blob: 132d1ddb2238179466f45a599b1ccd687f9ee3f8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51
Nick Piggin0f8053a2006-03-22 00:08:33 -080052#include "internal.h"
53
Mel Gorman33906bc2010-08-09 17:19:16 -070054#define CREATE_TRACE_POINTS
55#include <trace/events/vmscan.h>
56
Mel Gormanee64fc92011-01-13 15:45:55 -080057/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080058 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080063 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080065 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080066 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080067 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080068typedef unsigned __bitwise__ reclaim_mode_t;
69#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070074
Linus Torvalds1da177e2005-04-16 15:20:36 -070075struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
78
Rik van Riela79311c2009-01-06 14:40:01 -080079 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
81
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080082 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
84
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080085 unsigned long hibernation_mode;
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040088 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 int may_writepage;
91
Johannes Weinera6dc60f2009-03-31 15:19:30 -070092 /* Can mapped pages be reclaimed? */
93 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080094
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070095 /* Can pages be swapped as part of reclaim? */
96 int may_swap;
97
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070098 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080099
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700100 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700103 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800104 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700105
Balbir Singh66e17072008-02-07 00:13:56 -0800106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
108
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700109 /*
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
111 * are scanned.
112 */
113 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114};
115
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118#ifdef ARCH_HAS_PREFETCH
119#define prefetch_prev_lru_page(_page, _base, _field) \
120 do { \
121 if ((_page)->lru.prev != _base) { \
122 struct page *prev; \
123 \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
126 } \
127 } while (0)
128#else
129#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130#endif
131
132#ifdef ARCH_HAS_PREFETCHW
133#define prefetchw_prev_lru_page(_page, _base, _field) \
134 do { \
135 if ((_page)->lru.prev != _base) { \
136 struct page *prev; \
137 \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
140 } \
141 } while (0)
142#else
143#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144#endif
145
146/*
147 * From 0 .. 100. Higher means more swappy.
148 */
149int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700150long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
152static LIST_HEAD(shrinker_list);
153static DECLARE_RWSEM(shrinker_rwsem);
154
Balbir Singh00f0b822008-03-04 14:28:39 -0800155#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800156#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800157#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800158#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#endif
160
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800161static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
163{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800164 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800167 return &zone->reclaim_stat;
168}
169
Vincent Li0b217672009-09-21 17:03:09 -0700170static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800172{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800173 if (!scanning_global_lru(sc))
KAMEZAWA Hiroyukibb2a0de2011-07-26 16:08:22 -0700174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175 zone_to_nid(zone), zone_idx(zone), BIT(lru));
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800176
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800177 return zone_page_state(zone, NR_LRU_BASE + lru);
178}
179
180
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181/*
182 * Add a shrinker callback to be called from the vm
183 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700184void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700186 shrinker->nr = 0;
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700191EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
193/*
194 * Remove one
195 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700196void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197{
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700202EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
Ying Han1495f232011-05-24 17:12:27 -0700204static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
207{
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
210}
211
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212#define SHRINK_BATCH 128
213/*
214 * Call the shrink functions to age shrinkable caches
215 *
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
220 *
Simon Arlott183ff222007-10-20 01:27:18 +0200221 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 * slab to avoid swapping.
223 *
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
225 *
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700229 *
230 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700232unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700233 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700234 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235{
236 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800237 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
Ying Han1495f232011-05-24 17:12:27 -0700239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241
Minchan Kimf06590b2011-05-24 17:11:11 -0700242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
244 ret = 1;
245 goto out;
246 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
250 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000251 unsigned long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000252 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000253 long nr;
254 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000255 long batch_size = shrinker->batch ? shrinker->batch
256 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
Dave Chinneracf92b42011-07-08 14:14:35 +1000258 /*
259 * copy the current shrinker scan count into a local variable
260 * and zero it so that other concurrent shrinker invocations
261 * don't also do this scanning work.
262 */
263 do {
264 nr = shrinker->nr;
265 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
266
267 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700268 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800270 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000272 total_scan += delta;
273 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700274 printk(KERN_ERR "shrink_slab: %pF negative objects to "
275 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000276 shrinker->shrink, total_scan);
277 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800278 }
279
280 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000281 * We need to avoid excessive windup on filesystem shrinkers
282 * due to large numbers of GFP_NOFS allocations causing the
283 * shrinkers to return -1 all the time. This results in a large
284 * nr being built up so when a shrink that can do some work
285 * comes along it empties the entire cache due to nr >>>
286 * max_pass. This is bad for sustaining a working set in
287 * memory.
288 *
289 * Hence only allow the shrinker to scan the entire cache when
290 * a large delta change is calculated directly.
291 */
292 if (delta < max_pass / 4)
293 total_scan = min(total_scan, max_pass / 2);
294
295 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800296 * Avoid risking looping forever due to too large nr value:
297 * never try to free more than twice the estimate number of
298 * freeable entries.
299 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000300 if (total_scan > max_pass * 2)
301 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302
Dave Chinneracf92b42011-07-08 14:14:35 +1000303 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000304 nr_pages_scanned, lru_pages,
305 max_pass, delta, total_scan);
306
Dave Chinnere9299f52011-07-08 14:14:37 +1000307 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700308 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Ying Han1495f232011-05-24 17:12:27 -0700310 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000312 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 if (shrink_ret == -1)
314 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700315 if (shrink_ret < nr_before)
316 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000317 count_vm_events(SLABS_SCANNED, batch_size);
318 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
320 cond_resched();
321 }
322
Dave Chinneracf92b42011-07-08 14:14:35 +1000323 /*
324 * move the unused scan count back into the shrinker in a
325 * manner that handles concurrent updates. If we exhausted the
326 * scan, there is no need to do an update.
327 */
328 do {
329 nr = shrinker->nr;
330 new_nr = total_scan + nr;
331 if (total_scan <= 0)
332 break;
333 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
334
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 }
337 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700338out:
339 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700340 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
Mel Gormanf3a310b2011-01-13 15:46:00 -0800343static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700344 bool sync)
345{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700347
348 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800349 * Initially assume we are entering either lumpy reclaim or
350 * reclaim/compaction.Depending on the order, we will either set the
351 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700352 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800353 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800355 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700357
358 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800359 * Avoid using lumpy reclaim or reclaim/compaction if possible by
360 * restricting when its set to either costly allocations or when
361 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700362 */
363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800364 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700365 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800366 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700367 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700369}
370
Mel Gormanf3a310b2011-01-13 15:46:00 -0800371static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700372{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700374}
375
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376static inline int is_page_cache_freeable(struct page *page)
377{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700378 /*
379 * A freeable page cache page is referenced only by the caller
380 * that isolated the page, the page cache radix tree and
381 * optional buffer heads at page->private.
382 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700383 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384}
385
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700386static int may_write_to_queue(struct backing_dev_info *bdi,
387 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388{
Christoph Lameter930d9152006-01-08 01:00:47 -0800389 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 return 1;
391 if (!bdi_write_congested(bdi))
392 return 1;
393 if (bdi == current->backing_dev_info)
394 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700395
396 /* lumpy reclaim for hugepage often need a lot of write */
397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 return 0;
400}
401
402/*
403 * We detected a synchronous write error writing a page out. Probably
404 * -ENOSPC. We need to propagate that into the address_space for a subsequent
405 * fsync(), msync() or close().
406 *
407 * The tricky part is that after writepage we cannot touch the mapping: nothing
408 * prevents it from being freed up. But we have a ref on the page and once
409 * that page is locked, the mapping is pinned.
410 *
411 * We're allowed to run sleeping lock_page() here because we know the caller has
412 * __GFP_FS.
413 */
414static void handle_write_error(struct address_space *mapping,
415 struct page *page, int error)
416{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100417 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700418 if (page_mapping(page) == mapping)
419 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 unlock_page(page);
421}
422
Christoph Lameter04e62a22006-06-23 02:03:38 -0700423/* possible outcome of pageout() */
424typedef enum {
425 /* failed to write page out, page is locked */
426 PAGE_KEEP,
427 /* move page to the active list, page is locked */
428 PAGE_ACTIVATE,
429 /* page has been sent to the disk successfully, page is unlocked */
430 PAGE_SUCCESS,
431 /* page is clean and locked */
432 PAGE_CLEAN,
433} pageout_t;
434
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435/*
Andrew Morton1742f192006-03-22 00:08:21 -0800436 * pageout is called by shrink_page_list() for each dirty page.
437 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700439static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700440 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
442 /*
443 * If the page is dirty, only perform writeback if that write
444 * will be non-blocking. To prevent this allocation from being
445 * stalled by pagecache activity. But note that there may be
446 * stalls if we need to run get_block(). We could test
447 * PagePrivate for that.
448 *
Vincent Li6aceb532009-12-14 17:58:49 -0800449 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 * this page's queue, we can perform writeback even if that
451 * will block.
452 *
453 * If the page is swapcache, write it back even if that would
454 * block, for some throttling. This happens by accident, because
455 * swap_backing_dev_info is bust: it doesn't reflect the
456 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 */
458 if (!is_page_cache_freeable(page))
459 return PAGE_KEEP;
460 if (!mapping) {
461 /*
462 * Some data journaling orphaned pages can have
463 * page->mapping == NULL while being dirty with clean buffers.
464 */
David Howells266cf652009-04-03 16:42:36 +0100465 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 if (try_to_free_buffers(page)) {
467 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700468 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 return PAGE_CLEAN;
470 }
471 }
472 return PAGE_KEEP;
473 }
474 if (mapping->a_ops->writepage == NULL)
475 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700476 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return PAGE_KEEP;
478
479 if (clear_page_dirty_for_io(page)) {
480 int res;
481 struct writeback_control wbc = {
482 .sync_mode = WB_SYNC_NONE,
483 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700484 .range_start = 0,
485 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 .for_reclaim = 1,
487 };
488
489 SetPageReclaim(page);
490 res = mapping->a_ops->writepage(page, &wbc);
491 if (res < 0)
492 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800493 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 ClearPageReclaim(page);
495 return PAGE_ACTIVATE;
496 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700497
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 if (!PageWriteback(page)) {
499 /* synchronous write or broken a_ops? */
500 ClearPageReclaim(page);
501 }
Mel Gorman755f0222010-08-09 17:19:18 -0700502 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800503 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700504 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 return PAGE_SUCCESS;
506 }
507
508 return PAGE_CLEAN;
509}
510
Andrew Mortona649fd92006-10-17 00:09:36 -0700511/*
Nick Piggine2867812008-07-25 19:45:30 -0700512 * Same as remove_mapping, but if the page is removed from the mapping, it
513 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700514 */
Nick Piggine2867812008-07-25 19:45:30 -0700515static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800516{
Nick Piggin28e4d962006-09-25 23:31:23 -0700517 BUG_ON(!PageLocked(page));
518 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800519
Nick Piggin19fd6232008-07-25 19:45:32 -0700520 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800521 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700522 * The non racy check for a busy page.
523 *
524 * Must be careful with the order of the tests. When someone has
525 * a ref to the page, it may be possible that they dirty it then
526 * drop the reference. So if PageDirty is tested before page_count
527 * here, then the following race may occur:
528 *
529 * get_user_pages(&page);
530 * [user mapping goes away]
531 * write_to(page);
532 * !PageDirty(page) [good]
533 * SetPageDirty(page);
534 * put_page(page);
535 * !page_count(page) [good, discard it]
536 *
537 * [oops, our write_to data is lost]
538 *
539 * Reversing the order of the tests ensures such a situation cannot
540 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541 * load is not satisfied before that of page->_count.
542 *
543 * Note that if SetPageDirty is always performed via set_page_dirty,
544 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800545 */
Nick Piggine2867812008-07-25 19:45:30 -0700546 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800547 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700548 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549 if (unlikely(PageDirty(page))) {
550 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800551 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700552 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800553
554 if (PageSwapCache(page)) {
555 swp_entry_t swap = { .val = page_private(page) };
556 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700557 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700558 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700559 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500560 void (*freepage)(struct page *);
561
562 freepage = mapping->a_ops->freepage;
563
Minchan Kime64a7822011-03-22 16:32:44 -0700564 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700565 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700566 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500567
568 if (freepage != NULL)
569 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800570 }
571
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800572 return 1;
573
574cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700575 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800576 return 0;
577}
578
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579/*
Nick Piggine2867812008-07-25 19:45:30 -0700580 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
581 * someone else has a ref on the page, abort and return 0. If it was
582 * successfully detached, return 1. Assumes the caller has a single ref on
583 * this page.
584 */
585int remove_mapping(struct address_space *mapping, struct page *page)
586{
587 if (__remove_mapping(mapping, page)) {
588 /*
589 * Unfreezing the refcount with 1 rather than 2 effectively
590 * drops the pagecache ref for us without requiring another
591 * atomic operation.
592 */
593 page_unfreeze_refs(page, 1);
594 return 1;
595 }
596 return 0;
597}
598
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700599/**
600 * putback_lru_page - put previously isolated page onto appropriate LRU list
601 * @page: page to be put back to appropriate lru list
602 *
603 * Add previously isolated @page to appropriate LRU list.
604 * Page may still be unevictable for other reasons.
605 *
606 * lru_lock must not be held, interrupts must be enabled.
607 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700608void putback_lru_page(struct page *page)
609{
610 int lru;
611 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700612 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700613
614 VM_BUG_ON(PageLRU(page));
615
616redo:
617 ClearPageUnevictable(page);
618
619 if (page_evictable(page, NULL)) {
620 /*
621 * For evictable pages, we can use the cache.
622 * In event of a race, worst case is we end up with an
623 * unevictable page on [in]active list.
624 * We know how to handle that.
625 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700626 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700627 lru_cache_add_lru(page, lru);
628 } else {
629 /*
630 * Put unevictable pages directly on zone's unevictable
631 * list.
632 */
633 lru = LRU_UNEVICTABLE;
634 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700635 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700636 * When racing with an mlock or AS_UNEVICTABLE clearing
637 * (page is unlocked) make sure that if the other thread
638 * does not observe our setting of PG_lru and fails
639 * isolation/check_move_unevictable_page,
640 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700641 * the page back to the evictable list.
642 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700643 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700644 */
645 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700646 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700647
648 /*
649 * page's status can change while we move it among lru. If an evictable
650 * page is on unevictable list, it never be freed. To avoid that,
651 * check after we added it to the list, again.
652 */
653 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654 if (!isolate_lru_page(page)) {
655 put_page(page);
656 goto redo;
657 }
658 /* This means someone else dropped this page from LRU
659 * So, it will be freed or putback to LRU again. There is
660 * nothing to do here.
661 */
662 }
663
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700664 if (was_unevictable && lru != LRU_UNEVICTABLE)
665 count_vm_event(UNEVICTABLE_PGRESCUED);
666 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667 count_vm_event(UNEVICTABLE_PGCULLED);
668
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700669 put_page(page); /* drop ref from isolate */
670}
671
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800672enum page_references {
673 PAGEREF_RECLAIM,
674 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800675 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800676 PAGEREF_ACTIVATE,
677};
678
679static enum page_references page_check_references(struct page *page,
680 struct scan_control *sc)
681{
Johannes Weiner64574742010-03-05 13:42:22 -0800682 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800683 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800684
Johannes Weiner64574742010-03-05 13:42:22 -0800685 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800687
688 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800689 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800690 return PAGEREF_RECLAIM;
691
692 /*
693 * Mlock lost the isolation race with us. Let try_to_unmap()
694 * move the page to the unevictable list.
695 */
696 if (vm_flags & VM_LOCKED)
697 return PAGEREF_RECLAIM;
698
Johannes Weiner64574742010-03-05 13:42:22 -0800699 if (referenced_ptes) {
700 if (PageAnon(page))
701 return PAGEREF_ACTIVATE;
702 /*
703 * All mapped pages start out with page table
704 * references from the instantiating fault, so we need
705 * to look twice if a mapped file page is used more
706 * than once.
707 *
708 * Mark it and spare it for another trip around the
709 * inactive list. Another page table reference will
710 * lead to its activation.
711 *
712 * Note: the mark is set for activated pages as well
713 * so that recently deactivated but used pages are
714 * quickly recovered.
715 */
716 SetPageReferenced(page);
717
718 if (referenced_page)
719 return PAGEREF_ACTIVATE;
720
721 return PAGEREF_KEEP;
722 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800723
724 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700725 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner64574742010-03-05 13:42:22 -0800726 return PAGEREF_RECLAIM_CLEAN;
727
728 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800729}
730
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700731static noinline_for_stack void free_page_list(struct list_head *free_pages)
732{
733 struct pagevec freed_pvec;
734 struct page *page, *tmp;
735
736 pagevec_init(&freed_pvec, 1);
737
738 list_for_each_entry_safe(page, tmp, free_pages, lru) {
739 list_del(&page->lru);
740 if (!pagevec_add(&freed_pvec, page)) {
741 __pagevec_free(&freed_pvec);
742 pagevec_reinit(&freed_pvec);
743 }
744 }
745
746 pagevec_free(&freed_pvec);
747}
748
Nick Piggine2867812008-07-25 19:45:30 -0700749/*
Andrew Morton1742f192006-03-22 00:08:21 -0800750 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 */
Andrew Morton1742f192006-03-22 00:08:21 -0800752static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700753 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700754 struct scan_control *sc,
Mel Gorman92df3a72011-10-31 17:07:56 -0700755 int priority,
756 unsigned long *ret_nr_dirty,
757 unsigned long *ret_nr_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758{
759 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700760 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700762 unsigned long nr_dirty = 0;
763 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800764 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700765 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
767 cond_resched();
768
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800770 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 struct address_space *mapping;
772 struct page *page;
773 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 cond_resched();
776
777 page = lru_to_page(page_list);
778 list_del(&page->lru);
779
Nick Piggin529ae9a2008-08-02 12:01:03 +0200780 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 goto keep;
782
Nick Piggin725d7042006-09-25 23:30:55 -0700783 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700784 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
786 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800787
Nick Pigginb291f002008-10-18 20:26:44 -0700788 if (unlikely(!page_evictable(page, NULL)))
789 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700790
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700791 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800792 goto keep_locked;
793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 /* Double the slab pressure for mapped and swapcache pages */
795 if (page_mapped(page) || PageSwapCache(page))
796 sc->nr_scanned++;
797
Andy Whitcroftc661b072007-08-22 14:01:26 -0700798 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
799 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
800
801 if (PageWriteback(page)) {
Mel Gorman92df3a72011-10-31 17:07:56 -0700802 nr_writeback++;
Andy Whitcroftc661b072007-08-22 14:01:26 -0700803 /*
Mel Gormana18bba02011-10-31 17:07:42 -0700804 * Synchronous reclaim cannot queue pages for
805 * writeback due to the possibility of stack overflow
806 * but if it encounters a page under writeback, wait
807 * for the IO to complete.
Andy Whitcroftc661b072007-08-22 14:01:26 -0700808 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800809 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700810 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700811 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700812 else {
813 unlock_page(page);
814 goto keep_lumpy;
815 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700816 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800818 references = page_check_references(page, sc);
819 switch (references) {
820 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800822 case PAGEREF_KEEP:
823 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800824 case PAGEREF_RECLAIM:
825 case PAGEREF_RECLAIM_CLEAN:
826 ; /* try to reclaim the page below */
827 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 /*
830 * Anonymous process memory has backing store?
831 * Try to allocate it some swap space here.
832 */
Nick Pigginb291f002008-10-18 20:26:44 -0700833 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800834 if (!(sc->gfp_mask & __GFP_IO))
835 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800836 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800838 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700839 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840
841 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
843 /*
844 * The page is mapped into the page tables of one or more
845 * processes. Try to unmap it here.
846 */
847 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200848 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 case SWAP_FAIL:
850 goto activate_locked;
851 case SWAP_AGAIN:
852 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700853 case SWAP_MLOCK:
854 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 case SWAP_SUCCESS:
856 ; /* try to free the page below */
857 }
858 }
859
860 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700861 nr_dirty++;
862
Mel Gormanee728862011-10-31 17:07:38 -0700863 /*
864 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700865 * avoid risk of stack overflow but do not writeback
866 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700867 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700868 if (page_is_file_cache(page) &&
869 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700870 /*
871 * Immediately reclaim when written back.
872 * Similar in principal to deactivate_page()
873 * except we already have the page isolated
874 * and know it's dirty
875 */
876 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
877 SetPageReclaim(page);
878
Mel Gormanee728862011-10-31 17:07:38 -0700879 goto keep_locked;
880 }
881
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800882 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700884 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800886 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887 goto keep_locked;
888
889 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700890 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700892 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 goto keep_locked;
894 case PAGE_ACTIVATE:
895 goto activate_locked;
896 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700897 if (PageWriteback(page))
898 goto keep_lumpy;
899 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700901
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 /*
903 * A synchronous write - probably a ramdisk. Go
904 * ahead and try to reclaim the page.
905 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200906 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 goto keep;
908 if (PageDirty(page) || PageWriteback(page))
909 goto keep_locked;
910 mapping = page_mapping(page);
911 case PAGE_CLEAN:
912 ; /* try to free the page below */
913 }
914 }
915
916 /*
917 * If the page has buffers, try to free the buffer mappings
918 * associated with this page. If we succeed we try to free
919 * the page as well.
920 *
921 * We do this even if the page is PageDirty().
922 * try_to_release_page() does not perform I/O, but it is
923 * possible for a page to have PageDirty set, but it is actually
924 * clean (all its buffers are clean). This happens if the
925 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700926 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 * try_to_release_page() will discover that cleanness and will
928 * drop the buffers and mark the page clean - it can be freed.
929 *
930 * Rarely, pages can have buffers and no ->mapping. These are
931 * the pages which were not successfully invalidated in
932 * truncate_complete_page(). We try to drop those buffers here
933 * and if that worked, and the page is no longer mapped into
934 * process address space (page_count == 1) it can be freed.
935 * Otherwise, leave the page on the LRU so it is swappable.
936 */
David Howells266cf652009-04-03 16:42:36 +0100937 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 if (!try_to_release_page(page, sc->gfp_mask))
939 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700940 if (!mapping && page_count(page) == 1) {
941 unlock_page(page);
942 if (put_page_testzero(page))
943 goto free_it;
944 else {
945 /*
946 * rare race with speculative reference.
947 * the speculative reference will free
948 * this page shortly, so we may
949 * increment nr_reclaimed here (and
950 * leave it off the LRU).
951 */
952 nr_reclaimed++;
953 continue;
954 }
955 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 }
957
Nick Piggine2867812008-07-25 19:45:30 -0700958 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800959 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
Nick Piggina978d6f2008-10-18 20:26:58 -0700961 /*
962 * At this point, we have no other references and there is
963 * no way to pick any more up (removed from LRU, removed
964 * from pagecache). Can use non-atomic bitops now (and
965 * we obviously don't have to worry about waking up a process
966 * waiting on the page lock, because there are no references.
967 */
968 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700969free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800970 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700971
972 /*
973 * Is there need to periodically free_page_list? It would
974 * appear not as the counts should be low
975 */
976 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 continue;
978
Nick Pigginb291f002008-10-18 20:26:44 -0700979cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800980 if (PageSwapCache(page))
981 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700982 unlock_page(page);
983 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800984 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700985 continue;
986
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700988 /* Not a candidate for swapping, so reclaim swap space. */
989 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800990 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700991 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 SetPageActive(page);
993 pgactivate++;
994keep_locked:
995 unlock_page(page);
996keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -0800997 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700998keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001000 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001002
Mel Gorman0e093d992010-10-26 14:21:45 -07001003 /*
1004 * Tag a zone as congested if all the dirty pages encountered were
1005 * backed by a congested BDI. In this case, reclaimers should just
1006 * back off and wait for congestion to clear because further reclaim
1007 * will encounter the same problem
1008 */
KAMEZAWA Hiroyukid6c438b2011-05-17 15:44:10 -07001009 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
Mel Gorman0e093d992010-10-26 14:21:45 -07001010 zone_set_flag(zone, ZONE_CONGESTED);
1011
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001012 free_page_list(&free_pages);
1013
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001015 count_vm_events(PGACTIVATE, pgactivate);
Mel Gorman92df3a72011-10-31 17:07:56 -07001016 *ret_nr_dirty += nr_dirty;
1017 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -08001018 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019}
1020
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001021/*
1022 * Attempt to remove the specified page from its LRU. Only take this page
1023 * if it is of the appropriate PageActive status. Pages which are being
1024 * freed elsewhere are also ignored.
1025 *
1026 * page: page to consider
1027 * mode: one of the LRU isolation modes defined above
1028 *
1029 * returns 0 on success, -ve errno on failure.
1030 */
Minchan Kim4356f212011-10-31 17:06:47 -07001031int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001032{
Minchan Kim4356f212011-10-31 17:06:47 -07001033 bool all_lru_mode;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001034 int ret = -EINVAL;
1035
1036 /* Only take pages on the LRU. */
1037 if (!PageLRU(page))
1038 return ret;
1039
Minchan Kim4356f212011-10-31 17:06:47 -07001040 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1041 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1042
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001043 /*
1044 * When checking the active state, we need to be sure we are
1045 * dealing with comparible boolean values. Take the logical not
1046 * of each.
1047 */
Minchan Kim4356f212011-10-31 17:06:47 -07001048 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001049 return ret;
1050
Minchan Kim4356f212011-10-31 17:06:47 -07001051 if (!all_lru_mode && !!page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001052 return ret;
1053
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001054 /*
1055 * When this function is being called for lumpy reclaim, we
1056 * initially look into all LRU pages, active, inactive and
1057 * unevictable; only give shrink_page_list evictable pages.
1058 */
1059 if (PageUnevictable(page))
1060 return ret;
1061
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001062 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001063
Minchan Kim39deaf82011-10-31 17:06:51 -07001064 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1065 return ret;
1066
Minchan Kimf80c0672011-10-31 17:06:55 -07001067 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1068 return ret;
1069
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001070 if (likely(get_page_unless_zero(page))) {
1071 /*
1072 * Be careful not to clear PageLRU until after we're
1073 * sure the page is not being freed elsewhere -- the
1074 * page release code relies on it.
1075 */
1076 ClearPageLRU(page);
1077 ret = 0;
1078 }
1079
1080 return ret;
1081}
1082
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001083/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 * zone->lru_lock is heavily contended. Some of the functions that
1085 * shrink the lists perform better by taking out a batch of pages
1086 * and working on them outside the LRU lock.
1087 *
1088 * For pagecache intensive workloads, this function is the hottest
1089 * spot in the kernel (apart from copy_*_user functions).
1090 *
1091 * Appropriate locks must be held before calling this function.
1092 *
1093 * @nr_to_scan: The number of pages to look through on the list.
1094 * @src: The LRU list to pull pages off.
1095 * @dst: The temp list to put pages on to.
1096 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001097 * @order: The caller's attempted allocation order
1098 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001099 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 *
1101 * returns how many pages were moved onto *@dst.
1102 */
Andrew Morton69e05942006-03-22 00:08:19 -08001103static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1104 struct list_head *src, struct list_head *dst,
Minchan Kim4356f212011-10-31 17:06:47 -07001105 unsigned long *scanned, int order, isolate_mode_t mode,
1106 int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107{
Andrew Morton69e05942006-03-22 00:08:19 -08001108 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001109 unsigned long nr_lumpy_taken = 0;
1110 unsigned long nr_lumpy_dirty = 0;
1111 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001112 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001114 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001115 struct page *page;
1116 unsigned long pfn;
1117 unsigned long end_pfn;
1118 unsigned long page_pfn;
1119 int zone_id;
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 page = lru_to_page(src);
1122 prefetchw_prev_lru_page(page, src, flags);
1123
Nick Piggin725d7042006-09-25 23:30:55 -07001124 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001125
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001126 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001127 case 0:
1128 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001129 mem_cgroup_del_lru(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001130 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001131 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001132
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001133 case -EBUSY:
1134 /* else it is being freed elsewhere */
1135 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001136 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001137 continue;
1138
1139 default:
1140 BUG();
1141 }
1142
1143 if (!order)
1144 continue;
1145
1146 /*
1147 * Attempt to take all pages in the order aligned region
1148 * surrounding the tag page. Only take those pages of
1149 * the same active state as that tag page. We may safely
1150 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001151 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001152 * where that page is in a different zone we will detect
1153 * it from its zone id and abort this block scan.
1154 */
1155 zone_id = page_zone_id(page);
1156 page_pfn = page_to_pfn(page);
1157 pfn = page_pfn & ~((1 << order) - 1);
1158 end_pfn = pfn + (1 << order);
1159 for (; pfn < end_pfn; pfn++) {
1160 struct page *cursor_page;
1161
1162 /* The target page is in the block, ignore it. */
1163 if (unlikely(pfn == page_pfn))
1164 continue;
1165
1166 /* Avoid holes within the zone. */
1167 if (unlikely(!pfn_valid_within(pfn)))
1168 break;
1169
1170 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001171
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001172 /* Check that we have not crossed a zone boundary. */
1173 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001174 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001175
1176 /*
1177 * If we don't have enough swap space, reclaiming of
1178 * anon page which don't already have a swap slot is
1179 * pointless.
1180 */
1181 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001182 !PageSwapCache(cursor_page))
1183 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001184
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001185 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001186 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001187 mem_cgroup_del_lru(cursor_page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001188 nr_taken += hpage_nr_pages(page);
Mel Gormana8a94d12010-08-09 17:19:17 -07001189 nr_lumpy_taken++;
1190 if (PageDirty(cursor_page))
1191 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001192 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001193 } else {
Andrea Arcangelid179e842011-06-15 15:08:51 -07001194 /*
1195 * Check if the page is freed already.
1196 *
1197 * We can't use page_count() as that
1198 * requires compound_head and we don't
1199 * have a pin on the page here. If a
1200 * page is tail, we may or may not
1201 * have isolated the head, so assume
1202 * it's not free, it'd be tricky to
1203 * track the head status without a
1204 * page pin.
1205 */
1206 if (!PageTail(cursor_page) &&
1207 !atomic_read(&cursor_page->_count))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001208 continue;
1209 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001210 }
1211 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001212
1213 /* If we break out of the loop above, lumpy reclaim failed */
1214 if (pfn < end_pfn)
1215 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 }
1217
1218 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001219
1220 trace_mm_vmscan_lru_isolate(order,
1221 nr_to_scan, scan,
1222 nr_taken,
1223 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1224 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 return nr_taken;
1226}
1227
Balbir Singh66e17072008-02-07 00:13:56 -08001228static unsigned long isolate_pages_global(unsigned long nr,
1229 struct list_head *dst,
1230 unsigned long *scanned, int order,
Minchan Kim4356f212011-10-31 17:06:47 -07001231 isolate_mode_t mode,
1232 struct zone *z, int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001233{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001234 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001235 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001236 lru += LRU_ACTIVE;
1237 if (file)
1238 lru += LRU_FILE;
1239 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001240 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001241}
1242
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001244 * clear_active_flags() is a helper for shrink_active_list(), clearing
1245 * any active bits from the pages in the list.
1246 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001247static unsigned long clear_active_flags(struct list_head *page_list,
1248 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001249{
1250 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001251 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001252 struct page *page;
1253
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001254 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001255 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001256 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001257 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001258 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001259 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001260 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001261 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001262 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001263 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001264 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001265
1266 return nr_active;
1267}
1268
Nick Piggin62695a82008-10-18 20:26:09 -07001269/**
1270 * isolate_lru_page - tries to isolate a page from its LRU list
1271 * @page: page to isolate from its LRU list
1272 *
1273 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1274 * vmstat statistic corresponding to whatever LRU list the page was on.
1275 *
1276 * Returns 0 if the page was removed from an LRU list.
1277 * Returns -EBUSY if the page was not on an LRU list.
1278 *
1279 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001280 * the active list, it will have PageActive set. If it was found on
1281 * the unevictable list, it will have the PageUnevictable bit set. That flag
1282 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001283 *
1284 * The vmstat statistic corresponding to the list on which the page was
1285 * found will be decremented.
1286 *
1287 * Restrictions:
1288 * (1) Must be called with an elevated refcount on the page. This is a
1289 * fundamentnal difference from isolate_lru_pages (which is called
1290 * without a stable reference).
1291 * (2) the lru_lock must not be held.
1292 * (3) interrupts must be enabled.
1293 */
1294int isolate_lru_page(struct page *page)
1295{
1296 int ret = -EBUSY;
1297
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001298 VM_BUG_ON(!page_count(page));
1299
Nick Piggin62695a82008-10-18 20:26:09 -07001300 if (PageLRU(page)) {
1301 struct zone *zone = page_zone(page);
1302
1303 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001304 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001305 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001306 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001307 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001308 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001309
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001310 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001311 }
1312 spin_unlock_irq(&zone->lru_lock);
1313 }
1314 return ret;
1315}
1316
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001317/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001318 * Are there way too many processes in the direct reclaim path already?
1319 */
1320static int too_many_isolated(struct zone *zone, int file,
1321 struct scan_control *sc)
1322{
1323 unsigned long inactive, isolated;
1324
1325 if (current_is_kswapd())
1326 return 0;
1327
1328 if (!scanning_global_lru(sc))
1329 return 0;
1330
1331 if (file) {
1332 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1333 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1334 } else {
1335 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1336 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1337 }
1338
1339 return isolated > inactive;
1340}
1341
1342/*
Mel Gorman66635622010-08-09 17:19:30 -07001343 * TODO: Try merging with migrations version of putback_lru_pages
1344 */
1345static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001346putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001347 unsigned long nr_anon, unsigned long nr_file,
1348 struct list_head *page_list)
1349{
1350 struct page *page;
1351 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001352 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001353
1354 pagevec_init(&pvec, 1);
1355
1356 /*
1357 * Put back any unfreeable pages.
1358 */
1359 spin_lock(&zone->lru_lock);
1360 while (!list_empty(page_list)) {
1361 int lru;
1362 page = lru_to_page(page_list);
1363 VM_BUG_ON(PageLRU(page));
1364 list_del(&page->lru);
1365 if (unlikely(!page_evictable(page, NULL))) {
1366 spin_unlock_irq(&zone->lru_lock);
1367 putback_lru_page(page);
1368 spin_lock_irq(&zone->lru_lock);
1369 continue;
1370 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001371 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001372 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001373 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001374 if (is_active_lru(lru)) {
1375 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001376 int numpages = hpage_nr_pages(page);
1377 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001378 }
1379 if (!pagevec_add(&pvec, page)) {
1380 spin_unlock_irq(&zone->lru_lock);
1381 __pagevec_release(&pvec);
1382 spin_lock_irq(&zone->lru_lock);
1383 }
1384 }
1385 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1386 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1387
1388 spin_unlock_irq(&zone->lru_lock);
1389 pagevec_release(&pvec);
1390}
1391
Mel Gorman1489fa12010-08-09 17:19:33 -07001392static noinline_for_stack void update_isolated_counts(struct zone *zone,
1393 struct scan_control *sc,
1394 unsigned long *nr_anon,
1395 unsigned long *nr_file,
1396 struct list_head *isolated_list)
1397{
1398 unsigned long nr_active;
1399 unsigned int count[NR_LRU_LISTS] = { 0, };
1400 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1401
1402 nr_active = clear_active_flags(isolated_list, count);
1403 __count_vm_events(PGDEACTIVATE, nr_active);
1404
1405 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1406 -count[LRU_ACTIVE_FILE]);
1407 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1408 -count[LRU_INACTIVE_FILE]);
1409 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1410 -count[LRU_ACTIVE_ANON]);
1411 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1412 -count[LRU_INACTIVE_ANON]);
1413
1414 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1415 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1416 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1417 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1418
1419 reclaim_stat->recent_scanned[0] += *nr_anon;
1420 reclaim_stat->recent_scanned[1] += *nr_file;
1421}
1422
Mel Gorman66635622010-08-09 17:19:30 -07001423/*
Mel Gormana18bba02011-10-31 17:07:42 -07001424 * Returns true if a direct reclaim should wait on pages under writeback.
Wu Fengguange31f36982010-08-09 17:20:01 -07001425 *
1426 * If we are direct reclaiming for contiguous pages and we do not reclaim
1427 * everything in the list, try again and wait for writeback IO to complete.
1428 * This will stall high-order allocations noticeably. Only do that when really
1429 * need to free the pages under high memory pressure.
1430 */
1431static inline bool should_reclaim_stall(unsigned long nr_taken,
1432 unsigned long nr_freed,
1433 int priority,
1434 struct scan_control *sc)
1435{
1436 int lumpy_stall_priority;
1437
1438 /* kswapd should not stall on sync IO */
1439 if (current_is_kswapd())
1440 return false;
1441
1442 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001443 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001444 return false;
1445
Justin P. Mattock81d66c72011-08-23 09:28:02 -07001446 /* If we have reclaimed everything on the isolated list, no stall */
Wu Fengguange31f36982010-08-09 17:20:01 -07001447 if (nr_freed == nr_taken)
1448 return false;
1449
1450 /*
1451 * For high-order allocations, there are two stall thresholds.
1452 * High-cost allocations stall immediately where as lower
1453 * order allocations such as stacks require the scanning
1454 * priority to be much higher before stalling.
1455 */
1456 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1457 lumpy_stall_priority = DEF_PRIORITY;
1458 else
1459 lumpy_stall_priority = DEF_PRIORITY / 3;
1460
1461 return priority <= lumpy_stall_priority;
1462}
1463
1464/*
Andrew Morton1742f192006-03-22 00:08:21 -08001465 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1466 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 */
Mel Gorman66635622010-08-09 17:19:30 -07001468static noinline_for_stack unsigned long
1469shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1470 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471{
1472 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001473 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001474 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001475 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001476 unsigned long nr_anon;
1477 unsigned long nr_file;
Mel Gorman92df3a72011-10-31 17:07:56 -07001478 unsigned long nr_dirty = 0;
1479 unsigned long nr_writeback = 0;
Minchan Kim4356f212011-10-31 17:06:47 -07001480 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001481
Rik van Riel35cd7812009-09-21 17:01:38 -07001482 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001483 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001484
1485 /* We are about to die and free our memory. Return now. */
1486 if (fatal_signal_pending(current))
1487 return SWAP_CLUSTER_MAX;
1488 }
1489
Mel Gormanf3a310b2011-01-13 15:46:00 -08001490 set_reclaim_mode(priority, sc, false);
Minchan Kim4356f212011-10-31 17:06:47 -07001491 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1492 reclaim_mode |= ISOLATE_ACTIVE;
1493
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001495
1496 if (!sc->may_unmap)
1497 reclaim_mode |= ISOLATE_UNMAPPED;
1498 if (!sc->may_writepage)
1499 reclaim_mode |= ISOLATE_CLEAN;
1500
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001503 if (scanning_global_lru(sc)) {
Minchan Kim4356f212011-10-31 17:06:47 -07001504 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1505 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001506 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001507 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001508 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1509 nr_scanned);
1510 else
1511 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1512 nr_scanned);
1513 } else {
Minchan Kim4356f212011-10-31 17:06:47 -07001514 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1515 &nr_scanned, sc->order, reclaim_mode, zone,
1516 sc->mem_cgroup, 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001518 * mem_cgroup_isolate_pages() keeps track of
1519 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001521 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001522
Mel Gorman66635622010-08-09 17:19:30 -07001523 if (nr_taken == 0) {
1524 spin_unlock_irq(&zone->lru_lock);
1525 return 0;
1526 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001527
Mel Gorman1489fa12010-08-09 17:19:33 -07001528 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001529
1530 spin_unlock_irq(&zone->lru_lock);
1531
Mel Gorman92df3a72011-10-31 17:07:56 -07001532 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1533 &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001534
Wu Fengguange31f36982010-08-09 17:20:01 -07001535 /* Check if we should syncronously wait for writeback */
1536 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001537 set_reclaim_mode(priority, sc, true);
Mel Gorman92df3a72011-10-31 17:07:56 -07001538 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1539 priority, &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001540 }
1541
1542 local_irq_disable();
1543 if (current_is_kswapd())
1544 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1545 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1546
Mel Gorman1489fa12010-08-09 17:19:33 -07001547 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001548
Mel Gorman92df3a72011-10-31 17:07:56 -07001549 /*
1550 * If reclaim is isolating dirty pages under writeback, it implies
1551 * that the long-lived page allocation rate is exceeding the page
1552 * laundering rate. Either the global limits are not being effective
1553 * at throttling processes due to the page distribution throughout
1554 * zones or there is heavy usage of a slow backing device. The
1555 * only option is to throttle from reclaim context which is not ideal
1556 * as there is no guarantee the dirtying process is throttled in the
1557 * same way balance_dirty_pages() manages.
1558 *
1559 * This scales the number of dirty pages that must be under writeback
1560 * before throttling depending on priority. It is a simple backoff
1561 * function that has the most effect in the range DEF_PRIORITY to
1562 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1563 * in trouble and reclaim is considered to be in trouble.
1564 *
1565 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1566 * DEF_PRIORITY-1 50% must be PageWriteback
1567 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1568 * ...
1569 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1570 * isolated page is PageWriteback
1571 */
1572 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1573 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1574
Mel Gormane11da5b2010-10-26 14:21:40 -07001575 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1576 zone_idx(zone),
1577 nr_scanned, nr_reclaimed,
1578 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001579 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001580 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581}
1582
Martin Bligh3bb1a852006-10-28 10:38:24 -07001583/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 * This moves pages from the active list to the inactive list.
1585 *
1586 * We move them the other way if the page is referenced by one or more
1587 * processes, from rmap.
1588 *
1589 * If the pages are mostly unmapped, the processing is fast and it is
1590 * appropriate to hold zone->lru_lock across the whole operation. But if
1591 * the pages are mapped, the processing is slow (page_referenced()) so we
1592 * should drop zone->lru_lock around each page. It's impossible to balance
1593 * this, so instead we remove the pages from the LRU while processing them.
1594 * It is safe to rely on PG_active against the non-LRU pages in here because
1595 * nobody will play with that bit on a non-LRU page.
1596 *
1597 * The downside is that we have to touch page->_count against each page.
1598 * But we had to alter page->flags anyway.
1599 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001600
Wu Fengguang3eb41402009-06-16 15:33:13 -07001601static void move_active_pages_to_lru(struct zone *zone,
1602 struct list_head *list,
1603 enum lru_list lru)
1604{
1605 unsigned long pgmoved = 0;
1606 struct pagevec pvec;
1607 struct page *page;
1608
1609 pagevec_init(&pvec, 1);
1610
1611 while (!list_empty(list)) {
1612 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001613
1614 VM_BUG_ON(PageLRU(page));
1615 SetPageLRU(page);
1616
Wu Fengguang3eb41402009-06-16 15:33:13 -07001617 list_move(&page->lru, &zone->lru[lru].list);
1618 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001619 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001620
1621 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1622 spin_unlock_irq(&zone->lru_lock);
1623 if (buffer_heads_over_limit)
1624 pagevec_strip(&pvec);
1625 __pagevec_release(&pvec);
1626 spin_lock_irq(&zone->lru_lock);
1627 }
1628 }
1629 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1630 if (!is_active_lru(lru))
1631 __count_vm_events(PGDEACTIVATE, pgmoved);
1632}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001633
Andrew Morton1742f192006-03-22 00:08:21 -08001634static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001635 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001637 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001638 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001639 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001641 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001642 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001644 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001645 unsigned long nr_rotated = 0;
Minchan Kimf80c0672011-10-31 17:06:55 -07001646 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
1648 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001649
1650 if (!sc->may_unmap)
1651 reclaim_mode |= ISOLATE_UNMAPPED;
1652 if (!sc->may_writepage)
1653 reclaim_mode |= ISOLATE_CLEAN;
1654
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001656 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001657 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1658 &pgscanned, sc->order,
Minchan Kimf80c0672011-10-31 17:06:55 -07001659 reclaim_mode, zone,
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001660 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001661 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001662 } else {
1663 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1664 &pgscanned, sc->order,
Minchan Kimf80c0672011-10-31 17:06:55 -07001665 reclaim_mode, zone,
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001666 sc->mem_cgroup, 1, file);
1667 /*
1668 * mem_cgroup_isolate_pages() keeps track of
1669 * scanned pages on its own.
1670 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001671 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001672
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001673 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001674
Wu Fengguang3eb41402009-06-16 15:33:13 -07001675 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001676 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001677 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001678 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001679 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001680 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 spin_unlock_irq(&zone->lru_lock);
1682
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 while (!list_empty(&l_hold)) {
1684 cond_resched();
1685 page = lru_to_page(&l_hold);
1686 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001687
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001688 if (unlikely(!page_evictable(page, NULL))) {
1689 putback_lru_page(page);
1690 continue;
1691 }
1692
Johannes Weiner64574742010-03-05 13:42:22 -08001693 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001694 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001695 /*
1696 * Identify referenced, file-backed active pages and
1697 * give them one more trip around the active list. So
1698 * that executable code get better chances to stay in
1699 * memory under moderate memory pressure. Anon pages
1700 * are not likely to be evicted by use-once streaming
1701 * IO, plus JVM can create lots of anon VM_EXEC pages,
1702 * so we ignore them here.
1703 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001704 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001705 list_add(&page->lru, &l_active);
1706 continue;
1707 }
1708 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001709
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001710 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 list_add(&page->lru, &l_inactive);
1712 }
1713
Andrew Mortonb5557492009-01-06 14:40:13 -08001714 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001715 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001716 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001717 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001718 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001719 * Count referenced pages from currently used mappings as rotated,
1720 * even though only some of them are actually re-activated. This
1721 * helps balance scan pressure between file and anonymous pages in
1722 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001723 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001724 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001725
Wu Fengguang3eb41402009-06-16 15:33:13 -07001726 move_active_pages_to_lru(zone, &l_active,
1727 LRU_ACTIVE + file * LRU_FILE);
1728 move_active_pages_to_lru(zone, &l_inactive,
1729 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001730 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001731 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732}
1733
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001734#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001735static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001736{
1737 unsigned long active, inactive;
1738
1739 active = zone_page_state(zone, NR_ACTIVE_ANON);
1740 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1741
1742 if (inactive * zone->inactive_ratio < active)
1743 return 1;
1744
1745 return 0;
1746}
1747
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001748/**
1749 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1750 * @zone: zone to check
1751 * @sc: scan control of this context
1752 *
1753 * Returns true if the zone does not have enough inactive anon pages,
1754 * meaning some active anon pages need to be deactivated.
1755 */
1756static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1757{
1758 int low;
1759
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001760 /*
1761 * If we don't have swap space, anonymous page deactivation
1762 * is pointless.
1763 */
1764 if (!total_swap_pages)
1765 return 0;
1766
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001767 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001768 low = inactive_anon_is_low_global(zone);
1769 else
Johannes Weiner9b272972011-11-02 13:38:23 -07001770 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001771 return low;
1772}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001773#else
1774static inline int inactive_anon_is_low(struct zone *zone,
1775 struct scan_control *sc)
1776{
1777 return 0;
1778}
1779#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001780
Rik van Riel56e49d22009-06-16 15:32:28 -07001781static int inactive_file_is_low_global(struct zone *zone)
1782{
1783 unsigned long active, inactive;
1784
1785 active = zone_page_state(zone, NR_ACTIVE_FILE);
1786 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1787
1788 return (active > inactive);
1789}
1790
1791/**
1792 * inactive_file_is_low - check if file pages need to be deactivated
1793 * @zone: zone to check
1794 * @sc: scan control of this context
1795 *
1796 * When the system is doing streaming IO, memory pressure here
1797 * ensures that active file pages get deactivated, until more
1798 * than half of the file pages are on the inactive list.
1799 *
1800 * Once we get to that situation, protect the system's working
1801 * set from being evicted by disabling active file page aging.
1802 *
1803 * This uses a different ratio than the anonymous pages, because
1804 * the page cache uses a use-once replacement algorithm.
1805 */
1806static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1807{
1808 int low;
1809
1810 if (scanning_global_lru(sc))
1811 low = inactive_file_is_low_global(zone);
1812 else
Johannes Weiner9b272972011-11-02 13:38:23 -07001813 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
Rik van Riel56e49d22009-06-16 15:32:28 -07001814 return low;
1815}
1816
Rik van Rielb39415b2009-12-14 17:59:48 -08001817static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1818 int file)
1819{
1820 if (file)
1821 return inactive_file_is_low(zone, sc);
1822 else
1823 return inactive_anon_is_low(zone, sc);
1824}
1825
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001826static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001827 struct zone *zone, struct scan_control *sc, int priority)
1828{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001829 int file = is_file_lru(lru);
1830
Rik van Rielb39415b2009-12-14 17:59:48 -08001831 if (is_active_lru(lru)) {
1832 if (inactive_list_is_low(zone, sc, file))
1833 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001834 return 0;
1835 }
1836
Rik van Riel33c120e2008-10-18 20:26:36 -07001837 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001838}
1839
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001840static int vmscan_swappiness(struct scan_control *sc)
1841{
1842 if (scanning_global_lru(sc))
1843 return vm_swappiness;
1844 return mem_cgroup_swappiness(sc->mem_cgroup);
1845}
1846
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001848 * Determine how aggressively the anon and file LRU lists should be
1849 * scanned. The relative value of each set of LRU lists is determined
1850 * by looking at the fraction of the pages scanned we did rotate back
1851 * onto the active list instead of evict.
1852 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001853 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001854 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001855static void get_scan_count(struct zone *zone, struct scan_control *sc,
1856 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001857{
1858 unsigned long anon, file, free;
1859 unsigned long anon_prio, file_prio;
1860 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001861 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001862 u64 fraction[2], denominator;
1863 enum lru_list l;
1864 int noswap = 0;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001865 bool force_scan = false;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001866
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001867 /*
1868 * If the zone or memcg is small, nr[l] can be 0. This
1869 * results in no scanning on this priority and a potential
1870 * priority drop. Global direct reclaim can go to the next
1871 * zone and tends to have no problems. Global kswapd is for
1872 * zone balancing and it needs to scan a minimum amount. When
1873 * reclaiming for a memcg, a priority drop can cause high
1874 * latencies, so it's better to scan a minimum amount there as
1875 * well.
1876 */
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001877 if (scanning_global_lru(sc) && current_is_kswapd())
1878 force_scan = true;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001879 if (!scanning_global_lru(sc))
1880 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001881
1882 /* If we have no swap space, do not bother scanning anon pages. */
1883 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1884 noswap = 1;
1885 fraction[0] = 0;
1886 fraction[1] = 1;
1887 denominator = 1;
1888 goto out;
1889 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001890
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001891 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1892 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1893 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1894 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1895
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001896 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001897 free = zone_page_state(zone, NR_FREE_PAGES);
1898 /* If we have very few page cache pages,
1899 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001900 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001901 fraction[0] = 1;
1902 fraction[1] = 0;
1903 denominator = 1;
1904 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001905 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001906 }
1907
1908 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001909 * With swappiness at 100, anonymous and file have the same priority.
1910 * This scanning priority is essentially the inverse of IO cost.
1911 */
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001912 anon_prio = vmscan_swappiness(sc);
1913 file_prio = 200 - vmscan_swappiness(sc);
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001914
1915 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001916 * OK, so we have swap space and a fair amount of page cache
1917 * pages. We use the recently rotated / recently scanned
1918 * ratios to determine how valuable each cache is.
1919 *
1920 * Because workloads change over time (and to avoid overflow)
1921 * we keep these statistics as a floating average, which ends
1922 * up weighing recent references more than old ones.
1923 *
1924 * anon in [0], file in [1]
1925 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001926 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001927 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001928 reclaim_stat->recent_scanned[0] /= 2;
1929 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001930 }
1931
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001932 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001933 reclaim_stat->recent_scanned[1] /= 2;
1934 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001935 }
1936
1937 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001938 * The amount of pressure on anon vs file pages is inversely
1939 * proportional to the fraction of recently scanned pages on
1940 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001941 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001942 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1943 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001944
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001945 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1946 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001947 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001948
Shaohua Li76a33fc2010-05-24 14:32:36 -07001949 fraction[0] = ap;
1950 fraction[1] = fp;
1951 denominator = ap + fp + 1;
1952out:
1953 for_each_evictable_lru(l) {
1954 int file = is_file_lru(l);
1955 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001956
Shaohua Li76a33fc2010-05-24 14:32:36 -07001957 scan = zone_nr_lru_pages(zone, sc, l);
1958 if (priority || noswap) {
1959 scan >>= priority;
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001960 if (!scan && force_scan)
1961 scan = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001962 scan = div64_u64(scan * fraction[file], denominator);
1963 }
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001964 nr[l] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001965 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001966}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001967
1968/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001969 * Reclaim/compaction depends on a number of pages being freed. To avoid
1970 * disruption to the system, a small number of order-0 pages continue to be
1971 * rotated and reclaimed in the normal fashion. However, by the time we get
1972 * back to the allocator and call try_to_compact_zone(), we ensure that
1973 * there are enough free pages for it to be likely successful
1974 */
1975static inline bool should_continue_reclaim(struct zone *zone,
1976 unsigned long nr_reclaimed,
1977 unsigned long nr_scanned,
1978 struct scan_control *sc)
1979{
1980 unsigned long pages_for_compaction;
1981 unsigned long inactive_lru_pages;
1982
1983 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001984 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001985 return false;
1986
Mel Gorman28765922011-02-25 14:44:20 -08001987 /* Consider stopping depending on scan and reclaim activity */
1988 if (sc->gfp_mask & __GFP_REPEAT) {
1989 /*
1990 * For __GFP_REPEAT allocations, stop reclaiming if the
1991 * full LRU list has been scanned and we are still failing
1992 * to reclaim pages. This full LRU scan is potentially
1993 * expensive but a __GFP_REPEAT caller really wants to succeed
1994 */
1995 if (!nr_reclaimed && !nr_scanned)
1996 return false;
1997 } else {
1998 /*
1999 * For non-__GFP_REPEAT allocations which can presumably
2000 * fail without consequence, stop if we failed to reclaim
2001 * any pages from the last SWAP_CLUSTER_MAX number of
2002 * pages that were scanned. This will return to the
2003 * caller faster at the risk reclaim/compaction and
2004 * the resulting allocation attempt fails
2005 */
2006 if (!nr_reclaimed)
2007 return false;
2008 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002009
2010 /*
2011 * If we have not reclaimed enough pages for compaction and the
2012 * inactive lists are large enough, continue reclaiming
2013 */
2014 pages_for_compaction = (2UL << sc->order);
2015 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2016 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2017 if (sc->nr_reclaimed < pages_for_compaction &&
2018 inactive_lru_pages > pages_for_compaction)
2019 return true;
2020
2021 /* If compaction would go ahead or the allocation would succeed, stop */
2022 switch (compaction_suitable(zone, sc->order)) {
2023 case COMPACT_PARTIAL:
2024 case COMPACT_CONTINUE:
2025 return false;
2026 default:
2027 return true;
2028 }
2029}
2030
2031/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2033 */
Rik van Riela79311c2009-01-06 14:40:01 -08002034static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08002035 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036{
Christoph Lameterb69408e2008-10-18 20:26:14 -07002037 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08002038 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07002039 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002040 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002041 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Shaohua Li3da367c2011-10-31 17:07:03 -07002042 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Mel Gorman3e7d3442011-01-13 15:45:56 -08002044restart:
2045 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08002046 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002047 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002048
Shaohua Li3da367c2011-10-31 17:07:03 -07002049 blk_start_plug(&plug);
Rik van Riel556adec2008-10-18 20:26:34 -07002050 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2051 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002052 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07002053 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08002054 nr_to_scan = min_t(unsigned long,
2055 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002056 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002058 nr_reclaimed += shrink_list(l, nr_to_scan,
2059 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002060 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 }
Rik van Riela79311c2009-01-06 14:40:01 -08002062 /*
2063 * On large memory systems, scan >> priority can become
2064 * really large. This is fine for the starting priority;
2065 * we want to put equal scanning pressure on each zone.
2066 * However, if the VM has a harder time of freeing pages,
2067 * with multiple processes reclaiming pages, the total
2068 * freeing target can get unreasonably large.
2069 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08002070 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08002071 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 }
Shaohua Li3da367c2011-10-31 17:07:03 -07002073 blk_finish_plug(&plug);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002074 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08002075
Rik van Riel556adec2008-10-18 20:26:34 -07002076 /*
2077 * Even if we did not try to evict anon pages at all, we want to
2078 * rebalance the anon lru active/inactive ratio.
2079 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002080 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002081 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2082
Mel Gorman3e7d3442011-01-13 15:45:56 -08002083 /* reclaim/compaction might need reclaim to continue */
2084 if (should_continue_reclaim(zone, nr_reclaimed,
2085 sc->nr_scanned - nr_scanned, sc))
2086 goto restart;
2087
Andrew Morton232ea4d2007-02-28 20:13:21 -08002088 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089}
2090
2091/*
2092 * This is the direct reclaim path, for page-allocating processes. We only
2093 * try to reclaim pages from zones which will satisfy the caller's allocation
2094 * request.
2095 *
Mel Gorman41858962009-06-16 15:32:12 -07002096 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2097 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2099 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002100 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2101 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2102 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 * If a zone is deemed to be full of pinned pages then just give it a light
2105 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002106 *
2107 * This function returns true if a zone is being reclaimed for a costly
2108 * high-order allocation and compaction is either ready to begin or deferred.
2109 * This indicates to the caller that it should retry the allocation or fail.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 */
Mel Gormane0c23272011-10-31 17:09:33 -07002111static bool shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08002112 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113{
Mel Gormandd1a2392008-04-28 02:12:17 -07002114 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002115 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002116 unsigned long nr_soft_reclaimed;
2117 unsigned long nr_soft_scanned;
Mel Gormane0c23272011-10-31 17:09:33 -07002118 bool should_abort_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002119
Mel Gormand4debc62010-08-09 17:19:29 -07002120 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2121 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002122 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002124 /*
2125 * Take care memory controller reclaiming has small influence
2126 * to global LRU.
2127 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002128 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002129 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2130 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002131 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002132 continue; /* Let kswapd poll it */
Rik van Riele0887c12011-10-31 17:09:31 -07002133 if (COMPACTION_BUILD) {
2134 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002135 * If we already have plenty of memory free for
2136 * compaction in this zone, don't free any more.
2137 * Even though compaction is invoked for any
2138 * non-zero order, only frequent costly order
2139 * reclamation is disruptive enough to become a
2140 * noticable problem, like transparent huge page
2141 * allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002142 */
2143 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2144 (compaction_suitable(zone, sc->order) ||
Mel Gormane0c23272011-10-31 17:09:33 -07002145 compaction_deferred(zone))) {
2146 should_abort_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002147 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002148 }
Rik van Riele0887c12011-10-31 17:09:31 -07002149 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002150 /*
2151 * This steals pages from memory cgroups over softlimit
2152 * and returns the number of reclaimed pages and
2153 * scanned pages. This works for global memory pressure
2154 * and balancing, not for a memcg's limit.
2155 */
2156 nr_soft_scanned = 0;
2157 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2158 sc->order, sc->gfp_mask,
2159 &nr_soft_scanned);
2160 sc->nr_reclaimed += nr_soft_reclaimed;
2161 sc->nr_scanned += nr_soft_scanned;
2162 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002163 }
Nick Piggin408d8542006-09-25 23:31:27 -07002164
Rik van Riela79311c2009-01-06 14:40:01 -08002165 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 }
Mel Gormane0c23272011-10-31 17:09:33 -07002167
2168 return should_abort_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002169}
2170
2171static bool zone_reclaimable(struct zone *zone)
2172{
2173 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2174}
2175
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002176/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002177static bool all_unreclaimable(struct zonelist *zonelist,
2178 struct scan_control *sc)
2179{
2180 struct zoneref *z;
2181 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002182
2183 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2184 gfp_zone(sc->gfp_mask), sc->nodemask) {
2185 if (!populated_zone(zone))
2186 continue;
2187 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2188 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002189 if (!zone->all_unreclaimable)
2190 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002191 }
2192
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002193 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002195
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196/*
2197 * This is the main entry point to direct page reclaim.
2198 *
2199 * If a full scan of the inactive list fails to free enough memory then we
2200 * are "out of memory" and something needs to be killed.
2201 *
2202 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2203 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002204 * caller can't do much about. We kick the writeback threads and take explicit
2205 * naps in the hope that some of these pages can be written. But if the
2206 * allocating task holds filesystem locks which prevent writeout this might not
2207 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002208 *
2209 * returns: 0, if no pages reclaimed
2210 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 */
Mel Gormandac1d272008-04-28 02:12:12 -07002212static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002213 struct scan_control *sc,
2214 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215{
2216 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002217 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002219 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002220 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002221 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Miao Xiec0ff7452010-05-24 14:32:08 -07002223 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002224 delayacct_freepages_start();
2225
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002226 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002227 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228
2229 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002230 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002231 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002232 disable_swap_token(sc->mem_cgroup);
Mel Gormane0c23272011-10-31 17:09:33 -07002233 if (shrink_zones(priority, zonelist, sc))
2234 break;
2235
Balbir Singh66e17072008-02-07 00:13:56 -08002236 /*
2237 * Don't shrink slabs when reclaiming memory from
2238 * over limit cgroups
2239 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002240 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002241 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002242 for_each_zone_zonelist(zone, z, zonelist,
2243 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002244 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2245 continue;
2246
2247 lru_pages += zone_reclaimable_pages(zone);
2248 }
2249
Ying Han1495f232011-05-24 17:12:27 -07002250 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002251 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002252 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002253 reclaim_state->reclaimed_slab = 0;
2254 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 }
Balbir Singh66e17072008-02-07 00:13:56 -08002256 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002257 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260 /*
2261 * Try to write back as many pages as we just scanned. This
2262 * tends to cause slow streaming writers to write data to the
2263 * disk smoothly, at the dirtying rate, which is nice. But
2264 * that's undesirable in laptop mode, where we *want* lumpy
2265 * writeout. So in laptop mode, write out the whole world.
2266 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002267 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2268 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002269 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002270 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 }
2272
2273 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002274 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002275 priority < DEF_PRIORITY - 2) {
2276 struct zone *preferred_zone;
2277
2278 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002279 &cpuset_current_mems_allowed,
2280 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002281 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2282 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002284
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002286 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002287 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002288
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002289 if (sc->nr_reclaimed)
2290 return sc->nr_reclaimed;
2291
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002292 /*
2293 * As hibernation is going on, kswapd is freezed so that it can't mark
2294 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2295 * check.
2296 */
2297 if (oom_killer_disabled)
2298 return 0;
2299
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002300 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002301 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002302 return 1;
2303
2304 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305}
2306
Mel Gormandac1d272008-04-28 02:12:12 -07002307unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002308 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002309{
Mel Gorman33906bc2010-08-09 17:19:16 -07002310 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002311 struct scan_control sc = {
2312 .gfp_mask = gfp_mask,
2313 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002314 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002315 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002316 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002317 .order = order,
2318 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002319 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002320 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002321 struct shrink_control shrink = {
2322 .gfp_mask = sc.gfp_mask,
2323 };
Balbir Singh66e17072008-02-07 00:13:56 -08002324
Mel Gorman33906bc2010-08-09 17:19:16 -07002325 trace_mm_vmscan_direct_reclaim_begin(order,
2326 sc.may_writepage,
2327 gfp_mask);
2328
Ying Hana09ed5e2011-05-24 17:12:26 -07002329 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002330
2331 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2332
2333 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002334}
2335
Balbir Singh00f0b822008-03-04 14:28:39 -08002336#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002337
Balbir Singh4e416952009-09-23 15:56:39 -07002338unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2339 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002340 struct zone *zone,
2341 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002342{
2343 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002344 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002345 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002346 .may_writepage = !laptop_mode,
2347 .may_unmap = 1,
2348 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002349 .order = 0,
2350 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002351 };
Ying Han0ae5e892011-05-26 16:25:25 -07002352
Balbir Singh4e416952009-09-23 15:56:39 -07002353 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2354 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002355
2356 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2357 sc.may_writepage,
2358 sc.gfp_mask);
2359
Balbir Singh4e416952009-09-23 15:56:39 -07002360 /*
2361 * NOTE: Although we can get the priority field, using it
2362 * here is not a good idea, since it limits the pages we can scan.
2363 * if we don't reclaim here, the shrink_zone from balance_pgdat
2364 * will pick up pages from other mem cgroup's as well. We hack
2365 * the priority and make it zero.
2366 */
2367 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002368
2369 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2370
Ying Han0ae5e892011-05-26 16:25:25 -07002371 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002372 return sc.nr_reclaimed;
2373}
2374
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002375unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002376 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002377 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002378{
Balbir Singh4e416952009-09-23 15:56:39 -07002379 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002380 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002381 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002382 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002383 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002384 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002385 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002386 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002387 .order = 0,
2388 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002389 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002390 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2391 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2392 };
2393 struct shrink_control shrink = {
2394 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002395 };
Balbir Singh66e17072008-02-07 00:13:56 -08002396
Ying Han889976d2011-05-26 16:25:33 -07002397 /*
2398 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2399 * take care of from where we get pages. So the node where we start the
2400 * scan does not need to be the current node.
2401 */
2402 nid = mem_cgroup_select_victim_node(mem_cont);
2403
2404 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002405
2406 trace_mm_vmscan_memcg_reclaim_begin(0,
2407 sc.may_writepage,
2408 sc.gfp_mask);
2409
Ying Hana09ed5e2011-05-24 17:12:26 -07002410 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002411
2412 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2413
2414 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002415}
2416#endif
2417
Mel Gorman1741c872011-01-13 15:46:21 -08002418/*
2419 * pgdat_balanced is used when checking if a node is balanced for high-order
2420 * allocations. Only zones that meet watermarks and are in a zone allowed
2421 * by the callers classzone_idx are added to balanced_pages. The total of
2422 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2423 * for the node to be considered balanced. Forcing all zones to be balanced
2424 * for high orders can cause excessive reclaim when there are imbalanced zones.
2425 * The choice of 25% is due to
2426 * o a 16M DMA zone that is balanced will not balance a zone on any
2427 * reasonable sized machine
2428 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002429 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002430 * would need to be at least 256M for it to be balance a whole node.
2431 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2432 * to balance a node on its own. These seemed like reasonable ratios.
2433 */
2434static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2435 int classzone_idx)
2436{
2437 unsigned long present_pages = 0;
2438 int i;
2439
2440 for (i = 0; i <= classzone_idx; i++)
2441 present_pages += pgdat->node_zones[i].present_pages;
2442
Shaohua Li4746efd2011-07-19 08:49:26 -07002443 /* A special case here: if zone has no page, we think it's balanced */
2444 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002445}
2446
Mel Gormanf50de2d2009-12-14 17:58:53 -08002447/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002448static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2449 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002450{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002451 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002452 unsigned long balanced = 0;
2453 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002454
2455 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2456 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002457 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002458
Mel Gorman0abdee22011-01-13 15:46:22 -08002459 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002460 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002461 struct zone *zone = pgdat->node_zones + i;
2462
2463 if (!populated_zone(zone))
2464 continue;
2465
Mel Gorman355b09c2011-01-13 15:46:24 -08002466 /*
2467 * balance_pgdat() skips over all_unreclaimable after
2468 * DEF_PRIORITY. Effectively, it considers them balanced so
2469 * they must be considered balanced here as well if kswapd
2470 * is to sleep
2471 */
2472 if (zone->all_unreclaimable) {
2473 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002474 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002475 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002476
Mel Gorman88f5acf2011-01-13 15:45:41 -08002477 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002478 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002479 all_zones_ok = false;
2480 else
2481 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002482 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002483
Mel Gorman1741c872011-01-13 15:46:21 -08002484 /*
2485 * For high-order requests, the balanced zones must contain at least
2486 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2487 * must be balanced
2488 */
2489 if (order)
Johannes Weinerafc7e322011-05-24 17:11:09 -07002490 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002491 else
2492 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002493}
2494
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495/*
2496 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002497 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002499 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 *
2501 * There is special handling here for zones which are full of pinned pages.
2502 * This can happen if the pages are all mlocked, or if they are all used by
2503 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2504 * What we do is to detect the case where all pages in the zone have been
2505 * scanned twice and there has been zero successful reclaim. Mark the zone as
2506 * dead and from now on, only perform a short scan. Basically we're polling
2507 * the zone for when the problem goes away.
2508 *
2509 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002510 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2511 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2512 * lower zones regardless of the number of free pages in the lower zones. This
2513 * interoperates with the page allocator fallback scheme to ensure that aging
2514 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 */
Mel Gorman99504742011-01-13 15:46:20 -08002516static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002517 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002520 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 int priority;
2522 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002523 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002524 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002526 unsigned long nr_soft_reclaimed;
2527 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002528 struct scan_control sc = {
2529 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002530 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002531 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002532 /*
2533 * kswapd doesn't want to be bailed out while reclaim. because
2534 * we want to put equal scanning pressure on each zone.
2535 */
2536 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002537 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002538 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002539 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002540 struct shrink_control shrink = {
2541 .gfp_mask = sc.gfp_mask,
2542 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543loop_again:
2544 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002545 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002546 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002547 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002551 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002553 /* The swap token gets in the way of swapout... */
2554 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002555 disable_swap_token(NULL);
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002556
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002558 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002560 /*
2561 * Scan in the highmem->dma direction for the highest
2562 * zone which needs scanning
2563 */
2564 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2565 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002567 if (!populated_zone(zone))
2568 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002570 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002571 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572
Rik van Riel556adec2008-10-18 20:26:34 -07002573 /*
2574 * Do some background aging of the anon list, to give
2575 * pages a chance to be referenced before reclaiming.
2576 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002577 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002578 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2579 &sc, priority, 0);
2580
Mel Gorman88f5acf2011-01-13 15:45:41 -08002581 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002582 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002583 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002584 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002585 } else {
2586 /* If balanced, clear the congested flag */
2587 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002590 if (i < 0)
2591 goto out;
2592
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 for (i = 0; i <= end_zone; i++) {
2594 struct zone *zone = pgdat->node_zones + i;
2595
Wu Fengguangadea02a2009-09-21 17:01:42 -07002596 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 }
2598
2599 /*
2600 * Now scan the zone in the dma->highmem direction, stopping
2601 * at the last zone which needs scanning.
2602 *
2603 * We do this because the page allocator works in the opposite
2604 * direction. This prevents the page allocator from allocating
2605 * pages behind kswapd's direction of progress, which would
2606 * cause too much scanning of the lower zones.
2607 */
2608 for (i = 0; i <= end_zone; i++) {
2609 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002610 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002611 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612
Con Kolivasf3fe6512006-01-06 00:11:15 -08002613 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 continue;
2615
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002616 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 continue;
2618
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002620
Ying Han0ae5e892011-05-26 16:25:25 -07002621 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002622 /*
2623 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002624 */
Ying Han0ae5e892011-05-26 16:25:25 -07002625 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2626 order, sc.gfp_mask,
2627 &nr_soft_scanned);
2628 sc.nr_reclaimed += nr_soft_reclaimed;
2629 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002630
Rik van Riel32a43302007-10-16 01:24:50 -07002631 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002632 * We put equal pressure on every zone, unless
2633 * one zone has way too many pages free
2634 * already. The "too many pages" is defined
2635 * as the high wmark plus a "gap" where the
2636 * gap is either the low watermark or 1%
2637 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002638 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002639 balance_gap = min(low_wmark_pages(zone),
2640 (zone->present_pages +
2641 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2642 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002643 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002644 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002645 end_zone, 0)) {
Rik van Riela79311c2009-01-06 14:40:01 -08002646 shrink_zone(priority, zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002647
Mel Gormand7868da2011-07-08 15:39:38 -07002648 reclaim_state->reclaimed_slab = 0;
2649 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2650 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2651 total_scanned += sc.nr_scanned;
2652
2653 if (nr_slab == 0 && !zone_reclaimable(zone))
2654 zone->all_unreclaimable = 1;
2655 }
2656
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 /*
2658 * If we've done a decent amount of scanning and
2659 * the reclaim ratio is low, start doing writepage
2660 * even in laptop mode
2661 */
2662 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002663 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002665
Mel Gorman215ddd62011-07-08 15:39:40 -07002666 if (zone->all_unreclaimable) {
2667 if (end_zone && end_zone == i)
2668 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002669 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002670 }
Mel Gormand7868da2011-07-08 15:39:38 -07002671
Mel Gorman88f5acf2011-01-13 15:45:41 -08002672 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002673 high_wmark_pages(zone), end_zone, 0)) {
2674 all_zones_ok = 0;
2675 /*
2676 * We are still under min water mark. This
2677 * means that we have a GFP_ATOMIC allocation
2678 * failure risk. Hurry up!
2679 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002680 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002681 min_wmark_pages(zone), end_zone, 0))
2682 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002683 } else {
2684 /*
2685 * If a zone reaches its high watermark,
2686 * consider it to be no longer congested. It's
2687 * possible there are dirty pages backed by
2688 * congested BDIs but as pressure is relieved,
2689 * spectulatively avoid congestion waits
2690 */
2691 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002692 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002693 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002694 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002695
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002697 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 break; /* kswapd: all done */
2699 /*
2700 * OK, kswapd is getting into trouble. Take a nap, then take
2701 * another pass across the zones.
2702 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002703 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2704 if (has_under_min_watermark_zone)
2705 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2706 else
2707 congestion_wait(BLK_RW_ASYNC, HZ/10);
2708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709
2710 /*
2711 * We do this so kswapd doesn't build up large priorities for
2712 * example when it is freeing in parallel with allocators. It
2713 * matches the direct reclaim path behaviour in terms of impact
2714 * on zone->*_priority.
2715 */
Rik van Riela79311c2009-01-06 14:40:01 -08002716 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 break;
2718 }
2719out:
Mel Gorman99504742011-01-13 15:46:20 -08002720
2721 /*
2722 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002723 * high-order: Balanced zones must make up at least 25% of the node
2724 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002725 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002726 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002728
2729 try_to_freeze();
2730
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002731 /*
2732 * Fragmentation may mean that the system cannot be
2733 * rebalanced for high-order allocations in all zones.
2734 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2735 * it means the zones have been fully scanned and are still
2736 * not balanced. For high-order allocations, there is
2737 * little point trying all over again as kswapd may
2738 * infinite loop.
2739 *
2740 * Instead, recheck all watermarks at order-0 as they
2741 * are the most important. If watermarks are ok, kswapd will go
2742 * back to sleep. High-order users can still perform direct
2743 * reclaim if they wish.
2744 */
2745 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2746 order = sc.order = 0;
2747
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 goto loop_again;
2749 }
2750
Mel Gorman99504742011-01-13 15:46:20 -08002751 /*
2752 * If kswapd was reclaiming at a higher order, it has the option of
2753 * sleeping without all zones being balanced. Before it does, it must
2754 * ensure that the watermarks for order-0 on *all* zones are met and
2755 * that the congestion flags are cleared. The congestion flag must
2756 * be cleared as kswapd is the only mechanism that clears the flag
2757 * and it is potentially going to sleep here.
2758 */
2759 if (order) {
2760 for (i = 0; i <= end_zone; i++) {
2761 struct zone *zone = pgdat->node_zones + i;
2762
2763 if (!populated_zone(zone))
2764 continue;
2765
2766 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2767 continue;
2768
2769 /* Confirm the zone is balanced for order-0 */
2770 if (!zone_watermark_ok(zone, 0,
2771 high_wmark_pages(zone), 0, 0)) {
2772 order = sc.order = 0;
2773 goto loop_again;
2774 }
2775
2776 /* If balanced, clear the congested flag */
2777 zone_clear_flag(zone, ZONE_CONGESTED);
Shaohua Li16fb9512011-10-31 17:08:02 -07002778 if (i <= *classzone_idx)
2779 balanced += zone->present_pages;
Mel Gorman99504742011-01-13 15:46:20 -08002780 }
2781 }
2782
Mel Gorman0abdee22011-01-13 15:46:22 -08002783 /*
2784 * Return the order we were reclaiming at so sleeping_prematurely()
2785 * makes a decision on the order we were last reclaiming at. However,
2786 * if another caller entered the allocator slow path while kswapd
2787 * was awake, order will remain at the higher level
2788 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002789 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002790 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791}
2792
Mel Gormandc83edd2011-01-13 15:46:26 -08002793static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002794{
2795 long remaining = 0;
2796 DEFINE_WAIT(wait);
2797
2798 if (freezing(current) || kthread_should_stop())
2799 return;
2800
2801 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2802
2803 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002804 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002805 remaining = schedule_timeout(HZ/10);
2806 finish_wait(&pgdat->kswapd_wait, &wait);
2807 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2808 }
2809
2810 /*
2811 * After a short sleep, check if it was a premature sleep. If not, then
2812 * go fully to sleep until explicitly woken up.
2813 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002814 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002815 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2816
2817 /*
2818 * vmstat counters are not perfectly accurate and the estimated
2819 * value for counters such as NR_FREE_PAGES can deviate from the
2820 * true value by nr_online_cpus * threshold. To avoid the zone
2821 * watermarks being breached while under pressure, we reduce the
2822 * per-cpu vmstat threshold while kswapd is awake and restore
2823 * them before going back to sleep.
2824 */
2825 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2826 schedule();
2827 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2828 } else {
2829 if (remaining)
2830 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2831 else
2832 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2833 }
2834 finish_wait(&pgdat->kswapd_wait, &wait);
2835}
2836
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837/*
2838 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002839 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 *
2841 * This basically trickles out pages so that we have _some_
2842 * free memory available even if there is no other activity
2843 * that frees anything up. This is needed for things like routing
2844 * etc, where we otherwise might have all activity going on in
2845 * asynchronous contexts that cannot page things out.
2846 *
2847 * If there are applications that are active memory-allocators
2848 * (most normal use), this basically shouldn't matter.
2849 */
2850static int kswapd(void *p)
2851{
Mel Gorman215ddd62011-07-08 15:39:40 -07002852 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002853 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07002854 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002855 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 pg_data_t *pgdat = (pg_data_t*)p;
2857 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002858
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 struct reclaim_state reclaim_state = {
2860 .reclaimed_slab = 0,
2861 };
Rusty Russella70f7302009-03-13 14:49:46 +10302862 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863
Nick Piggincf40bd12009-01-21 08:12:39 +01002864 lockdep_set_current_reclaim_state(GFP_KERNEL);
2865
Rusty Russell174596a2009-01-01 10:12:29 +10302866 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002867 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 current->reclaim_state = &reclaim_state;
2869
2870 /*
2871 * Tell the memory management that we're a "memory allocator",
2872 * and that if we need more memory we should get access to it
2873 * regardless (see "__alloc_pages()"). "kswapd" should
2874 * never get caught in the normal page freeing logic.
2875 *
2876 * (Kswapd normally doesn't need memory anyway, but sometimes
2877 * you need a small amount of memory in order to be able to
2878 * page out something else, and this flag essentially protects
2879 * us from recursively trying to free more memory as we're
2880 * trying to free the first piece of memory in the first place).
2881 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002882 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002883 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884
Mel Gorman215ddd62011-07-08 15:39:40 -07002885 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002886 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002887 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002888 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08002890 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002891
Mel Gorman215ddd62011-07-08 15:39:40 -07002892 /*
2893 * If the last balance_pgdat was unsuccessful it's unlikely a
2894 * new request of a similar or harder type will succeed soon
2895 * so consider going to sleep on the basis we reclaimed at
2896 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002897 if (balanced_classzone_idx >= new_classzone_idx &&
2898 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07002899 new_order = pgdat->kswapd_max_order;
2900 new_classzone_idx = pgdat->classzone_idx;
2901 pgdat->kswapd_max_order = 0;
2902 pgdat->classzone_idx = pgdat->nr_zones - 1;
2903 }
2904
Mel Gorman99504742011-01-13 15:46:20 -08002905 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 /*
2907 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002908 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 */
2910 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002911 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002913 kswapd_try_to_sleep(pgdat, balanced_order,
2914 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002916 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07002917 new_order = order;
2918 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002919 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002920 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922
David Rientjes8fe23e02009-12-14 17:58:33 -08002923 ret = try_to_freeze();
2924 if (kthread_should_stop())
2925 break;
2926
2927 /*
2928 * We can speed up thawing tasks if we don't call balance_pgdat
2929 * after returning from the refrigerator
2930 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002931 if (!ret) {
2932 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002933 balanced_classzone_idx = classzone_idx;
2934 balanced_order = balance_pgdat(pgdat, order,
2935 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 }
2938 return 0;
2939}
2940
2941/*
2942 * A zone is low on free memory, so wake its kswapd task to service it.
2943 */
Mel Gorman99504742011-01-13 15:46:20 -08002944void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945{
2946 pg_data_t *pgdat;
2947
Con Kolivasf3fe6512006-01-06 00:11:15 -08002948 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 return;
2950
Paul Jackson02a0e532006-12-13 00:34:25 -08002951 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002953 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002954 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002955 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002956 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2957 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002958 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002960 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2961 return;
2962
2963 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002964 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965}
2966
Wu Fengguangadea02a2009-09-21 17:01:42 -07002967/*
2968 * The reclaimable count would be mostly accurate.
2969 * The less reclaimable pages may be
2970 * - mlocked pages, which will be moved to unevictable list when encountered
2971 * - mapped pages, which may require several travels to be reclaimed
2972 * - dirty pages, which is not "instantly" reclaimable
2973 */
2974unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002975{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002976 int nr;
2977
2978 nr = global_page_state(NR_ACTIVE_FILE) +
2979 global_page_state(NR_INACTIVE_FILE);
2980
2981 if (nr_swap_pages > 0)
2982 nr += global_page_state(NR_ACTIVE_ANON) +
2983 global_page_state(NR_INACTIVE_ANON);
2984
2985 return nr;
2986}
2987
2988unsigned long zone_reclaimable_pages(struct zone *zone)
2989{
2990 int nr;
2991
2992 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2993 zone_page_state(zone, NR_INACTIVE_FILE);
2994
2995 if (nr_swap_pages > 0)
2996 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2997 zone_page_state(zone, NR_INACTIVE_ANON);
2998
2999 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003000}
3001
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003002#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003004 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003005 * freed pages.
3006 *
3007 * Rather than trying to age LRUs the aim is to preserve the overall
3008 * LRU order by reclaiming preferentially
3009 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003011unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003013 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003014 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003015 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3016 .may_swap = 1,
3017 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003018 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003019 .nr_to_reclaim = nr_to_reclaim,
3020 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003021 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003023 struct shrink_control shrink = {
3024 .gfp_mask = sc.gfp_mask,
3025 };
3026 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003027 struct task_struct *p = current;
3028 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003030 p->flags |= PF_MEMALLOC;
3031 lockdep_set_current_reclaim_state(sc.gfp_mask);
3032 reclaim_state.reclaimed_slab = 0;
3033 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003034
Ying Hana09ed5e2011-05-24 17:12:26 -07003035 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003036
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003037 p->reclaim_state = NULL;
3038 lockdep_clear_current_reclaim_state();
3039 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003040
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003041 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003043#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045/* It's optimal to keep kswapds on the same CPUs as their memory, but
3046 not required for correctness. So if the last cpu in a node goes
3047 away, we get changed to run anywhere: as the first one comes back,
3048 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003049static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08003050 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003052 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003054 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003055 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003056 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303057 const struct cpumask *mask;
3058
3059 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003060
Rusty Russell3e597942009-01-01 10:12:24 +10303061 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003063 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 }
3065 }
3066 return NOTIFY_OK;
3067}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
Yasunori Goto3218ae12006-06-27 02:53:33 -07003069/*
3070 * This kswapd start function will be called by init and node-hot-add.
3071 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3072 */
3073int kswapd_run(int nid)
3074{
3075 pg_data_t *pgdat = NODE_DATA(nid);
3076 int ret = 0;
3077
3078 if (pgdat->kswapd)
3079 return 0;
3080
3081 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3082 if (IS_ERR(pgdat->kswapd)) {
3083 /* failure at boot is fatal */
3084 BUG_ON(system_state == SYSTEM_BOOTING);
3085 printk("Failed to start kswapd on node %d\n",nid);
3086 ret = -1;
3087 }
3088 return ret;
3089}
3090
David Rientjes8fe23e02009-12-14 17:58:33 -08003091/*
3092 * Called by memory hotplug when all memory in a node is offlined.
3093 */
3094void kswapd_stop(int nid)
3095{
3096 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3097
3098 if (kswapd)
3099 kthread_stop(kswapd);
3100}
3101
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102static int __init kswapd_init(void)
3103{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003104 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003105
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003107 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003108 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 hotcpu_notifier(cpu_callback, 0);
3110 return 0;
3111}
3112
3113module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003114
3115#ifdef CONFIG_NUMA
3116/*
3117 * Zone reclaim mode
3118 *
3119 * If non-zero call zone_reclaim when the number of free pages falls below
3120 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003121 */
3122int zone_reclaim_mode __read_mostly;
3123
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003124#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003125#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003126#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3127#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3128
Christoph Lameter9eeff232006-01-18 17:42:31 -08003129/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003130 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3131 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3132 * a zone.
3133 */
3134#define ZONE_RECLAIM_PRIORITY 4
3135
Christoph Lameter9eeff232006-01-18 17:42:31 -08003136/*
Christoph Lameter96146342006-07-03 00:24:13 -07003137 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3138 * occur.
3139 */
3140int sysctl_min_unmapped_ratio = 1;
3141
3142/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003143 * If the number of slab pages in a zone grows beyond this percentage then
3144 * slab reclaim needs to occur.
3145 */
3146int sysctl_min_slab_ratio = 5;
3147
Mel Gorman90afa5d2009-06-16 15:33:20 -07003148static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3149{
3150 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3151 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3152 zone_page_state(zone, NR_ACTIVE_FILE);
3153
3154 /*
3155 * It's possible for there to be more file mapped pages than
3156 * accounted for by the pages on the file LRU lists because
3157 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3158 */
3159 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3160}
3161
3162/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3163static long zone_pagecache_reclaimable(struct zone *zone)
3164{
3165 long nr_pagecache_reclaimable;
3166 long delta = 0;
3167
3168 /*
3169 * If RECLAIM_SWAP is set, then all file pages are considered
3170 * potentially reclaimable. Otherwise, we have to worry about
3171 * pages like swapcache and zone_unmapped_file_pages() provides
3172 * a better estimate
3173 */
3174 if (zone_reclaim_mode & RECLAIM_SWAP)
3175 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3176 else
3177 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3178
3179 /* If we can't clean pages, remove dirty pages from consideration */
3180 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3181 delta += zone_page_state(zone, NR_FILE_DIRTY);
3182
3183 /* Watch for any possible underflows due to delta */
3184 if (unlikely(delta > nr_pagecache_reclaimable))
3185 delta = nr_pagecache_reclaimable;
3186
3187 return nr_pagecache_reclaimable - delta;
3188}
3189
Christoph Lameter0ff38492006-09-25 23:31:52 -07003190/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003191 * Try to free up some pages from this zone through reclaim.
3192 */
Andrew Morton179e9632006-03-22 00:08:18 -08003193static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003194{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003195 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003196 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003197 struct task_struct *p = current;
3198 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08003199 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08003200 struct scan_control sc = {
3201 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07003202 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003203 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003204 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3205 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003206 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003207 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003208 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003209 struct shrink_control shrink = {
3210 .gfp_mask = sc.gfp_mask,
3211 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003212 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003213
Christoph Lameter9eeff232006-01-18 17:42:31 -08003214 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003215 /*
3216 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3217 * and we also need to be able to write out pages for RECLAIM_WRITE
3218 * and RECLAIM_SWAP.
3219 */
3220 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003221 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003222 reclaim_state.reclaimed_slab = 0;
3223 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003224
Mel Gorman90afa5d2009-06-16 15:33:20 -07003225 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003226 /*
3227 * Free memory by calling shrink zone with increasing
3228 * priorities until we have enough memory freed.
3229 */
3230 priority = ZONE_RECLAIM_PRIORITY;
3231 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003232 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003233 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003234 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003235 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003236
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003237 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3238 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003239 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003240 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003241 * many pages were freed in this zone. So we take the current
3242 * number of slab pages and shake the slab until it is reduced
3243 * by the same nr_pages that we used for reclaiming unmapped
3244 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003245 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003246 * Note that shrink_slab will free memory on all zones and may
3247 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003248 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003249 for (;;) {
3250 unsigned long lru_pages = zone_reclaimable_pages(zone);
3251
3252 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003253 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003254 break;
3255
3256 /* Freed enough memory */
3257 nr_slab_pages1 = zone_page_state(zone,
3258 NR_SLAB_RECLAIMABLE);
3259 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3260 break;
3261 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003262
3263 /*
3264 * Update nr_reclaimed by the number of slab pages we
3265 * reclaimed from this zone.
3266 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003267 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3268 if (nr_slab_pages1 < nr_slab_pages0)
3269 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003270 }
3271
Christoph Lameter9eeff232006-01-18 17:42:31 -08003272 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003273 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003274 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003275 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003276}
Andrew Morton179e9632006-03-22 00:08:18 -08003277
3278int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3279{
Andrew Morton179e9632006-03-22 00:08:18 -08003280 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003281 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003282
3283 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003284 * Zone reclaim reclaims unmapped file backed pages and
3285 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003286 *
Christoph Lameter96146342006-07-03 00:24:13 -07003287 * A small portion of unmapped file backed pages is needed for
3288 * file I/O otherwise pages read by file I/O will be immediately
3289 * thrown out if the zone is overallocated. So we do not reclaim
3290 * if less than a specified percentage of the zone is used by
3291 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003292 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003293 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3294 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003295 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003296
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003297 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003298 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003299
Andrew Morton179e9632006-03-22 00:08:18 -08003300 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003301 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003302 */
David Rientjesd773ed62007-10-16 23:26:01 -07003303 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003304 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003305
3306 /*
3307 * Only run zone reclaim on the local zone or on zones that do not
3308 * have associated processors. This will favor the local processor
3309 * over remote processors and spread off node memory allocations
3310 * as wide as possible.
3311 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003312 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003313 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003314 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003315
3316 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003317 return ZONE_RECLAIM_NOSCAN;
3318
David Rientjesd773ed62007-10-16 23:26:01 -07003319 ret = __zone_reclaim(zone, gfp_mask, order);
3320 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3321
Mel Gorman24cf725182009-06-16 15:33:23 -07003322 if (!ret)
3323 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3324
David Rientjesd773ed62007-10-16 23:26:01 -07003325 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003326}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003327#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003328
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003329/*
3330 * page_evictable - test whether a page is evictable
3331 * @page: the page to test
3332 * @vma: the VMA in which the page is or will be mapped, may be NULL
3333 *
3334 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003335 * lists vs unevictable list. The vma argument is !NULL when called from the
3336 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003337 *
3338 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003339 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003340 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003341 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003342 */
3343int page_evictable(struct page *page, struct vm_area_struct *vma)
3344{
3345
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003346 if (mapping_unevictable(page_mapping(page)))
3347 return 0;
3348
Nick Pigginb291f002008-10-18 20:26:44 -07003349 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3350 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003351
3352 return 1;
3353}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003354
3355/**
3356 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3357 * @page: page to check evictability and move to appropriate lru list
3358 * @zone: zone page is in
3359 *
3360 * Checks a page for evictability and moves the page to the appropriate
3361 * zone lru list.
3362 *
3363 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3364 * have PageUnevictable set.
3365 */
3366static void check_move_unevictable_page(struct page *page, struct zone *zone)
3367{
3368 VM_BUG_ON(PageActive(page));
3369
3370retry:
3371 ClearPageUnevictable(page);
3372 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003373 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003374
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003375 __dec_zone_state(zone, NR_UNEVICTABLE);
3376 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003377 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003378 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3379 __count_vm_event(UNEVICTABLE_PGRESCUED);
3380 } else {
3381 /*
3382 * rotate unevictable list
3383 */
3384 SetPageUnevictable(page);
3385 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003386 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003387 if (page_evictable(page, NULL))
3388 goto retry;
3389 }
3390}
3391
3392/**
3393 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3394 * @mapping: struct address_space to scan for evictable pages
3395 *
3396 * Scan all pages in mapping. Check unevictable pages for
3397 * evictability and move them to the appropriate zone lru list.
3398 */
3399void scan_mapping_unevictable_pages(struct address_space *mapping)
3400{
3401 pgoff_t next = 0;
3402 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3403 PAGE_CACHE_SHIFT;
3404 struct zone *zone;
3405 struct pagevec pvec;
3406
3407 if (mapping->nrpages == 0)
3408 return;
3409
3410 pagevec_init(&pvec, 0);
3411 while (next < end &&
3412 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3413 int i;
3414 int pg_scanned = 0;
3415
3416 zone = NULL;
3417
3418 for (i = 0; i < pagevec_count(&pvec); i++) {
3419 struct page *page = pvec.pages[i];
3420 pgoff_t page_index = page->index;
3421 struct zone *pagezone = page_zone(page);
3422
3423 pg_scanned++;
3424 if (page_index > next)
3425 next = page_index;
3426 next++;
3427
3428 if (pagezone != zone) {
3429 if (zone)
3430 spin_unlock_irq(&zone->lru_lock);
3431 zone = pagezone;
3432 spin_lock_irq(&zone->lru_lock);
3433 }
3434
3435 if (PageLRU(page) && PageUnevictable(page))
3436 check_move_unevictable_page(page, zone);
3437 }
3438 if (zone)
3439 spin_unlock_irq(&zone->lru_lock);
3440 pagevec_release(&pvec);
3441
3442 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3443 }
3444
3445}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003446
Johannes Weiner264e56d2011-10-31 17:09:13 -07003447static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003448{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003449 printk_once(KERN_WARNING
3450 "The scan_unevictable_pages sysctl/node-interface has been "
3451 "disabled for lack of a legitimate use case. If you have "
3452 "one, please send an email to linux-mm@kvack.org.\n");
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003453}
3454
3455/*
3456 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3457 * all nodes' unevictable lists for evictable pages
3458 */
3459unsigned long scan_unevictable_pages;
3460
3461int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003462 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003463 size_t *length, loff_t *ppos)
3464{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003465 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003466 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003467 scan_unevictable_pages = 0;
3468 return 0;
3469}
3470
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003471#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003472/*
3473 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3474 * a specified node's per zone unevictable lists for evictable pages.
3475 */
3476
3477static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3478 struct sysdev_attribute *attr,
3479 char *buf)
3480{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003481 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003482 return sprintf(buf, "0\n"); /* always zero; should fit... */
3483}
3484
3485static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3486 struct sysdev_attribute *attr,
3487 const char *buf, size_t count)
3488{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003489 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003490 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