blob: 7658fd6536dd241c78bca77415b49994339b8516 [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
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/backing-dev.h>
30#include <linux/rmap.h>
31#include <linux/topology.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080034#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070043#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070044#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
46#include <asm/tlbflush.h>
47#include <asm/div64.h>
48
49#include <linux/swapops.h>
50
Nick Piggin0f8053a2006-03-22 00:08:33 -080051#include "internal.h"
52
Mel Gorman33906bc2010-08-09 17:19:16 -070053#define CREATE_TRACE_POINTS
54#include <trace/events/vmscan.h>
55
Mel Gormanee64fc92011-01-13 15:45:55 -080056/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080057 * reclaim_mode determines how the inactive list is shrunk
58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59 * RECLAIM_MODE_ASYNC: Do not block
60 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
61 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080062 * page from the LRU and reclaim all pages within a
63 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080064 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080065 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080066 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080067typedef unsigned __bitwise__ reclaim_mode_t;
68#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
69#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
70#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
71#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
72#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070073
Linus Torvalds1da177e2005-04-16 15:20:36 -070074struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070075 /* Incremented by the number of inactive pages that were scanned */
76 unsigned long nr_scanned;
77
Rik van Riela79311c2009-01-06 14:40:01 -080078 /* Number of pages freed so far during a call to shrink_zones() */
79 unsigned long nr_reclaimed;
80
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080081 /* How many pages shrink_list() should reclaim */
82 unsigned long nr_to_reclaim;
83
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080084 unsigned long hibernation_mode;
85
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040087 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89 int may_writepage;
90
Johannes Weinera6dc60f82009-03-31 15:19:30 -070091 /* Can mapped pages be reclaimed? */
92 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080093
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070094 /* Can pages be swapped as part of reclaim? */
95 int may_swap;
96
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070097 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080098
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070099 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530100 * Intend to reclaim enough continuous memory rather than reclaim
101 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700102 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800103 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700104
Johannes Weinerf16015f2012-01-12 17:17:52 -0800105 /*
106 * The memory cgroup that hit its limit and as a result is the
107 * primary target of this reclaim invocation.
108 */
109 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -0800110
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700111 /*
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 * are scanned.
114 */
115 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
Johannes Weinerf16015f2012-01-12 17:17:52 -0800118struct mem_cgroup_zone {
119 struct mem_cgroup *mem_cgroup;
120 struct zone *zone;
121};
122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
124
125#ifdef ARCH_HAS_PREFETCH
126#define prefetch_prev_lru_page(_page, _base, _field) \
127 do { \
128 if ((_page)->lru.prev != _base) { \
129 struct page *prev; \
130 \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetch(&prev->_field); \
133 } \
134 } while (0)
135#else
136#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
137#endif
138
139#ifdef ARCH_HAS_PREFETCHW
140#define prefetchw_prev_lru_page(_page, _base, _field) \
141 do { \
142 if ((_page)->lru.prev != _base) { \
143 struct page *prev; \
144 \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetchw(&prev->_field); \
147 } \
148 } while (0)
149#else
150#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
151#endif
152
153/*
154 * From 0 .. 100. Higher means more swappy.
155 */
156int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700157long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158
159static LIST_HEAD(shrinker_list);
160static DECLARE_RWSEM(shrinker_rwsem);
161
Balbir Singh00f0b822008-03-04 14:28:39 -0800162#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800163static bool global_reclaim(struct scan_control *sc)
164{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800165 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800166}
167
Johannes Weinerf16015f2012-01-12 17:17:52 -0800168static bool scanning_global_lru(struct mem_cgroup_zone *mz)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800169{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800170 return !mz->mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800171}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800172#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800173static bool global_reclaim(struct scan_control *sc)
174{
175 return true;
176}
177
Johannes Weinerf16015f2012-01-12 17:17:52 -0800178static bool scanning_global_lru(struct mem_cgroup_zone *mz)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800179{
180 return true;
181}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800182#endif
183
Johannes Weinerf16015f2012-01-12 17:17:52 -0800184static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800185{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800186 if (!scanning_global_lru(mz))
187 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800188
Johannes Weinerf16015f2012-01-12 17:17:52 -0800189 return &mz->zone->reclaim_stat;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800190}
191
Johannes Weinerf16015f2012-01-12 17:17:52 -0800192static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
193 enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800194{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800195 if (!scanning_global_lru(mz))
196 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
197 zone_to_nid(mz->zone),
198 zone_idx(mz->zone),
199 BIT(lru));
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800200
Johannes Weinerf16015f2012-01-12 17:17:52 -0800201 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800202}
203
204
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205/*
206 * Add a shrinker callback to be called from the vm
207 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700208void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209{
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800210 atomic_long_set(&shrinker->nr_in_batch, 0);
Rusty Russell8e1f9362007-07-17 04:03:17 -0700211 down_write(&shrinker_rwsem);
212 list_add_tail(&shrinker->list, &shrinker_list);
213 up_write(&shrinker_rwsem);
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);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700226EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
Ying Han1495f232011-05-24 17:12:27 -0700228static inline int do_shrinker_shrink(struct shrinker *shrinker,
229 struct shrink_control *sc,
230 unsigned long nr_to_scan)
231{
232 sc->nr_to_scan = nr_to_scan;
233 return (*shrinker->shrink)(shrinker, sc);
234}
235
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236#define SHRINK_BATCH 128
237/*
238 * Call the shrink functions to age shrinkable caches
239 *
240 * Here we assume it costs one seek to replace a lru page and that it also
241 * takes a seek to recreate a cache object. With this in mind we age equal
242 * percentages of the lru and ageable caches. This should balance the seeks
243 * generated by these structures.
244 *
Simon Arlott183ff222007-10-20 01:27:18 +0200245 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 * slab to avoid swapping.
247 *
248 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
249 *
250 * `lru_pages' represents the number of on-LRU pages in all the zones which
251 * are eligible for the caller's allocation attempt. It is used for balancing
252 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700253 *
254 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700256unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700257 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700258 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259{
260 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800261 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Ying Han1495f232011-05-24 17:12:27 -0700263 if (nr_pages_scanned == 0)
264 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
Minchan Kimf06590b2011-05-24 17:11:11 -0700266 if (!down_read_trylock(&shrinker_rwsem)) {
267 /* Assume we'll be able to shrink next time */
268 ret = 1;
269 goto out;
270 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271
272 list_for_each_entry(shrinker, &shrinker_list, list) {
273 unsigned long long delta;
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800274 long total_scan;
275 long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000276 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000277 long nr;
278 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000279 long batch_size = shrinker->batch ? shrinker->batch
280 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800282 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
283 if (max_pass <= 0)
284 continue;
285
Dave Chinneracf92b42011-07-08 14:14:35 +1000286 /*
287 * copy the current shrinker scan count into a local variable
288 * and zero it so that other concurrent shrinker invocations
289 * don't also do this scanning work.
290 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800291 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
Dave Chinneracf92b42011-07-08 14:14:35 +1000292
293 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700294 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800295 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000297 total_scan += delta;
298 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700299 printk(KERN_ERR "shrink_slab: %pF negative objects to "
300 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000301 shrinker->shrink, total_scan);
302 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800303 }
304
305 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000306 * We need to avoid excessive windup on filesystem shrinkers
307 * due to large numbers of GFP_NOFS allocations causing the
308 * shrinkers to return -1 all the time. This results in a large
309 * nr being built up so when a shrink that can do some work
310 * comes along it empties the entire cache due to nr >>>
311 * max_pass. This is bad for sustaining a working set in
312 * memory.
313 *
314 * Hence only allow the shrinker to scan the entire cache when
315 * a large delta change is calculated directly.
316 */
317 if (delta < max_pass / 4)
318 total_scan = min(total_scan, max_pass / 2);
319
320 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800321 * Avoid risking looping forever due to too large nr value:
322 * never try to free more than twice the estimate number of
323 * freeable entries.
324 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000325 if (total_scan > max_pass * 2)
326 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327
Dave Chinneracf92b42011-07-08 14:14:35 +1000328 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000329 nr_pages_scanned, lru_pages,
330 max_pass, delta, total_scan);
331
Dave Chinnere9299f52011-07-08 14:14:37 +1000332 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700333 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334
Ying Han1495f232011-05-24 17:12:27 -0700335 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
336 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000337 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 if (shrink_ret == -1)
339 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700340 if (shrink_ret < nr_before)
341 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000342 count_vm_events(SLABS_SCANNED, batch_size);
343 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
345 cond_resched();
346 }
347
Dave Chinneracf92b42011-07-08 14:14:35 +1000348 /*
349 * move the unused scan count back into the shrinker in a
350 * manner that handles concurrent updates. If we exhausted the
351 * scan, there is no need to do an update.
352 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800353 if (total_scan > 0)
354 new_nr = atomic_long_add_return(total_scan,
355 &shrinker->nr_in_batch);
356 else
357 new_nr = atomic_long_read(&shrinker->nr_in_batch);
Dave Chinneracf92b42011-07-08 14:14:35 +1000358
359 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 }
361 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700362out:
363 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700364 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365}
366
Mel Gormanf3a310b2011-01-13 15:46:00 -0800367static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700368 bool sync)
369{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800370 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700371
372 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800373 * Initially assume we are entering either lumpy reclaim or
374 * reclaim/compaction.Depending on the order, we will either set the
375 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700376 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800377 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800378 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800379 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800380 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700381
382 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800383 * Avoid using lumpy reclaim or reclaim/compaction if possible by
384 * restricting when its set to either costly allocations or when
385 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700386 */
387 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800388 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700389 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800390 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700391 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800392 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700393}
394
Mel Gormanf3a310b2011-01-13 15:46:00 -0800395static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700396{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800397 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700398}
399
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400static inline int is_page_cache_freeable(struct page *page)
401{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700402 /*
403 * A freeable page cache page is referenced only by the caller
404 * that isolated the page, the page cache radix tree and
405 * optional buffer heads at page->private.
406 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700407 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408}
409
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700410static int may_write_to_queue(struct backing_dev_info *bdi,
411 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
Christoph Lameter930d9152006-01-08 01:00:47 -0800413 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 return 1;
415 if (!bdi_write_congested(bdi))
416 return 1;
417 if (bdi == current->backing_dev_info)
418 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700419
420 /* lumpy reclaim for hugepage often need a lot of write */
421 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
422 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 return 0;
424}
425
426/*
427 * We detected a synchronous write error writing a page out. Probably
428 * -ENOSPC. We need to propagate that into the address_space for a subsequent
429 * fsync(), msync() or close().
430 *
431 * The tricky part is that after writepage we cannot touch the mapping: nothing
432 * prevents it from being freed up. But we have a ref on the page and once
433 * that page is locked, the mapping is pinned.
434 *
435 * We're allowed to run sleeping lock_page() here because we know the caller has
436 * __GFP_FS.
437 */
438static void handle_write_error(struct address_space *mapping,
439 struct page *page, int error)
440{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100441 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700442 if (page_mapping(page) == mapping)
443 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 unlock_page(page);
445}
446
Christoph Lameter04e62a22006-06-23 02:03:38 -0700447/* possible outcome of pageout() */
448typedef enum {
449 /* failed to write page out, page is locked */
450 PAGE_KEEP,
451 /* move page to the active list, page is locked */
452 PAGE_ACTIVATE,
453 /* page has been sent to the disk successfully, page is unlocked */
454 PAGE_SUCCESS,
455 /* page is clean and locked */
456 PAGE_CLEAN,
457} pageout_t;
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/*
Andrew Morton1742f192006-03-22 00:08:21 -0800460 * pageout is called by shrink_page_list() for each dirty page.
461 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700463static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700464 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465{
466 /*
467 * If the page is dirty, only perform writeback if that write
468 * will be non-blocking. To prevent this allocation from being
469 * stalled by pagecache activity. But note that there may be
470 * stalls if we need to run get_block(). We could test
471 * PagePrivate for that.
472 *
Vincent Li6aceb532009-12-14 17:58:49 -0800473 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 * this page's queue, we can perform writeback even if that
475 * will block.
476 *
477 * If the page is swapcache, write it back even if that would
478 * block, for some throttling. This happens by accident, because
479 * swap_backing_dev_info is bust: it doesn't reflect the
480 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 */
482 if (!is_page_cache_freeable(page))
483 return PAGE_KEEP;
484 if (!mapping) {
485 /*
486 * Some data journaling orphaned pages can have
487 * page->mapping == NULL while being dirty with clean buffers.
488 */
David Howells266cf652009-04-03 16:42:36 +0100489 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 if (try_to_free_buffers(page)) {
491 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700492 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 return PAGE_CLEAN;
494 }
495 }
496 return PAGE_KEEP;
497 }
498 if (mapping->a_ops->writepage == NULL)
499 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700500 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 return PAGE_KEEP;
502
503 if (clear_page_dirty_for_io(page)) {
504 int res;
505 struct writeback_control wbc = {
506 .sync_mode = WB_SYNC_NONE,
507 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700508 .range_start = 0,
509 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 .for_reclaim = 1,
511 };
512
513 SetPageReclaim(page);
514 res = mapping->a_ops->writepage(page, &wbc);
515 if (res < 0)
516 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800517 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 ClearPageReclaim(page);
519 return PAGE_ACTIVATE;
520 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 if (!PageWriteback(page)) {
523 /* synchronous write or broken a_ops? */
524 ClearPageReclaim(page);
525 }
Mel Gorman755f0222010-08-09 17:19:18 -0700526 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800527 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700528 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 return PAGE_SUCCESS;
530 }
531
532 return PAGE_CLEAN;
533}
534
Andrew Mortona649fd92006-10-17 00:09:36 -0700535/*
Nick Piggine2867812008-07-25 19:45:30 -0700536 * Same as remove_mapping, but if the page is removed from the mapping, it
537 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700538 */
Nick Piggine2867812008-07-25 19:45:30 -0700539static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800540{
Nick Piggin28e4d962006-09-25 23:31:23 -0700541 BUG_ON(!PageLocked(page));
542 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800543
Nick Piggin19fd6232008-07-25 19:45:32 -0700544 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800545 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700546 * The non racy check for a busy page.
547 *
548 * Must be careful with the order of the tests. When someone has
549 * a ref to the page, it may be possible that they dirty it then
550 * drop the reference. So if PageDirty is tested before page_count
551 * here, then the following race may occur:
552 *
553 * get_user_pages(&page);
554 * [user mapping goes away]
555 * write_to(page);
556 * !PageDirty(page) [good]
557 * SetPageDirty(page);
558 * put_page(page);
559 * !page_count(page) [good, discard it]
560 *
561 * [oops, our write_to data is lost]
562 *
563 * Reversing the order of the tests ensures such a situation cannot
564 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
565 * load is not satisfied before that of page->_count.
566 *
567 * Note that if SetPageDirty is always performed via set_page_dirty,
568 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800569 */
Nick Piggine2867812008-07-25 19:45:30 -0700570 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800571 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700572 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
573 if (unlikely(PageDirty(page))) {
574 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800575 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700576 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800577
578 if (PageSwapCache(page)) {
579 swp_entry_t swap = { .val = page_private(page) };
580 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700581 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700582 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700583 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500584 void (*freepage)(struct page *);
585
586 freepage = mapping->a_ops->freepage;
587
Minchan Kime64a7822011-03-22 16:32:44 -0700588 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700589 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700590 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500591
592 if (freepage != NULL)
593 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800594 }
595
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800596 return 1;
597
598cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700599 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800600 return 0;
601}
602
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603/*
Nick Piggine2867812008-07-25 19:45:30 -0700604 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
605 * someone else has a ref on the page, abort and return 0. If it was
606 * successfully detached, return 1. Assumes the caller has a single ref on
607 * this page.
608 */
609int remove_mapping(struct address_space *mapping, struct page *page)
610{
611 if (__remove_mapping(mapping, page)) {
612 /*
613 * Unfreezing the refcount with 1 rather than 2 effectively
614 * drops the pagecache ref for us without requiring another
615 * atomic operation.
616 */
617 page_unfreeze_refs(page, 1);
618 return 1;
619 }
620 return 0;
621}
622
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700623/**
624 * putback_lru_page - put previously isolated page onto appropriate LRU list
625 * @page: page to be put back to appropriate lru list
626 *
627 * Add previously isolated @page to appropriate LRU list.
628 * Page may still be unevictable for other reasons.
629 *
630 * lru_lock must not be held, interrupts must be enabled.
631 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700632void putback_lru_page(struct page *page)
633{
634 int lru;
635 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700636 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700637
638 VM_BUG_ON(PageLRU(page));
639
640redo:
641 ClearPageUnevictable(page);
642
643 if (page_evictable(page, NULL)) {
644 /*
645 * For evictable pages, we can use the cache.
646 * In event of a race, worst case is we end up with an
647 * unevictable page on [in]active list.
648 * We know how to handle that.
649 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700650 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700651 lru_cache_add_lru(page, lru);
652 } else {
653 /*
654 * Put unevictable pages directly on zone's unevictable
655 * list.
656 */
657 lru = LRU_UNEVICTABLE;
658 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700659 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700660 * When racing with an mlock or AS_UNEVICTABLE clearing
661 * (page is unlocked) make sure that if the other thread
662 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800663 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700664 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700665 * the page back to the evictable list.
666 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700667 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700668 */
669 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700670 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700671
672 /*
673 * page's status can change while we move it among lru. If an evictable
674 * page is on unevictable list, it never be freed. To avoid that,
675 * check after we added it to the list, again.
676 */
677 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
678 if (!isolate_lru_page(page)) {
679 put_page(page);
680 goto redo;
681 }
682 /* This means someone else dropped this page from LRU
683 * So, it will be freed or putback to LRU again. There is
684 * nothing to do here.
685 */
686 }
687
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700688 if (was_unevictable && lru != LRU_UNEVICTABLE)
689 count_vm_event(UNEVICTABLE_PGRESCUED);
690 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
691 count_vm_event(UNEVICTABLE_PGCULLED);
692
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700693 put_page(page); /* drop ref from isolate */
694}
695
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800696enum page_references {
697 PAGEREF_RECLAIM,
698 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800699 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800700 PAGEREF_ACTIVATE,
701};
702
703static enum page_references page_check_references(struct page *page,
Johannes Weinerf16015f2012-01-12 17:17:52 -0800704 struct mem_cgroup_zone *mz,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800705 struct scan_control *sc)
706{
Johannes Weiner645747462010-03-05 13:42:22 -0800707 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800708 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800709
Johannes Weinerf16015f2012-01-12 17:17:52 -0800710 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800711 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800712
713 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800714 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800715 return PAGEREF_RECLAIM;
716
717 /*
718 * Mlock lost the isolation race with us. Let try_to_unmap()
719 * move the page to the unevictable list.
720 */
721 if (vm_flags & VM_LOCKED)
722 return PAGEREF_RECLAIM;
723
Johannes Weiner645747462010-03-05 13:42:22 -0800724 if (referenced_ptes) {
725 if (PageAnon(page))
726 return PAGEREF_ACTIVATE;
727 /*
728 * All mapped pages start out with page table
729 * references from the instantiating fault, so we need
730 * to look twice if a mapped file page is used more
731 * than once.
732 *
733 * Mark it and spare it for another trip around the
734 * inactive list. Another page table reference will
735 * lead to its activation.
736 *
737 * Note: the mark is set for activated pages as well
738 * so that recently deactivated but used pages are
739 * quickly recovered.
740 */
741 SetPageReferenced(page);
742
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800743 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800744 return PAGEREF_ACTIVATE;
745
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800746 /*
747 * Activate file-backed executable pages after first usage.
748 */
749 if (vm_flags & VM_EXEC)
750 return PAGEREF_ACTIVATE;
751
Johannes Weiner645747462010-03-05 13:42:22 -0800752 return PAGEREF_KEEP;
753 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800754
755 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700756 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800757 return PAGEREF_RECLAIM_CLEAN;
758
759 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800760}
761
Nick Piggine2867812008-07-25 19:45:30 -0700762/*
Andrew Morton1742f192006-03-22 00:08:21 -0800763 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 */
Andrew Morton1742f192006-03-22 00:08:21 -0800765static unsigned long shrink_page_list(struct list_head *page_list,
Johannes Weinerf16015f2012-01-12 17:17:52 -0800766 struct mem_cgroup_zone *mz,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700767 struct scan_control *sc,
Mel Gorman92df3a72011-10-31 17:07:56 -0700768 int priority,
769 unsigned long *ret_nr_dirty,
770 unsigned long *ret_nr_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771{
772 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700773 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700775 unsigned long nr_dirty = 0;
776 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800777 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700778 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
780 cond_resched();
781
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800783 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 struct address_space *mapping;
785 struct page *page;
786 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787
788 cond_resched();
789
790 page = lru_to_page(page_list);
791 list_del(&page->lru);
792
Nick Piggin529ae9a2008-08-02 12:01:03 +0200793 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 goto keep;
795
Nick Piggin725d7042006-09-25 23:30:55 -0700796 VM_BUG_ON(PageActive(page));
Johannes Weinerf16015f2012-01-12 17:17:52 -0800797 VM_BUG_ON(page_zone(page) != mz->zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
799 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800800
Nick Pigginb291f002008-10-18 20:26:44 -0700801 if (unlikely(!page_evictable(page, NULL)))
802 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700803
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700804 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800805 goto keep_locked;
806
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 /* Double the slab pressure for mapped and swapcache pages */
808 if (page_mapped(page) || PageSwapCache(page))
809 sc->nr_scanned++;
810
Andy Whitcroftc661b072007-08-22 14:01:26 -0700811 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
812 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
813
814 if (PageWriteback(page)) {
Mel Gorman92df3a72011-10-31 17:07:56 -0700815 nr_writeback++;
Andy Whitcroftc661b072007-08-22 14:01:26 -0700816 /*
Mel Gormana18bba02011-10-31 17:07:42 -0700817 * Synchronous reclaim cannot queue pages for
818 * writeback due to the possibility of stack overflow
819 * but if it encounters a page under writeback, wait
820 * for the IO to complete.
Andy Whitcroftc661b072007-08-22 14:01:26 -0700821 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800822 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700823 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700824 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700825 else {
826 unlock_page(page);
827 goto keep_lumpy;
828 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700829 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
Johannes Weinerf16015f2012-01-12 17:17:52 -0800831 references = page_check_references(page, mz, sc);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800832 switch (references) {
833 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800835 case PAGEREF_KEEP:
836 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800837 case PAGEREF_RECLAIM:
838 case PAGEREF_RECLAIM_CLEAN:
839 ; /* try to reclaim the page below */
840 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 /*
843 * Anonymous process memory has backing store?
844 * Try to allocate it some swap space here.
845 */
Nick Pigginb291f002008-10-18 20:26:44 -0700846 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800847 if (!(sc->gfp_mask & __GFP_IO))
848 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800849 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800851 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700852 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853
854 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
856 /*
857 * The page is mapped into the page tables of one or more
858 * processes. Try to unmap it here.
859 */
860 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200861 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 case SWAP_FAIL:
863 goto activate_locked;
864 case SWAP_AGAIN:
865 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700866 case SWAP_MLOCK:
867 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 case SWAP_SUCCESS:
869 ; /* try to free the page below */
870 }
871 }
872
873 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700874 nr_dirty++;
875
Mel Gormanee728862011-10-31 17:07:38 -0700876 /*
877 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700878 * avoid risk of stack overflow but do not writeback
879 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700880 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700881 if (page_is_file_cache(page) &&
882 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700883 /*
884 * Immediately reclaim when written back.
885 * Similar in principal to deactivate_page()
886 * except we already have the page isolated
887 * and know it's dirty
888 */
889 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
890 SetPageReclaim(page);
891
Mel Gormanee728862011-10-31 17:07:38 -0700892 goto keep_locked;
893 }
894
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800895 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700897 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800899 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 goto keep_locked;
901
902 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700903 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700905 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 goto keep_locked;
907 case PAGE_ACTIVATE:
908 goto activate_locked;
909 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700910 if (PageWriteback(page))
911 goto keep_lumpy;
912 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700914
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 /*
916 * A synchronous write - probably a ramdisk. Go
917 * ahead and try to reclaim the page.
918 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200919 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 goto keep;
921 if (PageDirty(page) || PageWriteback(page))
922 goto keep_locked;
923 mapping = page_mapping(page);
924 case PAGE_CLEAN:
925 ; /* try to free the page below */
926 }
927 }
928
929 /*
930 * If the page has buffers, try to free the buffer mappings
931 * associated with this page. If we succeed we try to free
932 * the page as well.
933 *
934 * We do this even if the page is PageDirty().
935 * try_to_release_page() does not perform I/O, but it is
936 * possible for a page to have PageDirty set, but it is actually
937 * clean (all its buffers are clean). This happens if the
938 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700939 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 * try_to_release_page() will discover that cleanness and will
941 * drop the buffers and mark the page clean - it can be freed.
942 *
943 * Rarely, pages can have buffers and no ->mapping. These are
944 * the pages which were not successfully invalidated in
945 * truncate_complete_page(). We try to drop those buffers here
946 * and if that worked, and the page is no longer mapped into
947 * process address space (page_count == 1) it can be freed.
948 * Otherwise, leave the page on the LRU so it is swappable.
949 */
David Howells266cf652009-04-03 16:42:36 +0100950 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 if (!try_to_release_page(page, sc->gfp_mask))
952 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700953 if (!mapping && page_count(page) == 1) {
954 unlock_page(page);
955 if (put_page_testzero(page))
956 goto free_it;
957 else {
958 /*
959 * rare race with speculative reference.
960 * the speculative reference will free
961 * this page shortly, so we may
962 * increment nr_reclaimed here (and
963 * leave it off the LRU).
964 */
965 nr_reclaimed++;
966 continue;
967 }
968 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 }
970
Nick Piggine2867812008-07-25 19:45:30 -0700971 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800972 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973
Nick Piggina978d6f2008-10-18 20:26:58 -0700974 /*
975 * At this point, we have no other references and there is
976 * no way to pick any more up (removed from LRU, removed
977 * from pagecache). Can use non-atomic bitops now (and
978 * we obviously don't have to worry about waking up a process
979 * waiting on the page lock, because there are no references.
980 */
981 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700982free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800983 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700984
985 /*
986 * Is there need to periodically free_page_list? It would
987 * appear not as the counts should be low
988 */
989 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 continue;
991
Nick Pigginb291f002008-10-18 20:26:44 -0700992cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800993 if (PageSwapCache(page))
994 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700995 unlock_page(page);
996 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800997 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700998 continue;
999
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001001 /* Not a candidate for swapping, so reclaim swap space. */
1002 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001003 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001004 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 SetPageActive(page);
1006 pgactivate++;
1007keep_locked:
1008 unlock_page(page);
1009keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -08001010 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001011keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001013 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001015
Mel Gorman0e093d992010-10-26 14:21:45 -07001016 /*
1017 * Tag a zone as congested if all the dirty pages encountered were
1018 * backed by a congested BDI. In this case, reclaimers should just
1019 * back off and wait for congestion to clear because further reclaim
1020 * will encounter the same problem
1021 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001022 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
Johannes Weinerf16015f2012-01-12 17:17:52 -08001023 zone_set_flag(mz->zone, ZONE_CONGESTED);
Mel Gorman0e093d992010-10-26 14:21:45 -07001024
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001025 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001026
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001028 count_vm_events(PGACTIVATE, pgactivate);
Mel Gorman92df3a72011-10-31 17:07:56 -07001029 *ret_nr_dirty += nr_dirty;
1030 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -08001031 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032}
1033
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001034/*
1035 * Attempt to remove the specified page from its LRU. Only take this page
1036 * if it is of the appropriate PageActive status. Pages which are being
1037 * freed elsewhere are also ignored.
1038 *
1039 * page: page to consider
1040 * mode: one of the LRU isolation modes defined above
1041 *
1042 * returns 0 on success, -ve errno on failure.
1043 */
Minchan Kim4356f212011-10-31 17:06:47 -07001044int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001045{
Minchan Kim4356f212011-10-31 17:06:47 -07001046 bool all_lru_mode;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001047 int ret = -EINVAL;
1048
1049 /* Only take pages on the LRU. */
1050 if (!PageLRU(page))
1051 return ret;
1052
Minchan Kim4356f212011-10-31 17:06:47 -07001053 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1054 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1055
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001056 /*
1057 * When checking the active state, we need to be sure we are
1058 * dealing with comparible boolean values. Take the logical not
1059 * of each.
1060 */
Minchan Kim4356f212011-10-31 17:06:47 -07001061 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001062 return ret;
1063
Minchan Kim4356f212011-10-31 17:06:47 -07001064 if (!all_lru_mode && !!page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001065 return ret;
1066
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001067 /*
1068 * When this function is being called for lumpy reclaim, we
1069 * initially look into all LRU pages, active, inactive and
1070 * unevictable; only give shrink_page_list evictable pages.
1071 */
1072 if (PageUnevictable(page))
1073 return ret;
1074
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001075 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001076
Mel Gormanc8244932012-01-12 17:19:38 -08001077 /*
1078 * To minimise LRU disruption, the caller can indicate that it only
1079 * wants to isolate pages it will be able to operate on without
1080 * blocking - clean pages for the most part.
1081 *
1082 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1083 * is used by reclaim when it is cannot write to backing storage
1084 *
1085 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1086 * that it is possible to migrate without blocking
1087 */
1088 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1089 /* All the caller can do on PageWriteback is block */
1090 if (PageWriteback(page))
1091 return ret;
1092
1093 if (PageDirty(page)) {
1094 struct address_space *mapping;
1095
1096 /* ISOLATE_CLEAN means only clean pages */
1097 if (mode & ISOLATE_CLEAN)
1098 return ret;
1099
1100 /*
1101 * Only pages without mappings or that have a
1102 * ->migratepage callback are possible to migrate
1103 * without blocking
1104 */
1105 mapping = page_mapping(page);
1106 if (mapping && !mapping->a_ops->migratepage)
1107 return ret;
1108 }
1109 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001110
Minchan Kimf80c0672011-10-31 17:06:55 -07001111 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1112 return ret;
1113
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001114 if (likely(get_page_unless_zero(page))) {
1115 /*
1116 * Be careful not to clear PageLRU until after we're
1117 * sure the page is not being freed elsewhere -- the
1118 * page release code relies on it.
1119 */
1120 ClearPageLRU(page);
1121 ret = 0;
1122 }
1123
1124 return ret;
1125}
1126
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001127/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 * zone->lru_lock is heavily contended. Some of the functions that
1129 * shrink the lists perform better by taking out a batch of pages
1130 * and working on them outside the LRU lock.
1131 *
1132 * For pagecache intensive workloads, this function is the hottest
1133 * spot in the kernel (apart from copy_*_user functions).
1134 *
1135 * Appropriate locks must be held before calling this function.
1136 *
1137 * @nr_to_scan: The number of pages to look through on the list.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001138 * @mz: The mem_cgroup_zone to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001140 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001141 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001142 * @mode: One of the LRU isolation modes
Hugh Dickinsf6260122012-01-12 17:20:06 -08001143 * @active: True [1] if isolating active pages
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001144 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 *
1146 * returns how many pages were moved onto *@dst.
1147 */
Andrew Morton69e05942006-03-22 00:08:19 -08001148static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Hugh Dickinsf6260122012-01-12 17:20:06 -08001149 struct mem_cgroup_zone *mz, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001150 unsigned long *nr_scanned, struct scan_control *sc,
1151 isolate_mode_t mode, int active, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152{
Hugh Dickinsf6260122012-01-12 17:20:06 -08001153 struct lruvec *lruvec;
1154 struct list_head *src;
Andrew Morton69e05942006-03-22 00:08:19 -08001155 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001156 unsigned long nr_lumpy_taken = 0;
1157 unsigned long nr_lumpy_dirty = 0;
1158 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001159 unsigned long scan;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001160 int lru = LRU_BASE;
1161
1162 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1163 if (active)
1164 lru += LRU_ACTIVE;
1165 if (file)
1166 lru += LRU_FILE;
1167 src = &lruvec->lists[lru];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001169 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001170 struct page *page;
1171 unsigned long pfn;
1172 unsigned long end_pfn;
1173 unsigned long page_pfn;
1174 int zone_id;
1175
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 page = lru_to_page(src);
1177 prefetchw_prev_lru_page(page, src, flags);
1178
Nick Piggin725d7042006-09-25 23:30:55 -07001179 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001180
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001181 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001182 case 0:
Johannes Weiner925b7672012-01-12 17:18:15 -08001183 mem_cgroup_lru_del(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001184 list_move(&page->lru, dst);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001185 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001186 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001187
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001188 case -EBUSY:
1189 /* else it is being freed elsewhere */
1190 list_move(&page->lru, src);
1191 continue;
1192
1193 default:
1194 BUG();
1195 }
1196
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001197 if (!sc->order || !(sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM))
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001198 continue;
1199
1200 /*
1201 * Attempt to take all pages in the order aligned region
1202 * surrounding the tag page. Only take those pages of
1203 * the same active state as that tag page. We may safely
1204 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001205 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001206 * where that page is in a different zone we will detect
1207 * it from its zone id and abort this block scan.
1208 */
1209 zone_id = page_zone_id(page);
1210 page_pfn = page_to_pfn(page);
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001211 pfn = page_pfn & ~((1 << sc->order) - 1);
1212 end_pfn = pfn + (1 << sc->order);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001213 for (; pfn < end_pfn; pfn++) {
1214 struct page *cursor_page;
1215
1216 /* The target page is in the block, ignore it. */
1217 if (unlikely(pfn == page_pfn))
1218 continue;
1219
1220 /* Avoid holes within the zone. */
1221 if (unlikely(!pfn_valid_within(pfn)))
1222 break;
1223
1224 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001225
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001226 /* Check that we have not crossed a zone boundary. */
1227 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001228 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001229
1230 /*
1231 * If we don't have enough swap space, reclaiming of
1232 * anon page which don't already have a swap slot is
1233 * pointless.
1234 */
Hugh Dickins043bcbe2012-01-10 15:08:33 -08001235 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001236 !PageSwapCache(cursor_page))
1237 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001238
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001239 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andrea Arcangeli50134732012-01-12 17:19:29 -08001240 unsigned int isolated_pages;
1241
Johannes Weiner925b7672012-01-12 17:18:15 -08001242 mem_cgroup_lru_del(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001243 list_move(&cursor_page->lru, dst);
Andrea Arcangeli50134732012-01-12 17:19:29 -08001244 isolated_pages = hpage_nr_pages(cursor_page);
1245 nr_taken += isolated_pages;
1246 nr_lumpy_taken += isolated_pages;
Mel Gormana8a94d12010-08-09 17:19:17 -07001247 if (PageDirty(cursor_page))
Andrea Arcangeli50134732012-01-12 17:19:29 -08001248 nr_lumpy_dirty += isolated_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001249 scan++;
Andrea Arcangeli50134732012-01-12 17:19:29 -08001250 pfn += isolated_pages - 1;
Mel Gormana8a94d12010-08-09 17:19:17 -07001251 } else {
Andrea Arcangelid179e842011-06-15 15:08:51 -07001252 /*
1253 * Check if the page is freed already.
1254 *
1255 * We can't use page_count() as that
1256 * requires compound_head and we don't
1257 * have a pin on the page here. If a
1258 * page is tail, we may or may not
1259 * have isolated the head, so assume
1260 * it's not free, it'd be tricky to
1261 * track the head status without a
1262 * page pin.
1263 */
1264 if (!PageTail(cursor_page) &&
1265 !atomic_read(&cursor_page->_count))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001266 continue;
1267 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001268 }
1269 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001270
1271 /* If we break out of the loop above, lumpy reclaim failed */
1272 if (pfn < end_pfn)
1273 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 }
1275
Hugh Dickinsf6260122012-01-12 17:20:06 -08001276 *nr_scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001277
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001278 trace_mm_vmscan_lru_isolate(sc->order,
Mel Gormana8a94d12010-08-09 17:19:17 -07001279 nr_to_scan, scan,
1280 nr_taken,
1281 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
Tao Maea4d3492012-01-12 17:19:20 -08001282 mode, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 return nr_taken;
1284}
1285
Nick Piggin62695a82008-10-18 20:26:09 -07001286/**
1287 * isolate_lru_page - tries to isolate a page from its LRU list
1288 * @page: page to isolate from its LRU list
1289 *
1290 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1291 * vmstat statistic corresponding to whatever LRU list the page was on.
1292 *
1293 * Returns 0 if the page was removed from an LRU list.
1294 * Returns -EBUSY if the page was not on an LRU list.
1295 *
1296 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001297 * the active list, it will have PageActive set. If it was found on
1298 * the unevictable list, it will have the PageUnevictable bit set. That flag
1299 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001300 *
1301 * The vmstat statistic corresponding to the list on which the page was
1302 * found will be decremented.
1303 *
1304 * Restrictions:
1305 * (1) Must be called with an elevated refcount on the page. This is a
1306 * fundamentnal difference from isolate_lru_pages (which is called
1307 * without a stable reference).
1308 * (2) the lru_lock must not be held.
1309 * (3) interrupts must be enabled.
1310 */
1311int isolate_lru_page(struct page *page)
1312{
1313 int ret = -EBUSY;
1314
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001315 VM_BUG_ON(!page_count(page));
1316
Nick Piggin62695a82008-10-18 20:26:09 -07001317 if (PageLRU(page)) {
1318 struct zone *zone = page_zone(page);
1319
1320 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001321 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001322 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001323 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001324 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001325 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001326
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001327 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001328 }
1329 spin_unlock_irq(&zone->lru_lock);
1330 }
1331 return ret;
1332}
1333
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001334/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001335 * Are there way too many processes in the direct reclaim path already?
1336 */
1337static int too_many_isolated(struct zone *zone, int file,
1338 struct scan_control *sc)
1339{
1340 unsigned long inactive, isolated;
1341
1342 if (current_is_kswapd())
1343 return 0;
1344
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001345 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001346 return 0;
1347
1348 if (file) {
1349 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1350 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1351 } else {
1352 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1353 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1354 }
1355
1356 return isolated > inactive;
1357}
1358
Mel Gorman66635622010-08-09 17:19:30 -07001359static noinline_for_stack void
Hugh Dickins3f797682012-01-12 17:20:07 -08001360putback_inactive_pages(struct mem_cgroup_zone *mz,
1361 struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001362{
Johannes Weinerf16015f2012-01-12 17:17:52 -08001363 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
Hugh Dickins3f797682012-01-12 17:20:07 -08001364 struct zone *zone = mz->zone;
1365 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001366
Mel Gorman66635622010-08-09 17:19:30 -07001367 /*
1368 * Put back any unfreeable pages.
1369 */
Mel Gorman66635622010-08-09 17:19:30 -07001370 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001371 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001372 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001373
Mel Gorman66635622010-08-09 17:19:30 -07001374 VM_BUG_ON(PageLRU(page));
1375 list_del(&page->lru);
1376 if (unlikely(!page_evictable(page, NULL))) {
1377 spin_unlock_irq(&zone->lru_lock);
1378 putback_lru_page(page);
1379 spin_lock_irq(&zone->lru_lock);
1380 continue;
1381 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001382 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001383 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001384 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001385 if (is_active_lru(lru)) {
1386 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001387 int numpages = hpage_nr_pages(page);
1388 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001389 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001390 if (put_page_testzero(page)) {
1391 __ClearPageLRU(page);
1392 __ClearPageActive(page);
1393 del_page_from_lru_list(zone, page, lru);
1394
1395 if (unlikely(PageCompound(page))) {
1396 spin_unlock_irq(&zone->lru_lock);
1397 (*get_compound_page_dtor(page))(page);
1398 spin_lock_irq(&zone->lru_lock);
1399 } else
1400 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001401 }
1402 }
Mel Gorman66635622010-08-09 17:19:30 -07001403
Hugh Dickins3f797682012-01-12 17:20:07 -08001404 /*
1405 * To save our caller's stack, now use input list for pages to free.
1406 */
1407 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001408}
1409
Johannes Weinerf16015f2012-01-12 17:17:52 -08001410static noinline_for_stack void
1411update_isolated_counts(struct mem_cgroup_zone *mz,
Hugh Dickins3f797682012-01-12 17:20:07 -08001412 struct list_head *page_list,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001413 unsigned long *nr_anon,
Hugh Dickins3f797682012-01-12 17:20:07 -08001414 unsigned long *nr_file)
Mel Gorman1489fa12010-08-09 17:19:33 -07001415{
Johannes Weinerf16015f2012-01-12 17:17:52 -08001416 struct zone *zone = mz->zone;
Mel Gorman1489fa12010-08-09 17:19:33 -07001417 unsigned int count[NR_LRU_LISTS] = { 0, };
Hugh Dickins3f797682012-01-12 17:20:07 -08001418 unsigned long nr_active = 0;
1419 struct page *page;
1420 int lru;
Mel Gorman1489fa12010-08-09 17:19:33 -07001421
Hugh Dickins3f797682012-01-12 17:20:07 -08001422 /*
1423 * Count pages and clear active flags
1424 */
1425 list_for_each_entry(page, page_list, lru) {
1426 int numpages = hpage_nr_pages(page);
1427 lru = page_lru_base_type(page);
1428 if (PageActive(page)) {
1429 lru += LRU_ACTIVE;
1430 ClearPageActive(page);
1431 nr_active += numpages;
1432 }
1433 count[lru] += numpages;
1434 }
1435
Hillf Dantond563c052012-03-21 16:34:02 -07001436 preempt_disable();
Mel Gorman1489fa12010-08-09 17:19:33 -07001437 __count_vm_events(PGDEACTIVATE, nr_active);
1438
1439 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1440 -count[LRU_ACTIVE_FILE]);
1441 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1442 -count[LRU_INACTIVE_FILE]);
1443 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1444 -count[LRU_ACTIVE_ANON]);
1445 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1446 -count[LRU_INACTIVE_ANON]);
1447
1448 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1449 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
Mel Gorman1489fa12010-08-09 17:19:33 -07001450
Hillf Dantond563c052012-03-21 16:34:02 -07001451 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1452 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1453 preempt_enable();
Mel Gorman1489fa12010-08-09 17:19:33 -07001454}
1455
Mel Gorman66635622010-08-09 17:19:30 -07001456/*
Mel Gormana18bba02011-10-31 17:07:42 -07001457 * Returns true if a direct reclaim should wait on pages under writeback.
Wu Fengguange31f36982010-08-09 17:20:01 -07001458 *
1459 * If we are direct reclaiming for contiguous pages and we do not reclaim
1460 * everything in the list, try again and wait for writeback IO to complete.
1461 * This will stall high-order allocations noticeably. Only do that when really
1462 * need to free the pages under high memory pressure.
1463 */
1464static inline bool should_reclaim_stall(unsigned long nr_taken,
1465 unsigned long nr_freed,
1466 int priority,
1467 struct scan_control *sc)
1468{
1469 int lumpy_stall_priority;
1470
1471 /* kswapd should not stall on sync IO */
1472 if (current_is_kswapd())
1473 return false;
1474
1475 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001476 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001477 return false;
1478
Justin P. Mattock81d66c72011-08-23 09:28:02 -07001479 /* If we have reclaimed everything on the isolated list, no stall */
Wu Fengguange31f36982010-08-09 17:20:01 -07001480 if (nr_freed == nr_taken)
1481 return false;
1482
1483 /*
1484 * For high-order allocations, there are two stall thresholds.
1485 * High-cost allocations stall immediately where as lower
1486 * order allocations such as stacks require the scanning
1487 * priority to be much higher before stalling.
1488 */
1489 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1490 lumpy_stall_priority = DEF_PRIORITY;
1491 else
1492 lumpy_stall_priority = DEF_PRIORITY / 3;
1493
1494 return priority <= lumpy_stall_priority;
1495}
1496
1497/*
Andrew Morton1742f192006-03-22 00:08:21 -08001498 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1499 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 */
Mel Gorman66635622010-08-09 17:19:30 -07001501static noinline_for_stack unsigned long
Johannes Weinerf16015f2012-01-12 17:17:52 -08001502shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1503 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504{
1505 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001506 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001507 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001508 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001509 unsigned long nr_anon;
1510 unsigned long nr_file;
Mel Gorman92df3a72011-10-31 17:07:56 -07001511 unsigned long nr_dirty = 0;
1512 unsigned long nr_writeback = 0;
Hillf Danton61317282012-03-21 16:33:48 -07001513 isolate_mode_t isolate_mode = ISOLATE_INACTIVE;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001514 struct zone *zone = mz->zone;
Hillf Dantond563c052012-03-21 16:34:02 -07001515 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001516
Rik van Riel35cd7812009-09-21 17:01:38 -07001517 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001518 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001519
1520 /* We are about to die and free our memory. Return now. */
1521 if (fatal_signal_pending(current))
1522 return SWAP_CLUSTER_MAX;
1523 }
1524
Mel Gormanf3a310b2011-01-13 15:46:00 -08001525 set_reclaim_mode(priority, sc, false);
Minchan Kim4356f212011-10-31 17:06:47 -07001526 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Hillf Danton61317282012-03-21 16:33:48 -07001527 isolate_mode |= ISOLATE_ACTIVE;
Minchan Kim4356f212011-10-31 17:06:47 -07001528
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001530
1531 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001532 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001533 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001534 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001535
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001538 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned,
1539 sc, isolate_mode, 0, file);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001540 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001541 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001542 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001543 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1544 nr_scanned);
1545 else
1546 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1547 nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001548 }
Hillf Dantond563c052012-03-21 16:34:02 -07001549 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001550
Hillf Dantond563c052012-03-21 16:34:02 -07001551 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001552 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001553
Hugh Dickins3f797682012-01-12 17:20:07 -08001554 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1555
Johannes Weinerf16015f2012-01-12 17:17:52 -08001556 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
Mel Gorman92df3a72011-10-31 17:07:56 -07001557 &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001558
Wu Fengguange31f36982010-08-09 17:20:01 -07001559 /* Check if we should syncronously wait for writeback */
1560 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001561 set_reclaim_mode(priority, sc, true);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001562 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
Mel Gorman92df3a72011-10-31 17:07:56 -07001563 priority, &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001564 }
1565
Hugh Dickins3f797682012-01-12 17:20:07 -08001566 spin_lock_irq(&zone->lru_lock);
1567
Hillf Dantond563c052012-03-21 16:34:02 -07001568 reclaim_stat->recent_scanned[0] += nr_anon;
1569 reclaim_stat->recent_scanned[1] += nr_file;
1570
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001571 if (current_is_kswapd())
1572 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1573 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1574
Hugh Dickins3f797682012-01-12 17:20:07 -08001575 putback_inactive_pages(mz, &page_list);
1576
1577 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1578 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1579
1580 spin_unlock_irq(&zone->lru_lock);
1581
1582 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001583
Mel Gorman92df3a72011-10-31 17:07:56 -07001584 /*
1585 * If reclaim is isolating dirty pages under writeback, it implies
1586 * that the long-lived page allocation rate is exceeding the page
1587 * laundering rate. Either the global limits are not being effective
1588 * at throttling processes due to the page distribution throughout
1589 * zones or there is heavy usage of a slow backing device. The
1590 * only option is to throttle from reclaim context which is not ideal
1591 * as there is no guarantee the dirtying process is throttled in the
1592 * same way balance_dirty_pages() manages.
1593 *
1594 * This scales the number of dirty pages that must be under writeback
1595 * before throttling depending on priority. It is a simple backoff
1596 * function that has the most effect in the range DEF_PRIORITY to
1597 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1598 * in trouble and reclaim is considered to be in trouble.
1599 *
1600 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1601 * DEF_PRIORITY-1 50% must be PageWriteback
1602 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1603 * ...
1604 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1605 * isolated page is PageWriteback
1606 */
1607 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1608 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1609
Mel Gormane11da5b2010-10-26 14:21:40 -07001610 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1611 zone_idx(zone),
1612 nr_scanned, nr_reclaimed,
1613 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001614 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001615 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616}
1617
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001618/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 * This moves pages from the active list to the inactive list.
1620 *
1621 * We move them the other way if the page is referenced by one or more
1622 * processes, from rmap.
1623 *
1624 * If the pages are mostly unmapped, the processing is fast and it is
1625 * appropriate to hold zone->lru_lock across the whole operation. But if
1626 * the pages are mapped, the processing is slow (page_referenced()) so we
1627 * should drop zone->lru_lock around each page. It's impossible to balance
1628 * this, so instead we remove the pages from the LRU while processing them.
1629 * It is safe to rely on PG_active against the non-LRU pages in here because
1630 * nobody will play with that bit on a non-LRU page.
1631 *
1632 * The downside is that we have to touch page->_count against each page.
1633 * But we had to alter page->flags anyway.
1634 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001635
Wu Fengguang3eb41402009-06-16 15:33:13 -07001636static void move_active_pages_to_lru(struct zone *zone,
1637 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001638 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001639 enum lru_list lru)
1640{
1641 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001642 struct page *page;
1643
Wu Fengguang3eb41402009-06-16 15:33:13 -07001644 while (!list_empty(list)) {
Johannes Weiner925b7672012-01-12 17:18:15 -08001645 struct lruvec *lruvec;
1646
Wu Fengguang3eb41402009-06-16 15:33:13 -07001647 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001648
1649 VM_BUG_ON(PageLRU(page));
1650 SetPageLRU(page);
1651
Johannes Weiner925b7672012-01-12 17:18:15 -08001652 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1653 list_move(&page->lru, &lruvec->lists[lru]);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001654 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001655
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001656 if (put_page_testzero(page)) {
1657 __ClearPageLRU(page);
1658 __ClearPageActive(page);
1659 del_page_from_lru_list(zone, page, lru);
1660
1661 if (unlikely(PageCompound(page))) {
1662 spin_unlock_irq(&zone->lru_lock);
1663 (*get_compound_page_dtor(page))(page);
1664 spin_lock_irq(&zone->lru_lock);
1665 } else
1666 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001667 }
1668 }
1669 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1670 if (!is_active_lru(lru))
1671 __count_vm_events(PGDEACTIVATE, pgmoved);
1672}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001673
Hugh Dickinsf6260122012-01-12 17:20:06 -08001674static void shrink_active_list(unsigned long nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001675 struct mem_cgroup_zone *mz,
1676 struct scan_control *sc,
1677 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001679 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001680 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001681 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001683 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001684 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 struct page *page;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001686 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001687 unsigned long nr_rotated = 0;
Hillf Danton61317282012-03-21 16:33:48 -07001688 isolate_mode_t isolate_mode = ISOLATE_ACTIVE;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001689 struct zone *zone = mz->zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
1691 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001692
Konstantin Khlebnikov1480de02012-03-21 16:34:17 -07001693 reset_reclaim_mode(sc);
1694
Minchan Kimf80c0672011-10-31 17:06:55 -07001695 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001696 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001697 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001698 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001699
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001701
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001702 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc,
Hillf Danton61317282012-03-21 16:33:48 -07001703 isolate_mode, 1, file);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001704 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001705 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001706
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001707 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001708
Hugh Dickinsf6260122012-01-12 17:20:06 -08001709 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001710 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001711 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001712 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001713 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001714 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 spin_unlock_irq(&zone->lru_lock);
1716
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 while (!list_empty(&l_hold)) {
1718 cond_resched();
1719 page = lru_to_page(&l_hold);
1720 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001721
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001722 if (unlikely(!page_evictable(page, NULL))) {
1723 putback_lru_page(page);
1724 continue;
1725 }
1726
Mel Gormancc715d92012-03-21 16:34:00 -07001727 if (unlikely(buffer_heads_over_limit)) {
1728 if (page_has_private(page) && trylock_page(page)) {
1729 if (page_has_private(page))
1730 try_to_release_page(page, 0);
1731 unlock_page(page);
1732 }
1733 }
1734
Johannes Weinerf16015f2012-01-12 17:17:52 -08001735 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001736 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001737 /*
1738 * Identify referenced, file-backed active pages and
1739 * give them one more trip around the active list. So
1740 * that executable code get better chances to stay in
1741 * memory under moderate memory pressure. Anon pages
1742 * are not likely to be evicted by use-once streaming
1743 * IO, plus JVM can create lots of anon VM_EXEC pages,
1744 * so we ignore them here.
1745 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001746 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001747 list_add(&page->lru, &l_active);
1748 continue;
1749 }
1750 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001751
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001752 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 list_add(&page->lru, &l_inactive);
1754 }
1755
Andrew Mortonb5557492009-01-06 14:40:13 -08001756 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001757 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001758 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001759 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001760 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001761 * Count referenced pages from currently used mappings as rotated,
1762 * even though only some of them are actually re-activated. This
1763 * helps balance scan pressure between file and anonymous pages in
1764 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001765 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001766 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001767
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001768 move_active_pages_to_lru(zone, &l_active, &l_hold,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001769 LRU_ACTIVE + file * LRU_FILE);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001770 move_active_pages_to_lru(zone, &l_inactive, &l_hold,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001771 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001772 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001773 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001774
1775 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776}
1777
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001778#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001779static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001780{
1781 unsigned long active, inactive;
1782
1783 active = zone_page_state(zone, NR_ACTIVE_ANON);
1784 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1785
1786 if (inactive * zone->inactive_ratio < active)
1787 return 1;
1788
1789 return 0;
1790}
1791
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001792/**
1793 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1794 * @zone: zone to check
1795 * @sc: scan control of this context
1796 *
1797 * Returns true if the zone does not have enough inactive anon pages,
1798 * meaning some active anon pages need to be deactivated.
1799 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001800static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001801{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001802 /*
1803 * If we don't have swap space, anonymous page deactivation
1804 * is pointless.
1805 */
1806 if (!total_swap_pages)
1807 return 0;
1808
Johannes Weinerf16015f2012-01-12 17:17:52 -08001809 if (!scanning_global_lru(mz))
1810 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1811 mz->zone);
1812
1813 return inactive_anon_is_low_global(mz->zone);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001814}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001815#else
Johannes Weinerf16015f2012-01-12 17:17:52 -08001816static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001817{
1818 return 0;
1819}
1820#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001821
Rik van Riel56e49d22009-06-16 15:32:28 -07001822static int inactive_file_is_low_global(struct zone *zone)
1823{
1824 unsigned long active, inactive;
1825
1826 active = zone_page_state(zone, NR_ACTIVE_FILE);
1827 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1828
1829 return (active > inactive);
1830}
1831
1832/**
1833 * inactive_file_is_low - check if file pages need to be deactivated
Johannes Weinerf16015f2012-01-12 17:17:52 -08001834 * @mz: memory cgroup and zone to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001835 *
1836 * When the system is doing streaming IO, memory pressure here
1837 * ensures that active file pages get deactivated, until more
1838 * than half of the file pages are on the inactive list.
1839 *
1840 * Once we get to that situation, protect the system's working
1841 * set from being evicted by disabling active file page aging.
1842 *
1843 * This uses a different ratio than the anonymous pages, because
1844 * the page cache uses a use-once replacement algorithm.
1845 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001846static int inactive_file_is_low(struct mem_cgroup_zone *mz)
Rik van Riel56e49d22009-06-16 15:32:28 -07001847{
Johannes Weinerf16015f2012-01-12 17:17:52 -08001848 if (!scanning_global_lru(mz))
1849 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1850 mz->zone);
Rik van Riel56e49d22009-06-16 15:32:28 -07001851
Johannes Weinerf16015f2012-01-12 17:17:52 -08001852 return inactive_file_is_low_global(mz->zone);
Rik van Riel56e49d22009-06-16 15:32:28 -07001853}
1854
Johannes Weinerf16015f2012-01-12 17:17:52 -08001855static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
Rik van Rielb39415b2009-12-14 17:59:48 -08001856{
1857 if (file)
Johannes Weinerf16015f2012-01-12 17:17:52 -08001858 return inactive_file_is_low(mz);
Rik van Rielb39415b2009-12-14 17:59:48 -08001859 else
Johannes Weinerf16015f2012-01-12 17:17:52 -08001860 return inactive_anon_is_low(mz);
Rik van Rielb39415b2009-12-14 17:59:48 -08001861}
1862
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001863static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001864 struct mem_cgroup_zone *mz,
1865 struct scan_control *sc, int priority)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001866{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001867 int file = is_file_lru(lru);
1868
Rik van Rielb39415b2009-12-14 17:59:48 -08001869 if (is_active_lru(lru)) {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001870 if (inactive_list_is_low(mz, file))
1871 shrink_active_list(nr_to_scan, mz, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001872 return 0;
1873 }
1874
Johannes Weinerf16015f2012-01-12 17:17:52 -08001875 return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001876}
1877
Johannes Weinerf16015f2012-01-12 17:17:52 -08001878static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1879 struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001880{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001881 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001882 return vm_swappiness;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001883 return mem_cgroup_swappiness(mz->mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001884}
1885
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001887 * Determine how aggressively the anon and file LRU lists should be
1888 * scanned. The relative value of each set of LRU lists is determined
1889 * by looking at the fraction of the pages scanned we did rotate back
1890 * onto the active list instead of evict.
1891 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001892 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001893 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001894static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1895 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001896{
1897 unsigned long anon, file, free;
1898 unsigned long anon_prio, file_prio;
1899 unsigned long ap, fp;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001900 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001901 u64 fraction[2], denominator;
Hugh Dickins41113042012-01-12 17:20:01 -08001902 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001903 int noswap = 0;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001904 bool force_scan = false;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001905
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001906 /*
1907 * If the zone or memcg is small, nr[l] can be 0. This
1908 * results in no scanning on this priority and a potential
1909 * priority drop. Global direct reclaim can go to the next
1910 * zone and tends to have no problems. Global kswapd is for
1911 * zone balancing and it needs to scan a minimum amount. When
1912 * reclaiming for a memcg, a priority drop can cause high
1913 * latencies, so it's better to scan a minimum amount there as
1914 * well.
1915 */
Johannes Weinerb95a2f22012-01-12 17:18:06 -08001916 if (current_is_kswapd() && mz->zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001917 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001918 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001919 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001920
1921 /* If we have no swap space, do not bother scanning anon pages. */
1922 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1923 noswap = 1;
1924 fraction[0] = 0;
1925 fraction[1] = 1;
1926 denominator = 1;
1927 goto out;
1928 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001929
Johannes Weinerf16015f2012-01-12 17:17:52 -08001930 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1931 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1932 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1933 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001934
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001935 if (global_reclaim(sc)) {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001936 free = zone_page_state(mz->zone, NR_FREE_PAGES);
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001937 /* If we have very few page cache pages,
1938 force-scan anon pages. */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001939 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001940 fraction[0] = 1;
1941 fraction[1] = 0;
1942 denominator = 1;
1943 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001944 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001945 }
1946
1947 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001948 * With swappiness at 100, anonymous and file have the same priority.
1949 * This scanning priority is essentially the inverse of IO cost.
1950 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001951 anon_prio = vmscan_swappiness(mz, sc);
1952 file_prio = 200 - vmscan_swappiness(mz, sc);
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001953
1954 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001955 * OK, so we have swap space and a fair amount of page cache
1956 * pages. We use the recently rotated / recently scanned
1957 * ratios to determine how valuable each cache is.
1958 *
1959 * Because workloads change over time (and to avoid overflow)
1960 * we keep these statistics as a floating average, which ends
1961 * up weighing recent references more than old ones.
1962 *
1963 * anon in [0], file in [1]
1964 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001965 spin_lock_irq(&mz->zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001966 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001967 reclaim_stat->recent_scanned[0] /= 2;
1968 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001969 }
1970
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001971 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001972 reclaim_stat->recent_scanned[1] /= 2;
1973 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001974 }
1975
1976 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001977 * The amount of pressure on anon vs file pages is inversely
1978 * proportional to the fraction of recently scanned pages on
1979 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001980 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001981 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1982 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001983
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001984 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1985 fp /= reclaim_stat->recent_rotated[1] + 1;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001986 spin_unlock_irq(&mz->zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001987
Shaohua Li76a33fc2010-05-24 14:32:36 -07001988 fraction[0] = ap;
1989 fraction[1] = fp;
1990 denominator = ap + fp + 1;
1991out:
Hugh Dickins41113042012-01-12 17:20:01 -08001992 for_each_evictable_lru(lru) {
1993 int file = is_file_lru(lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001994 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001995
Hugh Dickins41113042012-01-12 17:20:01 -08001996 scan = zone_nr_lru_pages(mz, lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001997 if (priority || noswap) {
1998 scan >>= priority;
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001999 if (!scan && force_scan)
2000 scan = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002001 scan = div64_u64(scan * fraction[file], denominator);
2002 }
Hugh Dickins41113042012-01-12 17:20:01 -08002003 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002004 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002005}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002006
2007/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08002008 * Reclaim/compaction depends on a number of pages being freed. To avoid
2009 * disruption to the system, a small number of order-0 pages continue to be
2010 * rotated and reclaimed in the normal fashion. However, by the time we get
2011 * back to the allocator and call try_to_compact_zone(), we ensure that
2012 * there are enough free pages for it to be likely successful
2013 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002014static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002015 unsigned long nr_reclaimed,
2016 unsigned long nr_scanned,
2017 struct scan_control *sc)
2018{
2019 unsigned long pages_for_compaction;
2020 unsigned long inactive_lru_pages;
2021
2022 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08002023 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002024 return false;
2025
Mel Gorman28765922011-02-25 14:44:20 -08002026 /* Consider stopping depending on scan and reclaim activity */
2027 if (sc->gfp_mask & __GFP_REPEAT) {
2028 /*
2029 * For __GFP_REPEAT allocations, stop reclaiming if the
2030 * full LRU list has been scanned and we are still failing
2031 * to reclaim pages. This full LRU scan is potentially
2032 * expensive but a __GFP_REPEAT caller really wants to succeed
2033 */
2034 if (!nr_reclaimed && !nr_scanned)
2035 return false;
2036 } else {
2037 /*
2038 * For non-__GFP_REPEAT allocations which can presumably
2039 * fail without consequence, stop if we failed to reclaim
2040 * any pages from the last SWAP_CLUSTER_MAX number of
2041 * pages that were scanned. This will return to the
2042 * caller faster at the risk reclaim/compaction and
2043 * the resulting allocation attempt fails
2044 */
2045 if (!nr_reclaimed)
2046 return false;
2047 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002048
2049 /*
2050 * If we have not reclaimed enough pages for compaction and the
2051 * inactive lists are large enough, continue reclaiming
2052 */
2053 pages_for_compaction = (2UL << sc->order);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002054 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
Minchan Kim86cfd3a2012-01-10 15:08:18 -08002055 if (nr_swap_pages > 0)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002056 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002057 if (sc->nr_reclaimed < pages_for_compaction &&
2058 inactive_lru_pages > pages_for_compaction)
2059 return true;
2060
2061 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002062 switch (compaction_suitable(mz->zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002063 case COMPACT_PARTIAL:
2064 case COMPACT_CONTINUE:
2065 return false;
2066 default:
2067 return true;
2068 }
2069}
2070
2071/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2073 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002074static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2075 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076{
Christoph Lameterb69408e2008-10-18 20:26:14 -07002077 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08002078 unsigned long nr_to_scan;
Hugh Dickins41113042012-01-12 17:20:01 -08002079 enum lru_list lru;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002080 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002081 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Shaohua Li3da367c2011-10-31 17:07:03 -07002082 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083
Mel Gorman3e7d3442011-01-13 15:45:56 -08002084restart:
2085 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002086 nr_scanned = sc->nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002087 get_scan_count(mz, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002088
Shaohua Li3da367c2011-10-31 17:07:03 -07002089 blk_start_plug(&plug);
Rik van Riel556adec2008-10-18 20:26:34 -07002090 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2091 nr[LRU_INACTIVE_FILE]) {
Hugh Dickins41113042012-01-12 17:20:01 -08002092 for_each_evictable_lru(lru) {
2093 if (nr[lru]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08002094 nr_to_scan = min_t(unsigned long,
Hugh Dickins41113042012-01-12 17:20:01 -08002095 nr[lru], SWAP_CLUSTER_MAX);
2096 nr[lru] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097
Hugh Dickins41113042012-01-12 17:20:01 -08002098 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002099 mz, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002100 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 }
Rik van Riela79311c2009-01-06 14:40:01 -08002102 /*
2103 * On large memory systems, scan >> priority can become
2104 * really large. This is fine for the starting priority;
2105 * we want to put equal scanning pressure on each zone.
2106 * However, if the VM has a harder time of freeing pages,
2107 * with multiple processes reclaiming pages, the total
2108 * freeing target can get unreasonably large.
2109 */
Hillf Dantonc38446c2012-03-21 16:33:50 -07002110 if (nr_reclaimed >= nr_to_reclaim)
2111 nr_to_reclaim = 0;
2112 else
2113 nr_to_reclaim -= nr_reclaimed;
2114
2115 if (!nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08002116 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 }
Shaohua Li3da367c2011-10-31 17:07:03 -07002118 blk_finish_plug(&plug);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002119 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002120
Rik van Riel556adec2008-10-18 20:26:34 -07002121 /*
2122 * Even if we did not try to evict anon pages at all, we want to
2123 * rebalance the anon lru active/inactive ratio.
2124 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002125 if (inactive_anon_is_low(mz))
2126 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
Rik van Riel556adec2008-10-18 20:26:34 -07002127
Mel Gorman3e7d3442011-01-13 15:45:56 -08002128 /* reclaim/compaction might need reclaim to continue */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002129 if (should_continue_reclaim(mz, nr_reclaimed,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002130 sc->nr_scanned - nr_scanned, sc))
2131 goto restart;
2132
Andrew Morton232ea4d2007-02-28 20:13:21 -08002133 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134}
2135
Johannes Weinerf16015f2012-01-12 17:17:52 -08002136static void shrink_zone(int priority, struct zone *zone,
2137 struct scan_control *sc)
2138{
Johannes Weiner56600482012-01-12 17:17:59 -08002139 struct mem_cgroup *root = sc->target_mem_cgroup;
2140 struct mem_cgroup_reclaim_cookie reclaim = {
Johannes Weinerf16015f2012-01-12 17:17:52 -08002141 .zone = zone,
Johannes Weiner56600482012-01-12 17:17:59 -08002142 .priority = priority,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002143 };
Johannes Weiner56600482012-01-12 17:17:59 -08002144 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002145
Johannes Weiner56600482012-01-12 17:17:59 -08002146 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2147 do {
2148 struct mem_cgroup_zone mz = {
2149 .mem_cgroup = memcg,
2150 .zone = zone,
2151 };
2152
2153 shrink_mem_cgroup_zone(priority, &mz, sc);
2154 /*
2155 * Limit reclaim has historically picked one memcg and
2156 * scanned it with decreasing priority levels until
2157 * nr_to_reclaim had been reclaimed. This priority
2158 * cycle is thus over after a single memcg.
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002159 *
2160 * Direct reclaim and kswapd, on the other hand, have
2161 * to scan all memory cgroups to fulfill the overall
2162 * scan target for the zone.
Johannes Weiner56600482012-01-12 17:17:59 -08002163 */
2164 if (!global_reclaim(sc)) {
2165 mem_cgroup_iter_break(root, memcg);
2166 break;
2167 }
2168 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2169 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002170}
2171
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002172/* Returns true if compaction should go ahead for a high-order request */
2173static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2174{
2175 unsigned long balance_gap, watermark;
2176 bool watermark_ok;
2177
2178 /* Do not consider compaction for orders reclaim is meant to satisfy */
2179 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2180 return false;
2181
2182 /*
2183 * Compaction takes time to run and there are potentially other
2184 * callers using the pages just freed. Continue reclaiming until
2185 * there is a buffer of free pages available to give compaction
2186 * a reasonable chance of completing and allocating the page
2187 */
2188 balance_gap = min(low_wmark_pages(zone),
2189 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2190 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2191 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2192 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2193
2194 /*
2195 * If compaction is deferred, reclaim up to a point where
2196 * compaction will have a chance of success when re-enabled
2197 */
Rik van Rielaff62242012-03-21 16:33:52 -07002198 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002199 return watermark_ok;
2200
2201 /* If compaction is not ready to start, keep reclaiming */
2202 if (!compaction_suitable(zone, sc->order))
2203 return false;
2204
2205 return watermark_ok;
2206}
2207
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208/*
2209 * This is the direct reclaim path, for page-allocating processes. We only
2210 * try to reclaim pages from zones which will satisfy the caller's allocation
2211 * request.
2212 *
Mel Gorman41858962009-06-16 15:32:12 -07002213 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2214 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2216 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002217 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2218 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2219 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 * If a zone is deemed to be full of pinned pages then just give it a light
2222 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002223 *
2224 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002225 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002226 * the caller that it should consider retrying the allocation instead of
2227 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 */
Mel Gormane0c23272011-10-31 17:09:33 -07002229static bool shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08002230 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231{
Mel Gormandd1a2392008-04-28 02:12:17 -07002232 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002233 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002234 unsigned long nr_soft_reclaimed;
2235 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002236 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002237
Mel Gormancc715d92012-03-21 16:34:00 -07002238 /*
2239 * If the number of buffer_heads in the machine exceeds the maximum
2240 * allowed level, force direct reclaim to scan the highmem zone as
2241 * highmem pages could be pinning lowmem pages storing buffer_heads
2242 */
2243 if (buffer_heads_over_limit)
2244 sc->gfp_mask |= __GFP_HIGHMEM;
2245
Mel Gormand4debc62010-08-09 17:19:29 -07002246 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2247 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002248 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002250 /*
2251 * Take care memory controller reclaiming has small influence
2252 * to global LRU.
2253 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002254 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002255 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2256 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002257 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002258 continue; /* Let kswapd poll it */
Rik van Riele0887c12011-10-31 17:09:31 -07002259 if (COMPACTION_BUILD) {
2260 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002261 * If we already have plenty of memory free for
2262 * compaction in this zone, don't free any more.
2263 * Even though compaction is invoked for any
2264 * non-zero order, only frequent costly order
2265 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002266 * noticeable problem, like transparent huge
2267 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002268 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002269 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002270 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002271 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002272 }
Rik van Riele0887c12011-10-31 17:09:31 -07002273 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002274 /*
2275 * This steals pages from memory cgroups over softlimit
2276 * and returns the number of reclaimed pages and
2277 * scanned pages. This works for global memory pressure
2278 * and balancing, not for a memcg's limit.
2279 */
2280 nr_soft_scanned = 0;
2281 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2282 sc->order, sc->gfp_mask,
2283 &nr_soft_scanned);
2284 sc->nr_reclaimed += nr_soft_reclaimed;
2285 sc->nr_scanned += nr_soft_scanned;
2286 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002287 }
Nick Piggin408d8542006-09-25 23:31:27 -07002288
Rik van Riela79311c2009-01-06 14:40:01 -08002289 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 }
Mel Gormane0c23272011-10-31 17:09:33 -07002291
Mel Gorman0cee34f2012-01-12 17:19:49 -08002292 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002293}
2294
2295static bool zone_reclaimable(struct zone *zone)
2296{
2297 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2298}
2299
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002300/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002301static bool all_unreclaimable(struct zonelist *zonelist,
2302 struct scan_control *sc)
2303{
2304 struct zoneref *z;
2305 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002306
2307 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2308 gfp_zone(sc->gfp_mask), sc->nodemask) {
2309 if (!populated_zone(zone))
2310 continue;
2311 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2312 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002313 if (!zone->all_unreclaimable)
2314 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002315 }
2316
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002317 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002319
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320/*
2321 * This is the main entry point to direct page reclaim.
2322 *
2323 * If a full scan of the inactive list fails to free enough memory then we
2324 * are "out of memory" and something needs to be killed.
2325 *
2326 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2327 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002328 * caller can't do much about. We kick the writeback threads and take explicit
2329 * naps in the hope that some of these pages can be written. But if the
2330 * allocating task holds filesystem locks which prevent writeout this might not
2331 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002332 *
2333 * returns: 0, if no pages reclaimed
2334 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 */
Mel Gormandac1d272008-04-28 02:12:12 -07002336static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002337 struct scan_control *sc,
2338 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339{
2340 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002341 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002343 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002344 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002345 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002346 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347
Keika Kobayashi873b4772008-07-25 01:48:52 -07002348 delayacct_freepages_start();
2349
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002350 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002351 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
2353 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002354 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002355 if (!priority)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002356 disable_swap_token(sc->target_mem_cgroup);
Mel Gorman0cee34f2012-01-12 17:19:49 -08002357 aborted_reclaim = shrink_zones(priority, zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002358
Balbir Singh66e17072008-02-07 00:13:56 -08002359 /*
2360 * Don't shrink slabs when reclaiming memory from
2361 * over limit cgroups
2362 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002363 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002364 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002365 for_each_zone_zonelist(zone, z, zonelist,
2366 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002367 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2368 continue;
2369
2370 lru_pages += zone_reclaimable_pages(zone);
2371 }
2372
Ying Han1495f232011-05-24 17:12:27 -07002373 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002374 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002375 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002376 reclaim_state->reclaimed_slab = 0;
2377 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 }
Balbir Singh66e17072008-02-07 00:13:56 -08002379 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002380 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382
2383 /*
2384 * Try to write back as many pages as we just scanned. This
2385 * tends to cause slow streaming writers to write data to the
2386 * disk smoothly, at the dirtying rate, which is nice. But
2387 * that's undesirable in laptop mode, where we *want* lumpy
2388 * writeout. So in laptop mode, write out the whole world.
2389 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002390 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2391 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002392 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2393 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002394 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 }
2396
2397 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002398 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002399 priority < DEF_PRIORITY - 2) {
2400 struct zone *preferred_zone;
2401
2402 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002403 &cpuset_current_mems_allowed,
2404 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002405 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2406 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002408
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002410 delayacct_freepages_end();
2411
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002412 if (sc->nr_reclaimed)
2413 return sc->nr_reclaimed;
2414
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002415 /*
2416 * As hibernation is going on, kswapd is freezed so that it can't mark
2417 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2418 * check.
2419 */
2420 if (oom_killer_disabled)
2421 return 0;
2422
Mel Gorman0cee34f2012-01-12 17:19:49 -08002423 /* Aborted reclaim to try compaction? don't OOM, then */
2424 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002425 return 1;
2426
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002427 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002428 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002429 return 1;
2430
2431 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432}
2433
Mel Gormandac1d272008-04-28 02:12:12 -07002434unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002435 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002436{
Mel Gorman33906bc2010-08-09 17:19:16 -07002437 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002438 struct scan_control sc = {
2439 .gfp_mask = gfp_mask,
2440 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002441 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002442 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002443 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002444 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002445 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002446 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002447 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002448 struct shrink_control shrink = {
2449 .gfp_mask = sc.gfp_mask,
2450 };
Balbir Singh66e17072008-02-07 00:13:56 -08002451
Mel Gorman33906bc2010-08-09 17:19:16 -07002452 trace_mm_vmscan_direct_reclaim_begin(order,
2453 sc.may_writepage,
2454 gfp_mask);
2455
Ying Hana09ed5e2011-05-24 17:12:26 -07002456 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002457
2458 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2459
2460 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002461}
2462
Balbir Singh00f0b822008-03-04 14:28:39 -08002463#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002464
Johannes Weiner72835c82012-01-12 17:18:32 -08002465unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002466 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002467 struct zone *zone,
2468 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002469{
2470 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002471 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002472 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002473 .may_writepage = !laptop_mode,
2474 .may_unmap = 1,
2475 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002476 .order = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002477 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002478 };
Johannes Weiner56600482012-01-12 17:17:59 -08002479 struct mem_cgroup_zone mz = {
Johannes Weiner72835c82012-01-12 17:18:32 -08002480 .mem_cgroup = memcg,
Johannes Weiner56600482012-01-12 17:17:59 -08002481 .zone = zone,
2482 };
Ying Han0ae5e892011-05-26 16:25:25 -07002483
Balbir Singh4e416952009-09-23 15:56:39 -07002484 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2485 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002486
2487 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2488 sc.may_writepage,
2489 sc.gfp_mask);
2490
Balbir Singh4e416952009-09-23 15:56:39 -07002491 /*
2492 * NOTE: Although we can get the priority field, using it
2493 * here is not a good idea, since it limits the pages we can scan.
2494 * if we don't reclaim here, the shrink_zone from balance_pgdat
2495 * will pick up pages from other mem cgroup's as well. We hack
2496 * the priority and make it zero.
2497 */
Johannes Weiner56600482012-01-12 17:17:59 -08002498 shrink_mem_cgroup_zone(0, &mz, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002499
2500 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2501
Ying Han0ae5e892011-05-26 16:25:25 -07002502 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002503 return sc.nr_reclaimed;
2504}
2505
Johannes Weiner72835c82012-01-12 17:18:32 -08002506unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002507 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002508 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002509{
Balbir Singh4e416952009-09-23 15:56:39 -07002510 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002511 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002512 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002513 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002514 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002515 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002516 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002517 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002518 .order = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002519 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002520 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002521 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2522 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2523 };
2524 struct shrink_control shrink = {
2525 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002526 };
Balbir Singh66e17072008-02-07 00:13:56 -08002527
Ying Han889976d2011-05-26 16:25:33 -07002528 /*
2529 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2530 * take care of from where we get pages. So the node where we start the
2531 * scan does not need to be the current node.
2532 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002533 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002534
2535 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002536
2537 trace_mm_vmscan_memcg_reclaim_begin(0,
2538 sc.may_writepage,
2539 sc.gfp_mask);
2540
Ying Hana09ed5e2011-05-24 17:12:26 -07002541 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002542
2543 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2544
2545 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002546}
2547#endif
2548
Johannes Weinerf16015f2012-01-12 17:17:52 -08002549static void age_active_anon(struct zone *zone, struct scan_control *sc,
2550 int priority)
2551{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002552 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002553
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002554 if (!total_swap_pages)
2555 return;
2556
2557 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2558 do {
2559 struct mem_cgroup_zone mz = {
2560 .mem_cgroup = memcg,
2561 .zone = zone,
2562 };
2563
2564 if (inactive_anon_is_low(&mz))
2565 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2566 sc, priority, 0);
2567
2568 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2569 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002570}
2571
Mel Gorman1741c872011-01-13 15:46:21 -08002572/*
2573 * pgdat_balanced is used when checking if a node is balanced for high-order
2574 * allocations. Only zones that meet watermarks and are in a zone allowed
2575 * by the callers classzone_idx are added to balanced_pages. The total of
2576 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2577 * for the node to be considered balanced. Forcing all zones to be balanced
2578 * for high orders can cause excessive reclaim when there are imbalanced zones.
2579 * The choice of 25% is due to
2580 * o a 16M DMA zone that is balanced will not balance a zone on any
2581 * reasonable sized machine
2582 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002583 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002584 * would need to be at least 256M for it to be balance a whole node.
2585 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2586 * to balance a node on its own. These seemed like reasonable ratios.
2587 */
2588static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2589 int classzone_idx)
2590{
2591 unsigned long present_pages = 0;
2592 int i;
2593
2594 for (i = 0; i <= classzone_idx; i++)
2595 present_pages += pgdat->node_zones[i].present_pages;
2596
Shaohua Li4746efd2011-07-19 08:49:26 -07002597 /* A special case here: if zone has no page, we think it's balanced */
2598 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002599}
2600
Mel Gormanf50de2d2009-12-14 17:58:53 -08002601/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002602static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2603 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002604{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002605 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002606 unsigned long balanced = 0;
2607 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002608
2609 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2610 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002611 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002612
Mel Gorman0abdee22011-01-13 15:46:22 -08002613 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002614 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002615 struct zone *zone = pgdat->node_zones + i;
2616
2617 if (!populated_zone(zone))
2618 continue;
2619
Mel Gorman355b09c2011-01-13 15:46:24 -08002620 /*
2621 * balance_pgdat() skips over all_unreclaimable after
2622 * DEF_PRIORITY. Effectively, it considers them balanced so
2623 * they must be considered balanced here as well if kswapd
2624 * is to sleep
2625 */
2626 if (zone->all_unreclaimable) {
2627 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002628 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002629 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002630
Mel Gorman88f5acf2011-01-13 15:45:41 -08002631 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002632 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002633 all_zones_ok = false;
2634 else
2635 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002636 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002637
Mel Gorman1741c872011-01-13 15:46:21 -08002638 /*
2639 * For high-order requests, the balanced zones must contain at least
2640 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2641 * must be balanced
2642 */
2643 if (order)
Johannes Weinerafc7e3262011-05-24 17:11:09 -07002644 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002645 else
2646 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002647}
2648
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649/*
2650 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002651 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002653 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 *
2655 * There is special handling here for zones which are full of pinned pages.
2656 * This can happen if the pages are all mlocked, or if they are all used by
2657 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2658 * What we do is to detect the case where all pages in the zone have been
2659 * scanned twice and there has been zero successful reclaim. Mark the zone as
2660 * dead and from now on, only perform a short scan. Basically we're polling
2661 * the zone for when the problem goes away.
2662 *
2663 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002664 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2665 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2666 * lower zones regardless of the number of free pages in the lower zones. This
2667 * interoperates with the page allocator fallback scheme to ensure that aging
2668 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 */
Mel Gorman99504742011-01-13 15:46:20 -08002670static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002671 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002674 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 int priority;
2676 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002677 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002678 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002680 unsigned long nr_soft_reclaimed;
2681 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002682 struct scan_control sc = {
2683 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002684 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002685 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002686 /*
2687 * kswapd doesn't want to be bailed out while reclaim. because
2688 * we want to put equal scanning pressure on each zone.
2689 */
2690 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002691 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002692 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002693 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002694 struct shrink_control shrink = {
2695 .gfp_mask = sc.gfp_mask,
2696 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697loop_again:
2698 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002699 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002700 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002701 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002705 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002707 /* The swap token gets in the way of swapout... */
2708 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002709 disable_swap_token(NULL);
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002710
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002712 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002714 /*
2715 * Scan in the highmem->dma direction for the highest
2716 * zone which needs scanning
2717 */
2718 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2719 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002721 if (!populated_zone(zone))
2722 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002724 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002725 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726
Rik van Riel556adec2008-10-18 20:26:34 -07002727 /*
2728 * Do some background aging of the anon list, to give
2729 * pages a chance to be referenced before reclaiming.
2730 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002731 age_active_anon(zone, &sc, priority);
Rik van Riel556adec2008-10-18 20:26:34 -07002732
Mel Gormancc715d92012-03-21 16:34:00 -07002733 /*
2734 * If the number of buffer_heads in the machine
2735 * exceeds the maximum allowed level and this node
2736 * has a highmem zone, force kswapd to reclaim from
2737 * it to relieve lowmem pressure.
2738 */
2739 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2740 end_zone = i;
2741 break;
2742 }
2743
Mel Gorman88f5acf2011-01-13 15:45:41 -08002744 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002745 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002746 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002747 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002748 } else {
2749 /* If balanced, clear the congested flag */
2750 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002753 if (i < 0)
2754 goto out;
2755
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 for (i = 0; i <= end_zone; i++) {
2757 struct zone *zone = pgdat->node_zones + i;
2758
Wu Fengguangadea02a2009-09-21 17:01:42 -07002759 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 }
2761
2762 /*
2763 * Now scan the zone in the dma->highmem direction, stopping
2764 * at the last zone which needs scanning.
2765 *
2766 * We do this because the page allocator works in the opposite
2767 * direction. This prevents the page allocator from allocating
2768 * pages behind kswapd's direction of progress, which would
2769 * cause too much scanning of the lower zones.
2770 */
2771 for (i = 0; i <= end_zone; i++) {
2772 struct zone *zone = pgdat->node_zones + i;
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002773 int nr_slab, testorder;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002774 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775
Con Kolivasf3fe6512006-01-06 00:11:15 -08002776 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 continue;
2778
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002779 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 continue;
2781
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002783
Ying Han0ae5e892011-05-26 16:25:25 -07002784 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002785 /*
2786 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002787 */
Ying Han0ae5e892011-05-26 16:25:25 -07002788 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2789 order, sc.gfp_mask,
2790 &nr_soft_scanned);
2791 sc.nr_reclaimed += nr_soft_reclaimed;
2792 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002793
Rik van Riel32a43302007-10-16 01:24:50 -07002794 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002795 * We put equal pressure on every zone, unless
2796 * one zone has way too many pages free
2797 * already. The "too many pages" is defined
2798 * as the high wmark plus a "gap" where the
2799 * gap is either the low watermark or 1%
2800 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002801 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002802 balance_gap = min(low_wmark_pages(zone),
2803 (zone->present_pages +
2804 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2805 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002806 /*
2807 * Kswapd reclaims only single pages with compaction
2808 * enabled. Trying too hard to reclaim until contiguous
2809 * free pages have become available can hurt performance
2810 * by evicting too much useful data from memory.
2811 * Do not reclaim more than needed for compaction.
2812 */
2813 testorder = order;
2814 if (COMPACTION_BUILD && order &&
2815 compaction_suitable(zone, order) !=
2816 COMPACT_SKIPPED)
2817 testorder = 0;
2818
Mel Gormancc715d92012-03-21 16:34:00 -07002819 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
Hugh Dickins643ac9f2012-03-23 02:57:31 -07002820 !zone_watermark_ok_safe(zone, testorder,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002821 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002822 end_zone, 0)) {
Rik van Riela79311c2009-01-06 14:40:01 -08002823 shrink_zone(priority, zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002824
Mel Gormand7868da2011-07-08 15:39:38 -07002825 reclaim_state->reclaimed_slab = 0;
2826 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2827 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2828 total_scanned += sc.nr_scanned;
2829
2830 if (nr_slab == 0 && !zone_reclaimable(zone))
2831 zone->all_unreclaimable = 1;
2832 }
2833
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 /*
2835 * If we've done a decent amount of scanning and
2836 * the reclaim ratio is low, start doing writepage
2837 * even in laptop mode
2838 */
2839 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002840 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002842
Mel Gorman215ddd62011-07-08 15:39:40 -07002843 if (zone->all_unreclaimable) {
2844 if (end_zone && end_zone == i)
2845 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002846 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002847 }
Mel Gormand7868da2011-07-08 15:39:38 -07002848
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002849 if (!zone_watermark_ok_safe(zone, testorder,
Minchan Kim45973d72010-03-05 13:41:45 -08002850 high_wmark_pages(zone), end_zone, 0)) {
2851 all_zones_ok = 0;
2852 /*
2853 * We are still under min water mark. This
2854 * means that we have a GFP_ATOMIC allocation
2855 * failure risk. Hurry up!
2856 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002857 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002858 min_wmark_pages(zone), end_zone, 0))
2859 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002860 } else {
2861 /*
2862 * If a zone reaches its high watermark,
2863 * consider it to be no longer congested. It's
2864 * possible there are dirty pages backed by
2865 * congested BDIs but as pressure is relieved,
2866 * spectulatively avoid congestion waits
2867 */
2868 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002869 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002870 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002871 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002872
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002874 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 break; /* kswapd: all done */
2876 /*
2877 * OK, kswapd is getting into trouble. Take a nap, then take
2878 * another pass across the zones.
2879 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002880 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2881 if (has_under_min_watermark_zone)
2882 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2883 else
2884 congestion_wait(BLK_RW_ASYNC, HZ/10);
2885 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886
2887 /*
2888 * We do this so kswapd doesn't build up large priorities for
2889 * example when it is freeing in parallel with allocators. It
2890 * matches the direct reclaim path behaviour in terms of impact
2891 * on zone->*_priority.
2892 */
Rik van Riela79311c2009-01-06 14:40:01 -08002893 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 break;
2895 }
2896out:
Mel Gorman99504742011-01-13 15:46:20 -08002897
2898 /*
2899 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002900 * high-order: Balanced zones must make up at least 25% of the node
2901 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002902 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002903 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002905
2906 try_to_freeze();
2907
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002908 /*
2909 * Fragmentation may mean that the system cannot be
2910 * rebalanced for high-order allocations in all zones.
2911 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2912 * it means the zones have been fully scanned and are still
2913 * not balanced. For high-order allocations, there is
2914 * little point trying all over again as kswapd may
2915 * infinite loop.
2916 *
2917 * Instead, recheck all watermarks at order-0 as they
2918 * are the most important. If watermarks are ok, kswapd will go
2919 * back to sleep. High-order users can still perform direct
2920 * reclaim if they wish.
2921 */
2922 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2923 order = sc.order = 0;
2924
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 goto loop_again;
2926 }
2927
Mel Gorman99504742011-01-13 15:46:20 -08002928 /*
2929 * If kswapd was reclaiming at a higher order, it has the option of
2930 * sleeping without all zones being balanced. Before it does, it must
2931 * ensure that the watermarks for order-0 on *all* zones are met and
2932 * that the congestion flags are cleared. The congestion flag must
2933 * be cleared as kswapd is the only mechanism that clears the flag
2934 * and it is potentially going to sleep here.
2935 */
2936 if (order) {
Rik van Riel7be62de2012-03-21 16:33:52 -07002937 int zones_need_compaction = 1;
2938
Mel Gorman99504742011-01-13 15:46:20 -08002939 for (i = 0; i <= end_zone; i++) {
2940 struct zone *zone = pgdat->node_zones + i;
2941
2942 if (!populated_zone(zone))
2943 continue;
2944
2945 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2946 continue;
2947
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002948 /* Would compaction fail due to lack of free memory? */
2949 if (compaction_suitable(zone, order) == COMPACT_SKIPPED)
2950 goto loop_again;
2951
Mel Gorman99504742011-01-13 15:46:20 -08002952 /* Confirm the zone is balanced for order-0 */
2953 if (!zone_watermark_ok(zone, 0,
2954 high_wmark_pages(zone), 0, 0)) {
2955 order = sc.order = 0;
2956 goto loop_again;
2957 }
2958
Rik van Riel7be62de2012-03-21 16:33:52 -07002959 /* Check if the memory needs to be defragmented. */
2960 if (zone_watermark_ok(zone, order,
2961 low_wmark_pages(zone), *classzone_idx, 0))
2962 zones_need_compaction = 0;
2963
Mel Gorman99504742011-01-13 15:46:20 -08002964 /* If balanced, clear the congested flag */
2965 zone_clear_flag(zone, ZONE_CONGESTED);
2966 }
Rik van Riel7be62de2012-03-21 16:33:52 -07002967
2968 if (zones_need_compaction)
2969 compact_pgdat(pgdat, order);
Mel Gorman99504742011-01-13 15:46:20 -08002970 }
2971
Mel Gorman0abdee22011-01-13 15:46:22 -08002972 /*
2973 * Return the order we were reclaiming at so sleeping_prematurely()
2974 * makes a decision on the order we were last reclaiming at. However,
2975 * if another caller entered the allocator slow path while kswapd
2976 * was awake, order will remain at the higher level
2977 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002978 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002979 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980}
2981
Mel Gormandc83edd2011-01-13 15:46:26 -08002982static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002983{
2984 long remaining = 0;
2985 DEFINE_WAIT(wait);
2986
2987 if (freezing(current) || kthread_should_stop())
2988 return;
2989
2990 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2991
2992 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002993 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002994 remaining = schedule_timeout(HZ/10);
2995 finish_wait(&pgdat->kswapd_wait, &wait);
2996 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2997 }
2998
2999 /*
3000 * After a short sleep, check if it was a premature sleep. If not, then
3001 * go fully to sleep until explicitly woken up.
3002 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003003 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003004 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3005
3006 /*
3007 * vmstat counters are not perfectly accurate and the estimated
3008 * value for counters such as NR_FREE_PAGES can deviate from the
3009 * true value by nr_online_cpus * threshold. To avoid the zone
3010 * watermarks being breached while under pressure, we reduce the
3011 * per-cpu vmstat threshold while kswapd is awake and restore
3012 * them before going back to sleep.
3013 */
3014 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3015 schedule();
3016 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3017 } else {
3018 if (remaining)
3019 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3020 else
3021 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3022 }
3023 finish_wait(&pgdat->kswapd_wait, &wait);
3024}
3025
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026/*
3027 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003028 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 *
3030 * This basically trickles out pages so that we have _some_
3031 * free memory available even if there is no other activity
3032 * that frees anything up. This is needed for things like routing
3033 * etc, where we otherwise might have all activity going on in
3034 * asynchronous contexts that cannot page things out.
3035 *
3036 * If there are applications that are active memory-allocators
3037 * (most normal use), this basically shouldn't matter.
3038 */
3039static int kswapd(void *p)
3040{
Mel Gorman215ddd62011-07-08 15:39:40 -07003041 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003042 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003043 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003044 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 pg_data_t *pgdat = (pg_data_t*)p;
3046 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003047
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 struct reclaim_state reclaim_state = {
3049 .reclaimed_slab = 0,
3050 };
Rusty Russella70f7302009-03-13 14:49:46 +10303051 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052
Nick Piggincf40bd12009-01-21 08:12:39 +01003053 lockdep_set_current_reclaim_state(GFP_KERNEL);
3054
Rusty Russell174596a2009-01-01 10:12:29 +10303055 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003056 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 current->reclaim_state = &reclaim_state;
3058
3059 /*
3060 * Tell the memory management that we're a "memory allocator",
3061 * and that if we need more memory we should get access to it
3062 * regardless (see "__alloc_pages()"). "kswapd" should
3063 * never get caught in the normal page freeing logic.
3064 *
3065 * (Kswapd normally doesn't need memory anyway, but sometimes
3066 * you need a small amount of memory in order to be able to
3067 * page out something else, and this flag essentially protects
3068 * us from recursively trying to free more memory as we're
3069 * trying to free the first piece of memory in the first place).
3070 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003071 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003072 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073
Mel Gorman215ddd62011-07-08 15:39:40 -07003074 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003075 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003076 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003077 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003079 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003080
Mel Gorman215ddd62011-07-08 15:39:40 -07003081 /*
3082 * If the last balance_pgdat was unsuccessful it's unlikely a
3083 * new request of a similar or harder type will succeed soon
3084 * so consider going to sleep on the basis we reclaimed at
3085 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003086 if (balanced_classzone_idx >= new_classzone_idx &&
3087 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003088 new_order = pgdat->kswapd_max_order;
3089 new_classzone_idx = pgdat->classzone_idx;
3090 pgdat->kswapd_max_order = 0;
3091 pgdat->classzone_idx = pgdat->nr_zones - 1;
3092 }
3093
Mel Gorman99504742011-01-13 15:46:20 -08003094 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095 /*
3096 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003097 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 */
3099 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003100 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003102 kswapd_try_to_sleep(pgdat, balanced_order,
3103 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003105 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003106 new_order = order;
3107 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003108 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003109 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111
David Rientjes8fe23e02009-12-14 17:58:33 -08003112 ret = try_to_freeze();
3113 if (kthread_should_stop())
3114 break;
3115
3116 /*
3117 * We can speed up thawing tasks if we don't call balance_pgdat
3118 * after returning from the refrigerator
3119 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003120 if (!ret) {
3121 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003122 balanced_classzone_idx = classzone_idx;
3123 balanced_order = balance_pgdat(pgdat, order,
3124 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003125 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 }
3127 return 0;
3128}
3129
3130/*
3131 * A zone is low on free memory, so wake its kswapd task to service it.
3132 */
Mel Gorman99504742011-01-13 15:46:20 -08003133void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134{
3135 pg_data_t *pgdat;
3136
Con Kolivasf3fe6512006-01-06 00:11:15 -08003137 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 return;
3139
Paul Jackson02a0e532006-12-13 00:34:25 -08003140 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003142 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003143 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003144 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003145 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3146 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003147 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003149 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3150 return;
3151
3152 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003153 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154}
3155
Wu Fengguangadea02a2009-09-21 17:01:42 -07003156/*
3157 * The reclaimable count would be mostly accurate.
3158 * The less reclaimable pages may be
3159 * - mlocked pages, which will be moved to unevictable list when encountered
3160 * - mapped pages, which may require several travels to be reclaimed
3161 * - dirty pages, which is not "instantly" reclaimable
3162 */
3163unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003164{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003165 int nr;
3166
3167 nr = global_page_state(NR_ACTIVE_FILE) +
3168 global_page_state(NR_INACTIVE_FILE);
3169
3170 if (nr_swap_pages > 0)
3171 nr += global_page_state(NR_ACTIVE_ANON) +
3172 global_page_state(NR_INACTIVE_ANON);
3173
3174 return nr;
3175}
3176
3177unsigned long zone_reclaimable_pages(struct zone *zone)
3178{
3179 int nr;
3180
3181 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3182 zone_page_state(zone, NR_INACTIVE_FILE);
3183
3184 if (nr_swap_pages > 0)
3185 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3186 zone_page_state(zone, NR_INACTIVE_ANON);
3187
3188 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003189}
3190
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003191#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003193 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003194 * freed pages.
3195 *
3196 * Rather than trying to age LRUs the aim is to preserve the overall
3197 * LRU order by reclaiming preferentially
3198 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003200unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003202 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003203 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003204 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3205 .may_swap = 1,
3206 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003207 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003208 .nr_to_reclaim = nr_to_reclaim,
3209 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003210 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003212 struct shrink_control shrink = {
3213 .gfp_mask = sc.gfp_mask,
3214 };
3215 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003216 struct task_struct *p = current;
3217 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003219 p->flags |= PF_MEMALLOC;
3220 lockdep_set_current_reclaim_state(sc.gfp_mask);
3221 reclaim_state.reclaimed_slab = 0;
3222 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003223
Ying Hana09ed5e2011-05-24 17:12:26 -07003224 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003225
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003226 p->reclaim_state = NULL;
3227 lockdep_clear_current_reclaim_state();
3228 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003229
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003230 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003232#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234/* It's optimal to keep kswapds on the same CPUs as their memory, but
3235 not required for correctness. So if the last cpu in a node goes
3236 away, we get changed to run anywhere: as the first one comes back,
3237 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003238static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08003239 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003241 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003243 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003244 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003245 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303246 const struct cpumask *mask;
3247
3248 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003249
Rusty Russell3e597942009-01-01 10:12:24 +10303250 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003252 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003253 }
3254 }
3255 return NOTIFY_OK;
3256}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257
Yasunori Goto3218ae12006-06-27 02:53:33 -07003258/*
3259 * This kswapd start function will be called by init and node-hot-add.
3260 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3261 */
3262int kswapd_run(int nid)
3263{
3264 pg_data_t *pgdat = NODE_DATA(nid);
3265 int ret = 0;
3266
3267 if (pgdat->kswapd)
3268 return 0;
3269
3270 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3271 if (IS_ERR(pgdat->kswapd)) {
3272 /* failure at boot is fatal */
3273 BUG_ON(system_state == SYSTEM_BOOTING);
3274 printk("Failed to start kswapd on node %d\n",nid);
3275 ret = -1;
3276 }
3277 return ret;
3278}
3279
David Rientjes8fe23e02009-12-14 17:58:33 -08003280/*
3281 * Called by memory hotplug when all memory in a node is offlined.
3282 */
3283void kswapd_stop(int nid)
3284{
3285 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3286
3287 if (kswapd)
3288 kthread_stop(kswapd);
3289}
3290
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291static int __init kswapd_init(void)
3292{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003293 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003294
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003296 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003297 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 hotcpu_notifier(cpu_callback, 0);
3299 return 0;
3300}
3301
3302module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003303
3304#ifdef CONFIG_NUMA
3305/*
3306 * Zone reclaim mode
3307 *
3308 * If non-zero call zone_reclaim when the number of free pages falls below
3309 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003310 */
3311int zone_reclaim_mode __read_mostly;
3312
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003313#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003314#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003315#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3316#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3317
Christoph Lameter9eeff232006-01-18 17:42:31 -08003318/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003319 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3320 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3321 * a zone.
3322 */
3323#define ZONE_RECLAIM_PRIORITY 4
3324
Christoph Lameter9eeff232006-01-18 17:42:31 -08003325/*
Christoph Lameter96146342006-07-03 00:24:13 -07003326 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3327 * occur.
3328 */
3329int sysctl_min_unmapped_ratio = 1;
3330
3331/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003332 * If the number of slab pages in a zone grows beyond this percentage then
3333 * slab reclaim needs to occur.
3334 */
3335int sysctl_min_slab_ratio = 5;
3336
Mel Gorman90afa5d2009-06-16 15:33:20 -07003337static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3338{
3339 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3340 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3341 zone_page_state(zone, NR_ACTIVE_FILE);
3342
3343 /*
3344 * It's possible for there to be more file mapped pages than
3345 * accounted for by the pages on the file LRU lists because
3346 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3347 */
3348 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3349}
3350
3351/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3352static long zone_pagecache_reclaimable(struct zone *zone)
3353{
3354 long nr_pagecache_reclaimable;
3355 long delta = 0;
3356
3357 /*
3358 * If RECLAIM_SWAP is set, then all file pages are considered
3359 * potentially reclaimable. Otherwise, we have to worry about
3360 * pages like swapcache and zone_unmapped_file_pages() provides
3361 * a better estimate
3362 */
3363 if (zone_reclaim_mode & RECLAIM_SWAP)
3364 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3365 else
3366 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3367
3368 /* If we can't clean pages, remove dirty pages from consideration */
3369 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3370 delta += zone_page_state(zone, NR_FILE_DIRTY);
3371
3372 /* Watch for any possible underflows due to delta */
3373 if (unlikely(delta > nr_pagecache_reclaimable))
3374 delta = nr_pagecache_reclaimable;
3375
3376 return nr_pagecache_reclaimable - delta;
3377}
3378
Christoph Lameter0ff38492006-09-25 23:31:52 -07003379/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003380 * Try to free up some pages from this zone through reclaim.
3381 */
Andrew Morton179e9632006-03-22 00:08:18 -08003382static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003383{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003384 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003385 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003386 struct task_struct *p = current;
3387 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08003388 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08003389 struct scan_control sc = {
3390 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003391 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003392 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003393 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3394 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003395 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003396 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003397 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003398 struct shrink_control shrink = {
3399 .gfp_mask = sc.gfp_mask,
3400 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003401 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003402
Christoph Lameter9eeff232006-01-18 17:42:31 -08003403 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003404 /*
3405 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3406 * and we also need to be able to write out pages for RECLAIM_WRITE
3407 * and RECLAIM_SWAP.
3408 */
3409 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003410 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003411 reclaim_state.reclaimed_slab = 0;
3412 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003413
Mel Gorman90afa5d2009-06-16 15:33:20 -07003414 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003415 /*
3416 * Free memory by calling shrink zone with increasing
3417 * priorities until we have enough memory freed.
3418 */
3419 priority = ZONE_RECLAIM_PRIORITY;
3420 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003421 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003422 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003423 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003424 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003425
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003426 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3427 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003428 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003429 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003430 * many pages were freed in this zone. So we take the current
3431 * number of slab pages and shake the slab until it is reduced
3432 * by the same nr_pages that we used for reclaiming unmapped
3433 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003434 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003435 * Note that shrink_slab will free memory on all zones and may
3436 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003437 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003438 for (;;) {
3439 unsigned long lru_pages = zone_reclaimable_pages(zone);
3440
3441 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003442 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003443 break;
3444
3445 /* Freed enough memory */
3446 nr_slab_pages1 = zone_page_state(zone,
3447 NR_SLAB_RECLAIMABLE);
3448 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3449 break;
3450 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003451
3452 /*
3453 * Update nr_reclaimed by the number of slab pages we
3454 * reclaimed from this zone.
3455 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003456 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3457 if (nr_slab_pages1 < nr_slab_pages0)
3458 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003459 }
3460
Christoph Lameter9eeff232006-01-18 17:42:31 -08003461 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003462 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003463 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003464 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003465}
Andrew Morton179e9632006-03-22 00:08:18 -08003466
3467int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3468{
Andrew Morton179e9632006-03-22 00:08:18 -08003469 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003470 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003471
3472 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003473 * Zone reclaim reclaims unmapped file backed pages and
3474 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003475 *
Christoph Lameter96146342006-07-03 00:24:13 -07003476 * A small portion of unmapped file backed pages is needed for
3477 * file I/O otherwise pages read by file I/O will be immediately
3478 * thrown out if the zone is overallocated. So we do not reclaim
3479 * if less than a specified percentage of the zone is used by
3480 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003481 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003482 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3483 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003484 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003485
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003486 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003487 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003488
Andrew Morton179e9632006-03-22 00:08:18 -08003489 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003490 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003491 */
David Rientjesd773ed62007-10-16 23:26:01 -07003492 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003493 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003494
3495 /*
3496 * Only run zone reclaim on the local zone or on zones that do not
3497 * have associated processors. This will favor the local processor
3498 * over remote processors and spread off node memory allocations
3499 * as wide as possible.
3500 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003501 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003502 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003503 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003504
3505 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003506 return ZONE_RECLAIM_NOSCAN;
3507
David Rientjesd773ed62007-10-16 23:26:01 -07003508 ret = __zone_reclaim(zone, gfp_mask, order);
3509 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3510
Mel Gorman24cf725182009-06-16 15:33:23 -07003511 if (!ret)
3512 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3513
David Rientjesd773ed62007-10-16 23:26:01 -07003514 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003515}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003516#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003517
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003518/*
3519 * page_evictable - test whether a page is evictable
3520 * @page: the page to test
3521 * @vma: the VMA in which the page is or will be mapped, may be NULL
3522 *
3523 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003524 * lists vs unevictable list. The vma argument is !NULL when called from the
3525 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003526 *
3527 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003528 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003529 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003530 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003531 */
3532int page_evictable(struct page *page, struct vm_area_struct *vma)
3533{
3534
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003535 if (mapping_unevictable(page_mapping(page)))
3536 return 0;
3537
Nick Pigginb291f002008-10-18 20:26:44 -07003538 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3539 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003540
3541 return 1;
3542}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003543
Hugh Dickins850465792012-01-20 14:34:19 -08003544#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003545/**
Hugh Dickins24513262012-01-20 14:34:21 -08003546 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3547 * @pages: array of pages to check
3548 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003549 *
Hugh Dickins24513262012-01-20 14:34:21 -08003550 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003551 *
3552 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003553 */
Hugh Dickins24513262012-01-20 14:34:21 -08003554void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003555{
Johannes Weiner925b7672012-01-12 17:18:15 -08003556 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003557 struct zone *zone = NULL;
3558 int pgscanned = 0;
3559 int pgrescued = 0;
3560 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003561
Hugh Dickins24513262012-01-20 14:34:21 -08003562 for (i = 0; i < nr_pages; i++) {
3563 struct page *page = pages[i];
3564 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003565
Hugh Dickins24513262012-01-20 14:34:21 -08003566 pgscanned++;
3567 pagezone = page_zone(page);
3568 if (pagezone != zone) {
3569 if (zone)
3570 spin_unlock_irq(&zone->lru_lock);
3571 zone = pagezone;
3572 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003573 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003574
Hugh Dickins24513262012-01-20 14:34:21 -08003575 if (!PageLRU(page) || !PageUnevictable(page))
3576 continue;
3577
3578 if (page_evictable(page, NULL)) {
3579 enum lru_list lru = page_lru_base_type(page);
3580
3581 VM_BUG_ON(PageActive(page));
3582 ClearPageUnevictable(page);
3583 __dec_zone_state(zone, NR_UNEVICTABLE);
3584 lruvec = mem_cgroup_lru_move_lists(zone, page,
3585 LRU_UNEVICTABLE, lru);
3586 list_move(&page->lru, &lruvec->lists[lru]);
3587 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3588 pgrescued++;
3589 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003590 }
Hugh Dickins24513262012-01-20 14:34:21 -08003591
3592 if (zone) {
3593 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3594 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3595 spin_unlock_irq(&zone->lru_lock);
3596 }
Hugh Dickins850465792012-01-20 14:34:19 -08003597}
3598#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003599
Johannes Weiner264e56d2011-10-31 17:09:13 -07003600static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003601{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003602 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003603 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003604 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003605 "one, please send an email to linux-mm@kvack.org.\n",
3606 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003607}
3608
3609/*
3610 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3611 * all nodes' unevictable lists for evictable pages
3612 */
3613unsigned long scan_unevictable_pages;
3614
3615int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003616 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003617 size_t *length, loff_t *ppos)
3618{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003619 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003620 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003621 scan_unevictable_pages = 0;
3622 return 0;
3623}
3624
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003625#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003626/*
3627 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3628 * a specified node's per zone unevictable lists for evictable pages.
3629 */
3630
Kay Sievers10fbcf42011-12-21 14:48:43 -08003631static ssize_t read_scan_unevictable_node(struct device *dev,
3632 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003633 char *buf)
3634{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003635 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003636 return sprintf(buf, "0\n"); /* always zero; should fit... */
3637}
3638
Kay Sievers10fbcf42011-12-21 14:48:43 -08003639static ssize_t write_scan_unevictable_node(struct device *dev,
3640 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003641 const char *buf, size_t count)
3642{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003643 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003644 return 1;
3645}
3646
3647
Kay Sievers10fbcf42011-12-21 14:48:43 -08003648static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003649 read_scan_unevictable_node,
3650 write_scan_unevictable_node);
3651
3652int scan_unevictable_register_node(struct node *node)
3653{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003654 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003655}
3656
3657void scan_unevictable_unregister_node(struct node *node)
3658{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003659 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003660}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003661#endif