blob: ab2505c3ef5460e23facf80725633e977882ec8b [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
94 unsigned int hibernation_mode:1;
95
96 /* One of the zones is ready for compaction */
97 unsigned int compaction_ready:1;
98
99 /* Incremented by the number of inactive pages that were scanned */
100 unsigned long nr_scanned;
101
102 /* Number of pages freed so far during a call to shrink_zones() */
103 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104};
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
107
108#ifdef ARCH_HAS_PREFETCH
109#define prefetch_prev_lru_page(_page, _base, _field) \
110 do { \
111 if ((_page)->lru.prev != _base) { \
112 struct page *prev; \
113 \
114 prev = lru_to_page(&(_page->lru)); \
115 prefetch(&prev->_field); \
116 } \
117 } while (0)
118#else
119#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
120#endif
121
122#ifdef ARCH_HAS_PREFETCHW
123#define prefetchw_prev_lru_page(_page, _base, _field) \
124 do { \
125 if ((_page)->lru.prev != _base) { \
126 struct page *prev; \
127 \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetchw(&prev->_field); \
130 } \
131 } while (0)
132#else
133#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
134#endif
135
136/*
137 * From 0 .. 100. Higher means more swappy.
138 */
139int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700140/*
141 * The total number of pages which are beyond the high watermark within all
142 * zones.
143 */
144unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146static LIST_HEAD(shrinker_list);
147static DECLARE_RWSEM(shrinker_rwsem);
148
Andrew Mortonc255a452012-07-31 16:43:02 -0700149#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800150static bool global_reclaim(struct scan_control *sc)
151{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800152 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800153}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800154#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155static bool global_reclaim(struct scan_control *sc)
156{
157 return true;
158}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#endif
160
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800161static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700162{
163 int nr;
164
165 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
166 zone_page_state(zone, NR_INACTIVE_FILE);
167
168 if (get_nr_swap_pages() > 0)
169 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
170 zone_page_state(zone, NR_INACTIVE_ANON);
171
172 return nr;
173}
174
175bool zone_reclaimable(struct zone *zone)
176{
Mel Gorman0d5d8232014-08-06 16:07:16 -0700177 return zone_page_state(zone, NR_PAGES_SCANNED) <
178 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700179}
180
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700181static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800182{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700183 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700184 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800185
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700186 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800187}
188
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000190 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000192int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000194 size_t size = sizeof(*shrinker->nr_deferred);
195
196 /*
197 * If we only have one possible node in the system anyway, save
198 * ourselves the trouble and disable NUMA aware behavior. This way we
199 * will save memory and some small loop time later.
200 */
201 if (nr_node_ids == 1)
202 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
203
204 if (shrinker->flags & SHRINKER_NUMA_AWARE)
205 size *= nr_node_ids;
206
207 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
208 if (!shrinker->nr_deferred)
209 return -ENOMEM;
210
Rusty Russell8e1f9362007-07-17 04:03:17 -0700211 down_write(&shrinker_rwsem);
212 list_add_tail(&shrinker->list, &shrinker_list);
213 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000214 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218/*
219 * Remove one
220 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700221void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222{
223 down_write(&shrinker_rwsem);
224 list_del(&shrinker->list);
225 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700226 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700228EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229
230#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000231
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800232static unsigned long shrink_slabs(struct shrink_control *shrinkctl,
233 struct shrinker *shrinker,
234 unsigned long nr_scanned,
235 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000237 unsigned long freed = 0;
238 unsigned long long delta;
239 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700240 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000241 long nr;
242 long new_nr;
243 int nid = shrinkctl->nid;
244 long batch_size = shrinker->batch ? shrinker->batch
245 : SHRINK_BATCH;
246
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700247 freeable = shrinker->count_objects(shrinker, shrinkctl);
248 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000249 return 0;
250
251 /*
252 * copy the current shrinker scan count into a local variable
253 * and zero it so that other concurrent shrinker invocations
254 * don't also do this scanning work.
255 */
256 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
257
258 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800259 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700260 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800261 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000262 total_scan += delta;
263 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800264 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000265 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700266 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000267 }
268
269 /*
270 * We need to avoid excessive windup on filesystem shrinkers
271 * due to large numbers of GFP_NOFS allocations causing the
272 * shrinkers to return -1 all the time. This results in a large
273 * nr being built up so when a shrink that can do some work
274 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700275 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000276 * memory.
277 *
278 * Hence only allow the shrinker to scan the entire cache when
279 * a large delta change is calculated directly.
280 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700281 if (delta < freeable / 4)
282 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000283
284 /*
285 * Avoid risking looping forever due to too large nr value:
286 * never try to free more than twice the estimate number of
287 * freeable entries.
288 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700289 if (total_scan > freeable * 2)
290 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000291
292 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800293 nr_scanned, nr_eligible,
294 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000295
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800296 /*
297 * Normally, we should not scan less than batch_size objects in one
298 * pass to avoid too frequent shrinker calls, but if the slab has less
299 * than batch_size objects in total and we are really tight on memory,
300 * we will try to reclaim all available objects, otherwise we can end
301 * up failing allocations although there are plenty of reclaimable
302 * objects spread over several slabs with usage less than the
303 * batch_size.
304 *
305 * We detect the "tight on memory" situations by looking at the total
306 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700307 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800308 * scanning at high prio and therefore should try to reclaim as much as
309 * possible.
310 */
311 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700312 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000313 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800314 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000315
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800316 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000317 ret = shrinker->scan_objects(shrinker, shrinkctl);
318 if (ret == SHRINK_STOP)
319 break;
320 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000321
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800322 count_vm_events(SLABS_SCANNED, nr_to_scan);
323 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000324
325 cond_resched();
326 }
327
328 /*
329 * move the unused scan count back into the shrinker in a
330 * manner that handles concurrent updates. If we exhausted the
331 * scan, there is no need to do an update.
332 */
333 if (total_scan > 0)
334 new_nr = atomic_long_add_return(total_scan,
335 &shrinker->nr_deferred[nid]);
336 else
337 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
338
Dave Hansendf9024a2014-06-04 16:08:07 -0700339 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000340 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800343/**
344 * shrink_node_slabs - shrink slab caches of a given node
345 * @gfp_mask: allocation context
346 * @nid: node whose slab caches to target
347 * @nr_scanned: pressure numerator
348 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800350 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800352 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
353 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800355 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
356 * the available objects should be scanned. Page reclaim for example
357 * passes the number of pages scanned and the number of pages on the
358 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
359 * when it encountered mapped pages. The ratio is further biased by
360 * the ->seeks setting of the shrink function, which indicates the
361 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800363 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800365unsigned long shrink_node_slabs(gfp_t gfp_mask, int nid,
366 unsigned long nr_scanned,
367 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
369 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000370 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800372 if (nr_scanned == 0)
373 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
Minchan Kimf06590b2011-05-24 17:11:11 -0700375 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000376 /*
377 * If we would return 0, our callers would understand that we
378 * have nothing else to shrink and give up trying. By returning
379 * 1 we keep it going and assume we'll be able to shrink next
380 * time.
381 */
382 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700383 goto out;
384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385
386 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800387 struct shrink_control sc = {
388 .gfp_mask = gfp_mask,
389 .nid = nid,
390 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800391
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800392 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
393 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000394
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800395 freed += shrink_slabs(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800397
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700399out:
400 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000401 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402}
403
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404static inline int is_page_cache_freeable(struct page *page)
405{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700406 /*
407 * A freeable page cache page is referenced only by the caller
408 * that isolated the page, the page cache radix tree and
409 * optional buffer heads at page->private.
410 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700411 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412}
413
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700414static int may_write_to_queue(struct backing_dev_info *bdi,
415 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
Christoph Lameter930d9152006-01-08 01:00:47 -0800417 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 return 1;
419 if (!bdi_write_congested(bdi))
420 return 1;
421 if (bdi == current->backing_dev_info)
422 return 1;
423 return 0;
424}
425
426/*
427 * We detected a synchronous write error writing a page out. Probably
428 * -ENOSPC. We need to propagate that into the address_space for a subsequent
429 * fsync(), msync() or close().
430 *
431 * The tricky part is that after writepage we cannot touch the mapping: nothing
432 * prevents it from being freed up. But we have a ref on the page and once
433 * that page is locked, the mapping is pinned.
434 *
435 * We're allowed to run sleeping lock_page() here because we know the caller has
436 * __GFP_FS.
437 */
438static void handle_write_error(struct address_space *mapping,
439 struct page *page, int error)
440{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100441 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700442 if (page_mapping(page) == mapping)
443 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 unlock_page(page);
445}
446
Christoph Lameter04e62a22006-06-23 02:03:38 -0700447/* possible outcome of pageout() */
448typedef enum {
449 /* failed to write page out, page is locked */
450 PAGE_KEEP,
451 /* move page to the active list, page is locked */
452 PAGE_ACTIVATE,
453 /* page has been sent to the disk successfully, page is unlocked */
454 PAGE_SUCCESS,
455 /* page is clean and locked */
456 PAGE_CLEAN,
457} pageout_t;
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/*
Andrew Morton1742f192006-03-22 00:08:21 -0800460 * pageout is called by shrink_page_list() for each dirty page.
461 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700463static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700464 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465{
466 /*
467 * If the page is dirty, only perform writeback if that write
468 * will be non-blocking. To prevent this allocation from being
469 * stalled by pagecache activity. But note that there may be
470 * stalls if we need to run get_block(). We could test
471 * PagePrivate for that.
472 *
Al Viro81742022014-04-03 03:17:43 -0400473 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 * this page's queue, we can perform writeback even if that
475 * will block.
476 *
477 * If the page is swapcache, write it back even if that would
478 * block, for some throttling. This happens by accident, because
479 * swap_backing_dev_info is bust: it doesn't reflect the
480 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 */
482 if (!is_page_cache_freeable(page))
483 return PAGE_KEEP;
484 if (!mapping) {
485 /*
486 * Some data journaling orphaned pages can have
487 * page->mapping == NULL while being dirty with clean buffers.
488 */
David Howells266cf652009-04-03 16:42:36 +0100489 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 if (try_to_free_buffers(page)) {
491 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700492 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 return PAGE_CLEAN;
494 }
495 }
496 return PAGE_KEEP;
497 }
498 if (mapping->a_ops->writepage == NULL)
499 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700500 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 return PAGE_KEEP;
502
503 if (clear_page_dirty_for_io(page)) {
504 int res;
505 struct writeback_control wbc = {
506 .sync_mode = WB_SYNC_NONE,
507 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700508 .range_start = 0,
509 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 .for_reclaim = 1,
511 };
512
513 SetPageReclaim(page);
514 res = mapping->a_ops->writepage(page, &wbc);
515 if (res < 0)
516 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800517 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 ClearPageReclaim(page);
519 return PAGE_ACTIVATE;
520 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 if (!PageWriteback(page)) {
523 /* synchronous write or broken a_ops? */
524 ClearPageReclaim(page);
525 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700526 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700527 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return PAGE_SUCCESS;
529 }
530
531 return PAGE_CLEAN;
532}
533
Andrew Mortona649fd92006-10-17 00:09:36 -0700534/*
Nick Piggine2867812008-07-25 19:45:30 -0700535 * Same as remove_mapping, but if the page is removed from the mapping, it
536 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700537 */
Johannes Weinera5289102014-04-03 14:47:51 -0700538static int __remove_mapping(struct address_space *mapping, struct page *page,
539 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800540{
Nick Piggin28e4d962006-09-25 23:31:23 -0700541 BUG_ON(!PageLocked(page));
542 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800543
Nick Piggin19fd6232008-07-25 19:45:32 -0700544 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800545 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700546 * The non racy check for a busy page.
547 *
548 * Must be careful with the order of the tests. When someone has
549 * a ref to the page, it may be possible that they dirty it then
550 * drop the reference. So if PageDirty is tested before page_count
551 * here, then the following race may occur:
552 *
553 * get_user_pages(&page);
554 * [user mapping goes away]
555 * write_to(page);
556 * !PageDirty(page) [good]
557 * SetPageDirty(page);
558 * put_page(page);
559 * !page_count(page) [good, discard it]
560 *
561 * [oops, our write_to data is lost]
562 *
563 * Reversing the order of the tests ensures such a situation cannot
564 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
565 * load is not satisfied before that of page->_count.
566 *
567 * Note that if SetPageDirty is always performed via set_page_dirty,
568 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800569 */
Nick Piggine2867812008-07-25 19:45:30 -0700570 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800571 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700572 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
573 if (unlikely(PageDirty(page))) {
574 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800575 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700576 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800577
578 if (PageSwapCache(page)) {
579 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700580 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800581 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700582 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700583 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700584 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500585 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700586 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500587
588 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700589 /*
590 * Remember a shadow entry for reclaimed file cache in
591 * order to detect refaults, thus thrashing, later on.
592 *
593 * But don't store shadows in an address space that is
594 * already exiting. This is not just an optizimation,
595 * inode reclaim needs to empty out the radix tree or
596 * the nodes are lost. Don't plant shadows behind its
597 * back.
598 */
599 if (reclaimed && page_is_file_cache(page) &&
600 !mapping_exiting(mapping))
601 shadow = workingset_eviction(mapping, page);
602 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700603 spin_unlock_irq(&mapping->tree_lock);
Linus Torvalds6072d132010-12-01 13:35:19 -0500604
605 if (freepage != NULL)
606 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800607 }
608
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800609 return 1;
610
611cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700612 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800613 return 0;
614}
615
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616/*
Nick Piggine2867812008-07-25 19:45:30 -0700617 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
618 * someone else has a ref on the page, abort and return 0. If it was
619 * successfully detached, return 1. Assumes the caller has a single ref on
620 * this page.
621 */
622int remove_mapping(struct address_space *mapping, struct page *page)
623{
Johannes Weinera5289102014-04-03 14:47:51 -0700624 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700625 /*
626 * Unfreezing the refcount with 1 rather than 2 effectively
627 * drops the pagecache ref for us without requiring another
628 * atomic operation.
629 */
630 page_unfreeze_refs(page, 1);
631 return 1;
632 }
633 return 0;
634}
635
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700636/**
637 * putback_lru_page - put previously isolated page onto appropriate LRU list
638 * @page: page to be put back to appropriate lru list
639 *
640 * Add previously isolated @page to appropriate LRU list.
641 * Page may still be unevictable for other reasons.
642 *
643 * lru_lock must not be held, interrupts must be enabled.
644 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700645void putback_lru_page(struct page *page)
646{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700647 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700648 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700649
Sasha Levin309381fea2014-01-23 15:52:54 -0800650 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700651
652redo:
653 ClearPageUnevictable(page);
654
Hugh Dickins39b5f292012-10-08 16:33:18 -0700655 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700656 /*
657 * For evictable pages, we can use the cache.
658 * In event of a race, worst case is we end up with an
659 * unevictable page on [in]active list.
660 * We know how to handle that.
661 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700662 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700663 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700664 } else {
665 /*
666 * Put unevictable pages directly on zone's unevictable
667 * list.
668 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700669 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700670 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700671 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700672 * When racing with an mlock or AS_UNEVICTABLE clearing
673 * (page is unlocked) make sure that if the other thread
674 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800675 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700676 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700677 * the page back to the evictable list.
678 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700679 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700680 */
681 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700682 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700683
684 /*
685 * page's status can change while we move it among lru. If an evictable
686 * page is on unevictable list, it never be freed. To avoid that,
687 * check after we added it to the list, again.
688 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700689 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700690 if (!isolate_lru_page(page)) {
691 put_page(page);
692 goto redo;
693 }
694 /* This means someone else dropped this page from LRU
695 * So, it will be freed or putback to LRU again. There is
696 * nothing to do here.
697 */
698 }
699
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700700 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700701 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700702 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700703 count_vm_event(UNEVICTABLE_PGCULLED);
704
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700705 put_page(page); /* drop ref from isolate */
706}
707
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800708enum page_references {
709 PAGEREF_RECLAIM,
710 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800711 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800712 PAGEREF_ACTIVATE,
713};
714
715static enum page_references page_check_references(struct page *page,
716 struct scan_control *sc)
717{
Johannes Weiner645747462010-03-05 13:42:22 -0800718 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800719 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800720
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700721 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
722 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800723 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800725 /*
726 * Mlock lost the isolation race with us. Let try_to_unmap()
727 * move the page to the unevictable list.
728 */
729 if (vm_flags & VM_LOCKED)
730 return PAGEREF_RECLAIM;
731
Johannes Weiner645747462010-03-05 13:42:22 -0800732 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700733 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800734 return PAGEREF_ACTIVATE;
735 /*
736 * All mapped pages start out with page table
737 * references from the instantiating fault, so we need
738 * to look twice if a mapped file page is used more
739 * than once.
740 *
741 * Mark it and spare it for another trip around the
742 * inactive list. Another page table reference will
743 * lead to its activation.
744 *
745 * Note: the mark is set for activated pages as well
746 * so that recently deactivated but used pages are
747 * quickly recovered.
748 */
749 SetPageReferenced(page);
750
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800751 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800752 return PAGEREF_ACTIVATE;
753
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800754 /*
755 * Activate file-backed executable pages after first usage.
756 */
757 if (vm_flags & VM_EXEC)
758 return PAGEREF_ACTIVATE;
759
Johannes Weiner645747462010-03-05 13:42:22 -0800760 return PAGEREF_KEEP;
761 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800762
763 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700764 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800765 return PAGEREF_RECLAIM_CLEAN;
766
767 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800768}
769
Mel Gormane2be15f2013-07-03 15:01:57 -0700770/* Check if a page is dirty or under writeback */
771static void page_check_dirty_writeback(struct page *page,
772 bool *dirty, bool *writeback)
773{
Mel Gormanb4597222013-07-03 15:02:05 -0700774 struct address_space *mapping;
775
Mel Gormane2be15f2013-07-03 15:01:57 -0700776 /*
777 * Anonymous pages are not handled by flushers and must be written
778 * from reclaim context. Do not stall reclaim based on them
779 */
780 if (!page_is_file_cache(page)) {
781 *dirty = false;
782 *writeback = false;
783 return;
784 }
785
786 /* By default assume that the page flags are accurate */
787 *dirty = PageDirty(page);
788 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700789
790 /* Verify dirty/writeback state if the filesystem supports it */
791 if (!page_has_private(page))
792 return;
793
794 mapping = page_mapping(page);
795 if (mapping && mapping->a_ops->is_dirty_writeback)
796 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700797}
798
Nick Piggine2867812008-07-25 19:45:30 -0700799/*
Andrew Morton1742f192006-03-22 00:08:21 -0800800 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 */
Andrew Morton1742f192006-03-22 00:08:21 -0800802static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700803 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700804 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700805 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700806 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700807 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700808 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700809 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700810 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700811 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812{
813 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700814 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700816 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700817 unsigned long nr_dirty = 0;
818 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800819 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700820 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700821 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
823 cond_resched();
824
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 while (!list_empty(page_list)) {
826 struct address_space *mapping;
827 struct page *page;
828 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700829 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700830 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 cond_resched();
833
834 page = lru_to_page(page_list);
835 list_del(&page->lru);
836
Nick Piggin529ae9a2008-08-02 12:01:03 +0200837 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 goto keep;
839
Sasha Levin309381fea2014-01-23 15:52:54 -0800840 VM_BUG_ON_PAGE(PageActive(page), page);
841 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
843 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800844
Hugh Dickins39b5f292012-10-08 16:33:18 -0700845 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700846 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700847
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700848 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800849 goto keep_locked;
850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 /* Double the slab pressure for mapped and swapcache pages */
852 if (page_mapped(page) || PageSwapCache(page))
853 sc->nr_scanned++;
854
Andy Whitcroftc661b072007-08-22 14:01:26 -0700855 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
856 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
857
Mel Gorman283aba92013-07-03 15:01:51 -0700858 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700859 * The number of dirty pages determines if a zone is marked
860 * reclaim_congested which affects wait_iff_congested. kswapd
861 * will stall and start writing pages if the tail of the LRU
862 * is all dirty unqueued pages.
863 */
864 page_check_dirty_writeback(page, &dirty, &writeback);
865 if (dirty || writeback)
866 nr_dirty++;
867
868 if (dirty && !writeback)
869 nr_unqueued_dirty++;
870
Mel Gormand04e8ac2013-07-03 15:02:03 -0700871 /*
872 * Treat this page as congested if the underlying BDI is or if
873 * pages are cycling through the LRU so quickly that the
874 * pages marked for immediate reclaim are making it to the
875 * end of the LRU a second time.
876 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700877 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800878 if (((dirty || writeback) && mapping &&
879 bdi_write_congested(mapping->backing_dev_info)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700880 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700881 nr_congested++;
882
883 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700884 * If a page at the tail of the LRU is under writeback, there
885 * are three cases to consider.
886 *
887 * 1) If reclaim is encountering an excessive number of pages
888 * under writeback and this page is both under writeback and
889 * PageReclaim then it indicates that pages are being queued
890 * for IO but are being recycled through the LRU before the
891 * IO can complete. Waiting on the page itself risks an
892 * indefinite stall if it is impossible to writeback the
893 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700894 * note that the LRU is being scanned too quickly and the
895 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700896 *
897 * 2) Global reclaim encounters a page, memcg encounters a
898 * page that is not marked for immediate reclaim or
899 * the caller does not have __GFP_IO. In this case mark
900 * the page for immediate reclaim and continue scanning.
901 *
902 * __GFP_IO is checked because a loop driver thread might
903 * enter reclaim, and deadlock if it waits on a page for
904 * which it is needed to do the write (loop masks off
905 * __GFP_IO|__GFP_FS for this reason); but more thought
906 * would probably show more reasons.
907 *
908 * Don't require __GFP_FS, since we're not going into the
909 * FS, just waiting on its writeback completion. Worryingly,
910 * ext4 gfs2 and xfs allocate pages with
911 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
912 * may_enter_fs here is liable to OOM on them.
913 *
914 * 3) memcg encounters a page that is not already marked
915 * PageReclaim. memcg does not have any dirty pages
916 * throttling so we could easily OOM just because too many
917 * pages are in writeback and there is nothing else to
918 * reclaim. Wait for the writeback to complete.
919 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700920 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700921 /* Case 1 above */
922 if (current_is_kswapd() &&
923 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700924 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700925 nr_immediate++;
926 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700927
928 /* Case 2 above */
929 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700930 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
931 /*
932 * This is slightly racy - end_page_writeback()
933 * might have just cleared PageReclaim, then
934 * setting PageReclaim here end up interpreted
935 * as PageReadahead - but that does not matter
936 * enough to care. What we do want is for this
937 * page to have PageReclaim set next time memcg
938 * reclaim reaches the tests above, so it will
939 * then wait_on_page_writeback() to avoid OOM;
940 * and it's also appropriate in global reclaim.
941 */
942 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700943 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700944
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700945 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700946
947 /* Case 3 above */
948 } else {
949 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700950 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700951 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
Minchan Kim02c6de82012-10-08 16:31:55 -0700953 if (!force_reclaim)
954 references = page_check_references(page, sc);
955
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800956 switch (references) {
957 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800959 case PAGEREF_KEEP:
960 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800961 case PAGEREF_RECLAIM:
962 case PAGEREF_RECLAIM_CLEAN:
963 ; /* try to reclaim the page below */
964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 /*
967 * Anonymous process memory has backing store?
968 * Try to allocate it some swap space here.
969 */
Nick Pigginb291f002008-10-18 20:26:44 -0700970 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800971 if (!(sc->gfp_mask & __GFP_IO))
972 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700973 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800975 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976
Mel Gormane2be15f2013-07-03 15:01:57 -0700977 /* Adding to swap updated mapping */
978 mapping = page_mapping(page);
979 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
981 /*
982 * The page is mapped into the page tables of one or more
983 * processes. Try to unmap it here.
984 */
985 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700986 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 case SWAP_FAIL:
988 goto activate_locked;
989 case SWAP_AGAIN:
990 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700991 case SWAP_MLOCK:
992 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 case SWAP_SUCCESS:
994 ; /* try to free the page below */
995 }
996 }
997
998 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700999 /*
1000 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001001 * avoid risk of stack overflow but only writeback
1002 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001003 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001004 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001005 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001006 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001007 /*
1008 * Immediately reclaim when written back.
1009 * Similar in principal to deactivate_page()
1010 * except we already have the page isolated
1011 * and know it's dirty
1012 */
1013 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1014 SetPageReclaim(page);
1015
Mel Gormanee728862011-10-31 17:07:38 -07001016 goto keep_locked;
1017 }
1018
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001019 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001021 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001023 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 goto keep_locked;
1025
1026 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001027 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 case PAGE_KEEP:
1029 goto keep_locked;
1030 case PAGE_ACTIVATE:
1031 goto activate_locked;
1032 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001033 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001034 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001035 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001037
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 /*
1039 * A synchronous write - probably a ramdisk. Go
1040 * ahead and try to reclaim the page.
1041 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001042 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 goto keep;
1044 if (PageDirty(page) || PageWriteback(page))
1045 goto keep_locked;
1046 mapping = page_mapping(page);
1047 case PAGE_CLEAN:
1048 ; /* try to free the page below */
1049 }
1050 }
1051
1052 /*
1053 * If the page has buffers, try to free the buffer mappings
1054 * associated with this page. If we succeed we try to free
1055 * the page as well.
1056 *
1057 * We do this even if the page is PageDirty().
1058 * try_to_release_page() does not perform I/O, but it is
1059 * possible for a page to have PageDirty set, but it is actually
1060 * clean (all its buffers are clean). This happens if the
1061 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001062 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 * try_to_release_page() will discover that cleanness and will
1064 * drop the buffers and mark the page clean - it can be freed.
1065 *
1066 * Rarely, pages can have buffers and no ->mapping. These are
1067 * the pages which were not successfully invalidated in
1068 * truncate_complete_page(). We try to drop those buffers here
1069 * and if that worked, and the page is no longer mapped into
1070 * process address space (page_count == 1) it can be freed.
1071 * Otherwise, leave the page on the LRU so it is swappable.
1072 */
David Howells266cf652009-04-03 16:42:36 +01001073 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 if (!try_to_release_page(page, sc->gfp_mask))
1075 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001076 if (!mapping && page_count(page) == 1) {
1077 unlock_page(page);
1078 if (put_page_testzero(page))
1079 goto free_it;
1080 else {
1081 /*
1082 * rare race with speculative reference.
1083 * the speculative reference will free
1084 * this page shortly, so we may
1085 * increment nr_reclaimed here (and
1086 * leave it off the LRU).
1087 */
1088 nr_reclaimed++;
1089 continue;
1090 }
1091 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 }
1093
Johannes Weinera5289102014-04-03 14:47:51 -07001094 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001095 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096
Nick Piggina978d6f2008-10-18 20:26:58 -07001097 /*
1098 * At this point, we have no other references and there is
1099 * no way to pick any more up (removed from LRU, removed
1100 * from pagecache). Can use non-atomic bitops now (and
1101 * we obviously don't have to worry about waking up a process
1102 * waiting on the page lock, because there are no references.
1103 */
1104 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001105free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001106 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001107
1108 /*
1109 * Is there need to periodically free_page_list? It would
1110 * appear not as the counts should be low
1111 */
1112 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 continue;
1114
Nick Pigginb291f002008-10-18 20:26:44 -07001115cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001116 if (PageSwapCache(page))
1117 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001118 unlock_page(page);
1119 putback_lru_page(page);
1120 continue;
1121
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001123 /* Not a candidate for swapping, so reclaim swap space. */
1124 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001125 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001126 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 SetPageActive(page);
1128 pgactivate++;
1129keep_locked:
1130 unlock_page(page);
1131keep:
1132 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001133 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001135
Johannes Weiner747db952014-08-08 14:19:24 -07001136 mem_cgroup_uncharge_list(&free_pages);
Mel Gormanb745bc82014-06-04 16:10:22 -07001137 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001140 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001141
Mel Gorman8e950282013-07-03 15:02:02 -07001142 *ret_nr_dirty += nr_dirty;
1143 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001144 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001145 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001146 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001147 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148}
1149
Minchan Kim02c6de82012-10-08 16:31:55 -07001150unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1151 struct list_head *page_list)
1152{
1153 struct scan_control sc = {
1154 .gfp_mask = GFP_KERNEL,
1155 .priority = DEF_PRIORITY,
1156 .may_unmap = 1,
1157 };
Mel Gorman8e950282013-07-03 15:02:02 -07001158 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001159 struct page *page, *next;
1160 LIST_HEAD(clean_pages);
1161
1162 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001163 if (page_is_file_cache(page) && !PageDirty(page) &&
1164 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001165 ClearPageActive(page);
1166 list_move(&page->lru, &clean_pages);
1167 }
1168 }
1169
1170 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001171 TTU_UNMAP|TTU_IGNORE_ACCESS,
1172 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001173 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001174 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001175 return ret;
1176}
1177
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001178/*
1179 * Attempt to remove the specified page from its LRU. Only take this page
1180 * if it is of the appropriate PageActive status. Pages which are being
1181 * freed elsewhere are also ignored.
1182 *
1183 * page: page to consider
1184 * mode: one of the LRU isolation modes defined above
1185 *
1186 * returns 0 on success, -ve errno on failure.
1187 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001188int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001189{
1190 int ret = -EINVAL;
1191
1192 /* Only take pages on the LRU. */
1193 if (!PageLRU(page))
1194 return ret;
1195
Minchan Kime46a2872012-10-08 16:33:48 -07001196 /* Compaction should not handle unevictable pages but CMA can do so */
1197 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001198 return ret;
1199
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001200 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001201
Mel Gormanc8244932012-01-12 17:19:38 -08001202 /*
1203 * To minimise LRU disruption, the caller can indicate that it only
1204 * wants to isolate pages it will be able to operate on without
1205 * blocking - clean pages for the most part.
1206 *
1207 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1208 * is used by reclaim when it is cannot write to backing storage
1209 *
1210 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1211 * that it is possible to migrate without blocking
1212 */
1213 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1214 /* All the caller can do on PageWriteback is block */
1215 if (PageWriteback(page))
1216 return ret;
1217
1218 if (PageDirty(page)) {
1219 struct address_space *mapping;
1220
1221 /* ISOLATE_CLEAN means only clean pages */
1222 if (mode & ISOLATE_CLEAN)
1223 return ret;
1224
1225 /*
1226 * Only pages without mappings or that have a
1227 * ->migratepage callback are possible to migrate
1228 * without blocking
1229 */
1230 mapping = page_mapping(page);
1231 if (mapping && !mapping->a_ops->migratepage)
1232 return ret;
1233 }
1234 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001235
Minchan Kimf80c0672011-10-31 17:06:55 -07001236 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1237 return ret;
1238
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001239 if (likely(get_page_unless_zero(page))) {
1240 /*
1241 * Be careful not to clear PageLRU until after we're
1242 * sure the page is not being freed elsewhere -- the
1243 * page release code relies on it.
1244 */
1245 ClearPageLRU(page);
1246 ret = 0;
1247 }
1248
1249 return ret;
1250}
1251
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001252/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 * zone->lru_lock is heavily contended. Some of the functions that
1254 * shrink the lists perform better by taking out a batch of pages
1255 * and working on them outside the LRU lock.
1256 *
1257 * For pagecache intensive workloads, this function is the hottest
1258 * spot in the kernel (apart from copy_*_user functions).
1259 *
1260 * Appropriate locks must be held before calling this function.
1261 *
1262 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001263 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001265 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001266 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001267 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001268 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 *
1270 * returns how many pages were moved onto *@dst.
1271 */
Andrew Morton69e05942006-03-22 00:08:19 -08001272static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001273 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001274 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001275 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001277 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001278 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001279 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001281 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001282 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001283 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001284
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 page = lru_to_page(src);
1286 prefetchw_prev_lru_page(page, src, flags);
1287
Sasha Levin309381fea2014-01-23 15:52:54 -08001288 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001289
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001290 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001291 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001292 nr_pages = hpage_nr_pages(page);
1293 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001294 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001295 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001296 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001297
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001298 case -EBUSY:
1299 /* else it is being freed elsewhere */
1300 list_move(&page->lru, src);
1301 continue;
1302
1303 default:
1304 BUG();
1305 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 }
1307
Hugh Dickinsf6260122012-01-12 17:20:06 -08001308 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001309 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1310 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 return nr_taken;
1312}
1313
Nick Piggin62695a82008-10-18 20:26:09 -07001314/**
1315 * isolate_lru_page - tries to isolate a page from its LRU list
1316 * @page: page to isolate from its LRU list
1317 *
1318 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1319 * vmstat statistic corresponding to whatever LRU list the page was on.
1320 *
1321 * Returns 0 if the page was removed from an LRU list.
1322 * Returns -EBUSY if the page was not on an LRU list.
1323 *
1324 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001325 * the active list, it will have PageActive set. If it was found on
1326 * the unevictable list, it will have the PageUnevictable bit set. That flag
1327 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001328 *
1329 * The vmstat statistic corresponding to the list on which the page was
1330 * found will be decremented.
1331 *
1332 * Restrictions:
1333 * (1) Must be called with an elevated refcount on the page. This is a
1334 * fundamentnal difference from isolate_lru_pages (which is called
1335 * without a stable reference).
1336 * (2) the lru_lock must not be held.
1337 * (3) interrupts must be enabled.
1338 */
1339int isolate_lru_page(struct page *page)
1340{
1341 int ret = -EBUSY;
1342
Sasha Levin309381fea2014-01-23 15:52:54 -08001343 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001344
Nick Piggin62695a82008-10-18 20:26:09 -07001345 if (PageLRU(page)) {
1346 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001347 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001348
1349 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001350 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001351 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001352 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001353 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001354 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001355 del_page_from_lru_list(page, lruvec, lru);
1356 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001357 }
1358 spin_unlock_irq(&zone->lru_lock);
1359 }
1360 return ret;
1361}
1362
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001363/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001364 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1365 * then get resheduled. When there are massive number of tasks doing page
1366 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1367 * the LRU list will go small and be scanned faster than necessary, leading to
1368 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001369 */
1370static int too_many_isolated(struct zone *zone, int file,
1371 struct scan_control *sc)
1372{
1373 unsigned long inactive, isolated;
1374
1375 if (current_is_kswapd())
1376 return 0;
1377
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001378 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001379 return 0;
1380
1381 if (file) {
1382 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1383 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1384 } else {
1385 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1386 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1387 }
1388
Fengguang Wu3cf23842012-12-18 14:23:31 -08001389 /*
1390 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1391 * won't get blocked by normal direct-reclaimers, forming a circular
1392 * deadlock.
1393 */
1394 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1395 inactive >>= 3;
1396
Rik van Riel35cd7812009-09-21 17:01:38 -07001397 return isolated > inactive;
1398}
1399
Mel Gorman66635622010-08-09 17:19:30 -07001400static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001401putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001402{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001403 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1404 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001405 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001406
Mel Gorman66635622010-08-09 17:19:30 -07001407 /*
1408 * Put back any unfreeable pages.
1409 */
Mel Gorman66635622010-08-09 17:19:30 -07001410 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001411 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001412 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001413
Sasha Levin309381fea2014-01-23 15:52:54 -08001414 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001415 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001416 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001417 spin_unlock_irq(&zone->lru_lock);
1418 putback_lru_page(page);
1419 spin_lock_irq(&zone->lru_lock);
1420 continue;
1421 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001422
1423 lruvec = mem_cgroup_page_lruvec(page, zone);
1424
Linus Torvalds7a608572011-01-17 14:42:19 -08001425 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001426 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001427 add_page_to_lru_list(page, lruvec, lru);
1428
Mel Gorman66635622010-08-09 17:19:30 -07001429 if (is_active_lru(lru)) {
1430 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001431 int numpages = hpage_nr_pages(page);
1432 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001433 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001434 if (put_page_testzero(page)) {
1435 __ClearPageLRU(page);
1436 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001437 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001438
1439 if (unlikely(PageCompound(page))) {
1440 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001441 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001442 (*get_compound_page_dtor(page))(page);
1443 spin_lock_irq(&zone->lru_lock);
1444 } else
1445 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001446 }
1447 }
Mel Gorman66635622010-08-09 17:19:30 -07001448
Hugh Dickins3f797682012-01-12 17:20:07 -08001449 /*
1450 * To save our caller's stack, now use input list for pages to free.
1451 */
1452 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001453}
1454
1455/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001456 * If a kernel thread (such as nfsd for loop-back mounts) services
1457 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1458 * In that case we should only throttle if the backing device it is
1459 * writing to is congested. In other cases it is safe to throttle.
1460 */
1461static int current_may_throttle(void)
1462{
1463 return !(current->flags & PF_LESS_THROTTLE) ||
1464 current->backing_dev_info == NULL ||
1465 bdi_write_congested(current->backing_dev_info);
1466}
1467
1468/*
Andrew Morton1742f192006-03-22 00:08:21 -08001469 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1470 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 */
Mel Gorman66635622010-08-09 17:19:30 -07001472static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001473shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001474 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475{
1476 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001477 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001478 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001479 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001480 unsigned long nr_dirty = 0;
1481 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001482 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001483 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001484 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001485 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001486 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001487 struct zone *zone = lruvec_zone(lruvec);
1488 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001489
Rik van Riel35cd7812009-09-21 17:01:38 -07001490 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001491 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001492
1493 /* We are about to die and free our memory. Return now. */
1494 if (fatal_signal_pending(current))
1495 return SWAP_CLUSTER_MAX;
1496 }
1497
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001499
1500 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001501 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001502 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001503 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001507 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1508 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001509
1510 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1511 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1512
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001513 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001514 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001515 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001516 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001517 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001518 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001519 }
Hillf Dantond563c052012-03-21 16:34:02 -07001520 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001521
Hillf Dantond563c052012-03-21 16:34:02 -07001522 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001523 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001524
Minchan Kim02c6de82012-10-08 16:31:55 -07001525 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001526 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1527 &nr_writeback, &nr_immediate,
1528 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001529
Hugh Dickins3f797682012-01-12 17:20:07 -08001530 spin_lock_irq(&zone->lru_lock);
1531
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001532 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001533
Ying Han904249a2012-04-25 16:01:48 -07001534 if (global_reclaim(sc)) {
1535 if (current_is_kswapd())
1536 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1537 nr_reclaimed);
1538 else
1539 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1540 nr_reclaimed);
1541 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001542
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001543 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001544
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001545 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001546
1547 spin_unlock_irq(&zone->lru_lock);
1548
Johannes Weiner747db952014-08-08 14:19:24 -07001549 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001550 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001551
Mel Gorman92df3a72011-10-31 17:07:56 -07001552 /*
1553 * If reclaim is isolating dirty pages under writeback, it implies
1554 * that the long-lived page allocation rate is exceeding the page
1555 * laundering rate. Either the global limits are not being effective
1556 * at throttling processes due to the page distribution throughout
1557 * zones or there is heavy usage of a slow backing device. The
1558 * only option is to throttle from reclaim context which is not ideal
1559 * as there is no guarantee the dirtying process is throttled in the
1560 * same way balance_dirty_pages() manages.
1561 *
Mel Gorman8e950282013-07-03 15:02:02 -07001562 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1563 * of pages under pages flagged for immediate reclaim and stall if any
1564 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001565 */
Mel Gorman918fc712013-07-08 16:00:25 -07001566 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001567 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001568
Mel Gormand43006d2013-07-03 15:01:50 -07001569 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001570 * memcg will stall in page writeback so only consider forcibly
1571 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001572 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001573 if (global_reclaim(sc)) {
1574 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001575 * Tag a zone as congested if all the dirty pages scanned were
1576 * backed by a congested BDI and wait_iff_congested will stall.
1577 */
1578 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001579 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001580
1581 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001582 * If dirty pages are scanned that are not queued for IO, it
1583 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001584 * the zone ZONE_DIRTY and kswapd will start writing pages from
1585 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001586 */
1587 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001588 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001589
1590 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001591 * If kswapd scans pages marked marked for immediate
1592 * reclaim and under writeback (nr_immediate), it implies
1593 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001594 * they are written so also forcibly stall.
1595 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001596 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001597 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001598 }
Mel Gormand43006d2013-07-03 15:01:50 -07001599
Mel Gorman8e950282013-07-03 15:02:02 -07001600 /*
1601 * Stall direct reclaim for IO completions if underlying BDIs or zone
1602 * is congested. Allow kswapd to continue until it starts encountering
1603 * unqueued dirty pages or cycling through the LRU too quickly.
1604 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001605 if (!sc->hibernation_mode && !current_is_kswapd() &&
1606 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001607 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1608
Mel Gormane11da5b2010-10-26 14:21:40 -07001609 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1610 zone_idx(zone),
1611 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001612 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001613 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001614 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615}
1616
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001617/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 * This moves pages from the active list to the inactive list.
1619 *
1620 * We move them the other way if the page is referenced by one or more
1621 * processes, from rmap.
1622 *
1623 * If the pages are mostly unmapped, the processing is fast and it is
1624 * appropriate to hold zone->lru_lock across the whole operation. But if
1625 * the pages are mapped, the processing is slow (page_referenced()) so we
1626 * should drop zone->lru_lock around each page. It's impossible to balance
1627 * this, so instead we remove the pages from the LRU while processing them.
1628 * It is safe to rely on PG_active against the non-LRU pages in here because
1629 * nobody will play with that bit on a non-LRU page.
1630 *
1631 * The downside is that we have to touch page->_count against each page.
1632 * But we had to alter page->flags anyway.
1633 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001634
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001635static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001636 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001637 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001638 enum lru_list lru)
1639{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001640 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001641 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001642 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001643 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001644
Wu Fengguang3eb41402009-06-16 15:33:13 -07001645 while (!list_empty(list)) {
1646 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001647 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001648
Sasha Levin309381fea2014-01-23 15:52:54 -08001649 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001650 SetPageLRU(page);
1651
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001652 nr_pages = hpage_nr_pages(page);
1653 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001654 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001655 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001656
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001657 if (put_page_testzero(page)) {
1658 __ClearPageLRU(page);
1659 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001660 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001661
1662 if (unlikely(PageCompound(page))) {
1663 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001664 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001665 (*get_compound_page_dtor(page))(page);
1666 spin_lock_irq(&zone->lru_lock);
1667 } else
1668 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001669 }
1670 }
1671 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1672 if (!is_active_lru(lru))
1673 __count_vm_events(PGDEACTIVATE, pgmoved);
1674}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001675
Hugh Dickinsf6260122012-01-12 17:20:06 -08001676static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001677 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001678 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001679 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001681 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001682 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001683 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001685 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001686 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001688 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001689 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001690 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001691 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001692 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693
1694 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001695
1696 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001697 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001698 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001699 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001702
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001703 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1704 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001705 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001706 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001707
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001708 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001709
Hugh Dickinsf6260122012-01-12 17:20:06 -08001710 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001711 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001712 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 spin_unlock_irq(&zone->lru_lock);
1714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 while (!list_empty(&l_hold)) {
1716 cond_resched();
1717 page = lru_to_page(&l_hold);
1718 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001719
Hugh Dickins39b5f292012-10-08 16:33:18 -07001720 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001721 putback_lru_page(page);
1722 continue;
1723 }
1724
Mel Gormancc715d92012-03-21 16:34:00 -07001725 if (unlikely(buffer_heads_over_limit)) {
1726 if (page_has_private(page) && trylock_page(page)) {
1727 if (page_has_private(page))
1728 try_to_release_page(page, 0);
1729 unlock_page(page);
1730 }
1731 }
1732
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001733 if (page_referenced(page, 0, sc->target_mem_cgroup,
1734 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001735 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001736 /*
1737 * Identify referenced, file-backed active pages and
1738 * give them one more trip around the active list. So
1739 * that executable code get better chances to stay in
1740 * memory under moderate memory pressure. Anon pages
1741 * are not likely to be evicted by use-once streaming
1742 * IO, plus JVM can create lots of anon VM_EXEC pages,
1743 * so we ignore them here.
1744 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001745 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001746 list_add(&page->lru, &l_active);
1747 continue;
1748 }
1749 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001750
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001751 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 list_add(&page->lru, &l_inactive);
1753 }
1754
Andrew Mortonb5557492009-01-06 14:40:13 -08001755 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001756 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001757 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001758 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001759 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001760 * Count referenced pages from currently used mappings as rotated,
1761 * even though only some of them are actually re-activated. This
1762 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001763 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001764 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001765 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001766
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001767 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1768 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001769 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001770 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001771
Johannes Weiner747db952014-08-08 14:19:24 -07001772 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001773 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774}
1775
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001776#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001777static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001778{
1779 unsigned long active, inactive;
1780
1781 active = zone_page_state(zone, NR_ACTIVE_ANON);
1782 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1783
1784 if (inactive * zone->inactive_ratio < active)
1785 return 1;
1786
1787 return 0;
1788}
1789
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001790/**
1791 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001792 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001793 *
1794 * Returns true if the zone does not have enough inactive anon pages,
1795 * meaning some active anon pages need to be deactivated.
1796 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001797static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001798{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001799 /*
1800 * If we don't have swap space, anonymous page deactivation
1801 * is pointless.
1802 */
1803 if (!total_swap_pages)
1804 return 0;
1805
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001806 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001807 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001808
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001809 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001810}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001811#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001812static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001813{
1814 return 0;
1815}
1816#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001817
Rik van Riel56e49d22009-06-16 15:32:28 -07001818/**
1819 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001820 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001821 *
1822 * When the system is doing streaming IO, memory pressure here
1823 * ensures that active file pages get deactivated, until more
1824 * than half of the file pages are on the inactive list.
1825 *
1826 * Once we get to that situation, protect the system's working
1827 * set from being evicted by disabling active file page aging.
1828 *
1829 * This uses a different ratio than the anonymous pages, because
1830 * the page cache uses a use-once replacement algorithm.
1831 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001832static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001833{
Johannes Weinere3790142013-02-22 16:35:19 -08001834 unsigned long inactive;
1835 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001836
Johannes Weinere3790142013-02-22 16:35:19 -08001837 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1838 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1839
1840 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001841}
1842
Hugh Dickins75b00af2012-05-29 15:07:09 -07001843static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001844{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001845 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001846 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001847 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001848 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001849}
1850
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001851static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001852 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001853{
Rik van Rielb39415b2009-12-14 17:59:48 -08001854 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001855 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001856 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001857 return 0;
1858 }
1859
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001860 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001861}
1862
Johannes Weiner9a265112013-02-22 16:32:17 -08001863enum scan_balance {
1864 SCAN_EQUAL,
1865 SCAN_FRACT,
1866 SCAN_ANON,
1867 SCAN_FILE,
1868};
1869
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001871 * Determine how aggressively the anon and file LRU lists should be
1872 * scanned. The relative value of each set of LRU lists is determined
1873 * by looking at the fraction of the pages scanned we did rotate back
1874 * onto the active list instead of evict.
1875 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001876 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1877 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001878 */
Johannes Weiner02695172014-08-06 16:06:17 -07001879static void get_scan_count(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001880 struct scan_control *sc, unsigned long *nr,
1881 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001882{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001883 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001884 u64 fraction[2];
1885 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001886 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001887 unsigned long anon_prio, file_prio;
1888 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001889 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001890 bool force_scan = false;
1891 unsigned long ap, fp;
1892 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001893 bool some_scanned;
1894 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001895
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001896 /*
1897 * If the zone or memcg is small, nr[l] can be 0. This
1898 * results in no scanning on this priority and a potential
1899 * priority drop. Global direct reclaim can go to the next
1900 * zone and tends to have no problems. Global kswapd is for
1901 * zone balancing and it needs to scan a minimum amount. When
1902 * reclaiming for a memcg, a priority drop can cause high
1903 * latencies, so it's better to scan a minimum amount there as
1904 * well.
1905 */
Lisa Du6e543d52013-09-11 14:22:36 -07001906 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001907 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001908 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001909 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001910
1911 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001912 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001913 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001914 goto out;
1915 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001916
Johannes Weiner10316b32013-02-22 16:32:14 -08001917 /*
1918 * Global reclaim will swap to prevent OOM even with no
1919 * swappiness, but memcg users want to use this knob to
1920 * disable swapping for individual groups completely when
1921 * using the memory controller's swap limit feature would be
1922 * too expensive.
1923 */
Johannes Weiner02695172014-08-06 16:06:17 -07001924 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001925 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001926 goto out;
1927 }
1928
1929 /*
1930 * Do not apply any pressure balancing cleverness when the
1931 * system is close to OOM, scan both anon and file equally
1932 * (unless the swappiness setting disagrees with swapping).
1933 */
Johannes Weiner02695172014-08-06 16:06:17 -07001934 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001935 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001936 goto out;
1937 }
1938
Johannes Weiner11d16c22013-02-22 16:32:15 -08001939 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001940 * Prevent the reclaimer from falling into the cache trap: as
1941 * cache pages start out inactive, every cache fault will tip
1942 * the scan balance towards the file LRU. And as the file LRU
1943 * shrinks, so does the window for rotation from references.
1944 * This means we have a runaway feedback loop where a tiny
1945 * thrashing file LRU becomes infinitely more attractive than
1946 * anon pages. Try to detect this based on file LRU size.
1947 */
1948 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001949 unsigned long zonefile;
1950 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07001951
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001952 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1953 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1954 zone_page_state(zone, NR_INACTIVE_FILE);
1955
1956 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07001957 scan_balance = SCAN_ANON;
1958 goto out;
1959 }
1960 }
1961
1962 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001963 * There is enough inactive page cache, do not reclaim
1964 * anything from the anonymous working set right now.
1965 */
1966 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001967 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001968 goto out;
1969 }
1970
Johannes Weiner9a265112013-02-22 16:32:17 -08001971 scan_balance = SCAN_FRACT;
1972
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001973 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001974 * With swappiness at 100, anonymous and file have the same priority.
1975 * This scanning priority is essentially the inverse of IO cost.
1976 */
Johannes Weiner02695172014-08-06 16:06:17 -07001977 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001978 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001979
1980 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001981 * OK, so we have swap space and a fair amount of page cache
1982 * pages. We use the recently rotated / recently scanned
1983 * ratios to determine how valuable each cache is.
1984 *
1985 * Because workloads change over time (and to avoid overflow)
1986 * we keep these statistics as a floating average, which ends
1987 * up weighing recent references more than old ones.
1988 *
1989 * anon in [0], file in [1]
1990 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001991
1992 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1993 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1994 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1995 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1996
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001997 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001998 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001999 reclaim_stat->recent_scanned[0] /= 2;
2000 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002001 }
2002
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002003 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002004 reclaim_stat->recent_scanned[1] /= 2;
2005 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002006 }
2007
2008 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002009 * The amount of pressure on anon vs file pages is inversely
2010 * proportional to the fraction of recently scanned pages on
2011 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002012 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002013 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002014 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002015
Satoru Moriyafe350042012-05-29 15:06:47 -07002016 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002017 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002018 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002019
Shaohua Li76a33fc2010-05-24 14:32:36 -07002020 fraction[0] = ap;
2021 fraction[1] = fp;
2022 denominator = ap + fp + 1;
2023out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002024 some_scanned = false;
2025 /* Only use force_scan on second pass. */
2026 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002027 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002028 for_each_evictable_lru(lru) {
2029 int file = is_file_lru(lru);
2030 unsigned long size;
2031 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002032
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002033 size = get_lru_size(lruvec, lru);
2034 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002035
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002036 if (!scan && pass && force_scan)
2037 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002038
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002039 switch (scan_balance) {
2040 case SCAN_EQUAL:
2041 /* Scan lists relative to size */
2042 break;
2043 case SCAN_FRACT:
2044 /*
2045 * Scan types proportional to swappiness and
2046 * their relative recent reclaim efficiency.
2047 */
2048 scan = div64_u64(scan * fraction[file],
2049 denominator);
2050 break;
2051 case SCAN_FILE:
2052 case SCAN_ANON:
2053 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002054 if ((scan_balance == SCAN_FILE) != file) {
2055 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002056 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002057 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002058 break;
2059 default:
2060 /* Look ma, no brain */
2061 BUG();
2062 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002063
2064 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002065 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002066
Johannes Weiner9a265112013-02-22 16:32:17 -08002067 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002068 * Skip the second pass and don't force_scan,
2069 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002070 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002071 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002072 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002073 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002074}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002075
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002076/*
2077 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2078 */
Johannes Weiner02695172014-08-06 16:06:17 -07002079static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002080 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002081{
2082 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002083 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002084 unsigned long nr_to_scan;
2085 enum lru_list lru;
2086 unsigned long nr_reclaimed = 0;
2087 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2088 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002089 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002090
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002091 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002092
Mel Gormane82e0562013-07-03 15:01:44 -07002093 /* Record the original scan target for proportional adjustments later */
2094 memcpy(targets, nr, sizeof(nr));
2095
Mel Gorman1a501902014-06-04 16:10:49 -07002096 /*
2097 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2098 * event that can occur when there is little memory pressure e.g.
2099 * multiple streaming readers/writers. Hence, we do not abort scanning
2100 * when the requested number of pages are reclaimed when scanning at
2101 * DEF_PRIORITY on the assumption that the fact we are direct
2102 * reclaiming implies that kswapd is not keeping up and it is best to
2103 * do a batch of work at once. For memcg reclaim one check is made to
2104 * abort proportional reclaim if either the file or anon lru has already
2105 * dropped to zero at the first pass.
2106 */
2107 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2108 sc->priority == DEF_PRIORITY);
2109
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002110 blk_start_plug(&plug);
2111 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2112 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002113 unsigned long nr_anon, nr_file, percentage;
2114 unsigned long nr_scanned;
2115
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002116 for_each_evictable_lru(lru) {
2117 if (nr[lru]) {
2118 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2119 nr[lru] -= nr_to_scan;
2120
2121 nr_reclaimed += shrink_list(lru, nr_to_scan,
2122 lruvec, sc);
2123 }
2124 }
Mel Gormane82e0562013-07-03 15:01:44 -07002125
2126 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2127 continue;
2128
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002129 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002130 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002131 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002132 * proportionally what was requested by get_scan_count(). We
2133 * stop reclaiming one LRU and reduce the amount scanning
2134 * proportional to the original scan target.
2135 */
2136 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2137 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2138
Mel Gorman1a501902014-06-04 16:10:49 -07002139 /*
2140 * It's just vindictive to attack the larger once the smaller
2141 * has gone to zero. And given the way we stop scanning the
2142 * smaller below, this makes sure that we only make one nudge
2143 * towards proportionality once we've got nr_to_reclaim.
2144 */
2145 if (!nr_file || !nr_anon)
2146 break;
2147
Mel Gormane82e0562013-07-03 15:01:44 -07002148 if (nr_file > nr_anon) {
2149 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2150 targets[LRU_ACTIVE_ANON] + 1;
2151 lru = LRU_BASE;
2152 percentage = nr_anon * 100 / scan_target;
2153 } else {
2154 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2155 targets[LRU_ACTIVE_FILE] + 1;
2156 lru = LRU_FILE;
2157 percentage = nr_file * 100 / scan_target;
2158 }
2159
2160 /* Stop scanning the smaller of the LRU */
2161 nr[lru] = 0;
2162 nr[lru + LRU_ACTIVE] = 0;
2163
2164 /*
2165 * Recalculate the other LRU scan count based on its original
2166 * scan target and the percentage scanning already complete
2167 */
2168 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2169 nr_scanned = targets[lru] - nr[lru];
2170 nr[lru] = targets[lru] * (100 - percentage) / 100;
2171 nr[lru] -= min(nr[lru], nr_scanned);
2172
2173 lru += LRU_ACTIVE;
2174 nr_scanned = targets[lru] - nr[lru];
2175 nr[lru] = targets[lru] * (100 - percentage) / 100;
2176 nr[lru] -= min(nr[lru], nr_scanned);
2177
2178 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002179 }
2180 blk_finish_plug(&plug);
2181 sc->nr_reclaimed += nr_reclaimed;
2182
2183 /*
2184 * Even if we did not try to evict anon pages at all, we want to
2185 * rebalance the anon lru active/inactive ratio.
2186 */
2187 if (inactive_anon_is_low(lruvec))
2188 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2189 sc, LRU_ACTIVE_ANON);
2190
2191 throttle_vm_writeout(sc->gfp_mask);
2192}
2193
Mel Gorman23b9da52012-05-29 15:06:20 -07002194/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002195static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002196{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002197 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002198 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002199 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002200 return true;
2201
2202 return false;
2203}
2204
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002205/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002206 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2207 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2208 * true if more pages should be reclaimed such that when the page allocator
2209 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2210 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002211 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002212static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002213 unsigned long nr_reclaimed,
2214 unsigned long nr_scanned,
2215 struct scan_control *sc)
2216{
2217 unsigned long pages_for_compaction;
2218 unsigned long inactive_lru_pages;
2219
2220 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002221 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002222 return false;
2223
Mel Gorman28765922011-02-25 14:44:20 -08002224 /* Consider stopping depending on scan and reclaim activity */
2225 if (sc->gfp_mask & __GFP_REPEAT) {
2226 /*
2227 * For __GFP_REPEAT allocations, stop reclaiming if the
2228 * full LRU list has been scanned and we are still failing
2229 * to reclaim pages. This full LRU scan is potentially
2230 * expensive but a __GFP_REPEAT caller really wants to succeed
2231 */
2232 if (!nr_reclaimed && !nr_scanned)
2233 return false;
2234 } else {
2235 /*
2236 * For non-__GFP_REPEAT allocations which can presumably
2237 * fail without consequence, stop if we failed to reclaim
2238 * any pages from the last SWAP_CLUSTER_MAX number of
2239 * pages that were scanned. This will return to the
2240 * caller faster at the risk reclaim/compaction and
2241 * the resulting allocation attempt fails
2242 */
2243 if (!nr_reclaimed)
2244 return false;
2245 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002246
2247 /*
2248 * If we have not reclaimed enough pages for compaction and the
2249 * inactive lists are large enough, continue reclaiming
2250 */
2251 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002252 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002253 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002254 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002255 if (sc->nr_reclaimed < pages_for_compaction &&
2256 inactive_lru_pages > pages_for_compaction)
2257 return true;
2258
2259 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002260 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002261 case COMPACT_PARTIAL:
2262 case COMPACT_CONTINUE:
2263 return false;
2264 default:
2265 return true;
2266 }
2267}
2268
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002269static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2270 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002271{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002272 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002273 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002274
Johannes Weiner56600482012-01-12 17:17:59 -08002275 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002276 struct mem_cgroup *root = sc->target_mem_cgroup;
2277 struct mem_cgroup_reclaim_cookie reclaim = {
2278 .zone = zone,
2279 .priority = sc->priority,
2280 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002281 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002282 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002283
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002284 nr_reclaimed = sc->nr_reclaimed;
2285 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002286
Andrew Morton694fbc02013-09-24 15:27:37 -07002287 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2288 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002289 unsigned long lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002290 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002291 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002292
2293 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002294 swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002295
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002296 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2297 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002298
2299 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002300 * Direct reclaim and kswapd have to scan all memory
2301 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002302 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002303 *
2304 * Limit reclaim, on the other hand, only cares about
2305 * nr_to_reclaim pages to be reclaimed and it will
2306 * retry with decreasing priority if one round over the
2307 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002308 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002309 if (!global_reclaim(sc) &&
2310 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002311 mem_cgroup_iter_break(root, memcg);
2312 break;
2313 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002314 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2315 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002316
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002317 /*
2318 * Shrink the slab caches in the same proportion that
2319 * the eligible LRU pages were scanned.
2320 */
2321 if (global_reclaim(sc) && is_classzone) {
2322 struct reclaim_state *reclaim_state;
2323
2324 shrink_node_slabs(sc->gfp_mask, zone_to_nid(zone),
2325 sc->nr_scanned - nr_scanned,
2326 zone_lru_pages);
2327
2328 reclaim_state = current->reclaim_state;
2329 if (reclaim_state) {
2330 sc->nr_reclaimed +=
2331 reclaim_state->reclaimed_slab;
2332 reclaim_state->reclaimed_slab = 0;
2333 }
2334 }
2335
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002336 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2337 sc->nr_scanned - nr_scanned,
2338 sc->nr_reclaimed - nr_reclaimed);
2339
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002340 if (sc->nr_reclaimed - nr_reclaimed)
2341 reclaimable = true;
2342
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002343 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2344 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002345
2346 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002347}
2348
Vlastimil Babka53853e22014-10-09 15:27:02 -07002349/*
2350 * Returns true if compaction should go ahead for a high-order request, or
2351 * the high-order allocation would succeed without compaction.
2352 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002353static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002354{
2355 unsigned long balance_gap, watermark;
2356 bool watermark_ok;
2357
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002358 /*
2359 * Compaction takes time to run and there are potentially other
2360 * callers using the pages just freed. Continue reclaiming until
2361 * there is a buffer of free pages available to give compaction
2362 * a reasonable chance of completing and allocating the page
2363 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002364 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2365 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002366 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002367 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2368
2369 /*
2370 * If compaction is deferred, reclaim up to a point where
2371 * compaction will have a chance of success when re-enabled
2372 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002373 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002374 return watermark_ok;
2375
Vlastimil Babka53853e22014-10-09 15:27:02 -07002376 /*
2377 * If compaction is not ready to start and allocation is not likely
2378 * to succeed without it, then keep reclaiming.
2379 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002380 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002381 return false;
2382
2383 return watermark_ok;
2384}
2385
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386/*
2387 * This is the direct reclaim path, for page-allocating processes. We only
2388 * try to reclaim pages from zones which will satisfy the caller's allocation
2389 * request.
2390 *
Mel Gorman41858962009-06-16 15:32:12 -07002391 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2392 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2394 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002395 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2396 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2397 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 * If a zone is deemed to be full of pinned pages then just give it a light
2400 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002401 *
2402 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002404static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405{
Mel Gormandd1a2392008-04-28 02:12:17 -07002406 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002407 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002408 unsigned long nr_soft_reclaimed;
2409 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002410 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002411 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002412 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002413
Mel Gormancc715d92012-03-21 16:34:00 -07002414 /*
2415 * If the number of buffer_heads in the machine exceeds the maximum
2416 * allowed level, force direct reclaim to scan the highmem zone as
2417 * highmem pages could be pinning lowmem pages storing buffer_heads
2418 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002419 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002420 if (buffer_heads_over_limit)
2421 sc->gfp_mask |= __GFP_HIGHMEM;
2422
Mel Gormand4debc62010-08-09 17:19:29 -07002423 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002424 requested_highidx, sc->nodemask) {
2425 enum zone_type classzone_idx;
2426
Con Kolivasf3fe6512006-01-06 00:11:15 -08002427 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002429
2430 classzone_idx = requested_highidx;
2431 while (!populated_zone(zone->zone_pgdat->node_zones +
2432 classzone_idx))
2433 classzone_idx--;
2434
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002435 /*
2436 * Take care memory controller reclaiming has small influence
2437 * to global LRU.
2438 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002439 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002440 if (!cpuset_zone_allowed(zone,
2441 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002442 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002443
Lisa Du6e543d52013-09-11 14:22:36 -07002444 if (sc->priority != DEF_PRIORITY &&
2445 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002446 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002447
2448 /*
2449 * If we already have plenty of memory free for
2450 * compaction in this zone, don't free any more.
2451 * Even though compaction is invoked for any
2452 * non-zero order, only frequent costly order
2453 * reclamation is disruptive enough to become a
2454 * noticeable problem, like transparent huge
2455 * page allocations.
2456 */
2457 if (IS_ENABLED(CONFIG_COMPACTION) &&
2458 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2459 zonelist_zone_idx(z) <= requested_highidx &&
2460 compaction_ready(zone, sc->order)) {
2461 sc->compaction_ready = true;
2462 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002463 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002464
Andrew Morton0608f432013-09-24 15:27:41 -07002465 /*
2466 * This steals pages from memory cgroups over softlimit
2467 * and returns the number of reclaimed pages and
2468 * scanned pages. This works for global memory pressure
2469 * and balancing, not for a memcg's limit.
2470 */
2471 nr_soft_scanned = 0;
2472 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2473 sc->order, sc->gfp_mask,
2474 &nr_soft_scanned);
2475 sc->nr_reclaimed += nr_soft_reclaimed;
2476 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002477 if (nr_soft_reclaimed)
2478 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002479 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002480 }
Nick Piggin408d8542006-09-25 23:31:27 -07002481
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002482 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002483 reclaimable = true;
2484
2485 if (global_reclaim(sc) &&
2486 !reclaimable && zone_reclaimable(zone))
2487 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 }
Mel Gormane0c23272011-10-31 17:09:33 -07002489
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002490 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002491 * Restore to original mask to avoid the impact on the caller if we
2492 * promoted it to __GFP_HIGHMEM.
2493 */
2494 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002495
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002496 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002498
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499/*
2500 * This is the main entry point to direct page reclaim.
2501 *
2502 * If a full scan of the inactive list fails to free enough memory then we
2503 * are "out of memory" and something needs to be killed.
2504 *
2505 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2506 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002507 * caller can't do much about. We kick the writeback threads and take explicit
2508 * naps in the hope that some of these pages can be written. But if the
2509 * allocating task holds filesystem locks which prevent writeout this might not
2510 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002511 *
2512 * returns: 0, if no pages reclaimed
2513 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 */
Mel Gormandac1d272008-04-28 02:12:12 -07002515static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002516 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517{
Andrew Morton69e05942006-03-22 00:08:19 -08002518 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002519 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002520 bool zones_reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521
Keika Kobayashi873b4772008-07-25 01:48:52 -07002522 delayacct_freepages_start();
2523
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002524 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002525 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002527 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002528 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2529 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002530 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002531 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002532
Balbir Singh66e17072008-02-07 00:13:56 -08002533 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002534 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002535 break;
2536
2537 if (sc->compaction_ready)
2538 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539
2540 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002541 * If we're getting trouble reclaiming, start doing
2542 * writepage even in laptop mode.
2543 */
2544 if (sc->priority < DEF_PRIORITY - 2)
2545 sc->may_writepage = 1;
2546
2547 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 * Try to write back as many pages as we just scanned. This
2549 * tends to cause slow streaming writers to write data to the
2550 * disk smoothly, at the dirtying rate, which is nice. But
2551 * that's undesirable in laptop mode, where we *want* lumpy
2552 * writeout. So in laptop mode, write out the whole world.
2553 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002554 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2555 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002556 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2557 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002558 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002560 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002561
Keika Kobayashi873b4772008-07-25 01:48:52 -07002562 delayacct_freepages_end();
2563
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002564 if (sc->nr_reclaimed)
2565 return sc->nr_reclaimed;
2566
Mel Gorman0cee34f2012-01-12 17:19:49 -08002567 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002568 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002569 return 1;
2570
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002571 /* Any of the zones still reclaimable? Don't OOM. */
2572 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002573 return 1;
2574
2575 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576}
2577
Mel Gorman55150612012-07-31 16:44:35 -07002578static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2579{
2580 struct zone *zone;
2581 unsigned long pfmemalloc_reserve = 0;
2582 unsigned long free_pages = 0;
2583 int i;
2584 bool wmark_ok;
2585
2586 for (i = 0; i <= ZONE_NORMAL; i++) {
2587 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002588 if (!populated_zone(zone))
2589 continue;
2590
Mel Gorman55150612012-07-31 16:44:35 -07002591 pfmemalloc_reserve += min_wmark_pages(zone);
2592 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2593 }
2594
Mel Gorman675becc2014-06-04 16:07:35 -07002595 /* If there are no reserves (unexpected config) then do not throttle */
2596 if (!pfmemalloc_reserve)
2597 return true;
2598
Mel Gorman55150612012-07-31 16:44:35 -07002599 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2600
2601 /* kswapd must be awake if processes are being throttled */
2602 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2603 pgdat->classzone_idx = min(pgdat->classzone_idx,
2604 (enum zone_type)ZONE_NORMAL);
2605 wake_up_interruptible(&pgdat->kswapd_wait);
2606 }
2607
2608 return wmark_ok;
2609}
2610
2611/*
2612 * Throttle direct reclaimers if backing storage is backed by the network
2613 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2614 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002615 * when the low watermark is reached.
2616 *
2617 * Returns true if a fatal signal was delivered during throttling. If this
2618 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002619 */
Mel Gorman50694c22012-11-26 16:29:48 -08002620static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002621 nodemask_t *nodemask)
2622{
Mel Gorman675becc2014-06-04 16:07:35 -07002623 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002624 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002625 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002626
2627 /*
2628 * Kernel threads should not be throttled as they may be indirectly
2629 * responsible for cleaning pages necessary for reclaim to make forward
2630 * progress. kjournald for example may enter direct reclaim while
2631 * committing a transaction where throttling it could forcing other
2632 * processes to block on log_wait_commit().
2633 */
2634 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002635 goto out;
2636
2637 /*
2638 * If a fatal signal is pending, this process should not throttle.
2639 * It should return quickly so it can exit and free its memory
2640 */
2641 if (fatal_signal_pending(current))
2642 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002643
Mel Gorman675becc2014-06-04 16:07:35 -07002644 /*
2645 * Check if the pfmemalloc reserves are ok by finding the first node
2646 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2647 * GFP_KERNEL will be required for allocating network buffers when
2648 * swapping over the network so ZONE_HIGHMEM is unusable.
2649 *
2650 * Throttling is based on the first usable node and throttled processes
2651 * wait on a queue until kswapd makes progress and wakes them. There
2652 * is an affinity then between processes waking up and where reclaim
2653 * progress has been made assuming the process wakes on the same node.
2654 * More importantly, processes running on remote nodes will not compete
2655 * for remote pfmemalloc reserves and processes on different nodes
2656 * should make reasonable progress.
2657 */
2658 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2659 gfp_mask, nodemask) {
2660 if (zone_idx(zone) > ZONE_NORMAL)
2661 continue;
2662
2663 /* Throttle based on the first usable node */
2664 pgdat = zone->zone_pgdat;
2665 if (pfmemalloc_watermark_ok(pgdat))
2666 goto out;
2667 break;
2668 }
2669
2670 /* If no zone was usable by the allocation flags then do not throttle */
2671 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002672 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002673
Mel Gorman68243e72012-07-31 16:44:39 -07002674 /* Account for the throttling */
2675 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2676
Mel Gorman55150612012-07-31 16:44:35 -07002677 /*
2678 * If the caller cannot enter the filesystem, it's possible that it
2679 * is due to the caller holding an FS lock or performing a journal
2680 * transaction in the case of a filesystem like ext[3|4]. In this case,
2681 * it is not safe to block on pfmemalloc_wait as kswapd could be
2682 * blocked waiting on the same lock. Instead, throttle for up to a
2683 * second before continuing.
2684 */
2685 if (!(gfp_mask & __GFP_FS)) {
2686 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2687 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002688
2689 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002690 }
2691
2692 /* Throttle until kswapd wakes the process */
2693 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2694 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002695
2696check_pending:
2697 if (fatal_signal_pending(current))
2698 return true;
2699
2700out:
2701 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002702}
2703
Mel Gormandac1d272008-04-28 02:12:12 -07002704unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002705 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002706{
Mel Gorman33906bc2010-08-09 17:19:16 -07002707 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002708 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002709 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002710 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2711 .order = order,
2712 .nodemask = nodemask,
2713 .priority = DEF_PRIORITY,
2714 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002715 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002716 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002717 };
2718
Mel Gorman55150612012-07-31 16:44:35 -07002719 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002720 * Do not enter reclaim if fatal signal was delivered while throttled.
2721 * 1 is returned so that the page allocator does not OOM kill at this
2722 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002723 */
Mel Gorman50694c22012-11-26 16:29:48 -08002724 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002725 return 1;
2726
Mel Gorman33906bc2010-08-09 17:19:16 -07002727 trace_mm_vmscan_direct_reclaim_begin(order,
2728 sc.may_writepage,
2729 gfp_mask);
2730
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002731 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002732
2733 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2734
2735 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002736}
2737
Andrew Mortonc255a452012-07-31 16:43:02 -07002738#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002739
Johannes Weiner72835c82012-01-12 17:18:32 -08002740unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002741 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002742 struct zone *zone,
2743 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002744{
2745 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002746 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002747 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002748 .may_writepage = !laptop_mode,
2749 .may_unmap = 1,
2750 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002751 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002752 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002753 int swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002754 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002755
Balbir Singh4e416952009-09-23 15:56:39 -07002756 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2757 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002758
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002759 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002760 sc.may_writepage,
2761 sc.gfp_mask);
2762
Balbir Singh4e416952009-09-23 15:56:39 -07002763 /*
2764 * NOTE: Although we can get the priority field, using it
2765 * here is not a good idea, since it limits the pages we can scan.
2766 * if we don't reclaim here, the shrink_zone from balance_pgdat
2767 * will pick up pages from other mem cgroup's as well. We hack
2768 * the priority and make it zero.
2769 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002770 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002771
2772 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2773
Ying Han0ae5e892011-05-26 16:25:25 -07002774 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002775 return sc.nr_reclaimed;
2776}
2777
Johannes Weiner72835c82012-01-12 17:18:32 -08002778unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002779 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002780 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002781 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002782{
Balbir Singh4e416952009-09-23 15:56:39 -07002783 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002784 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002785 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002786 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002787 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002788 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2789 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2790 .target_mem_cgroup = memcg,
2791 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002792 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002793 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002794 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002795 };
Balbir Singh66e17072008-02-07 00:13:56 -08002796
Ying Han889976d2011-05-26 16:25:33 -07002797 /*
2798 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2799 * take care of from where we get pages. So the node where we start the
2800 * scan does not need to be the current node.
2801 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002802 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002803
2804 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002805
2806 trace_mm_vmscan_memcg_reclaim_begin(0,
2807 sc.may_writepage,
2808 sc.gfp_mask);
2809
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002810 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002811
2812 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2813
2814 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002815}
2816#endif
2817
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002818static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002819{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002820 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002821
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002822 if (!total_swap_pages)
2823 return;
2824
2825 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2826 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002827 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002828
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002829 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002830 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002831 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002832
2833 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2834 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002835}
2836
Johannes Weiner60cefed2012-11-29 13:54:23 -08002837static bool zone_balanced(struct zone *zone, int order,
2838 unsigned long balance_gap, int classzone_idx)
2839{
2840 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2841 balance_gap, classzone_idx, 0))
2842 return false;
2843
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002844 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2845 order, 0, classzone_idx) == COMPACT_SKIPPED)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002846 return false;
2847
2848 return true;
2849}
2850
Mel Gorman1741c872011-01-13 15:46:21 -08002851/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002852 * pgdat_balanced() is used when checking if a node is balanced.
2853 *
2854 * For order-0, all zones must be balanced!
2855 *
2856 * For high-order allocations only zones that meet watermarks and are in a
2857 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2858 * total of balanced pages must be at least 25% of the zones allowed by
2859 * classzone_idx for the node to be considered balanced. Forcing all zones to
2860 * be balanced for high orders can cause excessive reclaim when there are
2861 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002862 * The choice of 25% is due to
2863 * o a 16M DMA zone that is balanced will not balance a zone on any
2864 * reasonable sized machine
2865 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002866 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002867 * would need to be at least 256M for it to be balance a whole node.
2868 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2869 * to balance a node on its own. These seemed like reasonable ratios.
2870 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002871static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002872{
Jiang Liub40da042013-02-22 16:33:52 -08002873 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002874 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002875 int i;
2876
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002877 /* Check the watermark levels */
2878 for (i = 0; i <= classzone_idx; i++) {
2879 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002880
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002881 if (!populated_zone(zone))
2882 continue;
2883
Jiang Liub40da042013-02-22 16:33:52 -08002884 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002885
2886 /*
2887 * A special case here:
2888 *
2889 * balance_pgdat() skips over all_unreclaimable after
2890 * DEF_PRIORITY. Effectively, it considers them balanced so
2891 * they must be considered balanced here as well!
2892 */
Lisa Du6e543d52013-09-11 14:22:36 -07002893 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002894 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002895 continue;
2896 }
2897
2898 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002899 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002900 else if (!order)
2901 return false;
2902 }
2903
2904 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002905 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002906 else
2907 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002908}
2909
Mel Gorman55150612012-07-31 16:44:35 -07002910/*
2911 * Prepare kswapd for sleeping. This verifies that there are no processes
2912 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2913 *
2914 * Returns true if kswapd is ready to sleep
2915 */
2916static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002917 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002918{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002919 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2920 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002921 return false;
2922
2923 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002924 * The throttled processes are normally woken up in balance_pgdat() as
2925 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2926 * race between when kswapd checks the watermarks and a process gets
2927 * throttled. There is also a potential race if processes get
2928 * throttled, kswapd wakes, a large process exits thereby balancing the
2929 * zones, which causes kswapd to exit balance_pgdat() before reaching
2930 * the wake up checks. If kswapd is going to sleep, no process should
2931 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2932 * the wake up is premature, processes will wake kswapd and get
2933 * throttled again. The difference from wake ups in balance_pgdat() is
2934 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07002935 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08002936 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2937 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002938
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002939 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002940}
2941
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942/*
Mel Gorman75485362013-07-03 15:01:42 -07002943 * kswapd shrinks the zone by the number of pages required to reach
2944 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002945 *
2946 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002947 * reclaim or if the lack of progress was due to pages under writeback.
2948 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002949 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002950static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002951 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002952 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002953 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002954{
Mel Gorman7c954f62013-07-03 15:01:54 -07002955 int testorder = sc->order;
2956 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07002957 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002958
2959 /* Reclaim above the high watermark. */
2960 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002961
2962 /*
2963 * Kswapd reclaims only single pages with compaction enabled. Trying
2964 * too hard to reclaim until contiguous free pages have become
2965 * available can hurt performance by evicting too much useful data
2966 * from memory. Do not reclaim more than needed for compaction.
2967 */
2968 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002969 compaction_suitable(zone, sc->order, 0, classzone_idx)
2970 != COMPACT_SKIPPED)
Mel Gorman7c954f62013-07-03 15:01:54 -07002971 testorder = 0;
2972
2973 /*
2974 * We put equal pressure on every zone, unless one zone has way too
2975 * many pages free already. The "too many pages" is defined as the
2976 * high wmark plus a "gap" where the gap is either the low
2977 * watermark or 1% of the zone, whichever is smaller.
2978 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002979 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2980 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07002981
2982 /*
2983 * If there is no low memory pressure or the zone is balanced then no
2984 * reclaim is necessary
2985 */
2986 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2987 if (!lowmem_pressure && zone_balanced(zone, testorder,
2988 balance_gap, classzone_idx))
2989 return true;
2990
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002991 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07002992
Mel Gorman2ab44f42013-07-03 15:01:47 -07002993 /* Account for the number of pages attempted to reclaim */
2994 *nr_attempted += sc->nr_to_reclaim;
2995
Johannes Weiner57054652014-10-09 15:28:17 -07002996 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07002997
Mel Gorman7c954f62013-07-03 15:01:54 -07002998 /*
2999 * If a zone reaches its high watermark, consider it to be no longer
3000 * congested. It's possible there are dirty pages backed by congested
3001 * BDIs but as pressure is relieved, speculatively avoid congestion
3002 * waits.
3003 */
Lisa Du6e543d52013-09-11 14:22:36 -07003004 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07003005 zone_balanced(zone, testorder, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003006 clear_bit(ZONE_CONGESTED, &zone->flags);
3007 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003008 }
3009
Mel Gormanb8e83b92013-07-03 15:01:45 -07003010 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003011}
3012
3013/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003015 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003017 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 *
3019 * There is special handling here for zones which are full of pinned pages.
3020 * This can happen if the pages are all mlocked, or if they are all used by
3021 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3022 * What we do is to detect the case where all pages in the zone have been
3023 * scanned twice and there has been zero successful reclaim. Mark the zone as
3024 * dead and from now on, only perform a short scan. Basically we're polling
3025 * the zone for when the problem goes away.
3026 *
3027 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003028 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3029 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3030 * lower zones regardless of the number of free pages in the lower zones. This
3031 * interoperates with the page allocator fallback scheme to ensure that aging
3032 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 */
Mel Gorman99504742011-01-13 15:46:20 -08003034static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003035 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003038 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003039 unsigned long nr_soft_reclaimed;
3040 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003041 struct scan_control sc = {
3042 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003043 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003044 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003045 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003046 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003047 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003048 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003049 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003051 do {
Mel Gorman2ab44f42013-07-03 15:01:47 -07003052 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003053 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003054 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003055
3056 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003058 /*
3059 * Scan in the highmem->dma direction for the highest
3060 * zone which needs scanning
3061 */
3062 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3063 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003065 if (!populated_zone(zone))
3066 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067
Lisa Du6e543d52013-09-11 14:22:36 -07003068 if (sc.priority != DEF_PRIORITY &&
3069 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003070 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071
Rik van Riel556adec2008-10-18 20:26:34 -07003072 /*
3073 * Do some background aging of the anon list, to give
3074 * pages a chance to be referenced before reclaiming.
3075 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003076 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003077
Mel Gormancc715d92012-03-21 16:34:00 -07003078 /*
3079 * If the number of buffer_heads in the machine
3080 * exceeds the maximum allowed level and this node
3081 * has a highmem zone, force kswapd to reclaim from
3082 * it to relieve lowmem pressure.
3083 */
3084 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3085 end_zone = i;
3086 break;
3087 }
3088
Johannes Weiner60cefed2012-11-29 13:54:23 -08003089 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003090 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003091 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003092 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003093 /*
3094 * If balanced, clear the dirty and congested
3095 * flags
3096 */
Johannes Weiner57054652014-10-09 15:28:17 -07003097 clear_bit(ZONE_CONGESTED, &zone->flags);
3098 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003101
Mel Gormanb8e83b92013-07-03 15:01:45 -07003102 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003103 goto out;
3104
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 for (i = 0; i <= end_zone; i++) {
3106 struct zone *zone = pgdat->node_zones + i;
3107
Mel Gorman2ab44f42013-07-03 15:01:47 -07003108 if (!populated_zone(zone))
3109 continue;
3110
Mel Gorman2ab44f42013-07-03 15:01:47 -07003111 /*
3112 * If any zone is currently balanced then kswapd will
3113 * not call compaction as it is expected that the
3114 * necessary pages are already available.
3115 */
3116 if (pgdat_needs_compaction &&
3117 zone_watermark_ok(zone, order,
3118 low_wmark_pages(zone),
3119 *classzone_idx, 0))
3120 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121 }
3122
3123 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003124 * If we're getting trouble reclaiming, start doing writepage
3125 * even in laptop mode.
3126 */
3127 if (sc.priority < DEF_PRIORITY - 2)
3128 sc.may_writepage = 1;
3129
3130 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 * Now scan the zone in the dma->highmem direction, stopping
3132 * at the last zone which needs scanning.
3133 *
3134 * We do this because the page allocator works in the opposite
3135 * direction. This prevents the page allocator from allocating
3136 * pages behind kswapd's direction of progress, which would
3137 * cause too much scanning of the lower zones.
3138 */
3139 for (i = 0; i <= end_zone; i++) {
3140 struct zone *zone = pgdat->node_zones + i;
3141
Con Kolivasf3fe6512006-01-06 00:11:15 -08003142 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 continue;
3144
Lisa Du6e543d52013-09-11 14:22:36 -07003145 if (sc.priority != DEF_PRIORITY &&
3146 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 continue;
3148
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003150
Andrew Morton0608f432013-09-24 15:27:41 -07003151 nr_soft_scanned = 0;
3152 /*
3153 * Call soft limit reclaim before calling shrink_zone.
3154 */
3155 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3156 order, sc.gfp_mask,
3157 &nr_soft_scanned);
3158 sc.nr_reclaimed += nr_soft_reclaimed;
3159
Rik van Riel32a43302007-10-16 01:24:50 -07003160 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003161 * There should be no need to raise the scanning
3162 * priority if enough pages are already being scanned
3163 * that that high watermark would be met at 100%
3164 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003165 */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003166 if (kswapd_shrink_zone(zone, end_zone,
3167 &sc, &nr_attempted))
Mel Gorman7c954f62013-07-03 15:01:54 -07003168 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 }
Mel Gorman55150612012-07-31 16:44:35 -07003170
3171 /*
3172 * If the low watermark is met there is no need for processes
3173 * to be throttled on pfmemalloc_wait as they should not be
3174 * able to safely make forward progress. Wake them
3175 */
3176 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3177 pfmemalloc_watermark_ok(pgdat))
3178 wake_up(&pgdat->pfmemalloc_wait);
3179
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003181 * Fragmentation may mean that the system cannot be rebalanced
3182 * for high-order allocations in all zones. If twice the
3183 * allocation size has been reclaimed and the zones are still
3184 * not balanced then recheck the watermarks at order-0 to
3185 * prevent kswapd reclaiming excessively. Assume that a
3186 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003188 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003189 order = sc.order = 0;
3190
Mel Gormanb8e83b92013-07-03 15:01:45 -07003191 /* Check if kswapd should be suspending */
3192 if (try_to_freeze() || kthread_should_stop())
3193 break;
3194
3195 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003196 * Compact if necessary and kswapd is reclaiming at least the
3197 * high watermark number of pages as requsted
3198 */
3199 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3200 compact_pgdat(pgdat, order);
3201
3202 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003203 * Raise priority if scanning rate is too low or there was no
3204 * progress in reclaiming pages
3205 */
3206 if (raise_priority || !sc.nr_reclaimed)
3207 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003208 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003209 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210
Mel Gormanb8e83b92013-07-03 15:01:45 -07003211out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003212 /*
Mel Gorman55150612012-07-31 16:44:35 -07003213 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003214 * makes a decision on the order we were last reclaiming at. However,
3215 * if another caller entered the allocator slow path while kswapd
3216 * was awake, order will remain at the higher level
3217 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003218 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003219 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220}
3221
Mel Gormandc83edd2011-01-13 15:46:26 -08003222static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003223{
3224 long remaining = 0;
3225 DEFINE_WAIT(wait);
3226
3227 if (freezing(current) || kthread_should_stop())
3228 return;
3229
3230 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3231
3232 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003233 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003234 remaining = schedule_timeout(HZ/10);
3235 finish_wait(&pgdat->kswapd_wait, &wait);
3236 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3237 }
3238
3239 /*
3240 * After a short sleep, check if it was a premature sleep. If not, then
3241 * go fully to sleep until explicitly woken up.
3242 */
Mel Gorman55150612012-07-31 16:44:35 -07003243 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003244 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3245
3246 /*
3247 * vmstat counters are not perfectly accurate and the estimated
3248 * value for counters such as NR_FREE_PAGES can deviate from the
3249 * true value by nr_online_cpus * threshold. To avoid the zone
3250 * watermarks being breached while under pressure, we reduce the
3251 * per-cpu vmstat threshold while kswapd is awake and restore
3252 * them before going back to sleep.
3253 */
3254 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003255
Mel Gorman62997022012-10-08 16:32:47 -07003256 /*
3257 * Compaction records what page blocks it recently failed to
3258 * isolate pages from and skips them in the future scanning.
3259 * When kswapd is going to sleep, it is reasonable to assume
3260 * that pages and compaction may succeed so reset the cache.
3261 */
3262 reset_isolation_suitable(pgdat);
3263
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003264 if (!kthread_should_stop())
3265 schedule();
3266
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003267 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3268 } else {
3269 if (remaining)
3270 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3271 else
3272 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3273 }
3274 finish_wait(&pgdat->kswapd_wait, &wait);
3275}
3276
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277/*
3278 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003279 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280 *
3281 * This basically trickles out pages so that we have _some_
3282 * free memory available even if there is no other activity
3283 * that frees anything up. This is needed for things like routing
3284 * etc, where we otherwise might have all activity going on in
3285 * asynchronous contexts that cannot page things out.
3286 *
3287 * If there are applications that are active memory-allocators
3288 * (most normal use), this basically shouldn't matter.
3289 */
3290static int kswapd(void *p)
3291{
Mel Gorman215ddd62011-07-08 15:39:40 -07003292 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003293 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003294 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003295 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 pg_data_t *pgdat = (pg_data_t*)p;
3297 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003298
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 struct reclaim_state reclaim_state = {
3300 .reclaimed_slab = 0,
3301 };
Rusty Russella70f7302009-03-13 14:49:46 +10303302 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303
Nick Piggincf40bd12009-01-21 08:12:39 +01003304 lockdep_set_current_reclaim_state(GFP_KERNEL);
3305
Rusty Russell174596a2009-01-01 10:12:29 +10303306 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003307 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 current->reclaim_state = &reclaim_state;
3309
3310 /*
3311 * Tell the memory management that we're a "memory allocator",
3312 * and that if we need more memory we should get access to it
3313 * regardless (see "__alloc_pages()"). "kswapd" should
3314 * never get caught in the normal page freeing logic.
3315 *
3316 * (Kswapd normally doesn't need memory anyway, but sometimes
3317 * you need a small amount of memory in order to be able to
3318 * page out something else, and this flag essentially protects
3319 * us from recursively trying to free more memory as we're
3320 * trying to free the first piece of memory in the first place).
3321 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003322 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003323 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
Mel Gorman215ddd62011-07-08 15:39:40 -07003325 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003326 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003327 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003328 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003330 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003331
Mel Gorman215ddd62011-07-08 15:39:40 -07003332 /*
3333 * If the last balance_pgdat was unsuccessful it's unlikely a
3334 * new request of a similar or harder type will succeed soon
3335 * so consider going to sleep on the basis we reclaimed at
3336 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003337 if (balanced_classzone_idx >= new_classzone_idx &&
3338 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003339 new_order = pgdat->kswapd_max_order;
3340 new_classzone_idx = pgdat->classzone_idx;
3341 pgdat->kswapd_max_order = 0;
3342 pgdat->classzone_idx = pgdat->nr_zones - 1;
3343 }
3344
Mel Gorman99504742011-01-13 15:46:20 -08003345 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 /*
3347 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003348 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 */
3350 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003351 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003353 kswapd_try_to_sleep(pgdat, balanced_order,
3354 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003356 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003357 new_order = order;
3358 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003359 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003360 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
David Rientjes8fe23e02009-12-14 17:58:33 -08003363 ret = try_to_freeze();
3364 if (kthread_should_stop())
3365 break;
3366
3367 /*
3368 * We can speed up thawing tasks if we don't call balance_pgdat
3369 * after returning from the refrigerator
3370 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003371 if (!ret) {
3372 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003373 balanced_classzone_idx = classzone_idx;
3374 balanced_order = balance_pgdat(pgdat, order,
3375 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003376 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003378
Johannes Weiner71abdc12014-06-06 14:35:35 -07003379 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003380 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003381 lockdep_clear_current_reclaim_state();
3382
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 return 0;
3384}
3385
3386/*
3387 * A zone is low on free memory, so wake its kswapd task to service it.
3388 */
Mel Gorman99504742011-01-13 15:46:20 -08003389void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390{
3391 pg_data_t *pgdat;
3392
Con Kolivasf3fe6512006-01-06 00:11:15 -08003393 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 return;
3395
Vladimir Davydov344736f2014-10-20 15:50:30 +04003396 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003398 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003399 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003400 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003401 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3402 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003403 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003405 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003406 return;
3407
3408 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003409 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410}
3411
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003412#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003414 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003415 * freed pages.
3416 *
3417 * Rather than trying to age LRUs the aim is to preserve the overall
3418 * LRU order by reclaiming preferentially
3419 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003421unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003423 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003424 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003425 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003426 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003427 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003428 .may_writepage = 1,
3429 .may_unmap = 1,
3430 .may_swap = 1,
3431 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003433 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003434 struct task_struct *p = current;
3435 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003437 p->flags |= PF_MEMALLOC;
3438 lockdep_set_current_reclaim_state(sc.gfp_mask);
3439 reclaim_state.reclaimed_slab = 0;
3440 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003441
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003442 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003443
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003444 p->reclaim_state = NULL;
3445 lockdep_clear_current_reclaim_state();
3446 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003447
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003448 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003450#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452/* It's optimal to keep kswapds on the same CPUs as their memory, but
3453 not required for correctness. So if the last cpu in a node goes
3454 away, we get changed to run anywhere: as the first one comes back,
3455 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003456static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3457 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003459 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003461 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003462 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003463 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303464 const struct cpumask *mask;
3465
3466 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003467
Rusty Russell3e597942009-01-01 10:12:24 +10303468 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003470 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 }
3472 }
3473 return NOTIFY_OK;
3474}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475
Yasunori Goto3218ae12006-06-27 02:53:33 -07003476/*
3477 * This kswapd start function will be called by init and node-hot-add.
3478 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3479 */
3480int kswapd_run(int nid)
3481{
3482 pg_data_t *pgdat = NODE_DATA(nid);
3483 int ret = 0;
3484
3485 if (pgdat->kswapd)
3486 return 0;
3487
3488 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3489 if (IS_ERR(pgdat->kswapd)) {
3490 /* failure at boot is fatal */
3491 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003492 pr_err("Failed to start kswapd on node %d\n", nid);
3493 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003494 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003495 }
3496 return ret;
3497}
3498
David Rientjes8fe23e02009-12-14 17:58:33 -08003499/*
Jiang Liud8adde12012-07-11 14:01:52 -07003500 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003501 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003502 */
3503void kswapd_stop(int nid)
3504{
3505 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3506
Jiang Liud8adde12012-07-11 14:01:52 -07003507 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003508 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003509 NODE_DATA(nid)->kswapd = NULL;
3510 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003511}
3512
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513static int __init kswapd_init(void)
3514{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003515 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003516
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003518 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003519 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 hotcpu_notifier(cpu_callback, 0);
3521 return 0;
3522}
3523
3524module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003525
3526#ifdef CONFIG_NUMA
3527/*
3528 * Zone reclaim mode
3529 *
3530 * If non-zero call zone_reclaim when the number of free pages falls below
3531 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003532 */
3533int zone_reclaim_mode __read_mostly;
3534
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003535#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003536#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003537#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3538#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3539
Christoph Lameter9eeff232006-01-18 17:42:31 -08003540/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003541 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3542 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3543 * a zone.
3544 */
3545#define ZONE_RECLAIM_PRIORITY 4
3546
Christoph Lameter9eeff232006-01-18 17:42:31 -08003547/*
Christoph Lameter96146342006-07-03 00:24:13 -07003548 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3549 * occur.
3550 */
3551int sysctl_min_unmapped_ratio = 1;
3552
3553/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003554 * If the number of slab pages in a zone grows beyond this percentage then
3555 * slab reclaim needs to occur.
3556 */
3557int sysctl_min_slab_ratio = 5;
3558
Mel Gorman90afa5d2009-06-16 15:33:20 -07003559static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3560{
3561 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3562 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3563 zone_page_state(zone, NR_ACTIVE_FILE);
3564
3565 /*
3566 * It's possible for there to be more file mapped pages than
3567 * accounted for by the pages on the file LRU lists because
3568 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3569 */
3570 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3571}
3572
3573/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3574static long zone_pagecache_reclaimable(struct zone *zone)
3575{
3576 long nr_pagecache_reclaimable;
3577 long delta = 0;
3578
3579 /*
3580 * If RECLAIM_SWAP is set, then all file pages are considered
3581 * potentially reclaimable. Otherwise, we have to worry about
3582 * pages like swapcache and zone_unmapped_file_pages() provides
3583 * a better estimate
3584 */
3585 if (zone_reclaim_mode & RECLAIM_SWAP)
3586 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3587 else
3588 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3589
3590 /* If we can't clean pages, remove dirty pages from consideration */
3591 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3592 delta += zone_page_state(zone, NR_FILE_DIRTY);
3593
3594 /* Watch for any possible underflows due to delta */
3595 if (unlikely(delta > nr_pagecache_reclaimable))
3596 delta = nr_pagecache_reclaimable;
3597
3598 return nr_pagecache_reclaimable - delta;
3599}
3600
Christoph Lameter0ff38492006-09-25 23:31:52 -07003601/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003602 * Try to free up some pages from this zone through reclaim.
3603 */
Andrew Morton179e9632006-03-22 00:08:18 -08003604static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003605{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003606 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003607 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003608 struct task_struct *p = current;
3609 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003610 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003611 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003612 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003613 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003614 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003615 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3616 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3617 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003618 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003619
Christoph Lameter9eeff232006-01-18 17:42:31 -08003620 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003621 /*
3622 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3623 * and we also need to be able to write out pages for RECLAIM_WRITE
3624 * and RECLAIM_SWAP.
3625 */
3626 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003627 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003628 reclaim_state.reclaimed_slab = 0;
3629 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003630
Mel Gorman90afa5d2009-06-16 15:33:20 -07003631 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003632 /*
3633 * Free memory by calling shrink zone with increasing
3634 * priorities until we have enough memory freed.
3635 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003636 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003637 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003638 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003639 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003640
Christoph Lameter9eeff232006-01-18 17:42:31 -08003641 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003642 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003643 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003644 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003645}
Andrew Morton179e9632006-03-22 00:08:18 -08003646
3647int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3648{
Andrew Morton179e9632006-03-22 00:08:18 -08003649 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003650 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003651
3652 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003653 * Zone reclaim reclaims unmapped file backed pages and
3654 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003655 *
Christoph Lameter96146342006-07-03 00:24:13 -07003656 * A small portion of unmapped file backed pages is needed for
3657 * file I/O otherwise pages read by file I/O will be immediately
3658 * thrown out if the zone is overallocated. So we do not reclaim
3659 * if less than a specified percentage of the zone is used by
3660 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003661 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003662 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3663 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003664 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003665
Lisa Du6e543d52013-09-11 14:22:36 -07003666 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003667 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003668
Andrew Morton179e9632006-03-22 00:08:18 -08003669 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003670 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003671 */
David Rientjesd773ed62007-10-16 23:26:01 -07003672 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003673 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003674
3675 /*
3676 * Only run zone reclaim on the local zone or on zones that do not
3677 * have associated processors. This will favor the local processor
3678 * over remote processors and spread off node memory allocations
3679 * as wide as possible.
3680 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003681 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003682 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003683 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003684
Johannes Weiner57054652014-10-09 15:28:17 -07003685 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003686 return ZONE_RECLAIM_NOSCAN;
3687
David Rientjesd773ed62007-10-16 23:26:01 -07003688 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003689 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003690
Mel Gorman24cf725182009-06-16 15:33:23 -07003691 if (!ret)
3692 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3693
David Rientjesd773ed62007-10-16 23:26:01 -07003694 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003695}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003696#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003697
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003698/*
3699 * page_evictable - test whether a page is evictable
3700 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003701 *
3702 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003703 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003704 *
3705 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003706 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003707 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003708 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003709 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003710int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003711{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003712 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003713}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003714
Hugh Dickins850465792012-01-20 14:34:19 -08003715#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003716/**
Hugh Dickins24513262012-01-20 14:34:21 -08003717 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3718 * @pages: array of pages to check
3719 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003720 *
Hugh Dickins24513262012-01-20 14:34:21 -08003721 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003722 *
3723 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003724 */
Hugh Dickins24513262012-01-20 14:34:21 -08003725void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003726{
Johannes Weiner925b7672012-01-12 17:18:15 -08003727 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003728 struct zone *zone = NULL;
3729 int pgscanned = 0;
3730 int pgrescued = 0;
3731 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003732
Hugh Dickins24513262012-01-20 14:34:21 -08003733 for (i = 0; i < nr_pages; i++) {
3734 struct page *page = pages[i];
3735 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003736
Hugh Dickins24513262012-01-20 14:34:21 -08003737 pgscanned++;
3738 pagezone = page_zone(page);
3739 if (pagezone != zone) {
3740 if (zone)
3741 spin_unlock_irq(&zone->lru_lock);
3742 zone = pagezone;
3743 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003744 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003745 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003746
Hugh Dickins24513262012-01-20 14:34:21 -08003747 if (!PageLRU(page) || !PageUnevictable(page))
3748 continue;
3749
Hugh Dickins39b5f292012-10-08 16:33:18 -07003750 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003751 enum lru_list lru = page_lru_base_type(page);
3752
Sasha Levin309381fea2014-01-23 15:52:54 -08003753 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003754 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003755 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3756 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003757 pgrescued++;
3758 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003759 }
Hugh Dickins24513262012-01-20 14:34:21 -08003760
3761 if (zone) {
3762 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3763 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3764 spin_unlock_irq(&zone->lru_lock);
3765 }
Hugh Dickins850465792012-01-20 14:34:19 -08003766}
3767#endif /* CONFIG_SHMEM */