determines if a zone is marked
907 * reclaim_congested which affects wait_iff_congested. kswapd
908 * will stall and start writing pages if the tail of the LRU
909 * is all dirty unqueued pages.
910 */
911 page_check_dirty_writeback(page, &dirty, &writeback);
912 if (dirty || writeback)
913 nr_dirty++;
914
915 if (dirty && !writeback)
916 nr_unqueued_dirty++;
917
Mel Gormand04e8ac2013-07-03 15:02:03 -0700918 /*
919 * Treat this page as congested if the underlying BDI is or if
920 * pages are cycling through the LRU so quickly that the
921 * pages marked for immediate reclaim are making it to the
922 * end of the LRU a second time.
923 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700924 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800925 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400926 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700927 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700928 nr_congested++;
929
930 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700931 * If a page at the tail of the LRU is under writeback, there
932 * are three cases to consider.
933 *
934 * 1) If reclaim is encountering an excessive number of pages
935 * under writeback and this page is both under writeback and
936 * PageReclaim then it indicates that pages are being queued
937 * for IO but are being recycled through the LRU before the
938 * IO can complete. Waiting on the page itself risks an
939 * indefinite stall if it is impossible to writeback the
940 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700941 * note that the LRU is being scanned too quickly and the
942 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700943 *
944 * 2) Global reclaim encounters a page, memcg encounters a
945 * page that is not marked for immediate reclaim or
946 * the caller does not have __GFP_IO. In this case mark
947 * the page for immediate reclaim and continue scanning.
948 *
949 * __GFP_IO is checked because a loop driver thread might
950 * enter reclaim, and deadlock if it waits on a page for
951 * which it is needed to do the write (loop masks off
952 * __GFP_IO|__GFP_FS for this reason); but more thought
953 * would probably show more reasons.
954 *
955 * Don't require __GFP_FS, since we're not going into the
956 * FS, just waiting on its writeback completion. Worryingly,
957 * ext4 gfs2 and xfs allocate pages with
958 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
959 * may_enter_fs here is liable to OOM on them.
960 *
961 * 3) memcg encounters a page that is not already marked
962 * PageReclaim. memcg does not have any dirty pages
963 * throttling so we could easily OOM just because too many
964 * pages are in writeback and there is nothing else to
965 * reclaim. Wait for the writeback to complete.
966 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700967 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700968 /* Case 1 above */
969 if (current_is_kswapd() &&
970 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700971 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700972 nr_immediate++;
973 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700974
975 /* Case 2 above */
976 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700977 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
978 /*
979 * This is slightly racy - end_page_writeback()
980 * might have just cleared PageReclaim, then
981 * setting PageReclaim here end up interpreted
982 * as PageReadahead - but that does not matter
983 * enough to care. What we do want is for this
984 * page to have PageReclaim set next time memcg
985 * reclaim reaches the tests above, so it will
986 * then wait_on_page_writeback() to avoid OOM;
987 * and it's also appropriate in global reclaim.
988 */
989 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700990 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700991
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700992 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700993
994 /* Case 3 above */
995 } else {
996 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700997 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700998 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999
Minchan Kim02c6de82012-10-08 16:31:55 -07001000 if (!force_reclaim)
1001 references = page_check_references(page, sc);
1002
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001003 switch (references) {
1004 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001006 case PAGEREF_KEEP:
1007 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001008 case PAGEREF_RECLAIM:
1009 case PAGEREF_RECLAIM_CLEAN:
1010 ; /* try to reclaim the page below */
1011 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 /*
1014 * Anonymous process memory has backing store?
1015 * Try to allocate it some swap space here.
1016 */
Nick Pigginb291f002008-10-18 20:26:44 -07001017 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001018 if (!(sc->gfp_mask & __GFP_IO))
1019 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001020 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001022 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023
Mel Gormane2be15f2013-07-03 15:01:57 -07001024 /* Adding to swap updated mapping */
1025 mapping = page_mapping(page);
1026 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027
1028 /*
1029 * The page is mapped into the page tables of one or more
1030 * processes. Try to unmap it here.
1031 */
1032 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001033 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 case SWAP_FAIL:
1035 goto activate_locked;
1036 case SWAP_AGAIN:
1037 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001038 case SWAP_MLOCK:
1039 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 case SWAP_SUCCESS:
1041 ; /* try to free the page below */
1042 }
1043 }
1044
1045 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001046 /*
1047 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001048 * avoid risk of stack overflow but only writeback
1049 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001050 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001051 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001052 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001053 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001054 /*
1055 * Immediately reclaim when written back.
1056 * Similar in principal to deactivate_page()
1057 * except we already have the page isolated
1058 * and know it's dirty
1059 */
1060 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1061 SetPageReclaim(page);
1062
Mel Gormanee728862011-10-31 17:07:38 -07001063 goto keep_locked;
1064 }
1065
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001066 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001068 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001070 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 goto keep_locked;
1072
1073 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001074 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 case PAGE_KEEP:
1076 goto keep_locked;
1077 case PAGE_ACTIVATE:
1078 goto activate_locked;
1079 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001080 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001081 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001082 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001084
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 /*
1086 * A synchronous write - probably a ramdisk. Go
1087 * ahead and try to reclaim the page.
1088 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001089 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 goto keep;
1091 if (PageDirty(page) || PageWriteback(page))
1092 goto keep_locked;
1093 mapping = page_mapping(page);
1094 case PAGE_CLEAN:
1095 ; /* try to free the page below */
1096 }
1097 }
1098
1099 /*
1100 * If the page has buffers, try to free the buffer mappings
1101 * associated with this page. If we succeed we try to free
1102 * the page as well.
1103 *
1104 * We do this even if the page is PageDirty().
1105 * try_to_release_page() does not perform I/O, but it is
1106 * possible for a page to have PageDirty set, but it is actually
1107 * clean (all its buffers are clean). This happens if the
1108 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001109 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 * try_to_release_page() will discover that cleanness and will
1111 * drop the buffers and mark the page clean - it can be freed.
1112 *
1113 * Rarely, pages can have buffers and no ->mapping. These are
1114 * the pages which were not successfully invalidated in
1115 * truncate_complete_page(). We try to drop those buffers here
1116 * and if that worked, and the page is no longer mapped into
1117 * process address space (page_count == 1) it can be freed.
1118 * Otherwise, leave the page on the LRU so it is swappable.
1119 */
David Howells266cf652009-04-03 16:42:36 +01001120 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 if (!try_to_release_page(page, sc->gfp_mask))
1122 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001123 if (!mapping && page_count(page) == 1) {
1124 unlock_page(page);
1125 if (put_page_testzero(page))
1126 goto free_it;
1127 else {
1128 /*
1129 * rare race with speculative reference.
1130 * the speculative reference will free
1131 * this page shortly, so we may
1132 * increment nr_reclaimed here (and
1133 * leave it off the LRU).
1134 */
1135 nr_reclaimed++;
1136 continue;
1137 }
1138 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 }
1140
Johannes Weinera5289102014-04-03 14:47:51 -07001141 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001142 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Nick Piggina978d6f2008-10-18 20:26:58 -07001144 /*
1145 * At this point, we have no other references and there is
1146 * no way to pick any more up (removed from LRU, removed
1147 * from pagecache). Can use non-atomic bitops now (and
1148 * we obviously don't have to worry about waking up a process
1149 * waiting on the page lock, because there are no references.
1150 */
1151 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001152free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001153 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001154
1155 /*
1156 * Is there need to periodically free_page_list? It would
1157 * appear not as the counts should be low
1158 */
1159 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 continue;
1161
Nick Pigginb291f002008-10-18 20:26:44 -07001162cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001163 if (PageSwapCache(page))
1164 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001165 unlock_page(page);
1166 putback_lru_page(page);
1167 continue;
1168
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001170 /* Not a candidate for swapping, so reclaim swap space. */
1171 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001172 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001173 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 SetPageActive(page);
1175 pgactivate++;
1176keep_locked:
1177 unlock_page(page);
1178keep:
1179 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001180 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001182
Johannes Weiner747db952014-08-08 14:19:24 -07001183 mem_cgroup_uncharge_list(&free_pages);
Mel Gormanb745bc82014-06-04 16:10:22 -07001184 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001185
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001187 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001188
Mel Gorman8e950282013-07-03 15:02:02 -07001189 *ret_nr_dirty += nr_dirty;
1190 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001191 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001192 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001193 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001194 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195}
1196
Minchan Kim02c6de82012-10-08 16:31:55 -07001197unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1198 struct list_head *page_list)
1199{
1200 struct scan_control sc = {
1201 .gfp_mask = GFP_KERNEL,
1202 .priority = DEF_PRIORITY,
1203 .may_unmap = 1,
1204 };
Mel Gorman8e950282013-07-03 15:02:02 -07001205 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001206 struct page *page, *next;
1207 LIST_HEAD(clean_pages);
1208
1209 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001210 if (page_is_file_cache(page) && !PageDirty(page) &&
1211 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001212 ClearPageActive(page);
1213 list_move(&page->lru, &clean_pages);
1214 }
1215 }
1216
1217 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001218 TTU_UNMAP|TTU_IGNORE_ACCESS,
1219 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001220 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001221 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001222 return ret;
1223}
1224
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001225/*
1226 * Attempt to remove the specified page from its LRU. Only take this page
1227 * if it is of the appropriate PageActive status. Pages which are being
1228 * freed elsewhere are also ignored.
1229 *
1230 * page: page to consider
1231 * mode: one of the LRU isolation modes defined above
1232 *
1233 * returns 0 on success, -ve errno on failure.
1234 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001235int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001236{
1237 int ret = -EINVAL;
1238
1239 /* Only take pages on the LRU. */
1240 if (!PageLRU(page))
1241 return ret;
1242
Minchan Kime46a2872012-10-08 16:33:48 -07001243 /* Compaction should not handle unevictable pages but CMA can do so */
1244 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001245 return ret;
1246
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001247 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001248
Mel Gormanc8244932012-01-12 17:19:38 -08001249 /*
1250 * To minimise LRU disruption, the caller can indicate that it only
1251 * wants to isolate pages it will be able to operate on without
1252 * blocking - clean pages for the most part.
1253 *
1254 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1255 * is used by reclaim when it is cannot write to backing storage
1256 *
1257 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1258 * that it is possible to migrate without blocking
1259 */
1260 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1261 /* All the caller can do on PageWriteback is block */
1262 if (PageWriteback(page))
1263 return ret;
1264
1265 if (PageDirty(page)) {
1266 struct address_space *mapping;
1267
1268 /* ISOLATE_CLEAN means only clean pages */
1269 if (mode & ISOLATE_CLEAN)
1270 return ret;
1271
1272 /*
1273 * Only pages without mappings or that have a
1274 * ->migratepage callback are possible to migrate
1275 * without blocking
1276 */
1277 mapping = page_mapping(page);
1278 if (mapping && !mapping->a_ops->migratepage)
1279 return ret;
1280 }
1281 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001282
Minchan Kimf80c0672011-10-31 17:06:55 -07001283 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1284 return ret;
1285
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001286 if (likely(get_page_unless_zero(page))) {
1287 /*
1288 * Be careful not to clear PageLRU until after we're
1289 * sure the page is not being freed elsewhere -- the
1290 * page release code relies on it.
1291 */
1292 ClearPageLRU(page);
1293 ret = 0;
1294 }
1295
1296 return ret;
1297}
1298
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001299/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 * zone->lru_lock is heavily contended. Some of the functions that
1301 * shrink the lists perform better by taking out a batch of pages
1302 * and working on them outside the LRU lock.
1303 *
1304 * For pagecache intensive workloads, this function is the hottest
1305 * spot in the kernel (apart from copy_*_user functions).
1306 *
1307 * Appropriate locks must be held before calling this function.
1308 *
1309 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001310 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001312 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001313 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001314 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001315 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 *
1317 * returns how many pages were moved onto *@dst.
1318 */
Andrew Morton69e05942006-03-22 00:08:19 -08001319static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001320 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001321 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001322 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001324 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001325 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001326 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001328 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001329 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001330 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001331
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 page = lru_to_page(src);
1333 prefetchw_prev_lru_page(page, src, flags);
1334
Sasha Levin309381fea2014-01-23 15:52:54 -08001335 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001336
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001337 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001338 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001339 nr_pages = hpage_nr_pages(page);
1340 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001341 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001342 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001343 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001344
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001345 case -EBUSY:
1346 /* else it is being freed elsewhere */
1347 list_move(&page->lru, src);
1348 continue;
1349
1350 default:
1351 BUG();
1352 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354
Hugh Dickinsf6260122012-01-12 17:20:06 -08001355 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001356 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1357 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 return nr_taken;
1359}
1360
Nick Piggin62695a82008-10-18 20:26:09 -07001361/**
1362 * isolate_lru_page - tries to isolate a page from its LRU list
1363 * @page: page to isolate from its LRU list
1364 *
1365 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1366 * vmstat statistic corresponding to whatever LRU list the page was on.
1367 *
1368 * Returns 0 if the page was removed from an LRU list.
1369 * Returns -EBUSY if the page was not on an LRU list.
1370 *
1371 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001372 * the active list, it will have PageActive set. If it was found on
1373 * the unevictable list, it will have the PageUnevictable bit set. That flag
1374 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001375 *
1376 * The vmstat statistic corresponding to the list on which the page was
1377 * found will be decremented.
1378 *
1379 * Restrictions:
1380 * (1) Must be called with an elevated refcount on the page. This is a
1381 * fundamentnal difference from isolate_lru_pages (which is called
1382 * without a stable reference).
1383 * (2) the lru_lock must not be held.
1384 * (3) interrupts must be enabled.
1385 */
1386int isolate_lru_page(struct page *page)
1387{
1388 int ret = -EBUSY;
1389
Sasha Levin309381fea2014-01-23 15:52:54 -08001390 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001391
Nick Piggin62695a82008-10-18 20:26:09 -07001392 if (PageLRU(page)) {
1393 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001394 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001395
1396 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001397 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001398 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001399 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001400 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001401 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001402 del_page_from_lru_list(page, lruvec, lru);
1403 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001404 }
1405 spin_unlock_irq(&zone->lru_lock);
1406 }
1407 return ret;
1408}
1409
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001410/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001411 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1412 * then get resheduled. When there are massive number of tasks doing page
1413 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1414 * the LRU list will go small and be scanned faster than necessary, leading to
1415 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001416 */
1417static int too_many_isolated(struct zone *zone, int file,
1418 struct scan_control *sc)
1419{
1420 unsigned long inactive, isolated;
1421
1422 if (current_is_kswapd())
1423 return 0;
1424
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001425 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001426 return 0;
1427
1428 if (file) {
1429 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1430 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1431 } else {
1432 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1433 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1434 }
1435
Fengguang Wu3cf23842012-12-18 14:23:31 -08001436 /*
1437 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1438 * won't get blocked by normal direct-reclaimers, forming a circular
1439 * deadlock.
1440 */
1441 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1442 inactive >>= 3;
1443
Rik van Riel35cd7812009-09-21 17:01:38 -07001444 return isolated > inactive;
1445}
1446
Mel Gorman66635622010-08-09 17:19:30 -07001447static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001448putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001449{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001450 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1451 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001452 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001453
Mel Gorman66635622010-08-09 17:19:30 -07001454 /*
1455 * Put back any unfreeable pages.
1456 */
Mel Gorman66635622010-08-09 17:19:30 -07001457 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001458 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001459 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001460
Sasha Levin309381fea2014-01-23 15:52:54 -08001461 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001462 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001463 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001464 spin_unlock_irq(&zone->lru_lock);
1465 putback_lru_page(page);
1466 spin_lock_irq(&zone->lru_lock);
1467 continue;
1468 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001469
1470 lruvec = mem_cgroup_page_lruvec(page, zone);
1471
Linus Torvalds7a608572011-01-17 14:42:19 -08001472 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001473 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001474 add_page_to_lru_list(page, lruvec, lru);
1475
Mel Gorman66635622010-08-09 17:19:30 -07001476 if (is_active_lru(lru)) {
1477 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001478 int numpages = hpage_nr_pages(page);
1479 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001480 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001481 if (put_page_testzero(page)) {
1482 __ClearPageLRU(page);
1483 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001484 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001485
1486 if (unlikely(PageCompound(page))) {
1487 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001488 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001489 (*get_compound_page_dtor(page))(page);
1490 spin_lock_irq(&zone->lru_lock);
1491 } else
1492 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001493 }
1494 }
Mel Gorman66635622010-08-09 17:19:30 -07001495
Hugh Dickins3f797682012-01-12 17:20:07 -08001496 /*
1497 * To save our caller's stack, now use input list for pages to free.
1498 */
1499 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001500}
1501
1502/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001503 * If a kernel thread (such as nfsd for loop-back mounts) services
1504 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1505 * In that case we should only throttle if the backing device it is
1506 * writing to is congested. In other cases it is safe to throttle.
1507 */
1508static int current_may_throttle(void)
1509{
1510 return !(current->flags & PF_LESS_THROTTLE) ||
1511 current->backing_dev_info == NULL ||
1512 bdi_write_congested(current->backing_dev_info);
1513}
1514
1515/*
Andrew Morton1742f192006-03-22 00:08:21 -08001516 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1517 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 */
Mel Gorman66635622010-08-09 17:19:30 -07001519static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001520shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001521 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522{
1523 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001524 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001525 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001526 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001527 unsigned long nr_dirty = 0;
1528 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001529 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001530 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001531 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001532 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001533 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001534 struct zone *zone = lruvec_zone(lruvec);
1535 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001536
Rik van Riel35cd7812009-09-21 17:01:38 -07001537 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001538 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001539
1540 /* We are about to die and free our memory. Return now. */
1541 if (fatal_signal_pending(current))
1542 return SWAP_CLUSTER_MAX;
1543 }
1544
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001546
1547 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001548 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001549 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001550 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001551
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001554 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1555 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001556
1557 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1558 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1559
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001560 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001561 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001562 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001563 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001564 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001565 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001566 }
Hillf Dantond563c052012-03-21 16:34:02 -07001567 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001568
Hillf Dantond563c052012-03-21 16:34:02 -07001569 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001570 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001571
Minchan Kim02c6de82012-10-08 16:31:55 -07001572 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001573 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1574 &nr_writeback, &nr_immediate,
1575 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001576
Hugh Dickins3f797682012-01-12 17:20:07 -08001577 spin_lock_irq(&zone->lru_lock);
1578
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001579 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001580
Ying Han904249a2012-04-25 16:01:48 -07001581 if (global_reclaim(sc)) {
1582 if (current_is_kswapd())
1583 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1584 nr_reclaimed);
1585 else
1586 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1587 nr_reclaimed);
1588 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001589
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001590 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001591
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001592 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001593
1594 spin_unlock_irq(&zone->lru_lock);
1595
Johannes Weiner747db952014-08-08 14:19:24 -07001596 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001597 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001598
Mel Gorman92df3a72011-10-31 17:07:56 -07001599 /*
1600 * If reclaim is isolating dirty pages under writeback, it implies
1601 * that the long-lived page allocation rate is exceeding the page
1602 * laundering rate. Either the global limits are not being effective
1603 * at throttling processes due to the page distribution throughout
1604 * zones or there is heavy usage of a slow backing device. The
1605 * only option is to throttle from reclaim context which is not ideal
1606 * as there is no guarantee the dirtying process is throttled in the
1607 * same way balance_dirty_pages() manages.
1608 *
Mel Gorman8e950282013-07-03 15:02:02 -07001609 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1610 * of pages under pages flagged for immediate reclaim and stall if any
1611 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001612 */
Mel Gorman918fc712013-07-08 16:00:25 -07001613 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001614 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001615
Mel Gormand43006d2013-07-03 15:01:50 -07001616 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001617 * memcg will stall in page writeback so only consider forcibly
1618 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001619 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001620 if (global_reclaim(sc)) {
1621 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001622 * Tag a zone as congested if all the dirty pages scanned were
1623 * backed by a congested BDI and wait_iff_congested will stall.
1624 */
1625 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001626 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001627
1628 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001629 * If dirty pages are scanned that are not queued for IO, it
1630 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001631 * the zone ZONE_DIRTY and kswapd will start writing pages from
1632 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001633 */
1634 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001635 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001636
1637 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001638 * If kswapd scans pages marked marked for immediate
1639 * reclaim and under writeback (nr_immediate), it implies
1640 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001641 * they are written so also forcibly stall.
1642 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001643 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001644 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001645 }
Mel Gormand43006d2013-07-03 15:01:50 -07001646
Mel Gorman8e950282013-07-03 15:02:02 -07001647 /*
1648 * Stall direct reclaim for IO completions if underlying BDIs or zone
1649 * is congested. Allow kswapd to continue until it starts encountering
1650 * unqueued dirty pages or cycling through the LRU too quickly.
1651 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001652 if (!sc->hibernation_mode && !current_is_kswapd() &&
1653 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001654 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1655
Mel Gormane11da5b2010-10-26 14:21:40 -07001656 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1657 zone_idx(zone),
1658 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001659 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001660 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001661 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662}
1663
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001664/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 * This moves pages from the active list to the inactive list.
1666 *
1667 * We move them the other way if the page is referenced by one or more
1668 * processes, from rmap.
1669 *
1670 * If the pages are mostly unmapped, the processing is fast and it is
1671 * appropriate to hold zone->lru_lock across the whole operation. But if
1672 * the pages are mapped, the processing is slow (page_referenced()) so we
1673 * should drop zone->lru_lock around each page. It's impossible to balance
1674 * this, so instead we remove the pages from the LRU while processing them.
1675 * It is safe to rely on PG_active against the non-LRU pages in here because
1676 * nobody will play with that bit on a non-LRU page.
1677 *
1678 * The downside is that we have to touch page->_count against each page.
1679 * But we had to alter page->flags anyway.
1680 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001681
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001682static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001683 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001684 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001685 enum lru_list lru)
1686{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001687 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001688 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001689 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001690 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001691
Wu Fengguang3eb41402009-06-16 15:33:13 -07001692 while (!list_empty(list)) {
1693 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001694 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001695
Sasha Levin309381fea2014-01-23 15:52:54 -08001696 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001697 SetPageLRU(page);
1698
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001699 nr_pages = hpage_nr_pages(page);
1700 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001701 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001702 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001703
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001704 if (put_page_testzero(page)) {
1705 __ClearPageLRU(page);
1706 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001707 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001708
1709 if (unlikely(PageCompound(page))) {
1710 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001711 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001712 (*get_compound_page_dtor(page))(page);
1713 spin_lock_irq(&zone->lru_lock);
1714 } else
1715 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001716 }
1717 }
1718 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1719 if (!is_active_lru(lru))
1720 __count_vm_events(PGDEACTIVATE, pgmoved);
1721}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001722
Hugh Dickinsf6260122012-01-12 17:20:06 -08001723static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001724 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001725 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001726 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001728 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001729 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001730 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001732 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001733 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001735 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001736 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001737 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001738 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001739 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740
1741 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001742
1743 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001744 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001745 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001746 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001747
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001749
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001750 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1751 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001752 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001753 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001754
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001755 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001756
Hugh Dickinsf6260122012-01-12 17:20:06 -08001757 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001758 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001759 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 spin_unlock_irq(&zone->lru_lock);
1761
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 while (!list_empty(&l_hold)) {
1763 cond_resched();
1764 page = lru_to_page(&l_hold);
1765 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001766
Hugh Dickins39b5f292012-10-08 16:33:18 -07001767 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001768 putback_lru_page(page);
1769 continue;
1770 }
1771
Mel Gormancc715d92012-03-21 16:34:00 -07001772 if (unlikely(buffer_heads_over_limit)) {
1773 if (page_has_private(page) && trylock_page(page)) {
1774 if (page_has_private(page))
1775 try_to_release_page(page, 0);
1776 unlock_page(page);
1777 }
1778 }
1779
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001780 if (page_referenced(page, 0, sc->target_mem_cgroup,
1781 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001782 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001783 /*
1784 * Identify referenced, file-backed active pages and
1785 * give them one more trip around the active list. So
1786 * that executable code get better chances to stay in
1787 * memory under moderate memory pressure. Anon pages
1788 * are not likely to be evicted by use-once streaming
1789 * IO, plus JVM can create lots of anon VM_EXEC pages,
1790 * so we ignore them here.
1791 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001792 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001793 list_add(&page->lru, &l_active);
1794 continue;
1795 }
1796 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001797
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001798 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 list_add(&page->lru, &l_inactive);
1800 }
1801
Andrew Mortonb5557492009-01-06 14:40:13 -08001802 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001803 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001804 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001805 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001806 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001807 * Count referenced pages from currently used mappings as rotated,
1808 * even though only some of them are actually re-activated. This
1809 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001810 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001811 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001812 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001813
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001814 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1815 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001816 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001817 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001818
Johannes Weiner747db952014-08-08 14:19:24 -07001819 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001820 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821}
1822
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001823#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001824static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001825{
1826 unsigned long active, inactive;
1827
1828 active = zone_page_state(zone, NR_ACTIVE_ANON);
1829 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1830
1831 if (inactive * zone->inactive_ratio < active)
1832 return 1;
1833
1834 return 0;
1835}
1836
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001837/**
1838 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001839 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001840 *
1841 * Returns true if the zone does not have enough inactive anon pages,
1842 * meaning some active anon pages need to be deactivated.
1843 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001844static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001845{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001846 /*
1847 * If we don't have swap space, anonymous page deactivation
1848 * is pointless.
1849 */
1850 if (!total_swap_pages)
1851 return 0;
1852
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001853 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001854 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001855
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001856 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001857}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001858#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001859static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001860{
1861 return 0;
1862}
1863#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001864
Rik van Riel56e49d22009-06-16 15:32:28 -07001865/**
1866 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001867 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001868 *
1869 * When the system is doing streaming IO, memory pressure here
1870 * ensures that active file pages get deactivated, until more
1871 * than half of the file pages are on the inactive list.
1872 *
1873 * Once we get to that situation, protect the system's working
1874 * set from being evicted by disabling active file page aging.
1875 *
1876 * This uses a different ratio than the anonymous pages, because
1877 * the page cache uses a use-once replacement algorithm.
1878 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001879static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001880{
Johannes Weinere3790142013-02-22 16:35:19 -08001881 unsigned long inactive;
1882 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001883
Johannes Weinere3790142013-02-22 16:35:19 -08001884 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1885 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1886
1887 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001888}
1889
Hugh Dickins75b00af2012-05-29 15:07:09 -07001890static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001891{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001892 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001893 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001894 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001895 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001896}
1897
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001898static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001899 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001900{
Rik van Rielb39415b2009-12-14 17:59:48 -08001901 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001902 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001903 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001904 return 0;
1905 }
1906
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001907 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001908}
1909
Johannes Weiner9a265112013-02-22 16:32:17 -08001910enum scan_balance {
1911 SCAN_EQUAL,
1912 SCAN_FRACT,
1913 SCAN_ANON,
1914 SCAN_FILE,
1915};
1916
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001918 * Determine how aggressively the anon and file LRU lists should be
1919 * scanned. The relative value of each set of LRU lists is determined
1920 * by looking at the fraction of the pages scanned we did rotate back
1921 * onto the active list instead of evict.
1922 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001923 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1924 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001925 */
Johannes Weiner02695172014-08-06 16:06:17 -07001926static void get_scan_count(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001927 struct scan_control *sc, unsigned long *nr,
1928 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001929{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001930 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001931 u64 fraction[2];
1932 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001933 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001934 unsigned long anon_prio, file_prio;
1935 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001936 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001937 bool force_scan = false;
1938 unsigned long ap, fp;
1939 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001940 bool some_scanned;
1941 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001942
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001943 /*
1944 * If the zone or memcg is small, nr[l] can be 0. This
1945 * results in no scanning on this priority and a potential
1946 * priority drop. Global direct reclaim can go to the next
1947 * zone and tends to have no problems. Global kswapd is for
1948 * zone balancing and it needs to scan a minimum amount. When
1949 * reclaiming for a memcg, a priority drop can cause high
1950 * latencies, so it's better to scan a minimum amount there as
1951 * well.
1952 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08001953 if (current_is_kswapd()) {
1954 if (!zone_reclaimable(zone))
1955 force_scan = true;
1956 if (!mem_cgroup_lruvec_online(lruvec))
1957 force_scan = true;
1958 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001959 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001960 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001961
1962 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001963 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001964 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001965 goto out;
1966 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001967
Johannes Weiner10316b32013-02-22 16:32:14 -08001968 /*
1969 * Global reclaim will swap to prevent OOM even with no
1970 * swappiness, but memcg users want to use this knob to
1971 * disable swapping for individual groups completely when
1972 * using the memory controller's swap limit feature would be
1973 * too expensive.
1974 */
Johannes Weiner02695172014-08-06 16:06:17 -07001975 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001976 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001977 goto out;
1978 }
1979
1980 /*
1981 * Do not apply any pressure balancing cleverness when the
1982 * system is close to OOM, scan both anon and file equally
1983 * (unless the swappiness setting disagrees with swapping).
1984 */
Johannes Weiner02695172014-08-06 16:06:17 -07001985 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001986 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001987 goto out;
1988 }
1989
Johannes Weiner11d16c22013-02-22 16:32:15 -08001990 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001991 * Prevent the reclaimer from falling into the cache trap: as
1992 * cache pages start out inactive, every cache fault will tip
1993 * the scan balance towards the file LRU. And as the file LRU
1994 * shrinks, so does the window for rotation from references.
1995 * This means we have a runaway feedback loop where a tiny
1996 * thrashing file LRU becomes infinitely more attractive than
1997 * anon pages. Try to detect this based on file LRU size.
1998 */
1999 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002000 unsigned long zonefile;
2001 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07002002
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002003 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2004 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2005 zone_page_state(zone, NR_INACTIVE_FILE);
2006
2007 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07002008 scan_balance = SCAN_ANON;
2009 goto out;
2010 }
2011 }
2012
2013 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002014 * There is enough inactive page cache, do not reclaim
2015 * anything from the anonymous working set right now.
2016 */
2017 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002018 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002019 goto out;
2020 }
2021
Johannes Weiner9a265112013-02-22 16:32:17 -08002022 scan_balance = SCAN_FRACT;
2023
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002024 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002025 * With swappiness at 100, anonymous and file have the same priority.
2026 * This scanning priority is essentially the inverse of IO cost.
2027 */
Johannes Weiner02695172014-08-06 16:06:17 -07002028 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002029 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002030
2031 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002032 * OK, so we have swap space and a fair amount of page cache
2033 * pages. We use the recently rotated / recently scanned
2034 * ratios to determine how valuable each cache is.
2035 *
2036 * Because workloads change over time (and to avoid overflow)
2037 * we keep these statistics as a floating average, which ends
2038 * up weighing recent references more than old ones.
2039 *
2040 * anon in [0], file in [1]
2041 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002042
2043 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2044 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2045 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2046 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2047
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002048 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002049 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002050 reclaim_stat->recent_scanned[0] /= 2;
2051 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002052 }
2053
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002054 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002055 reclaim_stat->recent_scanned[1] /= 2;
2056 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002057 }
2058
2059 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002060 * The amount of pressure on anon vs file pages is inversely
2061 * proportional to the fraction of recently scanned pages on
2062 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002063 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002064 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002065 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002066
Satoru Moriyafe350042012-05-29 15:06:47 -07002067 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002068 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002069 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002070
Shaohua Li76a33fc2010-05-24 14:32:36 -07002071 fraction[0] = ap;
2072 fraction[1] = fp;
2073 denominator = ap + fp + 1;
2074out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002075 some_scanned = false;
2076 /* Only use force_scan on second pass. */
2077 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002078 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002079 for_each_evictable_lru(lru) {
2080 int file = is_file_lru(lru);
2081 unsigned long size;
2082 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002083
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002084 size = get_lru_size(lruvec, lru);
2085 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002086
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002087 if (!scan && pass && force_scan)
2088 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002089
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002090 switch (scan_balance) {
2091 case SCAN_EQUAL:
2092 /* Scan lists relative to size */
2093 break;
2094 case SCAN_FRACT:
2095 /*
2096 * Scan types proportional to swappiness and
2097 * their relative recent reclaim efficiency.
2098 */
2099 scan = div64_u64(scan * fraction[file],
2100 denominator);
2101 break;
2102 case SCAN_FILE:
2103 case SCAN_ANON:
2104 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002105 if ((scan_balance == SCAN_FILE) != file) {
2106 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002107 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002108 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002109 break;
2110 default:
2111 /* Look ma, no brain */
2112 BUG();
2113 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002114
2115 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002116 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002117
Johannes Weiner9a265112013-02-22 16:32:17 -08002118 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002119 * Skip the second pass and don't force_scan,
2120 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002121 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002122 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002123 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002124 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002125}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002126
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002127/*
2128 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2129 */
Johannes Weiner02695172014-08-06 16:06:17 -07002130static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002131 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002132{
2133 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002134 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002135 unsigned long nr_to_scan;
2136 enum lru_list lru;
2137 unsigned long nr_reclaimed = 0;
2138 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2139 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002140 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002141
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002142 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002143
Mel Gormane82e0562013-07-03 15:01:44 -07002144 /* Record the original scan target for proportional adjustments later */
2145 memcpy(targets, nr, sizeof(nr));
2146
Mel Gorman1a501902014-06-04 16:10:49 -07002147 /*
2148 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2149 * event that can occur when there is little memory pressure e.g.
2150 * multiple streaming readers/writers. Hence, we do not abort scanning
2151 * when the requested number of pages are reclaimed when scanning at
2152 * DEF_PRIORITY on the assumption that the fact we are direct
2153 * reclaiming implies that kswapd is not keeping up and it is best to
2154 * do a batch of work at once. For memcg reclaim one check is made to
2155 * abort proportional reclaim if either the file or anon lru has already
2156 * dropped to zero at the first pass.
2157 */
2158 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2159 sc->priority == DEF_PRIORITY);
2160
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002161 blk_start_plug(&plug);
2162 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2163 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002164 unsigned long nr_anon, nr_file, percentage;
2165 unsigned long nr_scanned;
2166
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002167 for_each_evictable_lru(lru) {
2168 if (nr[lru]) {
2169 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2170 nr[lru] -= nr_to_scan;
2171
2172 nr_reclaimed += shrink_list(lru, nr_to_scan,
2173 lruvec, sc);
2174 }
2175 }
Mel Gormane82e0562013-07-03 15:01:44 -07002176
2177 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2178 continue;
2179
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002180 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002181 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002182 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002183 * proportionally what was requested by get_scan_count(). We
2184 * stop reclaiming one LRU and reduce the amount scanning
2185 * proportional to the original scan target.
2186 */
2187 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2188 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2189
Mel Gorman1a501902014-06-04 16:10:49 -07002190 /*
2191 * It's just vindictive to attack the larger once the smaller
2192 * has gone to zero. And given the way we stop scanning the
2193 * smaller below, this makes sure that we only make one nudge
2194 * towards proportionality once we've got nr_to_reclaim.
2195 */
2196 if (!nr_file || !nr_anon)
2197 break;
2198
Mel Gormane82e0562013-07-03 15:01:44 -07002199 if (nr_file > nr_anon) {
2200 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2201 targets[LRU_ACTIVE_ANON] + 1;
2202 lru = LRU_BASE;
2203 percentage = nr_anon * 100 / scan_target;
2204 } else {
2205 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2206 targets[LRU_ACTIVE_FILE] + 1;
2207 lru = LRU_FILE;
2208 percentage = nr_file * 100 / scan_target;
2209 }
2210
2211 /* Stop scanning the smaller of the LRU */
2212 nr[lru] = 0;
2213 nr[lru + LRU_ACTIVE] = 0;
2214
2215 /*
2216 * Recalculate the other LRU scan count based on its original
2217 * scan target and the percentage scanning already complete
2218 */
2219 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2220 nr_scanned = targets[lru] - nr[lru];
2221 nr[lru] = targets[lru] * (100 - percentage) / 100;
2222 nr[lru] -= min(nr[lru], nr_scanned);
2223
2224 lru += LRU_ACTIVE;
2225 nr_scanned = targets[lru] - nr[lru];
2226 nr[lru] = targets[lru] * (100 - percentage) / 100;
2227 nr[lru] -= min(nr[lru], nr_scanned);
2228
2229 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002230 }
2231 blk_finish_plug(&plug);
2232 sc->nr_reclaimed += nr_reclaimed;
2233
2234 /*
2235 * Even if we did not try to evict anon pages at all, we want to
2236 * rebalance the anon lru active/inactive ratio.
2237 */
2238 if (inactive_anon_is_low(lruvec))
2239 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2240 sc, LRU_ACTIVE_ANON);
2241
2242 throttle_vm_writeout(sc->gfp_mask);
2243}
2244
Mel Gorman23b9da52012-05-29 15:06:20 -07002245/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002246static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002247{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002248 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002249 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002250 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002251 return true;
2252
2253 return false;
2254}
2255
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002256/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002257 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2258 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2259 * true if more pages should be reclaimed such that when the page allocator
2260 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2261 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002262 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002263static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002264 unsigned long nr_reclaimed,
2265 unsigned long nr_scanned,
2266 struct scan_control *sc)
2267{
2268 unsigned long pages_for_compaction;
2269 unsigned long inactive_lru_pages;
2270
2271 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002272 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002273 return false;
2274
Mel Gorman28765922011-02-25 14:44:20 -08002275 /* Consider stopping depending on scan and reclaim activity */
2276 if (sc->gfp_mask & __GFP_REPEAT) {
2277 /*
2278 * For __GFP_REPEAT allocations, stop reclaiming if the
2279 * full LRU list has been scanned and we are still failing
2280 * to reclaim pages. This full LRU scan is potentially
2281 * expensive but a __GFP_REPEAT caller really wants to succeed
2282 */
2283 if (!nr_reclaimed && !nr_scanned)
2284 return false;
2285 } else {
2286 /*
2287 * For non-__GFP_REPEAT allocations which can presumably
2288 * fail without consequence, stop if we failed to reclaim
2289 * any pages from the last SWAP_CLUSTER_MAX number of
2290 * pages that were scanned. This will return to the
2291 * caller faster at the risk reclaim/compaction and
2292 * the resulting allocation attempt fails
2293 */
2294 if (!nr_reclaimed)
2295 return false;
2296 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002297
2298 /*
2299 * If we have not reclaimed enough pages for compaction and the
2300 * inactive lists are large enough, continue reclaiming
2301 */
2302 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002303 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002304 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002305 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002306 if (sc->nr_reclaimed < pages_for_compaction &&
2307 inactive_lru_pages > pages_for_compaction)
2308 return true;
2309
2310 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002311 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002312 case COMPACT_PARTIAL:
2313 case COMPACT_CONTINUE:
2314 return false;
2315 default:
2316 return true;
2317 }
2318}
2319
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002320static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2321 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002322{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002323 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002324 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002325 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002326
Johannes Weiner56600482012-01-12 17:17:59 -08002327 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002328 struct mem_cgroup *root = sc->target_mem_cgroup;
2329 struct mem_cgroup_reclaim_cookie reclaim = {
2330 .zone = zone,
2331 .priority = sc->priority,
2332 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002333 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002334 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002335
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002336 nr_reclaimed = sc->nr_reclaimed;
2337 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002338
Andrew Morton694fbc02013-09-24 15:27:37 -07002339 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2340 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002341 unsigned long lru_pages;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002342 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002343 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002344 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002345
Johannes Weiner241994e2015-02-11 15:26:06 -08002346 if (mem_cgroup_low(root, memcg)) {
2347 if (!sc->may_thrash)
2348 continue;
2349 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2350 }
2351
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002352 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002353 swappiness = mem_cgroup_swappiness(memcg);
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002354 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002355
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002356 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2357 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002358
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002359 if (memcg && is_classzone)
2360 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2361 memcg, sc->nr_scanned - scanned,
2362 lru_pages);
2363
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002364 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002365 * Direct reclaim and kswapd have to scan all memory
2366 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002367 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002368 *
2369 * Limit reclaim, on the other hand, only cares about
2370 * nr_to_reclaim pages to be reclaimed and it will
2371 * retry with decreasing priority if one round over the
2372 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002373 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002374 if (!global_reclaim(sc) &&
2375 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002376 mem_cgroup_iter_break(root, memcg);
2377 break;
2378 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002379 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002380
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002381 /*
2382 * Shrink the slab caches in the same proportion that
2383 * the eligible LRU pages were scanned.
2384 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002385 if (global_reclaim(sc) && is_classzone)
2386 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2387 sc->nr_scanned - nr_scanned,
2388 zone_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002389
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002390 if (reclaim_state) {
2391 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2392 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002393 }
2394
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002395 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2396 sc->nr_scanned - nr_scanned,
2397 sc->nr_reclaimed - nr_reclaimed);
2398
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002399 if (sc->nr_reclaimed - nr_reclaimed)
2400 reclaimable = true;
2401
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002402 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2403 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002404
2405 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002406}
2407
Vlastimil Babka53853e22014-10-09 15:27:02 -07002408/*
2409 * Returns true if compaction should go ahead for a high-order request, or
2410 * the high-order allocation would succeed without compaction.
2411 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002412static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002413{
2414 unsigned long balance_gap, watermark;
2415 bool watermark_ok;
2416
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002417 /*
2418 * Compaction takes time to run and there are potentially other
2419 * callers using the pages just freed. Continue reclaiming until
2420 * there is a buffer of free pages available to give compaction
2421 * a reasonable chance of completing and allocating the page
2422 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002423 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2424 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002425 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002426 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2427
2428 /*
2429 * If compaction is deferred, reclaim up to a point where
2430 * compaction will have a chance of success when re-enabled
2431 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002432 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002433 return watermark_ok;
2434
Vlastimil Babka53853e22014-10-09 15:27:02 -07002435 /*
2436 * If compaction is not ready to start and allocation is not likely
2437 * to succeed without it, then keep reclaiming.
2438 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002439 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002440 return false;
2441
2442 return watermark_ok;
2443}
2444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445/*
2446 * This is the direct reclaim path, for page-allocating processes. We only
2447 * try to reclaim pages from zones which will satisfy the caller's allocation
2448 * request.
2449 *
Mel Gorman41858962009-06-16 15:32:12 -07002450 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2451 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2453 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002454 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2455 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2456 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 * If a zone is deemed to be full of pinned pages then just give it a light
2459 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002460 *
2461 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002463static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464{
Mel Gormandd1a2392008-04-28 02:12:17 -07002465 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002466 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002467 unsigned long nr_soft_reclaimed;
2468 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002469 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002470 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002471 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002472
Mel Gormancc715d92012-03-21 16:34:00 -07002473 /*
2474 * If the number of buffer_heads in the machine exceeds the maximum
2475 * allowed level, force direct reclaim to scan the highmem zone as
2476 * highmem pages could be pinning lowmem pages storing buffer_heads
2477 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002478 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002479 if (buffer_heads_over_limit)
2480 sc->gfp_mask |= __GFP_HIGHMEM;
2481
Mel Gormand4debc62010-08-09 17:19:29 -07002482 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002483 requested_highidx, sc->nodemask) {
2484 enum zone_type classzone_idx;
2485
Con Kolivasf3fe6512006-01-06 00:11:15 -08002486 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002488
2489 classzone_idx = requested_highidx;
2490 while (!populated_zone(zone->zone_pgdat->node_zones +
2491 classzone_idx))
2492 classzone_idx--;
2493
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002494 /*
2495 * Take care memory controller reclaiming has small influence
2496 * to global LRU.
2497 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002498 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002499 if (!cpuset_zone_allowed(zone,
2500 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002501 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002502
Lisa Du6e543d52013-09-11 14:22:36 -07002503 if (sc->priority != DEF_PRIORITY &&
2504 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002505 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002506
2507 /*
2508 * If we already have plenty of memory free for
2509 * compaction in this zone, don't free any more.
2510 * Even though compaction is invoked for any
2511 * non-zero order, only frequent costly order
2512 * reclamation is disruptive enough to become a
2513 * noticeable problem, like transparent huge
2514 * page allocations.
2515 */
2516 if (IS_ENABLED(CONFIG_COMPACTION) &&
2517 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2518 zonelist_zone_idx(z) <= requested_highidx &&
2519 compaction_ready(zone, sc->order)) {
2520 sc->compaction_ready = true;
2521 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002522 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002523
Andrew Morton0608f432013-09-24 15:27:41 -07002524 /*
2525 * This steals pages from memory cgroups over softlimit
2526 * and returns the number of reclaimed pages and
2527 * scanned pages. This works for global memory pressure
2528 * and balancing, not for a memcg's limit.
2529 */
2530 nr_soft_scanned = 0;
2531 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2532 sc->order, sc->gfp_mask,
2533 &nr_soft_scanned);
2534 sc->nr_reclaimed += nr_soft_reclaimed;
2535 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002536 if (nr_soft_reclaimed)
2537 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002538 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002539 }
Nick Piggin408d8542006-09-25 23:31:27 -07002540
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002541 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002542 reclaimable = true;
2543
2544 if (global_reclaim(sc) &&
2545 !reclaimable && zone_reclaimable(zone))
2546 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 }
Mel Gormane0c23272011-10-31 17:09:33 -07002548
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002549 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002550 * Restore to original mask to avoid the impact on the caller if we
2551 * promoted it to __GFP_HIGHMEM.
2552 */
2553 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002554
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002555 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002557
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558/*
2559 * This is the main entry point to direct page reclaim.
2560 *
2561 * If a full scan of the inactive list fails to free enough memory then we
2562 * are "out of memory" and something needs to be killed.
2563 *
2564 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2565 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002566 * caller can't do much about. We kick the writeback threads and take explicit
2567 * naps in the hope that some of these pages can be written. But if the
2568 * allocating task holds filesystem locks which prevent writeout this might not
2569 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002570 *
2571 * returns: 0, if no pages reclaimed
2572 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 */
Mel Gormandac1d272008-04-28 02:12:12 -07002574static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002575 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576{
Johannes Weiner241994e2015-02-11 15:26:06 -08002577 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002578 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002579 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002580 bool zones_reclaimable;
Johannes Weiner241994e2015-02-11 15:26:06 -08002581retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002582 delayacct_freepages_start();
2583
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002584 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002585 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002587 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002588 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2589 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002590 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002591 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002592
Balbir Singh66e17072008-02-07 00:13:56 -08002593 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002594 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002595 break;
2596
2597 if (sc->compaction_ready)
2598 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
2600 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002601 * If we're getting trouble reclaiming, start doing
2602 * writepage even in laptop mode.
2603 */
2604 if (sc->priority < DEF_PRIORITY - 2)
2605 sc->may_writepage = 1;
2606
2607 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 * Try to write back as many pages as we just scanned. This
2609 * tends to cause slow streaming writers to write data to the
2610 * disk smoothly, at the dirtying rate, which is nice. But
2611 * that's undesirable in laptop mode, where we *want* lumpy
2612 * writeout. So in laptop mode, write out the whole world.
2613 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002614 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2615 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002616 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2617 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002618 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002620 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002621
Keika Kobayashi873b4772008-07-25 01:48:52 -07002622 delayacct_freepages_end();
2623
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002624 if (sc->nr_reclaimed)
2625 return sc->nr_reclaimed;
2626
Mel Gorman0cee34f2012-01-12 17:19:49 -08002627 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002628 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002629 return 1;
2630
Johannes Weiner241994e2015-02-11 15:26:06 -08002631 /* Untapped cgroup reserves? Don't OOM, retry. */
2632 if (!sc->may_thrash) {
2633 sc->priority = initial_priority;
2634 sc->may_thrash = 1;
2635 goto retry;
2636 }
2637
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002638 /* Any of the zones still reclaimable? Don't OOM. */
2639 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002640 return 1;
2641
2642 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643}
2644
Mel Gorman55150612012-07-31 16:44:35 -07002645static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2646{
2647 struct zone *zone;
2648 unsigned long pfmemalloc_reserve = 0;
2649 unsigned long free_pages = 0;
2650 int i;
2651 bool wmark_ok;
2652
2653 for (i = 0; i <= ZONE_NORMAL; i++) {
2654 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002655 if (!populated_zone(zone))
2656 continue;
2657
Mel Gorman55150612012-07-31 16:44:35 -07002658 pfmemalloc_reserve += min_wmark_pages(zone);
2659 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2660 }
2661
Mel Gorman675becc2014-06-04 16:07:35 -07002662 /* If there are no reserves (unexpected config) then do not throttle */
2663 if (!pfmemalloc_reserve)
2664 return true;
2665
Mel Gorman55150612012-07-31 16:44:35 -07002666 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2667
2668 /* kswapd must be awake if processes are being throttled */
2669 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2670 pgdat->classzone_idx = min(pgdat->classzone_idx,
2671 (enum zone_type)ZONE_NORMAL);
2672 wake_up_interruptible(&pgdat->kswapd_wait);
2673 }
2674
2675 return wmark_ok;
2676}
2677
2678/*
2679 * Throttle direct reclaimers if backing storage is backed by the network
2680 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2681 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002682 * when the low watermark is reached.
2683 *
2684 * Returns true if a fatal signal was delivered during throttling. If this
2685 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002686 */
Mel Gorman50694c22012-11-26 16:29:48 -08002687static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002688 nodemask_t *nodemask)
2689{
Mel Gorman675becc2014-06-04 16:07:35 -07002690 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002691 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002692 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002693
2694 /*
2695 * Kernel threads should not be throttled as they may be indirectly
2696 * responsible for cleaning pages necessary for reclaim to make forward
2697 * progress. kjournald for example may enter direct reclaim while
2698 * committing a transaction where throttling it could forcing other
2699 * processes to block on log_wait_commit().
2700 */
2701 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002702 goto out;
2703
2704 /*
2705 * If a fatal signal is pending, this process should not throttle.
2706 * It should return quickly so it can exit and free its memory
2707 */
2708 if (fatal_signal_pending(current))
2709 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002710
Mel Gorman675becc2014-06-04 16:07:35 -07002711 /*
2712 * Check if the pfmemalloc reserves are ok by finding the first node
2713 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2714 * GFP_KERNEL will be required for allocating network buffers when
2715 * swapping over the network so ZONE_HIGHMEM is unusable.
2716 *
2717 * Throttling is based on the first usable node and throttled processes
2718 * wait on a queue until kswapd makes progress and wakes them. There
2719 * is an affinity then between processes waking up and where reclaim
2720 * progress has been made assuming the process wakes on the same node.
2721 * More importantly, processes running on remote nodes will not compete
2722 * for remote pfmemalloc reserves and processes on different nodes
2723 * should make reasonable progress.
2724 */
2725 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002726 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002727 if (zone_idx(zone) > ZONE_NORMAL)
2728 continue;
2729
2730 /* Throttle based on the first usable node */
2731 pgdat = zone->zone_pgdat;
2732 if (pfmemalloc_watermark_ok(pgdat))
2733 goto out;
2734 break;
2735 }
2736
2737 /* If no zone was usable by the allocation flags then do not throttle */
2738 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002739 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002740
Mel Gorman68243e72012-07-31 16:44:39 -07002741 /* Account for the throttling */
2742 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2743
Mel Gorman55150612012-07-31 16:44:35 -07002744 /*
2745 * If the caller cannot enter the filesystem, it's possible that it
2746 * is due to the caller holding an FS lock or performing a journal
2747 * transaction in the case of a filesystem like ext[3|4]. In this case,
2748 * it is not safe to block on pfmemalloc_wait as kswapd could be
2749 * blocked waiting on the same lock. Instead, throttle for up to a
2750 * second before continuing.
2751 */
2752 if (!(gfp_mask & __GFP_FS)) {
2753 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2754 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002755
2756 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002757 }
2758
2759 /* Throttle until kswapd wakes the process */
2760 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2761 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002762
2763check_pending:
2764 if (fatal_signal_pending(current))
2765 return true;
2766
2767out:
2768 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002769}
2770
Mel Gormandac1d272008-04-28 02:12:12 -07002771unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002772 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002773{
Mel Gorman33906bc2010-08-09 17:19:16 -07002774 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002775 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002776 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002777 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2778 .order = order,
2779 .nodemask = nodemask,
2780 .priority = DEF_PRIORITY,
2781 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002782 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002783 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002784 };
2785
Mel Gorman55150612012-07-31 16:44:35 -07002786 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002787 * Do not enter reclaim if fatal signal was delivered while throttled.
2788 * 1 is returned so that the page allocator does not OOM kill at this
2789 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002790 */
Mel Gorman50694c22012-11-26 16:29:48 -08002791 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002792 return 1;
2793
Mel Gorman33906bc2010-08-09 17:19:16 -07002794 trace_mm_vmscan_direct_reclaim_begin(order,
2795 sc.may_writepage,
2796 gfp_mask);
2797
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002798 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002799
2800 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2801
2802 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002803}
2804
Andrew Mortonc255a452012-07-31 16:43:02 -07002805#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002806
Johannes Weiner72835c82012-01-12 17:18:32 -08002807unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002808 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002809 struct zone *zone,
2810 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002811{
2812 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002813 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002814 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002815 .may_writepage = !laptop_mode,
2816 .may_unmap = 1,
2817 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002818 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002819 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002820 int swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002821 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002822
Balbir Singh4e416952009-09-23 15:56:39 -07002823 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2824 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002825
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002826 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002827 sc.may_writepage,
2828 sc.gfp_mask);
2829
Balbir Singh4e416952009-09-23 15:56:39 -07002830 /*
2831 * NOTE: Although we can get the priority field, using it
2832 * here is not a good idea, since it limits the pages we can scan.
2833 * if we don't reclaim here, the shrink_zone from balance_pgdat
2834 * will pick up pages from other mem cgroup's as well. We hack
2835 * the priority and make it zero.
2836 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002837 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002838
2839 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2840
Ying Han0ae5e892011-05-26 16:25:25 -07002841 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002842 return sc.nr_reclaimed;
2843}
2844
Johannes Weiner72835c82012-01-12 17:18:32 -08002845unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002846 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002847 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002848 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002849{
Balbir Singh4e416952009-09-23 15:56:39 -07002850 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002851 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002852 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002853 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002854 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002855 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2856 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2857 .target_mem_cgroup = memcg,
2858 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002859 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002860 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002861 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002862 };
Balbir Singh66e17072008-02-07 00:13:56 -08002863
Ying Han889976d2011-05-26 16:25:33 -07002864 /*
2865 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2866 * take care of from where we get pages. So the node where we start the
2867 * scan does not need to be the current node.
2868 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002869 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002870
2871 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002872
2873 trace_mm_vmscan_memcg_reclaim_begin(0,
2874 sc.may_writepage,
2875 sc.gfp_mask);
2876
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002877 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002878
2879 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2880
2881 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002882}
2883#endif
2884
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002885static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002886{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002887 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002888
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002889 if (!total_swap_pages)
2890 return;
2891
2892 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2893 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002894 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002895
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002896 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002897 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002898 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002899
2900 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2901 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002902}
2903
Johannes Weiner60cefed2012-11-29 13:54:23 -08002904static bool zone_balanced(struct zone *zone, int order,
2905 unsigned long balance_gap, int classzone_idx)
2906{
2907 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2908 balance_gap, classzone_idx, 0))
2909 return false;
2910
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002911 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2912 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002913 return false;
2914
2915 return true;
2916}
2917
Mel Gorman1741c872011-01-13 15:46:21 -08002918/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002919 * pgdat_balanced() is used when checking if a node is balanced.
2920 *
2921 * For order-0, all zones must be balanced!
2922 *
2923 * For high-order allocations only zones that meet watermarks and are in a
2924 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2925 * total of balanced pages must be at least 25% of the zones allowed by
2926 * classzone_idx for the node to be considered balanced. Forcing all zones to
2927 * be balanced for high orders can cause excessive reclaim when there are
2928 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002929 * The choice of 25% is due to
2930 * o a 16M DMA zone that is balanced will not balance a zone on any
2931 * reasonable sized machine
2932 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002933 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002934 * would need to be at least 256M for it to be balance a whole node.
2935 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2936 * to balance a node on its own. These seemed like reasonable ratios.
2937 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002938static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002939{
Jiang Liub40da042013-02-22 16:33:52 -08002940 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002941 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002942 int i;
2943
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002944 /* Check the watermark levels */
2945 for (i = 0; i <= classzone_idx; i++) {
2946 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002947
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002948 if (!populated_zone(zone))
2949 continue;
2950
Jiang Liub40da042013-02-22 16:33:52 -08002951 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002952
2953 /*
2954 * A special case here:
2955 *
2956 * balance_pgdat() skips over all_unreclaimable after
2957 * DEF_PRIORITY. Effectively, it considers them balanced so
2958 * they must be considered balanced here as well!
2959 */
Lisa Du6e543d52013-09-11 14:22:36 -07002960 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002961 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002962 continue;
2963 }
2964
2965 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002966 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002967 else if (!order)
2968 return false;
2969 }
2970
2971 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002972 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002973 else
2974 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002975}
2976
Mel Gorman55150612012-07-31 16:44:35 -07002977/*
2978 * Prepare kswapd for sleeping. This verifies that there are no processes
2979 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2980 *
2981 * Returns true if kswapd is ready to sleep
2982 */
2983static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002984 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002985{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002986 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2987 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002988 return false;
2989
2990 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002991 * The throttled processes are normally woken up in balance_pgdat() as
2992 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2993 * race between when kswapd checks the watermarks and a process gets
2994 * throttled. There is also a potential race if processes get
2995 * throttled, kswapd wakes, a large process exits thereby balancing the
2996 * zones, which causes kswapd to exit balance_pgdat() before reaching
2997 * the wake up checks. If kswapd is going to sleep, no process should
2998 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2999 * the wake up is premature, processes will wake kswapd and get
3000 * throttled again. The difference from wake ups in balance_pgdat() is
3001 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003002 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003003 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3004 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003005
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003006 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003007}
3008
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009/*
Mel Gorman75485362013-07-03 15:01:42 -07003010 * kswapd shrinks the zone by the number of pages required to reach
3011 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003012 *
3013 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003014 * reclaim or if the lack of progress was due to pages under writeback.
3015 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003016 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003017static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07003018 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07003019 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07003020 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07003021{
Mel Gorman7c954f62013-07-03 15:01:54 -07003022 int testorder = sc->order;
3023 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07003024 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07003025
3026 /* Reclaim above the high watermark. */
3027 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07003028
3029 /*
3030 * Kswapd reclaims only single pages with compaction enabled. Trying
3031 * too hard to reclaim until contiguous free pages have become
3032 * available can hurt performance by evicting too much useful data
3033 * from memory. Do not reclaim more than needed for compaction.
3034 */
3035 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08003036 compaction_suitable(zone, sc->order, 0, classzone_idx)
3037 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07003038 testorder = 0;
3039
3040 /*
3041 * We put equal pressure on every zone, unless one zone has way too
3042 * many pages free already. The "too many pages" is defined as the
3043 * high wmark plus a "gap" where the gap is either the low
3044 * watermark or 1% of the zone, whichever is smaller.
3045 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07003046 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3047 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003048
3049 /*
3050 * If there is no low memory pressure or the zone is balanced then no
3051 * reclaim is necessary
3052 */
3053 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3054 if (!lowmem_pressure && zone_balanced(zone, testorder,
3055 balance_gap, classzone_idx))
3056 return true;
3057
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003058 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003059
Mel Gorman2ab44f42013-07-03 15:01:47 -07003060 /* Account for the number of pages attempted to reclaim */
3061 *nr_attempted += sc->nr_to_reclaim;
3062
Johannes Weiner57054652014-10-09 15:28:17 -07003063 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003064
Mel Gorman7c954f62013-07-03 15:01:54 -07003065 /*
3066 * If a zone reaches its high watermark, consider it to be no longer
3067 * congested. It's possible there are dirty pages backed by congested
3068 * BDIs but as pressure is relieved, speculatively avoid congestion
3069 * waits.
3070 */
Lisa Du6e543d52013-09-11 14:22:36 -07003071 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003072 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003073 clear_bit(ZONE_CONGESTED, &zone->flags);
3074 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003075 }
3076
Mel Gormanb8e83b92013-07-03 15:01:45 -07003077 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003078}
3079
3080/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003082 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003084 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 *
3086 * There is special handling here for zones which are full of pinned pages.
3087 * This can happen if the pages are all mlocked, or if they are all used by
3088 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3089 * What we do is to detect the case where all pages in the zone have been
3090 * scanned twice and there has been zero successful reclaim. Mark the zone as
3091 * dead and from now on, only perform a short scan. Basically we're polling
3092 * the zone for when the problem goes away.
3093 *
3094 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003095 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3096 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3097 * lower zones regardless of the number of free pages in the lower zones. This
3098 * interoperates with the page allocator fallback scheme to ensure that aging
3099 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100 */
Mel Gorman99504742011-01-13 15:46:20 -08003101static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003102 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003105 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003106 unsigned long nr_soft_reclaimed;
3107 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003108 struct scan_control sc = {
3109 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003110 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003111 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003112 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003113 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003114 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003115 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003116 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003118 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003119 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003120 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003121 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003122
3123 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003125 /*
3126 * Scan in the highmem->dma direction for the highest
3127 * zone which needs scanning
3128 */
3129 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3130 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003132 if (!populated_zone(zone))
3133 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134
Lisa Du6e543d52013-09-11 14:22:36 -07003135 if (sc.priority != DEF_PRIORITY &&
3136 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003137 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
Rik van Riel556adec2008-10-18 20:26:34 -07003139 /*
3140 * Do some background aging of the anon list, to give
3141 * pages a chance to be referenced before reclaiming.
3142 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003143 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003144
Mel Gormancc715d92012-03-21 16:34:00 -07003145 /*
3146 * If the number of buffer_heads in the machine
3147 * exceeds the maximum allowed level and this node
3148 * has a highmem zone, force kswapd to reclaim from
3149 * it to relieve lowmem pressure.
3150 */
3151 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3152 end_zone = i;
3153 break;
3154 }
3155
Johannes Weiner60cefed2012-11-29 13:54:23 -08003156 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003157 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003158 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003159 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003160 /*
3161 * If balanced, clear the dirty and congested
3162 * flags
3163 */
Johannes Weiner57054652014-10-09 15:28:17 -07003164 clear_bit(ZONE_CONGESTED, &zone->flags);
3165 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003168
Mel Gormanb8e83b92013-07-03 15:01:45 -07003169 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003170 goto out;
3171
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 for (i = 0; i <= end_zone; i++) {
3173 struct zone *zone = pgdat->node_zones + i;
3174
Mel Gorman2ab44f42013-07-03 15:01:47 -07003175 if (!populated_zone(zone))
3176 continue;
3177
Mel Gorman2ab44f42013-07-03 15:01:47 -07003178 /*
3179 * If any zone is currently balanced then kswapd will
3180 * not call compaction as it is expected that the
3181 * necessary pages are already available.
3182 */
3183 if (pgdat_needs_compaction &&
3184 zone_watermark_ok(zone, order,
3185 low_wmark_pages(zone),
3186 *classzone_idx, 0))
3187 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188 }
3189
3190 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003191 * If we're getting trouble reclaiming, start doing writepage
3192 * even in laptop mode.
3193 */
3194 if (sc.priority < DEF_PRIORITY - 2)
3195 sc.may_writepage = 1;
3196
3197 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 * Now scan the zone in the dma->highmem direction, stopping
3199 * at the last zone which needs scanning.
3200 *
3201 * We do this because the page allocator works in the opposite
3202 * direction. This prevents the page allocator from allocating
3203 * pages behind kswapd's direction of progress, which would
3204 * cause too much scanning of the lower zones.
3205 */
3206 for (i = 0; i <= end_zone; i++) {
3207 struct zone *zone = pgdat->node_zones + i;
3208
Con Kolivasf3fe6512006-01-06 00:11:15 -08003209 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 continue;
3211
Lisa Du6e543d52013-09-11 14:22:36 -07003212 if (sc.priority != DEF_PRIORITY &&
3213 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 continue;
3215
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003217
Andrew Morton0608f432013-09-24 15:27:41 -07003218 nr_soft_scanned = 0;
3219 /*
3220 * Call soft limit reclaim before calling shrink_zone.
3221 */
3222 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3223 order, sc.gfp_mask,
3224 &nr_soft_scanned);
3225 sc.nr_reclaimed += nr_soft_reclaimed;
3226
Rik van Riel32a43302007-10-16 01:24:50 -07003227 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003228 * There should be no need to raise the scanning
3229 * priority if enough pages are already being scanned
3230 * that that high watermark would be met at 100%
3231 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003232 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003233 if (kswapd_shrink_zone(zone, end_zone,
3234 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003235 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 }
Mel Gorman55150612012-07-31 16:44:35 -07003237
3238 /*
3239 * If the low watermark is met there is no need for processes
3240 * to be throttled on pfmemalloc_wait as they should not be
3241 * able to safely make forward progress. Wake them
3242 */
3243 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3244 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003245 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003246
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003248 * Fragmentation may mean that the system cannot be rebalanced
3249 * for high-order allocations in all zones. If twice the
3250 * allocation size has been reclaimed and the zones are still
3251 * not balanced then recheck the watermarks at order-0 to
3252 * prevent kswapd reclaiming excessively. Assume that a
3253 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003255 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003256 order = sc.order = 0;
3257
Mel Gormanb8e83b92013-07-03 15:01:45 -07003258 /* Check if kswapd should be suspending */
3259 if (try_to_freeze() || kthread_should_stop())
3260 break;
3261
3262 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003263 * Compact if necessary and kswapd is reclaiming at least the
3264 * high watermark number of pages as requsted
3265 */
3266 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3267 compact_pgdat(pgdat, order);
3268
3269 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003270 * Raise priority if scanning rate is too low or there was no
3271 * progress in reclaiming pages
3272 */
3273 if (raise_priority || !sc.nr_reclaimed)
3274 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003275 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003276 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277
Mel Gormanb8e83b92013-07-03 15:01:45 -07003278out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003279 /*
Mel Gorman55150612012-07-31 16:44:35 -07003280 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003281 * makes a decision on the order we were last reclaiming at. However,
3282 * if another caller entered the allocator slow path while kswapd
3283 * was awake, order will remain at the higher level
3284 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003285 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003286 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287}
3288
Mel Gormandc83edd2011-01-13 15:46:26 -08003289static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003290{
3291 long remaining = 0;
3292 DEFINE_WAIT(wait);
3293
3294 if (freezing(current) || kthread_should_stop())
3295 return;
3296
3297 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3298
3299 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003300 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003301 remaining = schedule_timeout(HZ/10);
3302 finish_wait(&pgdat->kswapd_wait, &wait);
3303 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3304 }
3305
3306 /*
3307 * After a short sleep, check if it was a premature sleep. If not, then
3308 * go fully to sleep until explicitly woken up.
3309 */
Mel Gorman55150612012-07-31 16:44:35 -07003310 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003311 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3312
3313 /*
3314 * vmstat counters are not perfectly accurate and the estimated
3315 * value for counters such as NR_FREE_PAGES can deviate from the
3316 * true value by nr_online_cpus * threshold. To avoid the zone
3317 * watermarks being breached while under pressure, we reduce the
3318 * per-cpu vmstat threshold while kswapd is awake and restore
3319 * them before going back to sleep.
3320 */
3321 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003322
Mel Gorman62997022012-10-08 16:32:47 -07003323 /*
3324 * Compaction records what page blocks it recently failed to
3325 * isolate pages from and skips them in the future scanning.
3326 * When kswapd is going to sleep, it is reasonable to assume
3327 * that pages and compaction may succeed so reset the cache.
3328 */
3329 reset_isolation_suitable(pgdat);
3330
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003331 if (!kthread_should_stop())
3332 schedule();
3333
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003334 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3335 } else {
3336 if (remaining)
3337 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3338 else
3339 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3340 }
3341 finish_wait(&pgdat->kswapd_wait, &wait);
3342}
3343
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344/*
3345 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003346 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 *
3348 * This basically trickles out pages so that we have _some_
3349 * free memory available even if there is no other activity
3350 * that frees anything up. This is needed for things like routing
3351 * etc, where we otherwise might have all activity going on in
3352 * asynchronous contexts that cannot page things out.
3353 *
3354 * If there are applications that are active memory-allocators
3355 * (most normal use), this basically shouldn't matter.
3356 */
3357static int kswapd(void *p)
3358{
Mel Gorman215ddd62011-07-08 15:39:40 -07003359 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003360 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003361 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003362 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363 pg_data_t *pgdat = (pg_data_t*)p;
3364 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003365
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 struct reclaim_state reclaim_state = {
3367 .reclaimed_slab = 0,
3368 };
Rusty Russella70f7302009-03-13 14:49:46 +10303369 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Nick Piggincf40bd12009-01-21 08:12:39 +01003371 lockdep_set_current_reclaim_state(GFP_KERNEL);
3372
Rusty Russell174596a2009-01-01 10:12:29 +10303373 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003374 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 current->reclaim_state = &reclaim_state;
3376
3377 /*
3378 * Tell the memory management that we're a "memory allocator",
3379 * and that if we need more memory we should get access to it
3380 * regardless (see "__alloc_pages()"). "kswapd" should
3381 * never get caught in the normal page freeing logic.
3382 *
3383 * (Kswapd normally doesn't need memory anyway, but sometimes
3384 * you need a small amount of memory in order to be able to
3385 * page out something else, and this flag essentially protects
3386 * us from recursively trying to free more memory as we're
3387 * trying to free the first piece of memory in the first place).
3388 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003389 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003390 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391
Mel Gorman215ddd62011-07-08 15:39:40 -07003392 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003393 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003394 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003395 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003397 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003398
Mel Gorman215ddd62011-07-08 15:39:40 -07003399 /*
3400 * If the last balance_pgdat was unsuccessful it's unlikely a
3401 * new request of a similar or harder type will succeed soon
3402 * so consider going to sleep on the basis we reclaimed at
3403 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003404 if (balanced_classzone_idx >= new_classzone_idx &&
3405 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003406 new_order = pgdat->kswapd_max_order;
3407 new_classzone_idx = pgdat->classzone_idx;
3408 pgdat->kswapd_max_order = 0;
3409 pgdat->classzone_idx = pgdat->nr_zones - 1;
3410 }
3411
Mel Gorman99504742011-01-13 15:46:20 -08003412 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 /*
3414 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003415 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 */
3417 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003418 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003420 kswapd_try_to_sleep(pgdat, balanced_order,
3421 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003423 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003424 new_order = order;
3425 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003426 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003427 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
David Rientjes8fe23e02009-12-14 17:58:33 -08003430 ret = try_to_freeze();
3431 if (kthread_should_stop())
3432 break;
3433
3434 /*
3435 * We can speed up thawing tasks if we don't call balance_pgdat
3436 * after returning from the refrigerator
3437 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003438 if (!ret) {
3439 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003440 balanced_classzone_idx = classzone_idx;
3441 balanced_order = balance_pgdat(pgdat, order,
3442 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003443 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003445
Johannes Weiner71abdc12014-06-06 14:35:35 -07003446 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003447 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003448 lockdep_clear_current_reclaim_state();
3449
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 return 0;
3451}
3452
3453/*
3454 * A zone is low on free memory, so wake its kswapd task to service it.
3455 */
Mel Gorman99504742011-01-13 15:46:20 -08003456void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457{
3458 pg_data_t *pgdat;
3459
Con Kolivasf3fe6512006-01-06 00:11:15 -08003460 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 return;
3462
Vladimir Davydov344736f2014-10-20 15:50:30 +04003463 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003465 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003466 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003467 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003468 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3469 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003470 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003472 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003473 return;
3474
3475 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003476 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477}
3478
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003479#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003481 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003482 * freed pages.
3483 *
3484 * Rather than trying to age LRUs the aim is to preserve the overall
3485 * LRU order by reclaiming preferentially
3486 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003488unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003490 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003491 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003492 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003493 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003494 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003495 .may_writepage = 1,
3496 .may_unmap = 1,
3497 .may_swap = 1,
3498 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003500 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003501 struct task_struct *p = current;
3502 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003504 p->flags |= PF_MEMALLOC;
3505 lockdep_set_current_reclaim_state(sc.gfp_mask);
3506 reclaim_state.reclaimed_slab = 0;
3507 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003508
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003509 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003510
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003511 p->reclaim_state = NULL;
3512 lockdep_clear_current_reclaim_state();
3513 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003514
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003515 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003517#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519/* It's optimal to keep kswapds on the same CPUs as their memory, but
3520 not required for correctness. So if the last cpu in a node goes
3521 away, we get changed to run anywhere: as the first one comes back,
3522 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003523static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3524 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003526 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003528 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003529 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003530 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303531 const struct cpumask *mask;
3532
3533 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003534
Rusty Russell3e597942009-01-01 10:12:24 +10303535 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003537 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 }
3539 }
3540 return NOTIFY_OK;
3541}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542
Yasunori Goto3218ae12006-06-27 02:53:33 -07003543/*
3544 * This kswapd start function will be called by init and node-hot-add.
3545 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3546 */
3547int kswapd_run(int nid)
3548{
3549 pg_data_t *pgdat = NODE_DATA(nid);
3550 int ret = 0;
3551
3552 if (pgdat->kswapd)
3553 return 0;
3554
3555 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3556 if (IS_ERR(pgdat->kswapd)) {
3557 /* failure at boot is fatal */
3558 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003559 pr_err("Failed to start kswapd on node %d\n", nid);
3560 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003561 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003562 }
3563 return ret;
3564}
3565
David Rientjes8fe23e02009-12-14 17:58:33 -08003566/*
Jiang Liud8adde12012-07-11 14:01:52 -07003567 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003568 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003569 */
3570void kswapd_stop(int nid)
3571{
3572 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3573
Jiang Liud8adde12012-07-11 14:01:52 -07003574 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003575 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003576 NODE_DATA(nid)->kswapd = NULL;
3577 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003578}
3579
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580static int __init kswapd_init(void)
3581{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003582 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003583
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003585 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003586 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 hotcpu_notifier(cpu_callback, 0);
3588 return 0;
3589}
3590
3591module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003592
3593#ifdef CONFIG_NUMA
3594/*
3595 * Zone reclaim mode
3596 *
3597 * If non-zero call zone_reclaim when the number of free pages falls below
3598 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003599 */
3600int zone_reclaim_mode __read_mostly;
3601
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003602#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003603#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003604#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3605#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3606
Christoph Lameter9eeff232006-01-18 17:42:31 -08003607/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003608 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3609 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3610 * a zone.
3611 */
3612#define ZONE_RECLAIM_PRIORITY 4
3613
Christoph Lameter9eeff232006-01-18 17:42:31 -08003614/*
Christoph Lameter96146342006-07-03 00:24:13 -07003615 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3616 * occur.
3617 */
3618int sysctl_min_unmapped_ratio = 1;
3619
3620/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003621 * If the number of slab pages in a zone grows beyond this percentage then
3622 * slab reclaim needs to occur.
3623 */
3624int sysctl_min_slab_ratio = 5;
3625
Mel Gorman90afa5d2009-06-16 15:33:20 -07003626static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3627{
3628 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3629 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3630 zone_page_state(zone, NR_ACTIVE_FILE);
3631
3632 /*
3633 * It's possible for there to be more file mapped pages than
3634 * accounted for by the pages on the file LRU lists because
3635 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3636 */
3637 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3638}
3639
3640/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3641static long zone_pagecache_reclaimable(struct zone *zone)
3642{
3643 long nr_pagecache_reclaimable;
3644 long delta = 0;
3645
3646 /*
3647 * If RECLAIM_SWAP is set, then all file pages are considered
3648 * potentially reclaimable. Otherwise, we have to worry about
3649 * pages like swapcache and zone_unmapped_file_pages() provides
3650 * a better estimate
3651 */
3652 if (zone_reclaim_mode & RECLAIM_SWAP)
3653 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3654 else
3655 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3656
3657 /* If we can't clean pages, remove dirty pages from consideration */
3658 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3659 delta += zone_page_state(zone, NR_FILE_DIRTY);
3660
3661 /* Watch for any possible underflows due to delta */
3662 if (unlikely(delta > nr_pagecache_reclaimable))
3663 delta = nr_pagecache_reclaimable;
3664
3665 return nr_pagecache_reclaimable - delta;
3666}
3667
Christoph Lameter0ff38492006-09-25 23:31:52 -07003668/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003669 * Try to free up some pages from this zone through reclaim.
3670 */
Andrew Morton179e9632006-03-22 00:08:18 -08003671static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003672{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003673 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003674 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003675 struct task_struct *p = current;
3676 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003677 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003678 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003679 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003680 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003681 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003682 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3683 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3684 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003685 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003686
Christoph Lameter9eeff232006-01-18 17:42:31 -08003687 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003688 /*
3689 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3690 * and we also need to be able to write out pages for RECLAIM_WRITE
3691 * and RECLAIM_SWAP.
3692 */
3693 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003694 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003695 reclaim_state.reclaimed_slab = 0;
3696 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003697
Mel Gorman90afa5d2009-06-16 15:33:20 -07003698 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003699 /*
3700 * Free memory by calling shrink zone with increasing
3701 * priorities until we have enough memory freed.
3702 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003703 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003704 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003705 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003706 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003707
Christoph Lameter9eeff232006-01-18 17:42:31 -08003708 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003709 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003710 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003711 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003712}
Andrew Morton179e9632006-03-22 00:08:18 -08003713
3714int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3715{
Andrew Morton179e9632006-03-22 00:08:18 -08003716 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003717 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003718
3719 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003720 * Zone reclaim reclaims unmapped file backed pages and
3721 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003722 *
Christoph Lameter96146342006-07-03 00:24:13 -07003723 * A small portion of unmapped file backed pages is needed for
3724 * file I/O otherwise pages read by file I/O will be immediately
3725 * thrown out if the zone is overallocated. So we do not reclaim
3726 * if less than a specified percentage of the zone is used by
3727 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003728 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003729 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3730 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003731 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003732
Lisa Du6e543d52013-09-11 14:22:36 -07003733 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003734 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003735
Andrew Morton179e9632006-03-22 00:08:18 -08003736 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003737 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003738 */
David Rientjesd773ed62007-10-16 23:26:01 -07003739 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003740 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003741
3742 /*
3743 * Only run zone reclaim on the local zone or on zones that do not
3744 * have associated processors. This will favor the local processor
3745 * over remote processors and spread off node memory allocations
3746 * as wide as possible.
3747 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003748 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003749 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003750 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003751
Johannes Weiner57054652014-10-09 15:28:17 -07003752 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003753 return ZONE_RECLAIM_NOSCAN;
3754
David Rientjesd773ed62007-10-16 23:26:01 -07003755 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003756 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003757
Mel Gorman24cf725182009-06-16 15:33:23 -07003758 if (!ret)
3759 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3760
David Rientjesd773ed62007-10-16 23:26:01 -07003761 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003762}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003763#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003764
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003765/*
3766 * page_evictable - test whether a page is evictable
3767 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003768 *
3769 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003770 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003771 *
3772 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003773 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003774 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003775 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003776 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003777int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003778{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003779 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003780}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003781
Hugh Dickins850465792012-01-20 14:34:19 -08003782#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003783/**
Hugh Dickins24513262012-01-20 14:34:21 -08003784 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3785 * @pages: array of pages to check
3786 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003787 *
Hugh Dickins24513262012-01-20 14:34:21 -08003788 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003789 *
3790 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003791 */
Hugh Dickins24513262012-01-20 14:34:21 -08003792void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003793{
Johannes Weiner925b7672012-01-12 17:18:15 -08003794 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003795 struct zone *zone = NULL;
3796 int pgscanned = 0;
3797 int pgrescued = 0;
3798 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003799
Hugh Dickins24513262012-01-20 14:34:21 -08003800 for (i = 0; i < nr_pages; i++) {
3801 struct page *page = pages[i];
3802 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003803
Hugh Dickins24513262012-01-20 14:34:21 -08003804 pgscanned++;
3805 pagezone = page_zone(page);
3806 if (pagezone != zone) {
3807 if (zone)
3808 spin_unlock_irq(&zone->lru_lock);
3809 zone = pagezone;
3810 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003811 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003812 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003813
Hugh Dickins24513262012-01-20 14:34:21 -08003814 if (!PageLRU(page) || !PageUnevictable(page))
3815 continue;
3816
Hugh Dickins39b5f292012-10-08 16:33:18 -07003817 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003818 enum lru_list lru = page_lru_base_type(page);
3819
Sasha Levin309381fea2014-01-23 15:52:54 -08003820 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003821 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003822 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3823 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003824 pgrescued++;
3825 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003826 }
Hugh Dickins24513262012-01-20 14:34:21 -08003827
3828 if (zone) {
3829 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3830 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3831 spin_unlock_irq(&zone->lru_lock);
3832 }
Hugh Dickins850465792012-01-20 14:34:19 -08003833}
3834#endif /* CONFIG_SHMEM */