blob: 5fec1ba9951fc92a3bff1e638b32e1b4a56bc7e3 [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{
177 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
178}
179
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700180static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800181{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700182 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700183 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800184
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700185 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800186}
187
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000189 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000191int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000193 size_t size = sizeof(*shrinker->nr_deferred);
194
195 /*
196 * If we only have one possible node in the system anyway, save
197 * ourselves the trouble and disable NUMA aware behavior. This way we
198 * will save memory and some small loop time later.
199 */
200 if (nr_node_ids == 1)
201 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
202
203 if (shrinker->flags & SHRINKER_NUMA_AWARE)
204 size *= nr_node_ids;
205
206 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
207 if (!shrinker->nr_deferred)
208 return -ENOMEM;
209
Rusty Russell8e1f9362007-07-17 04:03:17 -0700210 down_write(&shrinker_rwsem);
211 list_add_tail(&shrinker->list, &shrinker_list);
212 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000213 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700215EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
217/*
218 * Remove one
219 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700220void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221{
222 down_write(&shrinker_rwsem);
223 list_del(&shrinker->list);
224 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700225 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700227EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
229#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000230
231static unsigned long
232shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
233 unsigned long nr_pages_scanned, unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000235 unsigned long freed = 0;
236 unsigned long long delta;
237 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700238 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000239 long nr;
240 long new_nr;
241 int nid = shrinkctl->nid;
242 long batch_size = shrinker->batch ? shrinker->batch
243 : SHRINK_BATCH;
244
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700245 freeable = shrinker->count_objects(shrinker, shrinkctl);
246 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000247 return 0;
248
249 /*
250 * copy the current shrinker scan count into a local variable
251 * and zero it so that other concurrent shrinker invocations
252 * don't also do this scanning work.
253 */
254 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
255
256 total_scan = nr;
257 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700258 delta *= freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000259 do_div(delta, lru_pages + 1);
260 total_scan += delta;
261 if (total_scan < 0) {
262 printk(KERN_ERR
263 "shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000264 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700265 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000266 }
267
268 /*
269 * We need to avoid excessive windup on filesystem shrinkers
270 * due to large numbers of GFP_NOFS allocations causing the
271 * shrinkers to return -1 all the time. This results in a large
272 * nr being built up so when a shrink that can do some work
273 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700274 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000275 * memory.
276 *
277 * Hence only allow the shrinker to scan the entire cache when
278 * a large delta change is calculated directly.
279 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700280 if (delta < freeable / 4)
281 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000282
283 /*
284 * Avoid risking looping forever due to too large nr value:
285 * never try to free more than twice the estimate number of
286 * freeable entries.
287 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700288 if (total_scan > freeable * 2)
289 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000290
291 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
292 nr_pages_scanned, lru_pages,
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700293 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000294
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800295 /*
296 * Normally, we should not scan less than batch_size objects in one
297 * pass to avoid too frequent shrinker calls, but if the slab has less
298 * than batch_size objects in total and we are really tight on memory,
299 * we will try to reclaim all available objects, otherwise we can end
300 * up failing allocations although there are plenty of reclaimable
301 * objects spread over several slabs with usage less than the
302 * batch_size.
303 *
304 * We detect the "tight on memory" situations by looking at the total
305 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700306 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800307 * scanning at high prio and therefore should try to reclaim as much as
308 * possible.
309 */
310 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700311 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000312 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800313 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000314
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800315 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000316 ret = shrinker->scan_objects(shrinker, shrinkctl);
317 if (ret == SHRINK_STOP)
318 break;
319 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000320
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800321 count_vm_events(SLABS_SCANNED, nr_to_scan);
322 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000323
324 cond_resched();
325 }
326
327 /*
328 * move the unused scan count back into the shrinker in a
329 * manner that handles concurrent updates. If we exhausted the
330 * scan, there is no need to do an update.
331 */
332 if (total_scan > 0)
333 new_nr = atomic_long_add_return(total_scan,
334 &shrinker->nr_deferred[nid]);
335 else
336 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
337
Dave Hansendf9024a2014-06-04 16:08:07 -0700338 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000339 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340}
341
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342/*
343 * Call the shrink functions to age shrinkable caches
344 *
345 * Here we assume it costs one seek to replace a lru page and that it also
346 * takes a seek to recreate a cache object. With this in mind we age equal
347 * percentages of the lru and ageable caches. This should balance the seeks
348 * generated by these structures.
349 *
Simon Arlott183ff222007-10-20 01:27:18 +0200350 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351 * slab to avoid swapping.
352 *
353 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
354 *
355 * `lru_pages' represents the number of on-LRU pages in all the zones which
356 * are eligible for the caller's allocation attempt. It is used for balancing
357 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700358 *
359 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 */
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000361unsigned long shrink_slab(struct shrink_control *shrinkctl,
Ying Han1495f232011-05-24 17:12:27 -0700362 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700363 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000366 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
Ying Han1495f232011-05-24 17:12:27 -0700368 if (nr_pages_scanned == 0)
369 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370
Minchan Kimf06590b2011-05-24 17:11:11 -0700371 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000372 /*
373 * If we would return 0, our callers would understand that we
374 * have nothing else to shrink and give up trying. By returning
375 * 1 we keep it going and assume we'll be able to shrink next
376 * time.
377 */
378 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700379 goto out;
380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381
382 list_for_each_entry(shrinker, &shrinker_list, list) {
Vladimir Davydovec970972014-01-23 15:53:23 -0800383 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
384 shrinkctl->nid = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000385 freed += shrink_slab_node(shrinkctl, shrinker,
Vladimir Davydovec970972014-01-23 15:53:23 -0800386 nr_pages_scanned, lru_pages);
387 continue;
388 }
389
390 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
391 if (node_online(shrinkctl->nid))
392 freed += shrink_slab_node(shrinkctl, shrinker,
393 nr_pages_scanned, lru_pages);
Dave Chinneracf92b42011-07-08 14:14:35 +1000394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 }
397 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700398out:
399 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000400 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401}
402
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403static inline int is_page_cache_freeable(struct page *page)
404{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700405 /*
406 * A freeable page cache page is referenced only by the caller
407 * that isolated the page, the page cache radix tree and
408 * optional buffer heads at page->private.
409 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700410 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411}
412
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700413static int may_write_to_queue(struct backing_dev_info *bdi,
414 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415{
Christoph Lameter930d9152006-01-08 01:00:47 -0800416 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 return 1;
418 if (!bdi_write_congested(bdi))
419 return 1;
420 if (bdi == current->backing_dev_info)
421 return 1;
422 return 0;
423}
424
425/*
426 * We detected a synchronous write error writing a page out. Probably
427 * -ENOSPC. We need to propagate that into the address_space for a subsequent
428 * fsync(), msync() or close().
429 *
430 * The tricky part is that after writepage we cannot touch the mapping: nothing
431 * prevents it from being freed up. But we have a ref on the page and once
432 * that page is locked, the mapping is pinned.
433 *
434 * We're allowed to run sleeping lock_page() here because we know the caller has
435 * __GFP_FS.
436 */
437static void handle_write_error(struct address_space *mapping,
438 struct page *page, int error)
439{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100440 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700441 if (page_mapping(page) == mapping)
442 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 unlock_page(page);
444}
445
Christoph Lameter04e62a22006-06-23 02:03:38 -0700446/* possible outcome of pageout() */
447typedef enum {
448 /* failed to write page out, page is locked */
449 PAGE_KEEP,
450 /* move page to the active list, page is locked */
451 PAGE_ACTIVATE,
452 /* page has been sent to the disk successfully, page is unlocked */
453 PAGE_SUCCESS,
454 /* page is clean and locked */
455 PAGE_CLEAN,
456} pageout_t;
457
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458/*
Andrew Morton1742f192006-03-22 00:08:21 -0800459 * pageout is called by shrink_page_list() for each dirty page.
460 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700462static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700463 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464{
465 /*
466 * If the page is dirty, only perform writeback if that write
467 * will be non-blocking. To prevent this allocation from being
468 * stalled by pagecache activity. But note that there may be
469 * stalls if we need to run get_block(). We could test
470 * PagePrivate for that.
471 *
Al Viro81742022014-04-03 03:17:43 -0400472 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * this page's queue, we can perform writeback even if that
474 * will block.
475 *
476 * If the page is swapcache, write it back even if that would
477 * block, for some throttling. This happens by accident, because
478 * swap_backing_dev_info is bust: it doesn't reflect the
479 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 */
481 if (!is_page_cache_freeable(page))
482 return PAGE_KEEP;
483 if (!mapping) {
484 /*
485 * Some data journaling orphaned pages can have
486 * page->mapping == NULL while being dirty with clean buffers.
487 */
David Howells266cf652009-04-03 16:42:36 +0100488 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 if (try_to_free_buffers(page)) {
490 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700491 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 return PAGE_CLEAN;
493 }
494 }
495 return PAGE_KEEP;
496 }
497 if (mapping->a_ops->writepage == NULL)
498 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700499 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 return PAGE_KEEP;
501
502 if (clear_page_dirty_for_io(page)) {
503 int res;
504 struct writeback_control wbc = {
505 .sync_mode = WB_SYNC_NONE,
506 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700507 .range_start = 0,
508 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 .for_reclaim = 1,
510 };
511
512 SetPageReclaim(page);
513 res = mapping->a_ops->writepage(page, &wbc);
514 if (res < 0)
515 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800516 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 ClearPageReclaim(page);
518 return PAGE_ACTIVATE;
519 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700520
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 if (!PageWriteback(page)) {
522 /* synchronous write or broken a_ops? */
523 ClearPageReclaim(page);
524 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700525 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700526 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 return PAGE_SUCCESS;
528 }
529
530 return PAGE_CLEAN;
531}
532
Andrew Mortona649fd92006-10-17 00:09:36 -0700533/*
Nick Piggine2867812008-07-25 19:45:30 -0700534 * Same as remove_mapping, but if the page is removed from the mapping, it
535 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700536 */
Johannes Weinera5289102014-04-03 14:47:51 -0700537static int __remove_mapping(struct address_space *mapping, struct page *page,
538 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800539{
Nick Piggin28e4d962006-09-25 23:31:23 -0700540 BUG_ON(!PageLocked(page));
541 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800542
Nick Piggin19fd6232008-07-25 19:45:32 -0700543 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800544 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700545 * The non racy check for a busy page.
546 *
547 * Must be careful with the order of the tests. When someone has
548 * a ref to the page, it may be possible that they dirty it then
549 * drop the reference. So if PageDirty is tested before page_count
550 * here, then the following race may occur:
551 *
552 * get_user_pages(&page);
553 * [user mapping goes away]
554 * write_to(page);
555 * !PageDirty(page) [good]
556 * SetPageDirty(page);
557 * put_page(page);
558 * !page_count(page) [good, discard it]
559 *
560 * [oops, our write_to data is lost]
561 *
562 * Reversing the order of the tests ensures such a situation cannot
563 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
564 * load is not satisfied before that of page->_count.
565 *
566 * Note that if SetPageDirty is always performed via set_page_dirty,
567 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800568 */
Nick Piggine2867812008-07-25 19:45:30 -0700569 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800570 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700571 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
572 if (unlikely(PageDirty(page))) {
573 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800574 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700575 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800576
577 if (PageSwapCache(page)) {
578 swp_entry_t swap = { .val = page_private(page) };
579 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700580 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700581 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700582 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500583 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700584 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500585
586 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700587 /*
588 * Remember a shadow entry for reclaimed file cache in
589 * order to detect refaults, thus thrashing, later on.
590 *
591 * But don't store shadows in an address space that is
592 * already exiting. This is not just an optizimation,
593 * inode reclaim needs to empty out the radix tree or
594 * the nodes are lost. Don't plant shadows behind its
595 * back.
596 */
597 if (reclaimed && page_is_file_cache(page) &&
598 !mapping_exiting(mapping))
599 shadow = workingset_eviction(mapping, page);
600 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700601 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700602 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500603
604 if (freepage != NULL)
605 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800606 }
607
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800608 return 1;
609
610cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700611 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800612 return 0;
613}
614
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615/*
Nick Piggine2867812008-07-25 19:45:30 -0700616 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
617 * someone else has a ref on the page, abort and return 0. If it was
618 * successfully detached, return 1. Assumes the caller has a single ref on
619 * this page.
620 */
621int remove_mapping(struct address_space *mapping, struct page *page)
622{
Johannes Weinera5289102014-04-03 14:47:51 -0700623 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700624 /*
625 * Unfreezing the refcount with 1 rather than 2 effectively
626 * drops the pagecache ref for us without requiring another
627 * atomic operation.
628 */
629 page_unfreeze_refs(page, 1);
630 return 1;
631 }
632 return 0;
633}
634
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700635/**
636 * putback_lru_page - put previously isolated page onto appropriate LRU list
637 * @page: page to be put back to appropriate lru list
638 *
639 * Add previously isolated @page to appropriate LRU list.
640 * Page may still be unevictable for other reasons.
641 *
642 * lru_lock must not be held, interrupts must be enabled.
643 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700644void putback_lru_page(struct page *page)
645{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700646 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700647 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700648
Sasha Levin309381fea2014-01-23 15:52:54 -0800649 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700650
651redo:
652 ClearPageUnevictable(page);
653
Hugh Dickins39b5f292012-10-08 16:33:18 -0700654 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700655 /*
656 * For evictable pages, we can use the cache.
657 * In event of a race, worst case is we end up with an
658 * unevictable page on [in]active list.
659 * We know how to handle that.
660 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700661 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700662 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700663 } else {
664 /*
665 * Put unevictable pages directly on zone's unevictable
666 * list.
667 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700668 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700669 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700670 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700671 * When racing with an mlock or AS_UNEVICTABLE clearing
672 * (page is unlocked) make sure that if the other thread
673 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800674 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700675 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700676 * the page back to the evictable list.
677 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700678 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700679 */
680 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700681 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700682
683 /*
684 * page's status can change while we move it among lru. If an evictable
685 * page is on unevictable list, it never be freed. To avoid that,
686 * check after we added it to the list, again.
687 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700688 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700689 if (!isolate_lru_page(page)) {
690 put_page(page);
691 goto redo;
692 }
693 /* This means someone else dropped this page from LRU
694 * So, it will be freed or putback to LRU again. There is
695 * nothing to do here.
696 */
697 }
698
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700699 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700700 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700701 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700702 count_vm_event(UNEVICTABLE_PGCULLED);
703
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700704 put_page(page); /* drop ref from isolate */
705}
706
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800707enum page_references {
708 PAGEREF_RECLAIM,
709 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800710 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800711 PAGEREF_ACTIVATE,
712};
713
714static enum page_references page_check_references(struct page *page,
715 struct scan_control *sc)
716{
Johannes Weiner645747462010-03-05 13:42:22 -0800717 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800718 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800719
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700720 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
721 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800722 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800723
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724 /*
725 * Mlock lost the isolation race with us. Let try_to_unmap()
726 * move the page to the unevictable list.
727 */
728 if (vm_flags & VM_LOCKED)
729 return PAGEREF_RECLAIM;
730
Johannes Weiner645747462010-03-05 13:42:22 -0800731 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700732 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800733 return PAGEREF_ACTIVATE;
734 /*
735 * All mapped pages start out with page table
736 * references from the instantiating fault, so we need
737 * to look twice if a mapped file page is used more
738 * than once.
739 *
740 * Mark it and spare it for another trip around the
741 * inactive list. Another page table reference will
742 * lead to its activation.
743 *
744 * Note: the mark is set for activated pages as well
745 * so that recently deactivated but used pages are
746 * quickly recovered.
747 */
748 SetPageReferenced(page);
749
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800750 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800751 return PAGEREF_ACTIVATE;
752
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800753 /*
754 * Activate file-backed executable pages after first usage.
755 */
756 if (vm_flags & VM_EXEC)
757 return PAGEREF_ACTIVATE;
758
Johannes Weiner645747462010-03-05 13:42:22 -0800759 return PAGEREF_KEEP;
760 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800761
762 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700763 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800764 return PAGEREF_RECLAIM_CLEAN;
765
766 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800767}
768
Mel Gormane2be15f2013-07-03 15:01:57 -0700769/* Check if a page is dirty or under writeback */
770static void page_check_dirty_writeback(struct page *page,
771 bool *dirty, bool *writeback)
772{
Mel Gormanb4597222013-07-03 15:02:05 -0700773 struct address_space *mapping;
774
Mel Gormane2be15f2013-07-03 15:01:57 -0700775 /*
776 * Anonymous pages are not handled by flushers and must be written
777 * from reclaim context. Do not stall reclaim based on them
778 */
779 if (!page_is_file_cache(page)) {
780 *dirty = false;
781 *writeback = false;
782 return;
783 }
784
785 /* By default assume that the page flags are accurate */
786 *dirty = PageDirty(page);
787 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700788
789 /* Verify dirty/writeback state if the filesystem supports it */
790 if (!page_has_private(page))
791 return;
792
793 mapping = page_mapping(page);
794 if (mapping && mapping->a_ops->is_dirty_writeback)
795 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700796}
797
Nick Piggine2867812008-07-25 19:45:30 -0700798/*
Andrew Morton1742f192006-03-22 00:08:21 -0800799 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 */
Andrew Morton1742f192006-03-22 00:08:21 -0800801static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700802 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700803 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700804 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700805 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700806 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700807 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700808 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700809 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700810 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811{
812 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700813 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700815 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700816 unsigned long nr_dirty = 0;
817 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800818 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700819 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700820 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821
822 cond_resched();
823
Tim Chen69980e32012-07-31 16:46:08 -0700824 mem_cgroup_uncharge_start();
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);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700878 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
879 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700880 nr_congested++;
881
882 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700883 * If a page at the tail of the LRU is under writeback, there
884 * are three cases to consider.
885 *
886 * 1) If reclaim is encountering an excessive number of pages
887 * under writeback and this page is both under writeback and
888 * PageReclaim then it indicates that pages are being queued
889 * for IO but are being recycled through the LRU before the
890 * IO can complete. Waiting on the page itself risks an
891 * indefinite stall if it is impossible to writeback the
892 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700893 * note that the LRU is being scanned too quickly and the
894 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700895 *
896 * 2) Global reclaim encounters a page, memcg encounters a
897 * page that is not marked for immediate reclaim or
898 * the caller does not have __GFP_IO. In this case mark
899 * the page for immediate reclaim and continue scanning.
900 *
901 * __GFP_IO is checked because a loop driver thread might
902 * enter reclaim, and deadlock if it waits on a page for
903 * which it is needed to do the write (loop masks off
904 * __GFP_IO|__GFP_FS for this reason); but more thought
905 * would probably show more reasons.
906 *
907 * Don't require __GFP_FS, since we're not going into the
908 * FS, just waiting on its writeback completion. Worryingly,
909 * ext4 gfs2 and xfs allocate pages with
910 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
911 * may_enter_fs here is liable to OOM on them.
912 *
913 * 3) memcg encounters a page that is not already marked
914 * PageReclaim. memcg does not have any dirty pages
915 * throttling so we could easily OOM just because too many
916 * pages are in writeback and there is nothing else to
917 * reclaim. Wait for the writeback to complete.
918 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700919 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700920 /* Case 1 above */
921 if (current_is_kswapd() &&
922 PageReclaim(page) &&
923 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700924 nr_immediate++;
925 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700926
927 /* Case 2 above */
928 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700929 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
930 /*
931 * This is slightly racy - end_page_writeback()
932 * might have just cleared PageReclaim, then
933 * setting PageReclaim here end up interpreted
934 * as PageReadahead - but that does not matter
935 * enough to care. What we do want is for this
936 * page to have PageReclaim set next time memcg
937 * reclaim reaches the tests above, so it will
938 * then wait_on_page_writeback() to avoid OOM;
939 * and it's also appropriate in global reclaim.
940 */
941 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700942 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700943
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700944 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700945
946 /* Case 3 above */
947 } else {
948 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700949 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700950 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
Minchan Kim02c6de82012-10-08 16:31:55 -0700952 if (!force_reclaim)
953 references = page_check_references(page, sc);
954
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800955 switch (references) {
956 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800958 case PAGEREF_KEEP:
959 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800960 case PAGEREF_RECLAIM:
961 case PAGEREF_RECLAIM_CLEAN:
962 ; /* try to reclaim the page below */
963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 /*
966 * Anonymous process memory has backing store?
967 * Try to allocate it some swap space here.
968 */
Nick Pigginb291f002008-10-18 20:26:44 -0700969 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800970 if (!(sc->gfp_mask & __GFP_IO))
971 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700972 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800974 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
Mel Gormane2be15f2013-07-03 15:01:57 -0700976 /* Adding to swap updated mapping */
977 mapping = page_mapping(page);
978 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
980 /*
981 * The page is mapped into the page tables of one or more
982 * processes. Try to unmap it here.
983 */
984 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700985 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 case SWAP_FAIL:
987 goto activate_locked;
988 case SWAP_AGAIN:
989 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700990 case SWAP_MLOCK:
991 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 case SWAP_SUCCESS:
993 ; /* try to free the page below */
994 }
995 }
996
997 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700998 /*
999 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001000 * avoid risk of stack overflow but only writeback
1001 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001002 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001003 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001004 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -07001005 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001006 /*
1007 * Immediately reclaim when written back.
1008 * Similar in principal to deactivate_page()
1009 * except we already have the page isolated
1010 * and know it's dirty
1011 */
1012 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1013 SetPageReclaim(page);
1014
Mel Gormanee728862011-10-31 17:07:38 -07001015 goto keep_locked;
1016 }
1017
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001018 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001020 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001022 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 goto keep_locked;
1024
1025 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001026 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 case PAGE_KEEP:
1028 goto keep_locked;
1029 case PAGE_ACTIVATE:
1030 goto activate_locked;
1031 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001032 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001033 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001034 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001036
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 /*
1038 * A synchronous write - probably a ramdisk. Go
1039 * ahead and try to reclaim the page.
1040 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001041 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 goto keep;
1043 if (PageDirty(page) || PageWriteback(page))
1044 goto keep_locked;
1045 mapping = page_mapping(page);
1046 case PAGE_CLEAN:
1047 ; /* try to free the page below */
1048 }
1049 }
1050
1051 /*
1052 * If the page has buffers, try to free the buffer mappings
1053 * associated with this page. If we succeed we try to free
1054 * the page as well.
1055 *
1056 * We do this even if the page is PageDirty().
1057 * try_to_release_page() does not perform I/O, but it is
1058 * possible for a page to have PageDirty set, but it is actually
1059 * clean (all its buffers are clean). This happens if the
1060 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001061 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 * try_to_release_page() will discover that cleanness and will
1063 * drop the buffers and mark the page clean - it can be freed.
1064 *
1065 * Rarely, pages can have buffers and no ->mapping. These are
1066 * the pages which were not successfully invalidated in
1067 * truncate_complete_page(). We try to drop those buffers here
1068 * and if that worked, and the page is no longer mapped into
1069 * process address space (page_count == 1) it can be freed.
1070 * Otherwise, leave the page on the LRU so it is swappable.
1071 */
David Howells266cf652009-04-03 16:42:36 +01001072 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 if (!try_to_release_page(page, sc->gfp_mask))
1074 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001075 if (!mapping && page_count(page) == 1) {
1076 unlock_page(page);
1077 if (put_page_testzero(page))
1078 goto free_it;
1079 else {
1080 /*
1081 * rare race with speculative reference.
1082 * the speculative reference will free
1083 * this page shortly, so we may
1084 * increment nr_reclaimed here (and
1085 * leave it off the LRU).
1086 */
1087 nr_reclaimed++;
1088 continue;
1089 }
1090 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 }
1092
Johannes Weinera5289102014-04-03 14:47:51 -07001093 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001094 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095
Nick Piggina978d6f2008-10-18 20:26:58 -07001096 /*
1097 * At this point, we have no other references and there is
1098 * no way to pick any more up (removed from LRU, removed
1099 * from pagecache). Can use non-atomic bitops now (and
1100 * we obviously don't have to worry about waking up a process
1101 * waiting on the page lock, because there are no references.
1102 */
1103 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001104free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001105 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001106
1107 /*
1108 * Is there need to periodically free_page_list? It would
1109 * appear not as the counts should be low
1110 */
1111 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 continue;
1113
Nick Pigginb291f002008-10-18 20:26:44 -07001114cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001115 if (PageSwapCache(page))
1116 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001117 unlock_page(page);
1118 putback_lru_page(page);
1119 continue;
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001122 /* Not a candidate for swapping, so reclaim swap space. */
1123 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001124 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001125 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 SetPageActive(page);
1127 pgactivate++;
1128keep_locked:
1129 unlock_page(page);
1130keep:
1131 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001132 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001134
Mel Gormanb745bc82014-06-04 16:10:22 -07001135 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001136
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001138 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001139 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001140 *ret_nr_dirty += nr_dirty;
1141 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001142 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001143 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001144 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001145 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146}
1147
Minchan Kim02c6de82012-10-08 16:31:55 -07001148unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1149 struct list_head *page_list)
1150{
1151 struct scan_control sc = {
1152 .gfp_mask = GFP_KERNEL,
1153 .priority = DEF_PRIORITY,
1154 .may_unmap = 1,
1155 };
Mel Gorman8e950282013-07-03 15:02:02 -07001156 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001157 struct page *page, *next;
1158 LIST_HEAD(clean_pages);
1159
1160 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001161 if (page_is_file_cache(page) && !PageDirty(page) &&
1162 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001163 ClearPageActive(page);
1164 list_move(&page->lru, &clean_pages);
1165 }
1166 }
1167
1168 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001169 TTU_UNMAP|TTU_IGNORE_ACCESS,
1170 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001171 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001172 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001173 return ret;
1174}
1175
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001176/*
1177 * Attempt to remove the specified page from its LRU. Only take this page
1178 * if it is of the appropriate PageActive status. Pages which are being
1179 * freed elsewhere are also ignored.
1180 *
1181 * page: page to consider
1182 * mode: one of the LRU isolation modes defined above
1183 *
1184 * returns 0 on success, -ve errno on failure.
1185 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001186int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001187{
1188 int ret = -EINVAL;
1189
1190 /* Only take pages on the LRU. */
1191 if (!PageLRU(page))
1192 return ret;
1193
Minchan Kime46a2872012-10-08 16:33:48 -07001194 /* Compaction should not handle unevictable pages but CMA can do so */
1195 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001196 return ret;
1197
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001198 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001199
Mel Gormanc8244932012-01-12 17:19:38 -08001200 /*
1201 * To minimise LRU disruption, the caller can indicate that it only
1202 * wants to isolate pages it will be able to operate on without
1203 * blocking - clean pages for the most part.
1204 *
1205 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1206 * is used by reclaim when it is cannot write to backing storage
1207 *
1208 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1209 * that it is possible to migrate without blocking
1210 */
1211 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1212 /* All the caller can do on PageWriteback is block */
1213 if (PageWriteback(page))
1214 return ret;
1215
1216 if (PageDirty(page)) {
1217 struct address_space *mapping;
1218
1219 /* ISOLATE_CLEAN means only clean pages */
1220 if (mode & ISOLATE_CLEAN)
1221 return ret;
1222
1223 /*
1224 * Only pages without mappings or that have a
1225 * ->migratepage callback are possible to migrate
1226 * without blocking
1227 */
1228 mapping = page_mapping(page);
1229 if (mapping && !mapping->a_ops->migratepage)
1230 return ret;
1231 }
1232 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001233
Minchan Kimf80c0672011-10-31 17:06:55 -07001234 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1235 return ret;
1236
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001237 if (likely(get_page_unless_zero(page))) {
1238 /*
1239 * Be careful not to clear PageLRU until after we're
1240 * sure the page is not being freed elsewhere -- the
1241 * page release code relies on it.
1242 */
1243 ClearPageLRU(page);
1244 ret = 0;
1245 }
1246
1247 return ret;
1248}
1249
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001250/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 * zone->lru_lock is heavily contended. Some of the functions that
1252 * shrink the lists perform better by taking out a batch of pages
1253 * and working on them outside the LRU lock.
1254 *
1255 * For pagecache intensive workloads, this function is the hottest
1256 * spot in the kernel (apart from copy_*_user functions).
1257 *
1258 * Appropriate locks must be held before calling this function.
1259 *
1260 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001261 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001263 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001264 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001265 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001266 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 *
1268 * returns how many pages were moved onto *@dst.
1269 */
Andrew Morton69e05942006-03-22 00:08:19 -08001270static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001271 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001272 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001273 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001275 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001276 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001277 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001279 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001280 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001281 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001282
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 page = lru_to_page(src);
1284 prefetchw_prev_lru_page(page, src, flags);
1285
Sasha Levin309381fea2014-01-23 15:52:54 -08001286 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001287
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001288 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001289 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001290 nr_pages = hpage_nr_pages(page);
1291 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001292 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001293 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001294 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001295
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001296 case -EBUSY:
1297 /* else it is being freed elsewhere */
1298 list_move(&page->lru, src);
1299 continue;
1300
1301 default:
1302 BUG();
1303 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 }
1305
Hugh Dickinsf6260122012-01-12 17:20:06 -08001306 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001307 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1308 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 return nr_taken;
1310}
1311
Nick Piggin62695a82008-10-18 20:26:09 -07001312/**
1313 * isolate_lru_page - tries to isolate a page from its LRU list
1314 * @page: page to isolate from its LRU list
1315 *
1316 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1317 * vmstat statistic corresponding to whatever LRU list the page was on.
1318 *
1319 * Returns 0 if the page was removed from an LRU list.
1320 * Returns -EBUSY if the page was not on an LRU list.
1321 *
1322 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001323 * the active list, it will have PageActive set. If it was found on
1324 * the unevictable list, it will have the PageUnevictable bit set. That flag
1325 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001326 *
1327 * The vmstat statistic corresponding to the list on which the page was
1328 * found will be decremented.
1329 *
1330 * Restrictions:
1331 * (1) Must be called with an elevated refcount on the page. This is a
1332 * fundamentnal difference from isolate_lru_pages (which is called
1333 * without a stable reference).
1334 * (2) the lru_lock must not be held.
1335 * (3) interrupts must be enabled.
1336 */
1337int isolate_lru_page(struct page *page)
1338{
1339 int ret = -EBUSY;
1340
Sasha Levin309381fea2014-01-23 15:52:54 -08001341 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001342
Nick Piggin62695a82008-10-18 20:26:09 -07001343 if (PageLRU(page)) {
1344 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001345 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001346
1347 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001348 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001349 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001350 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001351 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001352 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001353 del_page_from_lru_list(page, lruvec, lru);
1354 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001355 }
1356 spin_unlock_irq(&zone->lru_lock);
1357 }
1358 return ret;
1359}
1360
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001361/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001362 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1363 * then get resheduled. When there are massive number of tasks doing page
1364 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1365 * the LRU list will go small and be scanned faster than necessary, leading to
1366 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001367 */
1368static int too_many_isolated(struct zone *zone, int file,
1369 struct scan_control *sc)
1370{
1371 unsigned long inactive, isolated;
1372
1373 if (current_is_kswapd())
1374 return 0;
1375
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001376 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001377 return 0;
1378
1379 if (file) {
1380 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1381 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1382 } else {
1383 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1384 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1385 }
1386
Fengguang Wu3cf23842012-12-18 14:23:31 -08001387 /*
1388 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1389 * won't get blocked by normal direct-reclaimers, forming a circular
1390 * deadlock.
1391 */
1392 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1393 inactive >>= 3;
1394
Rik van Riel35cd7812009-09-21 17:01:38 -07001395 return isolated > inactive;
1396}
1397
Mel Gorman66635622010-08-09 17:19:30 -07001398static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001399putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001400{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001401 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1402 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001403 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001404
Mel Gorman66635622010-08-09 17:19:30 -07001405 /*
1406 * Put back any unfreeable pages.
1407 */
Mel Gorman66635622010-08-09 17:19:30 -07001408 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001409 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001410 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001411
Sasha Levin309381fea2014-01-23 15:52:54 -08001412 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001413 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001414 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001415 spin_unlock_irq(&zone->lru_lock);
1416 putback_lru_page(page);
1417 spin_lock_irq(&zone->lru_lock);
1418 continue;
1419 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001420
1421 lruvec = mem_cgroup_page_lruvec(page, zone);
1422
Linus Torvalds7a608572011-01-17 14:42:19 -08001423 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001424 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001425 add_page_to_lru_list(page, lruvec, lru);
1426
Mel Gorman66635622010-08-09 17:19:30 -07001427 if (is_active_lru(lru)) {
1428 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001429 int numpages = hpage_nr_pages(page);
1430 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001431 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001432 if (put_page_testzero(page)) {
1433 __ClearPageLRU(page);
1434 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001435 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001436
1437 if (unlikely(PageCompound(page))) {
1438 spin_unlock_irq(&zone->lru_lock);
1439 (*get_compound_page_dtor(page))(page);
1440 spin_lock_irq(&zone->lru_lock);
1441 } else
1442 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001443 }
1444 }
Mel Gorman66635622010-08-09 17:19:30 -07001445
Hugh Dickins3f797682012-01-12 17:20:07 -08001446 /*
1447 * To save our caller's stack, now use input list for pages to free.
1448 */
1449 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001450}
1451
1452/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001453 * If a kernel thread (such as nfsd for loop-back mounts) services
1454 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1455 * In that case we should only throttle if the backing device it is
1456 * writing to is congested. In other cases it is safe to throttle.
1457 */
1458static int current_may_throttle(void)
1459{
1460 return !(current->flags & PF_LESS_THROTTLE) ||
1461 current->backing_dev_info == NULL ||
1462 bdi_write_congested(current->backing_dev_info);
1463}
1464
1465/*
Andrew Morton1742f192006-03-22 00:08:21 -08001466 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1467 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 */
Mel Gorman66635622010-08-09 17:19:30 -07001469static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001470shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001471 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
1473 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001474 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001475 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001476 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001477 unsigned long nr_dirty = 0;
1478 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001479 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001480 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001481 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001482 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001483 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001484 struct zone *zone = lruvec_zone(lruvec);
1485 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001486
Rik van Riel35cd7812009-09-21 17:01:38 -07001487 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001488 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001489
1490 /* We are about to die and free our memory. Return now. */
1491 if (fatal_signal_pending(current))
1492 return SWAP_CLUSTER_MAX;
1493 }
1494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001496
1497 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001498 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001499 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001500 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001501
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001504 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1505 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001506
1507 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1508 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1509
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001510 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001511 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001512 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001513 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001514 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001515 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001516 }
Hillf Dantond563c052012-03-21 16:34:02 -07001517 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001518
Hillf Dantond563c052012-03-21 16:34:02 -07001519 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001520 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001521
Minchan Kim02c6de82012-10-08 16:31:55 -07001522 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001523 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1524 &nr_writeback, &nr_immediate,
1525 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001526
Hugh Dickins3f797682012-01-12 17:20:07 -08001527 spin_lock_irq(&zone->lru_lock);
1528
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001529 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001530
Ying Han904249a2012-04-25 16:01:48 -07001531 if (global_reclaim(sc)) {
1532 if (current_is_kswapd())
1533 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1534 nr_reclaimed);
1535 else
1536 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1537 nr_reclaimed);
1538 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001539
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001540 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001541
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001542 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001543
1544 spin_unlock_irq(&zone->lru_lock);
1545
Mel Gormanb745bc82014-06-04 16:10:22 -07001546 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001547
Mel Gorman92df3a72011-10-31 17:07:56 -07001548 /*
1549 * If reclaim is isolating dirty pages under writeback, it implies
1550 * that the long-lived page allocation rate is exceeding the page
1551 * laundering rate. Either the global limits are not being effective
1552 * at throttling processes due to the page distribution throughout
1553 * zones or there is heavy usage of a slow backing device. The
1554 * only option is to throttle from reclaim context which is not ideal
1555 * as there is no guarantee the dirtying process is throttled in the
1556 * same way balance_dirty_pages() manages.
1557 *
Mel Gorman8e950282013-07-03 15:02:02 -07001558 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1559 * of pages under pages flagged for immediate reclaim and stall if any
1560 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001561 */
Mel Gorman918fc712013-07-08 16:00:25 -07001562 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001563 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001564
Mel Gormand43006d2013-07-03 15:01:50 -07001565 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001566 * memcg will stall in page writeback so only consider forcibly
1567 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001568 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001569 if (global_reclaim(sc)) {
1570 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001571 * Tag a zone as congested if all the dirty pages scanned were
1572 * backed by a congested BDI and wait_iff_congested will stall.
1573 */
1574 if (nr_dirty && nr_dirty == nr_congested)
1575 zone_set_flag(zone, ZONE_CONGESTED);
1576
1577 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001578 * If dirty pages are scanned that are not queued for IO, it
1579 * implies that flushers are not keeping up. In this case, flag
1580 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
Linus Torvaldsb738d762014-06-08 14:17:00 -07001581 * pages from reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001582 */
1583 if (nr_unqueued_dirty == nr_taken)
1584 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1585
1586 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001587 * If kswapd scans pages marked marked for immediate
1588 * reclaim and under writeback (nr_immediate), it implies
1589 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001590 * they are written so also forcibly stall.
1591 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001592 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001593 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001594 }
Mel Gormand43006d2013-07-03 15:01:50 -07001595
Mel Gorman8e950282013-07-03 15:02:02 -07001596 /*
1597 * Stall direct reclaim for IO completions if underlying BDIs or zone
1598 * is congested. Allow kswapd to continue until it starts encountering
1599 * unqueued dirty pages or cycling through the LRU too quickly.
1600 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001601 if (!sc->hibernation_mode && !current_is_kswapd() &&
1602 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001603 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1604
Mel Gormane11da5b2010-10-26 14:21:40 -07001605 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1606 zone_idx(zone),
1607 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001608 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001609 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001610 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611}
1612
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001613/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 * This moves pages from the active list to the inactive list.
1615 *
1616 * We move them the other way if the page is referenced by one or more
1617 * processes, from rmap.
1618 *
1619 * If the pages are mostly unmapped, the processing is fast and it is
1620 * appropriate to hold zone->lru_lock across the whole operation. But if
1621 * the pages are mapped, the processing is slow (page_referenced()) so we
1622 * should drop zone->lru_lock around each page. It's impossible to balance
1623 * this, so instead we remove the pages from the LRU while processing them.
1624 * It is safe to rely on PG_active against the non-LRU pages in here because
1625 * nobody will play with that bit on a non-LRU page.
1626 *
1627 * The downside is that we have to touch page->_count against each page.
1628 * But we had to alter page->flags anyway.
1629 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001630
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001631static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001632 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001633 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001634 enum lru_list lru)
1635{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001636 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001637 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001638 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001639 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001640
Wu Fengguang3eb41402009-06-16 15:33:13 -07001641 while (!list_empty(list)) {
1642 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001643 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001644
Sasha Levin309381fea2014-01-23 15:52:54 -08001645 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001646 SetPageLRU(page);
1647
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001648 nr_pages = hpage_nr_pages(page);
1649 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001650 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001651 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001652
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001653 if (put_page_testzero(page)) {
1654 __ClearPageLRU(page);
1655 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001656 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001657
1658 if (unlikely(PageCompound(page))) {
1659 spin_unlock_irq(&zone->lru_lock);
1660 (*get_compound_page_dtor(page))(page);
1661 spin_lock_irq(&zone->lru_lock);
1662 } else
1663 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001664 }
1665 }
1666 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1667 if (!is_active_lru(lru))
1668 __count_vm_events(PGDEACTIVATE, pgmoved);
1669}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001670
Hugh Dickinsf6260122012-01-12 17:20:06 -08001671static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001672 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001673 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001674 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001676 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001677 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001678 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001680 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001681 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001683 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001684 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001685 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001686 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001687 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688
1689 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001690
1691 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001692 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001693 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001694 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001695
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001697
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001698 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1699 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001700 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001701 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001702
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001703 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001704
Hugh Dickinsf6260122012-01-12 17:20:06 -08001705 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001706 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001707 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 spin_unlock_irq(&zone->lru_lock);
1709
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 while (!list_empty(&l_hold)) {
1711 cond_resched();
1712 page = lru_to_page(&l_hold);
1713 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001714
Hugh Dickins39b5f292012-10-08 16:33:18 -07001715 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001716 putback_lru_page(page);
1717 continue;
1718 }
1719
Mel Gormancc715d92012-03-21 16:34:00 -07001720 if (unlikely(buffer_heads_over_limit)) {
1721 if (page_has_private(page) && trylock_page(page)) {
1722 if (page_has_private(page))
1723 try_to_release_page(page, 0);
1724 unlock_page(page);
1725 }
1726 }
1727
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001728 if (page_referenced(page, 0, sc->target_mem_cgroup,
1729 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001730 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001731 /*
1732 * Identify referenced, file-backed active pages and
1733 * give them one more trip around the active list. So
1734 * that executable code get better chances to stay in
1735 * memory under moderate memory pressure. Anon pages
1736 * are not likely to be evicted by use-once streaming
1737 * IO, plus JVM can create lots of anon VM_EXEC pages,
1738 * so we ignore them here.
1739 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001740 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001741 list_add(&page->lru, &l_active);
1742 continue;
1743 }
1744 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001745
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001746 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 list_add(&page->lru, &l_inactive);
1748 }
1749
Andrew Mortonb5557492009-01-06 14:40:13 -08001750 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001751 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001752 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001753 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001754 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001755 * Count referenced pages from currently used mappings as rotated,
1756 * even though only some of them are actually re-activated. This
1757 * helps balance scan pressure between file and anonymous pages in
1758 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001759 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001760 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001761
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001762 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1763 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001764 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001765 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001766
Mel Gormanb745bc82014-06-04 16:10:22 -07001767 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768}
1769
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001770#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001771static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001772{
1773 unsigned long active, inactive;
1774
1775 active = zone_page_state(zone, NR_ACTIVE_ANON);
1776 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1777
1778 if (inactive * zone->inactive_ratio < active)
1779 return 1;
1780
1781 return 0;
1782}
1783
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001784/**
1785 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001786 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001787 *
1788 * Returns true if the zone does not have enough inactive anon pages,
1789 * meaning some active anon pages need to be deactivated.
1790 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001791static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001792{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001793 /*
1794 * If we don't have swap space, anonymous page deactivation
1795 * is pointless.
1796 */
1797 if (!total_swap_pages)
1798 return 0;
1799
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001800 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001801 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001802
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001803 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001804}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001805#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001806static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001807{
1808 return 0;
1809}
1810#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001811
Rik van Riel56e49d22009-06-16 15:32:28 -07001812/**
1813 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001814 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001815 *
1816 * When the system is doing streaming IO, memory pressure here
1817 * ensures that active file pages get deactivated, until more
1818 * than half of the file pages are on the inactive list.
1819 *
1820 * Once we get to that situation, protect the system's working
1821 * set from being evicted by disabling active file page aging.
1822 *
1823 * This uses a different ratio than the anonymous pages, because
1824 * the page cache uses a use-once replacement algorithm.
1825 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001826static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001827{
Johannes Weinere3790142013-02-22 16:35:19 -08001828 unsigned long inactive;
1829 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001830
Johannes Weinere3790142013-02-22 16:35:19 -08001831 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1832 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1833
1834 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001835}
1836
Hugh Dickins75b00af2012-05-29 15:07:09 -07001837static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001838{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001839 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001840 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001841 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001842 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001843}
1844
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001845static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001846 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001847{
Rik van Rielb39415b2009-12-14 17:59:48 -08001848 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001849 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001850 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001851 return 0;
1852 }
1853
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001854 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001855}
1856
Johannes Weiner9a265112013-02-22 16:32:17 -08001857enum scan_balance {
1858 SCAN_EQUAL,
1859 SCAN_FRACT,
1860 SCAN_ANON,
1861 SCAN_FILE,
1862};
1863
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001865 * Determine how aggressively the anon and file LRU lists should be
1866 * scanned. The relative value of each set of LRU lists is determined
1867 * by looking at the fraction of the pages scanned we did rotate back
1868 * onto the active list instead of evict.
1869 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001870 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1871 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001872 */
Johannes Weiner02695172014-08-06 16:06:17 -07001873static void get_scan_count(struct lruvec *lruvec, int swappiness,
1874 struct scan_control *sc, unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001875{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001876 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001877 u64 fraction[2];
1878 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001879 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001880 unsigned long anon_prio, file_prio;
1881 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001882 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001883 bool force_scan = false;
1884 unsigned long ap, fp;
1885 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001886 bool some_scanned;
1887 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001888
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001889 /*
1890 * If the zone or memcg is small, nr[l] can be 0. This
1891 * results in no scanning on this priority and a potential
1892 * priority drop. Global direct reclaim can go to the next
1893 * zone and tends to have no problems. Global kswapd is for
1894 * zone balancing and it needs to scan a minimum amount. When
1895 * reclaiming for a memcg, a priority drop can cause high
1896 * latencies, so it's better to scan a minimum amount there as
1897 * well.
1898 */
Lisa Du6e543d52013-09-11 14:22:36 -07001899 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001900 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001901 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001902 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001903
1904 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001905 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001906 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001907 goto out;
1908 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001909
Johannes Weiner10316b32013-02-22 16:32:14 -08001910 /*
1911 * Global reclaim will swap to prevent OOM even with no
1912 * swappiness, but memcg users want to use this knob to
1913 * disable swapping for individual groups completely when
1914 * using the memory controller's swap limit feature would be
1915 * too expensive.
1916 */
Johannes Weiner02695172014-08-06 16:06:17 -07001917 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001918 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001919 goto out;
1920 }
1921
1922 /*
1923 * Do not apply any pressure balancing cleverness when the
1924 * system is close to OOM, scan both anon and file equally
1925 * (unless the swappiness setting disagrees with swapping).
1926 */
Johannes Weiner02695172014-08-06 16:06:17 -07001927 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001928 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001929 goto out;
1930 }
1931
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001932 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1933 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1934 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1935 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001936
Johannes Weiner11d16c22013-02-22 16:32:15 -08001937 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001938 * Prevent the reclaimer from falling into the cache trap: as
1939 * cache pages start out inactive, every cache fault will tip
1940 * the scan balance towards the file LRU. And as the file LRU
1941 * shrinks, so does the window for rotation from references.
1942 * This means we have a runaway feedback loop where a tiny
1943 * thrashing file LRU becomes infinitely more attractive than
1944 * anon pages. Try to detect this based on file LRU size.
1945 */
1946 if (global_reclaim(sc)) {
1947 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1948
1949 if (unlikely(file + free <= high_wmark_pages(zone))) {
1950 scan_balance = SCAN_ANON;
1951 goto out;
1952 }
1953 }
1954
1955 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001956 * There is enough inactive page cache, do not reclaim
1957 * anything from the anonymous working set right now.
1958 */
1959 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001960 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001961 goto out;
1962 }
1963
Johannes Weiner9a265112013-02-22 16:32:17 -08001964 scan_balance = SCAN_FRACT;
1965
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001966 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001967 * With swappiness at 100, anonymous and file have the same priority.
1968 * This scanning priority is essentially the inverse of IO cost.
1969 */
Johannes Weiner02695172014-08-06 16:06:17 -07001970 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001971 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001972
1973 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001974 * OK, so we have swap space and a fair amount of page cache
1975 * pages. We use the recently rotated / recently scanned
1976 * ratios to determine how valuable each cache is.
1977 *
1978 * Because workloads change over time (and to avoid overflow)
1979 * we keep these statistics as a floating average, which ends
1980 * up weighing recent references more than old ones.
1981 *
1982 * anon in [0], file in [1]
1983 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001984 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001985 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001986 reclaim_stat->recent_scanned[0] /= 2;
1987 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001988 }
1989
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001990 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001991 reclaim_stat->recent_scanned[1] /= 2;
1992 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001993 }
1994
1995 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001996 * The amount of pressure on anon vs file pages is inversely
1997 * proportional to the fraction of recently scanned pages on
1998 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001999 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002000 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002001 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002002
Satoru Moriyafe350042012-05-29 15:06:47 -07002003 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002004 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002005 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002006
Shaohua Li76a33fc2010-05-24 14:32:36 -07002007 fraction[0] = ap;
2008 fraction[1] = fp;
2009 denominator = ap + fp + 1;
2010out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002011 some_scanned = false;
2012 /* Only use force_scan on second pass. */
2013 for (pass = 0; !some_scanned && pass < 2; pass++) {
2014 for_each_evictable_lru(lru) {
2015 int file = is_file_lru(lru);
2016 unsigned long size;
2017 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002018
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002019 size = get_lru_size(lruvec, lru);
2020 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002021
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002022 if (!scan && pass && force_scan)
2023 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002024
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002025 switch (scan_balance) {
2026 case SCAN_EQUAL:
2027 /* Scan lists relative to size */
2028 break;
2029 case SCAN_FRACT:
2030 /*
2031 * Scan types proportional to swappiness and
2032 * their relative recent reclaim efficiency.
2033 */
2034 scan = div64_u64(scan * fraction[file],
2035 denominator);
2036 break;
2037 case SCAN_FILE:
2038 case SCAN_ANON:
2039 /* Scan one type exclusively */
2040 if ((scan_balance == SCAN_FILE) != file)
2041 scan = 0;
2042 break;
2043 default:
2044 /* Look ma, no brain */
2045 BUG();
2046 }
2047 nr[lru] = scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002048 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002049 * Skip the second pass and don't force_scan,
2050 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002051 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002052 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002053 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002054 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002055}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002056
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002057/*
2058 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2059 */
Johannes Weiner02695172014-08-06 16:06:17 -07002060static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2061 struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002062{
2063 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002064 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002065 unsigned long nr_to_scan;
2066 enum lru_list lru;
2067 unsigned long nr_reclaimed = 0;
2068 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2069 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002070 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002071
Johannes Weiner02695172014-08-06 16:06:17 -07002072 get_scan_count(lruvec, swappiness, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002073
Mel Gormane82e0562013-07-03 15:01:44 -07002074 /* Record the original scan target for proportional adjustments later */
2075 memcpy(targets, nr, sizeof(nr));
2076
Mel Gorman1a501902014-06-04 16:10:49 -07002077 /*
2078 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2079 * event that can occur when there is little memory pressure e.g.
2080 * multiple streaming readers/writers. Hence, we do not abort scanning
2081 * when the requested number of pages are reclaimed when scanning at
2082 * DEF_PRIORITY on the assumption that the fact we are direct
2083 * reclaiming implies that kswapd is not keeping up and it is best to
2084 * do a batch of work at once. For memcg reclaim one check is made to
2085 * abort proportional reclaim if either the file or anon lru has already
2086 * dropped to zero at the first pass.
2087 */
2088 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2089 sc->priority == DEF_PRIORITY);
2090
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002091 blk_start_plug(&plug);
2092 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2093 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002094 unsigned long nr_anon, nr_file, percentage;
2095 unsigned long nr_scanned;
2096
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002097 for_each_evictable_lru(lru) {
2098 if (nr[lru]) {
2099 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2100 nr[lru] -= nr_to_scan;
2101
2102 nr_reclaimed += shrink_list(lru, nr_to_scan,
2103 lruvec, sc);
2104 }
2105 }
Mel Gormane82e0562013-07-03 15:01:44 -07002106
2107 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2108 continue;
2109
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002110 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002111 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002112 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002113 * proportionally what was requested by get_scan_count(). We
2114 * stop reclaiming one LRU and reduce the amount scanning
2115 * proportional to the original scan target.
2116 */
2117 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2118 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2119
Mel Gorman1a501902014-06-04 16:10:49 -07002120 /*
2121 * It's just vindictive to attack the larger once the smaller
2122 * has gone to zero. And given the way we stop scanning the
2123 * smaller below, this makes sure that we only make one nudge
2124 * towards proportionality once we've got nr_to_reclaim.
2125 */
2126 if (!nr_file || !nr_anon)
2127 break;
2128
Mel Gormane82e0562013-07-03 15:01:44 -07002129 if (nr_file > nr_anon) {
2130 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2131 targets[LRU_ACTIVE_ANON] + 1;
2132 lru = LRU_BASE;
2133 percentage = nr_anon * 100 / scan_target;
2134 } else {
2135 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2136 targets[LRU_ACTIVE_FILE] + 1;
2137 lru = LRU_FILE;
2138 percentage = nr_file * 100 / scan_target;
2139 }
2140
2141 /* Stop scanning the smaller of the LRU */
2142 nr[lru] = 0;
2143 nr[lru + LRU_ACTIVE] = 0;
2144
2145 /*
2146 * Recalculate the other LRU scan count based on its original
2147 * scan target and the percentage scanning already complete
2148 */
2149 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2150 nr_scanned = targets[lru] - nr[lru];
2151 nr[lru] = targets[lru] * (100 - percentage) / 100;
2152 nr[lru] -= min(nr[lru], nr_scanned);
2153
2154 lru += LRU_ACTIVE;
2155 nr_scanned = targets[lru] - nr[lru];
2156 nr[lru] = targets[lru] * (100 - percentage) / 100;
2157 nr[lru] -= min(nr[lru], nr_scanned);
2158
2159 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002160 }
2161 blk_finish_plug(&plug);
2162 sc->nr_reclaimed += nr_reclaimed;
2163
2164 /*
2165 * Even if we did not try to evict anon pages at all, we want to
2166 * rebalance the anon lru active/inactive ratio.
2167 */
2168 if (inactive_anon_is_low(lruvec))
2169 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2170 sc, LRU_ACTIVE_ANON);
2171
2172 throttle_vm_writeout(sc->gfp_mask);
2173}
2174
Mel Gorman23b9da52012-05-29 15:06:20 -07002175/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002176static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002177{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002178 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002179 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002180 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002181 return true;
2182
2183 return false;
2184}
2185
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002186/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002187 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2188 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2189 * true if more pages should be reclaimed such that when the page allocator
2190 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2191 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002192 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002193static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002194 unsigned long nr_reclaimed,
2195 unsigned long nr_scanned,
2196 struct scan_control *sc)
2197{
2198 unsigned long pages_for_compaction;
2199 unsigned long inactive_lru_pages;
2200
2201 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002202 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002203 return false;
2204
Mel Gorman28765922011-02-25 14:44:20 -08002205 /* Consider stopping depending on scan and reclaim activity */
2206 if (sc->gfp_mask & __GFP_REPEAT) {
2207 /*
2208 * For __GFP_REPEAT allocations, stop reclaiming if the
2209 * full LRU list has been scanned and we are still failing
2210 * to reclaim pages. This full LRU scan is potentially
2211 * expensive but a __GFP_REPEAT caller really wants to succeed
2212 */
2213 if (!nr_reclaimed && !nr_scanned)
2214 return false;
2215 } else {
2216 /*
2217 * For non-__GFP_REPEAT allocations which can presumably
2218 * fail without consequence, stop if we failed to reclaim
2219 * any pages from the last SWAP_CLUSTER_MAX number of
2220 * pages that were scanned. This will return to the
2221 * caller faster at the risk reclaim/compaction and
2222 * the resulting allocation attempt fails
2223 */
2224 if (!nr_reclaimed)
2225 return false;
2226 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002227
2228 /*
2229 * If we have not reclaimed enough pages for compaction and the
2230 * inactive lists are large enough, continue reclaiming
2231 */
2232 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002233 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002234 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002235 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002236 if (sc->nr_reclaimed < pages_for_compaction &&
2237 inactive_lru_pages > pages_for_compaction)
2238 return true;
2239
2240 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002241 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002242 case COMPACT_PARTIAL:
2243 case COMPACT_CONTINUE:
2244 return false;
2245 default:
2246 return true;
2247 }
2248}
2249
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002250static bool shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002251{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002252 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002253 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002254
Johannes Weiner56600482012-01-12 17:17:59 -08002255 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002256 struct mem_cgroup *root = sc->target_mem_cgroup;
2257 struct mem_cgroup_reclaim_cookie reclaim = {
2258 .zone = zone,
2259 .priority = sc->priority,
2260 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002261 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002262
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002263 nr_reclaimed = sc->nr_reclaimed;
2264 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002265
Andrew Morton694fbc02013-09-24 15:27:37 -07002266 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2267 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002268 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002269 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002270
2271 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002272 swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002273
Johannes Weiner02695172014-08-06 16:06:17 -07002274 shrink_lruvec(lruvec, swappiness, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002275
2276 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002277 * Direct reclaim and kswapd have to scan all memory
2278 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002279 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002280 *
2281 * Limit reclaim, on the other hand, only cares about
2282 * nr_to_reclaim pages to be reclaimed and it will
2283 * retry with decreasing priority if one round over the
2284 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002285 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002286 if (!global_reclaim(sc) &&
2287 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002288 mem_cgroup_iter_break(root, memcg);
2289 break;
2290 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002291 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2292 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002293
2294 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2295 sc->nr_scanned - nr_scanned,
2296 sc->nr_reclaimed - nr_reclaimed);
2297
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002298 if (sc->nr_reclaimed - nr_reclaimed)
2299 reclaimable = true;
2300
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002301 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2302 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002303
2304 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002305}
2306
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002307/* Returns true if compaction should go ahead for a high-order request */
Johannes Weiner0b064962014-08-06 16:06:12 -07002308static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002309{
2310 unsigned long balance_gap, watermark;
2311 bool watermark_ok;
2312
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002313 /*
2314 * Compaction takes time to run and there are potentially other
2315 * callers using the pages just freed. Continue reclaiming until
2316 * there is a buffer of free pages available to give compaction
2317 * a reasonable chance of completing and allocating the page
2318 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002319 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2320 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002321 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002322 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2323
2324 /*
2325 * If compaction is deferred, reclaim up to a point where
2326 * compaction will have a chance of success when re-enabled
2327 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002328 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002329 return watermark_ok;
2330
2331 /* If compaction is not ready to start, keep reclaiming */
Johannes Weiner0b064962014-08-06 16:06:12 -07002332 if (!compaction_suitable(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002333 return false;
2334
2335 return watermark_ok;
2336}
2337
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338/*
2339 * This is the direct reclaim path, for page-allocating processes. We only
2340 * try to reclaim pages from zones which will satisfy the caller's allocation
2341 * request.
2342 *
Mel Gorman41858962009-06-16 15:32:12 -07002343 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2344 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2346 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002347 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2348 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2349 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 * If a zone is deemed to be full of pinned pages then just give it a light
2352 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002353 *
2354 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002356static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357{
Mel Gormandd1a2392008-04-28 02:12:17 -07002358 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002359 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002360 unsigned long nr_soft_reclaimed;
2361 unsigned long nr_soft_scanned;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002362 unsigned long lru_pages = 0;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002363 struct reclaim_state *reclaim_state = current->reclaim_state;
Weijie Yang619d0d762014-04-07 15:36:59 -07002364 gfp_t orig_mask;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002365 struct shrink_control shrink = {
2366 .gfp_mask = sc->gfp_mask,
2367 };
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002368 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002369 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002370
Mel Gormancc715d92012-03-21 16:34:00 -07002371 /*
2372 * If the number of buffer_heads in the machine exceeds the maximum
2373 * allowed level, force direct reclaim to scan the highmem zone as
2374 * highmem pages could be pinning lowmem pages storing buffer_heads
2375 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002376 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002377 if (buffer_heads_over_limit)
2378 sc->gfp_mask |= __GFP_HIGHMEM;
2379
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002380 nodes_clear(shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002381
Mel Gormand4debc62010-08-09 17:19:29 -07002382 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2383 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002384 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002386 /*
2387 * Take care memory controller reclaiming has small influence
2388 * to global LRU.
2389 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002390 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002391 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2392 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002393
2394 lru_pages += zone_reclaimable_pages(zone);
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002395 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002396
Lisa Du6e543d52013-09-11 14:22:36 -07002397 if (sc->priority != DEF_PRIORITY &&
2398 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002399 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002400
2401 /*
2402 * If we already have plenty of memory free for
2403 * compaction in this zone, don't free any more.
2404 * Even though compaction is invoked for any
2405 * non-zero order, only frequent costly order
2406 * reclamation is disruptive enough to become a
2407 * noticeable problem, like transparent huge
2408 * page allocations.
2409 */
2410 if (IS_ENABLED(CONFIG_COMPACTION) &&
2411 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2412 zonelist_zone_idx(z) <= requested_highidx &&
2413 compaction_ready(zone, sc->order)) {
2414 sc->compaction_ready = true;
2415 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002416 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002417
Andrew Morton0608f432013-09-24 15:27:41 -07002418 /*
2419 * This steals pages from memory cgroups over softlimit
2420 * and returns the number of reclaimed pages and
2421 * scanned pages. This works for global memory pressure
2422 * and balancing, not for a memcg's limit.
2423 */
2424 nr_soft_scanned = 0;
2425 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2426 sc->order, sc->gfp_mask,
2427 &nr_soft_scanned);
2428 sc->nr_reclaimed += nr_soft_reclaimed;
2429 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002430 if (nr_soft_reclaimed)
2431 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002432 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002433 }
Nick Piggin408d8542006-09-25 23:31:27 -07002434
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002435 if (shrink_zone(zone, sc))
2436 reclaimable = true;
2437
2438 if (global_reclaim(sc) &&
2439 !reclaimable && zone_reclaimable(zone))
2440 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 }
Mel Gormane0c23272011-10-31 17:09:33 -07002442
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002443 /*
2444 * Don't shrink slabs when reclaiming memory from over limit cgroups
2445 * but do shrink slab at least once when aborting reclaim for
2446 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2447 * pages.
2448 */
2449 if (global_reclaim(sc)) {
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002450 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002451 if (reclaim_state) {
2452 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2453 reclaim_state->reclaimed_slab = 0;
2454 }
2455 }
2456
Weijie Yang619d0d762014-04-07 15:36:59 -07002457 /*
2458 * Restore to original mask to avoid the impact on the caller if we
2459 * promoted it to __GFP_HIGHMEM.
2460 */
2461 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002462
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002463 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002465
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466/*
2467 * This is the main entry point to direct page reclaim.
2468 *
2469 * If a full scan of the inactive list fails to free enough memory then we
2470 * are "out of memory" and something needs to be killed.
2471 *
2472 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2473 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002474 * caller can't do much about. We kick the writeback threads and take explicit
2475 * naps in the hope that some of these pages can be written. But if the
2476 * allocating task holds filesystem locks which prevent writeout this might not
2477 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002478 *
2479 * returns: 0, if no pages reclaimed
2480 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 */
Mel Gormandac1d272008-04-28 02:12:12 -07002482static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002483 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484{
Andrew Morton69e05942006-03-22 00:08:19 -08002485 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002486 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002487 bool zones_reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488
Keika Kobayashi873b4772008-07-25 01:48:52 -07002489 delayacct_freepages_start();
2490
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002491 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002492 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002494 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002495 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2496 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002497 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002498 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002499
Balbir Singh66e17072008-02-07 00:13:56 -08002500 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002501 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002502 break;
2503
2504 if (sc->compaction_ready)
2505 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
2507 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002508 * If we're getting trouble reclaiming, start doing
2509 * writepage even in laptop mode.
2510 */
2511 if (sc->priority < DEF_PRIORITY - 2)
2512 sc->may_writepage = 1;
2513
2514 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 * Try to write back as many pages as we just scanned. This
2516 * tends to cause slow streaming writers to write data to the
2517 * disk smoothly, at the dirtying rate, which is nice. But
2518 * that's undesirable in laptop mode, where we *want* lumpy
2519 * writeout. So in laptop mode, write out the whole world.
2520 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002521 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2522 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002523 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2524 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002525 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002527 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002528
Keika Kobayashi873b4772008-07-25 01:48:52 -07002529 delayacct_freepages_end();
2530
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002531 if (sc->nr_reclaimed)
2532 return sc->nr_reclaimed;
2533
Mel Gorman0cee34f2012-01-12 17:19:49 -08002534 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002535 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002536 return 1;
2537
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002538 /* Any of the zones still reclaimable? Don't OOM. */
2539 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002540 return 1;
2541
2542 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543}
2544
Mel Gorman55150612012-07-31 16:44:35 -07002545static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2546{
2547 struct zone *zone;
2548 unsigned long pfmemalloc_reserve = 0;
2549 unsigned long free_pages = 0;
2550 int i;
2551 bool wmark_ok;
2552
2553 for (i = 0; i <= ZONE_NORMAL; i++) {
2554 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002555 if (!populated_zone(zone))
2556 continue;
2557
Mel Gorman55150612012-07-31 16:44:35 -07002558 pfmemalloc_reserve += min_wmark_pages(zone);
2559 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2560 }
2561
Mel Gorman675becc2014-06-04 16:07:35 -07002562 /* If there are no reserves (unexpected config) then do not throttle */
2563 if (!pfmemalloc_reserve)
2564 return true;
2565
Mel Gorman55150612012-07-31 16:44:35 -07002566 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2567
2568 /* kswapd must be awake if processes are being throttled */
2569 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2570 pgdat->classzone_idx = min(pgdat->classzone_idx,
2571 (enum zone_type)ZONE_NORMAL);
2572 wake_up_interruptible(&pgdat->kswapd_wait);
2573 }
2574
2575 return wmark_ok;
2576}
2577
2578/*
2579 * Throttle direct reclaimers if backing storage is backed by the network
2580 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2581 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002582 * when the low watermark is reached.
2583 *
2584 * Returns true if a fatal signal was delivered during throttling. If this
2585 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002586 */
Mel Gorman50694c22012-11-26 16:29:48 -08002587static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002588 nodemask_t *nodemask)
2589{
Mel Gorman675becc2014-06-04 16:07:35 -07002590 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002591 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002592 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002593
2594 /*
2595 * Kernel threads should not be throttled as they may be indirectly
2596 * responsible for cleaning pages necessary for reclaim to make forward
2597 * progress. kjournald for example may enter direct reclaim while
2598 * committing a transaction where throttling it could forcing other
2599 * processes to block on log_wait_commit().
2600 */
2601 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002602 goto out;
2603
2604 /*
2605 * If a fatal signal is pending, this process should not throttle.
2606 * It should return quickly so it can exit and free its memory
2607 */
2608 if (fatal_signal_pending(current))
2609 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002610
Mel Gorman675becc2014-06-04 16:07:35 -07002611 /*
2612 * Check if the pfmemalloc reserves are ok by finding the first node
2613 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2614 * GFP_KERNEL will be required for allocating network buffers when
2615 * swapping over the network so ZONE_HIGHMEM is unusable.
2616 *
2617 * Throttling is based on the first usable node and throttled processes
2618 * wait on a queue until kswapd makes progress and wakes them. There
2619 * is an affinity then between processes waking up and where reclaim
2620 * progress has been made assuming the process wakes on the same node.
2621 * More importantly, processes running on remote nodes will not compete
2622 * for remote pfmemalloc reserves and processes on different nodes
2623 * should make reasonable progress.
2624 */
2625 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2626 gfp_mask, nodemask) {
2627 if (zone_idx(zone) > ZONE_NORMAL)
2628 continue;
2629
2630 /* Throttle based on the first usable node */
2631 pgdat = zone->zone_pgdat;
2632 if (pfmemalloc_watermark_ok(pgdat))
2633 goto out;
2634 break;
2635 }
2636
2637 /* If no zone was usable by the allocation flags then do not throttle */
2638 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002639 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002640
Mel Gorman68243e72012-07-31 16:44:39 -07002641 /* Account for the throttling */
2642 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2643
Mel Gorman55150612012-07-31 16:44:35 -07002644 /*
2645 * If the caller cannot enter the filesystem, it's possible that it
2646 * is due to the caller holding an FS lock or performing a journal
2647 * transaction in the case of a filesystem like ext[3|4]. In this case,
2648 * it is not safe to block on pfmemalloc_wait as kswapd could be
2649 * blocked waiting on the same lock. Instead, throttle for up to a
2650 * second before continuing.
2651 */
2652 if (!(gfp_mask & __GFP_FS)) {
2653 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2654 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002655
2656 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002657 }
2658
2659 /* Throttle until kswapd wakes the process */
2660 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2661 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002662
2663check_pending:
2664 if (fatal_signal_pending(current))
2665 return true;
2666
2667out:
2668 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002669}
2670
Mel Gormandac1d272008-04-28 02:12:12 -07002671unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002672 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002673{
Mel Gorman33906bc2010-08-09 17:19:16 -07002674 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002675 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002676 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002677 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2678 .order = order,
2679 .nodemask = nodemask,
2680 .priority = DEF_PRIORITY,
2681 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002682 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002683 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002684 };
2685
Mel Gorman55150612012-07-31 16:44:35 -07002686 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002687 * Do not enter reclaim if fatal signal was delivered while throttled.
2688 * 1 is returned so that the page allocator does not OOM kill at this
2689 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002690 */
Mel Gorman50694c22012-11-26 16:29:48 -08002691 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002692 return 1;
2693
Mel Gorman33906bc2010-08-09 17:19:16 -07002694 trace_mm_vmscan_direct_reclaim_begin(order,
2695 sc.may_writepage,
2696 gfp_mask);
2697
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002698 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002699
2700 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2701
2702 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002703}
2704
Andrew Mortonc255a452012-07-31 16:43:02 -07002705#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002706
Johannes Weiner72835c82012-01-12 17:18:32 -08002707unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002708 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002709 struct zone *zone,
2710 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002711{
2712 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002713 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002714 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002715 .may_writepage = !laptop_mode,
2716 .may_unmap = 1,
2717 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002718 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002719 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002720 int swappiness = mem_cgroup_swappiness(memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002721
Balbir Singh4e416952009-09-23 15:56:39 -07002722 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2723 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002724
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002725 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002726 sc.may_writepage,
2727 sc.gfp_mask);
2728
Balbir Singh4e416952009-09-23 15:56:39 -07002729 /*
2730 * NOTE: Although we can get the priority field, using it
2731 * here is not a good idea, since it limits the pages we can scan.
2732 * if we don't reclaim here, the shrink_zone from balance_pgdat
2733 * will pick up pages from other mem cgroup's as well. We hack
2734 * the priority and make it zero.
2735 */
Johannes Weiner02695172014-08-06 16:06:17 -07002736 shrink_lruvec(lruvec, swappiness, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002737
2738 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2739
Ying Han0ae5e892011-05-26 16:25:25 -07002740 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002741 return sc.nr_reclaimed;
2742}
2743
Johannes Weiner72835c82012-01-12 17:18:32 -08002744unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002745 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002746 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002747{
Balbir Singh4e416952009-09-23 15:56:39 -07002748 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002749 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002750 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002751 struct scan_control sc = {
Johannes Weineree814fe2014-08-06 16:06:19 -07002752 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2753 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2754 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2755 .target_mem_cgroup = memcg,
2756 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002757 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002758 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002759 .may_swap = !noswap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002760 };
Balbir Singh66e17072008-02-07 00:13:56 -08002761
Ying Han889976d2011-05-26 16:25:33 -07002762 /*
2763 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2764 * take care of from where we get pages. So the node where we start the
2765 * scan does not need to be the current node.
2766 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002767 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002768
2769 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002770
2771 trace_mm_vmscan_memcg_reclaim_begin(0,
2772 sc.may_writepage,
2773 sc.gfp_mask);
2774
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002775 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002776
2777 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2778
2779 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002780}
2781#endif
2782
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002783static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002784{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002785 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002786
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002787 if (!total_swap_pages)
2788 return;
2789
2790 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2791 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002792 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002793
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002794 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002795 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002796 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002797
2798 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2799 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002800}
2801
Johannes Weiner60cefed2012-11-29 13:54:23 -08002802static bool zone_balanced(struct zone *zone, int order,
2803 unsigned long balance_gap, int classzone_idx)
2804{
2805 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2806 balance_gap, classzone_idx, 0))
2807 return false;
2808
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002809 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2810 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002811 return false;
2812
2813 return true;
2814}
2815
Mel Gorman1741c872011-01-13 15:46:21 -08002816/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002817 * pgdat_balanced() is used when checking if a node is balanced.
2818 *
2819 * For order-0, all zones must be balanced!
2820 *
2821 * For high-order allocations only zones that meet watermarks and are in a
2822 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2823 * total of balanced pages must be at least 25% of the zones allowed by
2824 * classzone_idx for the node to be considered balanced. Forcing all zones to
2825 * be balanced for high orders can cause excessive reclaim when there are
2826 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002827 * The choice of 25% is due to
2828 * o a 16M DMA zone that is balanced will not balance a zone on any
2829 * reasonable sized machine
2830 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002831 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002832 * would need to be at least 256M for it to be balance a whole node.
2833 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2834 * to balance a node on its own. These seemed like reasonable ratios.
2835 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002836static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002837{
Jiang Liub40da042013-02-22 16:33:52 -08002838 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002839 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002840 int i;
2841
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002842 /* Check the watermark levels */
2843 for (i = 0; i <= classzone_idx; i++) {
2844 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002845
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002846 if (!populated_zone(zone))
2847 continue;
2848
Jiang Liub40da042013-02-22 16:33:52 -08002849 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002850
2851 /*
2852 * A special case here:
2853 *
2854 * balance_pgdat() skips over all_unreclaimable after
2855 * DEF_PRIORITY. Effectively, it considers them balanced so
2856 * they must be considered balanced here as well!
2857 */
Lisa Du6e543d52013-09-11 14:22:36 -07002858 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002859 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002860 continue;
2861 }
2862
2863 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002864 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002865 else if (!order)
2866 return false;
2867 }
2868
2869 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002870 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002871 else
2872 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002873}
2874
Mel Gorman55150612012-07-31 16:44:35 -07002875/*
2876 * Prepare kswapd for sleeping. This verifies that there are no processes
2877 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2878 *
2879 * Returns true if kswapd is ready to sleep
2880 */
2881static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002882 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002883{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002884 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2885 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002886 return false;
2887
2888 /*
2889 * There is a potential race between when kswapd checks its watermarks
2890 * and a process gets throttled. There is also a potential race if
2891 * processes get throttled, kswapd wakes, a large process exits therby
2892 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2893 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2894 * so wake them now if necessary. If necessary, processes will wake
2895 * kswapd and get throttled again
2896 */
2897 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2898 wake_up(&pgdat->pfmemalloc_wait);
2899 return false;
2900 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002901
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002902 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002903}
2904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905/*
Mel Gorman75485362013-07-03 15:01:42 -07002906 * kswapd shrinks the zone by the number of pages required to reach
2907 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002908 *
2909 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002910 * reclaim or if the lack of progress was due to pages under writeback.
2911 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002912 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002913static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002914 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002915 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002916 unsigned long lru_pages,
2917 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002918{
Mel Gorman7c954f62013-07-03 15:01:54 -07002919 int testorder = sc->order;
2920 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002921 struct reclaim_state *reclaim_state = current->reclaim_state;
2922 struct shrink_control shrink = {
2923 .gfp_mask = sc->gfp_mask,
2924 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002925 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002926
2927 /* Reclaim above the high watermark. */
2928 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002929
2930 /*
2931 * Kswapd reclaims only single pages with compaction enabled. Trying
2932 * too hard to reclaim until contiguous free pages have become
2933 * available can hurt performance by evicting too much useful data
2934 * from memory. Do not reclaim more than needed for compaction.
2935 */
2936 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2937 compaction_suitable(zone, sc->order) !=
2938 COMPACT_SKIPPED)
2939 testorder = 0;
2940
2941 /*
2942 * We put equal pressure on every zone, unless one zone has way too
2943 * many pages free already. The "too many pages" is defined as the
2944 * high wmark plus a "gap" where the gap is either the low
2945 * watermark or 1% of the zone, whichever is smaller.
2946 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002947 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2948 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07002949
2950 /*
2951 * If there is no low memory pressure or the zone is balanced then no
2952 * reclaim is necessary
2953 */
2954 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2955 if (!lowmem_pressure && zone_balanced(zone, testorder,
2956 balance_gap, classzone_idx))
2957 return true;
2958
Mel Gorman75485362013-07-03 15:01:42 -07002959 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002960 nodes_clear(shrink.nodes_to_scan);
2961 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002962
2963 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002964 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002965 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2966
Mel Gorman2ab44f42013-07-03 15:01:47 -07002967 /* Account for the number of pages attempted to reclaim */
2968 *nr_attempted += sc->nr_to_reclaim;
2969
Mel Gorman283aba92013-07-03 15:01:51 -07002970 zone_clear_flag(zone, ZONE_WRITEBACK);
2971
Mel Gorman7c954f62013-07-03 15:01:54 -07002972 /*
2973 * If a zone reaches its high watermark, consider it to be no longer
2974 * congested. It's possible there are dirty pages backed by congested
2975 * BDIs but as pressure is relieved, speculatively avoid congestion
2976 * waits.
2977 */
Lisa Du6e543d52013-09-11 14:22:36 -07002978 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002979 zone_balanced(zone, testorder, 0, classzone_idx)) {
2980 zone_clear_flag(zone, ZONE_CONGESTED);
2981 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2982 }
2983
Mel Gormanb8e83b92013-07-03 15:01:45 -07002984 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002985}
2986
2987/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002989 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002991 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 *
2993 * There is special handling here for zones which are full of pinned pages.
2994 * This can happen if the pages are all mlocked, or if they are all used by
2995 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2996 * What we do is to detect the case where all pages in the zone have been
2997 * scanned twice and there has been zero successful reclaim. Mark the zone as
2998 * dead and from now on, only perform a short scan. Basically we're polling
2999 * the zone for when the problem goes away.
3000 *
3001 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003002 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3003 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3004 * lower zones regardless of the number of free pages in the lower zones. This
3005 * interoperates with the page allocator fallback scheme to ensure that aging
3006 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 */
Mel Gorman99504742011-01-13 15:46:20 -08003008static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003009 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003012 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003013 unsigned long nr_soft_reclaimed;
3014 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003015 struct scan_control sc = {
3016 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003017 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003018 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003019 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003020 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003021 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003022 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003023 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003025 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003027 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003028 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003029 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003030
3031 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003033 /*
3034 * Scan in the highmem->dma direction for the highest
3035 * zone which needs scanning
3036 */
3037 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3038 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003040 if (!populated_zone(zone))
3041 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042
Lisa Du6e543d52013-09-11 14:22:36 -07003043 if (sc.priority != DEF_PRIORITY &&
3044 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003045 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046
Rik van Riel556adec2008-10-18 20:26:34 -07003047 /*
3048 * Do some background aging of the anon list, to give
3049 * pages a chance to be referenced before reclaiming.
3050 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003051 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003052
Mel Gormancc715d92012-03-21 16:34:00 -07003053 /*
3054 * If the number of buffer_heads in the machine
3055 * exceeds the maximum allowed level and this node
3056 * has a highmem zone, force kswapd to reclaim from
3057 * it to relieve lowmem pressure.
3058 */
3059 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3060 end_zone = i;
3061 break;
3062 }
3063
Johannes Weiner60cefed2012-11-29 13:54:23 -08003064 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003065 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003066 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003067 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003068 /*
3069 * If balanced, clear the dirty and congested
3070 * flags
3071 */
Shaohua Li439423f2011-08-25 15:59:12 -07003072 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003073 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003076
Mel Gormanb8e83b92013-07-03 15:01:45 -07003077 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003078 goto out;
3079
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 for (i = 0; i <= end_zone; i++) {
3081 struct zone *zone = pgdat->node_zones + i;
3082
Mel Gorman2ab44f42013-07-03 15:01:47 -07003083 if (!populated_zone(zone))
3084 continue;
3085
Wu Fengguangadea02a2009-09-21 17:01:42 -07003086 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003087
3088 /*
3089 * If any zone is currently balanced then kswapd will
3090 * not call compaction as it is expected that the
3091 * necessary pages are already available.
3092 */
3093 if (pgdat_needs_compaction &&
3094 zone_watermark_ok(zone, order,
3095 low_wmark_pages(zone),
3096 *classzone_idx, 0))
3097 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 }
3099
3100 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003101 * If we're getting trouble reclaiming, start doing writepage
3102 * even in laptop mode.
3103 */
3104 if (sc.priority < DEF_PRIORITY - 2)
3105 sc.may_writepage = 1;
3106
3107 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 * Now scan the zone in the dma->highmem direction, stopping
3109 * at the last zone which needs scanning.
3110 *
3111 * We do this because the page allocator works in the opposite
3112 * direction. This prevents the page allocator from allocating
3113 * pages behind kswapd's direction of progress, which would
3114 * cause too much scanning of the lower zones.
3115 */
3116 for (i = 0; i <= end_zone; i++) {
3117 struct zone *zone = pgdat->node_zones + i;
3118
Con Kolivasf3fe6512006-01-06 00:11:15 -08003119 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 continue;
3121
Lisa Du6e543d52013-09-11 14:22:36 -07003122 if (sc.priority != DEF_PRIORITY &&
3123 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124 continue;
3125
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003127
Andrew Morton0608f432013-09-24 15:27:41 -07003128 nr_soft_scanned = 0;
3129 /*
3130 * Call soft limit reclaim before calling shrink_zone.
3131 */
3132 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3133 order, sc.gfp_mask,
3134 &nr_soft_scanned);
3135 sc.nr_reclaimed += nr_soft_reclaimed;
3136
Rik van Riel32a43302007-10-16 01:24:50 -07003137 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003138 * There should be no need to raise the scanning
3139 * priority if enough pages are already being scanned
3140 * that that high watermark would be met at 100%
3141 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003142 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003143 if (kswapd_shrink_zone(zone, end_zone, &sc,
3144 lru_pages, &nr_attempted))
3145 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 }
Mel Gorman55150612012-07-31 16:44:35 -07003147
3148 /*
3149 * If the low watermark is met there is no need for processes
3150 * to be throttled on pfmemalloc_wait as they should not be
3151 * able to safely make forward progress. Wake them
3152 */
3153 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3154 pfmemalloc_watermark_ok(pgdat))
3155 wake_up(&pgdat->pfmemalloc_wait);
3156
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003158 * Fragmentation may mean that the system cannot be rebalanced
3159 * for high-order allocations in all zones. If twice the
3160 * allocation size has been reclaimed and the zones are still
3161 * not balanced then recheck the watermarks at order-0 to
3162 * prevent kswapd reclaiming excessively. Assume that a
3163 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003165 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003166 order = sc.order = 0;
3167
Mel Gormanb8e83b92013-07-03 15:01:45 -07003168 /* Check if kswapd should be suspending */
3169 if (try_to_freeze() || kthread_should_stop())
3170 break;
3171
3172 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003173 * Compact if necessary and kswapd is reclaiming at least the
3174 * high watermark number of pages as requsted
3175 */
3176 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3177 compact_pgdat(pgdat, order);
3178
3179 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003180 * Raise priority if scanning rate is too low or there was no
3181 * progress in reclaiming pages
3182 */
3183 if (raise_priority || !sc.nr_reclaimed)
3184 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003185 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003186 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187
Mel Gormanb8e83b92013-07-03 15:01:45 -07003188out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003189 /*
Mel Gorman55150612012-07-31 16:44:35 -07003190 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003191 * makes a decision on the order we were last reclaiming at. However,
3192 * if another caller entered the allocator slow path while kswapd
3193 * was awake, order will remain at the higher level
3194 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003195 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003196 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197}
3198
Mel Gormandc83edd2011-01-13 15:46:26 -08003199static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003200{
3201 long remaining = 0;
3202 DEFINE_WAIT(wait);
3203
3204 if (freezing(current) || kthread_should_stop())
3205 return;
3206
3207 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3208
3209 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003210 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003211 remaining = schedule_timeout(HZ/10);
3212 finish_wait(&pgdat->kswapd_wait, &wait);
3213 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3214 }
3215
3216 /*
3217 * After a short sleep, check if it was a premature sleep. If not, then
3218 * go fully to sleep until explicitly woken up.
3219 */
Mel Gorman55150612012-07-31 16:44:35 -07003220 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003221 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3222
3223 /*
3224 * vmstat counters are not perfectly accurate and the estimated
3225 * value for counters such as NR_FREE_PAGES can deviate from the
3226 * true value by nr_online_cpus * threshold. To avoid the zone
3227 * watermarks being breached while under pressure, we reduce the
3228 * per-cpu vmstat threshold while kswapd is awake and restore
3229 * them before going back to sleep.
3230 */
3231 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003232
Mel Gorman62997022012-10-08 16:32:47 -07003233 /*
3234 * Compaction records what page blocks it recently failed to
3235 * isolate pages from and skips them in the future scanning.
3236 * When kswapd is going to sleep, it is reasonable to assume
3237 * that pages and compaction may succeed so reset the cache.
3238 */
3239 reset_isolation_suitable(pgdat);
3240
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003241 if (!kthread_should_stop())
3242 schedule();
3243
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003244 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3245 } else {
3246 if (remaining)
3247 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3248 else
3249 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3250 }
3251 finish_wait(&pgdat->kswapd_wait, &wait);
3252}
3253
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254/*
3255 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003256 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 *
3258 * This basically trickles out pages so that we have _some_
3259 * free memory available even if there is no other activity
3260 * that frees anything up. This is needed for things like routing
3261 * etc, where we otherwise might have all activity going on in
3262 * asynchronous contexts that cannot page things out.
3263 *
3264 * If there are applications that are active memory-allocators
3265 * (most normal use), this basically shouldn't matter.
3266 */
3267static int kswapd(void *p)
3268{
Mel Gorman215ddd62011-07-08 15:39:40 -07003269 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003270 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003271 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003272 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 pg_data_t *pgdat = (pg_data_t*)p;
3274 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003275
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 struct reclaim_state reclaim_state = {
3277 .reclaimed_slab = 0,
3278 };
Rusty Russella70f7302009-03-13 14:49:46 +10303279 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280
Nick Piggincf40bd12009-01-21 08:12:39 +01003281 lockdep_set_current_reclaim_state(GFP_KERNEL);
3282
Rusty Russell174596a2009-01-01 10:12:29 +10303283 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003284 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 current->reclaim_state = &reclaim_state;
3286
3287 /*
3288 * Tell the memory management that we're a "memory allocator",
3289 * and that if we need more memory we should get access to it
3290 * regardless (see "__alloc_pages()"). "kswapd" should
3291 * never get caught in the normal page freeing logic.
3292 *
3293 * (Kswapd normally doesn't need memory anyway, but sometimes
3294 * you need a small amount of memory in order to be able to
3295 * page out something else, and this flag essentially protects
3296 * us from recursively trying to free more memory as we're
3297 * trying to free the first piece of memory in the first place).
3298 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003299 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003300 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301
Mel Gorman215ddd62011-07-08 15:39:40 -07003302 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003303 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003304 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003305 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003307 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003308
Mel Gorman215ddd62011-07-08 15:39:40 -07003309 /*
3310 * If the last balance_pgdat was unsuccessful it's unlikely a
3311 * new request of a similar or harder type will succeed soon
3312 * so consider going to sleep on the basis we reclaimed at
3313 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003314 if (balanced_classzone_idx >= new_classzone_idx &&
3315 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003316 new_order = pgdat->kswapd_max_order;
3317 new_classzone_idx = pgdat->classzone_idx;
3318 pgdat->kswapd_max_order = 0;
3319 pgdat->classzone_idx = pgdat->nr_zones - 1;
3320 }
3321
Mel Gorman99504742011-01-13 15:46:20 -08003322 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 /*
3324 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003325 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 */
3327 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003328 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003330 kswapd_try_to_sleep(pgdat, balanced_order,
3331 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003333 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003334 new_order = order;
3335 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003336 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003337 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339
David Rientjes8fe23e02009-12-14 17:58:33 -08003340 ret = try_to_freeze();
3341 if (kthread_should_stop())
3342 break;
3343
3344 /*
3345 * We can speed up thawing tasks if we don't call balance_pgdat
3346 * after returning from the refrigerator
3347 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003348 if (!ret) {
3349 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003350 balanced_classzone_idx = classzone_idx;
3351 balanced_order = balance_pgdat(pgdat, order,
3352 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003353 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003355
Johannes Weiner71abdc12014-06-06 14:35:35 -07003356 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003357 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003358 lockdep_clear_current_reclaim_state();
3359
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 return 0;
3361}
3362
3363/*
3364 * A zone is low on free memory, so wake its kswapd task to service it.
3365 */
Mel Gorman99504742011-01-13 15:46:20 -08003366void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367{
3368 pg_data_t *pgdat;
3369
Con Kolivasf3fe6512006-01-06 00:11:15 -08003370 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 return;
3372
Paul Jackson02a0e532006-12-13 00:34:25 -08003373 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003375 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003376 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003377 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003378 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3379 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003380 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003382 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003383 return;
3384
3385 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003386 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387}
3388
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003389#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003391 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003392 * freed pages.
3393 *
3394 * Rather than trying to age LRUs the aim is to preserve the overall
3395 * LRU order by reclaiming preferentially
3396 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003398unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003400 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003401 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003402 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003403 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003404 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003405 .may_writepage = 1,
3406 .may_unmap = 1,
3407 .may_swap = 1,
3408 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003410 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003411 struct task_struct *p = current;
3412 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003414 p->flags |= PF_MEMALLOC;
3415 lockdep_set_current_reclaim_state(sc.gfp_mask);
3416 reclaim_state.reclaimed_slab = 0;
3417 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003418
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003419 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003420
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003421 p->reclaim_state = NULL;
3422 lockdep_clear_current_reclaim_state();
3423 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003424
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003425 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003427#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429/* It's optimal to keep kswapds on the same CPUs as their memory, but
3430 not required for correctness. So if the last cpu in a node goes
3431 away, we get changed to run anywhere: as the first one comes back,
3432 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003433static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3434 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003436 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003438 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003439 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003440 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303441 const struct cpumask *mask;
3442
3443 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003444
Rusty Russell3e597942009-01-01 10:12:24 +10303445 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003447 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 }
3449 }
3450 return NOTIFY_OK;
3451}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452
Yasunori Goto3218ae12006-06-27 02:53:33 -07003453/*
3454 * This kswapd start function will be called by init and node-hot-add.
3455 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3456 */
3457int kswapd_run(int nid)
3458{
3459 pg_data_t *pgdat = NODE_DATA(nid);
3460 int ret = 0;
3461
3462 if (pgdat->kswapd)
3463 return 0;
3464
3465 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3466 if (IS_ERR(pgdat->kswapd)) {
3467 /* failure at boot is fatal */
3468 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003469 pr_err("Failed to start kswapd on node %d\n", nid);
3470 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003471 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003472 }
3473 return ret;
3474}
3475
David Rientjes8fe23e02009-12-14 17:58:33 -08003476/*
Jiang Liud8adde12012-07-11 14:01:52 -07003477 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003478 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003479 */
3480void kswapd_stop(int nid)
3481{
3482 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3483
Jiang Liud8adde12012-07-11 14:01:52 -07003484 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003485 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003486 NODE_DATA(nid)->kswapd = NULL;
3487 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003488}
3489
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490static int __init kswapd_init(void)
3491{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003492 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003493
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003495 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003496 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 hotcpu_notifier(cpu_callback, 0);
3498 return 0;
3499}
3500
3501module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003502
3503#ifdef CONFIG_NUMA
3504/*
3505 * Zone reclaim mode
3506 *
3507 * If non-zero call zone_reclaim when the number of free pages falls below
3508 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003509 */
3510int zone_reclaim_mode __read_mostly;
3511
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003512#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003513#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003514#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3515#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3516
Christoph Lameter9eeff232006-01-18 17:42:31 -08003517/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003518 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3519 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3520 * a zone.
3521 */
3522#define ZONE_RECLAIM_PRIORITY 4
3523
Christoph Lameter9eeff232006-01-18 17:42:31 -08003524/*
Christoph Lameter96146342006-07-03 00:24:13 -07003525 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3526 * occur.
3527 */
3528int sysctl_min_unmapped_ratio = 1;
3529
3530/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003531 * If the number of slab pages in a zone grows beyond this percentage then
3532 * slab reclaim needs to occur.
3533 */
3534int sysctl_min_slab_ratio = 5;
3535
Mel Gorman90afa5d2009-06-16 15:33:20 -07003536static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3537{
3538 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3539 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3540 zone_page_state(zone, NR_ACTIVE_FILE);
3541
3542 /*
3543 * It's possible for there to be more file mapped pages than
3544 * accounted for by the pages on the file LRU lists because
3545 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3546 */
3547 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3548}
3549
3550/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3551static long zone_pagecache_reclaimable(struct zone *zone)
3552{
3553 long nr_pagecache_reclaimable;
3554 long delta = 0;
3555
3556 /*
3557 * If RECLAIM_SWAP is set, then all file pages are considered
3558 * potentially reclaimable. Otherwise, we have to worry about
3559 * pages like swapcache and zone_unmapped_file_pages() provides
3560 * a better estimate
3561 */
3562 if (zone_reclaim_mode & RECLAIM_SWAP)
3563 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3564 else
3565 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3566
3567 /* If we can't clean pages, remove dirty pages from consideration */
3568 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3569 delta += zone_page_state(zone, NR_FILE_DIRTY);
3570
3571 /* Watch for any possible underflows due to delta */
3572 if (unlikely(delta > nr_pagecache_reclaimable))
3573 delta = nr_pagecache_reclaimable;
3574
3575 return nr_pagecache_reclaimable - delta;
3576}
3577
Christoph Lameter0ff38492006-09-25 23:31:52 -07003578/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003579 * Try to free up some pages from this zone through reclaim.
3580 */
Andrew Morton179e9632006-03-22 00:08:18 -08003581static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003582{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003583 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003584 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003585 struct task_struct *p = current;
3586 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003587 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003588 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003589 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003590 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003591 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003592 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3593 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3594 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003595 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003596 struct shrink_control shrink = {
3597 .gfp_mask = sc.gfp_mask,
3598 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003599 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003600
Christoph Lameter9eeff232006-01-18 17:42:31 -08003601 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003602 /*
3603 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3604 * and we also need to be able to write out pages for RECLAIM_WRITE
3605 * and RECLAIM_SWAP.
3606 */
3607 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003608 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003609 reclaim_state.reclaimed_slab = 0;
3610 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003611
Mel Gorman90afa5d2009-06-16 15:33:20 -07003612 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003613 /*
3614 * Free memory by calling shrink zone with increasing
3615 * priorities until we have enough memory freed.
3616 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003617 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003618 shrink_zone(zone, &sc);
3619 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003620 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003621
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003622 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3623 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003624 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003625 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003626 * many pages were freed in this zone. So we take the current
3627 * number of slab pages and shake the slab until it is reduced
3628 * by the same nr_pages that we used for reclaiming unmapped
3629 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003630 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003631 nodes_clear(shrink.nodes_to_scan);
3632 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003633 for (;;) {
3634 unsigned long lru_pages = zone_reclaimable_pages(zone);
3635
3636 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003637 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003638 break;
3639
3640 /* Freed enough memory */
3641 nr_slab_pages1 = zone_page_state(zone,
3642 NR_SLAB_RECLAIMABLE);
3643 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3644 break;
3645 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003646
3647 /*
3648 * Update nr_reclaimed by the number of slab pages we
3649 * reclaimed from this zone.
3650 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003651 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3652 if (nr_slab_pages1 < nr_slab_pages0)
3653 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003654 }
3655
Christoph Lameter9eeff232006-01-18 17:42:31 -08003656 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003657 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003658 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003659 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003660}
Andrew Morton179e9632006-03-22 00:08:18 -08003661
3662int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3663{
Andrew Morton179e9632006-03-22 00:08:18 -08003664 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003665 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003666
3667 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003668 * Zone reclaim reclaims unmapped file backed pages and
3669 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003670 *
Christoph Lameter96146342006-07-03 00:24:13 -07003671 * A small portion of unmapped file backed pages is needed for
3672 * file I/O otherwise pages read by file I/O will be immediately
3673 * thrown out if the zone is overallocated. So we do not reclaim
3674 * if less than a specified percentage of the zone is used by
3675 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003676 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003677 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3678 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003679 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003680
Lisa Du6e543d52013-09-11 14:22:36 -07003681 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003682 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003683
Andrew Morton179e9632006-03-22 00:08:18 -08003684 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003685 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003686 */
David Rientjesd773ed62007-10-16 23:26:01 -07003687 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003688 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003689
3690 /*
3691 * Only run zone reclaim on the local zone or on zones that do not
3692 * have associated processors. This will favor the local processor
3693 * over remote processors and spread off node memory allocations
3694 * as wide as possible.
3695 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003696 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003697 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003698 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003699
3700 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003701 return ZONE_RECLAIM_NOSCAN;
3702
David Rientjesd773ed62007-10-16 23:26:01 -07003703 ret = __zone_reclaim(zone, gfp_mask, order);
3704 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3705
Mel Gorman24cf725182009-06-16 15:33:23 -07003706 if (!ret)
3707 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3708
David Rientjesd773ed62007-10-16 23:26:01 -07003709 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003710}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003711#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003712
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003713/*
3714 * page_evictable - test whether a page is evictable
3715 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003716 *
3717 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003718 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003719 *
3720 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003721 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003722 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003723 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003724 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003725int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003726{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003727 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003728}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003729
Hugh Dickins850465792012-01-20 14:34:19 -08003730#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003731/**
Hugh Dickins24513262012-01-20 14:34:21 -08003732 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3733 * @pages: array of pages to check
3734 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003735 *
Hugh Dickins24513262012-01-20 14:34:21 -08003736 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003737 *
3738 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003739 */
Hugh Dickins24513262012-01-20 14:34:21 -08003740void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003741{
Johannes Weiner925b7672012-01-12 17:18:15 -08003742 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003743 struct zone *zone = NULL;
3744 int pgscanned = 0;
3745 int pgrescued = 0;
3746 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003747
Hugh Dickins24513262012-01-20 14:34:21 -08003748 for (i = 0; i < nr_pages; i++) {
3749 struct page *page = pages[i];
3750 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003751
Hugh Dickins24513262012-01-20 14:34:21 -08003752 pgscanned++;
3753 pagezone = page_zone(page);
3754 if (pagezone != zone) {
3755 if (zone)
3756 spin_unlock_irq(&zone->lru_lock);
3757 zone = pagezone;
3758 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003759 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003760 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003761
Hugh Dickins24513262012-01-20 14:34:21 -08003762 if (!PageLRU(page) || !PageUnevictable(page))
3763 continue;
3764
Hugh Dickins39b5f292012-10-08 16:33:18 -07003765 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003766 enum lru_list lru = page_lru_base_type(page);
3767
Sasha Levin309381fea2014-01-23 15:52:54 -08003768 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003769 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003770 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3771 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003772 pgrescued++;
3773 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003774 }
Hugh Dickins24513262012-01-20 14:34:21 -08003775
3776 if (zone) {
3777 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3778 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3779 spin_unlock_irq(&zone->lru_lock);
3780 }
Hugh Dickins850465792012-01-20 14:34:19 -08003781}
3782#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003783
Johannes Weiner264e56d2011-10-31 17:09:13 -07003784static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003785{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003786 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003787 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003788 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003789 "one, please send an email to linux-mm@kvack.org.\n",
3790 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003791}
3792
3793/*
3794 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3795 * all nodes' unevictable lists for evictable pages
3796 */
3797unsigned long scan_unevictable_pages;
3798
3799int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003800 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003801 size_t *length, loff_t *ppos)
3802{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003803 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003804 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003805 scan_unevictable_pages = 0;
3806 return 0;
3807}
3808
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003809#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003810/*
3811 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3812 * a specified node's per zone unevictable lists for evictable pages.
3813 */
3814
Kay Sievers10fbcf42011-12-21 14:48:43 -08003815static ssize_t read_scan_unevictable_node(struct device *dev,
3816 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003817 char *buf)
3818{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003819 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003820 return sprintf(buf, "0\n"); /* always zero; should fit... */
3821}
3822
Kay Sievers10fbcf42011-12-21 14:48:43 -08003823static ssize_t write_scan_unevictable_node(struct device *dev,
3824 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003825 const char *buf, size_t count)
3826{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003827 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003828 return 1;
3829}
3830
3831
Kay Sievers10fbcf42011-12-21 14:48:43 -08003832static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003833 read_scan_unevictable_node,
3834 write_scan_unevictable_node);
3835
3836int scan_unevictable_register_node(struct node *node)
3837{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003838 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003839}
3840
3841void scan_unevictable_unregister_node(struct node *node)
3842{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003843 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003844}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003845#endif