blob: 7ef69124fa3e5f4ef28baaed58a7e997d40155ab [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 Weinera6dc60f82009-03-31 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
Balbir Singh66e17072008-02-07 00:13:56 -0800106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -0700108 struct memcg_scanrecord *memcg_record;
Balbir Singh66e17072008-02-07 00:13:56 -0800109
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700110 /*
111 * Nodemask of nodes allowed by the caller. If NULL, all nodes
112 * are scanned.
113 */
114 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115};
116
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118
119#ifdef ARCH_HAS_PREFETCH
120#define prefetch_prev_lru_page(_page, _base, _field) \
121 do { \
122 if ((_page)->lru.prev != _base) { \
123 struct page *prev; \
124 \
125 prev = lru_to_page(&(_page->lru)); \
126 prefetch(&prev->_field); \
127 } \
128 } while (0)
129#else
130#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
131#endif
132
133#ifdef ARCH_HAS_PREFETCHW
134#define prefetchw_prev_lru_page(_page, _base, _field) \
135 do { \
136 if ((_page)->lru.prev != _base) { \
137 struct page *prev; \
138 \
139 prev = lru_to_page(&(_page->lru)); \
140 prefetchw(&prev->_field); \
141 } \
142 } while (0)
143#else
144#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
145#endif
146
147/*
148 * From 0 .. 100. Higher means more swappy.
149 */
150int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700151long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
153static LIST_HEAD(shrinker_list);
154static DECLARE_RWSEM(shrinker_rwsem);
155
Balbir Singh00f0b822008-03-04 14:28:39 -0800156#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800157#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800158#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800159#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800160#endif
161
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800162static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
163 struct scan_control *sc)
164{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800165 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800166 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800168 return &zone->reclaim_stat;
169}
170
Vincent Li0b217672009-09-21 17:03:09 -0700171static unsigned long zone_nr_lru_pages(struct zone *zone,
172 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800173{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800174 if (!scanning_global_lru(sc))
KAMEZAWA Hiroyukibb2a0de2011-07-26 16:08:22 -0700175 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
176 zone_to_nid(zone), zone_idx(zone), BIT(lru));
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800177
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800178 return zone_page_state(zone, NR_LRU_BASE + lru);
179}
180
181
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182/*
183 * Add a shrinker callback to be called from the vm
184 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700185void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700187 shrinker->nr = 0;
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700192EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194/*
195 * Remove one
196 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700197void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700203EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
Ying Han1495f232011-05-24 17:12:27 -0700205static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
208{
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
211}
212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213#define SHRINK_BATCH 128
214/*
215 * Call the shrink functions to age shrinkable caches
216 *
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
221 *
Simon Arlott183ff222007-10-20 01:27:18 +0200222 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 * slab to avoid swapping.
224 *
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 *
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700230 *
231 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700233unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700234 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700235 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800238 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Ying Han1495f232011-05-24 17:12:27 -0700240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Minchan Kimf06590b2011-05-24 17:11:11 -0700243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
245 ret = 1;
246 goto out;
247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000252 unsigned long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000253 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000254 long nr;
255 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000256 long batch_size = shrinker->batch ? shrinker->batch
257 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258
Dave Chinneracf92b42011-07-08 14:14:35 +1000259 /*
260 * copy the current shrinker scan count into a local variable
261 * and zero it so that other concurrent shrinker invocations
262 * don't also do this scanning work.
263 */
264 do {
265 nr = shrinker->nr;
266 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
267
268 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700269 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800271 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000273 total_scan += delta;
274 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
276 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800279 }
280
281 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
288 * memory.
289 *
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
292 */
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
295
296 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
299 * freeable entries.
300 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Dave Chinneracf92b42011-07-08 14:14:35 +1000304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
307
Dave Chinnere9299f52011-07-08 14:14:37 +1000308 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700309 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Ying Han1495f232011-05-24 17:12:27 -0700311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000313 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 if (shrink_ret == -1)
315 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
321 cond_resched();
322 }
323
Dave Chinneracf92b42011-07-08 14:14:35 +1000324 /*
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
328 */
329 do {
330 nr = shrinker->nr;
331 new_nr = total_scan + nr;
332 if (total_scan <= 0)
333 break;
334 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
335
336 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 }
338 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700339out:
340 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700341 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342}
343
Mel Gormanf3a310b2011-01-13 15:46:00 -0800344static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700345 bool sync)
346{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800347 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700348
349 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800350 * Initially assume we are entering either lumpy reclaim or
351 * reclaim/compaction.Depending on the order, we will either set the
352 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700353 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800354 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800355 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800356 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800357 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700358
359 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800360 * Avoid using lumpy reclaim or reclaim/compaction if possible by
361 * restricting when its set to either costly allocations or when
362 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700363 */
364 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800365 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700366 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800367 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700368 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800369 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700370}
371
Mel Gormanf3a310b2011-01-13 15:46:00 -0800372static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700373{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800374 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700375}
376
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377static inline int is_page_cache_freeable(struct page *page)
378{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700379 /*
380 * A freeable page cache page is referenced only by the caller
381 * that isolated the page, the page cache radix tree and
382 * optional buffer heads at page->private.
383 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700384 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385}
386
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700387static int may_write_to_queue(struct backing_dev_info *bdi,
388 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389{
Christoph Lameter930d9152006-01-08 01:00:47 -0800390 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 return 1;
392 if (!bdi_write_congested(bdi))
393 return 1;
394 if (bdi == current->backing_dev_info)
395 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700396
397 /* lumpy reclaim for hugepage often need a lot of write */
398 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
399 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 return 0;
401}
402
403/*
404 * We detected a synchronous write error writing a page out. Probably
405 * -ENOSPC. We need to propagate that into the address_space for a subsequent
406 * fsync(), msync() or close().
407 *
408 * The tricky part is that after writepage we cannot touch the mapping: nothing
409 * prevents it from being freed up. But we have a ref on the page and once
410 * that page is locked, the mapping is pinned.
411 *
412 * We're allowed to run sleeping lock_page() here because we know the caller has
413 * __GFP_FS.
414 */
415static void handle_write_error(struct address_space *mapping,
416 struct page *page, int error)
417{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100418 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700419 if (page_mapping(page) == mapping)
420 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 unlock_page(page);
422}
423
Christoph Lameter04e62a22006-06-23 02:03:38 -0700424/* possible outcome of pageout() */
425typedef enum {
426 /* failed to write page out, page is locked */
427 PAGE_KEEP,
428 /* move page to the active list, page is locked */
429 PAGE_ACTIVATE,
430 /* page has been sent to the disk successfully, page is unlocked */
431 PAGE_SUCCESS,
432 /* page is clean and locked */
433 PAGE_CLEAN,
434} pageout_t;
435
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436/*
Andrew Morton1742f192006-03-22 00:08:21 -0800437 * pageout is called by shrink_page_list() for each dirty page.
438 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700440static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700441 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442{
443 /*
444 * If the page is dirty, only perform writeback if that write
445 * will be non-blocking. To prevent this allocation from being
446 * stalled by pagecache activity. But note that there may be
447 * stalls if we need to run get_block(). We could test
448 * PagePrivate for that.
449 *
Vincent Li6aceb532009-12-14 17:58:49 -0800450 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 * this page's queue, we can perform writeback even if that
452 * will block.
453 *
454 * If the page is swapcache, write it back even if that would
455 * block, for some throttling. This happens by accident, because
456 * swap_backing_dev_info is bust: it doesn't reflect the
457 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 */
459 if (!is_page_cache_freeable(page))
460 return PAGE_KEEP;
461 if (!mapping) {
462 /*
463 * Some data journaling orphaned pages can have
464 * page->mapping == NULL while being dirty with clean buffers.
465 */
David Howells266cf652009-04-03 16:42:36 +0100466 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 if (try_to_free_buffers(page)) {
468 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700469 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 return PAGE_CLEAN;
471 }
472 }
473 return PAGE_KEEP;
474 }
475 if (mapping->a_ops->writepage == NULL)
476 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700477 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 return PAGE_KEEP;
479
480 if (clear_page_dirty_for_io(page)) {
481 int res;
482 struct writeback_control wbc = {
483 .sync_mode = WB_SYNC_NONE,
484 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700485 .range_start = 0,
486 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 .for_reclaim = 1,
488 };
489
490 SetPageReclaim(page);
491 res = mapping->a_ops->writepage(page, &wbc);
492 if (res < 0)
493 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800494 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 ClearPageReclaim(page);
496 return PAGE_ACTIVATE;
497 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700498
499 /*
500 * Wait on writeback if requested to. This happens when
501 * direct reclaiming a large contiguous area and the
502 * first attempt to free a range of pages fails.
503 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700504 if (PageWriteback(page) &&
Mel Gormanf3a310b2011-01-13 15:46:00 -0800505 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
Andy Whitcroftc661b072007-08-22 14:01:26 -0700506 wait_on_page_writeback(page);
507
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 if (!PageWriteback(page)) {
509 /* synchronous write or broken a_ops? */
510 ClearPageReclaim(page);
511 }
Mel Gorman755f0222010-08-09 17:19:18 -0700512 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800513 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700514 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 return PAGE_SUCCESS;
516 }
517
518 return PAGE_CLEAN;
519}
520
Andrew Mortona649fd92006-10-17 00:09:36 -0700521/*
Nick Piggine2867812008-07-25 19:45:30 -0700522 * Same as remove_mapping, but if the page is removed from the mapping, it
523 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700524 */
Nick Piggine2867812008-07-25 19:45:30 -0700525static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800526{
Nick Piggin28e4d962006-09-25 23:31:23 -0700527 BUG_ON(!PageLocked(page));
528 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800529
Nick Piggin19fd6232008-07-25 19:45:32 -0700530 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800531 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700532 * The non racy check for a busy page.
533 *
534 * Must be careful with the order of the tests. When someone has
535 * a ref to the page, it may be possible that they dirty it then
536 * drop the reference. So if PageDirty is tested before page_count
537 * here, then the following race may occur:
538 *
539 * get_user_pages(&page);
540 * [user mapping goes away]
541 * write_to(page);
542 * !PageDirty(page) [good]
543 * SetPageDirty(page);
544 * put_page(page);
545 * !page_count(page) [good, discard it]
546 *
547 * [oops, our write_to data is lost]
548 *
549 * Reversing the order of the tests ensures such a situation cannot
550 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
551 * load is not satisfied before that of page->_count.
552 *
553 * Note that if SetPageDirty is always performed via set_page_dirty,
554 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800555 */
Nick Piggine2867812008-07-25 19:45:30 -0700556 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800557 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700558 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
559 if (unlikely(PageDirty(page))) {
560 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800561 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700562 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800563
564 if (PageSwapCache(page)) {
565 swp_entry_t swap = { .val = page_private(page) };
566 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700567 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700568 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700569 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500570 void (*freepage)(struct page *);
571
572 freepage = mapping->a_ops->freepage;
573
Minchan Kime64a7822011-03-22 16:32:44 -0700574 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700575 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700576 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500577
578 if (freepage != NULL)
579 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800580 }
581
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800582 return 1;
583
584cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700585 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800586 return 0;
587}
588
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589/*
Nick Piggine2867812008-07-25 19:45:30 -0700590 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
591 * someone else has a ref on the page, abort and return 0. If it was
592 * successfully detached, return 1. Assumes the caller has a single ref on
593 * this page.
594 */
595int remove_mapping(struct address_space *mapping, struct page *page)
596{
597 if (__remove_mapping(mapping, page)) {
598 /*
599 * Unfreezing the refcount with 1 rather than 2 effectively
600 * drops the pagecache ref for us without requiring another
601 * atomic operation.
602 */
603 page_unfreeze_refs(page, 1);
604 return 1;
605 }
606 return 0;
607}
608
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700609/**
610 * putback_lru_page - put previously isolated page onto appropriate LRU list
611 * @page: page to be put back to appropriate lru list
612 *
613 * Add previously isolated @page to appropriate LRU list.
614 * Page may still be unevictable for other reasons.
615 *
616 * lru_lock must not be held, interrupts must be enabled.
617 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700618void putback_lru_page(struct page *page)
619{
620 int lru;
621 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700622 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700623
624 VM_BUG_ON(PageLRU(page));
625
626redo:
627 ClearPageUnevictable(page);
628
629 if (page_evictable(page, NULL)) {
630 /*
631 * For evictable pages, we can use the cache.
632 * In event of a race, worst case is we end up with an
633 * unevictable page on [in]active list.
634 * We know how to handle that.
635 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700636 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700637 lru_cache_add_lru(page, lru);
638 } else {
639 /*
640 * Put unevictable pages directly on zone's unevictable
641 * list.
642 */
643 lru = LRU_UNEVICTABLE;
644 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700645 /*
646 * When racing with an mlock clearing (page is
647 * unlocked), make sure that if the other thread does
648 * not observe our setting of PG_lru and fails
649 * isolation, we see PG_mlocked cleared below and move
650 * the page back to the evictable list.
651 *
652 * The other side is TestClearPageMlocked().
653 */
654 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700655 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700656
657 /*
658 * page's status can change while we move it among lru. If an evictable
659 * page is on unevictable list, it never be freed. To avoid that,
660 * check after we added it to the list, again.
661 */
662 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
663 if (!isolate_lru_page(page)) {
664 put_page(page);
665 goto redo;
666 }
667 /* This means someone else dropped this page from LRU
668 * So, it will be freed or putback to LRU again. There is
669 * nothing to do here.
670 */
671 }
672
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700673 if (was_unevictable && lru != LRU_UNEVICTABLE)
674 count_vm_event(UNEVICTABLE_PGRESCUED);
675 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
676 count_vm_event(UNEVICTABLE_PGCULLED);
677
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700678 put_page(page); /* drop ref from isolate */
679}
680
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800681enum page_references {
682 PAGEREF_RECLAIM,
683 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800684 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800685 PAGEREF_ACTIVATE,
686};
687
688static enum page_references page_check_references(struct page *page,
689 struct scan_control *sc)
690{
Johannes Weiner645747462010-03-05 13:42:22 -0800691 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800692 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800693
Johannes Weiner645747462010-03-05 13:42:22 -0800694 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
695 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800696
697 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800698 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800699 return PAGEREF_RECLAIM;
700
701 /*
702 * Mlock lost the isolation race with us. Let try_to_unmap()
703 * move the page to the unevictable list.
704 */
705 if (vm_flags & VM_LOCKED)
706 return PAGEREF_RECLAIM;
707
Johannes Weiner645747462010-03-05 13:42:22 -0800708 if (referenced_ptes) {
709 if (PageAnon(page))
710 return PAGEREF_ACTIVATE;
711 /*
712 * All mapped pages start out with page table
713 * references from the instantiating fault, so we need
714 * to look twice if a mapped file page is used more
715 * than once.
716 *
717 * Mark it and spare it for another trip around the
718 * inactive list. Another page table reference will
719 * lead to its activation.
720 *
721 * Note: the mark is set for activated pages as well
722 * so that recently deactivated but used pages are
723 * quickly recovered.
724 */
725 SetPageReferenced(page);
726
727 if (referenced_page)
728 return PAGEREF_ACTIVATE;
729
730 return PAGEREF_KEEP;
731 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800732
733 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700734 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800735 return PAGEREF_RECLAIM_CLEAN;
736
737 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800738}
739
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700740static noinline_for_stack void free_page_list(struct list_head *free_pages)
741{
742 struct pagevec freed_pvec;
743 struct page *page, *tmp;
744
745 pagevec_init(&freed_pvec, 1);
746
747 list_for_each_entry_safe(page, tmp, free_pages, lru) {
748 list_del(&page->lru);
749 if (!pagevec_add(&freed_pvec, page)) {
750 __pagevec_free(&freed_pvec);
751 pagevec_reinit(&freed_pvec);
752 }
753 }
754
755 pagevec_free(&freed_pvec);
756}
757
Nick Piggine2867812008-07-25 19:45:30 -0700758/*
Andrew Morton1742f192006-03-22 00:08:21 -0800759 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 */
Andrew Morton1742f192006-03-22 00:08:21 -0800761static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700762 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700763 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764{
765 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700766 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700768 unsigned long nr_dirty = 0;
769 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800770 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771
772 cond_resched();
773
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800775 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 struct address_space *mapping;
777 struct page *page;
778 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
780 cond_resched();
781
782 page = lru_to_page(page_list);
783 list_del(&page->lru);
784
Nick Piggin529ae9a2008-08-02 12:01:03 +0200785 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 goto keep;
787
Nick Piggin725d7042006-09-25 23:30:55 -0700788 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700789 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
791 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800792
Nick Pigginb291f002008-10-18 20:26:44 -0700793 if (unlikely(!page_evictable(page, NULL)))
794 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700795
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700796 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800797 goto keep_locked;
798
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 /* Double the slab pressure for mapped and swapcache pages */
800 if (page_mapped(page) || PageSwapCache(page))
801 sc->nr_scanned++;
802
Andy Whitcroftc661b072007-08-22 14:01:26 -0700803 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
804 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805
806 if (PageWriteback(page)) {
807 /*
808 * Synchronous reclaim is performed in two passes,
809 * first an asynchronous pass over the list to
810 * start parallel writeback, and a second synchronous
811 * pass to wait for the IO to complete. Wait here
812 * for any page for which writeback has already
813 * started.
814 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800815 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700816 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700817 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700818 else {
819 unlock_page(page);
820 goto keep_lumpy;
821 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700822 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800824 references = page_check_references(page, sc);
825 switch (references) {
826 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800828 case PAGEREF_KEEP:
829 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800830 case PAGEREF_RECLAIM:
831 case PAGEREF_RECLAIM_CLEAN:
832 ; /* try to reclaim the page below */
833 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 /*
836 * Anonymous process memory has backing store?
837 * Try to allocate it some swap space here.
838 */
Nick Pigginb291f002008-10-18 20:26:44 -0700839 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800840 if (!(sc->gfp_mask & __GFP_IO))
841 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800842 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800844 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700845 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846
847 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
849 /*
850 * The page is mapped into the page tables of one or more
851 * processes. Try to unmap it here.
852 */
853 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200854 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 case SWAP_FAIL:
856 goto activate_locked;
857 case SWAP_AGAIN:
858 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700859 case SWAP_MLOCK:
860 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 case SWAP_SUCCESS:
862 ; /* try to free the page below */
863 }
864 }
865
866 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700867 nr_dirty++;
868
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800869 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700871 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800873 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 goto keep_locked;
875
876 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700877 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700879 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 goto keep_locked;
881 case PAGE_ACTIVATE:
882 goto activate_locked;
883 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700884 if (PageWriteback(page))
885 goto keep_lumpy;
886 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 /*
890 * A synchronous write - probably a ramdisk. Go
891 * ahead and try to reclaim the page.
892 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200893 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 goto keep;
895 if (PageDirty(page) || PageWriteback(page))
896 goto keep_locked;
897 mapping = page_mapping(page);
898 case PAGE_CLEAN:
899 ; /* try to free the page below */
900 }
901 }
902
903 /*
904 * If the page has buffers, try to free the buffer mappings
905 * associated with this page. If we succeed we try to free
906 * the page as well.
907 *
908 * We do this even if the page is PageDirty().
909 * try_to_release_page() does not perform I/O, but it is
910 * possible for a page to have PageDirty set, but it is actually
911 * clean (all its buffers are clean). This happens if the
912 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700913 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 * try_to_release_page() will discover that cleanness and will
915 * drop the buffers and mark the page clean - it can be freed.
916 *
917 * Rarely, pages can have buffers and no ->mapping. These are
918 * the pages which were not successfully invalidated in
919 * truncate_complete_page(). We try to drop those buffers here
920 * and if that worked, and the page is no longer mapped into
921 * process address space (page_count == 1) it can be freed.
922 * Otherwise, leave the page on the LRU so it is swappable.
923 */
David Howells266cf652009-04-03 16:42:36 +0100924 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 if (!try_to_release_page(page, sc->gfp_mask))
926 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700927 if (!mapping && page_count(page) == 1) {
928 unlock_page(page);
929 if (put_page_testzero(page))
930 goto free_it;
931 else {
932 /*
933 * rare race with speculative reference.
934 * the speculative reference will free
935 * this page shortly, so we may
936 * increment nr_reclaimed here (and
937 * leave it off the LRU).
938 */
939 nr_reclaimed++;
940 continue;
941 }
942 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 }
944
Nick Piggine2867812008-07-25 19:45:30 -0700945 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800946 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Nick Piggina978d6f2008-10-18 20:26:58 -0700948 /*
949 * At this point, we have no other references and there is
950 * no way to pick any more up (removed from LRU, removed
951 * from pagecache). Can use non-atomic bitops now (and
952 * we obviously don't have to worry about waking up a process
953 * waiting on the page lock, because there are no references.
954 */
955 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700956free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800957 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700958
959 /*
960 * Is there need to periodically free_page_list? It would
961 * appear not as the counts should be low
962 */
963 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 continue;
965
Nick Pigginb291f002008-10-18 20:26:44 -0700966cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800967 if (PageSwapCache(page))
968 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700969 unlock_page(page);
970 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800971 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700972 continue;
973
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700975 /* Not a candidate for swapping, so reclaim swap space. */
976 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800977 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700978 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 SetPageActive(page);
980 pgactivate++;
981keep_locked:
982 unlock_page(page);
983keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -0800984 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700985keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700987 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700989
Mel Gorman0e093d992010-10-26 14:21:45 -0700990 /*
991 * Tag a zone as congested if all the dirty pages encountered were
992 * backed by a congested BDI. In this case, reclaimers should just
993 * back off and wait for congestion to clear because further reclaim
994 * will encounter the same problem
995 */
KAMEZAWA Hiroyukid6c438b2011-05-17 15:44:10 -0700996 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
Mel Gorman0e093d992010-10-26 14:21:45 -0700997 zone_set_flag(zone, ZONE_CONGESTED);
998
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700999 free_page_list(&free_pages);
1000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001002 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -08001003 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004}
1005
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001006/*
1007 * Attempt to remove the specified page from its LRU. Only take this page
1008 * if it is of the appropriate PageActive status. Pages which are being
1009 * freed elsewhere are also ignored.
1010 *
1011 * page: page to consider
1012 * mode: one of the LRU isolation modes defined above
1013 *
1014 * returns 0 on success, -ve errno on failure.
1015 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001016int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001017{
1018 int ret = -EINVAL;
1019
1020 /* Only take pages on the LRU. */
1021 if (!PageLRU(page))
1022 return ret;
1023
1024 /*
1025 * When checking the active state, we need to be sure we are
1026 * dealing with comparible boolean values. Take the logical not
1027 * of each.
1028 */
1029 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
1030 return ret;
1031
Johannes Weiner6c0b1352009-09-21 17:02:59 -07001032 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001033 return ret;
1034
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001035 /*
1036 * When this function is being called for lumpy reclaim, we
1037 * initially look into all LRU pages, active, inactive and
1038 * unevictable; only give shrink_page_list evictable pages.
1039 */
1040 if (PageUnevictable(page))
1041 return ret;
1042
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001043 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001044
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001045 if (likely(get_page_unless_zero(page))) {
1046 /*
1047 * Be careful not to clear PageLRU until after we're
1048 * sure the page is not being freed elsewhere -- the
1049 * page release code relies on it.
1050 */
1051 ClearPageLRU(page);
1052 ret = 0;
1053 }
1054
1055 return ret;
1056}
1057
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001058/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 * zone->lru_lock is heavily contended. Some of the functions that
1060 * shrink the lists perform better by taking out a batch of pages
1061 * and working on them outside the LRU lock.
1062 *
1063 * For pagecache intensive workloads, this function is the hottest
1064 * spot in the kernel (apart from copy_*_user functions).
1065 *
1066 * Appropriate locks must be held before calling this function.
1067 *
1068 * @nr_to_scan: The number of pages to look through on the list.
1069 * @src: The LRU list to pull pages off.
1070 * @dst: The temp list to put pages on to.
1071 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001072 * @order: The caller's attempted allocation order
1073 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001074 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 *
1076 * returns how many pages were moved onto *@dst.
1077 */
Andrew Morton69e05942006-03-22 00:08:19 -08001078static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1079 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001080 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081{
Andrew Morton69e05942006-03-22 00:08:19 -08001082 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001083 unsigned long nr_lumpy_taken = 0;
1084 unsigned long nr_lumpy_dirty = 0;
1085 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001086 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001088 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001089 struct page *page;
1090 unsigned long pfn;
1091 unsigned long end_pfn;
1092 unsigned long page_pfn;
1093 int zone_id;
1094
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 page = lru_to_page(src);
1096 prefetchw_prev_lru_page(page, src, flags);
1097
Nick Piggin725d7042006-09-25 23:30:55 -07001098 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001099
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001100 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001101 case 0:
1102 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001103 mem_cgroup_del_lru(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001104 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001105 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001106
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001107 case -EBUSY:
1108 /* else it is being freed elsewhere */
1109 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001110 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001111 continue;
1112
1113 default:
1114 BUG();
1115 }
1116
1117 if (!order)
1118 continue;
1119
1120 /*
1121 * Attempt to take all pages in the order aligned region
1122 * surrounding the tag page. Only take those pages of
1123 * the same active state as that tag page. We may safely
1124 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001125 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001126 * where that page is in a different zone we will detect
1127 * it from its zone id and abort this block scan.
1128 */
1129 zone_id = page_zone_id(page);
1130 page_pfn = page_to_pfn(page);
1131 pfn = page_pfn & ~((1 << order) - 1);
1132 end_pfn = pfn + (1 << order);
1133 for (; pfn < end_pfn; pfn++) {
1134 struct page *cursor_page;
1135
1136 /* The target page is in the block, ignore it. */
1137 if (unlikely(pfn == page_pfn))
1138 continue;
1139
1140 /* Avoid holes within the zone. */
1141 if (unlikely(!pfn_valid_within(pfn)))
1142 break;
1143
1144 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001145
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001146 /* Check that we have not crossed a zone boundary. */
1147 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001148 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001149
1150 /*
1151 * If we don't have enough swap space, reclaiming of
1152 * anon page which don't already have a swap slot is
1153 * pointless.
1154 */
1155 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001156 !PageSwapCache(cursor_page))
1157 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001158
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001159 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001160 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001161 mem_cgroup_del_lru(cursor_page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001162 nr_taken += hpage_nr_pages(page);
Mel Gormana8a94d12010-08-09 17:19:17 -07001163 nr_lumpy_taken++;
1164 if (PageDirty(cursor_page))
1165 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001166 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001167 } else {
Andrea Arcangelid179e842011-06-15 15:08:51 -07001168 /*
1169 * Check if the page is freed already.
1170 *
1171 * We can't use page_count() as that
1172 * requires compound_head and we don't
1173 * have a pin on the page here. If a
1174 * page is tail, we may or may not
1175 * have isolated the head, so assume
1176 * it's not free, it'd be tricky to
1177 * track the head status without a
1178 * page pin.
1179 */
1180 if (!PageTail(cursor_page) &&
1181 !atomic_read(&cursor_page->_count))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001182 continue;
1183 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001184 }
1185 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001186
1187 /* If we break out of the loop above, lumpy reclaim failed */
1188 if (pfn < end_pfn)
1189 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 }
1191
1192 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001193
1194 trace_mm_vmscan_lru_isolate(order,
1195 nr_to_scan, scan,
1196 nr_taken,
1197 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1198 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 return nr_taken;
1200}
1201
Balbir Singh66e17072008-02-07 00:13:56 -08001202static unsigned long isolate_pages_global(unsigned long nr,
1203 struct list_head *dst,
1204 unsigned long *scanned, int order,
1205 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001206 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001207{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001208 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001209 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001210 lru += LRU_ACTIVE;
1211 if (file)
1212 lru += LRU_FILE;
1213 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001214 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001215}
1216
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001218 * clear_active_flags() is a helper for shrink_active_list(), clearing
1219 * any active bits from the pages in the list.
1220 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001221static unsigned long clear_active_flags(struct list_head *page_list,
1222 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001223{
1224 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001225 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001226 struct page *page;
1227
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001228 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001229 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001230 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001231 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001232 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001233 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001234 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001235 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001236 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001237 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001238 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001239
1240 return nr_active;
1241}
1242
Nick Piggin62695a82008-10-18 20:26:09 -07001243/**
1244 * isolate_lru_page - tries to isolate a page from its LRU list
1245 * @page: page to isolate from its LRU list
1246 *
1247 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1248 * vmstat statistic corresponding to whatever LRU list the page was on.
1249 *
1250 * Returns 0 if the page was removed from an LRU list.
1251 * Returns -EBUSY if the page was not on an LRU list.
1252 *
1253 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001254 * the active list, it will have PageActive set. If it was found on
1255 * the unevictable list, it will have the PageUnevictable bit set. That flag
1256 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001257 *
1258 * The vmstat statistic corresponding to the list on which the page was
1259 * found will be decremented.
1260 *
1261 * Restrictions:
1262 * (1) Must be called with an elevated refcount on the page. This is a
1263 * fundamentnal difference from isolate_lru_pages (which is called
1264 * without a stable reference).
1265 * (2) the lru_lock must not be held.
1266 * (3) interrupts must be enabled.
1267 */
1268int isolate_lru_page(struct page *page)
1269{
1270 int ret = -EBUSY;
1271
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001272 VM_BUG_ON(!page_count(page));
1273
Nick Piggin62695a82008-10-18 20:26:09 -07001274 if (PageLRU(page)) {
1275 struct zone *zone = page_zone(page);
1276
1277 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001278 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001279 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001280 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001281 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001282 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001283
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001284 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001285 }
1286 spin_unlock_irq(&zone->lru_lock);
1287 }
1288 return ret;
1289}
1290
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001291/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001292 * Are there way too many processes in the direct reclaim path already?
1293 */
1294static int too_many_isolated(struct zone *zone, int file,
1295 struct scan_control *sc)
1296{
1297 unsigned long inactive, isolated;
1298
1299 if (current_is_kswapd())
1300 return 0;
1301
1302 if (!scanning_global_lru(sc))
1303 return 0;
1304
1305 if (file) {
1306 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1307 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1308 } else {
1309 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1310 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1311 }
1312
1313 return isolated > inactive;
1314}
1315
1316/*
Mel Gorman66635622010-08-09 17:19:30 -07001317 * TODO: Try merging with migrations version of putback_lru_pages
1318 */
1319static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001320putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001321 unsigned long nr_anon, unsigned long nr_file,
1322 struct list_head *page_list)
1323{
1324 struct page *page;
1325 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001326 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001327
1328 pagevec_init(&pvec, 1);
1329
1330 /*
1331 * Put back any unfreeable pages.
1332 */
1333 spin_lock(&zone->lru_lock);
1334 while (!list_empty(page_list)) {
1335 int lru;
1336 page = lru_to_page(page_list);
1337 VM_BUG_ON(PageLRU(page));
1338 list_del(&page->lru);
1339 if (unlikely(!page_evictable(page, NULL))) {
1340 spin_unlock_irq(&zone->lru_lock);
1341 putback_lru_page(page);
1342 spin_lock_irq(&zone->lru_lock);
1343 continue;
1344 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001345 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001346 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001347 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001348 if (is_active_lru(lru)) {
1349 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001350 int numpages = hpage_nr_pages(page);
1351 reclaim_stat->recent_rotated[file] += numpages;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07001352 if (!scanning_global_lru(sc))
1353 sc->memcg_record->nr_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001354 }
1355 if (!pagevec_add(&pvec, page)) {
1356 spin_unlock_irq(&zone->lru_lock);
1357 __pagevec_release(&pvec);
1358 spin_lock_irq(&zone->lru_lock);
1359 }
1360 }
1361 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1362 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1363
1364 spin_unlock_irq(&zone->lru_lock);
1365 pagevec_release(&pvec);
1366}
1367
Mel Gorman1489fa12010-08-09 17:19:33 -07001368static noinline_for_stack void update_isolated_counts(struct zone *zone,
1369 struct scan_control *sc,
1370 unsigned long *nr_anon,
1371 unsigned long *nr_file,
1372 struct list_head *isolated_list)
1373{
1374 unsigned long nr_active;
1375 unsigned int count[NR_LRU_LISTS] = { 0, };
1376 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1377
1378 nr_active = clear_active_flags(isolated_list, count);
1379 __count_vm_events(PGDEACTIVATE, nr_active);
1380
1381 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1382 -count[LRU_ACTIVE_FILE]);
1383 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1384 -count[LRU_INACTIVE_FILE]);
1385 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1386 -count[LRU_ACTIVE_ANON]);
1387 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1388 -count[LRU_INACTIVE_ANON]);
1389
1390 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1391 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1392 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1393 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1394
1395 reclaim_stat->recent_scanned[0] += *nr_anon;
1396 reclaim_stat->recent_scanned[1] += *nr_file;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07001397 if (!scanning_global_lru(sc)) {
1398 sc->memcg_record->nr_scanned[0] += *nr_anon;
1399 sc->memcg_record->nr_scanned[1] += *nr_file;
1400 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001401}
1402
Mel Gorman66635622010-08-09 17:19:30 -07001403/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001404 * Returns true if the caller should wait to clean dirty/writeback pages.
1405 *
1406 * If we are direct reclaiming for contiguous pages and we do not reclaim
1407 * everything in the list, try again and wait for writeback IO to complete.
1408 * This will stall high-order allocations noticeably. Only do that when really
1409 * need to free the pages under high memory pressure.
1410 */
1411static inline bool should_reclaim_stall(unsigned long nr_taken,
1412 unsigned long nr_freed,
1413 int priority,
1414 struct scan_control *sc)
1415{
1416 int lumpy_stall_priority;
1417
1418 /* kswapd should not stall on sync IO */
1419 if (current_is_kswapd())
1420 return false;
1421
1422 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001423 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001424 return false;
1425
1426 /* If we have relaimed everything on the isolated list, no stall */
1427 if (nr_freed == nr_taken)
1428 return false;
1429
1430 /*
1431 * For high-order allocations, there are two stall thresholds.
1432 * High-cost allocations stall immediately where as lower
1433 * order allocations such as stacks require the scanning
1434 * priority to be much higher before stalling.
1435 */
1436 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1437 lumpy_stall_priority = DEF_PRIORITY;
1438 else
1439 lumpy_stall_priority = DEF_PRIORITY / 3;
1440
1441 return priority <= lumpy_stall_priority;
1442}
1443
1444/*
Andrew Morton1742f192006-03-22 00:08:21 -08001445 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1446 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 */
Mel Gorman66635622010-08-09 17:19:30 -07001448static noinline_for_stack unsigned long
1449shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1450 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451{
1452 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001453 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001454 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001455 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001456 unsigned long nr_anon;
1457 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001458
Rik van Riel35cd7812009-09-21 17:01:38 -07001459 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001460 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001461
1462 /* We are about to die and free our memory. Return now. */
1463 if (fatal_signal_pending(current))
1464 return SWAP_CLUSTER_MAX;
1465 }
1466
Mel Gormanf3a310b2011-01-13 15:46:00 -08001467 set_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 lru_add_drain();
1469 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001471 if (scanning_global_lru(sc)) {
1472 nr_taken = isolate_pages_global(nr_to_scan,
1473 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001474 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001475 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001476 zone, 0, file);
1477 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001478 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001479 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1480 nr_scanned);
1481 else
1482 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1483 nr_scanned);
1484 } else {
1485 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1486 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001487 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001488 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001489 zone, sc->mem_cgroup,
1490 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001492 * mem_cgroup_isolate_pages() keeps track of
1493 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001495 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001496
Mel Gorman66635622010-08-09 17:19:30 -07001497 if (nr_taken == 0) {
1498 spin_unlock_irq(&zone->lru_lock);
1499 return 0;
1500 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001501
Mel Gorman1489fa12010-08-09 17:19:33 -07001502 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001503
1504 spin_unlock_irq(&zone->lru_lock);
1505
Mel Gorman0e093d992010-10-26 14:21:45 -07001506 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001507
Wu Fengguange31f36982010-08-09 17:20:01 -07001508 /* Check if we should syncronously wait for writeback */
1509 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001510 set_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001511 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001512 }
1513
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07001514 if (!scanning_global_lru(sc))
1515 sc->memcg_record->nr_freed[file] += nr_reclaimed;
1516
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001517 local_irq_disable();
1518 if (current_is_kswapd())
1519 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1520 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1521
Mel Gorman1489fa12010-08-09 17:19:33 -07001522 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001523
1524 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1525 zone_idx(zone),
1526 nr_scanned, nr_reclaimed,
1527 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001528 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001529 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530}
1531
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001532/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 * This moves pages from the active list to the inactive list.
1534 *
1535 * We move them the other way if the page is referenced by one or more
1536 * processes, from rmap.
1537 *
1538 * If the pages are mostly unmapped, the processing is fast and it is
1539 * appropriate to hold zone->lru_lock across the whole operation. But if
1540 * the pages are mapped, the processing is slow (page_referenced()) so we
1541 * should drop zone->lru_lock around each page. It's impossible to balance
1542 * this, so instead we remove the pages from the LRU while processing them.
1543 * It is safe to rely on PG_active against the non-LRU pages in here because
1544 * nobody will play with that bit on a non-LRU page.
1545 *
1546 * The downside is that we have to touch page->_count against each page.
1547 * But we had to alter page->flags anyway.
1548 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001549
Wu Fengguang3eb41402009-06-16 15:33:13 -07001550static void move_active_pages_to_lru(struct zone *zone,
1551 struct list_head *list,
1552 enum lru_list lru)
1553{
1554 unsigned long pgmoved = 0;
1555 struct pagevec pvec;
1556 struct page *page;
1557
1558 pagevec_init(&pvec, 1);
1559
1560 while (!list_empty(list)) {
1561 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001562
1563 VM_BUG_ON(PageLRU(page));
1564 SetPageLRU(page);
1565
Wu Fengguang3eb41402009-06-16 15:33:13 -07001566 list_move(&page->lru, &zone->lru[lru].list);
1567 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001568 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001569
1570 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1571 spin_unlock_irq(&zone->lru_lock);
1572 if (buffer_heads_over_limit)
1573 pagevec_strip(&pvec);
1574 __pagevec_release(&pvec);
1575 spin_lock_irq(&zone->lru_lock);
1576 }
1577 }
1578 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1579 if (!is_active_lru(lru))
1580 __count_vm_events(PGDEACTIVATE, pgmoved);
1581}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001582
Andrew Morton1742f192006-03-22 00:08:21 -08001583static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001584 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001586 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001587 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001588 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001590 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001591 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001593 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001594 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595
1596 lru_add_drain();
1597 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001598 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001599 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1600 &pgscanned, sc->order,
1601 ISOLATE_ACTIVE, zone,
1602 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001603 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001604 } else {
1605 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1606 &pgscanned, sc->order,
1607 ISOLATE_ACTIVE, zone,
1608 sc->mem_cgroup, 1, file);
1609 /*
1610 * mem_cgroup_isolate_pages() keeps track of
1611 * scanned pages on its own.
1612 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001613 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001614
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001615 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07001616 if (!scanning_global_lru(sc))
1617 sc->memcg_record->nr_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001618
Wu Fengguang3eb41402009-06-16 15:33:13 -07001619 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001620 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001621 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001622 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001623 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001624 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 spin_unlock_irq(&zone->lru_lock);
1626
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 while (!list_empty(&l_hold)) {
1628 cond_resched();
1629 page = lru_to_page(&l_hold);
1630 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001631
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001632 if (unlikely(!page_evictable(page, NULL))) {
1633 putback_lru_page(page);
1634 continue;
1635 }
1636
Johannes Weiner645747462010-03-05 13:42:22 -08001637 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001638 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001639 /*
1640 * Identify referenced, file-backed active pages and
1641 * give them one more trip around the active list. So
1642 * that executable code get better chances to stay in
1643 * memory under moderate memory pressure. Anon pages
1644 * are not likely to be evicted by use-once streaming
1645 * IO, plus JVM can create lots of anon VM_EXEC pages,
1646 * so we ignore them here.
1647 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001648 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001649 list_add(&page->lru, &l_active);
1650 continue;
1651 }
1652 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001653
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001654 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 list_add(&page->lru, &l_inactive);
1656 }
1657
Andrew Mortonb5557492009-01-06 14:40:13 -08001658 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001659 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001660 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001661 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001662 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001663 * Count referenced pages from currently used mappings as rotated,
1664 * even though only some of them are actually re-activated. This
1665 * helps balance scan pressure between file and anonymous pages in
1666 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001667 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001668 reclaim_stat->recent_rotated[file] += nr_rotated;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07001669 if (!scanning_global_lru(sc))
1670 sc->memcg_record->nr_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001671
Wu Fengguang3eb41402009-06-16 15:33:13 -07001672 move_active_pages_to_lru(zone, &l_active,
1673 LRU_ACTIVE + file * LRU_FILE);
1674 move_active_pages_to_lru(zone, &l_inactive,
1675 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001676 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001677 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678}
1679
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001680#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001681static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001682{
1683 unsigned long active, inactive;
1684
1685 active = zone_page_state(zone, NR_ACTIVE_ANON);
1686 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1687
1688 if (inactive * zone->inactive_ratio < active)
1689 return 1;
1690
1691 return 0;
1692}
1693
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001694/**
1695 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1696 * @zone: zone to check
1697 * @sc: scan control of this context
1698 *
1699 * Returns true if the zone does not have enough inactive anon pages,
1700 * meaning some active anon pages need to be deactivated.
1701 */
1702static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1703{
1704 int low;
1705
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001706 /*
1707 * If we don't have swap space, anonymous page deactivation
1708 * is pointless.
1709 */
1710 if (!total_swap_pages)
1711 return 0;
1712
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001713 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001714 low = inactive_anon_is_low_global(zone);
1715 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001716 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001717 return low;
1718}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001719#else
1720static inline int inactive_anon_is_low(struct zone *zone,
1721 struct scan_control *sc)
1722{
1723 return 0;
1724}
1725#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001726
Rik van Riel56e49d22009-06-16 15:32:28 -07001727static int inactive_file_is_low_global(struct zone *zone)
1728{
1729 unsigned long active, inactive;
1730
1731 active = zone_page_state(zone, NR_ACTIVE_FILE);
1732 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1733
1734 return (active > inactive);
1735}
1736
1737/**
1738 * inactive_file_is_low - check if file pages need to be deactivated
1739 * @zone: zone to check
1740 * @sc: scan control of this context
1741 *
1742 * When the system is doing streaming IO, memory pressure here
1743 * ensures that active file pages get deactivated, until more
1744 * than half of the file pages are on the inactive list.
1745 *
1746 * Once we get to that situation, protect the system's working
1747 * set from being evicted by disabling active file page aging.
1748 *
1749 * This uses a different ratio than the anonymous pages, because
1750 * the page cache uses a use-once replacement algorithm.
1751 */
1752static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1753{
1754 int low;
1755
1756 if (scanning_global_lru(sc))
1757 low = inactive_file_is_low_global(zone);
1758 else
1759 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1760 return low;
1761}
1762
Rik van Rielb39415b2009-12-14 17:59:48 -08001763static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1764 int file)
1765{
1766 if (file)
1767 return inactive_file_is_low(zone, sc);
1768 else
1769 return inactive_anon_is_low(zone, sc);
1770}
1771
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001772static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001773 struct zone *zone, struct scan_control *sc, int priority)
1774{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001775 int file = is_file_lru(lru);
1776
Rik van Rielb39415b2009-12-14 17:59:48 -08001777 if (is_active_lru(lru)) {
1778 if (inactive_list_is_low(zone, sc, file))
1779 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001780 return 0;
1781 }
1782
Rik van Riel33c120e2008-10-18 20:26:36 -07001783 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001784}
1785
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001786static int vmscan_swappiness(struct scan_control *sc)
1787{
1788 if (scanning_global_lru(sc))
1789 return vm_swappiness;
1790 return mem_cgroup_swappiness(sc->mem_cgroup);
1791}
1792
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001794 * Determine how aggressively the anon and file LRU lists should be
1795 * scanned. The relative value of each set of LRU lists is determined
1796 * by looking at the fraction of the pages scanned we did rotate back
1797 * onto the active list instead of evict.
1798 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001799 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001800 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001801static void get_scan_count(struct zone *zone, struct scan_control *sc,
1802 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001803{
1804 unsigned long anon, file, free;
1805 unsigned long anon_prio, file_prio;
1806 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001807 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001808 u64 fraction[2], denominator;
1809 enum lru_list l;
1810 int noswap = 0;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001811 int force_scan = 0;
KAMEZAWA Hiroyuki4508378b952011-07-26 16:08:24 -07001812 unsigned long nr_force_scan[2];
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001813
1814
1815 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1816 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1817 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1818 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1819
1820 if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1821 /* kswapd does zone balancing and need to scan this zone */
1822 if (scanning_global_lru(sc) && current_is_kswapd())
1823 force_scan = 1;
1824 /* memcg may have small limit and need to avoid priority drop */
1825 if (!scanning_global_lru(sc))
1826 force_scan = 1;
1827 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07001828
1829 /* If we have no swap space, do not bother scanning anon pages. */
1830 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1831 noswap = 1;
1832 fraction[0] = 0;
1833 fraction[1] = 1;
1834 denominator = 1;
KAMEZAWA Hiroyuki4508378b952011-07-26 16:08:24 -07001835 nr_force_scan[0] = 0;
1836 nr_force_scan[1] = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001837 goto out;
1838 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001839
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001840 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001841 free = zone_page_state(zone, NR_FREE_PAGES);
1842 /* If we have very few page cache pages,
1843 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001844 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001845 fraction[0] = 1;
1846 fraction[1] = 0;
1847 denominator = 1;
KAMEZAWA Hiroyuki4508378b952011-07-26 16:08:24 -07001848 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1849 nr_force_scan[1] = 0;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001850 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001851 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001852 }
1853
1854 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001855 * With swappiness at 100, anonymous and file have the same priority.
1856 * This scanning priority is essentially the inverse of IO cost.
1857 */
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001858 anon_prio = vmscan_swappiness(sc);
1859 file_prio = 200 - vmscan_swappiness(sc);
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001860
1861 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001862 * OK, so we have swap space and a fair amount of page cache
1863 * pages. We use the recently rotated / recently scanned
1864 * ratios to determine how valuable each cache is.
1865 *
1866 * Because workloads change over time (and to avoid overflow)
1867 * we keep these statistics as a floating average, which ends
1868 * up weighing recent references more than old ones.
1869 *
1870 * anon in [0], file in [1]
1871 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001872 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001873 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001874 reclaim_stat->recent_scanned[0] /= 2;
1875 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001876 }
1877
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001878 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001879 reclaim_stat->recent_scanned[1] /= 2;
1880 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001881 }
1882
1883 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001884 * The amount of pressure on anon vs file pages is inversely
1885 * proportional to the fraction of recently scanned pages on
1886 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001887 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001888 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1889 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001890
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001891 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1892 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001893 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001894
Shaohua Li76a33fc2010-05-24 14:32:36 -07001895 fraction[0] = ap;
1896 fraction[1] = fp;
1897 denominator = ap + fp + 1;
KAMEZAWA Hiroyuki4508378b952011-07-26 16:08:24 -07001898 if (force_scan) {
1899 unsigned long scan = SWAP_CLUSTER_MAX;
1900 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1901 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1902 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07001903out:
1904 for_each_evictable_lru(l) {
1905 int file = is_file_lru(l);
1906 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001907
Shaohua Li76a33fc2010-05-24 14:32:36 -07001908 scan = zone_nr_lru_pages(zone, sc, l);
1909 if (priority || noswap) {
1910 scan >>= priority;
1911 scan = div64_u64(scan * fraction[file], denominator);
1912 }
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001913
1914 /*
1915 * If zone is small or memcg is small, nr[l] can be 0.
1916 * This results no-scan on this priority and priority drop down.
1917 * For global direct reclaim, it can visit next zone and tend
1918 * not to have problems. For global kswapd, it's for zone
1919 * balancing and it need to scan a small amounts. When using
1920 * memcg, priority drop can cause big latency. So, it's better
1921 * to scan small amount. See may_noscan above.
1922 */
KAMEZAWA Hiroyuki4508378b952011-07-26 16:08:24 -07001923 if (!scan && force_scan)
1924 scan = nr_force_scan[file];
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001925 nr[l] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001926 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001927}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001928
1929/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001930 * Reclaim/compaction depends on a number of pages being freed. To avoid
1931 * disruption to the system, a small number of order-0 pages continue to be
1932 * rotated and reclaimed in the normal fashion. However, by the time we get
1933 * back to the allocator and call try_to_compact_zone(), we ensure that
1934 * there are enough free pages for it to be likely successful
1935 */
1936static inline bool should_continue_reclaim(struct zone *zone,
1937 unsigned long nr_reclaimed,
1938 unsigned long nr_scanned,
1939 struct scan_control *sc)
1940{
1941 unsigned long pages_for_compaction;
1942 unsigned long inactive_lru_pages;
1943
1944 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001945 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001946 return false;
1947
Mel Gorman28765922011-02-25 14:44:20 -08001948 /* Consider stopping depending on scan and reclaim activity */
1949 if (sc->gfp_mask & __GFP_REPEAT) {
1950 /*
1951 * For __GFP_REPEAT allocations, stop reclaiming if the
1952 * full LRU list has been scanned and we are still failing
1953 * to reclaim pages. This full LRU scan is potentially
1954 * expensive but a __GFP_REPEAT caller really wants to succeed
1955 */
1956 if (!nr_reclaimed && !nr_scanned)
1957 return false;
1958 } else {
1959 /*
1960 * For non-__GFP_REPEAT allocations which can presumably
1961 * fail without consequence, stop if we failed to reclaim
1962 * any pages from the last SWAP_CLUSTER_MAX number of
1963 * pages that were scanned. This will return to the
1964 * caller faster at the risk reclaim/compaction and
1965 * the resulting allocation attempt fails
1966 */
1967 if (!nr_reclaimed)
1968 return false;
1969 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001970
1971 /*
1972 * If we have not reclaimed enough pages for compaction and the
1973 * inactive lists are large enough, continue reclaiming
1974 */
1975 pages_for_compaction = (2UL << sc->order);
1976 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1977 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1978 if (sc->nr_reclaimed < pages_for_compaction &&
1979 inactive_lru_pages > pages_for_compaction)
1980 return true;
1981
1982 /* If compaction would go ahead or the allocation would succeed, stop */
1983 switch (compaction_suitable(zone, sc->order)) {
1984 case COMPACT_PARTIAL:
1985 case COMPACT_CONTINUE:
1986 return false;
1987 default:
1988 return true;
1989 }
1990}
1991
1992/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1994 */
Rik van Riela79311c2009-01-06 14:40:01 -08001995static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001996 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001998 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001999 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07002000 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002001 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002002 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003
Mel Gorman3e7d3442011-01-13 15:45:56 -08002004restart:
2005 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002006 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002007 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002008
Rik van Riel556adec2008-10-18 20:26:34 -07002009 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2010 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002011 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07002012 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08002013 nr_to_scan = min_t(unsigned long,
2014 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002015 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002017 nr_reclaimed += shrink_list(l, nr_to_scan,
2018 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002019 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 }
Rik van Riela79311c2009-01-06 14:40:01 -08002021 /*
2022 * On large memory systems, scan >> priority can become
2023 * really large. This is fine for the starting priority;
2024 * we want to put equal scanning pressure on each zone.
2025 * However, if the VM has a harder time of freeing pages,
2026 * with multiple processes reclaiming pages, the total
2027 * freeing target can get unreasonably large.
2028 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08002029 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08002030 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002032 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002033
Rik van Riel556adec2008-10-18 20:26:34 -07002034 /*
2035 * Even if we did not try to evict anon pages at all, we want to
2036 * rebalance the anon lru active/inactive ratio.
2037 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002038 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002039 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2040
Mel Gorman3e7d3442011-01-13 15:45:56 -08002041 /* reclaim/compaction might need reclaim to continue */
2042 if (should_continue_reclaim(zone, nr_reclaimed,
2043 sc->nr_scanned - nr_scanned, sc))
2044 goto restart;
2045
Andrew Morton232ea4d2007-02-28 20:13:21 -08002046 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047}
2048
2049/*
2050 * This is the direct reclaim path, for page-allocating processes. We only
2051 * try to reclaim pages from zones which will satisfy the caller's allocation
2052 * request.
2053 *
Mel Gorman41858962009-06-16 15:32:12 -07002054 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2055 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2057 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002058 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2059 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2060 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 * If a zone is deemed to be full of pinned pages then just give it a light
2063 * scan then give up on it.
2064 */
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002065static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08002066 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067{
Mel Gormandd1a2392008-04-28 02:12:17 -07002068 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002069 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002070 unsigned long nr_soft_reclaimed;
2071 unsigned long nr_soft_scanned;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002072
Mel Gormand4debc62010-08-09 17:19:29 -07002073 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2074 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002075 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002077 /*
2078 * Take care memory controller reclaiming has small influence
2079 * to global LRU.
2080 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002081 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002082 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2083 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002084 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002085 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002086 /*
2087 * This steals pages from memory cgroups over softlimit
2088 * and returns the number of reclaimed pages and
2089 * scanned pages. This works for global memory pressure
2090 * and balancing, not for a memcg's limit.
2091 */
2092 nr_soft_scanned = 0;
2093 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2094 sc->order, sc->gfp_mask,
2095 &nr_soft_scanned);
2096 sc->nr_reclaimed += nr_soft_reclaimed;
2097 sc->nr_scanned += nr_soft_scanned;
2098 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002099 }
Nick Piggin408d8542006-09-25 23:31:27 -07002100
Rik van Riela79311c2009-01-06 14:40:01 -08002101 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 }
Minchan Kimd1908362010-09-22 13:05:01 -07002103}
2104
2105static bool zone_reclaimable(struct zone *zone)
2106{
2107 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2108}
2109
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002110/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002111static bool all_unreclaimable(struct zonelist *zonelist,
2112 struct scan_control *sc)
2113{
2114 struct zoneref *z;
2115 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002116
2117 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2118 gfp_zone(sc->gfp_mask), sc->nodemask) {
2119 if (!populated_zone(zone))
2120 continue;
2121 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2122 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002123 if (!zone->all_unreclaimable)
2124 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002125 }
2126
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002127 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002129
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130/*
2131 * This is the main entry point to direct page reclaim.
2132 *
2133 * If a full scan of the inactive list fails to free enough memory then we
2134 * are "out of memory" and something needs to be killed.
2135 *
2136 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2137 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002138 * caller can't do much about. We kick the writeback threads and take explicit
2139 * naps in the hope that some of these pages can be written. But if the
2140 * allocating task holds filesystem locks which prevent writeout this might not
2141 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002142 *
2143 * returns: 0, if no pages reclaimed
2144 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 */
Mel Gormandac1d272008-04-28 02:12:12 -07002146static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002147 struct scan_control *sc,
2148 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149{
2150 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002151 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002153 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002154 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002155 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
Miao Xiec0ff7452010-05-24 14:32:08 -07002157 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002158 delayacct_freepages_start();
2159
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002160 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002161 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
2163 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002164 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002165 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002166 disable_swap_token(sc->mem_cgroup);
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002167 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08002168 /*
2169 * Don't shrink slabs when reclaiming memory from
2170 * over limit cgroups
2171 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002172 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002173 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002174 for_each_zone_zonelist(zone, z, zonelist,
2175 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002176 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2177 continue;
2178
2179 lru_pages += zone_reclaimable_pages(zone);
2180 }
2181
Ying Han1495f232011-05-24 17:12:27 -07002182 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002183 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002184 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002185 reclaim_state->reclaimed_slab = 0;
2186 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 }
Balbir Singh66e17072008-02-07 00:13:56 -08002188 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002189 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191
2192 /*
2193 * Try to write back as many pages as we just scanned. This
2194 * tends to cause slow streaming writers to write data to the
2195 * disk smoothly, at the dirtying rate, which is nice. But
2196 * that's undesirable in laptop mode, where we *want* lumpy
2197 * writeout. So in laptop mode, write out the whole world.
2198 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002199 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2200 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002201 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002202 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 }
2204
2205 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002206 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002207 priority < DEF_PRIORITY - 2) {
2208 struct zone *preferred_zone;
2209
2210 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002211 &cpuset_current_mems_allowed,
2212 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002213 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002216
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002218 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002219 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002220
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002221 if (sc->nr_reclaimed)
2222 return sc->nr_reclaimed;
2223
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002224 /*
2225 * As hibernation is going on, kswapd is freezed so that it can't mark
2226 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2227 * check.
2228 */
2229 if (oom_killer_disabled)
2230 return 0;
2231
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002232 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002233 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002234 return 1;
2235
2236 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237}
2238
Mel Gormandac1d272008-04-28 02:12:12 -07002239unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002240 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002241{
Mel Gorman33906bc2010-08-09 17:19:16 -07002242 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002243 struct scan_control sc = {
2244 .gfp_mask = gfp_mask,
2245 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002246 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002247 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002248 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002249 .order = order,
2250 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002251 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002252 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002253 struct shrink_control shrink = {
2254 .gfp_mask = sc.gfp_mask,
2255 };
Balbir Singh66e17072008-02-07 00:13:56 -08002256
Mel Gorman33906bc2010-08-09 17:19:16 -07002257 trace_mm_vmscan_direct_reclaim_begin(order,
2258 sc.may_writepage,
2259 gfp_mask);
2260
Ying Hana09ed5e2011-05-24 17:12:26 -07002261 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002262
2263 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2264
2265 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002266}
2267
Balbir Singh00f0b822008-03-04 14:28:39 -08002268#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002269
Balbir Singh4e416952009-09-23 15:56:39 -07002270unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002271 gfp_t gfp_mask, bool noswap,
2272 struct zone *zone,
2273 struct memcg_scanrecord *rec,
2274 unsigned long *scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002275{
2276 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002277 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002278 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002279 .may_writepage = !laptop_mode,
2280 .may_unmap = 1,
2281 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002282 .order = 0,
2283 .mem_cgroup = mem,
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002284 .memcg_record = rec,
Balbir Singh4e416952009-09-23 15:56:39 -07002285 };
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002286 unsigned long start, end;
Ying Han0ae5e892011-05-26 16:25:25 -07002287
Balbir Singh4e416952009-09-23 15:56:39 -07002288 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2289 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002290
2291 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2292 sc.may_writepage,
2293 sc.gfp_mask);
2294
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002295 start = sched_clock();
Balbir Singh4e416952009-09-23 15:56:39 -07002296 /*
2297 * NOTE: Although we can get the priority field, using it
2298 * here is not a good idea, since it limits the pages we can scan.
2299 * if we don't reclaim here, the shrink_zone from balance_pgdat
2300 * will pick up pages from other mem cgroup's as well. We hack
2301 * the priority and make it zero.
2302 */
2303 shrink_zone(0, zone, &sc);
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002304 end = sched_clock();
2305
2306 if (rec)
2307 rec->elapsed += end - start;
2308 *scanned = sc.nr_scanned;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002309
2310 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2311
Balbir Singh4e416952009-09-23 15:56:39 -07002312 return sc.nr_reclaimed;
2313}
2314
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002315unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002316 gfp_t gfp_mask,
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002317 bool noswap,
2318 struct memcg_scanrecord *rec)
Balbir Singh66e17072008-02-07 00:13:56 -08002319{
Balbir Singh4e416952009-09-23 15:56:39 -07002320 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002321 unsigned long nr_reclaimed;
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002322 unsigned long start, end;
Ying Han889976d2011-05-26 16:25:33 -07002323 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002324 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002325 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002326 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002327 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002328 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002329 .order = 0,
2330 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002331 .memcg_record = rec,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002332 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002333 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2334 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2335 };
2336 struct shrink_control shrink = {
2337 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002338 };
Balbir Singh66e17072008-02-07 00:13:56 -08002339
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002340 start = sched_clock();
Ying Han889976d2011-05-26 16:25:33 -07002341 /*
2342 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2343 * take care of from where we get pages. So the node where we start the
2344 * scan does not need to be the current node.
2345 */
2346 nid = mem_cgroup_select_victim_node(mem_cont);
2347
2348 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002349
2350 trace_mm_vmscan_memcg_reclaim_begin(0,
2351 sc.may_writepage,
2352 sc.gfp_mask);
2353
Ying Hana09ed5e2011-05-24 17:12:26 -07002354 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KAMEZAWA Hiroyuki82f9d482011-07-26 16:08:26 -07002355 end = sched_clock();
2356 if (rec)
2357 rec->elapsed += end - start;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002358
2359 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2360
2361 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002362}
2363#endif
2364
Mel Gorman1741c872011-01-13 15:46:21 -08002365/*
2366 * pgdat_balanced is used when checking if a node is balanced for high-order
2367 * allocations. Only zones that meet watermarks and are in a zone allowed
2368 * by the callers classzone_idx are added to balanced_pages. The total of
2369 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2370 * for the node to be considered balanced. Forcing all zones to be balanced
2371 * for high orders can cause excessive reclaim when there are imbalanced zones.
2372 * The choice of 25% is due to
2373 * o a 16M DMA zone that is balanced will not balance a zone on any
2374 * reasonable sized machine
2375 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002376 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002377 * would need to be at least 256M for it to be balance a whole node.
2378 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2379 * to balance a node on its own. These seemed like reasonable ratios.
2380 */
2381static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2382 int classzone_idx)
2383{
2384 unsigned long present_pages = 0;
2385 int i;
2386
2387 for (i = 0; i <= classzone_idx; i++)
2388 present_pages += pgdat->node_zones[i].present_pages;
2389
Shaohua Li4746efd2011-07-19 08:49:26 -07002390 /* A special case here: if zone has no page, we think it's balanced */
2391 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002392}
2393
Mel Gormanf50de2d2009-12-14 17:58:53 -08002394/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002395static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2396 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002397{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002398 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002399 unsigned long balanced = 0;
2400 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002401
2402 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2403 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002404 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002405
Mel Gorman0abdee22011-01-13 15:46:22 -08002406 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002407 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002408 struct zone *zone = pgdat->node_zones + i;
2409
2410 if (!populated_zone(zone))
2411 continue;
2412
Mel Gorman355b09c2011-01-13 15:46:24 -08002413 /*
2414 * balance_pgdat() skips over all_unreclaimable after
2415 * DEF_PRIORITY. Effectively, it considers them balanced so
2416 * they must be considered balanced here as well if kswapd
2417 * is to sleep
2418 */
2419 if (zone->all_unreclaimable) {
2420 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002421 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002422 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002423
Mel Gorman88f5acf2011-01-13 15:45:41 -08002424 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002425 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002426 all_zones_ok = false;
2427 else
2428 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002429 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002430
Mel Gorman1741c872011-01-13 15:46:21 -08002431 /*
2432 * For high-order requests, the balanced zones must contain at least
2433 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2434 * must be balanced
2435 */
2436 if (order)
Johannes Weinerafc7e322011-05-24 17:11:09 -07002437 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002438 else
2439 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002440}
2441
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442/*
2443 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002444 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002446 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 *
2448 * There is special handling here for zones which are full of pinned pages.
2449 * This can happen if the pages are all mlocked, or if they are all used by
2450 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2451 * What we do is to detect the case where all pages in the zone have been
2452 * scanned twice and there has been zero successful reclaim. Mark the zone as
2453 * dead and from now on, only perform a short scan. Basically we're polling
2454 * the zone for when the problem goes away.
2455 *
2456 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002457 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2458 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2459 * lower zones regardless of the number of free pages in the lower zones. This
2460 * interoperates with the page allocator fallback scheme to ensure that aging
2461 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 */
Mel Gorman99504742011-01-13 15:46:20 -08002463static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002464 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002467 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 int priority;
2469 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002470 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002471 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002473 unsigned long nr_soft_reclaimed;
2474 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002475 struct scan_control sc = {
2476 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002477 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002478 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002479 /*
2480 * kswapd doesn't want to be bailed out while reclaim. because
2481 * we want to put equal scanning pressure on each zone.
2482 */
2483 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002484 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002485 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002486 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002487 struct shrink_control shrink = {
2488 .gfp_mask = sc.gfp_mask,
2489 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490loop_again:
2491 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002492 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002493 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002494 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002498 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002500 /* The swap token gets in the way of swapout... */
2501 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002502 disable_swap_token(NULL);
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002503
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002505 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002507 /*
2508 * Scan in the highmem->dma direction for the highest
2509 * zone which needs scanning
2510 */
2511 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2512 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002514 if (!populated_zone(zone))
2515 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002517 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002518 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519
Rik van Riel556adec2008-10-18 20:26:34 -07002520 /*
2521 * Do some background aging of the anon list, to give
2522 * pages a chance to be referenced before reclaiming.
2523 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002524 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002525 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2526 &sc, priority, 0);
2527
Mel Gorman88f5acf2011-01-13 15:45:41 -08002528 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002529 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002530 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002531 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002534 if (i < 0)
2535 goto out;
2536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 for (i = 0; i <= end_zone; i++) {
2538 struct zone *zone = pgdat->node_zones + i;
2539
Wu Fengguangadea02a2009-09-21 17:01:42 -07002540 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 }
2542
2543 /*
2544 * Now scan the zone in the dma->highmem direction, stopping
2545 * at the last zone which needs scanning.
2546 *
2547 * We do this because the page allocator works in the opposite
2548 * direction. This prevents the page allocator from allocating
2549 * pages behind kswapd's direction of progress, which would
2550 * cause too much scanning of the lower zones.
2551 */
2552 for (i = 0; i <= end_zone; i++) {
2553 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002554 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002555 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556
Con Kolivasf3fe6512006-01-06 00:11:15 -08002557 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 continue;
2559
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002560 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 continue;
2562
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002564
Ying Han0ae5e892011-05-26 16:25:25 -07002565 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002566 /*
2567 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002568 */
Ying Han0ae5e892011-05-26 16:25:25 -07002569 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2570 order, sc.gfp_mask,
2571 &nr_soft_scanned);
2572 sc.nr_reclaimed += nr_soft_reclaimed;
2573 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002574
Rik van Riel32a43302007-10-16 01:24:50 -07002575 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002576 * We put equal pressure on every zone, unless
2577 * one zone has way too many pages free
2578 * already. The "too many pages" is defined
2579 * as the high wmark plus a "gap" where the
2580 * gap is either the low watermark or 1%
2581 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002582 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002583 balance_gap = min(low_wmark_pages(zone),
2584 (zone->present_pages +
2585 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2586 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002587 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002588 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002589 end_zone, 0)) {
Rik van Riela79311c2009-01-06 14:40:01 -08002590 shrink_zone(priority, zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002591
Mel Gormand7868da2011-07-08 15:39:38 -07002592 reclaim_state->reclaimed_slab = 0;
2593 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2594 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2595 total_scanned += sc.nr_scanned;
2596
2597 if (nr_slab == 0 && !zone_reclaimable(zone))
2598 zone->all_unreclaimable = 1;
2599 }
2600
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 /*
2602 * If we've done a decent amount of scanning and
2603 * the reclaim ratio is low, start doing writepage
2604 * even in laptop mode
2605 */
2606 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002607 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002609
Mel Gorman215ddd62011-07-08 15:39:40 -07002610 if (zone->all_unreclaimable) {
2611 if (end_zone && end_zone == i)
2612 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002613 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002614 }
Mel Gormand7868da2011-07-08 15:39:38 -07002615
Mel Gorman88f5acf2011-01-13 15:45:41 -08002616 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002617 high_wmark_pages(zone), end_zone, 0)) {
2618 all_zones_ok = 0;
2619 /*
2620 * We are still under min water mark. This
2621 * means that we have a GFP_ATOMIC allocation
2622 * failure risk. Hurry up!
2623 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002624 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002625 min_wmark_pages(zone), end_zone, 0))
2626 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002627 } else {
2628 /*
2629 * If a zone reaches its high watermark,
2630 * consider it to be no longer congested. It's
2631 * possible there are dirty pages backed by
2632 * congested BDIs but as pressure is relieved,
2633 * spectulatively avoid congestion waits
2634 */
2635 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002636 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002637 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002638 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002639
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002641 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 break; /* kswapd: all done */
2643 /*
2644 * OK, kswapd is getting into trouble. Take a nap, then take
2645 * another pass across the zones.
2646 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002647 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2648 if (has_under_min_watermark_zone)
2649 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2650 else
2651 congestion_wait(BLK_RW_ASYNC, HZ/10);
2652 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
2654 /*
2655 * We do this so kswapd doesn't build up large priorities for
2656 * example when it is freeing in parallel with allocators. It
2657 * matches the direct reclaim path behaviour in terms of impact
2658 * on zone->*_priority.
2659 */
Rik van Riela79311c2009-01-06 14:40:01 -08002660 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 break;
2662 }
2663out:
Mel Gorman99504742011-01-13 15:46:20 -08002664
2665 /*
2666 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002667 * high-order: Balanced zones must make up at least 25% of the node
2668 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002669 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002670 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002672
2673 try_to_freeze();
2674
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002675 /*
2676 * Fragmentation may mean that the system cannot be
2677 * rebalanced for high-order allocations in all zones.
2678 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2679 * it means the zones have been fully scanned and are still
2680 * not balanced. For high-order allocations, there is
2681 * little point trying all over again as kswapd may
2682 * infinite loop.
2683 *
2684 * Instead, recheck all watermarks at order-0 as they
2685 * are the most important. If watermarks are ok, kswapd will go
2686 * back to sleep. High-order users can still perform direct
2687 * reclaim if they wish.
2688 */
2689 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2690 order = sc.order = 0;
2691
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 goto loop_again;
2693 }
2694
Mel Gorman99504742011-01-13 15:46:20 -08002695 /*
2696 * If kswapd was reclaiming at a higher order, it has the option of
2697 * sleeping without all zones being balanced. Before it does, it must
2698 * ensure that the watermarks for order-0 on *all* zones are met and
2699 * that the congestion flags are cleared. The congestion flag must
2700 * be cleared as kswapd is the only mechanism that clears the flag
2701 * and it is potentially going to sleep here.
2702 */
2703 if (order) {
2704 for (i = 0; i <= end_zone; i++) {
2705 struct zone *zone = pgdat->node_zones + i;
2706
2707 if (!populated_zone(zone))
2708 continue;
2709
2710 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2711 continue;
2712
2713 /* Confirm the zone is balanced for order-0 */
2714 if (!zone_watermark_ok(zone, 0,
2715 high_wmark_pages(zone), 0, 0)) {
2716 order = sc.order = 0;
2717 goto loop_again;
2718 }
2719
2720 /* If balanced, clear the congested flag */
2721 zone_clear_flag(zone, ZONE_CONGESTED);
2722 }
2723 }
2724
Mel Gorman0abdee22011-01-13 15:46:22 -08002725 /*
2726 * Return the order we were reclaiming at so sleeping_prematurely()
2727 * makes a decision on the order we were last reclaiming at. However,
2728 * if another caller entered the allocator slow path while kswapd
2729 * was awake, order will remain at the higher level
2730 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002731 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002732 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733}
2734
Mel Gormandc83edd2011-01-13 15:46:26 -08002735static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002736{
2737 long remaining = 0;
2738 DEFINE_WAIT(wait);
2739
2740 if (freezing(current) || kthread_should_stop())
2741 return;
2742
2743 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2744
2745 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002746 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002747 remaining = schedule_timeout(HZ/10);
2748 finish_wait(&pgdat->kswapd_wait, &wait);
2749 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2750 }
2751
2752 /*
2753 * After a short sleep, check if it was a premature sleep. If not, then
2754 * go fully to sleep until explicitly woken up.
2755 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002756 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002757 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2758
2759 /*
2760 * vmstat counters are not perfectly accurate and the estimated
2761 * value for counters such as NR_FREE_PAGES can deviate from the
2762 * true value by nr_online_cpus * threshold. To avoid the zone
2763 * watermarks being breached while under pressure, we reduce the
2764 * per-cpu vmstat threshold while kswapd is awake and restore
2765 * them before going back to sleep.
2766 */
2767 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2768 schedule();
2769 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2770 } else {
2771 if (remaining)
2772 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2773 else
2774 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2775 }
2776 finish_wait(&pgdat->kswapd_wait, &wait);
2777}
2778
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779/*
2780 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002781 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 *
2783 * This basically trickles out pages so that we have _some_
2784 * free memory available even if there is no other activity
2785 * that frees anything up. This is needed for things like routing
2786 * etc, where we otherwise might have all activity going on in
2787 * asynchronous contexts that cannot page things out.
2788 *
2789 * If there are applications that are active memory-allocators
2790 * (most normal use), this basically shouldn't matter.
2791 */
2792static int kswapd(void *p)
2793{
Mel Gorman215ddd62011-07-08 15:39:40 -07002794 unsigned long order, new_order;
2795 int classzone_idx, new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 pg_data_t *pgdat = (pg_data_t*)p;
2797 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002798
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 struct reclaim_state reclaim_state = {
2800 .reclaimed_slab = 0,
2801 };
Rusty Russella70f7302009-03-13 14:49:46 +10302802 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803
Nick Piggincf40bd12009-01-21 08:12:39 +01002804 lockdep_set_current_reclaim_state(GFP_KERNEL);
2805
Rusty Russell174596a2009-01-01 10:12:29 +10302806 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002807 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 current->reclaim_state = &reclaim_state;
2809
2810 /*
2811 * Tell the memory management that we're a "memory allocator",
2812 * and that if we need more memory we should get access to it
2813 * regardless (see "__alloc_pages()"). "kswapd" should
2814 * never get caught in the normal page freeing logic.
2815 *
2816 * (Kswapd normally doesn't need memory anyway, but sometimes
2817 * you need a small amount of memory in order to be able to
2818 * page out something else, and this flag essentially protects
2819 * us from recursively trying to free more memory as we're
2820 * trying to free the first piece of memory in the first place).
2821 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002822 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002823 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
Mel Gorman215ddd62011-07-08 15:39:40 -07002825 order = new_order = 0;
2826 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08002828 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002829
Mel Gorman215ddd62011-07-08 15:39:40 -07002830 /*
2831 * If the last balance_pgdat was unsuccessful it's unlikely a
2832 * new request of a similar or harder type will succeed soon
2833 * so consider going to sleep on the basis we reclaimed at
2834 */
2835 if (classzone_idx >= new_classzone_idx && order == new_order) {
2836 new_order = pgdat->kswapd_max_order;
2837 new_classzone_idx = pgdat->classzone_idx;
2838 pgdat->kswapd_max_order = 0;
2839 pgdat->classzone_idx = pgdat->nr_zones - 1;
2840 }
2841
Mel Gorman99504742011-01-13 15:46:20 -08002842 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 /*
2844 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002845 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 */
2847 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002848 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 } else {
Mel Gormandc83edd2011-01-13 15:46:26 -08002850 kswapd_try_to_sleep(pgdat, order, classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002852 classzone_idx = pgdat->classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002853 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002854 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
David Rientjes8fe23e02009-12-14 17:58:33 -08002857 ret = try_to_freeze();
2858 if (kthread_should_stop())
2859 break;
2860
2861 /*
2862 * We can speed up thawing tasks if we don't call balance_pgdat
2863 * after returning from the refrigerator
2864 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002865 if (!ret) {
2866 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Mel Gormandc83edd2011-01-13 15:46:26 -08002867 order = balance_pgdat(pgdat, order, &classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002868 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 }
2870 return 0;
2871}
2872
2873/*
2874 * A zone is low on free memory, so wake its kswapd task to service it.
2875 */
Mel Gorman99504742011-01-13 15:46:20 -08002876void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877{
2878 pg_data_t *pgdat;
2879
Con Kolivasf3fe6512006-01-06 00:11:15 -08002880 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 return;
2882
Paul Jackson02a0e532006-12-13 00:34:25 -08002883 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002885 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002886 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002887 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002888 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2889 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002890 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002892 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2893 return;
2894
2895 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002896 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897}
2898
Wu Fengguangadea02a2009-09-21 17:01:42 -07002899/*
2900 * The reclaimable count would be mostly accurate.
2901 * The less reclaimable pages may be
2902 * - mlocked pages, which will be moved to unevictable list when encountered
2903 * - mapped pages, which may require several travels to be reclaimed
2904 * - dirty pages, which is not "instantly" reclaimable
2905 */
2906unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002907{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002908 int nr;
2909
2910 nr = global_page_state(NR_ACTIVE_FILE) +
2911 global_page_state(NR_INACTIVE_FILE);
2912
2913 if (nr_swap_pages > 0)
2914 nr += global_page_state(NR_ACTIVE_ANON) +
2915 global_page_state(NR_INACTIVE_ANON);
2916
2917 return nr;
2918}
2919
2920unsigned long zone_reclaimable_pages(struct zone *zone)
2921{
2922 int nr;
2923
2924 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2925 zone_page_state(zone, NR_INACTIVE_FILE);
2926
2927 if (nr_swap_pages > 0)
2928 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2929 zone_page_state(zone, NR_INACTIVE_ANON);
2930
2931 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002932}
2933
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002934#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002936 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002937 * freed pages.
2938 *
2939 * Rather than trying to age LRUs the aim is to preserve the overall
2940 * LRU order by reclaiming preferentially
2941 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002943unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002945 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002946 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002947 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2948 .may_swap = 1,
2949 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002950 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002951 .nr_to_reclaim = nr_to_reclaim,
2952 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002953 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002955 struct shrink_control shrink = {
2956 .gfp_mask = sc.gfp_mask,
2957 };
2958 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002959 struct task_struct *p = current;
2960 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002962 p->flags |= PF_MEMALLOC;
2963 lockdep_set_current_reclaim_state(sc.gfp_mask);
2964 reclaim_state.reclaimed_slab = 0;
2965 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002966
Ying Hana09ed5e2011-05-24 17:12:26 -07002967 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002968
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002969 p->reclaim_state = NULL;
2970 lockdep_clear_current_reclaim_state();
2971 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002972
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002973 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002975#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977/* It's optimal to keep kswapds on the same CPUs as their memory, but
2978 not required for correctness. So if the last cpu in a node goes
2979 away, we get changed to run anywhere: as the first one comes back,
2980 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002981static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002982 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002984 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002986 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002987 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002988 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302989 const struct cpumask *mask;
2990
2991 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002992
Rusty Russell3e597942009-01-01 10:12:24 +10302993 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002995 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 }
2997 }
2998 return NOTIFY_OK;
2999}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000
Yasunori Goto3218ae12006-06-27 02:53:33 -07003001/*
3002 * This kswapd start function will be called by init and node-hot-add.
3003 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3004 */
3005int kswapd_run(int nid)
3006{
3007 pg_data_t *pgdat = NODE_DATA(nid);
3008 int ret = 0;
3009
3010 if (pgdat->kswapd)
3011 return 0;
3012
3013 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3014 if (IS_ERR(pgdat->kswapd)) {
3015 /* failure at boot is fatal */
3016 BUG_ON(system_state == SYSTEM_BOOTING);
3017 printk("Failed to start kswapd on node %d\n",nid);
3018 ret = -1;
3019 }
3020 return ret;
3021}
3022
David Rientjes8fe23e02009-12-14 17:58:33 -08003023/*
3024 * Called by memory hotplug when all memory in a node is offlined.
3025 */
3026void kswapd_stop(int nid)
3027{
3028 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3029
3030 if (kswapd)
3031 kthread_stop(kswapd);
3032}
3033
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034static int __init kswapd_init(void)
3035{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003036 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003037
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003039 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003040 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 hotcpu_notifier(cpu_callback, 0);
3042 return 0;
3043}
3044
3045module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003046
3047#ifdef CONFIG_NUMA
3048/*
3049 * Zone reclaim mode
3050 *
3051 * If non-zero call zone_reclaim when the number of free pages falls below
3052 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003053 */
3054int zone_reclaim_mode __read_mostly;
3055
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003056#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003057#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003058#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3059#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3060
Christoph Lameter9eeff232006-01-18 17:42:31 -08003061/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003062 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3063 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3064 * a zone.
3065 */
3066#define ZONE_RECLAIM_PRIORITY 4
3067
Christoph Lameter9eeff232006-01-18 17:42:31 -08003068/*
Christoph Lameter96146342006-07-03 00:24:13 -07003069 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3070 * occur.
3071 */
3072int sysctl_min_unmapped_ratio = 1;
3073
3074/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003075 * If the number of slab pages in a zone grows beyond this percentage then
3076 * slab reclaim needs to occur.
3077 */
3078int sysctl_min_slab_ratio = 5;
3079
Mel Gorman90afa5d2009-06-16 15:33:20 -07003080static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3081{
3082 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3083 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3084 zone_page_state(zone, NR_ACTIVE_FILE);
3085
3086 /*
3087 * It's possible for there to be more file mapped pages than
3088 * accounted for by the pages on the file LRU lists because
3089 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3090 */
3091 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3092}
3093
3094/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3095static long zone_pagecache_reclaimable(struct zone *zone)
3096{
3097 long nr_pagecache_reclaimable;
3098 long delta = 0;
3099
3100 /*
3101 * If RECLAIM_SWAP is set, then all file pages are considered
3102 * potentially reclaimable. Otherwise, we have to worry about
3103 * pages like swapcache and zone_unmapped_file_pages() provides
3104 * a better estimate
3105 */
3106 if (zone_reclaim_mode & RECLAIM_SWAP)
3107 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3108 else
3109 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3110
3111 /* If we can't clean pages, remove dirty pages from consideration */
3112 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3113 delta += zone_page_state(zone, NR_FILE_DIRTY);
3114
3115 /* Watch for any possible underflows due to delta */
3116 if (unlikely(delta > nr_pagecache_reclaimable))
3117 delta = nr_pagecache_reclaimable;
3118
3119 return nr_pagecache_reclaimable - delta;
3120}
3121
Christoph Lameter0ff38492006-09-25 23:31:52 -07003122/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003123 * Try to free up some pages from this zone through reclaim.
3124 */
Andrew Morton179e9632006-03-22 00:08:18 -08003125static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003126{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003127 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003128 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003129 struct task_struct *p = current;
3130 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08003131 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08003132 struct scan_control sc = {
3133 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003134 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003135 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003136 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3137 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003138 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003139 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003140 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003141 struct shrink_control shrink = {
3142 .gfp_mask = sc.gfp_mask,
3143 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003144 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003145
Christoph Lameter9eeff232006-01-18 17:42:31 -08003146 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003147 /*
3148 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3149 * and we also need to be able to write out pages for RECLAIM_WRITE
3150 * and RECLAIM_SWAP.
3151 */
3152 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003153 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003154 reclaim_state.reclaimed_slab = 0;
3155 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003156
Mel Gorman90afa5d2009-06-16 15:33:20 -07003157 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003158 /*
3159 * Free memory by calling shrink zone with increasing
3160 * priorities until we have enough memory freed.
3161 */
3162 priority = ZONE_RECLAIM_PRIORITY;
3163 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003164 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003165 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003166 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003167 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003168
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003169 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3170 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003171 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003172 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003173 * many pages were freed in this zone. So we take the current
3174 * number of slab pages and shake the slab until it is reduced
3175 * by the same nr_pages that we used for reclaiming unmapped
3176 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003177 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003178 * Note that shrink_slab will free memory on all zones and may
3179 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003180 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003181 for (;;) {
3182 unsigned long lru_pages = zone_reclaimable_pages(zone);
3183
3184 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003185 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003186 break;
3187
3188 /* Freed enough memory */
3189 nr_slab_pages1 = zone_page_state(zone,
3190 NR_SLAB_RECLAIMABLE);
3191 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3192 break;
3193 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003194
3195 /*
3196 * Update nr_reclaimed by the number of slab pages we
3197 * reclaimed from this zone.
3198 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003199 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3200 if (nr_slab_pages1 < nr_slab_pages0)
3201 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003202 }
3203
Christoph Lameter9eeff232006-01-18 17:42:31 -08003204 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003205 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003206 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003207 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003208}
Andrew Morton179e9632006-03-22 00:08:18 -08003209
3210int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3211{
Andrew Morton179e9632006-03-22 00:08:18 -08003212 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003213 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003214
3215 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003216 * Zone reclaim reclaims unmapped file backed pages and
3217 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003218 *
Christoph Lameter96146342006-07-03 00:24:13 -07003219 * A small portion of unmapped file backed pages is needed for
3220 * file I/O otherwise pages read by file I/O will be immediately
3221 * thrown out if the zone is overallocated. So we do not reclaim
3222 * if less than a specified percentage of the zone is used by
3223 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003224 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003225 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3226 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003227 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003228
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003229 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003230 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003231
Andrew Morton179e9632006-03-22 00:08:18 -08003232 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003233 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003234 */
David Rientjesd773ed62007-10-16 23:26:01 -07003235 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003236 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003237
3238 /*
3239 * Only run zone reclaim on the local zone or on zones that do not
3240 * have associated processors. This will favor the local processor
3241 * over remote processors and spread off node memory allocations
3242 * as wide as possible.
3243 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003244 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003245 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003246 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003247
3248 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003249 return ZONE_RECLAIM_NOSCAN;
3250
David Rientjesd773ed62007-10-16 23:26:01 -07003251 ret = __zone_reclaim(zone, gfp_mask, order);
3252 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3253
Mel Gorman24cf725182009-06-16 15:33:23 -07003254 if (!ret)
3255 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3256
David Rientjesd773ed62007-10-16 23:26:01 -07003257 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003258}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003259#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003260
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003261/*
3262 * page_evictable - test whether a page is evictable
3263 * @page: the page to test
3264 * @vma: the VMA in which the page is or will be mapped, may be NULL
3265 *
3266 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003267 * lists vs unevictable list. The vma argument is !NULL when called from the
3268 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003269 *
3270 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003271 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003272 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003273 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003274 */
3275int page_evictable(struct page *page, struct vm_area_struct *vma)
3276{
3277
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003278 if (mapping_unevictable(page_mapping(page)))
3279 return 0;
3280
Nick Pigginb291f002008-10-18 20:26:44 -07003281 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3282 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003283
3284 return 1;
3285}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003286
3287/**
3288 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3289 * @page: page to check evictability and move to appropriate lru list
3290 * @zone: zone page is in
3291 *
3292 * Checks a page for evictability and moves the page to the appropriate
3293 * zone lru list.
3294 *
3295 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3296 * have PageUnevictable set.
3297 */
3298static void check_move_unevictable_page(struct page *page, struct zone *zone)
3299{
3300 VM_BUG_ON(PageActive(page));
3301
3302retry:
3303 ClearPageUnevictable(page);
3304 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003305 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003306
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003307 __dec_zone_state(zone, NR_UNEVICTABLE);
3308 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003309 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003310 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3311 __count_vm_event(UNEVICTABLE_PGRESCUED);
3312 } else {
3313 /*
3314 * rotate unevictable list
3315 */
3316 SetPageUnevictable(page);
3317 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003318 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003319 if (page_evictable(page, NULL))
3320 goto retry;
3321 }
3322}
3323
3324/**
3325 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3326 * @mapping: struct address_space to scan for evictable pages
3327 *
3328 * Scan all pages in mapping. Check unevictable pages for
3329 * evictability and move them to the appropriate zone lru list.
3330 */
3331void scan_mapping_unevictable_pages(struct address_space *mapping)
3332{
3333 pgoff_t next = 0;
3334 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3335 PAGE_CACHE_SHIFT;
3336 struct zone *zone;
3337 struct pagevec pvec;
3338
3339 if (mapping->nrpages == 0)
3340 return;
3341
3342 pagevec_init(&pvec, 0);
3343 while (next < end &&
3344 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3345 int i;
3346 int pg_scanned = 0;
3347
3348 zone = NULL;
3349
3350 for (i = 0; i < pagevec_count(&pvec); i++) {
3351 struct page *page = pvec.pages[i];
3352 pgoff_t page_index = page->index;
3353 struct zone *pagezone = page_zone(page);
3354
3355 pg_scanned++;
3356 if (page_index > next)
3357 next = page_index;
3358 next++;
3359
3360 if (pagezone != zone) {
3361 if (zone)
3362 spin_unlock_irq(&zone->lru_lock);
3363 zone = pagezone;
3364 spin_lock_irq(&zone->lru_lock);
3365 }
3366
3367 if (PageLRU(page) && PageUnevictable(page))
3368 check_move_unevictable_page(page, zone);
3369 }
3370 if (zone)
3371 spin_unlock_irq(&zone->lru_lock);
3372 pagevec_release(&pvec);
3373
3374 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3375 }
3376
3377}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003378
3379/**
3380 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3381 * @zone - zone of which to scan the unevictable list
3382 *
3383 * Scan @zone's unevictable LRU lists to check for pages that have become
3384 * evictable. Move those that have to @zone's inactive list where they
3385 * become candidates for reclaim, unless shrink_inactive_zone() decides
3386 * to reactivate them. Pages that are still unevictable are rotated
3387 * back onto @zone's unevictable list.
3388 */
3389#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08003390static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003391{
3392 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3393 unsigned long scan;
3394 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3395
3396 while (nr_to_scan > 0) {
3397 unsigned long batch_size = min(nr_to_scan,
3398 SCAN_UNEVICTABLE_BATCH_SIZE);
3399
3400 spin_lock_irq(&zone->lru_lock);
3401 for (scan = 0; scan < batch_size; scan++) {
3402 struct page *page = lru_to_page(l_unevictable);
3403
3404 if (!trylock_page(page))
3405 continue;
3406
3407 prefetchw_prev_lru_page(page, l_unevictable, flags);
3408
3409 if (likely(PageLRU(page) && PageUnevictable(page)))
3410 check_move_unevictable_page(page, zone);
3411
3412 unlock_page(page);
3413 }
3414 spin_unlock_irq(&zone->lru_lock);
3415
3416 nr_to_scan -= batch_size;
3417 }
3418}
3419
3420
3421/**
3422 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3423 *
3424 * A really big hammer: scan all zones' unevictable LRU lists to check for
3425 * pages that have become evictable. Move those back to the zones'
3426 * inactive list where they become candidates for reclaim.
3427 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3428 * and we add swap to the system. As such, it runs in the context of a task
3429 * that has possibly/probably made some previously unevictable pages
3430 * evictable.
3431 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003432static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003433{
3434 struct zone *zone;
3435
3436 for_each_zone(zone) {
3437 scan_zone_unevictable_pages(zone);
3438 }
3439}
3440
3441/*
3442 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3443 * all nodes' unevictable lists for evictable pages
3444 */
3445unsigned long scan_unevictable_pages;
3446
3447int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003448 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003449 size_t *length, loff_t *ppos)
3450{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003451 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003452
3453 if (write && *(unsigned long *)table->data)
3454 scan_all_zones_unevictable_pages();
3455
3456 scan_unevictable_pages = 0;
3457 return 0;
3458}
3459
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003460#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003461/*
3462 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3463 * a specified node's per zone unevictable lists for evictable pages.
3464 */
3465
3466static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3467 struct sysdev_attribute *attr,
3468 char *buf)
3469{
3470 return sprintf(buf, "0\n"); /* always zero; should fit... */
3471}
3472
3473static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3474 struct sysdev_attribute *attr,
3475 const char *buf, size_t count)
3476{
3477 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3478 struct zone *zone;
3479 unsigned long res;
3480 unsigned long req = strict_strtoul(buf, 10, &res);
3481
3482 if (!req)
3483 return 1; /* zero is no-op */
3484
3485 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3486 if (!populated_zone(zone))
3487 continue;
3488 scan_zone_unevictable_pages(zone);
3489 }
3490 return 1;
3491}
3492
3493
3494static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3495 read_scan_unevictable_node,
3496 write_scan_unevictable_node);
3497
3498int scan_unevictable_register_node(struct node *node)
3499{
3500 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3501}
3502
3503void scan_unevictable_unregister_node(struct node *node)
3504{
3505 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3506}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003507#endif