blob: 8e645ee520455432e583deb2614199d84e422066 [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
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800235static unsigned long shrink_slabs(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/**
347 * shrink_node_slabs - shrink slab caches of a given node
348 * @gfp_mask: allocation context
349 * @nid: node whose slab caches to target
350 * @nr_scanned: pressure numerator
351 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800353 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800355 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
356 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800358 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
359 * the available objects should be scanned. Page reclaim for example
360 * passes the number of pages scanned and the number of pages on the
361 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
362 * when it encountered mapped pages. The ratio is further biased by
363 * the ->seeks setting of the shrink function, which indicates the
364 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800366 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800368unsigned long shrink_node_slabs(gfp_t gfp_mask, int nid,
369 unsigned long nr_scanned,
370 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371{
372 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000373 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800375 if (nr_scanned == 0)
376 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
Minchan Kimf06590b2011-05-24 17:11:11 -0700378 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000379 /*
380 * If we would return 0, our callers would understand that we
381 * have nothing else to shrink and give up trying. By returning
382 * 1 we keep it going and assume we'll be able to shrink next
383 * time.
384 */
385 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700386 goto out;
387 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
389 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800390 struct shrink_control sc = {
391 .gfp_mask = gfp_mask,
392 .nid = nid,
393 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800394
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800395 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
396 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000397
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800398 freed += shrink_slabs(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800400
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700402out:
403 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000404 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405}
406
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407static inline int is_page_cache_freeable(struct page *page)
408{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700409 /*
410 * A freeable page cache page is referenced only by the caller
411 * that isolated the page, the page cache radix tree and
412 * optional buffer heads at page->private.
413 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700414 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415}
416
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700417static int may_write_to_queue(struct backing_dev_info *bdi,
418 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419{
Christoph Lameter930d9152006-01-08 01:00:47 -0800420 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 return 1;
422 if (!bdi_write_congested(bdi))
423 return 1;
424 if (bdi == current->backing_dev_info)
425 return 1;
426 return 0;
427}
428
429/*
430 * We detected a synchronous write error writing a page out. Probably
431 * -ENOSPC. We need to propagate that into the address_space for a subsequent
432 * fsync(), msync() or close().
433 *
434 * The tricky part is that after writepage we cannot touch the mapping: nothing
435 * prevents it from being freed up. But we have a ref on the page and once
436 * that page is locked, the mapping is pinned.
437 *
438 * We're allowed to run sleeping lock_page() here because we know the caller has
439 * __GFP_FS.
440 */
441static void handle_write_error(struct address_space *mapping,
442 struct page *page, int error)
443{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100444 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700445 if (page_mapping(page) == mapping)
446 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 unlock_page(page);
448}
449
Christoph Lameter04e62a22006-06-23 02:03:38 -0700450/* possible outcome of pageout() */
451typedef enum {
452 /* failed to write page out, page is locked */
453 PAGE_KEEP,
454 /* move page to the active list, page is locked */
455 PAGE_ACTIVATE,
456 /* page has been sent to the disk successfully, page is unlocked */
457 PAGE_SUCCESS,
458 /* page is clean and locked */
459 PAGE_CLEAN,
460} pageout_t;
461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462/*
Andrew Morton1742f192006-03-22 00:08:21 -0800463 * pageout is called by shrink_page_list() for each dirty page.
464 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700466static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700467 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468{
469 /*
470 * If the page is dirty, only perform writeback if that write
471 * will be non-blocking. To prevent this allocation from being
472 * stalled by pagecache activity. But note that there may be
473 * stalls if we need to run get_block(). We could test
474 * PagePrivate for that.
475 *
Al Viro81742022014-04-03 03:17:43 -0400476 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 * this page's queue, we can perform writeback even if that
478 * will block.
479 *
480 * If the page is swapcache, write it back even if that would
481 * block, for some throttling. This happens by accident, because
482 * swap_backing_dev_info is bust: it doesn't reflect the
483 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 */
485 if (!is_page_cache_freeable(page))
486 return PAGE_KEEP;
487 if (!mapping) {
488 /*
489 * Some data journaling orphaned pages can have
490 * page->mapping == NULL while being dirty with clean buffers.
491 */
David Howells266cf652009-04-03 16:42:36 +0100492 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 if (try_to_free_buffers(page)) {
494 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700495 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 return PAGE_CLEAN;
497 }
498 }
499 return PAGE_KEEP;
500 }
501 if (mapping->a_ops->writepage == NULL)
502 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700503 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 return PAGE_KEEP;
505
506 if (clear_page_dirty_for_io(page)) {
507 int res;
508 struct writeback_control wbc = {
509 .sync_mode = WB_SYNC_NONE,
510 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700511 .range_start = 0,
512 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 .for_reclaim = 1,
514 };
515
516 SetPageReclaim(page);
517 res = mapping->a_ops->writepage(page, &wbc);
518 if (res < 0)
519 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800520 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 ClearPageReclaim(page);
522 return PAGE_ACTIVATE;
523 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700524
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 if (!PageWriteback(page)) {
526 /* synchronous write or broken a_ops? */
527 ClearPageReclaim(page);
528 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700529 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700530 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 return PAGE_SUCCESS;
532 }
533
534 return PAGE_CLEAN;
535}
536
Andrew Mortona649fd92006-10-17 00:09:36 -0700537/*
Nick Piggine2867812008-07-25 19:45:30 -0700538 * Same as remove_mapping, but if the page is removed from the mapping, it
539 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700540 */
Johannes Weinera5289102014-04-03 14:47:51 -0700541static int __remove_mapping(struct address_space *mapping, struct page *page,
542 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800543{
Nick Piggin28e4d962006-09-25 23:31:23 -0700544 BUG_ON(!PageLocked(page));
545 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800546
Nick Piggin19fd6232008-07-25 19:45:32 -0700547 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800548 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700549 * The non racy check for a busy page.
550 *
551 * Must be careful with the order of the tests. When someone has
552 * a ref to the page, it may be possible that they dirty it then
553 * drop the reference. So if PageDirty is tested before page_count
554 * here, then the following race may occur:
555 *
556 * get_user_pages(&page);
557 * [user mapping goes away]
558 * write_to(page);
559 * !PageDirty(page) [good]
560 * SetPageDirty(page);
561 * put_page(page);
562 * !page_count(page) [good, discard it]
563 *
564 * [oops, our write_to data is lost]
565 *
566 * Reversing the order of the tests ensures such a situation cannot
567 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
568 * load is not satisfied before that of page->_count.
569 *
570 * Note that if SetPageDirty is always performed via set_page_dirty,
571 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800572 */
Nick Piggine2867812008-07-25 19:45:30 -0700573 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800574 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700575 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
576 if (unlikely(PageDirty(page))) {
577 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800578 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700579 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800580
581 if (PageSwapCache(page)) {
582 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700583 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800584 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700585 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700586 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700587 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500588 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700589 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500590
591 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700592 /*
593 * Remember a shadow entry for reclaimed file cache in
594 * order to detect refaults, thus thrashing, later on.
595 *
596 * But don't store shadows in an address space that is
597 * already exiting. This is not just an optizimation,
598 * inode reclaim needs to empty out the radix tree or
599 * the nodes are lost. Don't plant shadows behind its
600 * back.
601 */
602 if (reclaimed && page_is_file_cache(page) &&
603 !mapping_exiting(mapping))
604 shadow = workingset_eviction(mapping, page);
605 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700606 spin_unlock_irq(&mapping->tree_lock);
Linus Torvalds6072d132010-12-01 13:35:19 -0500607
608 if (freepage != NULL)
609 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800610 }
611
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800612 return 1;
613
614cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700615 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800616 return 0;
617}
618
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619/*
Nick Piggine2867812008-07-25 19:45:30 -0700620 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
621 * someone else has a ref on the page, abort and return 0. If it was
622 * successfully detached, return 1. Assumes the caller has a single ref on
623 * this page.
624 */
625int remove_mapping(struct address_space *mapping, struct page *page)
626{
Johannes Weinera5289102014-04-03 14:47:51 -0700627 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700628 /*
629 * Unfreezing the refcount with 1 rather than 2 effectively
630 * drops the pagecache ref for us without requiring another
631 * atomic operation.
632 */
633 page_unfreeze_refs(page, 1);
634 return 1;
635 }
636 return 0;
637}
638
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700639/**
640 * putback_lru_page - put previously isolated page onto appropriate LRU list
641 * @page: page to be put back to appropriate lru list
642 *
643 * Add previously isolated @page to appropriate LRU list.
644 * Page may still be unevictable for other reasons.
645 *
646 * lru_lock must not be held, interrupts must be enabled.
647 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700648void putback_lru_page(struct page *page)
649{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700650 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700651 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700652
Sasha Levin309381fea2014-01-23 15:52:54 -0800653 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700654
655redo:
656 ClearPageUnevictable(page);
657
Hugh Dickins39b5f292012-10-08 16:33:18 -0700658 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700659 /*
660 * For evictable pages, we can use the cache.
661 * In event of a race, worst case is we end up with an
662 * unevictable page on [in]active list.
663 * We know how to handle that.
664 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700665 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700666 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700667 } else {
668 /*
669 * Put unevictable pages directly on zone's unevictable
670 * list.
671 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700672 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700673 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700674 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700675 * When racing with an mlock or AS_UNEVICTABLE clearing
676 * (page is unlocked) make sure that if the other thread
677 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800678 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700679 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700680 * the page back to the evictable list.
681 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700682 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700683 */
684 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700685 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700686
687 /*
688 * page's status can change while we move it among lru. If an evictable
689 * page is on unevictable list, it never be freed. To avoid that,
690 * check after we added it to the list, again.
691 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700692 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700693 if (!isolate_lru_page(page)) {
694 put_page(page);
695 goto redo;
696 }
697 /* This means someone else dropped this page from LRU
698 * So, it will be freed or putback to LRU again. There is
699 * nothing to do here.
700 */
701 }
702
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700703 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700704 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700705 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700706 count_vm_event(UNEVICTABLE_PGCULLED);
707
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700708 put_page(page); /* drop ref from isolate */
709}
710
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800711enum page_references {
712 PAGEREF_RECLAIM,
713 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800714 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800715 PAGEREF_ACTIVATE,
716};
717
718static enum page_references page_check_references(struct page *page,
719 struct scan_control *sc)
720{
Johannes Weiner645747462010-03-05 13:42:22 -0800721 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800722 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800723
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700724 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
725 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800726 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800727
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800728 /*
729 * Mlock lost the isolation race with us. Let try_to_unmap()
730 * move the page to the unevictable list.
731 */
732 if (vm_flags & VM_LOCKED)
733 return PAGEREF_RECLAIM;
734
Johannes Weiner645747462010-03-05 13:42:22 -0800735 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700736 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800737 return PAGEREF_ACTIVATE;
738 /*
739 * All mapped pages start out with page table
740 * references from the instantiating fault, so we need
741 * to look twice if a mapped file page is used more
742 * than once.
743 *
744 * Mark it and spare it for another trip around the
745 * inactive list. Another page table reference will
746 * lead to its activation.
747 *
748 * Note: the mark is set for activated pages as well
749 * so that recently deactivated but used pages are
750 * quickly recovered.
751 */
752 SetPageReferenced(page);
753
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800754 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800755 return PAGEREF_ACTIVATE;
756
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800757 /*
758 * Activate file-backed executable pages after first usage.
759 */
760 if (vm_flags & VM_EXEC)
761 return PAGEREF_ACTIVATE;
762
Johannes Weiner645747462010-03-05 13:42:22 -0800763 return PAGEREF_KEEP;
764 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800765
766 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700767 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800768 return PAGEREF_RECLAIM_CLEAN;
769
770 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800771}
772
Mel Gormane2be15f2013-07-03 15:01:57 -0700773/* Check if a page is dirty or under writeback */
774static void page_check_dirty_writeback(struct page *page,
775 bool *dirty, bool *writeback)
776{
Mel Gormanb4597222013-07-03 15:02:05 -0700777 struct address_space *mapping;
778
Mel Gormane2be15f2013-07-03 15:01:57 -0700779 /*
780 * Anonymous pages are not handled by flushers and must be written
781 * from reclaim context. Do not stall reclaim based on them
782 */
783 if (!page_is_file_cache(page)) {
784 *dirty = false;
785 *writeback = false;
786 return;
787 }
788
789 /* By default assume that the page flags are accurate */
790 *dirty = PageDirty(page);
791 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700792
793 /* Verify dirty/writeback state if the filesystem supports it */
794 if (!page_has_private(page))
795 return;
796
797 mapping = page_mapping(page);
798 if (mapping && mapping->a_ops->is_dirty_writeback)
799 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700800}
801
Nick Piggine2867812008-07-25 19:45:30 -0700802/*
Andrew Morton1742f192006-03-22 00:08:21 -0800803 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 */
Andrew Morton1742f192006-03-22 00:08:21 -0800805static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700806 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700807 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700808 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700809 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700810 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700811 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700812 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700813 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700814 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815{
816 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700817 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700819 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700820 unsigned long nr_dirty = 0;
821 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800822 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700823 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700824 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825
826 cond_resched();
827
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 while (!list_empty(page_list)) {
829 struct address_space *mapping;
830 struct page *page;
831 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700832 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700833 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834
835 cond_resched();
836
837 page = lru_to_page(page_list);
838 list_del(&page->lru);
839
Nick Piggin529ae9a2008-08-02 12:01:03 +0200840 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 goto keep;
842
Sasha Levin309381fea2014-01-23 15:52:54 -0800843 VM_BUG_ON_PAGE(PageActive(page), page);
844 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845
846 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800847
Hugh Dickins39b5f292012-10-08 16:33:18 -0700848 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700849 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700850
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700851 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800852 goto keep_locked;
853
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 /* Double the slab pressure for mapped and swapcache pages */
855 if (page_mapped(page) || PageSwapCache(page))
856 sc->nr_scanned++;
857
Andy Whitcroftc661b072007-08-22 14:01:26 -0700858 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
859 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
860
Mel Gorman283aba92013-07-03 15:01:51 -0700861 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700862 * The number of dirty pages determines if a zone is marked
863 * reclaim_congested which affects wait_iff_congested. kswapd
864 * will stall and start writing pages if the tail of the LRU
865 * is all dirty unqueued pages.
866 */
867 page_check_dirty_writeback(page, &dirty, &writeback);
868 if (dirty || writeback)
869 nr_dirty++;
870
871 if (dirty && !writeback)
872 nr_unqueued_dirty++;
873
Mel Gormand04e8ac2013-07-03 15:02:03 -0700874 /*
875 * Treat this page as congested if the underlying BDI is or if
876 * pages are cycling through the LRU so quickly that the
877 * pages marked for immediate reclaim are making it to the
878 * end of the LRU a second time.
879 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700880 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800881 if (((dirty || writeback) && mapping &&
882 bdi_write_congested(mapping->backing_dev_info)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700883 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700884 nr_congested++;
885
886 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700887 * If a page at the tail of the LRU is under writeback, there
888 * are three cases to consider.
889 *
890 * 1) If reclaim is encountering an excessive number of pages
891 * under writeback and this page is both under writeback and
892 * PageReclaim then it indicates that pages are being queued
893 * for IO but are being recycled through the LRU before the
894 * IO can complete. Waiting on the page itself risks an
895 * indefinite stall if it is impossible to writeback the
896 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700897 * note that the LRU is being scanned too quickly and the
898 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700899 *
900 * 2) Global reclaim encounters a page, memcg encounters a
901 * page that is not marked for immediate reclaim or
902 * the caller does not have __GFP_IO. In this case mark
903 * the page for immediate reclaim and continue scanning.
904 *
905 * __GFP_IO is checked because a loop driver thread might
906 * enter reclaim, and deadlock if it waits on a page for
907 * which it is needed to do the write (loop masks off
908 * __GFP_IO|__GFP_FS for this reason); but more thought
909 * would probably show more reasons.
910 *
911 * Don't require __GFP_FS, since we're not going into the
912 * FS, just waiting on its writeback completion. Worryingly,
913 * ext4 gfs2 and xfs allocate pages with
914 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
915 * may_enter_fs here is liable to OOM on them.
916 *
917 * 3) memcg encounters a page that is not already marked
918 * PageReclaim. memcg does not have any dirty pages
919 * throttling so we could easily OOM just because too many
920 * pages are in writeback and there is nothing else to
921 * reclaim. Wait for the writeback to complete.
922 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700923 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700924 /* Case 1 above */
925 if (current_is_kswapd() &&
926 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700927 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700928 nr_immediate++;
929 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700930
931 /* Case 2 above */
932 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700933 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
934 /*
935 * This is slightly racy - end_page_writeback()
936 * might have just cleared PageReclaim, then
937 * setting PageReclaim here end up interpreted
938 * as PageReadahead - but that does not matter
939 * enough to care. What we do want is for this
940 * page to have PageReclaim set next time memcg
941 * reclaim reaches the tests above, so it will
942 * then wait_on_page_writeback() to avoid OOM;
943 * and it's also appropriate in global reclaim.
944 */
945 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700946 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700947
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700948 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700949
950 /* Case 3 above */
951 } else {
952 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700953 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
Minchan Kim02c6de82012-10-08 16:31:55 -0700956 if (!force_reclaim)
957 references = page_check_references(page, sc);
958
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800959 switch (references) {
960 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800962 case PAGEREF_KEEP:
963 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800964 case PAGEREF_RECLAIM:
965 case PAGEREF_RECLAIM_CLEAN:
966 ; /* try to reclaim the page below */
967 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 /*
970 * Anonymous process memory has backing store?
971 * Try to allocate it some swap space here.
972 */
Nick Pigginb291f002008-10-18 20:26:44 -0700973 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b932008-11-19 15:36:37 -0800974 if (!(sc->gfp_mask & __GFP_IO))
975 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700976 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 goto activate_locked;
Hugh Dickins63eb6b932008-11-19 15:36:37 -0800978 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
Mel Gormane2be15f2013-07-03 15:01:57 -0700980 /* Adding to swap updated mapping */
981 mapping = page_mapping(page);
982 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
984 /*
985 * The page is mapped into the page tables of one or more
986 * processes. Try to unmap it here.
987 */
988 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700989 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 case SWAP_FAIL:
991 goto activate_locked;
992 case SWAP_AGAIN:
993 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700994 case SWAP_MLOCK:
995 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 case SWAP_SUCCESS:
997 ; /* try to free the page below */
998 }
999 }
1000
1001 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001002 /*
1003 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001004 * avoid risk of stack overflow but only writeback
1005 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001006 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001007 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001008 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001009 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001010 /*
1011 * Immediately reclaim when written back.
1012 * Similar in principal to deactivate_page()
1013 * except we already have the page isolated
1014 * and know it's dirty
1015 */
1016 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1017 SetPageReclaim(page);
1018
Mel Gormanee728862011-10-31 17:07:38 -07001019 goto keep_locked;
1020 }
1021
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001022 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001024 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001026 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 goto keep_locked;
1028
1029 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001030 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 case PAGE_KEEP:
1032 goto keep_locked;
1033 case PAGE_ACTIVATE:
1034 goto activate_locked;
1035 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001036 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001037 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001038 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001040
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 /*
1042 * A synchronous write - probably a ramdisk. Go
1043 * ahead and try to reclaim the page.
1044 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001045 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 goto keep;
1047 if (PageDirty(page) || PageWriteback(page))
1048 goto keep_locked;
1049 mapping = page_mapping(page);
1050 case PAGE_CLEAN:
1051 ; /* try to free the page below */
1052 }
1053 }
1054
1055 /*
1056 * If the page has buffers, try to free the buffer mappings
1057 * associated with this page. If we succeed we try to free
1058 * the page as well.
1059 *
1060 * We do this even if the page is PageDirty().
1061 * try_to_release_page() does not perform I/O, but it is
1062 * possible for a page to have PageDirty set, but it is actually
1063 * clean (all its buffers are clean). This happens if the
1064 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001065 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 * try_to_release_page() will discover that cleanness and will
1067 * drop the buffers and mark the page clean - it can be freed.
1068 *
1069 * Rarely, pages can have buffers and no ->mapping. These are
1070 * the pages which were not successfully invalidated in
1071 * truncate_complete_page(). We try to drop those buffers here
1072 * and if that worked, and the page is no longer mapped into
1073 * process address space (page_count == 1) it can be freed.
1074 * Otherwise, leave the page on the LRU so it is swappable.
1075 */
David Howells266cf652009-04-03 16:42:36 +01001076 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 if (!try_to_release_page(page, sc->gfp_mask))
1078 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001079 if (!mapping && page_count(page) == 1) {
1080 unlock_page(page);
1081 if (put_page_testzero(page))
1082 goto free_it;
1083 else {
1084 /*
1085 * rare race with speculative reference.
1086 * the speculative reference will free
1087 * this page shortly, so we may
1088 * increment nr_reclaimed here (and
1089 * leave it off the LRU).
1090 */
1091 nr_reclaimed++;
1092 continue;
1093 }
1094 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 }
1096
Johannes Weinera5289102014-04-03 14:47:51 -07001097 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001098 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099
Nick Piggina978d6f2008-10-18 20:26:58 -07001100 /*
1101 * At this point, we have no other references and there is
1102 * no way to pick any more up (removed from LRU, removed
1103 * from pagecache). Can use non-atomic bitops now (and
1104 * we obviously don't have to worry about waking up a process
1105 * waiting on the page lock, because there are no references.
1106 */
1107 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001108free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001109 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001110
1111 /*
1112 * Is there need to periodically free_page_list? It would
1113 * appear not as the counts should be low
1114 */
1115 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 continue;
1117
Nick Pigginb291f002008-10-18 20:26:44 -07001118cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001119 if (PageSwapCache(page))
1120 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001121 unlock_page(page);
1122 putback_lru_page(page);
1123 continue;
1124
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001126 /* Not a candidate for swapping, so reclaim swap space. */
1127 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001128 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001129 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 SetPageActive(page);
1131 pgactivate++;
1132keep_locked:
1133 unlock_page(page);
1134keep:
1135 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001136 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001138
Johannes Weiner747db952014-08-08 14:19:24 -07001139 mem_cgroup_uncharge_list(&free_pages);
Mel Gormanb745bc82014-06-04 16:10:22 -07001140 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001141
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001143 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001144
Mel Gorman8e950282013-07-03 15:02:02 -07001145 *ret_nr_dirty += nr_dirty;
1146 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001147 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001148 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001149 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001150 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151}
1152
Minchan Kim02c6de82012-10-08 16:31:55 -07001153unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1154 struct list_head *page_list)
1155{
1156 struct scan_control sc = {
1157 .gfp_mask = GFP_KERNEL,
1158 .priority = DEF_PRIORITY,
1159 .may_unmap = 1,
1160 };
Mel Gorman8e950282013-07-03 15:02:02 -07001161 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001162 struct page *page, *next;
1163 LIST_HEAD(clean_pages);
1164
1165 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001166 if (page_is_file_cache(page) && !PageDirty(page) &&
1167 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001168 ClearPageActive(page);
1169 list_move(&page->lru, &clean_pages);
1170 }
1171 }
1172
1173 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001174 TTU_UNMAP|TTU_IGNORE_ACCESS,
1175 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001176 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001177 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001178 return ret;
1179}
1180
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001181/*
1182 * Attempt to remove the specified page from its LRU. Only take this page
1183 * if it is of the appropriate PageActive status. Pages which are being
1184 * freed elsewhere are also ignored.
1185 *
1186 * page: page to consider
1187 * mode: one of the LRU isolation modes defined above
1188 *
1189 * returns 0 on success, -ve errno on failure.
1190 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001191int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001192{
1193 int ret = -EINVAL;
1194
1195 /* Only take pages on the LRU. */
1196 if (!PageLRU(page))
1197 return ret;
1198
Minchan Kime46a2872012-10-08 16:33:48 -07001199 /* Compaction should not handle unevictable pages but CMA can do so */
1200 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001201 return ret;
1202
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001203 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001204
Mel Gormanc8244932012-01-12 17:19:38 -08001205 /*
1206 * To minimise LRU disruption, the caller can indicate that it only
1207 * wants to isolate pages it will be able to operate on without
1208 * blocking - clean pages for the most part.
1209 *
1210 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1211 * is used by reclaim when it is cannot write to backing storage
1212 *
1213 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1214 * that it is possible to migrate without blocking
1215 */
1216 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1217 /* All the caller can do on PageWriteback is block */
1218 if (PageWriteback(page))
1219 return ret;
1220
1221 if (PageDirty(page)) {
1222 struct address_space *mapping;
1223
1224 /* ISOLATE_CLEAN means only clean pages */
1225 if (mode & ISOLATE_CLEAN)
1226 return ret;
1227
1228 /*
1229 * Only pages without mappings or that have a
1230 * ->migratepage callback are possible to migrate
1231 * without blocking
1232 */
1233 mapping = page_mapping(page);
1234 if (mapping && !mapping->a_ops->migratepage)
1235 return ret;
1236 }
1237 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001238
Minchan Kimf80c0672011-10-31 17:06:55 -07001239 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1240 return ret;
1241
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001242 if (likely(get_page_unless_zero(page))) {
1243 /*
1244 * Be careful not to clear PageLRU until after we're
1245 * sure the page is not being freed elsewhere -- the
1246 * page release code relies on it.
1247 */
1248 ClearPageLRU(page);
1249 ret = 0;
1250 }
1251
1252 return ret;
1253}
1254
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001255/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 * zone->lru_lock is heavily contended. Some of the functions that
1257 * shrink the lists perform better by taking out a batch of pages
1258 * and working on them outside the LRU lock.
1259 *
1260 * For pagecache intensive workloads, this function is the hottest
1261 * spot in the kernel (apart from copy_*_user functions).
1262 *
1263 * Appropriate locks must be held before calling this function.
1264 *
1265 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001266 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001268 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001269 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001270 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001271 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 *
1273 * returns how many pages were moved onto *@dst.
1274 */
Andrew Morton69e05942006-03-22 00:08:19 -08001275static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001276 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001277 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001278 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001280 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001281 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001282 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001284 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001285 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001286 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001287
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 page = lru_to_page(src);
1289 prefetchw_prev_lru_page(page, src, flags);
1290
Sasha Levin309381fea2014-01-23 15:52:54 -08001291 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001292
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001293 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001294 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001295 nr_pages = hpage_nr_pages(page);
1296 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001297 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001298 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001299 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001300
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001301 case -EBUSY:
1302 /* else it is being freed elsewhere */
1303 list_move(&page->lru, src);
1304 continue;
1305
1306 default:
1307 BUG();
1308 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 }
1310
Hugh Dickinsf6260122012-01-12 17:20:06 -08001311 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001312 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1313 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 return nr_taken;
1315}
1316
Nick Piggin62695a82008-10-18 20:26:09 -07001317/**
1318 * isolate_lru_page - tries to isolate a page from its LRU list
1319 * @page: page to isolate from its LRU list
1320 *
1321 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1322 * vmstat statistic corresponding to whatever LRU list the page was on.
1323 *
1324 * Returns 0 if the page was removed from an LRU list.
1325 * Returns -EBUSY if the page was not on an LRU list.
1326 *
1327 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001328 * the active list, it will have PageActive set. If it was found on
1329 * the unevictable list, it will have the PageUnevictable bit set. That flag
1330 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001331 *
1332 * The vmstat statistic corresponding to the list on which the page was
1333 * found will be decremented.
1334 *
1335 * Restrictions:
1336 * (1) Must be called with an elevated refcount on the page. This is a
1337 * fundamentnal difference from isolate_lru_pages (which is called
1338 * without a stable reference).
1339 * (2) the lru_lock must not be held.
1340 * (3) interrupts must be enabled.
1341 */
1342int isolate_lru_page(struct page *page)
1343{
1344 int ret = -EBUSY;
1345
Sasha Levin309381fea2014-01-23 15:52:54 -08001346 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001347
Nick Piggin62695a82008-10-18 20:26:09 -07001348 if (PageLRU(page)) {
1349 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001350 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001351
1352 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001353 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001354 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001355 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001356 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001357 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001358 del_page_from_lru_list(page, lruvec, lru);
1359 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001360 }
1361 spin_unlock_irq(&zone->lru_lock);
1362 }
1363 return ret;
1364}
1365
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001366/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001367 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1368 * then get resheduled. When there are massive number of tasks doing page
1369 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1370 * the LRU list will go small and be scanned faster than necessary, leading to
1371 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001372 */
1373static int too_many_isolated(struct zone *zone, int file,
1374 struct scan_control *sc)
1375{
1376 unsigned long inactive, isolated;
1377
1378 if (current_is_kswapd())
1379 return 0;
1380
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001381 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001382 return 0;
1383
1384 if (file) {
1385 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1386 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1387 } else {
1388 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1389 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1390 }
1391
Fengguang Wu3cf23842012-12-18 14:23:31 -08001392 /*
1393 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1394 * won't get blocked by normal direct-reclaimers, forming a circular
1395 * deadlock.
1396 */
1397 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1398 inactive >>= 3;
1399
Rik van Riel35cd7812009-09-21 17:01:38 -07001400 return isolated > inactive;
1401}
1402
Mel Gorman66635622010-08-09 17:19:30 -07001403static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001404putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001405{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001406 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1407 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001408 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001409
Mel Gorman66635622010-08-09 17:19:30 -07001410 /*
1411 * Put back any unfreeable pages.
1412 */
Mel Gorman66635622010-08-09 17:19:30 -07001413 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001414 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001415 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001416
Sasha Levin309381fea2014-01-23 15:52:54 -08001417 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001418 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001419 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001420 spin_unlock_irq(&zone->lru_lock);
1421 putback_lru_page(page);
1422 spin_lock_irq(&zone->lru_lock);
1423 continue;
1424 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001425
1426 lruvec = mem_cgroup_page_lruvec(page, zone);
1427
Linus Torvalds7a608572011-01-17 14:42:19 -08001428 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001429 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001430 add_page_to_lru_list(page, lruvec, lru);
1431
Mel Gorman66635622010-08-09 17:19:30 -07001432 if (is_active_lru(lru)) {
1433 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001434 int numpages = hpage_nr_pages(page);
1435 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001436 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001437 if (put_page_testzero(page)) {
1438 __ClearPageLRU(page);
1439 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001440 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001441
1442 if (unlikely(PageCompound(page))) {
1443 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001444 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001445 (*get_compound_page_dtor(page))(page);
1446 spin_lock_irq(&zone->lru_lock);
1447 } else
1448 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001449 }
1450 }
Mel Gorman66635622010-08-09 17:19:30 -07001451
Hugh Dickins3f797682012-01-12 17:20:07 -08001452 /*
1453 * To save our caller's stack, now use input list for pages to free.
1454 */
1455 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001456}
1457
1458/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001459 * If a kernel thread (such as nfsd for loop-back mounts) services
1460 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1461 * In that case we should only throttle if the backing device it is
1462 * writing to is congested. In other cases it is safe to throttle.
1463 */
1464static int current_may_throttle(void)
1465{
1466 return !(current->flags & PF_LESS_THROTTLE) ||
1467 current->backing_dev_info == NULL ||
1468 bdi_write_congested(current->backing_dev_info);
1469}
1470
1471/*
Andrew Morton1742f192006-03-22 00:08:21 -08001472 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1473 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 */
Mel Gorman66635622010-08-09 17:19:30 -07001475static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001476shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001477 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478{
1479 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001480 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001481 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001482 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001483 unsigned long nr_dirty = 0;
1484 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001485 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001486 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001487 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001488 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001489 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001490 struct zone *zone = lruvec_zone(lruvec);
1491 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001492
Rik van Riel35cd7812009-09-21 17:01:38 -07001493 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001494 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001495
1496 /* We are about to die and free our memory. Return now. */
1497 if (fatal_signal_pending(current))
1498 return SWAP_CLUSTER_MAX;
1499 }
1500
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001502
1503 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001504 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001505 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001506 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001507
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001510 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1511 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001512
1513 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1514 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1515
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001516 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001517 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001518 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001519 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001520 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001521 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001522 }
Hillf Dantond563c052012-03-21 16:34:02 -07001523 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001524
Hillf Dantond563c052012-03-21 16:34:02 -07001525 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001526 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001527
Minchan Kim02c6de82012-10-08 16:31:55 -07001528 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001529 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1530 &nr_writeback, &nr_immediate,
1531 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001532
Hugh Dickins3f797682012-01-12 17:20:07 -08001533 spin_lock_irq(&zone->lru_lock);
1534
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001535 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001536
Ying Han904249a2012-04-25 16:01:48 -07001537 if (global_reclaim(sc)) {
1538 if (current_is_kswapd())
1539 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1540 nr_reclaimed);
1541 else
1542 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1543 nr_reclaimed);
1544 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001545
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001546 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001547
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001548 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001549
1550 spin_unlock_irq(&zone->lru_lock);
1551
Johannes Weiner747db952014-08-08 14:19:24 -07001552 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001553 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001554
Mel Gorman92df3a72011-10-31 17:07:56 -07001555 /*
1556 * If reclaim is isolating dirty pages under writeback, it implies
1557 * that the long-lived page allocation rate is exceeding the page
1558 * laundering rate. Either the global limits are not being effective
1559 * at throttling processes due to the page distribution throughout
1560 * zones or there is heavy usage of a slow backing device. The
1561 * only option is to throttle from reclaim context which is not ideal
1562 * as there is no guarantee the dirtying process is throttled in the
1563 * same way balance_dirty_pages() manages.
1564 *
Mel Gorman8e950282013-07-03 15:02:02 -07001565 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1566 * of pages under pages flagged for immediate reclaim and stall if any
1567 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001568 */
Mel Gorman918fc712013-07-08 16:00:25 -07001569 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001570 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001571
Mel Gormand43006d2013-07-03 15:01:50 -07001572 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001573 * memcg will stall in page writeback so only consider forcibly
1574 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001575 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001576 if (global_reclaim(sc)) {
1577 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001578 * Tag a zone as congested if all the dirty pages scanned were
1579 * backed by a congested BDI and wait_iff_congested will stall.
1580 */
1581 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001582 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001583
1584 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001585 * If dirty pages are scanned that are not queued for IO, it
1586 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001587 * the zone ZONE_DIRTY and kswapd will start writing pages from
1588 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001589 */
1590 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001591 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001592
1593 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001594 * If kswapd scans pages marked marked for immediate
1595 * reclaim and under writeback (nr_immediate), it implies
1596 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001597 * they are written so also forcibly stall.
1598 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001599 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001600 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001601 }
Mel Gormand43006d2013-07-03 15:01:50 -07001602
Mel Gorman8e950282013-07-03 15:02:02 -07001603 /*
1604 * Stall direct reclaim for IO completions if underlying BDIs or zone
1605 * is congested. Allow kswapd to continue until it starts encountering
1606 * unqueued dirty pages or cycling through the LRU too quickly.
1607 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001608 if (!sc->hibernation_mode && !current_is_kswapd() &&
1609 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001610 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1611
Mel Gormane11da5b2010-10-26 14:21:40 -07001612 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1613 zone_idx(zone),
1614 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001615 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001616 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001617 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618}
1619
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001620/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 * This moves pages from the active list to the inactive list.
1622 *
1623 * We move them the other way if the page is referenced by one or more
1624 * processes, from rmap.
1625 *
1626 * If the pages are mostly unmapped, the processing is fast and it is
1627 * appropriate to hold zone->lru_lock across the whole operation. But if
1628 * the pages are mapped, the processing is slow (page_referenced()) so we
1629 * should drop zone->lru_lock around each page. It's impossible to balance
1630 * this, so instead we remove the pages from the LRU while processing them.
1631 * It is safe to rely on PG_active against the non-LRU pages in here because
1632 * nobody will play with that bit on a non-LRU page.
1633 *
1634 * The downside is that we have to touch page->_count against each page.
1635 * But we had to alter page->flags anyway.
1636 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001637
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001638static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001639 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001640 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001641 enum lru_list lru)
1642{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001643 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001644 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001645 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001646 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001647
Wu Fengguang3eb41402009-06-16 15:33:13 -07001648 while (!list_empty(list)) {
1649 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001650 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001651
Sasha Levin309381fea2014-01-23 15:52:54 -08001652 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001653 SetPageLRU(page);
1654
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001655 nr_pages = hpage_nr_pages(page);
1656 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001657 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001658 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001659
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001660 if (put_page_testzero(page)) {
1661 __ClearPageLRU(page);
1662 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001663 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001664
1665 if (unlikely(PageCompound(page))) {
1666 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001667 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001668 (*get_compound_page_dtor(page))(page);
1669 spin_lock_irq(&zone->lru_lock);
1670 } else
1671 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001672 }
1673 }
1674 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1675 if (!is_active_lru(lru))
1676 __count_vm_events(PGDEACTIVATE, pgmoved);
1677}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001678
Hugh Dickinsf6260122012-01-12 17:20:06 -08001679static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001680 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001681 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001682 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001684 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001685 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001686 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001688 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001689 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001691 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001692 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001693 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001694 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001695 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696
1697 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001698
1699 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001700 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001701 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001702 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001703
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001705
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001706 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1707 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001708 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001709 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001710
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001711 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001712
Hugh Dickinsf6260122012-01-12 17:20:06 -08001713 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001714 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001715 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 spin_unlock_irq(&zone->lru_lock);
1717
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 while (!list_empty(&l_hold)) {
1719 cond_resched();
1720 page = lru_to_page(&l_hold);
1721 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001722
Hugh Dickins39b5f292012-10-08 16:33:18 -07001723 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001724 putback_lru_page(page);
1725 continue;
1726 }
1727
Mel Gormancc715d92012-03-21 16:34:00 -07001728 if (unlikely(buffer_heads_over_limit)) {
1729 if (page_has_private(page) && trylock_page(page)) {
1730 if (page_has_private(page))
1731 try_to_release_page(page, 0);
1732 unlock_page(page);
1733 }
1734 }
1735
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001736 if (page_referenced(page, 0, sc->target_mem_cgroup,
1737 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001738 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001739 /*
1740 * Identify referenced, file-backed active pages and
1741 * give them one more trip around the active list. So
1742 * that executable code get better chances to stay in
1743 * memory under moderate memory pressure. Anon pages
1744 * are not likely to be evicted by use-once streaming
1745 * IO, plus JVM can create lots of anon VM_EXEC pages,
1746 * so we ignore them here.
1747 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001748 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001749 list_add(&page->lru, &l_active);
1750 continue;
1751 }
1752 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001753
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001754 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 list_add(&page->lru, &l_inactive);
1756 }
1757
Andrew Mortonb5557492009-01-06 14:40:13 -08001758 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001759 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001760 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001761 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001762 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001763 * Count referenced pages from currently used mappings as rotated,
1764 * even though only some of them are actually re-activated. This
1765 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001766 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001767 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001768 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001769
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001770 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1771 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001772 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001773 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001774
Johannes Weiner747db952014-08-08 14:19:24 -07001775 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001776 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777}
1778
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001779#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001780static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001781{
1782 unsigned long active, inactive;
1783
1784 active = zone_page_state(zone, NR_ACTIVE_ANON);
1785 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1786
1787 if (inactive * zone->inactive_ratio < active)
1788 return 1;
1789
1790 return 0;
1791}
1792
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001793/**
1794 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001795 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001796 *
1797 * Returns true if the zone does not have enough inactive anon pages,
1798 * meaning some active anon pages need to be deactivated.
1799 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001800static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001801{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001802 /*
1803 * If we don't have swap space, anonymous page deactivation
1804 * is pointless.
1805 */
1806 if (!total_swap_pages)
1807 return 0;
1808
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001809 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001810 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001811
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001812 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001813}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001814#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001815static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001816{
1817 return 0;
1818}
1819#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001820
Rik van Riel56e49d22009-06-16 15:32:28 -07001821/**
1822 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001823 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001824 *
1825 * When the system is doing streaming IO, memory pressure here
1826 * ensures that active file pages get deactivated, until more
1827 * than half of the file pages are on the inactive list.
1828 *
1829 * Once we get to that situation, protect the system's working
1830 * set from being evicted by disabling active file page aging.
1831 *
1832 * This uses a different ratio than the anonymous pages, because
1833 * the page cache uses a use-once replacement algorithm.
1834 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001835static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001836{
Johannes Weinere3790142013-02-22 16:35:19 -08001837 unsigned long inactive;
1838 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001839
Johannes Weinere3790142013-02-22 16:35:19 -08001840 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1841 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1842
1843 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001844}
1845
Hugh Dickins75b00af2012-05-29 15:07:09 -07001846static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001847{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001848 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001849 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001850 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001851 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001852}
1853
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001854static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001855 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001856{
Rik van Rielb39415b2009-12-14 17:59:48 -08001857 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001858 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001859 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001860 return 0;
1861 }
1862
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001863 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001864}
1865
Johannes Weiner9a265112013-02-22 16:32:17 -08001866enum scan_balance {
1867 SCAN_EQUAL,
1868 SCAN_FRACT,
1869 SCAN_ANON,
1870 SCAN_FILE,
1871};
1872
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001874 * Determine how aggressively the anon and file LRU lists should be
1875 * scanned. The relative value of each set of LRU lists is determined
1876 * by looking at the fraction of the pages scanned we did rotate back
1877 * onto the active list instead of evict.
1878 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001879 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1880 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001881 */
Johannes Weiner02695172014-08-06 16:06:17 -07001882static void get_scan_count(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001883 struct scan_control *sc, unsigned long *nr,
1884 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001885{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001886 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001887 u64 fraction[2];
1888 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001889 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001890 unsigned long anon_prio, file_prio;
1891 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001892 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001893 bool force_scan = false;
1894 unsigned long ap, fp;
1895 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001896 bool some_scanned;
1897 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001898
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001899 /*
1900 * If the zone or memcg is small, nr[l] can be 0. This
1901 * results in no scanning on this priority and a potential
1902 * priority drop. Global direct reclaim can go to the next
1903 * zone and tends to have no problems. Global kswapd is for
1904 * zone balancing and it needs to scan a minimum amount. When
1905 * reclaiming for a memcg, a priority drop can cause high
1906 * latencies, so it's better to scan a minimum amount there as
1907 * well.
1908 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08001909 if (current_is_kswapd()) {
1910 if (!zone_reclaimable(zone))
1911 force_scan = true;
1912 if (!mem_cgroup_lruvec_online(lruvec))
1913 force_scan = true;
1914 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001915 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001916 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001917
1918 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001919 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001920 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001921 goto out;
1922 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001923
Johannes Weiner10316b32013-02-22 16:32:14 -08001924 /*
1925 * Global reclaim will swap to prevent OOM even with no
1926 * swappiness, but memcg users want to use this knob to
1927 * disable swapping for individual groups completely when
1928 * using the memory controller's swap limit feature would be
1929 * too expensive.
1930 */
Johannes Weiner02695172014-08-06 16:06:17 -07001931 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001932 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001933 goto out;
1934 }
1935
1936 /*
1937 * Do not apply any pressure balancing cleverness when the
1938 * system is close to OOM, scan both anon and file equally
1939 * (unless the swappiness setting disagrees with swapping).
1940 */
Johannes Weiner02695172014-08-06 16:06:17 -07001941 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001942 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001943 goto out;
1944 }
1945
Johannes Weiner11d16c22013-02-22 16:32:15 -08001946 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001947 * Prevent the reclaimer from falling into the cache trap: as
1948 * cache pages start out inactive, every cache fault will tip
1949 * the scan balance towards the file LRU. And as the file LRU
1950 * shrinks, so does the window for rotation from references.
1951 * This means we have a runaway feedback loop where a tiny
1952 * thrashing file LRU becomes infinitely more attractive than
1953 * anon pages. Try to detect this based on file LRU size.
1954 */
1955 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001956 unsigned long zonefile;
1957 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07001958
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001959 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1960 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1961 zone_page_state(zone, NR_INACTIVE_FILE);
1962
1963 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07001964 scan_balance = SCAN_ANON;
1965 goto out;
1966 }
1967 }
1968
1969 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001970 * There is enough inactive page cache, do not reclaim
1971 * anything from the anonymous working set right now.
1972 */
1973 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001974 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001975 goto out;
1976 }
1977
Johannes Weiner9a265112013-02-22 16:32:17 -08001978 scan_balance = SCAN_FRACT;
1979
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001980 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001981 * With swappiness at 100, anonymous and file have the same priority.
1982 * This scanning priority is essentially the inverse of IO cost.
1983 */
Johannes Weiner02695172014-08-06 16:06:17 -07001984 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001985 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001986
1987 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001988 * OK, so we have swap space and a fair amount of page cache
1989 * pages. We use the recently rotated / recently scanned
1990 * ratios to determine how valuable each cache is.
1991 *
1992 * Because workloads change over time (and to avoid overflow)
1993 * we keep these statistics as a floating average, which ends
1994 * up weighing recent references more than old ones.
1995 *
1996 * anon in [0], file in [1]
1997 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001998
1999 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2000 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2001 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2002 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2003
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002004 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002005 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002006 reclaim_stat->recent_scanned[0] /= 2;
2007 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002008 }
2009
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002010 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002011 reclaim_stat->recent_scanned[1] /= 2;
2012 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002013 }
2014
2015 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002016 * The amount of pressure on anon vs file pages is inversely
2017 * proportional to the fraction of recently scanned pages on
2018 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002019 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002020 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002021 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002022
Satoru Moriyafe350042012-05-29 15:06:47 -07002023 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002024 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002025 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002026
Shaohua Li76a33fc2010-05-24 14:32:36 -07002027 fraction[0] = ap;
2028 fraction[1] = fp;
2029 denominator = ap + fp + 1;
2030out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002031 some_scanned = false;
2032 /* Only use force_scan on second pass. */
2033 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002034 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002035 for_each_evictable_lru(lru) {
2036 int file = is_file_lru(lru);
2037 unsigned long size;
2038 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002039
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002040 size = get_lru_size(lruvec, lru);
2041 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002042
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002043 if (!scan && pass && force_scan)
2044 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002045
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002046 switch (scan_balance) {
2047 case SCAN_EQUAL:
2048 /* Scan lists relative to size */
2049 break;
2050 case SCAN_FRACT:
2051 /*
2052 * Scan types proportional to swappiness and
2053 * their relative recent reclaim efficiency.
2054 */
2055 scan = div64_u64(scan * fraction[file],
2056 denominator);
2057 break;
2058 case SCAN_FILE:
2059 case SCAN_ANON:
2060 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002061 if ((scan_balance == SCAN_FILE) != file) {
2062 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002063 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002064 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002065 break;
2066 default:
2067 /* Look ma, no brain */
2068 BUG();
2069 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002070
2071 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002072 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002073
Johannes Weiner9a265112013-02-22 16:32:17 -08002074 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002075 * Skip the second pass and don't force_scan,
2076 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002077 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002078 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002079 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002080 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002081}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002082
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002083/*
2084 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2085 */
Johannes Weiner02695172014-08-06 16:06:17 -07002086static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002087 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002088{
2089 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002090 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002091 unsigned long nr_to_scan;
2092 enum lru_list lru;
2093 unsigned long nr_reclaimed = 0;
2094 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2095 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002096 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002097
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002098 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002099
Mel Gormane82e0562013-07-03 15:01:44 -07002100 /* Record the original scan target for proportional adjustments later */
2101 memcpy(targets, nr, sizeof(nr));
2102
Mel Gorman1a501902014-06-04 16:10:49 -07002103 /*
2104 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2105 * event that can occur when there is little memory pressure e.g.
2106 * multiple streaming readers/writers. Hence, we do not abort scanning
2107 * when the requested number of pages are reclaimed when scanning at
2108 * DEF_PRIORITY on the assumption that the fact we are direct
2109 * reclaiming implies that kswapd is not keeping up and it is best to
2110 * do a batch of work at once. For memcg reclaim one check is made to
2111 * abort proportional reclaim if either the file or anon lru has already
2112 * dropped to zero at the first pass.
2113 */
2114 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2115 sc->priority == DEF_PRIORITY);
2116
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002117 blk_start_plug(&plug);
2118 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2119 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002120 unsigned long nr_anon, nr_file, percentage;
2121 unsigned long nr_scanned;
2122
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002123 for_each_evictable_lru(lru) {
2124 if (nr[lru]) {
2125 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2126 nr[lru] -= nr_to_scan;
2127
2128 nr_reclaimed += shrink_list(lru, nr_to_scan,
2129 lruvec, sc);
2130 }
2131 }
Mel Gormane82e0562013-07-03 15:01:44 -07002132
2133 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2134 continue;
2135
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002136 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002137 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002138 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002139 * proportionally what was requested by get_scan_count(). We
2140 * stop reclaiming one LRU and reduce the amount scanning
2141 * proportional to the original scan target.
2142 */
2143 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2144 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2145
Mel Gorman1a501902014-06-04 16:10:49 -07002146 /*
2147 * It's just vindictive to attack the larger once the smaller
2148 * has gone to zero. And given the way we stop scanning the
2149 * smaller below, this makes sure that we only make one nudge
2150 * towards proportionality once we've got nr_to_reclaim.
2151 */
2152 if (!nr_file || !nr_anon)
2153 break;
2154
Mel Gormane82e0562013-07-03 15:01:44 -07002155 if (nr_file > nr_anon) {
2156 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2157 targets[LRU_ACTIVE_ANON] + 1;
2158 lru = LRU_BASE;
2159 percentage = nr_anon * 100 / scan_target;
2160 } else {
2161 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2162 targets[LRU_ACTIVE_FILE] + 1;
2163 lru = LRU_FILE;
2164 percentage = nr_file * 100 / scan_target;
2165 }
2166
2167 /* Stop scanning the smaller of the LRU */
2168 nr[lru] = 0;
2169 nr[lru + LRU_ACTIVE] = 0;
2170
2171 /*
2172 * Recalculate the other LRU scan count based on its original
2173 * scan target and the percentage scanning already complete
2174 */
2175 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2176 nr_scanned = targets[lru] - nr[lru];
2177 nr[lru] = targets[lru] * (100 - percentage) / 100;
2178 nr[lru] -= min(nr[lru], nr_scanned);
2179
2180 lru += LRU_ACTIVE;
2181 nr_scanned = targets[lru] - nr[lru];
2182 nr[lru] = targets[lru] * (100 - percentage) / 100;
2183 nr[lru] -= min(nr[lru], nr_scanned);
2184
2185 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002186 }
2187 blk_finish_plug(&plug);
2188 sc->nr_reclaimed += nr_reclaimed;
2189
2190 /*
2191 * Even if we did not try to evict anon pages at all, we want to
2192 * rebalance the anon lru active/inactive ratio.
2193 */
2194 if (inactive_anon_is_low(lruvec))
2195 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2196 sc, LRU_ACTIVE_ANON);
2197
2198 throttle_vm_writeout(sc->gfp_mask);
2199}
2200
Mel Gorman23b9da52012-05-29 15:06:20 -07002201/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002202static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002203{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002204 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002205 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002206 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002207 return true;
2208
2209 return false;
2210}
2211
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002212/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002213 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2214 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2215 * true if more pages should be reclaimed such that when the page allocator
2216 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2217 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002218 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002219static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002220 unsigned long nr_reclaimed,
2221 unsigned long nr_scanned,
2222 struct scan_control *sc)
2223{
2224 unsigned long pages_for_compaction;
2225 unsigned long inactive_lru_pages;
2226
2227 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002228 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002229 return false;
2230
Mel Gorman28765922011-02-25 14:44:20 -08002231 /* Consider stopping depending on scan and reclaim activity */
2232 if (sc->gfp_mask & __GFP_REPEAT) {
2233 /*
2234 * For __GFP_REPEAT allocations, stop reclaiming if the
2235 * full LRU list has been scanned and we are still failing
2236 * to reclaim pages. This full LRU scan is potentially
2237 * expensive but a __GFP_REPEAT caller really wants to succeed
2238 */
2239 if (!nr_reclaimed && !nr_scanned)
2240 return false;
2241 } else {
2242 /*
2243 * For non-__GFP_REPEAT allocations which can presumably
2244 * fail without consequence, stop if we failed to reclaim
2245 * any pages from the last SWAP_CLUSTER_MAX number of
2246 * pages that were scanned. This will return to the
2247 * caller faster at the risk reclaim/compaction and
2248 * the resulting allocation attempt fails
2249 */
2250 if (!nr_reclaimed)
2251 return false;
2252 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002253
2254 /*
2255 * If we have not reclaimed enough pages for compaction and the
2256 * inactive lists are large enough, continue reclaiming
2257 */
2258 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002259 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002260 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002261 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002262 if (sc->nr_reclaimed < pages_for_compaction &&
2263 inactive_lru_pages > pages_for_compaction)
2264 return true;
2265
2266 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002267 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002268 case COMPACT_PARTIAL:
2269 case COMPACT_CONTINUE:
2270 return false;
2271 default:
2272 return true;
2273 }
2274}
2275
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002276static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2277 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002278{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002279 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002280 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002281
Johannes Weiner56600482012-01-12 17:17:59 -08002282 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002283 struct mem_cgroup *root = sc->target_mem_cgroup;
2284 struct mem_cgroup_reclaim_cookie reclaim = {
2285 .zone = zone,
2286 .priority = sc->priority,
2287 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002288 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002289 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002290
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002291 nr_reclaimed = sc->nr_reclaimed;
2292 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002293
Andrew Morton694fbc02013-09-24 15:27:37 -07002294 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2295 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002296 unsigned long lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002297 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002298 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002299
Johannes Weiner241994ed2015-02-11 15:26:06 -08002300 if (mem_cgroup_low(root, memcg)) {
2301 if (!sc->may_thrash)
2302 continue;
2303 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2304 }
2305
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002306 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002307 swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002308
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002309 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2310 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002311
2312 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002313 * Direct reclaim and kswapd have to scan all memory
2314 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002315 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002316 *
2317 * Limit reclaim, on the other hand, only cares about
2318 * nr_to_reclaim pages to be reclaimed and it will
2319 * retry with decreasing priority if one round over the
2320 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002321 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002322 if (!global_reclaim(sc) &&
2323 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002324 mem_cgroup_iter_break(root, memcg);
2325 break;
2326 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002327 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002328
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002329 /*
2330 * Shrink the slab caches in the same proportion that
2331 * the eligible LRU pages were scanned.
2332 */
2333 if (global_reclaim(sc) && is_classzone) {
2334 struct reclaim_state *reclaim_state;
2335
2336 shrink_node_slabs(sc->gfp_mask, zone_to_nid(zone),
2337 sc->nr_scanned - nr_scanned,
2338 zone_lru_pages);
2339
2340 reclaim_state = current->reclaim_state;
2341 if (reclaim_state) {
2342 sc->nr_reclaimed +=
2343 reclaim_state->reclaimed_slab;
2344 reclaim_state->reclaimed_slab = 0;
2345 }
2346 }
2347
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002348 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2349 sc->nr_scanned - nr_scanned,
2350 sc->nr_reclaimed - nr_reclaimed);
2351
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002352 if (sc->nr_reclaimed - nr_reclaimed)
2353 reclaimable = true;
2354
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002355 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2356 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002357
2358 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002359}
2360
Vlastimil Babka53853e22014-10-09 15:27:02 -07002361/*
2362 * Returns true if compaction should go ahead for a high-order request, or
2363 * the high-order allocation would succeed without compaction.
2364 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002365static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002366{
2367 unsigned long balance_gap, watermark;
2368 bool watermark_ok;
2369
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002370 /*
2371 * Compaction takes time to run and there are potentially other
2372 * callers using the pages just freed. Continue reclaiming until
2373 * there is a buffer of free pages available to give compaction
2374 * a reasonable chance of completing and allocating the page
2375 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002376 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2377 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002378 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002379 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2380
2381 /*
2382 * If compaction is deferred, reclaim up to a point where
2383 * compaction will have a chance of success when re-enabled
2384 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002385 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002386 return watermark_ok;
2387
Vlastimil Babka53853e22014-10-09 15:27:02 -07002388 /*
2389 * If compaction is not ready to start and allocation is not likely
2390 * to succeed without it, then keep reclaiming.
2391 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002392 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002393 return false;
2394
2395 return watermark_ok;
2396}
2397
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398/*
2399 * This is the direct reclaim path, for page-allocating processes. We only
2400 * try to reclaim pages from zones which will satisfy the caller's allocation
2401 * request.
2402 *
Mel Gorman41858962009-06-16 15:32:12 -07002403 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2404 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2406 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002407 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2408 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2409 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 * If a zone is deemed to be full of pinned pages then just give it a light
2412 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002413 *
2414 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002416static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417{
Mel Gormandd1a2392008-04-28 02:12:17 -07002418 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002419 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002420 unsigned long nr_soft_reclaimed;
2421 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002422 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002423 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002424 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002425
Mel Gormancc715d92012-03-21 16:34:00 -07002426 /*
2427 * If the number of buffer_heads in the machine exceeds the maximum
2428 * allowed level, force direct reclaim to scan the highmem zone as
2429 * highmem pages could be pinning lowmem pages storing buffer_heads
2430 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002431 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002432 if (buffer_heads_over_limit)
2433 sc->gfp_mask |= __GFP_HIGHMEM;
2434
Mel Gormand4debc62010-08-09 17:19:29 -07002435 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002436 requested_highidx, sc->nodemask) {
2437 enum zone_type classzone_idx;
2438
Con Kolivasf3fe6512006-01-06 00:11:15 -08002439 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002441
2442 classzone_idx = requested_highidx;
2443 while (!populated_zone(zone->zone_pgdat->node_zones +
2444 classzone_idx))
2445 classzone_idx--;
2446
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002447 /*
2448 * Take care memory controller reclaiming has small influence
2449 * to global LRU.
2450 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002451 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002452 if (!cpuset_zone_allowed(zone,
2453 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002454 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002455
Lisa Du6e543d52013-09-11 14:22:36 -07002456 if (sc->priority != DEF_PRIORITY &&
2457 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002458 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002459
2460 /*
2461 * If we already have plenty of memory free for
2462 * compaction in this zone, don't free any more.
2463 * Even though compaction is invoked for any
2464 * non-zero order, only frequent costly order
2465 * reclamation is disruptive enough to become a
2466 * noticeable problem, like transparent huge
2467 * page allocations.
2468 */
2469 if (IS_ENABLED(CONFIG_COMPACTION) &&
2470 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2471 zonelist_zone_idx(z) <= requested_highidx &&
2472 compaction_ready(zone, sc->order)) {
2473 sc->compaction_ready = true;
2474 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002475 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002476
Andrew Morton0608f432013-09-24 15:27:41 -07002477 /*
2478 * This steals pages from memory cgroups over softlimit
2479 * and returns the number of reclaimed pages and
2480 * scanned pages. This works for global memory pressure
2481 * and balancing, not for a memcg's limit.
2482 */
2483 nr_soft_scanned = 0;
2484 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2485 sc->order, sc->gfp_mask,
2486 &nr_soft_scanned);
2487 sc->nr_reclaimed += nr_soft_reclaimed;
2488 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002489 if (nr_soft_reclaimed)
2490 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002491 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002492 }
Nick Piggin408d8542006-09-25 23:31:27 -07002493
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002494 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002495 reclaimable = true;
2496
2497 if (global_reclaim(sc) &&
2498 !reclaimable && zone_reclaimable(zone))
2499 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 }
Mel Gormane0c23272011-10-31 17:09:33 -07002501
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002502 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002503 * Restore to original mask to avoid the impact on the caller if we
2504 * promoted it to __GFP_HIGHMEM.
2505 */
2506 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002507
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002508 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002510
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511/*
2512 * This is the main entry point to direct page reclaim.
2513 *
2514 * If a full scan of the inactive list fails to free enough memory then we
2515 * are "out of memory" and something needs to be killed.
2516 *
2517 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2518 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002519 * caller can't do much about. We kick the writeback threads and take explicit
2520 * naps in the hope that some of these pages can be written. But if the
2521 * allocating task holds filesystem locks which prevent writeout this might not
2522 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002523 *
2524 * returns: 0, if no pages reclaimed
2525 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 */
Mel Gormandac1d272008-04-28 02:12:12 -07002527static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002528 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002530 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002531 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002532 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002533 bool zones_reclaimable;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002534retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002535 delayacct_freepages_start();
2536
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002537 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002538 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002540 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002541 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2542 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002543 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002544 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002545
Balbir Singh66e17072008-02-07 00:13:56 -08002546 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002547 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002548 break;
2549
2550 if (sc->compaction_ready)
2551 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552
2553 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002554 * If we're getting trouble reclaiming, start doing
2555 * writepage even in laptop mode.
2556 */
2557 if (sc->priority < DEF_PRIORITY - 2)
2558 sc->may_writepage = 1;
2559
2560 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 * Try to write back as many pages as we just scanned. This
2562 * tends to cause slow streaming writers to write data to the
2563 * disk smoothly, at the dirtying rate, which is nice. But
2564 * that's undesirable in laptop mode, where we *want* lumpy
2565 * writeout. So in laptop mode, write out the whole world.
2566 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002567 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2568 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002569 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2570 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002571 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002573 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002574
Keika Kobayashi873b4772008-07-25 01:48:52 -07002575 delayacct_freepages_end();
2576
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002577 if (sc->nr_reclaimed)
2578 return sc->nr_reclaimed;
2579
Mel Gorman0cee34f2012-01-12 17:19:49 -08002580 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002581 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002582 return 1;
2583
Johannes Weiner241994ed2015-02-11 15:26:06 -08002584 /* Untapped cgroup reserves? Don't OOM, retry. */
2585 if (!sc->may_thrash) {
2586 sc->priority = initial_priority;
2587 sc->may_thrash = 1;
2588 goto retry;
2589 }
2590
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002591 /* Any of the zones still reclaimable? Don't OOM. */
2592 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002593 return 1;
2594
2595 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596}
2597
Mel Gorman55150612012-07-31 16:44:35 -07002598static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2599{
2600 struct zone *zone;
2601 unsigned long pfmemalloc_reserve = 0;
2602 unsigned long free_pages = 0;
2603 int i;
2604 bool wmark_ok;
2605
2606 for (i = 0; i <= ZONE_NORMAL; i++) {
2607 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002608 if (!populated_zone(zone))
2609 continue;
2610
Mel Gorman55150612012-07-31 16:44:35 -07002611 pfmemalloc_reserve += min_wmark_pages(zone);
2612 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2613 }
2614
Mel Gorman675becc2014-06-04 16:07:35 -07002615 /* If there are no reserves (unexpected config) then do not throttle */
2616 if (!pfmemalloc_reserve)
2617 return true;
2618
Mel Gorman55150612012-07-31 16:44:35 -07002619 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2620
2621 /* kswapd must be awake if processes are being throttled */
2622 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2623 pgdat->classzone_idx = min(pgdat->classzone_idx,
2624 (enum zone_type)ZONE_NORMAL);
2625 wake_up_interruptible(&pgdat->kswapd_wait);
2626 }
2627
2628 return wmark_ok;
2629}
2630
2631/*
2632 * Throttle direct reclaimers if backing storage is backed by the network
2633 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2634 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002635 * when the low watermark is reached.
2636 *
2637 * Returns true if a fatal signal was delivered during throttling. If this
2638 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002639 */
Mel Gorman50694c22012-11-26 16:29:48 -08002640static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002641 nodemask_t *nodemask)
2642{
Mel Gorman675becc2014-06-04 16:07:35 -07002643 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002644 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002645 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002646
2647 /*
2648 * Kernel threads should not be throttled as they may be indirectly
2649 * responsible for cleaning pages necessary for reclaim to make forward
2650 * progress. kjournald for example may enter direct reclaim while
2651 * committing a transaction where throttling it could forcing other
2652 * processes to block on log_wait_commit().
2653 */
2654 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002655 goto out;
2656
2657 /*
2658 * If a fatal signal is pending, this process should not throttle.
2659 * It should return quickly so it can exit and free its memory
2660 */
2661 if (fatal_signal_pending(current))
2662 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002663
Mel Gorman675becc2014-06-04 16:07:35 -07002664 /*
2665 * Check if the pfmemalloc reserves are ok by finding the first node
2666 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2667 * GFP_KERNEL will be required for allocating network buffers when
2668 * swapping over the network so ZONE_HIGHMEM is unusable.
2669 *
2670 * Throttling is based on the first usable node and throttled processes
2671 * wait on a queue until kswapd makes progress and wakes them. There
2672 * is an affinity then between processes waking up and where reclaim
2673 * progress has been made assuming the process wakes on the same node.
2674 * More importantly, processes running on remote nodes will not compete
2675 * for remote pfmemalloc reserves and processes on different nodes
2676 * should make reasonable progress.
2677 */
2678 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002679 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002680 if (zone_idx(zone) > ZONE_NORMAL)
2681 continue;
2682
2683 /* Throttle based on the first usable node */
2684 pgdat = zone->zone_pgdat;
2685 if (pfmemalloc_watermark_ok(pgdat))
2686 goto out;
2687 break;
2688 }
2689
2690 /* If no zone was usable by the allocation flags then do not throttle */
2691 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002692 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002693
Mel Gorman68243e72012-07-31 16:44:39 -07002694 /* Account for the throttling */
2695 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2696
Mel Gorman55150612012-07-31 16:44:35 -07002697 /*
2698 * If the caller cannot enter the filesystem, it's possible that it
2699 * is due to the caller holding an FS lock or performing a journal
2700 * transaction in the case of a filesystem like ext[3|4]. In this case,
2701 * it is not safe to block on pfmemalloc_wait as kswapd could be
2702 * blocked waiting on the same lock. Instead, throttle for up to a
2703 * second before continuing.
2704 */
2705 if (!(gfp_mask & __GFP_FS)) {
2706 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2707 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002708
2709 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002710 }
2711
2712 /* Throttle until kswapd wakes the process */
2713 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2714 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002715
2716check_pending:
2717 if (fatal_signal_pending(current))
2718 return true;
2719
2720out:
2721 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002722}
2723
Mel Gormandac1d272008-04-28 02:12:12 -07002724unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002725 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002726{
Mel Gorman33906bc2010-08-09 17:19:16 -07002727 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002728 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002729 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002730 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2731 .order = order,
2732 .nodemask = nodemask,
2733 .priority = DEF_PRIORITY,
2734 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002735 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002736 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002737 };
2738
Mel Gorman55150612012-07-31 16:44:35 -07002739 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002740 * Do not enter reclaim if fatal signal was delivered while throttled.
2741 * 1 is returned so that the page allocator does not OOM kill at this
2742 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002743 */
Mel Gorman50694c22012-11-26 16:29:48 -08002744 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002745 return 1;
2746
Mel Gorman33906bc2010-08-09 17:19:16 -07002747 trace_mm_vmscan_direct_reclaim_begin(order,
2748 sc.may_writepage,
2749 gfp_mask);
2750
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002751 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002752
2753 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2754
2755 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002756}
2757
Andrew Mortonc255a452012-07-31 16:43:02 -07002758#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002759
Johannes Weiner72835c82012-01-12 17:18:32 -08002760unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002761 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002762 struct zone *zone,
2763 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002764{
2765 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002766 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002767 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002768 .may_writepage = !laptop_mode,
2769 .may_unmap = 1,
2770 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002771 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002772 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002773 int swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002774 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002775
Balbir Singh4e416952009-09-23 15:56:39 -07002776 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2777 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002778
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002779 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002780 sc.may_writepage,
2781 sc.gfp_mask);
2782
Balbir Singh4e416952009-09-23 15:56:39 -07002783 /*
2784 * NOTE: Although we can get the priority field, using it
2785 * here is not a good idea, since it limits the pages we can scan.
2786 * if we don't reclaim here, the shrink_zone from balance_pgdat
2787 * will pick up pages from other mem cgroup's as well. We hack
2788 * the priority and make it zero.
2789 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002790 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002791
2792 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2793
Ying Han0ae5e892011-05-26 16:25:25 -07002794 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002795 return sc.nr_reclaimed;
2796}
2797
Johannes Weiner72835c82012-01-12 17:18:32 -08002798unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002799 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002800 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002801 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002802{
Balbir Singh4e416952009-09-23 15:56:39 -07002803 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002804 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002805 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002806 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002807 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002808 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2809 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2810 .target_mem_cgroup = memcg,
2811 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002812 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002813 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002814 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002815 };
Balbir Singh66e17072008-02-07 00:13:56 -08002816
Ying Han889976d2011-05-26 16:25:33 -07002817 /*
2818 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2819 * take care of from where we get pages. So the node where we start the
2820 * scan does not need to be the current node.
2821 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002822 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002823
2824 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002825
2826 trace_mm_vmscan_memcg_reclaim_begin(0,
2827 sc.may_writepage,
2828 sc.gfp_mask);
2829
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002830 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002831
2832 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2833
2834 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002835}
2836#endif
2837
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002838static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002839{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002840 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002841
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002842 if (!total_swap_pages)
2843 return;
2844
2845 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2846 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002847 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002848
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002849 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002850 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002851 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002852
2853 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2854 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002855}
2856
Johannes Weiner60cefed2012-11-29 13:54:23 -08002857static bool zone_balanced(struct zone *zone, int order,
2858 unsigned long balance_gap, int classzone_idx)
2859{
2860 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2861 balance_gap, classzone_idx, 0))
2862 return false;
2863
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002864 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2865 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002866 return false;
2867
2868 return true;
2869}
2870
Mel Gorman1741c872011-01-13 15:46:21 -08002871/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002872 * pgdat_balanced() is used when checking if a node is balanced.
2873 *
2874 * For order-0, all zones must be balanced!
2875 *
2876 * For high-order allocations only zones that meet watermarks and are in a
2877 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2878 * total of balanced pages must be at least 25% of the zones allowed by
2879 * classzone_idx for the node to be considered balanced. Forcing all zones to
2880 * be balanced for high orders can cause excessive reclaim when there are
2881 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002882 * The choice of 25% is due to
2883 * o a 16M DMA zone that is balanced will not balance a zone on any
2884 * reasonable sized machine
2885 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002886 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002887 * would need to be at least 256M for it to be balance a whole node.
2888 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2889 * to balance a node on its own. These seemed like reasonable ratios.
2890 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002891static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002892{
Jiang Liub40da042013-02-22 16:33:52 -08002893 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002894 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002895 int i;
2896
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002897 /* Check the watermark levels */
2898 for (i = 0; i <= classzone_idx; i++) {
2899 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002900
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002901 if (!populated_zone(zone))
2902 continue;
2903
Jiang Liub40da042013-02-22 16:33:52 -08002904 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002905
2906 /*
2907 * A special case here:
2908 *
2909 * balance_pgdat() skips over all_unreclaimable after
2910 * DEF_PRIORITY. Effectively, it considers them balanced so
2911 * they must be considered balanced here as well!
2912 */
Lisa Du6e543d52013-09-11 14:22:36 -07002913 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002914 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002915 continue;
2916 }
2917
2918 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002919 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002920 else if (!order)
2921 return false;
2922 }
2923
2924 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002925 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002926 else
2927 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002928}
2929
Mel Gorman55150612012-07-31 16:44:35 -07002930/*
2931 * Prepare kswapd for sleeping. This verifies that there are no processes
2932 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2933 *
2934 * Returns true if kswapd is ready to sleep
2935 */
2936static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002937 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002938{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002939 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2940 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002941 return false;
2942
2943 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002944 * The throttled processes are normally woken up in balance_pgdat() as
2945 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2946 * race between when kswapd checks the watermarks and a process gets
2947 * throttled. There is also a potential race if processes get
2948 * throttled, kswapd wakes, a large process exits thereby balancing the
2949 * zones, which causes kswapd to exit balance_pgdat() before reaching
2950 * the wake up checks. If kswapd is going to sleep, no process should
2951 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2952 * the wake up is premature, processes will wake kswapd and get
2953 * throttled again. The difference from wake ups in balance_pgdat() is
2954 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07002955 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002956 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2957 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002958
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002959 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002960}
2961
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962/*
Mel Gorman75485362013-07-03 15:01:42 -07002963 * kswapd shrinks the zone by the number of pages required to reach
2964 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002965 *
2966 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002967 * reclaim or if the lack of progress was due to pages under writeback.
2968 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002969 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002970static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002971 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002972 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002973 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002974{
Mel Gorman7c954f62013-07-03 15:01:54 -07002975 int testorder = sc->order;
2976 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07002977 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002978
2979 /* Reclaim above the high watermark. */
2980 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002981
2982 /*
2983 * Kswapd reclaims only single pages with compaction enabled. Trying
2984 * too hard to reclaim until contiguous free pages have become
2985 * available can hurt performance by evicting too much useful data
2986 * from memory. Do not reclaim more than needed for compaction.
2987 */
2988 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002989 compaction_suitable(zone, sc->order, 0, classzone_idx)
2990 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07002991 testorder = 0;
2992
2993 /*
2994 * We put equal pressure on every zone, unless one zone has way too
2995 * many pages free already. The "too many pages" is defined as the
2996 * high wmark plus a "gap" where the gap is either the low
2997 * watermark or 1% of the zone, whichever is smaller.
2998 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002999 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3000 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003001
3002 /*
3003 * If there is no low memory pressure or the zone is balanced then no
3004 * reclaim is necessary
3005 */
3006 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3007 if (!lowmem_pressure && zone_balanced(zone, testorder,
3008 balance_gap, classzone_idx))
3009 return true;
3010
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003011 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003012
Mel Gorman2ab44f42013-07-03 15:01:47 -07003013 /* Account for the number of pages attempted to reclaim */
3014 *nr_attempted += sc->nr_to_reclaim;
3015
Johannes Weiner57054652014-10-09 15:28:17 -07003016 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003017
Mel Gorman7c954f62013-07-03 15:01:54 -07003018 /*
3019 * If a zone reaches its high watermark, consider it to be no longer
3020 * congested. It's possible there are dirty pages backed by congested
3021 * BDIs but as pressure is relieved, speculatively avoid congestion
3022 * waits.
3023 */
Lisa Du6e543d52013-09-11 14:22:36 -07003024 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003025 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003026 clear_bit(ZONE_CONGESTED, &zone->flags);
3027 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003028 }
3029
Mel Gormanb8e83b92013-07-03 15:01:45 -07003030 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003031}
3032
3033/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003035 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003037 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 *
3039 * There is special handling here for zones which are full of pinned pages.
3040 * This can happen if the pages are all mlocked, or if they are all used by
3041 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3042 * What we do is to detect the case where all pages in the zone have been
3043 * scanned twice and there has been zero successful reclaim. Mark the zone as
3044 * dead and from now on, only perform a short scan. Basically we're polling
3045 * the zone for when the problem goes away.
3046 *
3047 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003048 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3049 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3050 * lower zones regardless of the number of free pages in the lower zones. This
3051 * interoperates with the page allocator fallback scheme to ensure that aging
3052 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 */
Mel Gorman99504742011-01-13 15:46:20 -08003054static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003055 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003058 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003059 unsigned long nr_soft_reclaimed;
3060 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003061 struct scan_control sc = {
3062 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003063 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003064 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003065 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003066 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003067 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003068 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003069 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003071 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003072 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003073 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003074 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003075
3076 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003078 /*
3079 * Scan in the highmem->dma direction for the highest
3080 * zone which needs scanning
3081 */
3082 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3083 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003085 if (!populated_zone(zone))
3086 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087
Lisa Du6e543d52013-09-11 14:22:36 -07003088 if (sc.priority != DEF_PRIORITY &&
3089 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003090 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091
Rik van Riel556adec2008-10-18 20:26:34 -07003092 /*
3093 * Do some background aging of the anon list, to give
3094 * pages a chance to be referenced before reclaiming.
3095 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003096 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003097
Mel Gormancc715d92012-03-21 16:34:00 -07003098 /*
3099 * If the number of buffer_heads in the machine
3100 * exceeds the maximum allowed level and this node
3101 * has a highmem zone, force kswapd to reclaim from
3102 * it to relieve lowmem pressure.
3103 */
3104 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3105 end_zone = i;
3106 break;
3107 }
3108
Johannes Weiner60cefed2012-11-29 13:54:23 -08003109 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003110 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003111 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003112 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003113 /*
3114 * If balanced, clear the dirty and congested
3115 * flags
3116 */
Johannes Weiner57054652014-10-09 15:28:17 -07003117 clear_bit(ZONE_CONGESTED, &zone->flags);
3118 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003121
Mel Gormanb8e83b92013-07-03 15:01:45 -07003122 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003123 goto out;
3124
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125 for (i = 0; i <= end_zone; i++) {
3126 struct zone *zone = pgdat->node_zones + i;
3127
Mel Gorman2ab44f42013-07-03 15:01:47 -07003128 if (!populated_zone(zone))
3129 continue;
3130
Mel Gorman2ab44f42013-07-03 15:01:47 -07003131 /*
3132 * If any zone is currently balanced then kswapd will
3133 * not call compaction as it is expected that the
3134 * necessary pages are already available.
3135 */
3136 if (pgdat_needs_compaction &&
3137 zone_watermark_ok(zone, order,
3138 low_wmark_pages(zone),
3139 *classzone_idx, 0))
3140 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 }
3142
3143 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003144 * If we're getting trouble reclaiming, start doing writepage
3145 * even in laptop mode.
3146 */
3147 if (sc.priority < DEF_PRIORITY - 2)
3148 sc.may_writepage = 1;
3149
3150 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151 * Now scan the zone in the dma->highmem direction, stopping
3152 * at the last zone which needs scanning.
3153 *
3154 * We do this because the page allocator works in the opposite
3155 * direction. This prevents the page allocator from allocating
3156 * pages behind kswapd's direction of progress, which would
3157 * cause too much scanning of the lower zones.
3158 */
3159 for (i = 0; i <= end_zone; i++) {
3160 struct zone *zone = pgdat->node_zones + i;
3161
Con Kolivasf3fe6512006-01-06 00:11:15 -08003162 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 continue;
3164
Lisa Du6e543d52013-09-11 14:22:36 -07003165 if (sc.priority != DEF_PRIORITY &&
3166 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 continue;
3168
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003170
Andrew Morton0608f432013-09-24 15:27:41 -07003171 nr_soft_scanned = 0;
3172 /*
3173 * Call soft limit reclaim before calling shrink_zone.
3174 */
3175 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3176 order, sc.gfp_mask,
3177 &nr_soft_scanned);
3178 sc.nr_reclaimed += nr_soft_reclaimed;
3179
Rik van Riel32a43302007-10-16 01:24:50 -07003180 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003181 * There should be no need to raise the scanning
3182 * priority if enough pages are already being scanned
3183 * that that high watermark would be met at 100%
3184 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003185 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003186 if (kswapd_shrink_zone(zone, end_zone,
3187 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003188 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 }
Mel Gorman55150612012-07-31 16:44:35 -07003190
3191 /*
3192 * If the low watermark is met there is no need for processes
3193 * to be throttled on pfmemalloc_wait as they should not be
3194 * able to safely make forward progress. Wake them
3195 */
3196 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3197 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003198 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003199
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003201 * Fragmentation may mean that the system cannot be rebalanced
3202 * for high-order allocations in all zones. If twice the
3203 * allocation size has been reclaimed and the zones are still
3204 * not balanced then recheck the watermarks at order-0 to
3205 * prevent kswapd reclaiming excessively. Assume that a
3206 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003208 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003209 order = sc.order = 0;
3210
Mel Gormanb8e83b92013-07-03 15:01:45 -07003211 /* Check if kswapd should be suspending */
3212 if (try_to_freeze() || kthread_should_stop())
3213 break;
3214
3215 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003216 * Compact if necessary and kswapd is reclaiming at least the
3217 * high watermark number of pages as requsted
3218 */
3219 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3220 compact_pgdat(pgdat, order);
3221
3222 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003223 * Raise priority if scanning rate is too low or there was no
3224 * progress in reclaiming pages
3225 */
3226 if (raise_priority || !sc.nr_reclaimed)
3227 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003228 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003229 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230
Mel Gormanb8e83b92013-07-03 15:01:45 -07003231out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003232 /*
Mel Gorman55150612012-07-31 16:44:35 -07003233 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003234 * makes a decision on the order we were last reclaiming at. However,
3235 * if another caller entered the allocator slow path while kswapd
3236 * was awake, order will remain at the higher level
3237 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003238 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003239 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240}
3241
Mel Gormandc83edd2011-01-13 15:46:26 -08003242static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003243{
3244 long remaining = 0;
3245 DEFINE_WAIT(wait);
3246
3247 if (freezing(current) || kthread_should_stop())
3248 return;
3249
3250 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3251
3252 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003253 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003254 remaining = schedule_timeout(HZ/10);
3255 finish_wait(&pgdat->kswapd_wait, &wait);
3256 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3257 }
3258
3259 /*
3260 * After a short sleep, check if it was a premature sleep. If not, then
3261 * go fully to sleep until explicitly woken up.
3262 */
Mel Gorman55150612012-07-31 16:44:35 -07003263 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003264 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3265
3266 /*
3267 * vmstat counters are not perfectly accurate and the estimated
3268 * value for counters such as NR_FREE_PAGES can deviate from the
3269 * true value by nr_online_cpus * threshold. To avoid the zone
3270 * watermarks being breached while under pressure, we reduce the
3271 * per-cpu vmstat threshold while kswapd is awake and restore
3272 * them before going back to sleep.
3273 */
3274 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003275
Mel Gorman62997022012-10-08 16:32:47 -07003276 /*
3277 * Compaction records what page blocks it recently failed to
3278 * isolate pages from and skips them in the future scanning.
3279 * When kswapd is going to sleep, it is reasonable to assume
3280 * that pages and compaction may succeed so reset the cache.
3281 */
3282 reset_isolation_suitable(pgdat);
3283
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003284 if (!kthread_should_stop())
3285 schedule();
3286
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003287 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3288 } else {
3289 if (remaining)
3290 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3291 else
3292 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3293 }
3294 finish_wait(&pgdat->kswapd_wait, &wait);
3295}
3296
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297/*
3298 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003299 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 *
3301 * This basically trickles out pages so that we have _some_
3302 * free memory available even if there is no other activity
3303 * that frees anything up. This is needed for things like routing
3304 * etc, where we otherwise might have all activity going on in
3305 * asynchronous contexts that cannot page things out.
3306 *
3307 * If there are applications that are active memory-allocators
3308 * (most normal use), this basically shouldn't matter.
3309 */
3310static int kswapd(void *p)
3311{
Mel Gorman215ddd62011-07-08 15:39:40 -07003312 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003313 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003314 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003315 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 pg_data_t *pgdat = (pg_data_t*)p;
3317 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003318
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 struct reclaim_state reclaim_state = {
3320 .reclaimed_slab = 0,
3321 };
Rusty Russella70f7302009-03-13 14:49:46 +10303322 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323
Nick Piggincf40bd12009-01-21 08:12:39 +01003324 lockdep_set_current_reclaim_state(GFP_KERNEL);
3325
Rusty Russell174596a2009-01-01 10:12:29 +10303326 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003327 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 current->reclaim_state = &reclaim_state;
3329
3330 /*
3331 * Tell the memory management that we're a "memory allocator",
3332 * and that if we need more memory we should get access to it
3333 * regardless (see "__alloc_pages()"). "kswapd" should
3334 * never get caught in the normal page freeing logic.
3335 *
3336 * (Kswapd normally doesn't need memory anyway, but sometimes
3337 * you need a small amount of memory in order to be able to
3338 * page out something else, and this flag essentially protects
3339 * us from recursively trying to free more memory as we're
3340 * trying to free the first piece of memory in the first place).
3341 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003342 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003343 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344
Mel Gorman215ddd62011-07-08 15:39:40 -07003345 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003346 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003347 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003348 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003350 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003351
Mel Gorman215ddd62011-07-08 15:39:40 -07003352 /*
3353 * If the last balance_pgdat was unsuccessful it's unlikely a
3354 * new request of a similar or harder type will succeed soon
3355 * so consider going to sleep on the basis we reclaimed at
3356 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003357 if (balanced_classzone_idx >= new_classzone_idx &&
3358 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003359 new_order = pgdat->kswapd_max_order;
3360 new_classzone_idx = pgdat->classzone_idx;
3361 pgdat->kswapd_max_order = 0;
3362 pgdat->classzone_idx = pgdat->nr_zones - 1;
3363 }
3364
Mel Gorman99504742011-01-13 15:46:20 -08003365 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 /*
3367 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003368 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369 */
3370 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003371 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003373 kswapd_try_to_sleep(pgdat, balanced_order,
3374 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003376 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003377 new_order = order;
3378 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003379 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003380 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382
David Rientjes8fe23e02009-12-14 17:58:33 -08003383 ret = try_to_freeze();
3384 if (kthread_should_stop())
3385 break;
3386
3387 /*
3388 * We can speed up thawing tasks if we don't call balance_pgdat
3389 * after returning from the refrigerator
3390 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003391 if (!ret) {
3392 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003393 balanced_classzone_idx = classzone_idx;
3394 balanced_order = balance_pgdat(pgdat, order,
3395 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003396 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003398
Johannes Weiner71abdc12014-06-06 14:35:35 -07003399 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003400 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003401 lockdep_clear_current_reclaim_state();
3402
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 return 0;
3404}
3405
3406/*
3407 * A zone is low on free memory, so wake its kswapd task to service it.
3408 */
Mel Gorman99504742011-01-13 15:46:20 -08003409void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410{
3411 pg_data_t *pgdat;
3412
Con Kolivasf3fe6512006-01-06 00:11:15 -08003413 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 return;
3415
Vladimir Davydov344736f2014-10-20 15:50:30 +04003416 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003418 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003419 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003420 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003421 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3422 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003423 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003425 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003426 return;
3427
3428 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003429 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430}
3431
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003432#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003434 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003435 * freed pages.
3436 *
3437 * Rather than trying to age LRUs the aim is to preserve the overall
3438 * LRU order by reclaiming preferentially
3439 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003441unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003443 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003444 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003445 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003446 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003447 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003448 .may_writepage = 1,
3449 .may_unmap = 1,
3450 .may_swap = 1,
3451 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003453 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003454 struct task_struct *p = current;
3455 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003457 p->flags |= PF_MEMALLOC;
3458 lockdep_set_current_reclaim_state(sc.gfp_mask);
3459 reclaim_state.reclaimed_slab = 0;
3460 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003461
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003462 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003463
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003464 p->reclaim_state = NULL;
3465 lockdep_clear_current_reclaim_state();
3466 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003467
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003468 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003470#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472/* It's optimal to keep kswapds on the same CPUs as their memory, but
3473 not required for correctness. So if the last cpu in a node goes
3474 away, we get changed to run anywhere: as the first one comes back,
3475 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003476static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3477 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003479 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003481 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003482 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003483 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303484 const struct cpumask *mask;
3485
3486 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003487
Rusty Russell3e597942009-01-01 10:12:24 +10303488 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003490 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 }
3492 }
3493 return NOTIFY_OK;
3494}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495
Yasunori Goto3218ae12006-06-27 02:53:33 -07003496/*
3497 * This kswapd start function will be called by init and node-hot-add.
3498 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3499 */
3500int kswapd_run(int nid)
3501{
3502 pg_data_t *pgdat = NODE_DATA(nid);
3503 int ret = 0;
3504
3505 if (pgdat->kswapd)
3506 return 0;
3507
3508 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3509 if (IS_ERR(pgdat->kswapd)) {
3510 /* failure at boot is fatal */
3511 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003512 pr_err("Failed to start kswapd on node %d\n", nid);
3513 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003514 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003515 }
3516 return ret;
3517}
3518
David Rientjes8fe23e02009-12-14 17:58:33 -08003519/*
Jiang Liud8adde12012-07-11 14:01:52 -07003520 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003521 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003522 */
3523void kswapd_stop(int nid)
3524{
3525 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3526
Jiang Liud8adde12012-07-11 14:01:52 -07003527 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003528 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003529 NODE_DATA(nid)->kswapd = NULL;
3530 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003531}
3532
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533static int __init kswapd_init(void)
3534{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003535 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003536
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003538 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003539 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 hotcpu_notifier(cpu_callback, 0);
3541 return 0;
3542}
3543
3544module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003545
3546#ifdef CONFIG_NUMA
3547/*
3548 * Zone reclaim mode
3549 *
3550 * If non-zero call zone_reclaim when the number of free pages falls below
3551 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003552 */
3553int zone_reclaim_mode __read_mostly;
3554
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003555#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003556#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003557#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3558#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3559
Christoph Lameter9eeff232006-01-18 17:42:31 -08003560/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003561 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3562 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3563 * a zone.
3564 */
3565#define ZONE_RECLAIM_PRIORITY 4
3566
Christoph Lameter9eeff232006-01-18 17:42:31 -08003567/*
Christoph Lameter96146342006-07-03 00:24:13 -07003568 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3569 * occur.
3570 */
3571int sysctl_min_unmapped_ratio = 1;
3572
3573/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003574 * If the number of slab pages in a zone grows beyond this percentage then
3575 * slab reclaim needs to occur.
3576 */
3577int sysctl_min_slab_ratio = 5;
3578
Mel Gorman90afa5de2009-06-16 15:33:20 -07003579static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3580{
3581 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3582 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3583 zone_page_state(zone, NR_ACTIVE_FILE);
3584
3585 /*
3586 * It's possible for there to be more file mapped pages than
3587 * accounted for by the pages on the file LRU lists because
3588 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3589 */
3590 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3591}
3592
3593/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3594static long zone_pagecache_reclaimable(struct zone *zone)
3595{
3596 long nr_pagecache_reclaimable;
3597 long delta = 0;
3598
3599 /*
3600 * If RECLAIM_SWAP is set, then all file pages are considered
3601 * potentially reclaimable. Otherwise, we have to worry about
3602 * pages like swapcache and zone_unmapped_file_pages() provides
3603 * a better estimate
3604 */
3605 if (zone_reclaim_mode & RECLAIM_SWAP)
3606 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3607 else
3608 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3609
3610 /* If we can't clean pages, remove dirty pages from consideration */
3611 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3612 delta += zone_page_state(zone, NR_FILE_DIRTY);
3613
3614 /* Watch for any possible underflows due to delta */
3615 if (unlikely(delta > nr_pagecache_reclaimable))
3616 delta = nr_pagecache_reclaimable;
3617
3618 return nr_pagecache_reclaimable - delta;
3619}
3620
Christoph Lameter0ff38492006-09-25 23:31:52 -07003621/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003622 * Try to free up some pages from this zone through reclaim.
3623 */
Andrew Morton179e9632006-03-22 00:08:18 -08003624static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003625{
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08003626 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003627 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003628 struct task_struct *p = current;
3629 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003630 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003631 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003632 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003633 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003634 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003635 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3636 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3637 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003638 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003639
Christoph Lameter9eeff232006-01-18 17:42:31 -08003640 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003641 /*
3642 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3643 * and we also need to be able to write out pages for RECLAIM_WRITE
3644 * and RECLAIM_SWAP.
3645 */
3646 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003647 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003648 reclaim_state.reclaimed_slab = 0;
3649 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003650
Mel Gorman90afa5de2009-06-16 15:33:20 -07003651 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003652 /*
3653 * Free memory by calling shrink zone with increasing
3654 * priorities until we have enough memory freed.
3655 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003656 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003657 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003658 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003659 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003660
Christoph Lameter9eeff232006-01-18 17:42:31 -08003661 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003662 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003663 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003664 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003665}
Andrew Morton179e9632006-03-22 00:08:18 -08003666
3667int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3668{
Andrew Morton179e9632006-03-22 00:08:18 -08003669 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003670 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003671
3672 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003673 * Zone reclaim reclaims unmapped file backed pages and
3674 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003675 *
Christoph Lameter96146342006-07-03 00:24:13 -07003676 * A small portion of unmapped file backed pages is needed for
3677 * file I/O otherwise pages read by file I/O will be immediately
3678 * thrown out if the zone is overallocated. So we do not reclaim
3679 * if less than a specified percentage of the zone is used by
3680 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003681 */
Mel Gorman90afa5de2009-06-16 15:33:20 -07003682 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3683 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003684 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003685
Lisa Du6e543d52013-09-11 14:22:36 -07003686 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003687 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003688
Andrew Morton179e9632006-03-22 00:08:18 -08003689 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003690 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003691 */
David Rientjesd773ed62007-10-16 23:26:01 -07003692 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003693 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003694
3695 /*
3696 * Only run zone reclaim on the local zone or on zones that do not
3697 * have associated processors. This will favor the local processor
3698 * over remote processors and spread off node memory allocations
3699 * as wide as possible.
3700 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003701 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003702 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003703 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003704
Johannes Weiner57054652014-10-09 15:28:17 -07003705 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003706 return ZONE_RECLAIM_NOSCAN;
3707
David Rientjesd773ed62007-10-16 23:26:01 -07003708 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003709 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003710
Mel Gorman24cf725182009-06-16 15:33:23 -07003711 if (!ret)
3712 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3713
David Rientjesd773ed62007-10-16 23:26:01 -07003714 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003715}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003716#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003717
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003718/*
3719 * page_evictable - test whether a page is evictable
3720 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003721 *
3722 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003723 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003724 *
3725 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003726 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003727 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003728 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003729 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003730int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003731{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003732 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003733}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003734
Hugh Dickins850465792012-01-20 14:34:19 -08003735#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003736/**
Hugh Dickins24513262012-01-20 14:34:21 -08003737 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3738 * @pages: array of pages to check
3739 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003740 *
Hugh Dickins24513262012-01-20 14:34:21 -08003741 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003742 *
3743 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003744 */
Hugh Dickins24513262012-01-20 14:34:21 -08003745void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003746{
Johannes Weiner925b7672012-01-12 17:18:15 -08003747 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003748 struct zone *zone = NULL;
3749 int pgscanned = 0;
3750 int pgrescued = 0;
3751 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003752
Hugh Dickins24513262012-01-20 14:34:21 -08003753 for (i = 0; i < nr_pages; i++) {
3754 struct page *page = pages[i];
3755 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003756
Hugh Dickins24513262012-01-20 14:34:21 -08003757 pgscanned++;
3758 pagezone = page_zone(page);
3759 if (pagezone != zone) {
3760 if (zone)
3761 spin_unlock_irq(&zone->lru_lock);
3762 zone = pagezone;
3763 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003764 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003765 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003766
Hugh Dickins24513262012-01-20 14:34:21 -08003767 if (!PageLRU(page) || !PageUnevictable(page))
3768 continue;
3769
Hugh Dickins39b5f292012-10-08 16:33:18 -07003770 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003771 enum lru_list lru = page_lru_base_type(page);
3772
Sasha Levin309381fea2014-01-23 15:52:54 -08003773 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003774 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003775 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3776 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003777 pgrescued++;
3778 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003779 }
Hugh Dickins24513262012-01-20 14:34:21 -08003780
3781 if (zone) {
3782 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3783 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3784 spin_unlock_irq(&zone->lru_lock);
3785 }
Hugh Dickins850465792012-01-20 14:34:19 -08003786}
3787#endif /* CONFIG_SHMEM */