blob: 2880396f7953b03476db86957c66a3787b40cb8c [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>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51
Nick Piggin0f8053a2006-03-22 00:08:33 -080052#include "internal.h"
53
Mel Gorman33906bc2010-08-09 17:19:16 -070054#define CREATE_TRACE_POINTS
55#include <trace/events/vmscan.h>
56
Mel Gormanee64fc92011-01-13 15:45:55 -080057/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080058 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080063 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080065 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080066 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080067 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080068typedef unsigned __bitwise__ reclaim_mode_t;
69#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070074
Linus Torvalds1da177e2005-04-16 15:20:36 -070075struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
78
Rik van Riela79311c2009-01-06 14:40:01 -080079 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
81
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080082 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
84
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080085 unsigned long hibernation_mode;
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040088 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 int may_writepage;
91
Johannes Weinera6dc60f2009-03-31 15:19:30 -070092 /* Can mapped pages be reclaimed? */
93 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080094
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070095 /* Can pages be swapped as part of reclaim? */
96 int may_swap;
97
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070098 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080099
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700100 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700103 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800104 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700105
Johannes Weinerf16015f2012-01-12 17:17:52 -0800106 /*
107 * The memory cgroup that hit its limit and as a result is the
108 * primary target of this reclaim invocation.
109 */
110 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -0800111
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700112 /*
113 * Nodemask of nodes allowed by the caller. If NULL, all nodes
114 * are scanned.
115 */
116 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117};
118
Johannes Weinerf16015f2012-01-12 17:17:52 -0800119struct mem_cgroup_zone {
120 struct mem_cgroup *mem_cgroup;
121 struct zone *zone;
122};
123
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
125
126#ifdef ARCH_HAS_PREFETCH
127#define prefetch_prev_lru_page(_page, _base, _field) \
128 do { \
129 if ((_page)->lru.prev != _base) { \
130 struct page *prev; \
131 \
132 prev = lru_to_page(&(_page->lru)); \
133 prefetch(&prev->_field); \
134 } \
135 } while (0)
136#else
137#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
138#endif
139
140#ifdef ARCH_HAS_PREFETCHW
141#define prefetchw_prev_lru_page(_page, _base, _field) \
142 do { \
143 if ((_page)->lru.prev != _base) { \
144 struct page *prev; \
145 \
146 prev = lru_to_page(&(_page->lru)); \
147 prefetchw(&prev->_field); \
148 } \
149 } while (0)
150#else
151#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
152#endif
153
154/*
155 * From 0 .. 100. Higher means more swappy.
156 */
157int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700158long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
160static LIST_HEAD(shrinker_list);
161static DECLARE_RWSEM(shrinker_rwsem);
162
Balbir Singh00f0b822008-03-04 14:28:39 -0800163#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800164static bool global_reclaim(struct scan_control *sc)
165{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800166 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800167}
168
Johannes Weinerf16015f2012-01-12 17:17:52 -0800169static bool scanning_global_lru(struct mem_cgroup_zone *mz)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800170{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800171 return !mz->mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800172}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800173#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800174static bool global_reclaim(struct scan_control *sc)
175{
176 return true;
177}
178
Johannes Weinerf16015f2012-01-12 17:17:52 -0800179static bool scanning_global_lru(struct mem_cgroup_zone *mz)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800180{
181 return true;
182}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800183#endif
184
Johannes Weinerf16015f2012-01-12 17:17:52 -0800185static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800186{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800187 if (!scanning_global_lru(mz))
188 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800189
Johannes Weinerf16015f2012-01-12 17:17:52 -0800190 return &mz->zone->reclaim_stat;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800191}
192
Johannes Weinerf16015f2012-01-12 17:17:52 -0800193static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
194 enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800195{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800196 if (!scanning_global_lru(mz))
197 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
198 zone_to_nid(mz->zone),
199 zone_idx(mz->zone),
200 BIT(lru));
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800201
Johannes Weinerf16015f2012-01-12 17:17:52 -0800202 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800203}
204
205
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206/*
207 * Add a shrinker callback to be called from the vm
208 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700209void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800211 atomic_long_set(&shrinker->nr_in_batch, 0);
Rusty Russell8e1f9362007-07-17 04:03:17 -0700212 down_write(&shrinker_rwsem);
213 list_add_tail(&shrinker->list, &shrinker_list);
214 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218/*
219 * Remove one
220 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700221void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222{
223 down_write(&shrinker_rwsem);
224 list_del(&shrinker->list);
225 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700227EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
Ying Han1495f232011-05-24 17:12:27 -0700229static inline int do_shrinker_shrink(struct shrinker *shrinker,
230 struct shrink_control *sc,
231 unsigned long nr_to_scan)
232{
233 sc->nr_to_scan = nr_to_scan;
234 return (*shrinker->shrink)(shrinker, sc);
235}
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237#define SHRINK_BATCH 128
238/*
239 * Call the shrink functions to age shrinkable caches
240 *
241 * Here we assume it costs one seek to replace a lru page and that it also
242 * takes a seek to recreate a cache object. With this in mind we age equal
243 * percentages of the lru and ageable caches. This should balance the seeks
244 * generated by these structures.
245 *
Simon Arlott183ff222007-10-20 01:27:18 +0200246 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 * slab to avoid swapping.
248 *
249 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
250 *
251 * `lru_pages' represents the number of on-LRU pages in all the zones which
252 * are eligible for the caller's allocation attempt. It is used for balancing
253 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700254 *
255 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700257unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700258 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700259 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260{
261 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800262 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
Ying Han1495f232011-05-24 17:12:27 -0700264 if (nr_pages_scanned == 0)
265 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
Minchan Kimf06590b2011-05-24 17:11:11 -0700267 if (!down_read_trylock(&shrinker_rwsem)) {
268 /* Assume we'll be able to shrink next time */
269 ret = 1;
270 goto out;
271 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272
273 list_for_each_entry(shrinker, &shrinker_list, list) {
274 unsigned long long delta;
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800275 long total_scan;
276 long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000277 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000278 long nr;
279 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000280 long batch_size = shrinker->batch ? shrinker->batch
281 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800283 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
284 if (max_pass <= 0)
285 continue;
286
Dave Chinneracf92b42011-07-08 14:14:35 +1000287 /*
288 * copy the current shrinker scan count into a local variable
289 * and zero it so that other concurrent shrinker invocations
290 * don't also do this scanning work.
291 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800292 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
Dave Chinneracf92b42011-07-08 14:14:35 +1000293
294 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700295 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800296 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000298 total_scan += delta;
299 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700300 printk(KERN_ERR "shrink_slab: %pF negative objects to "
301 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000302 shrinker->shrink, total_scan);
303 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800304 }
305
306 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000307 * We need to avoid excessive windup on filesystem shrinkers
308 * due to large numbers of GFP_NOFS allocations causing the
309 * shrinkers to return -1 all the time. This results in a large
310 * nr being built up so when a shrink that can do some work
311 * comes along it empties the entire cache due to nr >>>
312 * max_pass. This is bad for sustaining a working set in
313 * memory.
314 *
315 * Hence only allow the shrinker to scan the entire cache when
316 * a large delta change is calculated directly.
317 */
318 if (delta < max_pass / 4)
319 total_scan = min(total_scan, max_pass / 2);
320
321 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800322 * Avoid risking looping forever due to too large nr value:
323 * never try to free more than twice the estimate number of
324 * freeable entries.
325 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000326 if (total_scan > max_pass * 2)
327 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
Dave Chinneracf92b42011-07-08 14:14:35 +1000329 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000330 nr_pages_scanned, lru_pages,
331 max_pass, delta, total_scan);
332
Dave Chinnere9299f52011-07-08 14:14:37 +1000333 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700334 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335
Ying Han1495f232011-05-24 17:12:27 -0700336 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
337 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000338 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 if (shrink_ret == -1)
340 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700341 if (shrink_ret < nr_before)
342 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000343 count_vm_events(SLABS_SCANNED, batch_size);
344 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345
346 cond_resched();
347 }
348
Dave Chinneracf92b42011-07-08 14:14:35 +1000349 /*
350 * move the unused scan count back into the shrinker in a
351 * manner that handles concurrent updates. If we exhausted the
352 * scan, there is no need to do an update.
353 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800354 if (total_scan > 0)
355 new_nr = atomic_long_add_return(total_scan,
356 &shrinker->nr_in_batch);
357 else
358 new_nr = atomic_long_read(&shrinker->nr_in_batch);
Dave Chinneracf92b42011-07-08 14:14:35 +1000359
360 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 }
362 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700363out:
364 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700365 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366}
367
Mel Gormanf3a310b2011-01-13 15:46:00 -0800368static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700369 bool sync)
370{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800371 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700372
373 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800374 * Initially assume we are entering either lumpy reclaim or
375 * reclaim/compaction.Depending on the order, we will either set the
376 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700377 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800378 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800379 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800380 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800381 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700382
383 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800384 * Avoid using lumpy reclaim or reclaim/compaction if possible by
385 * restricting when its set to either costly allocations or when
386 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700387 */
388 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800389 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700390 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800391 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700392 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800393 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700394}
395
Mel Gormanf3a310b2011-01-13 15:46:00 -0800396static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700397{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800398 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700399}
400
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401static inline int is_page_cache_freeable(struct page *page)
402{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700403 /*
404 * A freeable page cache page is referenced only by the caller
405 * that isolated the page, the page cache radix tree and
406 * optional buffer heads at page->private.
407 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700408 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409}
410
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700411static int may_write_to_queue(struct backing_dev_info *bdi,
412 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413{
Christoph Lameter930d9152006-01-08 01:00:47 -0800414 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 return 1;
416 if (!bdi_write_congested(bdi))
417 return 1;
418 if (bdi == current->backing_dev_info)
419 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700420
421 /* lumpy reclaim for hugepage often need a lot of write */
422 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
423 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 return 0;
425}
426
427/*
428 * We detected a synchronous write error writing a page out. Probably
429 * -ENOSPC. We need to propagate that into the address_space for a subsequent
430 * fsync(), msync() or close().
431 *
432 * The tricky part is that after writepage we cannot touch the mapping: nothing
433 * prevents it from being freed up. But we have a ref on the page and once
434 * that page is locked, the mapping is pinned.
435 *
436 * We're allowed to run sleeping lock_page() here because we know the caller has
437 * __GFP_FS.
438 */
439static void handle_write_error(struct address_space *mapping,
440 struct page *page, int error)
441{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100442 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700443 if (page_mapping(page) == mapping)
444 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 unlock_page(page);
446}
447
Christoph Lameter04e62a22006-06-23 02:03:38 -0700448/* possible outcome of pageout() */
449typedef enum {
450 /* failed to write page out, page is locked */
451 PAGE_KEEP,
452 /* move page to the active list, page is locked */
453 PAGE_ACTIVATE,
454 /* page has been sent to the disk successfully, page is unlocked */
455 PAGE_SUCCESS,
456 /* page is clean and locked */
457 PAGE_CLEAN,
458} pageout_t;
459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460/*
Andrew Morton1742f192006-03-22 00:08:21 -0800461 * pageout is called by shrink_page_list() for each dirty page.
462 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700464static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700465 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466{
467 /*
468 * If the page is dirty, only perform writeback if that write
469 * will be non-blocking. To prevent this allocation from being
470 * stalled by pagecache activity. But note that there may be
471 * stalls if we need to run get_block(). We could test
472 * PagePrivate for that.
473 *
Vincent Li6aceb532009-12-14 17:58:49 -0800474 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 * this page's queue, we can perform writeback even if that
476 * will block.
477 *
478 * If the page is swapcache, write it back even if that would
479 * block, for some throttling. This happens by accident, because
480 * swap_backing_dev_info is bust: it doesn't reflect the
481 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
483 if (!is_page_cache_freeable(page))
484 return PAGE_KEEP;
485 if (!mapping) {
486 /*
487 * Some data journaling orphaned pages can have
488 * page->mapping == NULL while being dirty with clean buffers.
489 */
David Howells266cf652009-04-03 16:42:36 +0100490 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 if (try_to_free_buffers(page)) {
492 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700493 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 return PAGE_CLEAN;
495 }
496 }
497 return PAGE_KEEP;
498 }
499 if (mapping->a_ops->writepage == NULL)
500 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700501 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 return PAGE_KEEP;
503
504 if (clear_page_dirty_for_io(page)) {
505 int res;
506 struct writeback_control wbc = {
507 .sync_mode = WB_SYNC_NONE,
508 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700509 .range_start = 0,
510 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 .for_reclaim = 1,
512 };
513
514 SetPageReclaim(page);
515 res = mapping->a_ops->writepage(page, &wbc);
516 if (res < 0)
517 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800518 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 ClearPageReclaim(page);
520 return PAGE_ACTIVATE;
521 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700522
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 if (!PageWriteback(page)) {
524 /* synchronous write or broken a_ops? */
525 ClearPageReclaim(page);
526 }
Mel Gorman755f0222010-08-09 17:19:18 -0700527 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800528 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700529 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 return PAGE_SUCCESS;
531 }
532
533 return PAGE_CLEAN;
534}
535
Andrew Mortona649fd92006-10-17 00:09:36 -0700536/*
Nick Piggine2867812008-07-25 19:45:30 -0700537 * Same as remove_mapping, but if the page is removed from the mapping, it
538 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700539 */
Nick Piggine2867812008-07-25 19:45:30 -0700540static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800541{
Nick Piggin28e4d962006-09-25 23:31:23 -0700542 BUG_ON(!PageLocked(page));
543 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800544
Nick Piggin19fd6232008-07-25 19:45:32 -0700545 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800546 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700547 * The non racy check for a busy page.
548 *
549 * Must be careful with the order of the tests. When someone has
550 * a ref to the page, it may be possible that they dirty it then
551 * drop the reference. So if PageDirty is tested before page_count
552 * here, then the following race may occur:
553 *
554 * get_user_pages(&page);
555 * [user mapping goes away]
556 * write_to(page);
557 * !PageDirty(page) [good]
558 * SetPageDirty(page);
559 * put_page(page);
560 * !page_count(page) [good, discard it]
561 *
562 * [oops, our write_to data is lost]
563 *
564 * Reversing the order of the tests ensures such a situation cannot
565 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
566 * load is not satisfied before that of page->_count.
567 *
568 * Note that if SetPageDirty is always performed via set_page_dirty,
569 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800570 */
Nick Piggine2867812008-07-25 19:45:30 -0700571 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800572 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700573 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
574 if (unlikely(PageDirty(page))) {
575 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800576 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700577 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800578
579 if (PageSwapCache(page)) {
580 swp_entry_t swap = { .val = page_private(page) };
581 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700582 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700583 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700584 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500585 void (*freepage)(struct page *);
586
587 freepage = mapping->a_ops->freepage;
588
Minchan Kime64a7822011-03-22 16:32:44 -0700589 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700590 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700591 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500592
593 if (freepage != NULL)
594 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800595 }
596
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800597 return 1;
598
599cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700600 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800601 return 0;
602}
603
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604/*
Nick Piggine2867812008-07-25 19:45:30 -0700605 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
606 * someone else has a ref on the page, abort and return 0. If it was
607 * successfully detached, return 1. Assumes the caller has a single ref on
608 * this page.
609 */
610int remove_mapping(struct address_space *mapping, struct page *page)
611{
612 if (__remove_mapping(mapping, page)) {
613 /*
614 * Unfreezing the refcount with 1 rather than 2 effectively
615 * drops the pagecache ref for us without requiring another
616 * atomic operation.
617 */
618 page_unfreeze_refs(page, 1);
619 return 1;
620 }
621 return 0;
622}
623
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700624/**
625 * putback_lru_page - put previously isolated page onto appropriate LRU list
626 * @page: page to be put back to appropriate lru list
627 *
628 * Add previously isolated @page to appropriate LRU list.
629 * Page may still be unevictable for other reasons.
630 *
631 * lru_lock must not be held, interrupts must be enabled.
632 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700633void putback_lru_page(struct page *page)
634{
635 int lru;
636 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700637 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700638
639 VM_BUG_ON(PageLRU(page));
640
641redo:
642 ClearPageUnevictable(page);
643
644 if (page_evictable(page, NULL)) {
645 /*
646 * For evictable pages, we can use the cache.
647 * In event of a race, worst case is we end up with an
648 * unevictable page on [in]active list.
649 * We know how to handle that.
650 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700651 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700652 lru_cache_add_lru(page, lru);
653 } else {
654 /*
655 * Put unevictable pages directly on zone's unevictable
656 * list.
657 */
658 lru = LRU_UNEVICTABLE;
659 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700660 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700661 * When racing with an mlock or AS_UNEVICTABLE clearing
662 * (page is unlocked) make sure that if the other thread
663 * does not observe our setting of PG_lru and fails
664 * isolation/check_move_unevictable_page,
665 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700666 * the page back to the evictable list.
667 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700668 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700669 */
670 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700671 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700672
673 /*
674 * page's status can change while we move it among lru. If an evictable
675 * page is on unevictable list, it never be freed. To avoid that,
676 * check after we added it to the list, again.
677 */
678 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
679 if (!isolate_lru_page(page)) {
680 put_page(page);
681 goto redo;
682 }
683 /* This means someone else dropped this page from LRU
684 * So, it will be freed or putback to LRU again. There is
685 * nothing to do here.
686 */
687 }
688
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700689 if (was_unevictable && lru != LRU_UNEVICTABLE)
690 count_vm_event(UNEVICTABLE_PGRESCUED);
691 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
692 count_vm_event(UNEVICTABLE_PGCULLED);
693
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700694 put_page(page); /* drop ref from isolate */
695}
696
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800697enum page_references {
698 PAGEREF_RECLAIM,
699 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800700 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800701 PAGEREF_ACTIVATE,
702};
703
704static enum page_references page_check_references(struct page *page,
Johannes Weinerf16015f2012-01-12 17:17:52 -0800705 struct mem_cgroup_zone *mz,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800706 struct scan_control *sc)
707{
Johannes Weiner64574742010-03-05 13:42:22 -0800708 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800709 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800710
Johannes Weinerf16015f2012-01-12 17:17:52 -0800711 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
Johannes Weiner64574742010-03-05 13:42:22 -0800712 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800713
714 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800715 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800716 return PAGEREF_RECLAIM;
717
718 /*
719 * Mlock lost the isolation race with us. Let try_to_unmap()
720 * move the page to the unevictable list.
721 */
722 if (vm_flags & VM_LOCKED)
723 return PAGEREF_RECLAIM;
724
Johannes Weiner64574742010-03-05 13:42:22 -0800725 if (referenced_ptes) {
726 if (PageAnon(page))
727 return PAGEREF_ACTIVATE;
728 /*
729 * All mapped pages start out with page table
730 * references from the instantiating fault, so we need
731 * to look twice if a mapped file page is used more
732 * than once.
733 *
734 * Mark it and spare it for another trip around the
735 * inactive list. Another page table reference will
736 * lead to its activation.
737 *
738 * Note: the mark is set for activated pages as well
739 * so that recently deactivated but used pages are
740 * quickly recovered.
741 */
742 SetPageReferenced(page);
743
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800744 if (referenced_page || referenced_ptes > 1)
Johannes Weiner64574742010-03-05 13:42:22 -0800745 return PAGEREF_ACTIVATE;
746
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800747 /*
748 * Activate file-backed executable pages after first usage.
749 */
750 if (vm_flags & VM_EXEC)
751 return PAGEREF_ACTIVATE;
752
Johannes Weiner64574742010-03-05 13:42:22 -0800753 return PAGEREF_KEEP;
754 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800755
756 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700757 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner64574742010-03-05 13:42:22 -0800758 return PAGEREF_RECLAIM_CLEAN;
759
760 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800761}
762
Nick Piggine2867812008-07-25 19:45:30 -0700763/*
Andrew Morton1742f192006-03-22 00:08:21 -0800764 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 */
Andrew Morton1742f192006-03-22 00:08:21 -0800766static unsigned long shrink_page_list(struct list_head *page_list,
Johannes Weinerf16015f2012-01-12 17:17:52 -0800767 struct mem_cgroup_zone *mz,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700768 struct scan_control *sc,
Mel Gorman92df3a72011-10-31 17:07:56 -0700769 int priority,
770 unsigned long *ret_nr_dirty,
771 unsigned long *ret_nr_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772{
773 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700774 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700776 unsigned long nr_dirty = 0;
777 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800778 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700779 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
781 cond_resched();
782
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800784 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 struct address_space *mapping;
786 struct page *page;
787 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
789 cond_resched();
790
791 page = lru_to_page(page_list);
792 list_del(&page->lru);
793
Nick Piggin529ae9a2008-08-02 12:01:03 +0200794 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 goto keep;
796
Nick Piggin725d7042006-09-25 23:30:55 -0700797 VM_BUG_ON(PageActive(page));
Johannes Weinerf16015f2012-01-12 17:17:52 -0800798 VM_BUG_ON(page_zone(page) != mz->zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
800 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800801
Nick Pigginb291f002008-10-18 20:26:44 -0700802 if (unlikely(!page_evictable(page, NULL)))
803 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700804
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700805 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800806 goto keep_locked;
807
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 /* Double the slab pressure for mapped and swapcache pages */
809 if (page_mapped(page) || PageSwapCache(page))
810 sc->nr_scanned++;
811
Andy Whitcroftc661b072007-08-22 14:01:26 -0700812 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
813 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
814
815 if (PageWriteback(page)) {
Mel Gorman92df3a72011-10-31 17:07:56 -0700816 nr_writeback++;
Andy Whitcroftc661b072007-08-22 14:01:26 -0700817 /*
Mel Gormana18bba02011-10-31 17:07:42 -0700818 * Synchronous reclaim cannot queue pages for
819 * writeback due to the possibility of stack overflow
820 * but if it encounters a page under writeback, wait
821 * for the IO to complete.
Andy Whitcroftc661b072007-08-22 14:01:26 -0700822 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800823 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700824 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700825 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700826 else {
827 unlock_page(page);
828 goto keep_lumpy;
829 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700830 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
Johannes Weinerf16015f2012-01-12 17:17:52 -0800832 references = page_check_references(page, mz, sc);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800833 switch (references) {
834 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800836 case PAGEREF_KEEP:
837 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800838 case PAGEREF_RECLAIM:
839 case PAGEREF_RECLAIM_CLEAN:
840 ; /* try to reclaim the page below */
841 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 /*
844 * Anonymous process memory has backing store?
845 * Try to allocate it some swap space here.
846 */
Nick Pigginb291f002008-10-18 20:26:44 -0700847 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800848 if (!(sc->gfp_mask & __GFP_IO))
849 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800850 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800852 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700853 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854
855 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
857 /*
858 * The page is mapped into the page tables of one or more
859 * processes. Try to unmap it here.
860 */
861 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200862 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 case SWAP_FAIL:
864 goto activate_locked;
865 case SWAP_AGAIN:
866 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700867 case SWAP_MLOCK:
868 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 case SWAP_SUCCESS:
870 ; /* try to free the page below */
871 }
872 }
873
874 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700875 nr_dirty++;
876
Mel Gormanee728862011-10-31 17:07:38 -0700877 /*
878 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700879 * avoid risk of stack overflow but do not writeback
880 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700881 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700882 if (page_is_file_cache(page) &&
883 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700884 /*
885 * Immediately reclaim when written back.
886 * Similar in principal to deactivate_page()
887 * except we already have the page isolated
888 * and know it's dirty
889 */
890 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
891 SetPageReclaim(page);
892
Mel Gormanee728862011-10-31 17:07:38 -0700893 goto keep_locked;
894 }
895
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800896 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700898 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800900 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 goto keep_locked;
902
903 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700904 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700906 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 goto keep_locked;
908 case PAGE_ACTIVATE:
909 goto activate_locked;
910 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700911 if (PageWriteback(page))
912 goto keep_lumpy;
913 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700915
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 /*
917 * A synchronous write - probably a ramdisk. Go
918 * ahead and try to reclaim the page.
919 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200920 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 goto keep;
922 if (PageDirty(page) || PageWriteback(page))
923 goto keep_locked;
924 mapping = page_mapping(page);
925 case PAGE_CLEAN:
926 ; /* try to free the page below */
927 }
928 }
929
930 /*
931 * If the page has buffers, try to free the buffer mappings
932 * associated with this page. If we succeed we try to free
933 * the page as well.
934 *
935 * We do this even if the page is PageDirty().
936 * try_to_release_page() does not perform I/O, but it is
937 * possible for a page to have PageDirty set, but it is actually
938 * clean (all its buffers are clean). This happens if the
939 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700940 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 * try_to_release_page() will discover that cleanness and will
942 * drop the buffers and mark the page clean - it can be freed.
943 *
944 * Rarely, pages can have buffers and no ->mapping. These are
945 * the pages which were not successfully invalidated in
946 * truncate_complete_page(). We try to drop those buffers here
947 * and if that worked, and the page is no longer mapped into
948 * process address space (page_count == 1) it can be freed.
949 * Otherwise, leave the page on the LRU so it is swappable.
950 */
David Howells266cf652009-04-03 16:42:36 +0100951 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 if (!try_to_release_page(page, sc->gfp_mask))
953 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700954 if (!mapping && page_count(page) == 1) {
955 unlock_page(page);
956 if (put_page_testzero(page))
957 goto free_it;
958 else {
959 /*
960 * rare race with speculative reference.
961 * the speculative reference will free
962 * this page shortly, so we may
963 * increment nr_reclaimed here (and
964 * leave it off the LRU).
965 */
966 nr_reclaimed++;
967 continue;
968 }
969 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 }
971
Nick Piggine2867812008-07-25 19:45:30 -0700972 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800973 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974
Nick Piggina978d6f2008-10-18 20:26:58 -0700975 /*
976 * At this point, we have no other references and there is
977 * no way to pick any more up (removed from LRU, removed
978 * from pagecache). Can use non-atomic bitops now (and
979 * we obviously don't have to worry about waking up a process
980 * waiting on the page lock, because there are no references.
981 */
982 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700983free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800984 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700985
986 /*
987 * Is there need to periodically free_page_list? It would
988 * appear not as the counts should be low
989 */
990 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 continue;
992
Nick Pigginb291f002008-10-18 20:26:44 -0700993cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800994 if (PageSwapCache(page))
995 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700996 unlock_page(page);
997 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800998 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700999 continue;
1000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001002 /* Not a candidate for swapping, so reclaim swap space. */
1003 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001004 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001005 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 SetPageActive(page);
1007 pgactivate++;
1008keep_locked:
1009 unlock_page(page);
1010keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -08001011 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001012keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001014 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001016
Mel Gorman0e093d992010-10-26 14:21:45 -07001017 /*
1018 * Tag a zone as congested if all the dirty pages encountered were
1019 * backed by a congested BDI. In this case, reclaimers should just
1020 * back off and wait for congestion to clear because further reclaim
1021 * will encounter the same problem
1022 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001023 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
Johannes Weinerf16015f2012-01-12 17:17:52 -08001024 zone_set_flag(mz->zone, ZONE_CONGESTED);
Mel Gorman0e093d992010-10-26 14:21:45 -07001025
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001026 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001027
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001029 count_vm_events(PGACTIVATE, pgactivate);
Mel Gorman92df3a72011-10-31 17:07:56 -07001030 *ret_nr_dirty += nr_dirty;
1031 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -08001032 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033}
1034
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001035/*
1036 * Attempt to remove the specified page from its LRU. Only take this page
1037 * if it is of the appropriate PageActive status. Pages which are being
1038 * freed elsewhere are also ignored.
1039 *
1040 * page: page to consider
1041 * mode: one of the LRU isolation modes defined above
1042 *
1043 * returns 0 on success, -ve errno on failure.
1044 */
Minchan Kim4356f212011-10-31 17:06:47 -07001045int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001046{
Minchan Kim4356f212011-10-31 17:06:47 -07001047 bool all_lru_mode;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001048 int ret = -EINVAL;
1049
1050 /* Only take pages on the LRU. */
1051 if (!PageLRU(page))
1052 return ret;
1053
Minchan Kim4356f212011-10-31 17:06:47 -07001054 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1055 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1056
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001057 /*
1058 * When checking the active state, we need to be sure we are
1059 * dealing with comparible boolean values. Take the logical not
1060 * of each.
1061 */
Minchan Kim4356f212011-10-31 17:06:47 -07001062 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001063 return ret;
1064
Minchan Kim4356f212011-10-31 17:06:47 -07001065 if (!all_lru_mode && !!page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001066 return ret;
1067
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001068 /*
1069 * When this function is being called for lumpy reclaim, we
1070 * initially look into all LRU pages, active, inactive and
1071 * unevictable; only give shrink_page_list evictable pages.
1072 */
1073 if (PageUnevictable(page))
1074 return ret;
1075
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001076 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001077
Mel Gormanc8244932012-01-12 17:19:38 -08001078 /*
1079 * To minimise LRU disruption, the caller can indicate that it only
1080 * wants to isolate pages it will be able to operate on without
1081 * blocking - clean pages for the most part.
1082 *
1083 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1084 * is used by reclaim when it is cannot write to backing storage
1085 *
1086 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1087 * that it is possible to migrate without blocking
1088 */
1089 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1090 /* All the caller can do on PageWriteback is block */
1091 if (PageWriteback(page))
1092 return ret;
1093
1094 if (PageDirty(page)) {
1095 struct address_space *mapping;
1096
1097 /* ISOLATE_CLEAN means only clean pages */
1098 if (mode & ISOLATE_CLEAN)
1099 return ret;
1100
1101 /*
1102 * Only pages without mappings or that have a
1103 * ->migratepage callback are possible to migrate
1104 * without blocking
1105 */
1106 mapping = page_mapping(page);
1107 if (mapping && !mapping->a_ops->migratepage)
1108 return ret;
1109 }
1110 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001111
Minchan Kimf80c0672011-10-31 17:06:55 -07001112 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1113 return ret;
1114
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001115 if (likely(get_page_unless_zero(page))) {
1116 /*
1117 * Be careful not to clear PageLRU until after we're
1118 * sure the page is not being freed elsewhere -- the
1119 * page release code relies on it.
1120 */
1121 ClearPageLRU(page);
1122 ret = 0;
1123 }
1124
1125 return ret;
1126}
1127
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001128/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 * zone->lru_lock is heavily contended. Some of the functions that
1130 * shrink the lists perform better by taking out a batch of pages
1131 * and working on them outside the LRU lock.
1132 *
1133 * For pagecache intensive workloads, this function is the hottest
1134 * spot in the kernel (apart from copy_*_user functions).
1135 *
1136 * Appropriate locks must be held before calling this function.
1137 *
1138 * @nr_to_scan: The number of pages to look through on the list.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001139 * @mz: The mem_cgroup_zone to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001141 * @nr_scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001142 * @order: The caller's attempted allocation order
1143 * @mode: One of the LRU isolation modes
Hugh Dickinsf6260122012-01-12 17:20:06 -08001144 * @active: True [1] if isolating active pages
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001145 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 *
1147 * returns how many pages were moved onto *@dst.
1148 */
Andrew Morton69e05942006-03-22 00:08:19 -08001149static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Hugh Dickinsf6260122012-01-12 17:20:06 -08001150 struct mem_cgroup_zone *mz, struct list_head *dst,
1151 unsigned long *nr_scanned, int order, isolate_mode_t mode,
1152 int active, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153{
Hugh Dickinsf6260122012-01-12 17:20:06 -08001154 struct lruvec *lruvec;
1155 struct list_head *src;
Andrew Morton69e05942006-03-22 00:08:19 -08001156 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001157 unsigned long nr_lumpy_taken = 0;
1158 unsigned long nr_lumpy_dirty = 0;
1159 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001160 unsigned long scan;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001161 int lru = LRU_BASE;
1162
1163 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1164 if (active)
1165 lru += LRU_ACTIVE;
1166 if (file)
1167 lru += LRU_FILE;
1168 src = &lruvec->lists[lru];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001170 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001171 struct page *page;
1172 unsigned long pfn;
1173 unsigned long end_pfn;
1174 unsigned long page_pfn;
1175 int zone_id;
1176
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 page = lru_to_page(src);
1178 prefetchw_prev_lru_page(page, src, flags);
1179
Nick Piggin725d7042006-09-25 23:30:55 -07001180 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001181
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001182 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001183 case 0:
Johannes Weiner925b7672012-01-12 17:18:15 -08001184 mem_cgroup_lru_del(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001185 list_move(&page->lru, dst);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001186 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001187 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001188
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001189 case -EBUSY:
1190 /* else it is being freed elsewhere */
1191 list_move(&page->lru, src);
1192 continue;
1193
1194 default:
1195 BUG();
1196 }
1197
1198 if (!order)
1199 continue;
1200
1201 /*
1202 * Attempt to take all pages in the order aligned region
1203 * surrounding the tag page. Only take those pages of
1204 * the same active state as that tag page. We may safely
1205 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001206 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001207 * where that page is in a different zone we will detect
1208 * it from its zone id and abort this block scan.
1209 */
1210 zone_id = page_zone_id(page);
1211 page_pfn = page_to_pfn(page);
1212 pfn = page_pfn & ~((1 << order) - 1);
1213 end_pfn = pfn + (1 << order);
1214 for (; pfn < end_pfn; pfn++) {
1215 struct page *cursor_page;
1216
1217 /* The target page is in the block, ignore it. */
1218 if (unlikely(pfn == page_pfn))
1219 continue;
1220
1221 /* Avoid holes within the zone. */
1222 if (unlikely(!pfn_valid_within(pfn)))
1223 break;
1224
1225 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001226
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001227 /* Check that we have not crossed a zone boundary. */
1228 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001229 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001230
1231 /*
1232 * If we don't have enough swap space, reclaiming of
1233 * anon page which don't already have a swap slot is
1234 * pointless.
1235 */
Hugh Dickins043bcbe2012-01-10 15:08:33 -08001236 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001237 !PageSwapCache(cursor_page))
1238 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001239
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001240 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andrea Arcangeli50134732012-01-12 17:19:29 -08001241 unsigned int isolated_pages;
1242
Johannes Weiner925b7672012-01-12 17:18:15 -08001243 mem_cgroup_lru_del(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001244 list_move(&cursor_page->lru, dst);
Andrea Arcangeli50134732012-01-12 17:19:29 -08001245 isolated_pages = hpage_nr_pages(cursor_page);
1246 nr_taken += isolated_pages;
1247 nr_lumpy_taken += isolated_pages;
Mel Gormana8a94d12010-08-09 17:19:17 -07001248 if (PageDirty(cursor_page))
Andrea Arcangeli50134732012-01-12 17:19:29 -08001249 nr_lumpy_dirty += isolated_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001250 scan++;
Andrea Arcangeli50134732012-01-12 17:19:29 -08001251 pfn += isolated_pages - 1;
Mel Gormana8a94d12010-08-09 17:19:17 -07001252 } else {
Andrea Arcangelid179e842011-06-15 15:08:51 -07001253 /*
1254 * Check if the page is freed already.
1255 *
1256 * We can't use page_count() as that
1257 * requires compound_head and we don't
1258 * have a pin on the page here. If a
1259 * page is tail, we may or may not
1260 * have isolated the head, so assume
1261 * it's not free, it'd be tricky to
1262 * track the head status without a
1263 * page pin.
1264 */
1265 if (!PageTail(cursor_page) &&
1266 !atomic_read(&cursor_page->_count))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001267 continue;
1268 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001269 }
1270 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001271
1272 /* If we break out of the loop above, lumpy reclaim failed */
1273 if (pfn < end_pfn)
1274 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 }
1276
Hugh Dickinsf6260122012-01-12 17:20:06 -08001277 *nr_scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001278
1279 trace_mm_vmscan_lru_isolate(order,
1280 nr_to_scan, scan,
1281 nr_taken,
1282 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
Tao Maea4d3492012-01-12 17:19:20 -08001283 mode, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 return nr_taken;
1285}
1286
Nick Piggin62695a82008-10-18 20:26:09 -07001287/**
1288 * isolate_lru_page - tries to isolate a page from its LRU list
1289 * @page: page to isolate from its LRU list
1290 *
1291 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1292 * vmstat statistic corresponding to whatever LRU list the page was on.
1293 *
1294 * Returns 0 if the page was removed from an LRU list.
1295 * Returns -EBUSY if the page was not on an LRU list.
1296 *
1297 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001298 * the active list, it will have PageActive set. If it was found on
1299 * the unevictable list, it will have the PageUnevictable bit set. That flag
1300 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001301 *
1302 * The vmstat statistic corresponding to the list on which the page was
1303 * found will be decremented.
1304 *
1305 * Restrictions:
1306 * (1) Must be called with an elevated refcount on the page. This is a
1307 * fundamentnal difference from isolate_lru_pages (which is called
1308 * without a stable reference).
1309 * (2) the lru_lock must not be held.
1310 * (3) interrupts must be enabled.
1311 */
1312int isolate_lru_page(struct page *page)
1313{
1314 int ret = -EBUSY;
1315
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001316 VM_BUG_ON(!page_count(page));
1317
Nick Piggin62695a82008-10-18 20:26:09 -07001318 if (PageLRU(page)) {
1319 struct zone *zone = page_zone(page);
1320
1321 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001322 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001323 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001324 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001325 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001326 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001327
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001328 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001329 }
1330 spin_unlock_irq(&zone->lru_lock);
1331 }
1332 return ret;
1333}
1334
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001335/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001336 * Are there way too many processes in the direct reclaim path already?
1337 */
1338static int too_many_isolated(struct zone *zone, int file,
1339 struct scan_control *sc)
1340{
1341 unsigned long inactive, isolated;
1342
1343 if (current_is_kswapd())
1344 return 0;
1345
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001346 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001347 return 0;
1348
1349 if (file) {
1350 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1351 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1352 } else {
1353 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1354 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1355 }
1356
1357 return isolated > inactive;
1358}
1359
Mel Gorman66635622010-08-09 17:19:30 -07001360static noinline_for_stack void
Hugh Dickins3f797682012-01-12 17:20:07 -08001361putback_inactive_pages(struct mem_cgroup_zone *mz,
1362 struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001363{
Johannes Weinerf16015f2012-01-12 17:17:52 -08001364 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
Hugh Dickins3f797682012-01-12 17:20:07 -08001365 struct zone *zone = mz->zone;
1366 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001367
Mel Gorman66635622010-08-09 17:19:30 -07001368 /*
1369 * Put back any unfreeable pages.
1370 */
Mel Gorman66635622010-08-09 17:19:30 -07001371 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001372 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001373 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001374
Mel Gorman66635622010-08-09 17:19:30 -07001375 VM_BUG_ON(PageLRU(page));
1376 list_del(&page->lru);
1377 if (unlikely(!page_evictable(page, NULL))) {
1378 spin_unlock_irq(&zone->lru_lock);
1379 putback_lru_page(page);
1380 spin_lock_irq(&zone->lru_lock);
1381 continue;
1382 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001383 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001384 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001385 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001386 if (is_active_lru(lru)) {
1387 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001388 int numpages = hpage_nr_pages(page);
1389 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001390 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001391 if (put_page_testzero(page)) {
1392 __ClearPageLRU(page);
1393 __ClearPageActive(page);
1394 del_page_from_lru_list(zone, page, lru);
1395
1396 if (unlikely(PageCompound(page))) {
1397 spin_unlock_irq(&zone->lru_lock);
1398 (*get_compound_page_dtor(page))(page);
1399 spin_lock_irq(&zone->lru_lock);
1400 } else
1401 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001402 }
1403 }
Mel Gorman66635622010-08-09 17:19:30 -07001404
Hugh Dickins3f797682012-01-12 17:20:07 -08001405 /*
1406 * To save our caller's stack, now use input list for pages to free.
1407 */
1408 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001409}
1410
Johannes Weinerf16015f2012-01-12 17:17:52 -08001411static noinline_for_stack void
1412update_isolated_counts(struct mem_cgroup_zone *mz,
Hugh Dickins3f797682012-01-12 17:20:07 -08001413 struct list_head *page_list,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001414 unsigned long *nr_anon,
Hugh Dickins3f797682012-01-12 17:20:07 -08001415 unsigned long *nr_file)
Mel Gorman1489fa12010-08-09 17:19:33 -07001416{
Hugh Dickins3f797682012-01-12 17:20:07 -08001417 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001418 struct zone *zone = mz->zone;
Mel Gorman1489fa12010-08-09 17:19:33 -07001419 unsigned int count[NR_LRU_LISTS] = { 0, };
Hugh Dickins3f797682012-01-12 17:20:07 -08001420 unsigned long nr_active = 0;
1421 struct page *page;
1422 int lru;
Mel Gorman1489fa12010-08-09 17:19:33 -07001423
Hugh Dickins3f797682012-01-12 17:20:07 -08001424 /*
1425 * Count pages and clear active flags
1426 */
1427 list_for_each_entry(page, page_list, lru) {
1428 int numpages = hpage_nr_pages(page);
1429 lru = page_lru_base_type(page);
1430 if (PageActive(page)) {
1431 lru += LRU_ACTIVE;
1432 ClearPageActive(page);
1433 nr_active += numpages;
1434 }
1435 count[lru] += numpages;
1436 }
1437
Mel Gorman1489fa12010-08-09 17:19:33 -07001438 __count_vm_events(PGDEACTIVATE, nr_active);
1439
1440 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1441 -count[LRU_ACTIVE_FILE]);
1442 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1443 -count[LRU_INACTIVE_FILE]);
1444 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1445 -count[LRU_ACTIVE_ANON]);
1446 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1447 -count[LRU_INACTIVE_ANON]);
1448
1449 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1450 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
Mel Gorman1489fa12010-08-09 17:19:33 -07001451
1452 reclaim_stat->recent_scanned[0] += *nr_anon;
1453 reclaim_stat->recent_scanned[1] += *nr_file;
1454}
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;
Minchan Kim4356f212011-10-31 17:06:47 -07001513 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001514 struct zone *zone = mz->zone;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001515
Rik van Riel35cd7812009-09-21 17:01:38 -07001516 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001517 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001518
1519 /* We are about to die and free our memory. Return now. */
1520 if (fatal_signal_pending(current))
1521 return SWAP_CLUSTER_MAX;
1522 }
1523
Mel Gormanf3a310b2011-01-13 15:46:00 -08001524 set_reclaim_mode(priority, sc, false);
Minchan Kim4356f212011-10-31 17:06:47 -07001525 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1526 reclaim_mode |= ISOLATE_ACTIVE;
1527
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001529
1530 if (!sc->may_unmap)
1531 reclaim_mode |= ISOLATE_UNMAPPED;
1532 if (!sc->may_writepage)
1533 reclaim_mode |= ISOLATE_CLEAN;
1534
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Hugh Dickinsf6260122012-01-12 17:20:06 -08001537 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list,
1538 &nr_scanned, sc->order,
1539 reclaim_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 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001549
Mel Gorman66635622010-08-09 17:19:30 -07001550 if (nr_taken == 0) {
1551 spin_unlock_irq(&zone->lru_lock);
1552 return 0;
1553 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001554
Hugh Dickins3f797682012-01-12 17:20:07 -08001555 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1556
1557 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1558 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001559
1560 spin_unlock_irq(&zone->lru_lock);
1561
Johannes Weinerf16015f2012-01-12 17:17:52 -08001562 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
Mel Gorman92df3a72011-10-31 17:07:56 -07001563 &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001564
Wu Fengguange31f36982010-08-09 17:20:01 -07001565 /* Check if we should syncronously wait for writeback */
1566 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001567 set_reclaim_mode(priority, sc, true);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001568 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
Mel Gorman92df3a72011-10-31 17:07:56 -07001569 priority, &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001570 }
1571
Hugh Dickins3f797682012-01-12 17:20:07 -08001572 spin_lock_irq(&zone->lru_lock);
1573
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001574 if (current_is_kswapd())
1575 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1576 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1577
Hugh Dickins3f797682012-01-12 17:20:07 -08001578 putback_inactive_pages(mz, &page_list);
1579
1580 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1581 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1582
1583 spin_unlock_irq(&zone->lru_lock);
1584
1585 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001586
Mel Gorman92df3a72011-10-31 17:07:56 -07001587 /*
1588 * If reclaim is isolating dirty pages under writeback, it implies
1589 * that the long-lived page allocation rate is exceeding the page
1590 * laundering rate. Either the global limits are not being effective
1591 * at throttling processes due to the page distribution throughout
1592 * zones or there is heavy usage of a slow backing device. The
1593 * only option is to throttle from reclaim context which is not ideal
1594 * as there is no guarantee the dirtying process is throttled in the
1595 * same way balance_dirty_pages() manages.
1596 *
1597 * This scales the number of dirty pages that must be under writeback
1598 * before throttling depending on priority. It is a simple backoff
1599 * function that has the most effect in the range DEF_PRIORITY to
1600 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1601 * in trouble and reclaim is considered to be in trouble.
1602 *
1603 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1604 * DEF_PRIORITY-1 50% must be PageWriteback
1605 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1606 * ...
1607 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1608 * isolated page is PageWriteback
1609 */
1610 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1611 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1612
Mel Gormane11da5b2010-10-26 14:21:40 -07001613 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1614 zone_idx(zone),
1615 nr_scanned, nr_reclaimed,
1616 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001617 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001618 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619}
1620
Martin Bligh3bb1a852006-10-28 10:38:24 -07001621/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 * This moves pages from the active list to the inactive list.
1623 *
1624 * We move them the other way if the page is referenced by one or more
1625 * processes, from rmap.
1626 *
1627 * If the pages are mostly unmapped, the processing is fast and it is
1628 * appropriate to hold zone->lru_lock across the whole operation. But if
1629 * the pages are mapped, the processing is slow (page_referenced()) so we
1630 * should drop zone->lru_lock around each page. It's impossible to balance
1631 * this, so instead we remove the pages from the LRU while processing them.
1632 * It is safe to rely on PG_active against the non-LRU pages in here because
1633 * nobody will play with that bit on a non-LRU page.
1634 *
1635 * The downside is that we have to touch page->_count against each page.
1636 * But we had to alter page->flags anyway.
1637 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001638
Wu Fengguang3eb41402009-06-16 15:33:13 -07001639static void move_active_pages_to_lru(struct zone *zone,
1640 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001641 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001642 enum lru_list lru)
1643{
1644 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001645 struct page *page;
1646
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001647 if (buffer_heads_over_limit) {
1648 spin_unlock_irq(&zone->lru_lock);
1649 list_for_each_entry(page, list, lru) {
1650 if (page_has_private(page) && trylock_page(page)) {
1651 if (page_has_private(page))
1652 try_to_release_page(page, 0);
1653 unlock_page(page);
1654 }
1655 }
1656 spin_lock_irq(&zone->lru_lock);
1657 }
Wu Fengguang3eb41402009-06-16 15:33:13 -07001658
1659 while (!list_empty(list)) {
Johannes Weiner925b7672012-01-12 17:18:15 -08001660 struct lruvec *lruvec;
1661
Wu Fengguang3eb41402009-06-16 15:33:13 -07001662 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001663
1664 VM_BUG_ON(PageLRU(page));
1665 SetPageLRU(page);
1666
Johannes Weiner925b7672012-01-12 17:18:15 -08001667 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1668 list_move(&page->lru, &lruvec->lists[lru]);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001669 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001670
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001671 if (put_page_testzero(page)) {
1672 __ClearPageLRU(page);
1673 __ClearPageActive(page);
1674 del_page_from_lru_list(zone, page, lru);
1675
1676 if (unlikely(PageCompound(page))) {
1677 spin_unlock_irq(&zone->lru_lock);
1678 (*get_compound_page_dtor(page))(page);
1679 spin_lock_irq(&zone->lru_lock);
1680 } else
1681 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001682 }
1683 }
1684 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1685 if (!is_active_lru(lru))
1686 __count_vm_events(PGDEACTIVATE, pgmoved);
1687}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001688
Hugh Dickinsf6260122012-01-12 17:20:06 -08001689static void shrink_active_list(unsigned long nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001690 struct mem_cgroup_zone *mz,
1691 struct scan_control *sc,
1692 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001694 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001695 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001696 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001698 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001699 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 struct page *page;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001701 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001702 unsigned long nr_rotated = 0;
Minchan Kimf80c0672011-10-31 17:06:55 -07001703 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001704 struct zone *zone = mz->zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705
1706 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001707
1708 if (!sc->may_unmap)
1709 reclaim_mode |= ISOLATE_UNMAPPED;
1710 if (!sc->may_writepage)
1711 reclaim_mode |= ISOLATE_CLEAN;
1712
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001714
Hugh Dickinsf6260122012-01-12 17:20:06 -08001715 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold,
1716 &nr_scanned, sc->order,
1717 reclaim_mode, 1, file);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001718 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001719 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001720
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001721 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001722
Hugh Dickinsf6260122012-01-12 17:20:06 -08001723 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001724 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001725 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001726 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001727 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001728 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 spin_unlock_irq(&zone->lru_lock);
1730
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 while (!list_empty(&l_hold)) {
1732 cond_resched();
1733 page = lru_to_page(&l_hold);
1734 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001735
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001736 if (unlikely(!page_evictable(page, NULL))) {
1737 putback_lru_page(page);
1738 continue;
1739 }
1740
Johannes Weinerf16015f2012-01-12 17:17:52 -08001741 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001742 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001743 /*
1744 * Identify referenced, file-backed active pages and
1745 * give them one more trip around the active list. So
1746 * that executable code get better chances to stay in
1747 * memory under moderate memory pressure. Anon pages
1748 * are not likely to be evicted by use-once streaming
1749 * IO, plus JVM can create lots of anon VM_EXEC pages,
1750 * so we ignore them here.
1751 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001752 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001753 list_add(&page->lru, &l_active);
1754 continue;
1755 }
1756 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001757
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001758 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 list_add(&page->lru, &l_inactive);
1760 }
1761
Andrew Mortonb5557492009-01-06 14:40:13 -08001762 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001763 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001764 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001765 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001766 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001767 * Count referenced pages from currently used mappings as rotated,
1768 * even though only some of them are actually re-activated. This
1769 * helps balance scan pressure between file and anonymous pages in
1770 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001771 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001772 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001773
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001774 move_active_pages_to_lru(zone, &l_active, &l_hold,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001775 LRU_ACTIVE + file * LRU_FILE);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001776 move_active_pages_to_lru(zone, &l_inactive, &l_hold,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001777 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001778 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001779 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001780
1781 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782}
1783
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001784#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001785static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001786{
1787 unsigned long active, inactive;
1788
1789 active = zone_page_state(zone, NR_ACTIVE_ANON);
1790 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1791
1792 if (inactive * zone->inactive_ratio < active)
1793 return 1;
1794
1795 return 0;
1796}
1797
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001798/**
1799 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1800 * @zone: zone to check
1801 * @sc: scan control of this context
1802 *
1803 * Returns true if the zone does not have enough inactive anon pages,
1804 * meaning some active anon pages need to be deactivated.
1805 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001806static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001807{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001808 /*
1809 * If we don't have swap space, anonymous page deactivation
1810 * is pointless.
1811 */
1812 if (!total_swap_pages)
1813 return 0;
1814
Johannes Weinerf16015f2012-01-12 17:17:52 -08001815 if (!scanning_global_lru(mz))
1816 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1817 mz->zone);
1818
1819 return inactive_anon_is_low_global(mz->zone);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001820}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001821#else
Johannes Weinerf16015f2012-01-12 17:17:52 -08001822static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001823{
1824 return 0;
1825}
1826#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001827
Rik van Riel56e49d22009-06-16 15:32:28 -07001828static int inactive_file_is_low_global(struct zone *zone)
1829{
1830 unsigned long active, inactive;
1831
1832 active = zone_page_state(zone, NR_ACTIVE_FILE);
1833 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1834
1835 return (active > inactive);
1836}
1837
1838/**
1839 * inactive_file_is_low - check if file pages need to be deactivated
Johannes Weinerf16015f2012-01-12 17:17:52 -08001840 * @mz: memory cgroup and zone to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001841 *
1842 * When the system is doing streaming IO, memory pressure here
1843 * ensures that active file pages get deactivated, until more
1844 * than half of the file pages are on the inactive list.
1845 *
1846 * Once we get to that situation, protect the system's working
1847 * set from being evicted by disabling active file page aging.
1848 *
1849 * This uses a different ratio than the anonymous pages, because
1850 * the page cache uses a use-once replacement algorithm.
1851 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001852static int inactive_file_is_low(struct mem_cgroup_zone *mz)
Rik van Riel56e49d22009-06-16 15:32:28 -07001853{
Johannes Weinerf16015f2012-01-12 17:17:52 -08001854 if (!scanning_global_lru(mz))
1855 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1856 mz->zone);
Rik van Riel56e49d22009-06-16 15:32:28 -07001857
Johannes Weinerf16015f2012-01-12 17:17:52 -08001858 return inactive_file_is_low_global(mz->zone);
Rik van Riel56e49d22009-06-16 15:32:28 -07001859}
1860
Johannes Weinerf16015f2012-01-12 17:17:52 -08001861static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
Rik van Rielb39415b2009-12-14 17:59:48 -08001862{
1863 if (file)
Johannes Weinerf16015f2012-01-12 17:17:52 -08001864 return inactive_file_is_low(mz);
Rik van Rielb39415b2009-12-14 17:59:48 -08001865 else
Johannes Weinerf16015f2012-01-12 17:17:52 -08001866 return inactive_anon_is_low(mz);
Rik van Rielb39415b2009-12-14 17:59:48 -08001867}
1868
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001869static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001870 struct mem_cgroup_zone *mz,
1871 struct scan_control *sc, int priority)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001872{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001873 int file = is_file_lru(lru);
1874
Rik van Rielb39415b2009-12-14 17:59:48 -08001875 if (is_active_lru(lru)) {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001876 if (inactive_list_is_low(mz, file))
1877 shrink_active_list(nr_to_scan, mz, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001878 return 0;
1879 }
1880
Johannes Weinerf16015f2012-01-12 17:17:52 -08001881 return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001882}
1883
Johannes Weinerf16015f2012-01-12 17:17:52 -08001884static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1885 struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001886{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001887 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001888 return vm_swappiness;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001889 return mem_cgroup_swappiness(mz->mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001890}
1891
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001893 * Determine how aggressively the anon and file LRU lists should be
1894 * scanned. The relative value of each set of LRU lists is determined
1895 * by looking at the fraction of the pages scanned we did rotate back
1896 * onto the active list instead of evict.
1897 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001898 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001899 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001900static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1901 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001902{
1903 unsigned long anon, file, free;
1904 unsigned long anon_prio, file_prio;
1905 unsigned long ap, fp;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001906 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001907 u64 fraction[2], denominator;
Hugh Dickins41113042012-01-12 17:20:01 -08001908 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001909 int noswap = 0;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001910 bool force_scan = false;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001911
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001912 /*
1913 * If the zone or memcg is small, nr[l] can be 0. This
1914 * results in no scanning on this priority and a potential
1915 * priority drop. Global direct reclaim can go to the next
1916 * zone and tends to have no problems. Global kswapd is for
1917 * zone balancing and it needs to scan a minimum amount. When
1918 * reclaiming for a memcg, a priority drop can cause high
1919 * latencies, so it's better to scan a minimum amount there as
1920 * well.
1921 */
Johannes Weinerb95a2f22012-01-12 17:18:06 -08001922 if (current_is_kswapd() && mz->zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001923 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001924 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001925 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001926
1927 /* If we have no swap space, do not bother scanning anon pages. */
1928 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1929 noswap = 1;
1930 fraction[0] = 0;
1931 fraction[1] = 1;
1932 denominator = 1;
1933 goto out;
1934 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001935
Johannes Weinerf16015f2012-01-12 17:17:52 -08001936 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1937 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1938 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1939 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001940
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001941 if (global_reclaim(sc)) {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001942 free = zone_page_state(mz->zone, NR_FREE_PAGES);
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001943 /* If we have very few page cache pages,
1944 force-scan anon pages. */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001945 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001946 fraction[0] = 1;
1947 fraction[1] = 0;
1948 denominator = 1;
1949 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001950 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001951 }
1952
1953 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001954 * With swappiness at 100, anonymous and file have the same priority.
1955 * This scanning priority is essentially the inverse of IO cost.
1956 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001957 anon_prio = vmscan_swappiness(mz, sc);
1958 file_prio = 200 - vmscan_swappiness(mz, sc);
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001959
1960 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001961 * OK, so we have swap space and a fair amount of page cache
1962 * pages. We use the recently rotated / recently scanned
1963 * ratios to determine how valuable each cache is.
1964 *
1965 * Because workloads change over time (and to avoid overflow)
1966 * we keep these statistics as a floating average, which ends
1967 * up weighing recent references more than old ones.
1968 *
1969 * anon in [0], file in [1]
1970 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08001971 spin_lock_irq(&mz->zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001972 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001973 reclaim_stat->recent_scanned[0] /= 2;
1974 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001975 }
1976
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001977 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001978 reclaim_stat->recent_scanned[1] /= 2;
1979 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001980 }
1981
1982 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001983 * The amount of pressure on anon vs file pages is inversely
1984 * proportional to the fraction of recently scanned pages on
1985 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001986 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001987 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1988 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001989
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001990 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1991 fp /= reclaim_stat->recent_rotated[1] + 1;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001992 spin_unlock_irq(&mz->zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001993
Shaohua Li76a33fc2010-05-24 14:32:36 -07001994 fraction[0] = ap;
1995 fraction[1] = fp;
1996 denominator = ap + fp + 1;
1997out:
Hugh Dickins41113042012-01-12 17:20:01 -08001998 for_each_evictable_lru(lru) {
1999 int file = is_file_lru(lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07002000 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002001
Hugh Dickins41113042012-01-12 17:20:01 -08002002 scan = zone_nr_lru_pages(mz, lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07002003 if (priority || noswap) {
2004 scan >>= priority;
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002005 if (!scan && force_scan)
2006 scan = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002007 scan = div64_u64(scan * fraction[file], denominator);
2008 }
Hugh Dickins41113042012-01-12 17:20:01 -08002009 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002010 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002011}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002012
2013/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08002014 * Reclaim/compaction depends on a number of pages being freed. To avoid
2015 * disruption to the system, a small number of order-0 pages continue to be
2016 * rotated and reclaimed in the normal fashion. However, by the time we get
2017 * back to the allocator and call try_to_compact_zone(), we ensure that
2018 * there are enough free pages for it to be likely successful
2019 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002020static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002021 unsigned long nr_reclaimed,
2022 unsigned long nr_scanned,
2023 struct scan_control *sc)
2024{
2025 unsigned long pages_for_compaction;
2026 unsigned long inactive_lru_pages;
2027
2028 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08002029 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002030 return false;
2031
Mel Gorman28765922011-02-25 14:44:20 -08002032 /* Consider stopping depending on scan and reclaim activity */
2033 if (sc->gfp_mask & __GFP_REPEAT) {
2034 /*
2035 * For __GFP_REPEAT allocations, stop reclaiming if the
2036 * full LRU list has been scanned and we are still failing
2037 * to reclaim pages. This full LRU scan is potentially
2038 * expensive but a __GFP_REPEAT caller really wants to succeed
2039 */
2040 if (!nr_reclaimed && !nr_scanned)
2041 return false;
2042 } else {
2043 /*
2044 * For non-__GFP_REPEAT allocations which can presumably
2045 * fail without consequence, stop if we failed to reclaim
2046 * any pages from the last SWAP_CLUSTER_MAX number of
2047 * pages that were scanned. This will return to the
2048 * caller faster at the risk reclaim/compaction and
2049 * the resulting allocation attempt fails
2050 */
2051 if (!nr_reclaimed)
2052 return false;
2053 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002054
2055 /*
2056 * If we have not reclaimed enough pages for compaction and the
2057 * inactive lists are large enough, continue reclaiming
2058 */
2059 pages_for_compaction = (2UL << sc->order);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002060 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
Minchan Kim86cfd3a2012-01-10 15:08:18 -08002061 if (nr_swap_pages > 0)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002062 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002063 if (sc->nr_reclaimed < pages_for_compaction &&
2064 inactive_lru_pages > pages_for_compaction)
2065 return true;
2066
2067 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002068 switch (compaction_suitable(mz->zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002069 case COMPACT_PARTIAL:
2070 case COMPACT_CONTINUE:
2071 return false;
2072 default:
2073 return true;
2074 }
2075}
2076
2077/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2079 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002080static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2081 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082{
Christoph Lameterb69408e2008-10-18 20:26:14 -07002083 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08002084 unsigned long nr_to_scan;
Hugh Dickins41113042012-01-12 17:20:01 -08002085 enum lru_list lru;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002086 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002087 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Shaohua Li3da367c2011-10-31 17:07:03 -07002088 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089
Mel Gorman3e7d3442011-01-13 15:45:56 -08002090restart:
2091 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002092 nr_scanned = sc->nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002093 get_scan_count(mz, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002094
Shaohua Li3da367c2011-10-31 17:07:03 -07002095 blk_start_plug(&plug);
Rik van Riel556adec2008-10-18 20:26:34 -07002096 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2097 nr[LRU_INACTIVE_FILE]) {
Hugh Dickins41113042012-01-12 17:20:01 -08002098 for_each_evictable_lru(lru) {
2099 if (nr[lru]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08002100 nr_to_scan = min_t(unsigned long,
Hugh Dickins41113042012-01-12 17:20:01 -08002101 nr[lru], SWAP_CLUSTER_MAX);
2102 nr[lru] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103
Hugh Dickins41113042012-01-12 17:20:01 -08002104 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002105 mz, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 }
Rik van Riela79311c2009-01-06 14:40:01 -08002108 /*
2109 * On large memory systems, scan >> priority can become
2110 * really large. This is fine for the starting priority;
2111 * we want to put equal scanning pressure on each zone.
2112 * However, if the VM has a harder time of freeing pages,
2113 * with multiple processes reclaiming pages, the total
2114 * freeing target can get unreasonably large.
2115 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08002116 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08002117 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 }
Shaohua Li3da367c2011-10-31 17:07:03 -07002119 blk_finish_plug(&plug);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002120 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002121
Rik van Riel556adec2008-10-18 20:26:34 -07002122 /*
2123 * Even if we did not try to evict anon pages at all, we want to
2124 * rebalance the anon lru active/inactive ratio.
2125 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002126 if (inactive_anon_is_low(mz))
2127 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
Rik van Riel556adec2008-10-18 20:26:34 -07002128
Mel Gorman3e7d3442011-01-13 15:45:56 -08002129 /* reclaim/compaction might need reclaim to continue */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002130 if (should_continue_reclaim(mz, nr_reclaimed,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002131 sc->nr_scanned - nr_scanned, sc))
2132 goto restart;
2133
Andrew Morton232ea4d2007-02-28 20:13:21 -08002134 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135}
2136
Johannes Weinerf16015f2012-01-12 17:17:52 -08002137static void shrink_zone(int priority, struct zone *zone,
2138 struct scan_control *sc)
2139{
Johannes Weiner56600482012-01-12 17:17:59 -08002140 struct mem_cgroup *root = sc->target_mem_cgroup;
2141 struct mem_cgroup_reclaim_cookie reclaim = {
Johannes Weinerf16015f2012-01-12 17:17:52 -08002142 .zone = zone,
Johannes Weiner56600482012-01-12 17:17:59 -08002143 .priority = priority,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002144 };
Johannes Weiner56600482012-01-12 17:17:59 -08002145 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002146
Johannes Weiner56600482012-01-12 17:17:59 -08002147 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2148 do {
2149 struct mem_cgroup_zone mz = {
2150 .mem_cgroup = memcg,
2151 .zone = zone,
2152 };
2153
2154 shrink_mem_cgroup_zone(priority, &mz, sc);
2155 /*
2156 * Limit reclaim has historically picked one memcg and
2157 * scanned it with decreasing priority levels until
2158 * nr_to_reclaim had been reclaimed. This priority
2159 * cycle is thus over after a single memcg.
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002160 *
2161 * Direct reclaim and kswapd, on the other hand, have
2162 * to scan all memory cgroups to fulfill the overall
2163 * scan target for the zone.
Johannes Weiner56600482012-01-12 17:17:59 -08002164 */
2165 if (!global_reclaim(sc)) {
2166 mem_cgroup_iter_break(root, memcg);
2167 break;
2168 }
2169 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2170 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002171}
2172
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002173/* Returns true if compaction should go ahead for a high-order request */
2174static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2175{
2176 unsigned long balance_gap, watermark;
2177 bool watermark_ok;
2178
2179 /* Do not consider compaction for orders reclaim is meant to satisfy */
2180 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2181 return false;
2182
2183 /*
2184 * Compaction takes time to run and there are potentially other
2185 * callers using the pages just freed. Continue reclaiming until
2186 * there is a buffer of free pages available to give compaction
2187 * a reasonable chance of completing and allocating the page
2188 */
2189 balance_gap = min(low_wmark_pages(zone),
2190 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2191 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2192 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2193 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2194
2195 /*
2196 * If compaction is deferred, reclaim up to a point where
2197 * compaction will have a chance of success when re-enabled
2198 */
2199 if (compaction_deferred(zone))
2200 return watermark_ok;
2201
2202 /* If compaction is not ready to start, keep reclaiming */
2203 if (!compaction_suitable(zone, sc->order))
2204 return false;
2205
2206 return watermark_ok;
2207}
2208
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209/*
2210 * This is the direct reclaim path, for page-allocating processes. We only
2211 * try to reclaim pages from zones which will satisfy the caller's allocation
2212 * request.
2213 *
Mel Gorman41858962009-06-16 15:32:12 -07002214 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2215 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2217 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002218 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2219 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2220 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 * If a zone is deemed to be full of pinned pages then just give it a light
2223 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002224 *
2225 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002226 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002227 * the caller that it should consider retrying the allocation instead of
2228 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 */
Mel Gormane0c23272011-10-31 17:09:33 -07002230static bool shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08002231 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232{
Mel Gormandd1a2392008-04-28 02:12:17 -07002233 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002234 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002235 unsigned long nr_soft_reclaimed;
2236 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002237 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002238
Mel Gormand4debc62010-08-09 17:19:29 -07002239 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2240 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002241 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002243 /*
2244 * Take care memory controller reclaiming has small influence
2245 * to global LRU.
2246 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002247 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002248 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2249 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002250 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002251 continue; /* Let kswapd poll it */
Rik van Riele0887c12011-10-31 17:09:31 -07002252 if (COMPACTION_BUILD) {
2253 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002254 * If we already have plenty of memory free for
2255 * compaction in this zone, don't free any more.
2256 * Even though compaction is invoked for any
2257 * non-zero order, only frequent costly order
2258 * reclamation is disruptive enough to become a
2259 * noticable problem, like transparent huge page
2260 * allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002261 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002262 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002263 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002264 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002265 }
Rik van Riele0887c12011-10-31 17:09:31 -07002266 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002267 /*
2268 * This steals pages from memory cgroups over softlimit
2269 * and returns the number of reclaimed pages and
2270 * scanned pages. This works for global memory pressure
2271 * and balancing, not for a memcg's limit.
2272 */
2273 nr_soft_scanned = 0;
2274 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2275 sc->order, sc->gfp_mask,
2276 &nr_soft_scanned);
2277 sc->nr_reclaimed += nr_soft_reclaimed;
2278 sc->nr_scanned += nr_soft_scanned;
2279 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002280 }
Nick Piggin408d8542006-09-25 23:31:27 -07002281
Rik van Riela79311c2009-01-06 14:40:01 -08002282 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 }
Mel Gormane0c23272011-10-31 17:09:33 -07002284
Mel Gorman0cee34f2012-01-12 17:19:49 -08002285 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002286}
2287
2288static bool zone_reclaimable(struct zone *zone)
2289{
2290 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2291}
2292
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002293/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002294static bool all_unreclaimable(struct zonelist *zonelist,
2295 struct scan_control *sc)
2296{
2297 struct zoneref *z;
2298 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002299
2300 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2301 gfp_zone(sc->gfp_mask), sc->nodemask) {
2302 if (!populated_zone(zone))
2303 continue;
2304 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2305 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002306 if (!zone->all_unreclaimable)
2307 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002308 }
2309
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002310 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002312
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313/*
2314 * This is the main entry point to direct page reclaim.
2315 *
2316 * If a full scan of the inactive list fails to free enough memory then we
2317 * are "out of memory" and something needs to be killed.
2318 *
2319 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2320 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002321 * caller can't do much about. We kick the writeback threads and take explicit
2322 * naps in the hope that some of these pages can be written. But if the
2323 * allocating task holds filesystem locks which prevent writeout this might not
2324 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002325 *
2326 * returns: 0, if no pages reclaimed
2327 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 */
Mel Gormandac1d272008-04-28 02:12:12 -07002329static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002330 struct scan_control *sc,
2331 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332{
2333 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002334 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002336 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002337 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002338 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002339 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340
Miao Xiec0ff7452010-05-24 14:32:08 -07002341 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002342 delayacct_freepages_start();
2343
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002344 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002345 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
2347 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002348 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002349 if (!priority)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002350 disable_swap_token(sc->target_mem_cgroup);
Mel Gorman0cee34f2012-01-12 17:19:49 -08002351 aborted_reclaim = shrink_zones(priority, zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002352
Balbir Singh66e17072008-02-07 00:13:56 -08002353 /*
2354 * Don't shrink slabs when reclaiming memory from
2355 * over limit cgroups
2356 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002357 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002358 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002359 for_each_zone_zonelist(zone, z, zonelist,
2360 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002361 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2362 continue;
2363
2364 lru_pages += zone_reclaimable_pages(zone);
2365 }
2366
Ying Han1495f232011-05-24 17:12:27 -07002367 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002368 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002369 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002370 reclaim_state->reclaimed_slab = 0;
2371 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 }
Balbir Singh66e17072008-02-07 00:13:56 -08002373 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002374 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
2377 /*
2378 * Try to write back as many pages as we just scanned. This
2379 * tends to cause slow streaming writers to write data to the
2380 * disk smoothly, at the dirtying rate, which is nice. But
2381 * that's undesirable in laptop mode, where we *want* lumpy
2382 * writeout. So in laptop mode, write out the whole world.
2383 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002384 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2385 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002386 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2387 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002388 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 }
2390
2391 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002392 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002393 priority < DEF_PRIORITY - 2) {
2394 struct zone *preferred_zone;
2395
2396 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002397 &cpuset_current_mems_allowed,
2398 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002399 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2400 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002404 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002405 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002406
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002407 if (sc->nr_reclaimed)
2408 return sc->nr_reclaimed;
2409
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002410 /*
2411 * As hibernation is going on, kswapd is freezed so that it can't mark
2412 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2413 * check.
2414 */
2415 if (oom_killer_disabled)
2416 return 0;
2417
Mel Gorman0cee34f2012-01-12 17:19:49 -08002418 /* Aborted reclaim to try compaction? don't OOM, then */
2419 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002420 return 1;
2421
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002422 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002423 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002424 return 1;
2425
2426 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427}
2428
Mel Gormandac1d272008-04-28 02:12:12 -07002429unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002430 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002431{
Mel Gorman33906bc2010-08-09 17:19:16 -07002432 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002433 struct scan_control sc = {
2434 .gfp_mask = gfp_mask,
2435 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002436 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002437 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002438 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002439 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002440 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002441 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002442 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002443 struct shrink_control shrink = {
2444 .gfp_mask = sc.gfp_mask,
2445 };
Balbir Singh66e17072008-02-07 00:13:56 -08002446
Mel Gorman33906bc2010-08-09 17:19:16 -07002447 trace_mm_vmscan_direct_reclaim_begin(order,
2448 sc.may_writepage,
2449 gfp_mask);
2450
Ying Hana09ed5e2011-05-24 17:12:26 -07002451 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002452
2453 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2454
2455 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002456}
2457
Balbir Singh00f0b822008-03-04 14:28:39 -08002458#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002459
Johannes Weiner72835c82012-01-12 17:18:32 -08002460unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002461 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002462 struct zone *zone,
2463 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002464{
2465 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002466 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002467 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002468 .may_writepage = !laptop_mode,
2469 .may_unmap = 1,
2470 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002471 .order = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002472 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002473 };
Johannes Weiner56600482012-01-12 17:17:59 -08002474 struct mem_cgroup_zone mz = {
Johannes Weiner72835c82012-01-12 17:18:32 -08002475 .mem_cgroup = memcg,
Johannes Weiner56600482012-01-12 17:17:59 -08002476 .zone = zone,
2477 };
Ying Han0ae5e892011-05-26 16:25:25 -07002478
Balbir Singh4e416952009-09-23 15:56:39 -07002479 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2480 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002481
2482 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2483 sc.may_writepage,
2484 sc.gfp_mask);
2485
Balbir Singh4e416952009-09-23 15:56:39 -07002486 /*
2487 * NOTE: Although we can get the priority field, using it
2488 * here is not a good idea, since it limits the pages we can scan.
2489 * if we don't reclaim here, the shrink_zone from balance_pgdat
2490 * will pick up pages from other mem cgroup's as well. We hack
2491 * the priority and make it zero.
2492 */
Johannes Weiner56600482012-01-12 17:17:59 -08002493 shrink_mem_cgroup_zone(0, &mz, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002494
2495 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2496
Ying Han0ae5e892011-05-26 16:25:25 -07002497 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002498 return sc.nr_reclaimed;
2499}
2500
Johannes Weiner72835c82012-01-12 17:18:32 -08002501unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002502 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002503 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002504{
Balbir Singh4e416952009-09-23 15:56:39 -07002505 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002506 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002507 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002508 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002509 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002510 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002511 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002512 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002513 .order = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002514 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002515 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002516 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2517 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2518 };
2519 struct shrink_control shrink = {
2520 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002521 };
Balbir Singh66e17072008-02-07 00:13:56 -08002522
Ying Han889976d2011-05-26 16:25:33 -07002523 /*
2524 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2525 * take care of from where we get pages. So the node where we start the
2526 * scan does not need to be the current node.
2527 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002528 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002529
2530 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002531
2532 trace_mm_vmscan_memcg_reclaim_begin(0,
2533 sc.may_writepage,
2534 sc.gfp_mask);
2535
Ying Hana09ed5e2011-05-24 17:12:26 -07002536 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002537
2538 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2539
2540 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002541}
2542#endif
2543
Johannes Weinerf16015f2012-01-12 17:17:52 -08002544static void age_active_anon(struct zone *zone, struct scan_control *sc,
2545 int priority)
2546{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002547 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002548
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002549 if (!total_swap_pages)
2550 return;
2551
2552 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2553 do {
2554 struct mem_cgroup_zone mz = {
2555 .mem_cgroup = memcg,
2556 .zone = zone,
2557 };
2558
2559 if (inactive_anon_is_low(&mz))
2560 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2561 sc, priority, 0);
2562
2563 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2564 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002565}
2566
Mel Gorman1741c872011-01-13 15:46:21 -08002567/*
2568 * pgdat_balanced is used when checking if a node is balanced for high-order
2569 * allocations. Only zones that meet watermarks and are in a zone allowed
2570 * by the callers classzone_idx are added to balanced_pages. The total of
2571 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2572 * for the node to be considered balanced. Forcing all zones to be balanced
2573 * for high orders can cause excessive reclaim when there are imbalanced zones.
2574 * The choice of 25% is due to
2575 * o a 16M DMA zone that is balanced will not balance a zone on any
2576 * reasonable sized machine
2577 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002578 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002579 * would need to be at least 256M for it to be balance a whole node.
2580 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2581 * to balance a node on its own. These seemed like reasonable ratios.
2582 */
2583static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2584 int classzone_idx)
2585{
2586 unsigned long present_pages = 0;
2587 int i;
2588
2589 for (i = 0; i <= classzone_idx; i++)
2590 present_pages += pgdat->node_zones[i].present_pages;
2591
Shaohua Li4746efd2011-07-19 08:49:26 -07002592 /* A special case here: if zone has no page, we think it's balanced */
2593 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002594}
2595
Mel Gormanf50de2d2009-12-14 17:58:53 -08002596/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002597static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2598 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002599{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002600 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002601 unsigned long balanced = 0;
2602 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002603
2604 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2605 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002606 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002607
Mel Gorman0abdee22011-01-13 15:46:22 -08002608 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002609 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002610 struct zone *zone = pgdat->node_zones + i;
2611
2612 if (!populated_zone(zone))
2613 continue;
2614
Mel Gorman355b09c2011-01-13 15:46:24 -08002615 /*
2616 * balance_pgdat() skips over all_unreclaimable after
2617 * DEF_PRIORITY. Effectively, it considers them balanced so
2618 * they must be considered balanced here as well if kswapd
2619 * is to sleep
2620 */
2621 if (zone->all_unreclaimable) {
2622 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002623 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002624 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002625
Mel Gorman88f5acf2011-01-13 15:45:41 -08002626 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002627 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002628 all_zones_ok = false;
2629 else
2630 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002631 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002632
Mel Gorman1741c872011-01-13 15:46:21 -08002633 /*
2634 * For high-order requests, the balanced zones must contain at least
2635 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2636 * must be balanced
2637 */
2638 if (order)
Johannes Weinerafc7e322011-05-24 17:11:09 -07002639 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002640 else
2641 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002642}
2643
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644/*
2645 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002646 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002648 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 *
2650 * There is special handling here for zones which are full of pinned pages.
2651 * This can happen if the pages are all mlocked, or if they are all used by
2652 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2653 * What we do is to detect the case where all pages in the zone have been
2654 * scanned twice and there has been zero successful reclaim. Mark the zone as
2655 * dead and from now on, only perform a short scan. Basically we're polling
2656 * the zone for when the problem goes away.
2657 *
2658 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002659 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2660 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2661 * lower zones regardless of the number of free pages in the lower zones. This
2662 * interoperates with the page allocator fallback scheme to ensure that aging
2663 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 */
Mel Gorman99504742011-01-13 15:46:20 -08002665static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002666 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002669 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 int priority;
2671 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002672 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002673 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002675 unsigned long nr_soft_reclaimed;
2676 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002677 struct scan_control sc = {
2678 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002679 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002680 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002681 /*
2682 * kswapd doesn't want to be bailed out while reclaim. because
2683 * we want to put equal scanning pressure on each zone.
2684 */
2685 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002686 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002687 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002688 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002689 struct shrink_control shrink = {
2690 .gfp_mask = sc.gfp_mask,
2691 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692loop_again:
2693 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002694 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002695 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002696 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002700 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002702 /* The swap token gets in the way of swapout... */
2703 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002704 disable_swap_token(NULL);
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002705
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002707 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002709 /*
2710 * Scan in the highmem->dma direction for the highest
2711 * zone which needs scanning
2712 */
2713 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2714 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002716 if (!populated_zone(zone))
2717 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002719 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002720 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
Rik van Riel556adec2008-10-18 20:26:34 -07002722 /*
2723 * Do some background aging of the anon list, to give
2724 * pages a chance to be referenced before reclaiming.
2725 */
Johannes Weinerf16015f2012-01-12 17:17:52 -08002726 age_active_anon(zone, &sc, priority);
Rik van Riel556adec2008-10-18 20:26:34 -07002727
Mel Gorman88f5acf2011-01-13 15:45:41 -08002728 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002729 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002730 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002731 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002732 } else {
2733 /* If balanced, clear the congested flag */
2734 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002737 if (i < 0)
2738 goto out;
2739
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 for (i = 0; i <= end_zone; i++) {
2741 struct zone *zone = pgdat->node_zones + i;
2742
Wu Fengguangadea02a2009-09-21 17:01:42 -07002743 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 }
2745
2746 /*
2747 * Now scan the zone in the dma->highmem direction, stopping
2748 * at the last zone which needs scanning.
2749 *
2750 * We do this because the page allocator works in the opposite
2751 * direction. This prevents the page allocator from allocating
2752 * pages behind kswapd's direction of progress, which would
2753 * cause too much scanning of the lower zones.
2754 */
2755 for (i = 0; i <= end_zone; i++) {
2756 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002757 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002758 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759
Con Kolivasf3fe6512006-01-06 00:11:15 -08002760 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 continue;
2762
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002763 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 continue;
2765
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002767
Ying Han0ae5e892011-05-26 16:25:25 -07002768 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002769 /*
2770 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002771 */
Ying Han0ae5e892011-05-26 16:25:25 -07002772 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2773 order, sc.gfp_mask,
2774 &nr_soft_scanned);
2775 sc.nr_reclaimed += nr_soft_reclaimed;
2776 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002777
Rik van Riel32a43302007-10-16 01:24:50 -07002778 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002779 * We put equal pressure on every zone, unless
2780 * one zone has way too many pages free
2781 * already. The "too many pages" is defined
2782 * as the high wmark plus a "gap" where the
2783 * gap is either the low watermark or 1%
2784 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002785 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002786 balance_gap = min(low_wmark_pages(zone),
2787 (zone->present_pages +
2788 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2789 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002790 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002791 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002792 end_zone, 0)) {
Rik van Riela79311c2009-01-06 14:40:01 -08002793 shrink_zone(priority, zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002794
Mel Gormand7868da2011-07-08 15:39:38 -07002795 reclaim_state->reclaimed_slab = 0;
2796 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2797 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2798 total_scanned += sc.nr_scanned;
2799
2800 if (nr_slab == 0 && !zone_reclaimable(zone))
2801 zone->all_unreclaimable = 1;
2802 }
2803
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 /*
2805 * If we've done a decent amount of scanning and
2806 * the reclaim ratio is low, start doing writepage
2807 * even in laptop mode
2808 */
2809 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002810 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002812
Mel Gorman215ddd62011-07-08 15:39:40 -07002813 if (zone->all_unreclaimable) {
2814 if (end_zone && end_zone == i)
2815 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002816 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002817 }
Mel Gormand7868da2011-07-08 15:39:38 -07002818
Mel Gorman88f5acf2011-01-13 15:45:41 -08002819 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002820 high_wmark_pages(zone), end_zone, 0)) {
2821 all_zones_ok = 0;
2822 /*
2823 * We are still under min water mark. This
2824 * means that we have a GFP_ATOMIC allocation
2825 * failure risk. Hurry up!
2826 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002827 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002828 min_wmark_pages(zone), end_zone, 0))
2829 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002830 } else {
2831 /*
2832 * If a zone reaches its high watermark,
2833 * consider it to be no longer congested. It's
2834 * possible there are dirty pages backed by
2835 * congested BDIs but as pressure is relieved,
2836 * spectulatively avoid congestion waits
2837 */
2838 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002839 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002840 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002841 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002842
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002844 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 break; /* kswapd: all done */
2846 /*
2847 * OK, kswapd is getting into trouble. Take a nap, then take
2848 * another pass across the zones.
2849 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002850 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2851 if (has_under_min_watermark_zone)
2852 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2853 else
2854 congestion_wait(BLK_RW_ASYNC, HZ/10);
2855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
2857 /*
2858 * We do this so kswapd doesn't build up large priorities for
2859 * example when it is freeing in parallel with allocators. It
2860 * matches the direct reclaim path behaviour in terms of impact
2861 * on zone->*_priority.
2862 */
Rik van Riela79311c2009-01-06 14:40:01 -08002863 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 break;
2865 }
2866out:
Mel Gorman99504742011-01-13 15:46:20 -08002867
2868 /*
2869 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002870 * high-order: Balanced zones must make up at least 25% of the node
2871 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002872 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002873 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002875
2876 try_to_freeze();
2877
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002878 /*
2879 * Fragmentation may mean that the system cannot be
2880 * rebalanced for high-order allocations in all zones.
2881 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2882 * it means the zones have been fully scanned and are still
2883 * not balanced. For high-order allocations, there is
2884 * little point trying all over again as kswapd may
2885 * infinite loop.
2886 *
2887 * Instead, recheck all watermarks at order-0 as they
2888 * are the most important. If watermarks are ok, kswapd will go
2889 * back to sleep. High-order users can still perform direct
2890 * reclaim if they wish.
2891 */
2892 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2893 order = sc.order = 0;
2894
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 goto loop_again;
2896 }
2897
Mel Gorman99504742011-01-13 15:46:20 -08002898 /*
2899 * If kswapd was reclaiming at a higher order, it has the option of
2900 * sleeping without all zones being balanced. Before it does, it must
2901 * ensure that the watermarks for order-0 on *all* zones are met and
2902 * that the congestion flags are cleared. The congestion flag must
2903 * be cleared as kswapd is the only mechanism that clears the flag
2904 * and it is potentially going to sleep here.
2905 */
2906 if (order) {
2907 for (i = 0; i <= end_zone; i++) {
2908 struct zone *zone = pgdat->node_zones + i;
2909
2910 if (!populated_zone(zone))
2911 continue;
2912
2913 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2914 continue;
2915
2916 /* Confirm the zone is balanced for order-0 */
2917 if (!zone_watermark_ok(zone, 0,
2918 high_wmark_pages(zone), 0, 0)) {
2919 order = sc.order = 0;
2920 goto loop_again;
2921 }
2922
2923 /* If balanced, clear the congested flag */
2924 zone_clear_flag(zone, ZONE_CONGESTED);
Shaohua Li16fb9512011-10-31 17:08:02 -07002925 if (i <= *classzone_idx)
2926 balanced += zone->present_pages;
Mel Gorman99504742011-01-13 15:46:20 -08002927 }
2928 }
2929
Mel Gorman0abdee22011-01-13 15:46:22 -08002930 /*
2931 * Return the order we were reclaiming at so sleeping_prematurely()
2932 * makes a decision on the order we were last reclaiming at. However,
2933 * if another caller entered the allocator slow path while kswapd
2934 * was awake, order will remain at the higher level
2935 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002936 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002937 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938}
2939
Mel Gormandc83edd2011-01-13 15:46:26 -08002940static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002941{
2942 long remaining = 0;
2943 DEFINE_WAIT(wait);
2944
2945 if (freezing(current) || kthread_should_stop())
2946 return;
2947
2948 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2949
2950 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002951 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002952 remaining = schedule_timeout(HZ/10);
2953 finish_wait(&pgdat->kswapd_wait, &wait);
2954 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2955 }
2956
2957 /*
2958 * After a short sleep, check if it was a premature sleep. If not, then
2959 * go fully to sleep until explicitly woken up.
2960 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002961 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002962 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2963
2964 /*
2965 * vmstat counters are not perfectly accurate and the estimated
2966 * value for counters such as NR_FREE_PAGES can deviate from the
2967 * true value by nr_online_cpus * threshold. To avoid the zone
2968 * watermarks being breached while under pressure, we reduce the
2969 * per-cpu vmstat threshold while kswapd is awake and restore
2970 * them before going back to sleep.
2971 */
2972 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2973 schedule();
2974 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2975 } else {
2976 if (remaining)
2977 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2978 else
2979 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2980 }
2981 finish_wait(&pgdat->kswapd_wait, &wait);
2982}
2983
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984/*
2985 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002986 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 *
2988 * This basically trickles out pages so that we have _some_
2989 * free memory available even if there is no other activity
2990 * that frees anything up. This is needed for things like routing
2991 * etc, where we otherwise might have all activity going on in
2992 * asynchronous contexts that cannot page things out.
2993 *
2994 * If there are applications that are active memory-allocators
2995 * (most normal use), this basically shouldn't matter.
2996 */
2997static int kswapd(void *p)
2998{
Mel Gorman215ddd62011-07-08 15:39:40 -07002999 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003000 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003001 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003002 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 pg_data_t *pgdat = (pg_data_t*)p;
3004 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003005
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 struct reclaim_state reclaim_state = {
3007 .reclaimed_slab = 0,
3008 };
Rusty Russella70f7302009-03-13 14:49:46 +10303009 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010
Nick Piggincf40bd12009-01-21 08:12:39 +01003011 lockdep_set_current_reclaim_state(GFP_KERNEL);
3012
Rusty Russell174596a2009-01-01 10:12:29 +10303013 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003014 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 current->reclaim_state = &reclaim_state;
3016
3017 /*
3018 * Tell the memory management that we're a "memory allocator",
3019 * and that if we need more memory we should get access to it
3020 * regardless (see "__alloc_pages()"). "kswapd" should
3021 * never get caught in the normal page freeing logic.
3022 *
3023 * (Kswapd normally doesn't need memory anyway, but sometimes
3024 * you need a small amount of memory in order to be able to
3025 * page out something else, and this flag essentially protects
3026 * us from recursively trying to free more memory as we're
3027 * trying to free the first piece of memory in the first place).
3028 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003029 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003030 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031
Mel Gorman215ddd62011-07-08 15:39:40 -07003032 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003033 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003034 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003035 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003037 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003038
Mel Gorman215ddd62011-07-08 15:39:40 -07003039 /*
3040 * If the last balance_pgdat was unsuccessful it's unlikely a
3041 * new request of a similar or harder type will succeed soon
3042 * so consider going to sleep on the basis we reclaimed at
3043 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003044 if (balanced_classzone_idx >= new_classzone_idx &&
3045 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003046 new_order = pgdat->kswapd_max_order;
3047 new_classzone_idx = pgdat->classzone_idx;
3048 pgdat->kswapd_max_order = 0;
3049 pgdat->classzone_idx = pgdat->nr_zones - 1;
3050 }
3051
Mel Gorman99504742011-01-13 15:46:20 -08003052 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 /*
3054 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003055 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 */
3057 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003058 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003060 kswapd_try_to_sleep(pgdat, balanced_order,
3061 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003063 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003064 new_order = order;
3065 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003066 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003067 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069
David Rientjes8fe23e02009-12-14 17:58:33 -08003070 ret = try_to_freeze();
3071 if (kthread_should_stop())
3072 break;
3073
3074 /*
3075 * We can speed up thawing tasks if we don't call balance_pgdat
3076 * after returning from the refrigerator
3077 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003078 if (!ret) {
3079 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003080 balanced_classzone_idx = classzone_idx;
3081 balanced_order = balance_pgdat(pgdat, order,
3082 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003083 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 }
3085 return 0;
3086}
3087
3088/*
3089 * A zone is low on free memory, so wake its kswapd task to service it.
3090 */
Mel Gorman99504742011-01-13 15:46:20 -08003091void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092{
3093 pg_data_t *pgdat;
3094
Con Kolivasf3fe6512006-01-06 00:11:15 -08003095 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 return;
3097
Paul Jackson02a0e532006-12-13 00:34:25 -08003098 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003100 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003101 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003102 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003103 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3104 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003105 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003107 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3108 return;
3109
3110 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003111 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112}
3113
Wu Fengguangadea02a2009-09-21 17:01:42 -07003114/*
3115 * The reclaimable count would be mostly accurate.
3116 * The less reclaimable pages may be
3117 * - mlocked pages, which will be moved to unevictable list when encountered
3118 * - mapped pages, which may require several travels to be reclaimed
3119 * - dirty pages, which is not "instantly" reclaimable
3120 */
3121unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003122{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003123 int nr;
3124
3125 nr = global_page_state(NR_ACTIVE_FILE) +
3126 global_page_state(NR_INACTIVE_FILE);
3127
3128 if (nr_swap_pages > 0)
3129 nr += global_page_state(NR_ACTIVE_ANON) +
3130 global_page_state(NR_INACTIVE_ANON);
3131
3132 return nr;
3133}
3134
3135unsigned long zone_reclaimable_pages(struct zone *zone)
3136{
3137 int nr;
3138
3139 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3140 zone_page_state(zone, NR_INACTIVE_FILE);
3141
3142 if (nr_swap_pages > 0)
3143 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3144 zone_page_state(zone, NR_INACTIVE_ANON);
3145
3146 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003147}
3148
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003149#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003151 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003152 * freed pages.
3153 *
3154 * Rather than trying to age LRUs the aim is to preserve the overall
3155 * LRU order by reclaiming preferentially
3156 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003158unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003160 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003161 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003162 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3163 .may_swap = 1,
3164 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003165 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003166 .nr_to_reclaim = nr_to_reclaim,
3167 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003168 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003170 struct shrink_control shrink = {
3171 .gfp_mask = sc.gfp_mask,
3172 };
3173 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003174 struct task_struct *p = current;
3175 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003177 p->flags |= PF_MEMALLOC;
3178 lockdep_set_current_reclaim_state(sc.gfp_mask);
3179 reclaim_state.reclaimed_slab = 0;
3180 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003181
Ying Hana09ed5e2011-05-24 17:12:26 -07003182 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003183
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003184 p->reclaim_state = NULL;
3185 lockdep_clear_current_reclaim_state();
3186 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003187
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003188 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003190#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192/* It's optimal to keep kswapds on the same CPUs as their memory, but
3193 not required for correctness. So if the last cpu in a node goes
3194 away, we get changed to run anywhere: as the first one comes back,
3195 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003196static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08003197 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003199 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003201 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003202 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003203 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303204 const struct cpumask *mask;
3205
3206 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003207
Rusty Russell3e597942009-01-01 10:12:24 +10303208 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003210 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 }
3212 }
3213 return NOTIFY_OK;
3214}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
Yasunori Goto3218ae12006-06-27 02:53:33 -07003216/*
3217 * This kswapd start function will be called by init and node-hot-add.
3218 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3219 */
3220int kswapd_run(int nid)
3221{
3222 pg_data_t *pgdat = NODE_DATA(nid);
3223 int ret = 0;
3224
3225 if (pgdat->kswapd)
3226 return 0;
3227
3228 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3229 if (IS_ERR(pgdat->kswapd)) {
3230 /* failure at boot is fatal */
3231 BUG_ON(system_state == SYSTEM_BOOTING);
3232 printk("Failed to start kswapd on node %d\n",nid);
3233 ret = -1;
3234 }
3235 return ret;
3236}
3237
David Rientjes8fe23e02009-12-14 17:58:33 -08003238/*
3239 * Called by memory hotplug when all memory in a node is offlined.
3240 */
3241void kswapd_stop(int nid)
3242{
3243 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3244
3245 if (kswapd)
3246 kthread_stop(kswapd);
3247}
3248
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249static int __init kswapd_init(void)
3250{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003251 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003252
Linus Torvalds1da177e2005-04-16 15:20:36 -07003253 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003254 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003255 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 hotcpu_notifier(cpu_callback, 0);
3257 return 0;
3258}
3259
3260module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003261
3262#ifdef CONFIG_NUMA
3263/*
3264 * Zone reclaim mode
3265 *
3266 * If non-zero call zone_reclaim when the number of free pages falls below
3267 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003268 */
3269int zone_reclaim_mode __read_mostly;
3270
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003271#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003272#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003273#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3274#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3275
Christoph Lameter9eeff232006-01-18 17:42:31 -08003276/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003277 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3278 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3279 * a zone.
3280 */
3281#define ZONE_RECLAIM_PRIORITY 4
3282
Christoph Lameter9eeff232006-01-18 17:42:31 -08003283/*
Christoph Lameter96146342006-07-03 00:24:13 -07003284 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3285 * occur.
3286 */
3287int sysctl_min_unmapped_ratio = 1;
3288
3289/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003290 * If the number of slab pages in a zone grows beyond this percentage then
3291 * slab reclaim needs to occur.
3292 */
3293int sysctl_min_slab_ratio = 5;
3294
Mel Gorman90afa5d2009-06-16 15:33:20 -07003295static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3296{
3297 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3298 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3299 zone_page_state(zone, NR_ACTIVE_FILE);
3300
3301 /*
3302 * It's possible for there to be more file mapped pages than
3303 * accounted for by the pages on the file LRU lists because
3304 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3305 */
3306 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3307}
3308
3309/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3310static long zone_pagecache_reclaimable(struct zone *zone)
3311{
3312 long nr_pagecache_reclaimable;
3313 long delta = 0;
3314
3315 /*
3316 * If RECLAIM_SWAP is set, then all file pages are considered
3317 * potentially reclaimable. Otherwise, we have to worry about
3318 * pages like swapcache and zone_unmapped_file_pages() provides
3319 * a better estimate
3320 */
3321 if (zone_reclaim_mode & RECLAIM_SWAP)
3322 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3323 else
3324 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3325
3326 /* If we can't clean pages, remove dirty pages from consideration */
3327 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3328 delta += zone_page_state(zone, NR_FILE_DIRTY);
3329
3330 /* Watch for any possible underflows due to delta */
3331 if (unlikely(delta > nr_pagecache_reclaimable))
3332 delta = nr_pagecache_reclaimable;
3333
3334 return nr_pagecache_reclaimable - delta;
3335}
3336
Christoph Lameter0ff38492006-09-25 23:31:52 -07003337/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003338 * Try to free up some pages from this zone through reclaim.
3339 */
Andrew Morton179e9632006-03-22 00:08:18 -08003340static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003341{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003342 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003343 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003344 struct task_struct *p = current;
3345 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08003346 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08003347 struct scan_control sc = {
3348 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07003349 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003350 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003351 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3352 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003353 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003354 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003355 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003356 struct shrink_control shrink = {
3357 .gfp_mask = sc.gfp_mask,
3358 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003359 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003360
Christoph Lameter9eeff232006-01-18 17:42:31 -08003361 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003362 /*
3363 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3364 * and we also need to be able to write out pages for RECLAIM_WRITE
3365 * and RECLAIM_SWAP.
3366 */
3367 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003368 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003369 reclaim_state.reclaimed_slab = 0;
3370 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003371
Mel Gorman90afa5d2009-06-16 15:33:20 -07003372 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003373 /*
3374 * Free memory by calling shrink zone with increasing
3375 * priorities until we have enough memory freed.
3376 */
3377 priority = ZONE_RECLAIM_PRIORITY;
3378 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003379 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003380 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003381 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003382 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003383
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003384 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3385 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003386 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003387 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003388 * many pages were freed in this zone. So we take the current
3389 * number of slab pages and shake the slab until it is reduced
3390 * by the same nr_pages that we used for reclaiming unmapped
3391 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003392 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003393 * Note that shrink_slab will free memory on all zones and may
3394 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003395 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003396 for (;;) {
3397 unsigned long lru_pages = zone_reclaimable_pages(zone);
3398
3399 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003400 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003401 break;
3402
3403 /* Freed enough memory */
3404 nr_slab_pages1 = zone_page_state(zone,
3405 NR_SLAB_RECLAIMABLE);
3406 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3407 break;
3408 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003409
3410 /*
3411 * Update nr_reclaimed by the number of slab pages we
3412 * reclaimed from this zone.
3413 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003414 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3415 if (nr_slab_pages1 < nr_slab_pages0)
3416 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003417 }
3418
Christoph Lameter9eeff232006-01-18 17:42:31 -08003419 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003420 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003421 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003422 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003423}
Andrew Morton179e9632006-03-22 00:08:18 -08003424
3425int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3426{
Andrew Morton179e9632006-03-22 00:08:18 -08003427 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003428 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003429
3430 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003431 * Zone reclaim reclaims unmapped file backed pages and
3432 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003433 *
Christoph Lameter96146342006-07-03 00:24:13 -07003434 * A small portion of unmapped file backed pages is needed for
3435 * file I/O otherwise pages read by file I/O will be immediately
3436 * thrown out if the zone is overallocated. So we do not reclaim
3437 * if less than a specified percentage of the zone is used by
3438 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003439 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003440 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3441 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003442 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003443
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003444 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003445 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003446
Andrew Morton179e9632006-03-22 00:08:18 -08003447 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003448 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003449 */
David Rientjesd773ed62007-10-16 23:26:01 -07003450 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003451 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003452
3453 /*
3454 * Only run zone reclaim on the local zone or on zones that do not
3455 * have associated processors. This will favor the local processor
3456 * over remote processors and spread off node memory allocations
3457 * as wide as possible.
3458 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003459 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003460 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003461 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003462
3463 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003464 return ZONE_RECLAIM_NOSCAN;
3465
David Rientjesd773ed62007-10-16 23:26:01 -07003466 ret = __zone_reclaim(zone, gfp_mask, order);
3467 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3468
Mel Gorman24cf725182009-06-16 15:33:23 -07003469 if (!ret)
3470 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3471
David Rientjesd773ed62007-10-16 23:26:01 -07003472 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003473}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003474#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003475
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003476/*
3477 * page_evictable - test whether a page is evictable
3478 * @page: the page to test
3479 * @vma: the VMA in which the page is or will be mapped, may be NULL
3480 *
3481 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003482 * lists vs unevictable list. The vma argument is !NULL when called from the
3483 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003484 *
3485 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003486 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003487 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003488 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003489 */
3490int page_evictable(struct page *page, struct vm_area_struct *vma)
3491{
3492
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003493 if (mapping_unevictable(page_mapping(page)))
3494 return 0;
3495
Nick Pigginb291f002008-10-18 20:26:44 -07003496 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3497 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003498
3499 return 1;
3500}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003501
3502/**
3503 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3504 * @page: page to check evictability and move to appropriate lru list
3505 * @zone: zone page is in
3506 *
3507 * Checks a page for evictability and moves the page to the appropriate
3508 * zone lru list.
3509 *
3510 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3511 * have PageUnevictable set.
3512 */
3513static void check_move_unevictable_page(struct page *page, struct zone *zone)
3514{
Johannes Weiner925b7672012-01-12 17:18:15 -08003515 struct lruvec *lruvec;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003516
Johannes Weiner925b7672012-01-12 17:18:15 -08003517 VM_BUG_ON(PageActive(page));
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003518retry:
3519 ClearPageUnevictable(page);
3520 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003521 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003522
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003523 __dec_zone_state(zone, NR_UNEVICTABLE);
Johannes Weiner925b7672012-01-12 17:18:15 -08003524 lruvec = mem_cgroup_lru_move_lists(zone, page,
3525 LRU_UNEVICTABLE, l);
3526 list_move(&page->lru, &lruvec->lists[l]);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003527 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3528 __count_vm_event(UNEVICTABLE_PGRESCUED);
3529 } else {
3530 /*
3531 * rotate unevictable list
3532 */
3533 SetPageUnevictable(page);
Johannes Weiner925b7672012-01-12 17:18:15 -08003534 lruvec = mem_cgroup_lru_move_lists(zone, page, LRU_UNEVICTABLE,
3535 LRU_UNEVICTABLE);
3536 list_move(&page->lru, &lruvec->lists[LRU_UNEVICTABLE]);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003537 if (page_evictable(page, NULL))
3538 goto retry;
3539 }
3540}
3541
3542/**
3543 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3544 * @mapping: struct address_space to scan for evictable pages
3545 *
3546 * Scan all pages in mapping. Check unevictable pages for
3547 * evictability and move them to the appropriate zone lru list.
3548 */
3549void scan_mapping_unevictable_pages(struct address_space *mapping)
3550{
3551 pgoff_t next = 0;
3552 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3553 PAGE_CACHE_SHIFT;
3554 struct zone *zone;
3555 struct pagevec pvec;
3556
3557 if (mapping->nrpages == 0)
3558 return;
3559
3560 pagevec_init(&pvec, 0);
3561 while (next < end &&
3562 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3563 int i;
3564 int pg_scanned = 0;
3565
3566 zone = NULL;
3567
3568 for (i = 0; i < pagevec_count(&pvec); i++) {
3569 struct page *page = pvec.pages[i];
3570 pgoff_t page_index = page->index;
3571 struct zone *pagezone = page_zone(page);
3572
3573 pg_scanned++;
3574 if (page_index > next)
3575 next = page_index;
3576 next++;
3577
3578 if (pagezone != zone) {
3579 if (zone)
3580 spin_unlock_irq(&zone->lru_lock);
3581 zone = pagezone;
3582 spin_lock_irq(&zone->lru_lock);
3583 }
3584
3585 if (PageLRU(page) && PageUnevictable(page))
3586 check_move_unevictable_page(page, zone);
3587 }
3588 if (zone)
3589 spin_unlock_irq(&zone->lru_lock);
3590 pagevec_release(&pvec);
3591
3592 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3593 }
3594
3595}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003596
Johannes Weiner264e56d2011-10-31 17:09:13 -07003597static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003598{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003599 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003600 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003601 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003602 "one, please send an email to linux-mm@kvack.org.\n",
3603 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003604}
3605
3606/*
3607 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3608 * all nodes' unevictable lists for evictable pages
3609 */
3610unsigned long scan_unevictable_pages;
3611
3612int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003613 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003614 size_t *length, loff_t *ppos)
3615{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003616 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003617 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003618 scan_unevictable_pages = 0;
3619 return 0;
3620}
3621
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003622#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003623/*
3624 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3625 * a specified node's per zone unevictable lists for evictable pages.
3626 */
3627
Kay Sievers10fbcf42011-12-21 14:48:43 -08003628static ssize_t read_scan_unevictable_node(struct device *dev,
3629 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003630 char *buf)
3631{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003632 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003633 return sprintf(buf, "0\n"); /* always zero; should fit... */
3634}
3635
Kay Sievers10fbcf42011-12-21 14:48:43 -08003636static ssize_t write_scan_unevictable_node(struct device *dev,
3637 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003638 const char *buf, size_t count)
3639{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003640 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003641 return 1;
3642}
3643
3644
Kay Sievers10fbcf42011-12-21 14:48:43 -08003645static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003646 read_scan_unevictable_node,
3647 write_scan_unevictable_node);
3648
3649int scan_unevictable_register_node(struct node *node)
3650{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003651 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003652}
3653
3654void scan_unevictable_unregister_node(struct node *node)
3655{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003656 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003657}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003658#endif