blob: 5e8eadd71bac71bee1dd9a121a3d44f3a4373c56 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049
50#include <asm/tlbflush.h>
51#include <asm/div64.h>
52
53#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070054#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Nick Piggin0f8053a2006-03-22 00:08:33 -080056#include "internal.h"
57
Mel Gorman33906bc2010-08-09 17:19:16 -070058#define CREATE_TRACE_POINTS
59#include <trace/events/vmscan.h>
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080062 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
64
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040066 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Johannes Weineree814fe2014-08-06 16:06:19 -070068 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070069 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080070
Johannes Weineree814fe2014-08-06 16:06:19 -070071 /*
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
73 * are scanned.
74 */
75 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070076
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070077 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080078 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
80 */
81 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080082
Johannes Weineree814fe2014-08-06 16:06:19 -070083 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
86 unsigned int may_writepage:1;
87
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
90
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
93
Johannes Weiner241994ed2015-02-11 15:26:06 -080094 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
96
Johannes Weineree814fe2014-08-06 16:06:19 -070097 unsigned int hibernation_mode:1;
98
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
101
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
104
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107};
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
110
111#ifdef ARCH_HAS_PREFETCH
112#define prefetch_prev_lru_page(_page, _base, _field) \
113 do { \
114 if ((_page)->lru.prev != _base) { \
115 struct page *prev; \
116 \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetch(&prev->_field); \
119 } \
120 } while (0)
121#else
122#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
123#endif
124
125#ifdef ARCH_HAS_PREFETCHW
126#define prefetchw_prev_lru_page(_page, _base, _field) \
127 do { \
128 if ((_page)->lru.prev != _base) { \
129 struct page *prev; \
130 \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetchw(&prev->_field); \
133 } \
134 } while (0)
135#else
136#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137#endif
138
139/*
140 * From 0 .. 100. Higher means more swappy.
141 */
142int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700143/*
144 * The total number of pages which are beyond the high watermark within all
145 * zones.
146 */
147unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
149static LIST_HEAD(shrinker_list);
150static DECLARE_RWSEM(shrinker_rwsem);
151
Andrew Mortonc255a452012-07-31 16:43:02 -0700152#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800153static bool global_reclaim(struct scan_control *sc)
154{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800155 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800156}
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
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700455static int may_write_to_queue(struct backing_dev_info *bdi,
456 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457{
Christoph Lameter930d9152006-01-08 01:00:47 -0800458 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 return 1;
460 if (!bdi_write_congested(bdi))
461 return 1;
462 if (bdi == current->backing_dev_info)
463 return 1;
464 return 0;
465}
466
467/*
468 * We detected a synchronous write error writing a page out. Probably
469 * -ENOSPC. We need to propagate that into the address_space for a subsequent
470 * fsync(), msync() or close().
471 *
472 * The tricky part is that after writepage we cannot touch the mapping: nothing
473 * prevents it from being freed up. But we have a ref on the page and once
474 * that page is locked, the mapping is pinned.
475 *
476 * We're allowed to run sleeping lock_page() here because we know the caller has
477 * __GFP_FS.
478 */
479static void handle_write_error(struct address_space *mapping,
480 struct page *page, int error)
481{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100482 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700483 if (page_mapping(page) == mapping)
484 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 unlock_page(page);
486}
487
Christoph Lameter04e62a22006-06-23 02:03:38 -0700488/* possible outcome of pageout() */
489typedef enum {
490 /* failed to write page out, page is locked */
491 PAGE_KEEP,
492 /* move page to the active list, page is locked */
493 PAGE_ACTIVATE,
494 /* page has been sent to the disk successfully, page is unlocked */
495 PAGE_SUCCESS,
496 /* page is clean and locked */
497 PAGE_CLEAN,
498} pageout_t;
499
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500/*
Andrew Morton1742f192006-03-22 00:08:21 -0800501 * pageout is called by shrink_page_list() for each dirty page.
502 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700504static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700505 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506{
507 /*
508 * If the page is dirty, only perform writeback if that write
509 * will be non-blocking. To prevent this allocation from being
510 * stalled by pagecache activity. But note that there may be
511 * stalls if we need to run get_block(). We could test
512 * PagePrivate for that.
513 *
Al Viro81742022014-04-03 03:17:43 -0400514 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 * this page's queue, we can perform writeback even if that
516 * will block.
517 *
518 * If the page is swapcache, write it back even if that would
519 * block, for some throttling. This happens by accident, because
520 * swap_backing_dev_info is bust: it doesn't reflect the
521 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 */
523 if (!is_page_cache_freeable(page))
524 return PAGE_KEEP;
525 if (!mapping) {
526 /*
527 * Some data journaling orphaned pages can have
528 * page->mapping == NULL while being dirty with clean buffers.
529 */
David Howells266cf652009-04-03 16:42:36 +0100530 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 if (try_to_free_buffers(page)) {
532 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700533 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 return PAGE_CLEAN;
535 }
536 }
537 return PAGE_KEEP;
538 }
539 if (mapping->a_ops->writepage == NULL)
540 return PAGE_ACTIVATE;
Christoph Hellwigde1414a2015-01-14 10:42:36 +0100541 if (!may_write_to_queue(inode_to_bdi(mapping->host), sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 return PAGE_KEEP;
543
544 if (clear_page_dirty_for_io(page)) {
545 int res;
546 struct writeback_control wbc = {
547 .sync_mode = WB_SYNC_NONE,
548 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700549 .range_start = 0,
550 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 .for_reclaim = 1,
552 };
553
554 SetPageReclaim(page);
555 res = mapping->a_ops->writepage(page, &wbc);
556 if (res < 0)
557 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800558 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 ClearPageReclaim(page);
560 return PAGE_ACTIVATE;
561 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700562
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 if (!PageWriteback(page)) {
564 /* synchronous write or broken a_ops? */
565 ClearPageReclaim(page);
566 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700567 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700568 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 return PAGE_SUCCESS;
570 }
571
572 return PAGE_CLEAN;
573}
574
Andrew Mortona649fd92006-10-17 00:09:36 -0700575/*
Nick Piggine2867812008-07-25 19:45:30 -0700576 * Same as remove_mapping, but if the page is removed from the mapping, it
577 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700578 */
Johannes Weinera5289102014-04-03 14:47:51 -0700579static int __remove_mapping(struct address_space *mapping, struct page *page,
580 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800581{
Nick Piggin28e4d962006-09-25 23:31:23 -0700582 BUG_ON(!PageLocked(page));
583 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800584
Nick Piggin19fd6232008-07-25 19:45:32 -0700585 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800586 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700587 * The non racy check for a busy page.
588 *
589 * Must be careful with the order of the tests. When someone has
590 * a ref to the page, it may be possible that they dirty it then
591 * drop the reference. So if PageDirty is tested before page_count
592 * here, then the following race may occur:
593 *
594 * get_user_pages(&page);
595 * [user mapping goes away]
596 * write_to(page);
597 * !PageDirty(page) [good]
598 * SetPageDirty(page);
599 * put_page(page);
600 * !page_count(page) [good, discard it]
601 *
602 * [oops, our write_to data is lost]
603 *
604 * Reversing the order of the tests ensures such a situation cannot
605 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
606 * load is not satisfied before that of page->_count.
607 *
608 * Note that if SetPageDirty is always performed via set_page_dirty,
609 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800610 */
Nick Piggine2867812008-07-25 19:45:30 -0700611 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800612 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700613 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
614 if (unlikely(PageDirty(page))) {
615 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800616 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700617 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800618
619 if (PageSwapCache(page)) {
620 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700621 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800622 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700623 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700624 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700625 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500626 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700627 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500628
629 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700630 /*
631 * Remember a shadow entry for reclaimed file cache in
632 * order to detect refaults, thus thrashing, later on.
633 *
634 * But don't store shadows in an address space that is
635 * already exiting. This is not just an optizimation,
636 * inode reclaim needs to empty out the radix tree or
637 * the nodes are lost. Don't plant shadows behind its
638 * back.
639 */
640 if (reclaimed && page_is_file_cache(page) &&
641 !mapping_exiting(mapping))
642 shadow = workingset_eviction(mapping, page);
643 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700644 spin_unlock_irq(&mapping->tree_lock);
Linus Torvalds6072d132010-12-01 13:35:19 -0500645
646 if (freepage != NULL)
647 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800648 }
649
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800650 return 1;
651
652cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700653 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800654 return 0;
655}
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657/*
Nick Piggine2867812008-07-25 19:45:30 -0700658 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
659 * someone else has a ref on the page, abort and return 0. If it was
660 * successfully detached, return 1. Assumes the caller has a single ref on
661 * this page.
662 */
663int remove_mapping(struct address_space *mapping, struct page *page)
664{
Johannes Weinera5289102014-04-03 14:47:51 -0700665 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700666 /*
667 * Unfreezing the refcount with 1 rather than 2 effectively
668 * drops the pagecache ref for us without requiring another
669 * atomic operation.
670 */
671 page_unfreeze_refs(page, 1);
672 return 1;
673 }
674 return 0;
675}
676
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700677/**
678 * putback_lru_page - put previously isolated page onto appropriate LRU list
679 * @page: page to be put back to appropriate lru list
680 *
681 * Add previously isolated @page to appropriate LRU list.
682 * Page may still be unevictable for other reasons.
683 *
684 * lru_lock must not be held, interrupts must be enabled.
685 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700686void putback_lru_page(struct page *page)
687{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700688 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700689 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700690
Sasha Levin309381fea2014-01-23 15:52:54 -0800691 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700692
693redo:
694 ClearPageUnevictable(page);
695
Hugh Dickins39b5f292012-10-08 16:33:18 -0700696 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700697 /*
698 * For evictable pages, we can use the cache.
699 * In event of a race, worst case is we end up with an
700 * unevictable page on [in]active list.
701 * We know how to handle that.
702 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700703 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700704 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700705 } else {
706 /*
707 * Put unevictable pages directly on zone's unevictable
708 * list.
709 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700710 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700711 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700712 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700713 * When racing with an mlock or AS_UNEVICTABLE clearing
714 * (page is unlocked) make sure that if the other thread
715 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800716 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700717 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700718 * the page back to the evictable list.
719 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700720 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700721 */
722 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700723 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700724
725 /*
726 * page's status can change while we move it among lru. If an evictable
727 * page is on unevictable list, it never be freed. To avoid that,
728 * check after we added it to the list, again.
729 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700730 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700731 if (!isolate_lru_page(page)) {
732 put_page(page);
733 goto redo;
734 }
735 /* This means someone else dropped this page from LRU
736 * So, it will be freed or putback to LRU again. There is
737 * nothing to do here.
738 */
739 }
740
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700741 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700742 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700743 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700744 count_vm_event(UNEVICTABLE_PGCULLED);
745
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700746 put_page(page); /* drop ref from isolate */
747}
748
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800749enum page_references {
750 PAGEREF_RECLAIM,
751 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800752 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800753 PAGEREF_ACTIVATE,
754};
755
756static enum page_references page_check_references(struct page *page,
757 struct scan_control *sc)
758{
Johannes Weiner645747462010-03-05 13:42:22 -0800759 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800760 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800761
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700762 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
763 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800764 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800765
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800766 /*
767 * Mlock lost the isolation race with us. Let try_to_unmap()
768 * move the page to the unevictable list.
769 */
770 if (vm_flags & VM_LOCKED)
771 return PAGEREF_RECLAIM;
772
Johannes Weiner645747462010-03-05 13:42:22 -0800773 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700774 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800775 return PAGEREF_ACTIVATE;
776 /*
777 * All mapped pages start out with page table
778 * references from the instantiating fault, so we need
779 * to look twice if a mapped file page is used more
780 * than once.
781 *
782 * Mark it and spare it for another trip around the
783 * inactive list. Another page table reference will
784 * lead to its activation.
785 *
786 * Note: the mark is set for activated pages as well
787 * so that recently deactivated but used pages are
788 * quickly recovered.
789 */
790 SetPageReferenced(page);
791
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800792 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800793 return PAGEREF_ACTIVATE;
794
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800795 /*
796 * Activate file-backed executable pages after first usage.
797 */
798 if (vm_flags & VM_EXEC)
799 return PAGEREF_ACTIVATE;
800
Johannes Weiner645747462010-03-05 13:42:22 -0800801 return PAGEREF_KEEP;
802 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800803
804 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700805 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800806 return PAGEREF_RECLAIM_CLEAN;
807
808 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800809}
810
Mel Gormane2be15f2013-07-03 15:01:57 -0700811/* Check if a page is dirty or under writeback */
812static void page_check_dirty_writeback(struct page *page,
813 bool *dirty, bool *writeback)
814{
Mel Gormanb4597222013-07-03 15:02:05 -0700815 struct address_space *mapping;
816
Mel Gormane2be15f2013-07-03 15:01:57 -0700817 /*
818 * Anonymous pages are not handled by flushers and must be written
819 * from reclaim context. Do not stall reclaim based on them
820 */
821 if (!page_is_file_cache(page)) {
822 *dirty = false;
823 *writeback = false;
824 return;
825 }
826
827 /* By default assume that the page flags are accurate */
828 *dirty = PageDirty(page);
829 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700830
831 /* Verify dirty/writeback state if the filesystem supports it */
832 if (!page_has_private(page))
833 return;
834
835 mapping = page_mapping(page);
836 if (mapping && mapping->a_ops->is_dirty_writeback)
837 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700838}
839
Nick Piggine2867812008-07-25 19:45:30 -0700840/*
Andrew Morton1742f192006-03-22 00:08:21 -0800841 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 */
Andrew Morton1742f192006-03-22 00:08:21 -0800843static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700844 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700845 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700846 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700847 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700848 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700849 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700850 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700851 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700852 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853{
854 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700855 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700857 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700858 unsigned long nr_dirty = 0;
859 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800860 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700861 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700862 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
864 cond_resched();
865
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 while (!list_empty(page_list)) {
867 struct address_space *mapping;
868 struct page *page;
869 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700870 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700871 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872
873 cond_resched();
874
875 page = lru_to_page(page_list);
876 list_del(&page->lru);
877
Nick Piggin529ae9a2008-08-02 12:01:03 +0200878 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 goto keep;
880
Sasha Levin309381fea2014-01-23 15:52:54 -0800881 VM_BUG_ON_PAGE(PageActive(page), page);
882 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
884 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800885
Hugh Dickins39b5f292012-10-08 16:33:18 -0700886 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700887 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700888
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700889 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800890 goto keep_locked;
891
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 /* Double the slab pressure for mapped and swapcache pages */
893 if (page_mapped(page) || PageSwapCache(page))
894 sc->nr_scanned++;
895
Andy Whitcroftc661b072007-08-22 14:01:26 -0700896 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
897 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
898
Mel Gorman283aba92013-07-03 15:01:51 -0700899 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700900 * The number of dirty pages determines if a zone is marked
901 * reclaim_congested which affects wait_iff_congested. kswapd
902 * will stall and start writing pages if the tail of the LRU
903 * is all dirty unqueued pages.
904 */
905 page_check_dirty_writeback(page, &dirty, &writeback);
906 if (dirty || writeback)
907 nr_dirty++;
908
909 if (dirty && !writeback)
910 nr_unqueued_dirty++;
911
Mel Gormand04e8ac2013-07-03 15:02:03 -0700912 /*
913 * Treat this page as congested if the underlying BDI is or if
914 * pages are cycling through the LRU so quickly that the
915 * pages marked for immediate reclaim are making it to the
916 * end of the LRU a second time.
917 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700918 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800919 if (((dirty || writeback) && mapping &&
Christoph Hellwigde1414a2015-01-14 10:42:36 +0100920 bdi_write_congested(inode_to_bdi(mapping->host))) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700921 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700922 nr_congested++;
923
924 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700925 * If a page at the tail of the LRU is under writeback, there
926 * are three cases to consider.
927 *
928 * 1) If reclaim is encountering an excessive number of pages
929 * under writeback and this page is both under writeback and
930 * PageReclaim then it indicates that pages are being queued
931 * for IO but are being recycled through the LRU before the
932 * IO can complete. Waiting on the page itself risks an
933 * indefinite stall if it is impossible to writeback the
934 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700935 * note that the LRU is being scanned too quickly and the
936 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700937 *
938 * 2) Global reclaim encounters a page, memcg encounters a
939 * page that is not marked for immediate reclaim or
940 * the caller does not have __GFP_IO. In this case mark
941 * the page for immediate reclaim and continue scanning.
942 *
943 * __GFP_IO is checked because a loop driver thread might
944 * enter reclaim, and deadlock if it waits on a page for
945 * which it is needed to do the write (loop masks off
946 * __GFP_IO|__GFP_FS for this reason); but more thought
947 * would probably show more reasons.
948 *
949 * Don't require __GFP_FS, since we're not going into the
950 * FS, just waiting on its writeback completion. Worryingly,
951 * ext4 gfs2 and xfs allocate pages with
952 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
953 * may_enter_fs here is liable to OOM on them.
954 *
955 * 3) memcg encounters a page that is not already marked
956 * PageReclaim. memcg does not have any dirty pages
957 * throttling so we could easily OOM just because too many
958 * pages are in writeback and there is nothing else to
959 * reclaim. Wait for the writeback to complete.
960 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700961 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700962 /* Case 1 above */
963 if (current_is_kswapd() &&
964 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700965 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700966 nr_immediate++;
967 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700968
969 /* Case 2 above */
970 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700971 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
972 /*
973 * This is slightly racy - end_page_writeback()
974 * might have just cleared PageReclaim, then
975 * setting PageReclaim here end up interpreted
976 * as PageReadahead - but that does not matter
977 * enough to care. What we do want is for this
978 * page to have PageReclaim set next time memcg
979 * reclaim reaches the tests above, so it will
980 * then wait_on_page_writeback() to avoid OOM;
981 * and it's also appropriate in global reclaim.
982 */
983 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700984 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700985
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700986 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700987
988 /* Case 3 above */
989 } else {
990 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700991 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700992 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993
Minchan Kim02c6de82012-10-08 16:31:55 -0700994 if (!force_reclaim)
995 references = page_check_references(page, sc);
996
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800997 switch (references) {
998 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001000 case PAGEREF_KEEP:
1001 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001002 case PAGEREF_RECLAIM:
1003 case PAGEREF_RECLAIM_CLEAN:
1004 ; /* try to reclaim the page below */
1005 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 /*
1008 * Anonymous process memory has backing store?
1009 * Try to allocate it some swap space here.
1010 */
Nick Pigginb291f002008-10-18 20:26:44 -07001011 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001012 if (!(sc->gfp_mask & __GFP_IO))
1013 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001014 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001016 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
Mel Gormane2be15f2013-07-03 15:01:57 -07001018 /* Adding to swap updated mapping */
1019 mapping = page_mapping(page);
1020 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021
1022 /*
1023 * The page is mapped into the page tables of one or more
1024 * processes. Try to unmap it here.
1025 */
1026 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001027 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 case SWAP_FAIL:
1029 goto activate_locked;
1030 case SWAP_AGAIN:
1031 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001032 case SWAP_MLOCK:
1033 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 case SWAP_SUCCESS:
1035 ; /* try to free the page below */
1036 }
1037 }
1038
1039 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001040 /*
1041 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001042 * avoid risk of stack overflow but only writeback
1043 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001044 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001045 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001046 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001047 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001048 /*
1049 * Immediately reclaim when written back.
1050 * Similar in principal to deactivate_page()
1051 * except we already have the page isolated
1052 * and know it's dirty
1053 */
1054 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1055 SetPageReclaim(page);
1056
Mel Gormanee728862011-10-31 17:07:38 -07001057 goto keep_locked;
1058 }
1059
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001060 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001062 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001064 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 goto keep_locked;
1066
1067 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001068 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 case PAGE_KEEP:
1070 goto keep_locked;
1071 case PAGE_ACTIVATE:
1072 goto activate_locked;
1073 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001074 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001075 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001076 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001078
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 /*
1080 * A synchronous write - probably a ramdisk. Go
1081 * ahead and try to reclaim the page.
1082 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001083 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 goto keep;
1085 if (PageDirty(page) || PageWriteback(page))
1086 goto keep_locked;
1087 mapping = page_mapping(page);
1088 case PAGE_CLEAN:
1089 ; /* try to free the page below */
1090 }
1091 }
1092
1093 /*
1094 * If the page has buffers, try to free the buffer mappings
1095 * associated with this page. If we succeed we try to free
1096 * the page as well.
1097 *
1098 * We do this even if the page is PageDirty().
1099 * try_to_release_page() does not perform I/O, but it is
1100 * possible for a page to have PageDirty set, but it is actually
1101 * clean (all its buffers are clean). This happens if the
1102 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001103 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 * try_to_release_page() will discover that cleanness and will
1105 * drop the buffers and mark the page clean - it can be freed.
1106 *
1107 * Rarely, pages can have buffers and no ->mapping. These are
1108 * the pages which were not successfully invalidated in
1109 * truncate_complete_page(). We try to drop those buffers here
1110 * and if that worked, and the page is no longer mapped into
1111 * process address space (page_count == 1) it can be freed.
1112 * Otherwise, leave the page on the LRU so it is swappable.
1113 */
David Howells266cf652009-04-03 16:42:36 +01001114 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 if (!try_to_release_page(page, sc->gfp_mask))
1116 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001117 if (!mapping && page_count(page) == 1) {
1118 unlock_page(page);
1119 if (put_page_testzero(page))
1120 goto free_it;
1121 else {
1122 /*
1123 * rare race with speculative reference.
1124 * the speculative reference will free
1125 * this page shortly, so we may
1126 * increment nr_reclaimed here (and
1127 * leave it off the LRU).
1128 */
1129 nr_reclaimed++;
1130 continue;
1131 }
1132 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
1134
Johannes Weinera5289102014-04-03 14:47:51 -07001135 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001136 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Nick Piggina978d6f2008-10-18 20:26:58 -07001138 /*
1139 * At this point, we have no other references and there is
1140 * no way to pick any more up (removed from LRU, removed
1141 * from pagecache). Can use non-atomic bitops now (and
1142 * we obviously don't have to worry about waking up a process
1143 * waiting on the page lock, because there are no references.
1144 */
1145 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001146free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001147 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001148
1149 /*
1150 * Is there need to periodically free_page_list? It would
1151 * appear not as the counts should be low
1152 */
1153 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 continue;
1155
Nick Pigginb291f002008-10-18 20:26:44 -07001156cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001157 if (PageSwapCache(page))
1158 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001159 unlock_page(page);
1160 putback_lru_page(page);
1161 continue;
1162
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001164 /* Not a candidate for swapping, so reclaim swap space. */
1165 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001166 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001167 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 SetPageActive(page);
1169 pgactivate++;
1170keep_locked:
1171 unlock_page(page);
1172keep:
1173 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001174 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001176
Johannes Weiner747db952014-08-08 14:19:24 -07001177 mem_cgroup_uncharge_list(&free_pages);
Mel Gormanb745bc82014-06-04 16:10:22 -07001178 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001179
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001181 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001182
Mel Gorman8e950282013-07-03 15:02:02 -07001183 *ret_nr_dirty += nr_dirty;
1184 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001185 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001186 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001187 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001188 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189}
1190
Minchan Kim02c6de82012-10-08 16:31:55 -07001191unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1192 struct list_head *page_list)
1193{
1194 struct scan_control sc = {
1195 .gfp_mask = GFP_KERNEL,
1196 .priority = DEF_PRIORITY,
1197 .may_unmap = 1,
1198 };
Mel Gorman8e950282013-07-03 15:02:02 -07001199 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001200 struct page *page, *next;
1201 LIST_HEAD(clean_pages);
1202
1203 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001204 if (page_is_file_cache(page) && !PageDirty(page) &&
1205 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001206 ClearPageActive(page);
1207 list_move(&page->lru, &clean_pages);
1208 }
1209 }
1210
1211 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001212 TTU_UNMAP|TTU_IGNORE_ACCESS,
1213 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001214 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001215 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001216 return ret;
1217}
1218
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001219/*
1220 * Attempt to remove the specified page from its LRU. Only take this page
1221 * if it is of the appropriate PageActive status. Pages which are being
1222 * freed elsewhere are also ignored.
1223 *
1224 * page: page to consider
1225 * mode: one of the LRU isolation modes defined above
1226 *
1227 * returns 0 on success, -ve errno on failure.
1228 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001229int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001230{
1231 int ret = -EINVAL;
1232
1233 /* Only take pages on the LRU. */
1234 if (!PageLRU(page))
1235 return ret;
1236
Minchan Kime46a2872012-10-08 16:33:48 -07001237 /* Compaction should not handle unevictable pages but CMA can do so */
1238 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001239 return ret;
1240
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001241 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001242
Mel Gormanc8244932012-01-12 17:19:38 -08001243 /*
1244 * To minimise LRU disruption, the caller can indicate that it only
1245 * wants to isolate pages it will be able to operate on without
1246 * blocking - clean pages for the most part.
1247 *
1248 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1249 * is used by reclaim when it is cannot write to backing storage
1250 *
1251 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1252 * that it is possible to migrate without blocking
1253 */
1254 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1255 /* All the caller can do on PageWriteback is block */
1256 if (PageWriteback(page))
1257 return ret;
1258
1259 if (PageDirty(page)) {
1260 struct address_space *mapping;
1261
1262 /* ISOLATE_CLEAN means only clean pages */
1263 if (mode & ISOLATE_CLEAN)
1264 return ret;
1265
1266 /*
1267 * Only pages without mappings or that have a
1268 * ->migratepage callback are possible to migrate
1269 * without blocking
1270 */
1271 mapping = page_mapping(page);
1272 if (mapping && !mapping->a_ops->migratepage)
1273 return ret;
1274 }
1275 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001276
Minchan Kimf80c0672011-10-31 17:06:55 -07001277 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1278 return ret;
1279
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001280 if (likely(get_page_unless_zero(page))) {
1281 /*
1282 * Be careful not to clear PageLRU until after we're
1283 * sure the page is not being freed elsewhere -- the
1284 * page release code relies on it.
1285 */
1286 ClearPageLRU(page);
1287 ret = 0;
1288 }
1289
1290 return ret;
1291}
1292
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001293/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 * zone->lru_lock is heavily contended. Some of the functions that
1295 * shrink the lists perform better by taking out a batch of pages
1296 * and working on them outside the LRU lock.
1297 *
1298 * For pagecache intensive workloads, this function is the hottest
1299 * spot in the kernel (apart from copy_*_user functions).
1300 *
1301 * Appropriate locks must be held before calling this function.
1302 *
1303 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001304 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001306 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001307 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001308 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001309 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 *
1311 * returns how many pages were moved onto *@dst.
1312 */
Andrew Morton69e05942006-03-22 00:08:19 -08001313static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001314 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001315 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001316 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001318 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001319 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001320 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001322 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001323 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001324 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001325
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 page = lru_to_page(src);
1327 prefetchw_prev_lru_page(page, src, flags);
1328
Sasha Levin309381fea2014-01-23 15:52:54 -08001329 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001330
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001331 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001332 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001333 nr_pages = hpage_nr_pages(page);
1334 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001335 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001336 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001337 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001338
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001339 case -EBUSY:
1340 /* else it is being freed elsewhere */
1341 list_move(&page->lru, src);
1342 continue;
1343
1344 default:
1345 BUG();
1346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 }
1348
Hugh Dickinsf6260122012-01-12 17:20:06 -08001349 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001350 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1351 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 return nr_taken;
1353}
1354
Nick Piggin62695a82008-10-18 20:26:09 -07001355/**
1356 * isolate_lru_page - tries to isolate a page from its LRU list
1357 * @page: page to isolate from its LRU list
1358 *
1359 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1360 * vmstat statistic corresponding to whatever LRU list the page was on.
1361 *
1362 * Returns 0 if the page was removed from an LRU list.
1363 * Returns -EBUSY if the page was not on an LRU list.
1364 *
1365 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001366 * the active list, it will have PageActive set. If it was found on
1367 * the unevictable list, it will have the PageUnevictable bit set. That flag
1368 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001369 *
1370 * The vmstat statistic corresponding to the list on which the page was
1371 * found will be decremented.
1372 *
1373 * Restrictions:
1374 * (1) Must be called with an elevated refcount on the page. This is a
1375 * fundamentnal difference from isolate_lru_pages (which is called
1376 * without a stable reference).
1377 * (2) the lru_lock must not be held.
1378 * (3) interrupts must be enabled.
1379 */
1380int isolate_lru_page(struct page *page)
1381{
1382 int ret = -EBUSY;
1383
Sasha Levin309381fea2014-01-23 15:52:54 -08001384 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001385
Nick Piggin62695a82008-10-18 20:26:09 -07001386 if (PageLRU(page)) {
1387 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001388 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001389
1390 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001391 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001392 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001393 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001394 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001395 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001396 del_page_from_lru_list(page, lruvec, lru);
1397 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001398 }
1399 spin_unlock_irq(&zone->lru_lock);
1400 }
1401 return ret;
1402}
1403
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001404/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001405 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1406 * then get resheduled. When there are massive number of tasks doing page
1407 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1408 * the LRU list will go small and be scanned faster than necessary, leading to
1409 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001410 */
1411static int too_many_isolated(struct zone *zone, int file,
1412 struct scan_control *sc)
1413{
1414 unsigned long inactive, isolated;
1415
1416 if (current_is_kswapd())
1417 return 0;
1418
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001419 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001420 return 0;
1421
1422 if (file) {
1423 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1424 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1425 } else {
1426 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1427 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1428 }
1429
Fengguang Wu3cf23842012-12-18 14:23:31 -08001430 /*
1431 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1432 * won't get blocked by normal direct-reclaimers, forming a circular
1433 * deadlock.
1434 */
1435 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1436 inactive >>= 3;
1437
Rik van Riel35cd7812009-09-21 17:01:38 -07001438 return isolated > inactive;
1439}
1440
Mel Gorman66635622010-08-09 17:19:30 -07001441static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001442putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001443{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001444 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1445 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001446 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001447
Mel Gorman66635622010-08-09 17:19:30 -07001448 /*
1449 * Put back any unfreeable pages.
1450 */
Mel Gorman66635622010-08-09 17:19:30 -07001451 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001452 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001453 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001454
Sasha Levin309381fea2014-01-23 15:52:54 -08001455 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001456 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001457 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001458 spin_unlock_irq(&zone->lru_lock);
1459 putback_lru_page(page);
1460 spin_lock_irq(&zone->lru_lock);
1461 continue;
1462 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001463
1464 lruvec = mem_cgroup_page_lruvec(page, zone);
1465
Linus Torvalds7a608572011-01-17 14:42:19 -08001466 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001467 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001468 add_page_to_lru_list(page, lruvec, lru);
1469
Mel Gorman66635622010-08-09 17:19:30 -07001470 if (is_active_lru(lru)) {
1471 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001472 int numpages = hpage_nr_pages(page);
1473 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001474 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001475 if (put_page_testzero(page)) {
1476 __ClearPageLRU(page);
1477 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001478 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001479
1480 if (unlikely(PageCompound(page))) {
1481 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001482 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001483 (*get_compound_page_dtor(page))(page);
1484 spin_lock_irq(&zone->lru_lock);
1485 } else
1486 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001487 }
1488 }
Mel Gorman66635622010-08-09 17:19:30 -07001489
Hugh Dickins3f797682012-01-12 17:20:07 -08001490 /*
1491 * To save our caller's stack, now use input list for pages to free.
1492 */
1493 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001494}
1495
1496/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001497 * If a kernel thread (such as nfsd for loop-back mounts) services
1498 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1499 * In that case we should only throttle if the backing device it is
1500 * writing to is congested. In other cases it is safe to throttle.
1501 */
1502static int current_may_throttle(void)
1503{
1504 return !(current->flags & PF_LESS_THROTTLE) ||
1505 current->backing_dev_info == NULL ||
1506 bdi_write_congested(current->backing_dev_info);
1507}
1508
1509/*
Andrew Morton1742f192006-03-22 00:08:21 -08001510 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1511 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 */
Mel Gorman66635622010-08-09 17:19:30 -07001513static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001514shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001515 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516{
1517 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001518 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001519 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001520 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001521 unsigned long nr_dirty = 0;
1522 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001523 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001524 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001525 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001526 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001527 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001528 struct zone *zone = lruvec_zone(lruvec);
1529 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001530
Rik van Riel35cd7812009-09-21 17:01:38 -07001531 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001532 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001533
1534 /* We are about to die and free our memory. Return now. */
1535 if (fatal_signal_pending(current))
1536 return SWAP_CLUSTER_MAX;
1537 }
1538
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001540
1541 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001542 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001543 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001544 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001545
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001548 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1549 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001550
1551 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1552 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1553
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001554 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001555 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001556 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001557 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001558 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001559 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001560 }
Hillf Dantond563c052012-03-21 16:34:02 -07001561 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001562
Hillf Dantond563c052012-03-21 16:34:02 -07001563 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001564 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001565
Minchan Kim02c6de82012-10-08 16:31:55 -07001566 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001567 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1568 &nr_writeback, &nr_immediate,
1569 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001570
Hugh Dickins3f797682012-01-12 17:20:07 -08001571 spin_lock_irq(&zone->lru_lock);
1572
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001573 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001574
Ying Han904249a2012-04-25 16:01:48 -07001575 if (global_reclaim(sc)) {
1576 if (current_is_kswapd())
1577 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1578 nr_reclaimed);
1579 else
1580 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1581 nr_reclaimed);
1582 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001583
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001584 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001585
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001586 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001587
1588 spin_unlock_irq(&zone->lru_lock);
1589
Johannes Weiner747db952014-08-08 14:19:24 -07001590 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001591 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001592
Mel Gorman92df3a72011-10-31 17:07:56 -07001593 /*
1594 * If reclaim is isolating dirty pages under writeback, it implies
1595 * that the long-lived page allocation rate is exceeding the page
1596 * laundering rate. Either the global limits are not being effective
1597 * at throttling processes due to the page distribution throughout
1598 * zones or there is heavy usage of a slow backing device. The
1599 * only option is to throttle from reclaim context which is not ideal
1600 * as there is no guarantee the dirtying process is throttled in the
1601 * same way balance_dirty_pages() manages.
1602 *
Mel Gorman8e950282013-07-03 15:02:02 -07001603 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1604 * of pages under pages flagged for immediate reclaim and stall if any
1605 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001606 */
Mel Gorman918fc712013-07-08 16:00:25 -07001607 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001608 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001609
Mel Gormand43006d2013-07-03 15:01:50 -07001610 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001611 * memcg will stall in page writeback so only consider forcibly
1612 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001613 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001614 if (global_reclaim(sc)) {
1615 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001616 * Tag a zone as congested if all the dirty pages scanned were
1617 * backed by a congested BDI and wait_iff_congested will stall.
1618 */
1619 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001620 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001621
1622 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001623 * If dirty pages are scanned that are not queued for IO, it
1624 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001625 * the zone ZONE_DIRTY and kswapd will start writing pages from
1626 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001627 */
1628 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001629 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001630
1631 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001632 * If kswapd scans pages marked marked for immediate
1633 * reclaim and under writeback (nr_immediate), it implies
1634 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001635 * they are written so also forcibly stall.
1636 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001637 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001638 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001639 }
Mel Gormand43006d2013-07-03 15:01:50 -07001640
Mel Gorman8e950282013-07-03 15:02:02 -07001641 /*
1642 * Stall direct reclaim for IO completions if underlying BDIs or zone
1643 * is congested. Allow kswapd to continue until it starts encountering
1644 * unqueued dirty pages or cycling through the LRU too quickly.
1645 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001646 if (!sc->hibernation_mode && !current_is_kswapd() &&
1647 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001648 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1649
Mel Gormane11da5b2010-10-26 14:21:40 -07001650 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1651 zone_idx(zone),
1652 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001653 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001654 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001655 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656}
1657
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001658/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 * This moves pages from the active list to the inactive list.
1660 *
1661 * We move them the other way if the page is referenced by one or more
1662 * processes, from rmap.
1663 *
1664 * If the pages are mostly unmapped, the processing is fast and it is
1665 * appropriate to hold zone->lru_lock across the whole operation. But if
1666 * the pages are mapped, the processing is slow (page_referenced()) so we
1667 * should drop zone->lru_lock around each page. It's impossible to balance
1668 * this, so instead we remove the pages from the LRU while processing them.
1669 * It is safe to rely on PG_active against the non-LRU pages in here because
1670 * nobody will play with that bit on a non-LRU page.
1671 *
1672 * The downside is that we have to touch page->_count against each page.
1673 * But we had to alter page->flags anyway.
1674 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001675
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001676static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001677 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001678 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001679 enum lru_list lru)
1680{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001681 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001682 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001683 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001684 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001685
Wu Fengguang3eb41402009-06-16 15:33:13 -07001686 while (!list_empty(list)) {
1687 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001688 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001689
Sasha Levin309381fea2014-01-23 15:52:54 -08001690 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001691 SetPageLRU(page);
1692
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001693 nr_pages = hpage_nr_pages(page);
1694 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001695 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001696 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001697
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001698 if (put_page_testzero(page)) {
1699 __ClearPageLRU(page);
1700 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001701 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001702
1703 if (unlikely(PageCompound(page))) {
1704 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001705 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001706 (*get_compound_page_dtor(page))(page);
1707 spin_lock_irq(&zone->lru_lock);
1708 } else
1709 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001710 }
1711 }
1712 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1713 if (!is_active_lru(lru))
1714 __count_vm_events(PGDEACTIVATE, pgmoved);
1715}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001716
Hugh Dickinsf6260122012-01-12 17:20:06 -08001717static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001718 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001719 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001720 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001722 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001723 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001724 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001726 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001727 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001729 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001730 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001731 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001732 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001733 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734
1735 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001736
1737 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001738 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001739 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001740 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001741
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001743
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001744 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1745 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001746 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001747 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001748
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001749 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001750
Hugh Dickinsf6260122012-01-12 17:20:06 -08001751 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001752 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001753 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 spin_unlock_irq(&zone->lru_lock);
1755
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 while (!list_empty(&l_hold)) {
1757 cond_resched();
1758 page = lru_to_page(&l_hold);
1759 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001760
Hugh Dickins39b5f292012-10-08 16:33:18 -07001761 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001762 putback_lru_page(page);
1763 continue;
1764 }
1765
Mel Gormancc715d92012-03-21 16:34:00 -07001766 if (unlikely(buffer_heads_over_limit)) {
1767 if (page_has_private(page) && trylock_page(page)) {
1768 if (page_has_private(page))
1769 try_to_release_page(page, 0);
1770 unlock_page(page);
1771 }
1772 }
1773
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001774 if (page_referenced(page, 0, sc->target_mem_cgroup,
1775 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001776 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001777 /*
1778 * Identify referenced, file-backed active pages and
1779 * give them one more trip around the active list. So
1780 * that executable code get better chances to stay in
1781 * memory under moderate memory pressure. Anon pages
1782 * are not likely to be evicted by use-once streaming
1783 * IO, plus JVM can create lots of anon VM_EXEC pages,
1784 * so we ignore them here.
1785 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001786 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001787 list_add(&page->lru, &l_active);
1788 continue;
1789 }
1790 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001791
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001792 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 list_add(&page->lru, &l_inactive);
1794 }
1795
Andrew Mortonb5557492009-01-06 14:40:13 -08001796 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001797 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001798 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001799 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001800 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001801 * Count referenced pages from currently used mappings as rotated,
1802 * even though only some of them are actually re-activated. This
1803 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001804 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001805 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001806 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001807
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001808 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1809 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001810 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001811 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001812
Johannes Weiner747db952014-08-08 14:19:24 -07001813 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001814 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815}
1816
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001817#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001818static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001819{
1820 unsigned long active, inactive;
1821
1822 active = zone_page_state(zone, NR_ACTIVE_ANON);
1823 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1824
1825 if (inactive * zone->inactive_ratio < active)
1826 return 1;
1827
1828 return 0;
1829}
1830
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001831/**
1832 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001833 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001834 *
1835 * Returns true if the zone does not have enough inactive anon pages,
1836 * meaning some active anon pages need to be deactivated.
1837 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001838static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001839{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001840 /*
1841 * If we don't have swap space, anonymous page deactivation
1842 * is pointless.
1843 */
1844 if (!total_swap_pages)
1845 return 0;
1846
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001847 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001848 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001849
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001850 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001851}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001852#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001853static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001854{
1855 return 0;
1856}
1857#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001858
Rik van Riel56e49d22009-06-16 15:32:28 -07001859/**
1860 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001861 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001862 *
1863 * When the system is doing streaming IO, memory pressure here
1864 * ensures that active file pages get deactivated, until more
1865 * than half of the file pages are on the inactive list.
1866 *
1867 * Once we get to that situation, protect the system's working
1868 * set from being evicted by disabling active file page aging.
1869 *
1870 * This uses a different ratio than the anonymous pages, because
1871 * the page cache uses a use-once replacement algorithm.
1872 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001873static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001874{
Johannes Weinere3790142013-02-22 16:35:19 -08001875 unsigned long inactive;
1876 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001877
Johannes Weinere3790142013-02-22 16:35:19 -08001878 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1879 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1880
1881 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001882}
1883
Hugh Dickins75b00af2012-05-29 15:07:09 -07001884static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001885{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001886 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001887 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001888 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001889 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001890}
1891
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001892static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001893 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001894{
Rik van Rielb39415b2009-12-14 17:59:48 -08001895 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001896 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001897 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001898 return 0;
1899 }
1900
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001901 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001902}
1903
Johannes Weiner9a265112013-02-22 16:32:17 -08001904enum scan_balance {
1905 SCAN_EQUAL,
1906 SCAN_FRACT,
1907 SCAN_ANON,
1908 SCAN_FILE,
1909};
1910
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001912 * Determine how aggressively the anon and file LRU lists should be
1913 * scanned. The relative value of each set of LRU lists is determined
1914 * by looking at the fraction of the pages scanned we did rotate back
1915 * onto the active list instead of evict.
1916 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001917 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1918 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001919 */
Johannes Weiner02695172014-08-06 16:06:17 -07001920static void get_scan_count(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001921 struct scan_control *sc, unsigned long *nr,
1922 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001923{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001924 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001925 u64 fraction[2];
1926 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001927 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001928 unsigned long anon_prio, file_prio;
1929 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001930 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001931 bool force_scan = false;
1932 unsigned long ap, fp;
1933 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001934 bool some_scanned;
1935 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001936
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001937 /*
1938 * If the zone or memcg is small, nr[l] can be 0. This
1939 * results in no scanning on this priority and a potential
1940 * priority drop. Global direct reclaim can go to the next
1941 * zone and tends to have no problems. Global kswapd is for
1942 * zone balancing and it needs to scan a minimum amount. When
1943 * reclaiming for a memcg, a priority drop can cause high
1944 * latencies, so it's better to scan a minimum amount there as
1945 * well.
1946 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08001947 if (current_is_kswapd()) {
1948 if (!zone_reclaimable(zone))
1949 force_scan = true;
1950 if (!mem_cgroup_lruvec_online(lruvec))
1951 force_scan = true;
1952 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001953 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001954 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001955
1956 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001957 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001958 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001959 goto out;
1960 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001961
Johannes Weiner10316b32013-02-22 16:32:14 -08001962 /*
1963 * Global reclaim will swap to prevent OOM even with no
1964 * swappiness, but memcg users want to use this knob to
1965 * disable swapping for individual groups completely when
1966 * using the memory controller's swap limit feature would be
1967 * too expensive.
1968 */
Johannes Weiner02695172014-08-06 16:06:17 -07001969 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001970 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001971 goto out;
1972 }
1973
1974 /*
1975 * Do not apply any pressure balancing cleverness when the
1976 * system is close to OOM, scan both anon and file equally
1977 * (unless the swappiness setting disagrees with swapping).
1978 */
Johannes Weiner02695172014-08-06 16:06:17 -07001979 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001980 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001981 goto out;
1982 }
1983
Johannes Weiner11d16c22013-02-22 16:32:15 -08001984 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001985 * Prevent the reclaimer from falling into the cache trap: as
1986 * cache pages start out inactive, every cache fault will tip
1987 * the scan balance towards the file LRU. And as the file LRU
1988 * shrinks, so does the window for rotation from references.
1989 * This means we have a runaway feedback loop where a tiny
1990 * thrashing file LRU becomes infinitely more attractive than
1991 * anon pages. Try to detect this based on file LRU size.
1992 */
1993 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001994 unsigned long zonefile;
1995 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07001996
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001997 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1998 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1999 zone_page_state(zone, NR_INACTIVE_FILE);
2000
2001 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07002002 scan_balance = SCAN_ANON;
2003 goto out;
2004 }
2005 }
2006
2007 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002008 * There is enough inactive page cache, do not reclaim
2009 * anything from the anonymous working set right now.
2010 */
2011 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002012 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002013 goto out;
2014 }
2015
Johannes Weiner9a265112013-02-22 16:32:17 -08002016 scan_balance = SCAN_FRACT;
2017
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002018 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002019 * With swappiness at 100, anonymous and file have the same priority.
2020 * This scanning priority is essentially the inverse of IO cost.
2021 */
Johannes Weiner02695172014-08-06 16:06:17 -07002022 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002023 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002024
2025 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002026 * OK, so we have swap space and a fair amount of page cache
2027 * pages. We use the recently rotated / recently scanned
2028 * ratios to determine how valuable each cache is.
2029 *
2030 * Because workloads change over time (and to avoid overflow)
2031 * we keep these statistics as a floating average, which ends
2032 * up weighing recent references more than old ones.
2033 *
2034 * anon in [0], file in [1]
2035 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002036
2037 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2038 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2039 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2040 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2041
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002042 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002043 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002044 reclaim_stat->recent_scanned[0] /= 2;
2045 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002046 }
2047
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002048 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002049 reclaim_stat->recent_scanned[1] /= 2;
2050 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002051 }
2052
2053 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002054 * The amount of pressure on anon vs file pages is inversely
2055 * proportional to the fraction of recently scanned pages on
2056 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002057 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002058 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002059 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002060
Satoru Moriyafe350042012-05-29 15:06:47 -07002061 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002062 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002063 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002064
Shaohua Li76a33fc2010-05-24 14:32:36 -07002065 fraction[0] = ap;
2066 fraction[1] = fp;
2067 denominator = ap + fp + 1;
2068out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002069 some_scanned = false;
2070 /* Only use force_scan on second pass. */
2071 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002072 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002073 for_each_evictable_lru(lru) {
2074 int file = is_file_lru(lru);
2075 unsigned long size;
2076 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002077
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002078 size = get_lru_size(lruvec, lru);
2079 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002080
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002081 if (!scan && pass && force_scan)
2082 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002083
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002084 switch (scan_balance) {
2085 case SCAN_EQUAL:
2086 /* Scan lists relative to size */
2087 break;
2088 case SCAN_FRACT:
2089 /*
2090 * Scan types proportional to swappiness and
2091 * their relative recent reclaim efficiency.
2092 */
2093 scan = div64_u64(scan * fraction[file],
2094 denominator);
2095 break;
2096 case SCAN_FILE:
2097 case SCAN_ANON:
2098 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002099 if ((scan_balance == SCAN_FILE) != file) {
2100 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002101 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002102 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002103 break;
2104 default:
2105 /* Look ma, no brain */
2106 BUG();
2107 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002108
2109 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002110 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002111
Johannes Weiner9a265112013-02-22 16:32:17 -08002112 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002113 * Skip the second pass and don't force_scan,
2114 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002115 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002116 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002117 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002118 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002119}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002120
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002121/*
2122 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2123 */
Johannes Weiner02695172014-08-06 16:06:17 -07002124static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002125 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002126{
2127 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002128 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002129 unsigned long nr_to_scan;
2130 enum lru_list lru;
2131 unsigned long nr_reclaimed = 0;
2132 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2133 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002134 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002135
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002136 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002137
Mel Gormane82e0562013-07-03 15:01:44 -07002138 /* Record the original scan target for proportional adjustments later */
2139 memcpy(targets, nr, sizeof(nr));
2140
Mel Gorman1a501902014-06-04 16:10:49 -07002141 /*
2142 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2143 * event that can occur when there is little memory pressure e.g.
2144 * multiple streaming readers/writers. Hence, we do not abort scanning
2145 * when the requested number of pages are reclaimed when scanning at
2146 * DEF_PRIORITY on the assumption that the fact we are direct
2147 * reclaiming implies that kswapd is not keeping up and it is best to
2148 * do a batch of work at once. For memcg reclaim one check is made to
2149 * abort proportional reclaim if either the file or anon lru has already
2150 * dropped to zero at the first pass.
2151 */
2152 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2153 sc->priority == DEF_PRIORITY);
2154
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002155 blk_start_plug(&plug);
2156 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2157 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002158 unsigned long nr_anon, nr_file, percentage;
2159 unsigned long nr_scanned;
2160
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002161 for_each_evictable_lru(lru) {
2162 if (nr[lru]) {
2163 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2164 nr[lru] -= nr_to_scan;
2165
2166 nr_reclaimed += shrink_list(lru, nr_to_scan,
2167 lruvec, sc);
2168 }
2169 }
Mel Gormane82e0562013-07-03 15:01:44 -07002170
2171 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2172 continue;
2173
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002174 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002175 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002176 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002177 * proportionally what was requested by get_scan_count(). We
2178 * stop reclaiming one LRU and reduce the amount scanning
2179 * proportional to the original scan target.
2180 */
2181 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2182 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2183
Mel Gorman1a501902014-06-04 16:10:49 -07002184 /*
2185 * It's just vindictive to attack the larger once the smaller
2186 * has gone to zero. And given the way we stop scanning the
2187 * smaller below, this makes sure that we only make one nudge
2188 * towards proportionality once we've got nr_to_reclaim.
2189 */
2190 if (!nr_file || !nr_anon)
2191 break;
2192
Mel Gormane82e0562013-07-03 15:01:44 -07002193 if (nr_file > nr_anon) {
2194 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2195 targets[LRU_ACTIVE_ANON] + 1;
2196 lru = LRU_BASE;
2197 percentage = nr_anon * 100 / scan_target;
2198 } else {
2199 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2200 targets[LRU_ACTIVE_FILE] + 1;
2201 lru = LRU_FILE;
2202 percentage = nr_file * 100 / scan_target;
2203 }
2204
2205 /* Stop scanning the smaller of the LRU */
2206 nr[lru] = 0;
2207 nr[lru + LRU_ACTIVE] = 0;
2208
2209 /*
2210 * Recalculate the other LRU scan count based on its original
2211 * scan target and the percentage scanning already complete
2212 */
2213 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2214 nr_scanned = targets[lru] - nr[lru];
2215 nr[lru] = targets[lru] * (100 - percentage) / 100;
2216 nr[lru] -= min(nr[lru], nr_scanned);
2217
2218 lru += LRU_ACTIVE;
2219 nr_scanned = targets[lru] - nr[lru];
2220 nr[lru] = targets[lru] * (100 - percentage) / 100;
2221 nr[lru] -= min(nr[lru], nr_scanned);
2222
2223 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002224 }
2225 blk_finish_plug(&plug);
2226 sc->nr_reclaimed += nr_reclaimed;
2227
2228 /*
2229 * Even if we did not try to evict anon pages at all, we want to
2230 * rebalance the anon lru active/inactive ratio.
2231 */
2232 if (inactive_anon_is_low(lruvec))
2233 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2234 sc, LRU_ACTIVE_ANON);
2235
2236 throttle_vm_writeout(sc->gfp_mask);
2237}
2238
Mel Gorman23b9da52012-05-29 15:06:20 -07002239/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002240static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002241{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002242 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002243 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002244 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002245 return true;
2246
2247 return false;
2248}
2249
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002250/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002251 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2252 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2253 * true if more pages should be reclaimed such that when the page allocator
2254 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2255 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002256 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002257static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002258 unsigned long nr_reclaimed,
2259 unsigned long nr_scanned,
2260 struct scan_control *sc)
2261{
2262 unsigned long pages_for_compaction;
2263 unsigned long inactive_lru_pages;
2264
2265 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002266 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002267 return false;
2268
Mel Gorman28765922011-02-25 14:44:20 -08002269 /* Consider stopping depending on scan and reclaim activity */
2270 if (sc->gfp_mask & __GFP_REPEAT) {
2271 /*
2272 * For __GFP_REPEAT allocations, stop reclaiming if the
2273 * full LRU list has been scanned and we are still failing
2274 * to reclaim pages. This full LRU scan is potentially
2275 * expensive but a __GFP_REPEAT caller really wants to succeed
2276 */
2277 if (!nr_reclaimed && !nr_scanned)
2278 return false;
2279 } else {
2280 /*
2281 * For non-__GFP_REPEAT allocations which can presumably
2282 * fail without consequence, stop if we failed to reclaim
2283 * any pages from the last SWAP_CLUSTER_MAX number of
2284 * pages that were scanned. This will return to the
2285 * caller faster at the risk reclaim/compaction and
2286 * the resulting allocation attempt fails
2287 */
2288 if (!nr_reclaimed)
2289 return false;
2290 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002291
2292 /*
2293 * If we have not reclaimed enough pages for compaction and the
2294 * inactive lists are large enough, continue reclaiming
2295 */
2296 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002297 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002298 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002299 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002300 if (sc->nr_reclaimed < pages_for_compaction &&
2301 inactive_lru_pages > pages_for_compaction)
2302 return true;
2303
2304 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002305 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002306 case COMPACT_PARTIAL:
2307 case COMPACT_CONTINUE:
2308 return false;
2309 default:
2310 return true;
2311 }
2312}
2313
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002314static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2315 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002316{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002317 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002318 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002319 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002320
Johannes Weiner56600482012-01-12 17:17:59 -08002321 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002322 struct mem_cgroup *root = sc->target_mem_cgroup;
2323 struct mem_cgroup_reclaim_cookie reclaim = {
2324 .zone = zone,
2325 .priority = sc->priority,
2326 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002327 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002328 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002329
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002330 nr_reclaimed = sc->nr_reclaimed;
2331 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002332
Andrew Morton694fbc02013-09-24 15:27:37 -07002333 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2334 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002335 unsigned long lru_pages;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002336 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002337 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002338 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002339
Johannes Weiner241994ed2015-02-11 15:26:06 -08002340 if (mem_cgroup_low(root, memcg)) {
2341 if (!sc->may_thrash)
2342 continue;
2343 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2344 }
2345
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002346 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002347 swappiness = mem_cgroup_swappiness(memcg);
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002348 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002349
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002350 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2351 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002352
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002353 if (memcg && is_classzone)
2354 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2355 memcg, sc->nr_scanned - scanned,
2356 lru_pages);
2357
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002358 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002359 * Direct reclaim and kswapd have to scan all memory
2360 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002361 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002362 *
2363 * Limit reclaim, on the other hand, only cares about
2364 * nr_to_reclaim pages to be reclaimed and it will
2365 * retry with decreasing priority if one round over the
2366 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002367 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002368 if (!global_reclaim(sc) &&
2369 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002370 mem_cgroup_iter_break(root, memcg);
2371 break;
2372 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002373 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002374
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002375 /*
2376 * Shrink the slab caches in the same proportion that
2377 * the eligible LRU pages were scanned.
2378 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002379 if (global_reclaim(sc) && is_classzone)
2380 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2381 sc->nr_scanned - nr_scanned,
2382 zone_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002383
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002384 if (reclaim_state) {
2385 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2386 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002387 }
2388
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002389 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2390 sc->nr_scanned - nr_scanned,
2391 sc->nr_reclaimed - nr_reclaimed);
2392
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002393 if (sc->nr_reclaimed - nr_reclaimed)
2394 reclaimable = true;
2395
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002396 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2397 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002398
2399 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002400}
2401
Vlastimil Babka53853e22014-10-09 15:27:02 -07002402/*
2403 * Returns true if compaction should go ahead for a high-order request, or
2404 * the high-order allocation would succeed without compaction.
2405 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002406static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002407{
2408 unsigned long balance_gap, watermark;
2409 bool watermark_ok;
2410
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002411 /*
2412 * Compaction takes time to run and there are potentially other
2413 * callers using the pages just freed. Continue reclaiming until
2414 * there is a buffer of free pages available to give compaction
2415 * a reasonable chance of completing and allocating the page
2416 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002417 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2418 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002419 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002420 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2421
2422 /*
2423 * If compaction is deferred, reclaim up to a point where
2424 * compaction will have a chance of success when re-enabled
2425 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002426 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002427 return watermark_ok;
2428
Vlastimil Babka53853e22014-10-09 15:27:02 -07002429 /*
2430 * If compaction is not ready to start and allocation is not likely
2431 * to succeed without it, then keep reclaiming.
2432 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002433 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002434 return false;
2435
2436 return watermark_ok;
2437}
2438
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439/*
2440 * This is the direct reclaim path, for page-allocating processes. We only
2441 * try to reclaim pages from zones which will satisfy the caller's allocation
2442 * request.
2443 *
Mel Gorman41858962009-06-16 15:32:12 -07002444 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2445 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2447 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002448 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2449 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2450 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 * If a zone is deemed to be full of pinned pages then just give it a light
2453 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002454 *
2455 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002457static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458{
Mel Gormandd1a2392008-04-28 02:12:17 -07002459 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002460 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002461 unsigned long nr_soft_reclaimed;
2462 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002463 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002464 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002465 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002466
Mel Gormancc715d92012-03-21 16:34:00 -07002467 /*
2468 * If the number of buffer_heads in the machine exceeds the maximum
2469 * allowed level, force direct reclaim to scan the highmem zone as
2470 * highmem pages could be pinning lowmem pages storing buffer_heads
2471 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002472 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002473 if (buffer_heads_over_limit)
2474 sc->gfp_mask |= __GFP_HIGHMEM;
2475
Mel Gormand4debc62010-08-09 17:19:29 -07002476 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002477 requested_highidx, sc->nodemask) {
2478 enum zone_type classzone_idx;
2479
Con Kolivasf3fe6512006-01-06 00:11:15 -08002480 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002482
2483 classzone_idx = requested_highidx;
2484 while (!populated_zone(zone->zone_pgdat->node_zones +
2485 classzone_idx))
2486 classzone_idx--;
2487
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002488 /*
2489 * Take care memory controller reclaiming has small influence
2490 * to global LRU.
2491 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002492 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002493 if (!cpuset_zone_allowed(zone,
2494 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002495 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002496
Lisa Du6e543d52013-09-11 14:22:36 -07002497 if (sc->priority != DEF_PRIORITY &&
2498 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002499 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002500
2501 /*
2502 * If we already have plenty of memory free for
2503 * compaction in this zone, don't free any more.
2504 * Even though compaction is invoked for any
2505 * non-zero order, only frequent costly order
2506 * reclamation is disruptive enough to become a
2507 * noticeable problem, like transparent huge
2508 * page allocations.
2509 */
2510 if (IS_ENABLED(CONFIG_COMPACTION) &&
2511 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2512 zonelist_zone_idx(z) <= requested_highidx &&
2513 compaction_ready(zone, sc->order)) {
2514 sc->compaction_ready = true;
2515 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002516 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002517
Andrew Morton0608f432013-09-24 15:27:41 -07002518 /*
2519 * This steals pages from memory cgroups over softlimit
2520 * and returns the number of reclaimed pages and
2521 * scanned pages. This works for global memory pressure
2522 * and balancing, not for a memcg's limit.
2523 */
2524 nr_soft_scanned = 0;
2525 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2526 sc->order, sc->gfp_mask,
2527 &nr_soft_scanned);
2528 sc->nr_reclaimed += nr_soft_reclaimed;
2529 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002530 if (nr_soft_reclaimed)
2531 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002532 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002533 }
Nick Piggin408d8542006-09-25 23:31:27 -07002534
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002535 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002536 reclaimable = true;
2537
2538 if (global_reclaim(sc) &&
2539 !reclaimable && zone_reclaimable(zone))
2540 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 }
Mel Gormane0c23272011-10-31 17:09:33 -07002542
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002543 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002544 * Restore to original mask to avoid the impact on the caller if we
2545 * promoted it to __GFP_HIGHMEM.
2546 */
2547 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002548
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002549 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002551
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552/*
2553 * This is the main entry point to direct page reclaim.
2554 *
2555 * If a full scan of the inactive list fails to free enough memory then we
2556 * are "out of memory" and something needs to be killed.
2557 *
2558 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2559 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002560 * caller can't do much about. We kick the writeback threads and take explicit
2561 * naps in the hope that some of these pages can be written. But if the
2562 * allocating task holds filesystem locks which prevent writeout this might not
2563 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002564 *
2565 * returns: 0, if no pages reclaimed
2566 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 */
Mel Gormandac1d272008-04-28 02:12:12 -07002568static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002569 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002571 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002572 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002573 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002574 bool zones_reclaimable;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002575retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002576 delayacct_freepages_start();
2577
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002578 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002579 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002581 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002582 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2583 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002584 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002585 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002586
Balbir Singh66e17072008-02-07 00:13:56 -08002587 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002588 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002589 break;
2590
2591 if (sc->compaction_ready)
2592 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
2594 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002595 * If we're getting trouble reclaiming, start doing
2596 * writepage even in laptop mode.
2597 */
2598 if (sc->priority < DEF_PRIORITY - 2)
2599 sc->may_writepage = 1;
2600
2601 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 * Try to write back as many pages as we just scanned. This
2603 * tends to cause slow streaming writers to write data to the
2604 * disk smoothly, at the dirtying rate, which is nice. But
2605 * that's undesirable in laptop mode, where we *want* lumpy
2606 * writeout. So in laptop mode, write out the whole world.
2607 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002608 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2609 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002610 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2611 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002612 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002614 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002615
Keika Kobayashi873b4772008-07-25 01:48:52 -07002616 delayacct_freepages_end();
2617
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002618 if (sc->nr_reclaimed)
2619 return sc->nr_reclaimed;
2620
Mel Gorman0cee34f2012-01-12 17:19:49 -08002621 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002622 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002623 return 1;
2624
Johannes Weiner241994ed2015-02-11 15:26:06 -08002625 /* Untapped cgroup reserves? Don't OOM, retry. */
2626 if (!sc->may_thrash) {
2627 sc->priority = initial_priority;
2628 sc->may_thrash = 1;
2629 goto retry;
2630 }
2631
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002632 /* Any of the zones still reclaimable? Don't OOM. */
2633 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002634 return 1;
2635
2636 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637}
2638
Mel Gorman55150612012-07-31 16:44:35 -07002639static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2640{
2641 struct zone *zone;
2642 unsigned long pfmemalloc_reserve = 0;
2643 unsigned long free_pages = 0;
2644 int i;
2645 bool wmark_ok;
2646
2647 for (i = 0; i <= ZONE_NORMAL; i++) {
2648 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002649 if (!populated_zone(zone))
2650 continue;
2651
Mel Gorman55150612012-07-31 16:44:35 -07002652 pfmemalloc_reserve += min_wmark_pages(zone);
2653 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2654 }
2655
Mel Gorman675becc2014-06-04 16:07:35 -07002656 /* If there are no reserves (unexpected config) then do not throttle */
2657 if (!pfmemalloc_reserve)
2658 return true;
2659
Mel Gorman55150612012-07-31 16:44:35 -07002660 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2661
2662 /* kswapd must be awake if processes are being throttled */
2663 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2664 pgdat->classzone_idx = min(pgdat->classzone_idx,
2665 (enum zone_type)ZONE_NORMAL);
2666 wake_up_interruptible(&pgdat->kswapd_wait);
2667 }
2668
2669 return wmark_ok;
2670}
2671
2672/*
2673 * Throttle direct reclaimers if backing storage is backed by the network
2674 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2675 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002676 * when the low watermark is reached.
2677 *
2678 * Returns true if a fatal signal was delivered during throttling. If this
2679 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002680 */
Mel Gorman50694c22012-11-26 16:29:48 -08002681static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002682 nodemask_t *nodemask)
2683{
Mel Gorman675becc2014-06-04 16:07:35 -07002684 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002685 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002686 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002687
2688 /*
2689 * Kernel threads should not be throttled as they may be indirectly
2690 * responsible for cleaning pages necessary for reclaim to make forward
2691 * progress. kjournald for example may enter direct reclaim while
2692 * committing a transaction where throttling it could forcing other
2693 * processes to block on log_wait_commit().
2694 */
2695 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002696 goto out;
2697
2698 /*
2699 * If a fatal signal is pending, this process should not throttle.
2700 * It should return quickly so it can exit and free its memory
2701 */
2702 if (fatal_signal_pending(current))
2703 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002704
Mel Gorman675becc2014-06-04 16:07:35 -07002705 /*
2706 * Check if the pfmemalloc reserves are ok by finding the first node
2707 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2708 * GFP_KERNEL will be required for allocating network buffers when
2709 * swapping over the network so ZONE_HIGHMEM is unusable.
2710 *
2711 * Throttling is based on the first usable node and throttled processes
2712 * wait on a queue until kswapd makes progress and wakes them. There
2713 * is an affinity then between processes waking up and where reclaim
2714 * progress has been made assuming the process wakes on the same node.
2715 * More importantly, processes running on remote nodes will not compete
2716 * for remote pfmemalloc reserves and processes on different nodes
2717 * should make reasonable progress.
2718 */
2719 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002720 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002721 if (zone_idx(zone) > ZONE_NORMAL)
2722 continue;
2723
2724 /* Throttle based on the first usable node */
2725 pgdat = zone->zone_pgdat;
2726 if (pfmemalloc_watermark_ok(pgdat))
2727 goto out;
2728 break;
2729 }
2730
2731 /* If no zone was usable by the allocation flags then do not throttle */
2732 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002733 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002734
Mel Gorman68243e72012-07-31 16:44:39 -07002735 /* Account for the throttling */
2736 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2737
Mel Gorman55150612012-07-31 16:44:35 -07002738 /*
2739 * If the caller cannot enter the filesystem, it's possible that it
2740 * is due to the caller holding an FS lock or performing a journal
2741 * transaction in the case of a filesystem like ext[3|4]. In this case,
2742 * it is not safe to block on pfmemalloc_wait as kswapd could be
2743 * blocked waiting on the same lock. Instead, throttle for up to a
2744 * second before continuing.
2745 */
2746 if (!(gfp_mask & __GFP_FS)) {
2747 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2748 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002749
2750 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002751 }
2752
2753 /* Throttle until kswapd wakes the process */
2754 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2755 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002756
2757check_pending:
2758 if (fatal_signal_pending(current))
2759 return true;
2760
2761out:
2762 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002763}
2764
Mel Gormandac1d272008-04-28 02:12:12 -07002765unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002766 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002767{
Mel Gorman33906bc2010-08-09 17:19:16 -07002768 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002769 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002770 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002771 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2772 .order = order,
2773 .nodemask = nodemask,
2774 .priority = DEF_PRIORITY,
2775 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002776 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002777 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002778 };
2779
Mel Gorman55150612012-07-31 16:44:35 -07002780 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002781 * Do not enter reclaim if fatal signal was delivered while throttled.
2782 * 1 is returned so that the page allocator does not OOM kill at this
2783 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002784 */
Mel Gorman50694c22012-11-26 16:29:48 -08002785 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002786 return 1;
2787
Mel Gorman33906bc2010-08-09 17:19:16 -07002788 trace_mm_vmscan_direct_reclaim_begin(order,
2789 sc.may_writepage,
2790 gfp_mask);
2791
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002792 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002793
2794 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2795
2796 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002797}
2798
Andrew Mortonc255a452012-07-31 16:43:02 -07002799#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002800
Johannes Weiner72835c82012-01-12 17:18:32 -08002801unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002802 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002803 struct zone *zone,
2804 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002805{
2806 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002807 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002808 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002809 .may_writepage = !laptop_mode,
2810 .may_unmap = 1,
2811 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002812 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002813 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002814 int swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002815 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002816
Balbir Singh4e416952009-09-23 15:56:39 -07002817 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2818 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002819
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002820 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002821 sc.may_writepage,
2822 sc.gfp_mask);
2823
Balbir Singh4e416952009-09-23 15:56:39 -07002824 /*
2825 * NOTE: Although we can get the priority field, using it
2826 * here is not a good idea, since it limits the pages we can scan.
2827 * if we don't reclaim here, the shrink_zone from balance_pgdat
2828 * will pick up pages from other mem cgroup's as well. We hack
2829 * the priority and make it zero.
2830 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002831 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002832
2833 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2834
Ying Han0ae5e892011-05-26 16:25:25 -07002835 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002836 return sc.nr_reclaimed;
2837}
2838
Johannes Weiner72835c82012-01-12 17:18:32 -08002839unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002840 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002841 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002842 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002843{
Balbir Singh4e416952009-09-23 15:56:39 -07002844 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002845 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002846 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002847 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002848 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002849 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2850 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2851 .target_mem_cgroup = memcg,
2852 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002853 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002854 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002855 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002856 };
Balbir Singh66e17072008-02-07 00:13:56 -08002857
Ying Han889976d2011-05-26 16:25:33 -07002858 /*
2859 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2860 * take care of from where we get pages. So the node where we start the
2861 * scan does not need to be the current node.
2862 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002863 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002864
2865 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002866
2867 trace_mm_vmscan_memcg_reclaim_begin(0,
2868 sc.may_writepage,
2869 sc.gfp_mask);
2870
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002871 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002872
2873 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2874
2875 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002876}
2877#endif
2878
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002879static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002880{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002881 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002882
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002883 if (!total_swap_pages)
2884 return;
2885
2886 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2887 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002888 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002889
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002890 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002891 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002892 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002893
2894 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2895 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002896}
2897
Johannes Weiner60cefed2012-11-29 13:54:23 -08002898static bool zone_balanced(struct zone *zone, int order,
2899 unsigned long balance_gap, int classzone_idx)
2900{
2901 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2902 balance_gap, classzone_idx, 0))
2903 return false;
2904
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002905 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2906 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002907 return false;
2908
2909 return true;
2910}
2911
Mel Gorman1741c872011-01-13 15:46:21 -08002912/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002913 * pgdat_balanced() is used when checking if a node is balanced.
2914 *
2915 * For order-0, all zones must be balanced!
2916 *
2917 * For high-order allocations only zones that meet watermarks and are in a
2918 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2919 * total of balanced pages must be at least 25% of the zones allowed by
2920 * classzone_idx for the node to be considered balanced. Forcing all zones to
2921 * be balanced for high orders can cause excessive reclaim when there are
2922 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002923 * The choice of 25% is due to
2924 * o a 16M DMA zone that is balanced will not balance a zone on any
2925 * reasonable sized machine
2926 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002927 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002928 * would need to be at least 256M for it to be balance a whole node.
2929 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2930 * to balance a node on its own. These seemed like reasonable ratios.
2931 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002932static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002933{
Jiang Liub40da042013-02-22 16:33:52 -08002934 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002935 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002936 int i;
2937
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002938 /* Check the watermark levels */
2939 for (i = 0; i <= classzone_idx; i++) {
2940 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002941
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002942 if (!populated_zone(zone))
2943 continue;
2944
Jiang Liub40da042013-02-22 16:33:52 -08002945 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002946
2947 /*
2948 * A special case here:
2949 *
2950 * balance_pgdat() skips over all_unreclaimable after
2951 * DEF_PRIORITY. Effectively, it considers them balanced so
2952 * they must be considered balanced here as well!
2953 */
Lisa Du6e543d52013-09-11 14:22:36 -07002954 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002955 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002956 continue;
2957 }
2958
2959 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002960 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002961 else if (!order)
2962 return false;
2963 }
2964
2965 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002966 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002967 else
2968 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002969}
2970
Mel Gorman55150612012-07-31 16:44:35 -07002971/*
2972 * Prepare kswapd for sleeping. This verifies that there are no processes
2973 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2974 *
2975 * Returns true if kswapd is ready to sleep
2976 */
2977static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002978 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002979{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002980 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2981 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002982 return false;
2983
2984 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002985 * The throttled processes are normally woken up in balance_pgdat() as
2986 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2987 * race between when kswapd checks the watermarks and a process gets
2988 * throttled. There is also a potential race if processes get
2989 * throttled, kswapd wakes, a large process exits thereby balancing the
2990 * zones, which causes kswapd to exit balance_pgdat() before reaching
2991 * the wake up checks. If kswapd is going to sleep, no process should
2992 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2993 * the wake up is premature, processes will wake kswapd and get
2994 * throttled again. The difference from wake ups in balance_pgdat() is
2995 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07002996 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002997 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2998 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002999
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003000 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003001}
3002
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003/*
Mel Gorman75485362013-07-03 15:01:42 -07003004 * kswapd shrinks the zone by the number of pages required to reach
3005 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003006 *
3007 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003008 * reclaim or if the lack of progress was due to pages under writeback.
3009 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003010 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003011static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07003012 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07003013 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07003014 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07003015{
Mel Gorman7c954f62013-07-03 15:01:54 -07003016 int testorder = sc->order;
3017 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07003018 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07003019
3020 /* Reclaim above the high watermark. */
3021 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07003022
3023 /*
3024 * Kswapd reclaims only single pages with compaction enabled. Trying
3025 * too hard to reclaim until contiguous free pages have become
3026 * available can hurt performance by evicting too much useful data
3027 * from memory. Do not reclaim more than needed for compaction.
3028 */
3029 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08003030 compaction_suitable(zone, sc->order, 0, classzone_idx)
3031 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07003032 testorder = 0;
3033
3034 /*
3035 * We put equal pressure on every zone, unless one zone has way too
3036 * many pages free already. The "too many pages" is defined as the
3037 * high wmark plus a "gap" where the gap is either the low
3038 * watermark or 1% of the zone, whichever is smaller.
3039 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07003040 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3041 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003042
3043 /*
3044 * If there is no low memory pressure or the zone is balanced then no
3045 * reclaim is necessary
3046 */
3047 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3048 if (!lowmem_pressure && zone_balanced(zone, testorder,
3049 balance_gap, classzone_idx))
3050 return true;
3051
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003052 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003053
Mel Gorman2ab44f42013-07-03 15:01:47 -07003054 /* Account for the number of pages attempted to reclaim */
3055 *nr_attempted += sc->nr_to_reclaim;
3056
Johannes Weiner57054652014-10-09 15:28:17 -07003057 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003058
Mel Gorman7c954f62013-07-03 15:01:54 -07003059 /*
3060 * If a zone reaches its high watermark, consider it to be no longer
3061 * congested. It's possible there are dirty pages backed by congested
3062 * BDIs but as pressure is relieved, speculatively avoid congestion
3063 * waits.
3064 */
Lisa Du6e543d52013-09-11 14:22:36 -07003065 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003066 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003067 clear_bit(ZONE_CONGESTED, &zone->flags);
3068 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003069 }
3070
Mel Gormanb8e83b92013-07-03 15:01:45 -07003071 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003072}
3073
3074/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003076 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003078 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 *
3080 * There is special handling here for zones which are full of pinned pages.
3081 * This can happen if the pages are all mlocked, or if they are all used by
3082 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3083 * What we do is to detect the case where all pages in the zone have been
3084 * scanned twice and there has been zero successful reclaim. Mark the zone as
3085 * dead and from now on, only perform a short scan. Basically we're polling
3086 * the zone for when the problem goes away.
3087 *
3088 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003089 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3090 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3091 * lower zones regardless of the number of free pages in the lower zones. This
3092 * interoperates with the page allocator fallback scheme to ensure that aging
3093 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 */
Mel Gorman99504742011-01-13 15:46:20 -08003095static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003096 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003099 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003100 unsigned long nr_soft_reclaimed;
3101 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003102 struct scan_control sc = {
3103 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003104 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003105 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003106 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003107 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003108 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003109 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003110 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003112 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003113 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003114 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003115 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003116
3117 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003119 /*
3120 * Scan in the highmem->dma direction for the highest
3121 * zone which needs scanning
3122 */
3123 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3124 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003126 if (!populated_zone(zone))
3127 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128
Lisa Du6e543d52013-09-11 14:22:36 -07003129 if (sc.priority != DEF_PRIORITY &&
3130 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003131 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132
Rik van Riel556adec2008-10-18 20:26:34 -07003133 /*
3134 * Do some background aging of the anon list, to give
3135 * pages a chance to be referenced before reclaiming.
3136 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003137 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003138
Mel Gormancc715d92012-03-21 16:34:00 -07003139 /*
3140 * If the number of buffer_heads in the machine
3141 * exceeds the maximum allowed level and this node
3142 * has a highmem zone, force kswapd to reclaim from
3143 * it to relieve lowmem pressure.
3144 */
3145 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3146 end_zone = i;
3147 break;
3148 }
3149
Johannes Weiner60cefed2012-11-29 13:54:23 -08003150 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003151 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003152 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003153 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003154 /*
3155 * If balanced, clear the dirty and congested
3156 * flags
3157 */
Johannes Weiner57054652014-10-09 15:28:17 -07003158 clear_bit(ZONE_CONGESTED, &zone->flags);
3159 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003162
Mel Gormanb8e83b92013-07-03 15:01:45 -07003163 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003164 goto out;
3165
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 for (i = 0; i <= end_zone; i++) {
3167 struct zone *zone = pgdat->node_zones + i;
3168
Mel Gorman2ab44f42013-07-03 15:01:47 -07003169 if (!populated_zone(zone))
3170 continue;
3171
Mel Gorman2ab44f42013-07-03 15:01:47 -07003172 /*
3173 * If any zone is currently balanced then kswapd will
3174 * not call compaction as it is expected that the
3175 * necessary pages are already available.
3176 */
3177 if (pgdat_needs_compaction &&
3178 zone_watermark_ok(zone, order,
3179 low_wmark_pages(zone),
3180 *classzone_idx, 0))
3181 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 }
3183
3184 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003185 * If we're getting trouble reclaiming, start doing writepage
3186 * even in laptop mode.
3187 */
3188 if (sc.priority < DEF_PRIORITY - 2)
3189 sc.may_writepage = 1;
3190
3191 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192 * Now scan the zone in the dma->highmem direction, stopping
3193 * at the last zone which needs scanning.
3194 *
3195 * We do this because the page allocator works in the opposite
3196 * direction. This prevents the page allocator from allocating
3197 * pages behind kswapd's direction of progress, which would
3198 * cause too much scanning of the lower zones.
3199 */
3200 for (i = 0; i <= end_zone; i++) {
3201 struct zone *zone = pgdat->node_zones + i;
3202
Con Kolivasf3fe6512006-01-06 00:11:15 -08003203 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 continue;
3205
Lisa Du6e543d52013-09-11 14:22:36 -07003206 if (sc.priority != DEF_PRIORITY &&
3207 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 continue;
3209
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003211
Andrew Morton0608f432013-09-24 15:27:41 -07003212 nr_soft_scanned = 0;
3213 /*
3214 * Call soft limit reclaim before calling shrink_zone.
3215 */
3216 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3217 order, sc.gfp_mask,
3218 &nr_soft_scanned);
3219 sc.nr_reclaimed += nr_soft_reclaimed;
3220
Rik van Riel32a43302007-10-16 01:24:50 -07003221 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003222 * There should be no need to raise the scanning
3223 * priority if enough pages are already being scanned
3224 * that that high watermark would be met at 100%
3225 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003226 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003227 if (kswapd_shrink_zone(zone, end_zone,
3228 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003229 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 }
Mel Gorman55150612012-07-31 16:44:35 -07003231
3232 /*
3233 * If the low watermark is met there is no need for processes
3234 * to be throttled on pfmemalloc_wait as they should not be
3235 * able to safely make forward progress. Wake them
3236 */
3237 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3238 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003239 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003240
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003242 * Fragmentation may mean that the system cannot be rebalanced
3243 * for high-order allocations in all zones. If twice the
3244 * allocation size has been reclaimed and the zones are still
3245 * not balanced then recheck the watermarks at order-0 to
3246 * prevent kswapd reclaiming excessively. Assume that a
3247 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003249 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003250 order = sc.order = 0;
3251
Mel Gormanb8e83b92013-07-03 15:01:45 -07003252 /* Check if kswapd should be suspending */
3253 if (try_to_freeze() || kthread_should_stop())
3254 break;
3255
3256 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003257 * Compact if necessary and kswapd is reclaiming at least the
3258 * high watermark number of pages as requsted
3259 */
3260 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3261 compact_pgdat(pgdat, order);
3262
3263 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003264 * Raise priority if scanning rate is too low or there was no
3265 * progress in reclaiming pages
3266 */
3267 if (raise_priority || !sc.nr_reclaimed)
3268 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003269 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003270 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271
Mel Gormanb8e83b92013-07-03 15:01:45 -07003272out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003273 /*
Mel Gorman55150612012-07-31 16:44:35 -07003274 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003275 * makes a decision on the order we were last reclaiming at. However,
3276 * if another caller entered the allocator slow path while kswapd
3277 * was awake, order will remain at the higher level
3278 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003279 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003280 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281}
3282
Mel Gormandc83edd2011-01-13 15:46:26 -08003283static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003284{
3285 long remaining = 0;
3286 DEFINE_WAIT(wait);
3287
3288 if (freezing(current) || kthread_should_stop())
3289 return;
3290
3291 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3292
3293 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003294 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003295 remaining = schedule_timeout(HZ/10);
3296 finish_wait(&pgdat->kswapd_wait, &wait);
3297 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3298 }
3299
3300 /*
3301 * After a short sleep, check if it was a premature sleep. If not, then
3302 * go fully to sleep until explicitly woken up.
3303 */
Mel Gorman55150612012-07-31 16:44:35 -07003304 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003305 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3306
3307 /*
3308 * vmstat counters are not perfectly accurate and the estimated
3309 * value for counters such as NR_FREE_PAGES can deviate from the
3310 * true value by nr_online_cpus * threshold. To avoid the zone
3311 * watermarks being breached while under pressure, we reduce the
3312 * per-cpu vmstat threshold while kswapd is awake and restore
3313 * them before going back to sleep.
3314 */
3315 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003316
Mel Gorman62997022012-10-08 16:32:47 -07003317 /*
3318 * Compaction records what page blocks it recently failed to
3319 * isolate pages from and skips them in the future scanning.
3320 * When kswapd is going to sleep, it is reasonable to assume
3321 * that pages and compaction may succeed so reset the cache.
3322 */
3323 reset_isolation_suitable(pgdat);
3324
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003325 if (!kthread_should_stop())
3326 schedule();
3327
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003328 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3329 } else {
3330 if (remaining)
3331 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3332 else
3333 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3334 }
3335 finish_wait(&pgdat->kswapd_wait, &wait);
3336}
3337
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338/*
3339 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003340 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341 *
3342 * This basically trickles out pages so that we have _some_
3343 * free memory available even if there is no other activity
3344 * that frees anything up. This is needed for things like routing
3345 * etc, where we otherwise might have all activity going on in
3346 * asynchronous contexts that cannot page things out.
3347 *
3348 * If there are applications that are active memory-allocators
3349 * (most normal use), this basically shouldn't matter.
3350 */
3351static int kswapd(void *p)
3352{
Mel Gorman215ddd62011-07-08 15:39:40 -07003353 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003354 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003355 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003356 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 pg_data_t *pgdat = (pg_data_t*)p;
3358 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003359
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 struct reclaim_state reclaim_state = {
3361 .reclaimed_slab = 0,
3362 };
Rusty Russella70f7302009-03-13 14:49:46 +10303363 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364
Nick Piggincf40bd12009-01-21 08:12:39 +01003365 lockdep_set_current_reclaim_state(GFP_KERNEL);
3366
Rusty Russell174596a2009-01-01 10:12:29 +10303367 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003368 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369 current->reclaim_state = &reclaim_state;
3370
3371 /*
3372 * Tell the memory management that we're a "memory allocator",
3373 * and that if we need more memory we should get access to it
3374 * regardless (see "__alloc_pages()"). "kswapd" should
3375 * never get caught in the normal page freeing logic.
3376 *
3377 * (Kswapd normally doesn't need memory anyway, but sometimes
3378 * you need a small amount of memory in order to be able to
3379 * page out something else, and this flag essentially protects
3380 * us from recursively trying to free more memory as we're
3381 * trying to free the first piece of memory in the first place).
3382 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003383 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003384 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385
Mel Gorman215ddd62011-07-08 15:39:40 -07003386 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003387 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003388 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003389 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003391 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003392
Mel Gorman215ddd62011-07-08 15:39:40 -07003393 /*
3394 * If the last balance_pgdat was unsuccessful it's unlikely a
3395 * new request of a similar or harder type will succeed soon
3396 * so consider going to sleep on the basis we reclaimed at
3397 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003398 if (balanced_classzone_idx >= new_classzone_idx &&
3399 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003400 new_order = pgdat->kswapd_max_order;
3401 new_classzone_idx = pgdat->classzone_idx;
3402 pgdat->kswapd_max_order = 0;
3403 pgdat->classzone_idx = pgdat->nr_zones - 1;
3404 }
3405
Mel Gorman99504742011-01-13 15:46:20 -08003406 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 /*
3408 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003409 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 */
3411 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003412 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003414 kswapd_try_to_sleep(pgdat, balanced_order,
3415 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003417 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003418 new_order = order;
3419 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003420 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003421 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
David Rientjes8fe23e02009-12-14 17:58:33 -08003424 ret = try_to_freeze();
3425 if (kthread_should_stop())
3426 break;
3427
3428 /*
3429 * We can speed up thawing tasks if we don't call balance_pgdat
3430 * after returning from the refrigerator
3431 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003432 if (!ret) {
3433 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003434 balanced_classzone_idx = classzone_idx;
3435 balanced_order = balance_pgdat(pgdat, order,
3436 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003439
Johannes Weiner71abdc12014-06-06 14:35:35 -07003440 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003441 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003442 lockdep_clear_current_reclaim_state();
3443
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 return 0;
3445}
3446
3447/*
3448 * A zone is low on free memory, so wake its kswapd task to service it.
3449 */
Mel Gorman99504742011-01-13 15:46:20 -08003450void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451{
3452 pg_data_t *pgdat;
3453
Con Kolivasf3fe6512006-01-06 00:11:15 -08003454 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455 return;
3456
Vladimir Davydov344736f2014-10-20 15:50:30 +04003457 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003459 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003460 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003461 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003462 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3463 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003464 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003466 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003467 return;
3468
3469 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003470 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471}
3472
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003473#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003475 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003476 * freed pages.
3477 *
3478 * Rather than trying to age LRUs the aim is to preserve the overall
3479 * LRU order by reclaiming preferentially
3480 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003482unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003484 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003485 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003486 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003487 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003488 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003489 .may_writepage = 1,
3490 .may_unmap = 1,
3491 .may_swap = 1,
3492 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003494 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003495 struct task_struct *p = current;
3496 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003498 p->flags |= PF_MEMALLOC;
3499 lockdep_set_current_reclaim_state(sc.gfp_mask);
3500 reclaim_state.reclaimed_slab = 0;
3501 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003502
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003503 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003504
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003505 p->reclaim_state = NULL;
3506 lockdep_clear_current_reclaim_state();
3507 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003508
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003509 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003511#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513/* It's optimal to keep kswapds on the same CPUs as their memory, but
3514 not required for correctness. So if the last cpu in a node goes
3515 away, we get changed to run anywhere: as the first one comes back,
3516 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003517static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3518 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003520 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003522 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003523 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003524 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303525 const struct cpumask *mask;
3526
3527 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003528
Rusty Russell3e597942009-01-01 10:12:24 +10303529 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003531 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 }
3533 }
3534 return NOTIFY_OK;
3535}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
Yasunori Goto3218ae12006-06-27 02:53:33 -07003537/*
3538 * This kswapd start function will be called by init and node-hot-add.
3539 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3540 */
3541int kswapd_run(int nid)
3542{
3543 pg_data_t *pgdat = NODE_DATA(nid);
3544 int ret = 0;
3545
3546 if (pgdat->kswapd)
3547 return 0;
3548
3549 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3550 if (IS_ERR(pgdat->kswapd)) {
3551 /* failure at boot is fatal */
3552 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003553 pr_err("Failed to start kswapd on node %d\n", nid);
3554 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003555 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003556 }
3557 return ret;
3558}
3559
David Rientjes8fe23e02009-12-14 17:58:33 -08003560/*
Jiang Liud8adde12012-07-11 14:01:52 -07003561 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003562 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003563 */
3564void kswapd_stop(int nid)
3565{
3566 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3567
Jiang Liud8adde12012-07-11 14:01:52 -07003568 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003569 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003570 NODE_DATA(nid)->kswapd = NULL;
3571 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003572}
3573
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574static int __init kswapd_init(void)
3575{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003576 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003577
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003579 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003580 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 hotcpu_notifier(cpu_callback, 0);
3582 return 0;
3583}
3584
3585module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003586
3587#ifdef CONFIG_NUMA
3588/*
3589 * Zone reclaim mode
3590 *
3591 * If non-zero call zone_reclaim when the number of free pages falls below
3592 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003593 */
3594int zone_reclaim_mode __read_mostly;
3595
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003596#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003597#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003598#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3599#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3600
Christoph Lameter9eeff232006-01-18 17:42:31 -08003601/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003602 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3603 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3604 * a zone.
3605 */
3606#define ZONE_RECLAIM_PRIORITY 4
3607
Christoph Lameter9eeff232006-01-18 17:42:31 -08003608/*
Christoph Lameter96146342006-07-03 00:24:13 -07003609 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3610 * occur.
3611 */
3612int sysctl_min_unmapped_ratio = 1;
3613
3614/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003615 * If the number of slab pages in a zone grows beyond this percentage then
3616 * slab reclaim needs to occur.
3617 */
3618int sysctl_min_slab_ratio = 5;
3619
Mel Gorman90afa5d2009-06-16 15:33:20 -07003620static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3621{
3622 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3623 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3624 zone_page_state(zone, NR_ACTIVE_FILE);
3625
3626 /*
3627 * It's possible for there to be more file mapped pages than
3628 * accounted for by the pages on the file LRU lists because
3629 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3630 */
3631 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3632}
3633
3634/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3635static long zone_pagecache_reclaimable(struct zone *zone)
3636{
3637 long nr_pagecache_reclaimable;
3638 long delta = 0;
3639
3640 /*
3641 * If RECLAIM_SWAP is set, then all file pages are considered
3642 * potentially reclaimable. Otherwise, we have to worry about
3643 * pages like swapcache and zone_unmapped_file_pages() provides
3644 * a better estimate
3645 */
3646 if (zone_reclaim_mode & RECLAIM_SWAP)
3647 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3648 else
3649 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3650
3651 /* If we can't clean pages, remove dirty pages from consideration */
3652 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3653 delta += zone_page_state(zone, NR_FILE_DIRTY);
3654
3655 /* Watch for any possible underflows due to delta */
3656 if (unlikely(delta > nr_pagecache_reclaimable))
3657 delta = nr_pagecache_reclaimable;
3658
3659 return nr_pagecache_reclaimable - delta;
3660}
3661
Christoph Lameter0ff38492006-09-25 23:31:52 -07003662/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003663 * Try to free up some pages from this zone through reclaim.
3664 */
Andrew Morton179e9632006-03-22 00:08:18 -08003665static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003666{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003667 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003668 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003669 struct task_struct *p = current;
3670 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003671 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003672 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003673 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003674 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003675 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003676 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3677 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3678 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003679 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003680
Christoph Lameter9eeff232006-01-18 17:42:31 -08003681 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003682 /*
3683 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3684 * and we also need to be able to write out pages for RECLAIM_WRITE
3685 * and RECLAIM_SWAP.
3686 */
3687 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003688 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003689 reclaim_state.reclaimed_slab = 0;
3690 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003691
Mel Gorman90afa5d2009-06-16 15:33:20 -07003692 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003693 /*
3694 * Free memory by calling shrink zone with increasing
3695 * priorities until we have enough memory freed.
3696 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003697 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003698 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003699 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003700 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003701
Christoph Lameter9eeff232006-01-18 17:42:31 -08003702 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003703 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003704 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003705 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003706}
Andrew Morton179e9632006-03-22 00:08:18 -08003707
3708int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3709{
Andrew Morton179e9632006-03-22 00:08:18 -08003710 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003711 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003712
3713 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003714 * Zone reclaim reclaims unmapped file backed pages and
3715 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003716 *
Christoph Lameter96146342006-07-03 00:24:13 -07003717 * A small portion of unmapped file backed pages is needed for
3718 * file I/O otherwise pages read by file I/O will be immediately
3719 * thrown out if the zone is overallocated. So we do not reclaim
3720 * if less than a specified percentage of the zone is used by
3721 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003722 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003723 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3724 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003725 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003726
Lisa Du6e543d52013-09-11 14:22:36 -07003727 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003728 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003729
Andrew Morton179e9632006-03-22 00:08:18 -08003730 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003731 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003732 */
David Rientjesd773ed62007-10-16 23:26:01 -07003733 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003734 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003735
3736 /*
3737 * Only run zone reclaim on the local zone or on zones that do not
3738 * have associated processors. This will favor the local processor
3739 * over remote processors and spread off node memory allocations
3740 * as wide as possible.
3741 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003742 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003743 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003744 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003745
Johannes Weiner57054652014-10-09 15:28:17 -07003746 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003747 return ZONE_RECLAIM_NOSCAN;
3748
David Rientjesd773ed62007-10-16 23:26:01 -07003749 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003750 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003751
Mel Gorman24cf725182009-06-16 15:33:23 -07003752 if (!ret)
3753 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3754
David Rientjesd773ed62007-10-16 23:26:01 -07003755 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003756}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003757#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003758
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003759/*
3760 * page_evictable - test whether a page is evictable
3761 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003762 *
3763 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003764 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003765 *
3766 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003767 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003768 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003769 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003770 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003771int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003772{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003773 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003774}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003775
Hugh Dickins850465792012-01-20 14:34:19 -08003776#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003777/**
Hugh Dickins24513262012-01-20 14:34:21 -08003778 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3779 * @pages: array of pages to check
3780 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003781 *
Hugh Dickins24513262012-01-20 14:34:21 -08003782 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003783 *
3784 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003785 */
Hugh Dickins24513262012-01-20 14:34:21 -08003786void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003787{
Johannes Weiner925b7672012-01-12 17:18:15 -08003788 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003789 struct zone *zone = NULL;
3790 int pgscanned = 0;
3791 int pgrescued = 0;
3792 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003793
Hugh Dickins24513262012-01-20 14:34:21 -08003794 for (i = 0; i < nr_pages; i++) {
3795 struct page *page = pages[i];
3796 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003797
Hugh Dickins24513262012-01-20 14:34:21 -08003798 pgscanned++;
3799 pagezone = page_zone(page);
3800 if (pagezone != zone) {
3801 if (zone)
3802 spin_unlock_irq(&zone->lru_lock);
3803 zone = pagezone;
3804 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003805 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003806 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003807
Hugh Dickins24513262012-01-20 14:34:21 -08003808 if (!PageLRU(page) || !PageUnevictable(page))
3809 continue;
3810
Hugh Dickins39b5f292012-10-08 16:33:18 -07003811 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003812 enum lru_list lru = page_lru_base_type(page);
3813
Sasha Levin309381fea2014-01-23 15:52:54 -08003814 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003815 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003816 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3817 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003818 pgrescued++;
3819 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003820 }
Hugh Dickins24513262012-01-20 14:34:21 -08003821
3822 if (zone) {
3823 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3824 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3825 spin_unlock_irq(&zone->lru_lock);
3826 }
Hugh Dickins850465792012-01-20 14:34:19 -08003827}
3828#endif /* CONFIG_SHMEM */