blob: 7e0116150dc73a02b762cb8939c5db34688aa194 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51
Nick Piggin0f8053a2006-03-22 00:08:33 -080052#include "internal.h"
53
Mel Gorman33906bc2010-08-09 17:19:16 -070054#define CREATE_TRACE_POINTS
55#include <trace/events/vmscan.h>
56
Mel Gormanee64fc92011-01-13 15:45:55 -080057/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080058 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080063 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080065 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080066 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080067 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080068typedef unsigned __bitwise__ reclaim_mode_t;
69#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070074
Linus Torvalds1da177e2005-04-16 15:20:36 -070075struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
78
Rik van Riela79311c2009-01-06 14:40:01 -080079 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
81
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080082 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
84
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080085 unsigned long hibernation_mode;
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040088 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 int may_writepage;
91
Johannes Weinera6dc60f82009-03-31 15:19:30 -070092 /* Can mapped pages be reclaimed? */
93 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080094
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070095 /* Can pages be swapped as part of reclaim? */
96 int may_swap;
97
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070098 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070099
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700100 int order;
Balbir Singh66e17072008-02-07 00:13:56 -0800101
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700102 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530103 * Intend to reclaim enough continuous memory rather than reclaim
104 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700105 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800106 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700107
Balbir Singh66e17072008-02-07 00:13:56 -0800108 /* Which cgroup do we reclaim from */
109 struct mem_cgroup *mem_cgroup;
110
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700111 /*
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 * are scanned.
114 */
115 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120#ifdef ARCH_HAS_PREFETCH
121#define prefetch_prev_lru_page(_page, _base, _field) \
122 do { \
123 if ((_page)->lru.prev != _base) { \
124 struct page *prev; \
125 \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
128 } \
129 } while (0)
130#else
131#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132#endif
133
134#ifdef ARCH_HAS_PREFETCHW
135#define prefetchw_prev_lru_page(_page, _base, _field) \
136 do { \
137 if ((_page)->lru.prev != _base) { \
138 struct page *prev; \
139 \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
142 } \
143 } while (0)
144#else
145#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146#endif
147
148/*
149 * From 0 .. 100. Higher means more swappy.
150 */
151int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700152long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153
154static LIST_HEAD(shrinker_list);
155static DECLARE_RWSEM(shrinker_rwsem);
156
Balbir Singh00f0b822008-03-04 14:28:39 -0800157#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800158#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800160#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800161#endif
162
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800163static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164 struct scan_control *sc)
165{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800166 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800167 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800169 return &zone->reclaim_stat;
170}
171
Vincent Li0b217672009-09-21 17:03:09 -0700172static unsigned long zone_nr_lru_pages(struct zone *zone,
173 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800174{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800175 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800176 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
177
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800178 return zone_page_state(zone, NR_LRU_BASE + lru);
179}
180
181
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182/*
183 * Add a shrinker callback to be called from the vm
184 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700185void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700187 shrinker->nr = 0;
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700192EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194/*
195 * Remove one
196 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700197void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700203EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
Ying Han1495f232011-05-24 17:12:27 -0700205static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
208{
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
211}
212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213#define SHRINK_BATCH 128
214/*
215 * Call the shrink functions to age shrinkable caches
216 *
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
221 *
Simon Arlott183ff222007-10-20 01:27:18 +0200222 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 * slab to avoid swapping.
224 *
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 *
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700230 *
231 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700233unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700234 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700235 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800238 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Ying Han1495f232011-05-24 17:12:27 -0700240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Minchan Kimf06590b2011-05-24 17:11:11 -0700243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
245 ret = 1;
246 goto out;
247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000252 unsigned long max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
Ying Han1495f232011-05-24 17:12:27 -0700254 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
255 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800256 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 do_div(delta, lru_pages + 1);
258 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800259 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700260 printk(KERN_ERR "shrink_slab: %pF negative objects to "
261 "delete nr=%ld\n",
262 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800263 shrinker->nr = max_pass;
264 }
265
266 /*
267 * Avoid risking looping forever due to too large nr value:
268 * never try to free more than twice the estimate number of
269 * freeable entries.
270 */
271 if (shrinker->nr > max_pass * 2)
272 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274 total_scan = shrinker->nr;
275 shrinker->nr = 0;
276
277 while (total_scan >= SHRINK_BATCH) {
278 long this_scan = SHRINK_BATCH;
279 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700280 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
Ying Han1495f232011-05-24 17:12:27 -0700282 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
283 shrink_ret = do_shrinker_shrink(shrinker, shrink,
284 this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285 if (shrink_ret == -1)
286 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700287 if (shrink_ret < nr_before)
288 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700289 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 total_scan -= this_scan;
291
292 cond_resched();
293 }
294
295 shrinker->nr += total_scan;
296 }
297 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700298out:
299 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700300 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301}
302
Mel Gormanf3a310b2011-01-13 15:46:00 -0800303static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700304 bool sync)
305{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800306 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700307
308 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800309 * Initially assume we are entering either lumpy reclaim or
310 * reclaim/compaction.Depending on the order, we will either set the
311 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700312 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800313 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800314 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800315 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800316 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700317
318 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800319 * Avoid using lumpy reclaim or reclaim/compaction if possible by
320 * restricting when its set to either costly allocations or when
321 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700322 */
323 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800324 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700325 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800326 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700327 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800328 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700329}
330
Mel Gormanf3a310b2011-01-13 15:46:00 -0800331static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700332{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800333 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700334}
335
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336static inline int is_page_cache_freeable(struct page *page)
337{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700338 /*
339 * A freeable page cache page is referenced only by the caller
340 * that isolated the page, the page cache radix tree and
341 * optional buffer heads at page->private.
342 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700343 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344}
345
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700346static int may_write_to_queue(struct backing_dev_info *bdi,
347 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
Christoph Lameter930d9152006-01-08 01:00:47 -0800349 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 return 1;
351 if (!bdi_write_congested(bdi))
352 return 1;
353 if (bdi == current->backing_dev_info)
354 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700355
356 /* lumpy reclaim for hugepage often need a lot of write */
357 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
358 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 return 0;
360}
361
362/*
363 * We detected a synchronous write error writing a page out. Probably
364 * -ENOSPC. We need to propagate that into the address_space for a subsequent
365 * fsync(), msync() or close().
366 *
367 * The tricky part is that after writepage we cannot touch the mapping: nothing
368 * prevents it from being freed up. But we have a ref on the page and once
369 * that page is locked, the mapping is pinned.
370 *
371 * We're allowed to run sleeping lock_page() here because we know the caller has
372 * __GFP_FS.
373 */
374static void handle_write_error(struct address_space *mapping,
375 struct page *page, int error)
376{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100377 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700378 if (page_mapping(page) == mapping)
379 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 unlock_page(page);
381}
382
Christoph Lameter04e62a22006-06-23 02:03:38 -0700383/* possible outcome of pageout() */
384typedef enum {
385 /* failed to write page out, page is locked */
386 PAGE_KEEP,
387 /* move page to the active list, page is locked */
388 PAGE_ACTIVATE,
389 /* page has been sent to the disk successfully, page is unlocked */
390 PAGE_SUCCESS,
391 /* page is clean and locked */
392 PAGE_CLEAN,
393} pageout_t;
394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395/*
Andrew Morton1742f192006-03-22 00:08:21 -0800396 * pageout is called by shrink_page_list() for each dirty page.
397 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700399static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700400 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
402 /*
403 * If the page is dirty, only perform writeback if that write
404 * will be non-blocking. To prevent this allocation from being
405 * stalled by pagecache activity. But note that there may be
406 * stalls if we need to run get_block(). We could test
407 * PagePrivate for that.
408 *
Vincent Li6aceb532009-12-14 17:58:49 -0800409 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 * this page's queue, we can perform writeback even if that
411 * will block.
412 *
413 * If the page is swapcache, write it back even if that would
414 * block, for some throttling. This happens by accident, because
415 * swap_backing_dev_info is bust: it doesn't reflect the
416 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 */
418 if (!is_page_cache_freeable(page))
419 return PAGE_KEEP;
420 if (!mapping) {
421 /*
422 * Some data journaling orphaned pages can have
423 * page->mapping == NULL while being dirty with clean buffers.
424 */
David Howells266cf652009-04-03 16:42:36 +0100425 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 if (try_to_free_buffers(page)) {
427 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700428 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 return PAGE_CLEAN;
430 }
431 }
432 return PAGE_KEEP;
433 }
434 if (mapping->a_ops->writepage == NULL)
435 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700436 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 return PAGE_KEEP;
438
439 if (clear_page_dirty_for_io(page)) {
440 int res;
441 struct writeback_control wbc = {
442 .sync_mode = WB_SYNC_NONE,
443 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700444 .range_start = 0,
445 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 .for_reclaim = 1,
447 };
448
449 SetPageReclaim(page);
450 res = mapping->a_ops->writepage(page, &wbc);
451 if (res < 0)
452 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800453 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 ClearPageReclaim(page);
455 return PAGE_ACTIVATE;
456 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700457
458 /*
459 * Wait on writeback if requested to. This happens when
460 * direct reclaiming a large contiguous area and the
461 * first attempt to free a range of pages fails.
462 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700463 if (PageWriteback(page) &&
Mel Gormanf3a310b2011-01-13 15:46:00 -0800464 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
Andy Whitcroftc661b072007-08-22 14:01:26 -0700465 wait_on_page_writeback(page);
466
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 if (!PageWriteback(page)) {
468 /* synchronous write or broken a_ops? */
469 ClearPageReclaim(page);
470 }
Mel Gorman755f0222010-08-09 17:19:18 -0700471 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800472 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700473 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 return PAGE_SUCCESS;
475 }
476
477 return PAGE_CLEAN;
478}
479
Andrew Mortona649fd92006-10-17 00:09:36 -0700480/*
Nick Piggine2867812008-07-25 19:45:30 -0700481 * Same as remove_mapping, but if the page is removed from the mapping, it
482 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700483 */
Nick Piggine2867812008-07-25 19:45:30 -0700484static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800485{
Nick Piggin28e4d962006-09-25 23:31:23 -0700486 BUG_ON(!PageLocked(page));
487 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800488
Nick Piggin19fd6232008-07-25 19:45:32 -0700489 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800490 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700491 * The non racy check for a busy page.
492 *
493 * Must be careful with the order of the tests. When someone has
494 * a ref to the page, it may be possible that they dirty it then
495 * drop the reference. So if PageDirty is tested before page_count
496 * here, then the following race may occur:
497 *
498 * get_user_pages(&page);
499 * [user mapping goes away]
500 * write_to(page);
501 * !PageDirty(page) [good]
502 * SetPageDirty(page);
503 * put_page(page);
504 * !page_count(page) [good, discard it]
505 *
506 * [oops, our write_to data is lost]
507 *
508 * Reversing the order of the tests ensures such a situation cannot
509 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
510 * load is not satisfied before that of page->_count.
511 *
512 * Note that if SetPageDirty is always performed via set_page_dirty,
513 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800514 */
Nick Piggine2867812008-07-25 19:45:30 -0700515 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800516 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700517 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
518 if (unlikely(PageDirty(page))) {
519 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800520 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700521 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800522
523 if (PageSwapCache(page)) {
524 swp_entry_t swap = { .val = page_private(page) };
525 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700526 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700527 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700528 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500529 void (*freepage)(struct page *);
530
531 freepage = mapping->a_ops->freepage;
532
Minchan Kime64a7822011-03-22 16:32:44 -0700533 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700534 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700535 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500536
537 if (freepage != NULL)
538 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800539 }
540
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800541 return 1;
542
543cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700544 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800545 return 0;
546}
547
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548/*
Nick Piggine2867812008-07-25 19:45:30 -0700549 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
550 * someone else has a ref on the page, abort and return 0. If it was
551 * successfully detached, return 1. Assumes the caller has a single ref on
552 * this page.
553 */
554int remove_mapping(struct address_space *mapping, struct page *page)
555{
556 if (__remove_mapping(mapping, page)) {
557 /*
558 * Unfreezing the refcount with 1 rather than 2 effectively
559 * drops the pagecache ref for us without requiring another
560 * atomic operation.
561 */
562 page_unfreeze_refs(page, 1);
563 return 1;
564 }
565 return 0;
566}
567
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700568/**
569 * putback_lru_page - put previously isolated page onto appropriate LRU list
570 * @page: page to be put back to appropriate lru list
571 *
572 * Add previously isolated @page to appropriate LRU list.
573 * Page may still be unevictable for other reasons.
574 *
575 * lru_lock must not be held, interrupts must be enabled.
576 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700577void putback_lru_page(struct page *page)
578{
579 int lru;
580 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700581 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700582
583 VM_BUG_ON(PageLRU(page));
584
585redo:
586 ClearPageUnevictable(page);
587
588 if (page_evictable(page, NULL)) {
589 /*
590 * For evictable pages, we can use the cache.
591 * In event of a race, worst case is we end up with an
592 * unevictable page on [in]active list.
593 * We know how to handle that.
594 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700595 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700596 lru_cache_add_lru(page, lru);
597 } else {
598 /*
599 * Put unevictable pages directly on zone's unevictable
600 * list.
601 */
602 lru = LRU_UNEVICTABLE;
603 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700604 /*
605 * When racing with an mlock clearing (page is
606 * unlocked), make sure that if the other thread does
607 * not observe our setting of PG_lru and fails
608 * isolation, we see PG_mlocked cleared below and move
609 * the page back to the evictable list.
610 *
611 * The other side is TestClearPageMlocked().
612 */
613 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700614 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700615
616 /*
617 * page's status can change while we move it among lru. If an evictable
618 * page is on unevictable list, it never be freed. To avoid that,
619 * check after we added it to the list, again.
620 */
621 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
622 if (!isolate_lru_page(page)) {
623 put_page(page);
624 goto redo;
625 }
626 /* This means someone else dropped this page from LRU
627 * So, it will be freed or putback to LRU again. There is
628 * nothing to do here.
629 */
630 }
631
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700632 if (was_unevictable && lru != LRU_UNEVICTABLE)
633 count_vm_event(UNEVICTABLE_PGRESCUED);
634 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
635 count_vm_event(UNEVICTABLE_PGCULLED);
636
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700637 put_page(page); /* drop ref from isolate */
638}
639
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800640enum page_references {
641 PAGEREF_RECLAIM,
642 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800643 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800644 PAGEREF_ACTIVATE,
645};
646
647static enum page_references page_check_references(struct page *page,
648 struct scan_control *sc)
649{
Johannes Weiner645747462010-03-05 13:42:22 -0800650 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800651 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800652
Johannes Weiner645747462010-03-05 13:42:22 -0800653 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
654 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800655
656 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800657 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800658 return PAGEREF_RECLAIM;
659
660 /*
661 * Mlock lost the isolation race with us. Let try_to_unmap()
662 * move the page to the unevictable list.
663 */
664 if (vm_flags & VM_LOCKED)
665 return PAGEREF_RECLAIM;
666
Johannes Weiner645747462010-03-05 13:42:22 -0800667 if (referenced_ptes) {
668 if (PageAnon(page))
669 return PAGEREF_ACTIVATE;
670 /*
671 * All mapped pages start out with page table
672 * references from the instantiating fault, so we need
673 * to look twice if a mapped file page is used more
674 * than once.
675 *
676 * Mark it and spare it for another trip around the
677 * inactive list. Another page table reference will
678 * lead to its activation.
679 *
680 * Note: the mark is set for activated pages as well
681 * so that recently deactivated but used pages are
682 * quickly recovered.
683 */
684 SetPageReferenced(page);
685
686 if (referenced_page)
687 return PAGEREF_ACTIVATE;
688
689 return PAGEREF_KEEP;
690 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800691
692 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700693 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800694 return PAGEREF_RECLAIM_CLEAN;
695
696 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800697}
698
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700699static noinline_for_stack void free_page_list(struct list_head *free_pages)
700{
701 struct pagevec freed_pvec;
702 struct page *page, *tmp;
703
704 pagevec_init(&freed_pvec, 1);
705
706 list_for_each_entry_safe(page, tmp, free_pages, lru) {
707 list_del(&page->lru);
708 if (!pagevec_add(&freed_pvec, page)) {
709 __pagevec_free(&freed_pvec);
710 pagevec_reinit(&freed_pvec);
711 }
712 }
713
714 pagevec_free(&freed_pvec);
715}
716
Nick Piggine2867812008-07-25 19:45:30 -0700717/*
Andrew Morton1742f192006-03-22 00:08:21 -0800718 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 */
Andrew Morton1742f192006-03-22 00:08:21 -0800720static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700721 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700722 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723{
724 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700725 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700727 unsigned long nr_dirty = 0;
728 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800729 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
731 cond_resched();
732
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800734 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 struct address_space *mapping;
736 struct page *page;
737 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738
739 cond_resched();
740
741 page = lru_to_page(page_list);
742 list_del(&page->lru);
743
Nick Piggin529ae9a2008-08-02 12:01:03 +0200744 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 goto keep;
746
Nick Piggin725d7042006-09-25 23:30:55 -0700747 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700748 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
750 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800751
Nick Pigginb291f002008-10-18 20:26:44 -0700752 if (unlikely(!page_evictable(page, NULL)))
753 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700754
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700755 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800756 goto keep_locked;
757
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 /* Double the slab pressure for mapped and swapcache pages */
759 if (page_mapped(page) || PageSwapCache(page))
760 sc->nr_scanned++;
761
Andy Whitcroftc661b072007-08-22 14:01:26 -0700762 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
763 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
764
765 if (PageWriteback(page)) {
766 /*
767 * Synchronous reclaim is performed in two passes,
768 * first an asynchronous pass over the list to
769 * start parallel writeback, and a second synchronous
770 * pass to wait for the IO to complete. Wait here
771 * for any page for which writeback has already
772 * started.
773 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800774 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700775 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700776 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700777 else {
778 unlock_page(page);
779 goto keep_lumpy;
780 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700781 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800783 references = page_check_references(page, sc);
784 switch (references) {
785 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800787 case PAGEREF_KEEP:
788 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800789 case PAGEREF_RECLAIM:
790 case PAGEREF_RECLAIM_CLEAN:
791 ; /* try to reclaim the page below */
792 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 /*
795 * Anonymous process memory has backing store?
796 * Try to allocate it some swap space here.
797 */
Nick Pigginb291f002008-10-18 20:26:44 -0700798 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800799 if (!(sc->gfp_mask & __GFP_IO))
800 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800801 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800803 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700804 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805
806 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
808 /*
809 * The page is mapped into the page tables of one or more
810 * processes. Try to unmap it here.
811 */
812 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200813 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 case SWAP_FAIL:
815 goto activate_locked;
816 case SWAP_AGAIN:
817 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700818 case SWAP_MLOCK:
819 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 case SWAP_SUCCESS:
821 ; /* try to free the page below */
822 }
823 }
824
825 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700826 nr_dirty++;
827
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800828 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700830 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800832 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 goto keep_locked;
834
835 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700836 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700838 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 goto keep_locked;
840 case PAGE_ACTIVATE:
841 goto activate_locked;
842 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700843 if (PageWriteback(page))
844 goto keep_lumpy;
845 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700847
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 /*
849 * A synchronous write - probably a ramdisk. Go
850 * ahead and try to reclaim the page.
851 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200852 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 goto keep;
854 if (PageDirty(page) || PageWriteback(page))
855 goto keep_locked;
856 mapping = page_mapping(page);
857 case PAGE_CLEAN:
858 ; /* try to free the page below */
859 }
860 }
861
862 /*
863 * If the page has buffers, try to free the buffer mappings
864 * associated with this page. If we succeed we try to free
865 * the page as well.
866 *
867 * We do this even if the page is PageDirty().
868 * try_to_release_page() does not perform I/O, but it is
869 * possible for a page to have PageDirty set, but it is actually
870 * clean (all its buffers are clean). This happens if the
871 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700872 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 * try_to_release_page() will discover that cleanness and will
874 * drop the buffers and mark the page clean - it can be freed.
875 *
876 * Rarely, pages can have buffers and no ->mapping. These are
877 * the pages which were not successfully invalidated in
878 * truncate_complete_page(). We try to drop those buffers here
879 * and if that worked, and the page is no longer mapped into
880 * process address space (page_count == 1) it can be freed.
881 * Otherwise, leave the page on the LRU so it is swappable.
882 */
David Howells266cf652009-04-03 16:42:36 +0100883 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 if (!try_to_release_page(page, sc->gfp_mask))
885 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700886 if (!mapping && page_count(page) == 1) {
887 unlock_page(page);
888 if (put_page_testzero(page))
889 goto free_it;
890 else {
891 /*
892 * rare race with speculative reference.
893 * the speculative reference will free
894 * this page shortly, so we may
895 * increment nr_reclaimed here (and
896 * leave it off the LRU).
897 */
898 nr_reclaimed++;
899 continue;
900 }
901 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903
Nick Piggine2867812008-07-25 19:45:30 -0700904 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800905 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
Nick Piggina978d6f2008-10-18 20:26:58 -0700907 /*
908 * At this point, we have no other references and there is
909 * no way to pick any more up (removed from LRU, removed
910 * from pagecache). Can use non-atomic bitops now (and
911 * we obviously don't have to worry about waking up a process
912 * waiting on the page lock, because there are no references.
913 */
914 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700915free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800916 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700917
918 /*
919 * Is there need to periodically free_page_list? It would
920 * appear not as the counts should be low
921 */
922 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 continue;
924
Nick Pigginb291f002008-10-18 20:26:44 -0700925cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800926 if (PageSwapCache(page))
927 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700928 unlock_page(page);
929 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800930 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700931 continue;
932
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700934 /* Not a candidate for swapping, so reclaim swap space. */
935 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800936 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700937 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 SetPageActive(page);
939 pgactivate++;
940keep_locked:
941 unlock_page(page);
942keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -0800943 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700944keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700946 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700948
Mel Gorman0e093d992010-10-26 14:21:45 -0700949 /*
950 * Tag a zone as congested if all the dirty pages encountered were
951 * backed by a congested BDI. In this case, reclaimers should just
952 * back off and wait for congestion to clear because further reclaim
953 * will encounter the same problem
954 */
KAMEZAWA Hiroyukid6c438b2011-05-17 15:44:10 -0700955 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
Mel Gorman0e093d992010-10-26 14:21:45 -0700956 zone_set_flag(zone, ZONE_CONGESTED);
957
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700958 free_page_list(&free_pages);
959
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700961 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800962 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963}
964
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700965/*
966 * Attempt to remove the specified page from its LRU. Only take this page
967 * if it is of the appropriate PageActive status. Pages which are being
968 * freed elsewhere are also ignored.
969 *
970 * page: page to consider
971 * mode: one of the LRU isolation modes defined above
972 *
973 * returns 0 on success, -ve errno on failure.
974 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700975int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700976{
977 int ret = -EINVAL;
978
979 /* Only take pages on the LRU. */
980 if (!PageLRU(page))
981 return ret;
982
983 /*
984 * When checking the active state, we need to be sure we are
985 * dealing with comparible boolean values. Take the logical not
986 * of each.
987 */
988 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
989 return ret;
990
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700991 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700992 return ret;
993
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700994 /*
995 * When this function is being called for lumpy reclaim, we
996 * initially look into all LRU pages, active, inactive and
997 * unevictable; only give shrink_page_list evictable pages.
998 */
999 if (PageUnevictable(page))
1000 return ret;
1001
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001002 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001003
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001004 if (likely(get_page_unless_zero(page))) {
1005 /*
1006 * Be careful not to clear PageLRU until after we're
1007 * sure the page is not being freed elsewhere -- the
1008 * page release code relies on it.
1009 */
1010 ClearPageLRU(page);
1011 ret = 0;
1012 }
1013
1014 return ret;
1015}
1016
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001017/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 * zone->lru_lock is heavily contended. Some of the functions that
1019 * shrink the lists perform better by taking out a batch of pages
1020 * and working on them outside the LRU lock.
1021 *
1022 * For pagecache intensive workloads, this function is the hottest
1023 * spot in the kernel (apart from copy_*_user functions).
1024 *
1025 * Appropriate locks must be held before calling this function.
1026 *
1027 * @nr_to_scan: The number of pages to look through on the list.
1028 * @src: The LRU list to pull pages off.
1029 * @dst: The temp list to put pages on to.
1030 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001031 * @order: The caller's attempted allocation order
1032 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001033 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 *
1035 * returns how many pages were moved onto *@dst.
1036 */
Andrew Morton69e05942006-03-22 00:08:19 -08001037static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1038 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001039 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040{
Andrew Morton69e05942006-03-22 00:08:19 -08001041 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001042 unsigned long nr_lumpy_taken = 0;
1043 unsigned long nr_lumpy_dirty = 0;
1044 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001045 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001047 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001048 struct page *page;
1049 unsigned long pfn;
1050 unsigned long end_pfn;
1051 unsigned long page_pfn;
1052 int zone_id;
1053
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 page = lru_to_page(src);
1055 prefetchw_prev_lru_page(page, src, flags);
1056
Nick Piggin725d7042006-09-25 23:30:55 -07001057 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001058
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001059 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001060 case 0:
1061 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001062 mem_cgroup_del_lru(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001063 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001064 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001065
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001066 case -EBUSY:
1067 /* else it is being freed elsewhere */
1068 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001069 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001070 continue;
1071
1072 default:
1073 BUG();
1074 }
1075
1076 if (!order)
1077 continue;
1078
1079 /*
1080 * Attempt to take all pages in the order aligned region
1081 * surrounding the tag page. Only take those pages of
1082 * the same active state as that tag page. We may safely
1083 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001084 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001085 * where that page is in a different zone we will detect
1086 * it from its zone id and abort this block scan.
1087 */
1088 zone_id = page_zone_id(page);
1089 page_pfn = page_to_pfn(page);
1090 pfn = page_pfn & ~((1 << order) - 1);
1091 end_pfn = pfn + (1 << order);
1092 for (; pfn < end_pfn; pfn++) {
1093 struct page *cursor_page;
1094
1095 /* The target page is in the block, ignore it. */
1096 if (unlikely(pfn == page_pfn))
1097 continue;
1098
1099 /* Avoid holes within the zone. */
1100 if (unlikely(!pfn_valid_within(pfn)))
1101 break;
1102
1103 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001104
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001105 /* Check that we have not crossed a zone boundary. */
1106 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001107 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001108
1109 /*
1110 * If we don't have enough swap space, reclaiming of
1111 * anon page which don't already have a swap slot is
1112 * pointless.
1113 */
1114 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001115 !PageSwapCache(cursor_page))
1116 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001117
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001118 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001119 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001120 mem_cgroup_del_lru(cursor_page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001121 nr_taken += hpage_nr_pages(page);
Mel Gormana8a94d12010-08-09 17:19:17 -07001122 nr_lumpy_taken++;
1123 if (PageDirty(cursor_page))
1124 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001125 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001126 } else {
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001127 /* the page is freed already. */
1128 if (!page_count(cursor_page))
1129 continue;
1130 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001131 }
1132 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001133
1134 /* If we break out of the loop above, lumpy reclaim failed */
1135 if (pfn < end_pfn)
1136 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 }
1138
1139 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001140
1141 trace_mm_vmscan_lru_isolate(order,
1142 nr_to_scan, scan,
1143 nr_taken,
1144 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1145 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 return nr_taken;
1147}
1148
Balbir Singh66e17072008-02-07 00:13:56 -08001149static unsigned long isolate_pages_global(unsigned long nr,
1150 struct list_head *dst,
1151 unsigned long *scanned, int order,
1152 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001153 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001154{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001155 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001156 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001157 lru += LRU_ACTIVE;
1158 if (file)
1159 lru += LRU_FILE;
1160 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001161 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001162}
1163
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001165 * clear_active_flags() is a helper for shrink_active_list(), clearing
1166 * any active bits from the pages in the list.
1167 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001168static unsigned long clear_active_flags(struct list_head *page_list,
1169 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001170{
1171 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001172 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001173 struct page *page;
1174
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001175 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001176 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001177 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001178 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001179 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001180 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001181 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001182 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001183 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001184 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001185 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001186
1187 return nr_active;
1188}
1189
Nick Piggin62695a82008-10-18 20:26:09 -07001190/**
1191 * isolate_lru_page - tries to isolate a page from its LRU list
1192 * @page: page to isolate from its LRU list
1193 *
1194 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1195 * vmstat statistic corresponding to whatever LRU list the page was on.
1196 *
1197 * Returns 0 if the page was removed from an LRU list.
1198 * Returns -EBUSY if the page was not on an LRU list.
1199 *
1200 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001201 * the active list, it will have PageActive set. If it was found on
1202 * the unevictable list, it will have the PageUnevictable bit set. That flag
1203 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001204 *
1205 * The vmstat statistic corresponding to the list on which the page was
1206 * found will be decremented.
1207 *
1208 * Restrictions:
1209 * (1) Must be called with an elevated refcount on the page. This is a
1210 * fundamentnal difference from isolate_lru_pages (which is called
1211 * without a stable reference).
1212 * (2) the lru_lock must not be held.
1213 * (3) interrupts must be enabled.
1214 */
1215int isolate_lru_page(struct page *page)
1216{
1217 int ret = -EBUSY;
1218
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001219 VM_BUG_ON(!page_count(page));
1220
Nick Piggin62695a82008-10-18 20:26:09 -07001221 if (PageLRU(page)) {
1222 struct zone *zone = page_zone(page);
1223
1224 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001225 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001226 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001227 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001228 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001229 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001230
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001231 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001232 }
1233 spin_unlock_irq(&zone->lru_lock);
1234 }
1235 return ret;
1236}
1237
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001238/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001239 * Are there way too many processes in the direct reclaim path already?
1240 */
1241static int too_many_isolated(struct zone *zone, int file,
1242 struct scan_control *sc)
1243{
1244 unsigned long inactive, isolated;
1245
1246 if (current_is_kswapd())
1247 return 0;
1248
1249 if (!scanning_global_lru(sc))
1250 return 0;
1251
1252 if (file) {
1253 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1254 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1255 } else {
1256 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1257 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1258 }
1259
1260 return isolated > inactive;
1261}
1262
1263/*
Mel Gorman66635622010-08-09 17:19:30 -07001264 * TODO: Try merging with migrations version of putback_lru_pages
1265 */
1266static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001267putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001268 unsigned long nr_anon, unsigned long nr_file,
1269 struct list_head *page_list)
1270{
1271 struct page *page;
1272 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001273 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001274
1275 pagevec_init(&pvec, 1);
1276
1277 /*
1278 * Put back any unfreeable pages.
1279 */
1280 spin_lock(&zone->lru_lock);
1281 while (!list_empty(page_list)) {
1282 int lru;
1283 page = lru_to_page(page_list);
1284 VM_BUG_ON(PageLRU(page));
1285 list_del(&page->lru);
1286 if (unlikely(!page_evictable(page, NULL))) {
1287 spin_unlock_irq(&zone->lru_lock);
1288 putback_lru_page(page);
1289 spin_lock_irq(&zone->lru_lock);
1290 continue;
1291 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001292 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001293 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001294 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001295 if (is_active_lru(lru)) {
1296 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001297 int numpages = hpage_nr_pages(page);
1298 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001299 }
1300 if (!pagevec_add(&pvec, page)) {
1301 spin_unlock_irq(&zone->lru_lock);
1302 __pagevec_release(&pvec);
1303 spin_lock_irq(&zone->lru_lock);
1304 }
1305 }
1306 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1307 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1308
1309 spin_unlock_irq(&zone->lru_lock);
1310 pagevec_release(&pvec);
1311}
1312
Mel Gorman1489fa12010-08-09 17:19:33 -07001313static noinline_for_stack void update_isolated_counts(struct zone *zone,
1314 struct scan_control *sc,
1315 unsigned long *nr_anon,
1316 unsigned long *nr_file,
1317 struct list_head *isolated_list)
1318{
1319 unsigned long nr_active;
1320 unsigned int count[NR_LRU_LISTS] = { 0, };
1321 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1322
1323 nr_active = clear_active_flags(isolated_list, count);
1324 __count_vm_events(PGDEACTIVATE, nr_active);
1325
1326 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1327 -count[LRU_ACTIVE_FILE]);
1328 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1329 -count[LRU_INACTIVE_FILE]);
1330 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1331 -count[LRU_ACTIVE_ANON]);
1332 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1333 -count[LRU_INACTIVE_ANON]);
1334
1335 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1336 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1337 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1338 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1339
1340 reclaim_stat->recent_scanned[0] += *nr_anon;
1341 reclaim_stat->recent_scanned[1] += *nr_file;
1342}
1343
Mel Gorman66635622010-08-09 17:19:30 -07001344/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001345 * Returns true if the caller should wait to clean dirty/writeback pages.
1346 *
1347 * If we are direct reclaiming for contiguous pages and we do not reclaim
1348 * everything in the list, try again and wait for writeback IO to complete.
1349 * This will stall high-order allocations noticeably. Only do that when really
1350 * need to free the pages under high memory pressure.
1351 */
1352static inline bool should_reclaim_stall(unsigned long nr_taken,
1353 unsigned long nr_freed,
1354 int priority,
1355 struct scan_control *sc)
1356{
1357 int lumpy_stall_priority;
1358
1359 /* kswapd should not stall on sync IO */
1360 if (current_is_kswapd())
1361 return false;
1362
1363 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001364 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001365 return false;
1366
1367 /* If we have relaimed everything on the isolated list, no stall */
1368 if (nr_freed == nr_taken)
1369 return false;
1370
1371 /*
1372 * For high-order allocations, there are two stall thresholds.
1373 * High-cost allocations stall immediately where as lower
1374 * order allocations such as stacks require the scanning
1375 * priority to be much higher before stalling.
1376 */
1377 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1378 lumpy_stall_priority = DEF_PRIORITY;
1379 else
1380 lumpy_stall_priority = DEF_PRIORITY / 3;
1381
1382 return priority <= lumpy_stall_priority;
1383}
1384
1385/*
Andrew Morton1742f192006-03-22 00:08:21 -08001386 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1387 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 */
Mel Gorman66635622010-08-09 17:19:30 -07001389static noinline_for_stack unsigned long
1390shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1391 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392{
1393 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001394 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001395 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001396 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001397 unsigned long nr_anon;
1398 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001399
Rik van Riel35cd7812009-09-21 17:01:38 -07001400 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001401 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001402
1403 /* We are about to die and free our memory. Return now. */
1404 if (fatal_signal_pending(current))
1405 return SWAP_CLUSTER_MAX;
1406 }
1407
Mel Gormanf3a310b2011-01-13 15:46:00 -08001408 set_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 lru_add_drain();
1410 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001412 if (scanning_global_lru(sc)) {
1413 nr_taken = isolate_pages_global(nr_to_scan,
1414 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001415 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001416 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001417 zone, 0, file);
1418 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001419 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001420 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1421 nr_scanned);
1422 else
1423 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1424 nr_scanned);
1425 } else {
1426 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1427 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001428 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001429 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001430 zone, sc->mem_cgroup,
1431 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001433 * mem_cgroup_isolate_pages() keeps track of
1434 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001436 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001437
Mel Gorman66635622010-08-09 17:19:30 -07001438 if (nr_taken == 0) {
1439 spin_unlock_irq(&zone->lru_lock);
1440 return 0;
1441 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001442
Mel Gorman1489fa12010-08-09 17:19:33 -07001443 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001444
1445 spin_unlock_irq(&zone->lru_lock);
1446
Mel Gorman0e093d992010-10-26 14:21:45 -07001447 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001448
Wu Fengguange31f36982010-08-09 17:20:01 -07001449 /* Check if we should syncronously wait for writeback */
1450 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001451 set_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001452 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001453 }
1454
1455 local_irq_disable();
1456 if (current_is_kswapd())
1457 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1458 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1459
Mel Gorman1489fa12010-08-09 17:19:33 -07001460 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001461
1462 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1463 zone_idx(zone),
1464 nr_scanned, nr_reclaimed,
1465 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001466 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001467 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468}
1469
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001470/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 * This moves pages from the active list to the inactive list.
1472 *
1473 * We move them the other way if the page is referenced by one or more
1474 * processes, from rmap.
1475 *
1476 * If the pages are mostly unmapped, the processing is fast and it is
1477 * appropriate to hold zone->lru_lock across the whole operation. But if
1478 * the pages are mapped, the processing is slow (page_referenced()) so we
1479 * should drop zone->lru_lock around each page. It's impossible to balance
1480 * this, so instead we remove the pages from the LRU while processing them.
1481 * It is safe to rely on PG_active against the non-LRU pages in here because
1482 * nobody will play with that bit on a non-LRU page.
1483 *
1484 * The downside is that we have to touch page->_count against each page.
1485 * But we had to alter page->flags anyway.
1486 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001487
Wu Fengguang3eb41402009-06-16 15:33:13 -07001488static void move_active_pages_to_lru(struct zone *zone,
1489 struct list_head *list,
1490 enum lru_list lru)
1491{
1492 unsigned long pgmoved = 0;
1493 struct pagevec pvec;
1494 struct page *page;
1495
1496 pagevec_init(&pvec, 1);
1497
1498 while (!list_empty(list)) {
1499 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001500
1501 VM_BUG_ON(PageLRU(page));
1502 SetPageLRU(page);
1503
Wu Fengguang3eb41402009-06-16 15:33:13 -07001504 list_move(&page->lru, &zone->lru[lru].list);
1505 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001506 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001507
1508 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1509 spin_unlock_irq(&zone->lru_lock);
1510 if (buffer_heads_over_limit)
1511 pagevec_strip(&pvec);
1512 __pagevec_release(&pvec);
1513 spin_lock_irq(&zone->lru_lock);
1514 }
1515 }
1516 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1517 if (!is_active_lru(lru))
1518 __count_vm_events(PGDEACTIVATE, pgmoved);
1519}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001520
Andrew Morton1742f192006-03-22 00:08:21 -08001521static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001522 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001524 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001525 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001526 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001528 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001529 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001531 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001532 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533
1534 lru_add_drain();
1535 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001536 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001537 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1538 &pgscanned, sc->order,
1539 ISOLATE_ACTIVE, zone,
1540 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001541 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001542 } else {
1543 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1544 &pgscanned, sc->order,
1545 ISOLATE_ACTIVE, zone,
1546 sc->mem_cgroup, 1, file);
1547 /*
1548 * mem_cgroup_isolate_pages() keeps track of
1549 * scanned pages on its own.
1550 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001551 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001552
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001553 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001554
Wu Fengguang3eb41402009-06-16 15:33:13 -07001555 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001556 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001557 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001558 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001559 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001560 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 spin_unlock_irq(&zone->lru_lock);
1562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 while (!list_empty(&l_hold)) {
1564 cond_resched();
1565 page = lru_to_page(&l_hold);
1566 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001567
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001568 if (unlikely(!page_evictable(page, NULL))) {
1569 putback_lru_page(page);
1570 continue;
1571 }
1572
Johannes Weiner645747462010-03-05 13:42:22 -08001573 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001574 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001575 /*
1576 * Identify referenced, file-backed active pages and
1577 * give them one more trip around the active list. So
1578 * that executable code get better chances to stay in
1579 * memory under moderate memory pressure. Anon pages
1580 * are not likely to be evicted by use-once streaming
1581 * IO, plus JVM can create lots of anon VM_EXEC pages,
1582 * so we ignore them here.
1583 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001584 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001585 list_add(&page->lru, &l_active);
1586 continue;
1587 }
1588 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001589
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001590 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 list_add(&page->lru, &l_inactive);
1592 }
1593
Andrew Mortonb5557492009-01-06 14:40:13 -08001594 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001595 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001596 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001597 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001598 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001599 * Count referenced pages from currently used mappings as rotated,
1600 * even though only some of them are actually re-activated. This
1601 * helps balance scan pressure between file and anonymous pages in
1602 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001603 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001604 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001605
Wu Fengguang3eb41402009-06-16 15:33:13 -07001606 move_active_pages_to_lru(zone, &l_active,
1607 LRU_ACTIVE + file * LRU_FILE);
1608 move_active_pages_to_lru(zone, &l_inactive,
1609 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001610 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001611 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612}
1613
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001614#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001615static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001616{
1617 unsigned long active, inactive;
1618
1619 active = zone_page_state(zone, NR_ACTIVE_ANON);
1620 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1621
1622 if (inactive * zone->inactive_ratio < active)
1623 return 1;
1624
1625 return 0;
1626}
1627
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001628/**
1629 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1630 * @zone: zone to check
1631 * @sc: scan control of this context
1632 *
1633 * Returns true if the zone does not have enough inactive anon pages,
1634 * meaning some active anon pages need to be deactivated.
1635 */
1636static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1637{
1638 int low;
1639
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001640 /*
1641 * If we don't have swap space, anonymous page deactivation
1642 * is pointless.
1643 */
1644 if (!total_swap_pages)
1645 return 0;
1646
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001647 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001648 low = inactive_anon_is_low_global(zone);
1649 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001650 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001651 return low;
1652}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001653#else
1654static inline int inactive_anon_is_low(struct zone *zone,
1655 struct scan_control *sc)
1656{
1657 return 0;
1658}
1659#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001660
Rik van Riel56e49d22009-06-16 15:32:28 -07001661static int inactive_file_is_low_global(struct zone *zone)
1662{
1663 unsigned long active, inactive;
1664
1665 active = zone_page_state(zone, NR_ACTIVE_FILE);
1666 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1667
1668 return (active > inactive);
1669}
1670
1671/**
1672 * inactive_file_is_low - check if file pages need to be deactivated
1673 * @zone: zone to check
1674 * @sc: scan control of this context
1675 *
1676 * When the system is doing streaming IO, memory pressure here
1677 * ensures that active file pages get deactivated, until more
1678 * than half of the file pages are on the inactive list.
1679 *
1680 * Once we get to that situation, protect the system's working
1681 * set from being evicted by disabling active file page aging.
1682 *
1683 * This uses a different ratio than the anonymous pages, because
1684 * the page cache uses a use-once replacement algorithm.
1685 */
1686static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1687{
1688 int low;
1689
1690 if (scanning_global_lru(sc))
1691 low = inactive_file_is_low_global(zone);
1692 else
1693 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1694 return low;
1695}
1696
Rik van Rielb39415b2009-12-14 17:59:48 -08001697static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1698 int file)
1699{
1700 if (file)
1701 return inactive_file_is_low(zone, sc);
1702 else
1703 return inactive_anon_is_low(zone, sc);
1704}
1705
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001706static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001707 struct zone *zone, struct scan_control *sc, int priority)
1708{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001709 int file = is_file_lru(lru);
1710
Rik van Rielb39415b2009-12-14 17:59:48 -08001711 if (is_active_lru(lru)) {
1712 if (inactive_list_is_low(zone, sc, file))
1713 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001714 return 0;
1715 }
1716
Rik van Riel33c120e2008-10-18 20:26:36 -07001717 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001718}
1719
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001721 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1722 * until we collected @swap_cluster_max pages to scan.
1723 */
1724static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1725 unsigned long *nr_saved_scan)
1726{
1727 unsigned long nr;
1728
1729 *nr_saved_scan += nr_to_scan;
1730 nr = *nr_saved_scan;
1731
1732 if (nr >= SWAP_CLUSTER_MAX)
1733 *nr_saved_scan = 0;
1734 else
1735 nr = 0;
1736
1737 return nr;
1738}
1739
1740/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001741 * Determine how aggressively the anon and file LRU lists should be
1742 * scanned. The relative value of each set of LRU lists is determined
1743 * by looking at the fraction of the pages scanned we did rotate back
1744 * onto the active list instead of evict.
1745 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001746 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001747 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001748static void get_scan_count(struct zone *zone, struct scan_control *sc,
1749 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001750{
1751 unsigned long anon, file, free;
1752 unsigned long anon_prio, file_prio;
1753 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001754 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001755 u64 fraction[2], denominator;
1756 enum lru_list l;
1757 int noswap = 0;
1758
1759 /* If we have no swap space, do not bother scanning anon pages. */
1760 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1761 noswap = 1;
1762 fraction[0] = 0;
1763 fraction[1] = 1;
1764 denominator = 1;
1765 goto out;
1766 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001767
Vincent Li0b217672009-09-21 17:03:09 -07001768 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1769 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1770 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1771 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001772
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001773 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001774 free = zone_page_state(zone, NR_FREE_PAGES);
1775 /* If we have very few page cache pages,
1776 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001777 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001778 fraction[0] = 1;
1779 fraction[1] = 0;
1780 denominator = 1;
1781 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001782 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001783 }
1784
1785 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001786 * With swappiness at 100, anonymous and file have the same priority.
1787 * This scanning priority is essentially the inverse of IO cost.
1788 */
1789 anon_prio = sc->swappiness;
1790 file_prio = 200 - sc->swappiness;
1791
1792 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001793 * OK, so we have swap space and a fair amount of page cache
1794 * pages. We use the recently rotated / recently scanned
1795 * ratios to determine how valuable each cache is.
1796 *
1797 * Because workloads change over time (and to avoid overflow)
1798 * we keep these statistics as a floating average, which ends
1799 * up weighing recent references more than old ones.
1800 *
1801 * anon in [0], file in [1]
1802 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001803 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001804 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001805 reclaim_stat->recent_scanned[0] /= 2;
1806 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001807 }
1808
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001809 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001810 reclaim_stat->recent_scanned[1] /= 2;
1811 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001812 }
1813
1814 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001815 * The amount of pressure on anon vs file pages is inversely
1816 * proportional to the fraction of recently scanned pages on
1817 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001818 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001819 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1820 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001821
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001822 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1823 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001824 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001825
Shaohua Li76a33fc2010-05-24 14:32:36 -07001826 fraction[0] = ap;
1827 fraction[1] = fp;
1828 denominator = ap + fp + 1;
1829out:
1830 for_each_evictable_lru(l) {
1831 int file = is_file_lru(l);
1832 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001833
Shaohua Li76a33fc2010-05-24 14:32:36 -07001834 scan = zone_nr_lru_pages(zone, sc, l);
1835 if (priority || noswap) {
1836 scan >>= priority;
1837 scan = div64_u64(scan * fraction[file], denominator);
1838 }
1839 nr[l] = nr_scan_try_batch(scan,
1840 &reclaim_stat->nr_saved_scan[l]);
1841 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001842}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001843
1844/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001845 * Reclaim/compaction depends on a number of pages being freed. To avoid
1846 * disruption to the system, a small number of order-0 pages continue to be
1847 * rotated and reclaimed in the normal fashion. However, by the time we get
1848 * back to the allocator and call try_to_compact_zone(), we ensure that
1849 * there are enough free pages for it to be likely successful
1850 */
1851static inline bool should_continue_reclaim(struct zone *zone,
1852 unsigned long nr_reclaimed,
1853 unsigned long nr_scanned,
1854 struct scan_control *sc)
1855{
1856 unsigned long pages_for_compaction;
1857 unsigned long inactive_lru_pages;
1858
1859 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001860 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001861 return false;
1862
Mel Gorman28765922011-02-25 14:44:20 -08001863 /* Consider stopping depending on scan and reclaim activity */
1864 if (sc->gfp_mask & __GFP_REPEAT) {
1865 /*
1866 * For __GFP_REPEAT allocations, stop reclaiming if the
1867 * full LRU list has been scanned and we are still failing
1868 * to reclaim pages. This full LRU scan is potentially
1869 * expensive but a __GFP_REPEAT caller really wants to succeed
1870 */
1871 if (!nr_reclaimed && !nr_scanned)
1872 return false;
1873 } else {
1874 /*
1875 * For non-__GFP_REPEAT allocations which can presumably
1876 * fail without consequence, stop if we failed to reclaim
1877 * any pages from the last SWAP_CLUSTER_MAX number of
1878 * pages that were scanned. This will return to the
1879 * caller faster at the risk reclaim/compaction and
1880 * the resulting allocation attempt fails
1881 */
1882 if (!nr_reclaimed)
1883 return false;
1884 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001885
1886 /*
1887 * If we have not reclaimed enough pages for compaction and the
1888 * inactive lists are large enough, continue reclaiming
1889 */
1890 pages_for_compaction = (2UL << sc->order);
1891 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1892 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1893 if (sc->nr_reclaimed < pages_for_compaction &&
1894 inactive_lru_pages > pages_for_compaction)
1895 return true;
1896
1897 /* If compaction would go ahead or the allocation would succeed, stop */
1898 switch (compaction_suitable(zone, sc->order)) {
1899 case COMPACT_PARTIAL:
1900 case COMPACT_CONTINUE:
1901 return false;
1902 default:
1903 return true;
1904 }
1905}
1906
1907/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1909 */
Rik van Riela79311c2009-01-06 14:40:01 -08001910static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001911 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001913 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001914 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001915 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001916 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001917 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
Mel Gorman3e7d3442011-01-13 15:45:56 -08001919restart:
1920 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001921 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001922 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001923
Rik van Riel556adec2008-10-18 20:26:34 -07001924 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1925 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001926 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001927 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001928 nr_to_scan = min_t(unsigned long,
1929 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001930 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001932 nr_reclaimed += shrink_list(l, nr_to_scan,
1933 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001934 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 }
Rik van Riela79311c2009-01-06 14:40:01 -08001936 /*
1937 * On large memory systems, scan >> priority can become
1938 * really large. This is fine for the starting priority;
1939 * we want to put equal scanning pressure on each zone.
1940 * However, if the VM has a harder time of freeing pages,
1941 * with multiple processes reclaiming pages, the total
1942 * freeing target can get unreasonably large.
1943 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001944 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001945 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001947 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001948
Rik van Riel556adec2008-10-18 20:26:34 -07001949 /*
1950 * Even if we did not try to evict anon pages at all, we want to
1951 * rebalance the anon lru active/inactive ratio.
1952 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001953 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001954 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1955
Mel Gorman3e7d3442011-01-13 15:45:56 -08001956 /* reclaim/compaction might need reclaim to continue */
1957 if (should_continue_reclaim(zone, nr_reclaimed,
1958 sc->nr_scanned - nr_scanned, sc))
1959 goto restart;
1960
Andrew Morton232ea4d2007-02-28 20:13:21 -08001961 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962}
1963
1964/*
1965 * This is the direct reclaim path, for page-allocating processes. We only
1966 * try to reclaim pages from zones which will satisfy the caller's allocation
1967 * request.
1968 *
Mel Gorman41858962009-06-16 15:32:12 -07001969 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1970 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1972 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001973 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1974 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1975 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 * If a zone is deemed to be full of pinned pages then just give it a light
1978 * scan then give up on it.
1979 */
Minchan Kimd1908362010-09-22 13:05:01 -07001980static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001981 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982{
Mel Gormandd1a2392008-04-28 02:12:17 -07001983 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001984 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001985
Mel Gormand4debc62010-08-09 17:19:29 -07001986 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1987 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001988 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001990 /*
1991 * Take care memory controller reclaiming has small influence
1992 * to global LRU.
1993 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001994 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001995 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1996 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001997 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001998 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001999 }
Nick Piggin408d8542006-09-25 23:31:27 -07002000
Rik van Riela79311c2009-01-06 14:40:01 -08002001 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 }
Minchan Kimd1908362010-09-22 13:05:01 -07002003}
2004
2005static bool zone_reclaimable(struct zone *zone)
2006{
2007 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2008}
2009
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002010/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002011static bool all_unreclaimable(struct zonelist *zonelist,
2012 struct scan_control *sc)
2013{
2014 struct zoneref *z;
2015 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002016
2017 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2018 gfp_zone(sc->gfp_mask), sc->nodemask) {
2019 if (!populated_zone(zone))
2020 continue;
2021 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2022 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002023 if (!zone->all_unreclaimable)
2024 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002025 }
2026
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002027 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030/*
2031 * This is the main entry point to direct page reclaim.
2032 *
2033 * If a full scan of the inactive list fails to free enough memory then we
2034 * are "out of memory" and something needs to be killed.
2035 *
2036 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2037 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002038 * caller can't do much about. We kick the writeback threads and take explicit
2039 * naps in the hope that some of these pages can be written. But if the
2040 * allocating task holds filesystem locks which prevent writeout this might not
2041 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002042 *
2043 * returns: 0, if no pages reclaimed
2044 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 */
Mel Gormandac1d272008-04-28 02:12:12 -07002046static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002047 struct scan_control *sc,
2048 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049{
2050 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002051 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002053 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002054 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002055 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056
Miao Xiec0ff7452010-05-24 14:32:08 -07002057 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002058 delayacct_freepages_start();
2059
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002060 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002061 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062
2063 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002064 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002065 if (!priority)
2066 disable_swap_token();
Minchan Kimd1908362010-09-22 13:05:01 -07002067 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08002068 /*
2069 * Don't shrink slabs when reclaiming memory from
2070 * over limit cgroups
2071 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002072 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002073 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002074 for_each_zone_zonelist(zone, z, zonelist,
2075 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002076 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2077 continue;
2078
2079 lru_pages += zone_reclaimable_pages(zone);
2080 }
2081
Ying Han1495f232011-05-24 17:12:27 -07002082 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002083 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002084 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002085 reclaim_state->reclaimed_slab = 0;
2086 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 }
Balbir Singh66e17072008-02-07 00:13:56 -08002088 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002089 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091
2092 /*
2093 * Try to write back as many pages as we just scanned. This
2094 * tends to cause slow streaming writers to write data to the
2095 * disk smoothly, at the dirtying rate, which is nice. But
2096 * that's undesirable in laptop mode, where we *want* lumpy
2097 * writeout. So in laptop mode, write out the whole world.
2098 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002099 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2100 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002101 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002102 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 }
2104
2105 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002106 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002107 priority < DEF_PRIORITY - 2) {
2108 struct zone *preferred_zone;
2109
2110 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002111 &cpuset_current_mems_allowed,
2112 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002113 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2114 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002116
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002118 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002119 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002120
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002121 if (sc->nr_reclaimed)
2122 return sc->nr_reclaimed;
2123
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002124 /*
2125 * As hibernation is going on, kswapd is freezed so that it can't mark
2126 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2127 * check.
2128 */
2129 if (oom_killer_disabled)
2130 return 0;
2131
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002132 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002133 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002134 return 1;
2135
2136 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137}
2138
Mel Gormandac1d272008-04-28 02:12:12 -07002139unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002140 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002141{
Mel Gorman33906bc2010-08-09 17:19:16 -07002142 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002143 struct scan_control sc = {
2144 .gfp_mask = gfp_mask,
2145 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002146 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002147 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002148 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002149 .swappiness = vm_swappiness,
2150 .order = order,
2151 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002152 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002153 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002154 struct shrink_control shrink = {
2155 .gfp_mask = sc.gfp_mask,
2156 };
Balbir Singh66e17072008-02-07 00:13:56 -08002157
Mel Gorman33906bc2010-08-09 17:19:16 -07002158 trace_mm_vmscan_direct_reclaim_begin(order,
2159 sc.may_writepage,
2160 gfp_mask);
2161
Ying Hana09ed5e2011-05-24 17:12:26 -07002162 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002163
2164 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2165
2166 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002167}
2168
Balbir Singh00f0b822008-03-04 14:28:39 -08002169#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002170
Balbir Singh4e416952009-09-23 15:56:39 -07002171unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2172 gfp_t gfp_mask, bool noswap,
2173 unsigned int swappiness,
KOSAKI Motohiro14fec792010-08-10 18:03:05 -07002174 struct zone *zone)
Balbir Singh4e416952009-09-23 15:56:39 -07002175{
2176 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002177 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002178 .may_writepage = !laptop_mode,
2179 .may_unmap = 1,
2180 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002181 .swappiness = swappiness,
2182 .order = 0,
2183 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002184 };
Balbir Singh4e416952009-09-23 15:56:39 -07002185 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2186 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002187
2188 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2189 sc.may_writepage,
2190 sc.gfp_mask);
2191
Balbir Singh4e416952009-09-23 15:56:39 -07002192 /*
2193 * NOTE: Although we can get the priority field, using it
2194 * here is not a good idea, since it limits the pages we can scan.
2195 * if we don't reclaim here, the shrink_zone from balance_pgdat
2196 * will pick up pages from other mem cgroup's as well. We hack
2197 * the priority and make it zero.
2198 */
2199 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002200
2201 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2202
Balbir Singh4e416952009-09-23 15:56:39 -07002203 return sc.nr_reclaimed;
2204}
2205
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002206unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002207 gfp_t gfp_mask,
2208 bool noswap,
2209 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002210{
Balbir Singh4e416952009-09-23 15:56:39 -07002211 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002212 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002213 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002214 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002215 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002216 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002217 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002218 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002219 .order = 0,
2220 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002221 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002222 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2223 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2224 };
2225 struct shrink_control shrink = {
2226 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002227 };
Balbir Singh66e17072008-02-07 00:13:56 -08002228
Mel Gormandd1a2392008-04-28 02:12:17 -07002229 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002230
2231 trace_mm_vmscan_memcg_reclaim_begin(0,
2232 sc.may_writepage,
2233 sc.gfp_mask);
2234
Ying Hana09ed5e2011-05-24 17:12:26 -07002235 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002236
2237 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2238
2239 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002240}
2241#endif
2242
Mel Gorman1741c872011-01-13 15:46:21 -08002243/*
2244 * pgdat_balanced is used when checking if a node is balanced for high-order
2245 * allocations. Only zones that meet watermarks and are in a zone allowed
2246 * by the callers classzone_idx are added to balanced_pages. The total of
2247 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2248 * for the node to be considered balanced. Forcing all zones to be balanced
2249 * for high orders can cause excessive reclaim when there are imbalanced zones.
2250 * The choice of 25% is due to
2251 * o a 16M DMA zone that is balanced will not balance a zone on any
2252 * reasonable sized machine
2253 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002254 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002255 * would need to be at least 256M for it to be balance a whole node.
2256 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2257 * to balance a node on its own. These seemed like reasonable ratios.
2258 */
2259static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2260 int classzone_idx)
2261{
2262 unsigned long present_pages = 0;
2263 int i;
2264
2265 for (i = 0; i <= classzone_idx; i++)
2266 present_pages += pgdat->node_zones[i].present_pages;
2267
2268 return balanced_pages > (present_pages >> 2);
2269}
2270
Mel Gormanf50de2d2009-12-14 17:58:53 -08002271/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002272static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2273 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002274{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002275 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002276 unsigned long balanced = 0;
2277 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002278
2279 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2280 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002281 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002282
Mel Gorman0abdee22011-01-13 15:46:22 -08002283 /* Check the watermark levels */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002284 for (i = 0; i < pgdat->nr_zones; i++) {
2285 struct zone *zone = pgdat->node_zones + i;
2286
2287 if (!populated_zone(zone))
2288 continue;
2289
Mel Gorman355b09c2011-01-13 15:46:24 -08002290 /*
2291 * balance_pgdat() skips over all_unreclaimable after
2292 * DEF_PRIORITY. Effectively, it considers them balanced so
2293 * they must be considered balanced here as well if kswapd
2294 * is to sleep
2295 */
2296 if (zone->all_unreclaimable) {
2297 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002298 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002299 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002300
Mel Gorman88f5acf2011-01-13 15:45:41 -08002301 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormandc83edd2011-01-13 15:46:26 -08002302 classzone_idx, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002303 all_zones_ok = false;
2304 else
2305 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002306 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002307
Mel Gorman1741c872011-01-13 15:46:21 -08002308 /*
2309 * For high-order requests, the balanced zones must contain at least
2310 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2311 * must be balanced
2312 */
2313 if (order)
Johannes Weinerafc7e3262011-05-24 17:11:09 -07002314 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002315 else
2316 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002317}
2318
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319/*
2320 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002321 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002323 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 *
2325 * There is special handling here for zones which are full of pinned pages.
2326 * This can happen if the pages are all mlocked, or if they are all used by
2327 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2328 * What we do is to detect the case where all pages in the zone have been
2329 * scanned twice and there has been zero successful reclaim. Mark the zone as
2330 * dead and from now on, only perform a short scan. Basically we're polling
2331 * the zone for when the problem goes away.
2332 *
2333 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002334 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2335 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2336 * lower zones regardless of the number of free pages in the lower zones. This
2337 * interoperates with the page allocator fallback scheme to ensure that aging
2338 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 */
Mel Gorman99504742011-01-13 15:46:20 -08002340static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002341 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002344 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 int priority;
2346 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002347 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002348 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002350 struct scan_control sc = {
2351 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002352 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002353 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002354 /*
2355 * kswapd doesn't want to be bailed out while reclaim. because
2356 * we want to put equal scanning pressure on each zone.
2357 */
2358 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002359 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002360 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002361 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002362 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002363 struct shrink_control shrink = {
2364 .gfp_mask = sc.gfp_mask,
2365 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366loop_again:
2367 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002368 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002369 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002370 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002374 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002376 /* The swap token gets in the way of swapout... */
2377 if (!priority)
2378 disable_swap_token();
2379
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002381 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002383 /*
2384 * Scan in the highmem->dma direction for the highest
2385 * zone which needs scanning
2386 */
2387 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2388 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002390 if (!populated_zone(zone))
2391 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002393 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002394 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395
Rik van Riel556adec2008-10-18 20:26:34 -07002396 /*
2397 * Do some background aging of the anon list, to give
2398 * pages a chance to be referenced before reclaiming.
2399 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002400 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002401 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2402 &sc, priority, 0);
2403
Mel Gorman88f5acf2011-01-13 15:45:41 -08002404 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002405 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002406 end_zone = i;
Mel Gormandc83edd2011-01-13 15:46:26 -08002407 *classzone_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002408 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002411 if (i < 0)
2412 goto out;
2413
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 for (i = 0; i <= end_zone; i++) {
2415 struct zone *zone = pgdat->node_zones + i;
2416
Wu Fengguangadea02a2009-09-21 17:01:42 -07002417 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 }
2419
2420 /*
2421 * Now scan the zone in the dma->highmem direction, stopping
2422 * at the last zone which needs scanning.
2423 *
2424 * We do this because the page allocator works in the opposite
2425 * direction. This prevents the page allocator from allocating
2426 * pages behind kswapd's direction of progress, which would
2427 * cause too much scanning of the lower zones.
2428 */
2429 for (i = 0; i <= end_zone; i++) {
2430 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002431 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002432 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433
Con Kolivasf3fe6512006-01-06 00:11:15 -08002434 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 continue;
2436
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002437 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 continue;
2439
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002441
Balbir Singh4e416952009-09-23 15:56:39 -07002442 /*
2443 * Call soft limit reclaim before calling shrink_zone.
2444 * For now we ignore the return value
2445 */
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002446 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2447
Rik van Riel32a43302007-10-16 01:24:50 -07002448 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002449 * We put equal pressure on every zone, unless
2450 * one zone has way too many pages free
2451 * already. The "too many pages" is defined
2452 * as the high wmark plus a "gap" where the
2453 * gap is either the low watermark or 1%
2454 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002455 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002456 balance_gap = min(low_wmark_pages(zone),
2457 (zone->present_pages +
2458 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2459 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002460 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002461 high_wmark_pages(zone) + balance_gap,
2462 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002463 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 reclaim_state->reclaimed_slab = 0;
Ying Han1495f232011-05-24 17:12:27 -07002465 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002466 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 total_scanned += sc.nr_scanned;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002468
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002469 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 continue;
Andrea Arcangelid527caf2011-03-22 16:30:38 -07002471 if (nr_slab == 0 &&
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002472 !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002473 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 /*
2475 * If we've done a decent amount of scanning and
2476 * the reclaim ratio is low, start doing writepage
2477 * even in laptop mode
2478 */
2479 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002480 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002482
Mel Gorman88f5acf2011-01-13 15:45:41 -08002483 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002484 high_wmark_pages(zone), end_zone, 0)) {
2485 all_zones_ok = 0;
2486 /*
2487 * We are still under min water mark. This
2488 * means that we have a GFP_ATOMIC allocation
2489 * failure risk. Hurry up!
2490 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002491 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002492 min_wmark_pages(zone), end_zone, 0))
2493 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002494 } else {
2495 /*
2496 * If a zone reaches its high watermark,
2497 * consider it to be no longer congested. It's
2498 * possible there are dirty pages backed by
2499 * congested BDIs but as pressure is relieved,
2500 * spectulatively avoid congestion waits
2501 */
2502 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002503 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002504 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002505 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002506
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002508 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 break; /* kswapd: all done */
2510 /*
2511 * OK, kswapd is getting into trouble. Take a nap, then take
2512 * another pass across the zones.
2513 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002514 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2515 if (has_under_min_watermark_zone)
2516 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2517 else
2518 congestion_wait(BLK_RW_ASYNC, HZ/10);
2519 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
2521 /*
2522 * We do this so kswapd doesn't build up large priorities for
2523 * example when it is freeing in parallel with allocators. It
2524 * matches the direct reclaim path behaviour in terms of impact
2525 * on zone->*_priority.
2526 */
Rik van Riela79311c2009-01-06 14:40:01 -08002527 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 break;
2529 }
2530out:
Mel Gorman99504742011-01-13 15:46:20 -08002531
2532 /*
2533 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002534 * high-order: Balanced zones must make up at least 25% of the node
2535 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002536 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002537 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002539
2540 try_to_freeze();
2541
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002542 /*
2543 * Fragmentation may mean that the system cannot be
2544 * rebalanced for high-order allocations in all zones.
2545 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2546 * it means the zones have been fully scanned and are still
2547 * not balanced. For high-order allocations, there is
2548 * little point trying all over again as kswapd may
2549 * infinite loop.
2550 *
2551 * Instead, recheck all watermarks at order-0 as they
2552 * are the most important. If watermarks are ok, kswapd will go
2553 * back to sleep. High-order users can still perform direct
2554 * reclaim if they wish.
2555 */
2556 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2557 order = sc.order = 0;
2558
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 goto loop_again;
2560 }
2561
Mel Gorman99504742011-01-13 15:46:20 -08002562 /*
2563 * If kswapd was reclaiming at a higher order, it has the option of
2564 * sleeping without all zones being balanced. Before it does, it must
2565 * ensure that the watermarks for order-0 on *all* zones are met and
2566 * that the congestion flags are cleared. The congestion flag must
2567 * be cleared as kswapd is the only mechanism that clears the flag
2568 * and it is potentially going to sleep here.
2569 */
2570 if (order) {
2571 for (i = 0; i <= end_zone; i++) {
2572 struct zone *zone = pgdat->node_zones + i;
2573
2574 if (!populated_zone(zone))
2575 continue;
2576
2577 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2578 continue;
2579
2580 /* Confirm the zone is balanced for order-0 */
2581 if (!zone_watermark_ok(zone, 0,
2582 high_wmark_pages(zone), 0, 0)) {
2583 order = sc.order = 0;
2584 goto loop_again;
2585 }
2586
2587 /* If balanced, clear the congested flag */
2588 zone_clear_flag(zone, ZONE_CONGESTED);
2589 }
2590 }
2591
Mel Gorman0abdee22011-01-13 15:46:22 -08002592 /*
2593 * Return the order we were reclaiming at so sleeping_prematurely()
2594 * makes a decision on the order we were last reclaiming at. However,
2595 * if another caller entered the allocator slow path while kswapd
2596 * was awake, order will remain at the higher level
2597 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002598 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002599 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600}
2601
Mel Gormandc83edd2011-01-13 15:46:26 -08002602static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002603{
2604 long remaining = 0;
2605 DEFINE_WAIT(wait);
2606
2607 if (freezing(current) || kthread_should_stop())
2608 return;
2609
2610 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2611
2612 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002613 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002614 remaining = schedule_timeout(HZ/10);
2615 finish_wait(&pgdat->kswapd_wait, &wait);
2616 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2617 }
2618
2619 /*
2620 * After a short sleep, check if it was a premature sleep. If not, then
2621 * go fully to sleep until explicitly woken up.
2622 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002623 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002624 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2625
2626 /*
2627 * vmstat counters are not perfectly accurate and the estimated
2628 * value for counters such as NR_FREE_PAGES can deviate from the
2629 * true value by nr_online_cpus * threshold. To avoid the zone
2630 * watermarks being breached while under pressure, we reduce the
2631 * per-cpu vmstat threshold while kswapd is awake and restore
2632 * them before going back to sleep.
2633 */
2634 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2635 schedule();
2636 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2637 } else {
2638 if (remaining)
2639 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2640 else
2641 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2642 }
2643 finish_wait(&pgdat->kswapd_wait, &wait);
2644}
2645
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646/*
2647 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002648 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 *
2650 * This basically trickles out pages so that we have _some_
2651 * free memory available even if there is no other activity
2652 * that frees anything up. This is needed for things like routing
2653 * etc, where we otherwise might have all activity going on in
2654 * asynchronous contexts that cannot page things out.
2655 *
2656 * If there are applications that are active memory-allocators
2657 * (most normal use), this basically shouldn't matter.
2658 */
2659static int kswapd(void *p)
2660{
2661 unsigned long order;
Mel Gorman99504742011-01-13 15:46:20 -08002662 int classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 pg_data_t *pgdat = (pg_data_t*)p;
2664 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002665
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 struct reclaim_state reclaim_state = {
2667 .reclaimed_slab = 0,
2668 };
Rusty Russella70f7302009-03-13 14:49:46 +10302669 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670
Nick Piggincf40bd12009-01-21 08:12:39 +01002671 lockdep_set_current_reclaim_state(GFP_KERNEL);
2672
Rusty Russell174596a2009-01-01 10:12:29 +10302673 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002674 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 current->reclaim_state = &reclaim_state;
2676
2677 /*
2678 * Tell the memory management that we're a "memory allocator",
2679 * and that if we need more memory we should get access to it
2680 * regardless (see "__alloc_pages()"). "kswapd" should
2681 * never get caught in the normal page freeing logic.
2682 *
2683 * (Kswapd normally doesn't need memory anyway, but sometimes
2684 * you need a small amount of memory in order to be able to
2685 * page out something else, and this flag essentially protects
2686 * us from recursively trying to free more memory as we're
2687 * trying to free the first piece of memory in the first place).
2688 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002689 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002690 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691
2692 order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002693 classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 for ( ; ; ) {
2695 unsigned long new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002696 int new_classzone_idx;
David Rientjes8fe23e02009-12-14 17:58:33 -08002697 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002698
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 new_order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002700 new_classzone_idx = pgdat->classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 pgdat->kswapd_max_order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002702 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2703 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 /*
2705 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002706 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 */
2708 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002709 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 } else {
Mel Gormandc83edd2011-01-13 15:46:26 -08002711 kswapd_try_to_sleep(pgdat, order, classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002713 classzone_idx = pgdat->classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002714 pgdat->kswapd_max_order = 0;
2715 pgdat->classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
David Rientjes8fe23e02009-12-14 17:58:33 -08002718 ret = try_to_freeze();
2719 if (kthread_should_stop())
2720 break;
2721
2722 /*
2723 * We can speed up thawing tasks if we don't call balance_pgdat
2724 * after returning from the refrigerator
2725 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002726 if (!ret) {
2727 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Mel Gormandc83edd2011-01-13 15:46:26 -08002728 order = balance_pgdat(pgdat, order, &classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002729 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 }
2731 return 0;
2732}
2733
2734/*
2735 * A zone is low on free memory, so wake its kswapd task to service it.
2736 */
Mel Gorman99504742011-01-13 15:46:20 -08002737void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738{
2739 pg_data_t *pgdat;
2740
Con Kolivasf3fe6512006-01-06 00:11:15 -08002741 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 return;
2743
Paul Jackson02a0e532006-12-13 00:34:25 -08002744 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002746 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002747 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002748 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002749 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2750 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002751 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002753 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2754 return;
2755
2756 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002757 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759
Wu Fengguangadea02a2009-09-21 17:01:42 -07002760/*
2761 * The reclaimable count would be mostly accurate.
2762 * The less reclaimable pages may be
2763 * - mlocked pages, which will be moved to unevictable list when encountered
2764 * - mapped pages, which may require several travels to be reclaimed
2765 * - dirty pages, which is not "instantly" reclaimable
2766 */
2767unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002768{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002769 int nr;
2770
2771 nr = global_page_state(NR_ACTIVE_FILE) +
2772 global_page_state(NR_INACTIVE_FILE);
2773
2774 if (nr_swap_pages > 0)
2775 nr += global_page_state(NR_ACTIVE_ANON) +
2776 global_page_state(NR_INACTIVE_ANON);
2777
2778 return nr;
2779}
2780
2781unsigned long zone_reclaimable_pages(struct zone *zone)
2782{
2783 int nr;
2784
2785 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2786 zone_page_state(zone, NR_INACTIVE_FILE);
2787
2788 if (nr_swap_pages > 0)
2789 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2790 zone_page_state(zone, NR_INACTIVE_ANON);
2791
2792 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002793}
2794
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002795#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002797 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002798 * freed pages.
2799 *
2800 * Rather than trying to age LRUs the aim is to preserve the overall
2801 * LRU order by reclaiming preferentially
2802 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002804unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002806 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002807 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002808 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2809 .may_swap = 1,
2810 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002811 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002812 .nr_to_reclaim = nr_to_reclaim,
2813 .hibernation_mode = 1,
2814 .swappiness = vm_swappiness,
2815 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002817 struct shrink_control shrink = {
2818 .gfp_mask = sc.gfp_mask,
2819 };
2820 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002821 struct task_struct *p = current;
2822 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002824 p->flags |= PF_MEMALLOC;
2825 lockdep_set_current_reclaim_state(sc.gfp_mask);
2826 reclaim_state.reclaimed_slab = 0;
2827 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002828
Ying Hana09ed5e2011-05-24 17:12:26 -07002829 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002830
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002831 p->reclaim_state = NULL;
2832 lockdep_clear_current_reclaim_state();
2833 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002834
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002835 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002837#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839/* It's optimal to keep kswapds on the same CPUs as their memory, but
2840 not required for correctness. So if the last cpu in a node goes
2841 away, we get changed to run anywhere: as the first one comes back,
2842 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002843static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002844 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002846 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002848 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002849 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002850 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302851 const struct cpumask *mask;
2852
2853 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002854
Rusty Russell3e597942009-01-01 10:12:24 +10302855 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002857 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 }
2859 }
2860 return NOTIFY_OK;
2861}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862
Yasunori Goto3218ae12006-06-27 02:53:33 -07002863/*
2864 * This kswapd start function will be called by init and node-hot-add.
2865 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2866 */
2867int kswapd_run(int nid)
2868{
2869 pg_data_t *pgdat = NODE_DATA(nid);
2870 int ret = 0;
2871
2872 if (pgdat->kswapd)
2873 return 0;
2874
2875 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2876 if (IS_ERR(pgdat->kswapd)) {
2877 /* failure at boot is fatal */
2878 BUG_ON(system_state == SYSTEM_BOOTING);
2879 printk("Failed to start kswapd on node %d\n",nid);
2880 ret = -1;
2881 }
2882 return ret;
2883}
2884
David Rientjes8fe23e02009-12-14 17:58:33 -08002885/*
2886 * Called by memory hotplug when all memory in a node is offlined.
2887 */
2888void kswapd_stop(int nid)
2889{
2890 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2891
2892 if (kswapd)
2893 kthread_stop(kswapd);
2894}
2895
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896static int __init kswapd_init(void)
2897{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002898 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002899
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002901 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002902 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 hotcpu_notifier(cpu_callback, 0);
2904 return 0;
2905}
2906
2907module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002908
2909#ifdef CONFIG_NUMA
2910/*
2911 * Zone reclaim mode
2912 *
2913 * If non-zero call zone_reclaim when the number of free pages falls below
2914 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002915 */
2916int zone_reclaim_mode __read_mostly;
2917
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002918#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002919#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002920#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2921#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2922
Christoph Lameter9eeff232006-01-18 17:42:31 -08002923/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002924 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2925 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2926 * a zone.
2927 */
2928#define ZONE_RECLAIM_PRIORITY 4
2929
Christoph Lameter9eeff232006-01-18 17:42:31 -08002930/*
Christoph Lameter96146342006-07-03 00:24:13 -07002931 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2932 * occur.
2933 */
2934int sysctl_min_unmapped_ratio = 1;
2935
2936/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002937 * If the number of slab pages in a zone grows beyond this percentage then
2938 * slab reclaim needs to occur.
2939 */
2940int sysctl_min_slab_ratio = 5;
2941
Mel Gorman90afa5d2009-06-16 15:33:20 -07002942static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2943{
2944 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2945 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2946 zone_page_state(zone, NR_ACTIVE_FILE);
2947
2948 /*
2949 * It's possible for there to be more file mapped pages than
2950 * accounted for by the pages on the file LRU lists because
2951 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2952 */
2953 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2954}
2955
2956/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2957static long zone_pagecache_reclaimable(struct zone *zone)
2958{
2959 long nr_pagecache_reclaimable;
2960 long delta = 0;
2961
2962 /*
2963 * If RECLAIM_SWAP is set, then all file pages are considered
2964 * potentially reclaimable. Otherwise, we have to worry about
2965 * pages like swapcache and zone_unmapped_file_pages() provides
2966 * a better estimate
2967 */
2968 if (zone_reclaim_mode & RECLAIM_SWAP)
2969 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2970 else
2971 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2972
2973 /* If we can't clean pages, remove dirty pages from consideration */
2974 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2975 delta += zone_page_state(zone, NR_FILE_DIRTY);
2976
2977 /* Watch for any possible underflows due to delta */
2978 if (unlikely(delta > nr_pagecache_reclaimable))
2979 delta = nr_pagecache_reclaimable;
2980
2981 return nr_pagecache_reclaimable - delta;
2982}
2983
Christoph Lameter0ff38492006-09-25 23:31:52 -07002984/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002985 * Try to free up some pages from this zone through reclaim.
2986 */
Andrew Morton179e9632006-03-22 00:08:18 -08002987static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002988{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002989 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002990 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002991 struct task_struct *p = current;
2992 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002993 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002994 struct scan_control sc = {
2995 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002996 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002997 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002998 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2999 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003000 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003001 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003002 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003003 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003004 struct shrink_control shrink = {
3005 .gfp_mask = sc.gfp_mask,
3006 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003007 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003008
Christoph Lameter9eeff232006-01-18 17:42:31 -08003009 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003010 /*
3011 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3012 * and we also need to be able to write out pages for RECLAIM_WRITE
3013 * and RECLAIM_SWAP.
3014 */
3015 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003016 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003017 reclaim_state.reclaimed_slab = 0;
3018 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003019
Mel Gorman90afa5d2009-06-16 15:33:20 -07003020 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003021 /*
3022 * Free memory by calling shrink zone with increasing
3023 * priorities until we have enough memory freed.
3024 */
3025 priority = ZONE_RECLAIM_PRIORITY;
3026 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003027 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003028 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003029 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003030 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003031
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003032 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3033 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003034 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003035 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003036 * many pages were freed in this zone. So we take the current
3037 * number of slab pages and shake the slab until it is reduced
3038 * by the same nr_pages that we used for reclaiming unmapped
3039 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003040 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003041 * Note that shrink_slab will free memory on all zones and may
3042 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003043 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003044 for (;;) {
3045 unsigned long lru_pages = zone_reclaimable_pages(zone);
3046
3047 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003048 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003049 break;
3050
3051 /* Freed enough memory */
3052 nr_slab_pages1 = zone_page_state(zone,
3053 NR_SLAB_RECLAIMABLE);
3054 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3055 break;
3056 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003057
3058 /*
3059 * Update nr_reclaimed by the number of slab pages we
3060 * reclaimed from this zone.
3061 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003062 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3063 if (nr_slab_pages1 < nr_slab_pages0)
3064 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003065 }
3066
Christoph Lameter9eeff232006-01-18 17:42:31 -08003067 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003068 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003069 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003070 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003071}
Andrew Morton179e9632006-03-22 00:08:18 -08003072
3073int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3074{
Andrew Morton179e9632006-03-22 00:08:18 -08003075 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003076 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003077
3078 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003079 * Zone reclaim reclaims unmapped file backed pages and
3080 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003081 *
Christoph Lameter96146342006-07-03 00:24:13 -07003082 * A small portion of unmapped file backed pages is needed for
3083 * file I/O otherwise pages read by file I/O will be immediately
3084 * thrown out if the zone is overallocated. So we do not reclaim
3085 * if less than a specified percentage of the zone is used by
3086 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003087 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003088 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3089 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003090 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003091
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003092 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003093 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003094
Andrew Morton179e9632006-03-22 00:08:18 -08003095 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003096 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003097 */
David Rientjesd773ed62007-10-16 23:26:01 -07003098 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003099 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003100
3101 /*
3102 * Only run zone reclaim on the local zone or on zones that do not
3103 * have associated processors. This will favor the local processor
3104 * over remote processors and spread off node memory allocations
3105 * as wide as possible.
3106 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003107 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003108 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003109 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003110
3111 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003112 return ZONE_RECLAIM_NOSCAN;
3113
David Rientjesd773ed62007-10-16 23:26:01 -07003114 ret = __zone_reclaim(zone, gfp_mask, order);
3115 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3116
Mel Gorman24cf725182009-06-16 15:33:23 -07003117 if (!ret)
3118 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3119
David Rientjesd773ed62007-10-16 23:26:01 -07003120 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003121}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003122#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003123
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003124/*
3125 * page_evictable - test whether a page is evictable
3126 * @page: the page to test
3127 * @vma: the VMA in which the page is or will be mapped, may be NULL
3128 *
3129 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003130 * lists vs unevictable list. The vma argument is !NULL when called from the
3131 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003132 *
3133 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003134 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003135 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003136 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003137 */
3138int page_evictable(struct page *page, struct vm_area_struct *vma)
3139{
3140
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003141 if (mapping_unevictable(page_mapping(page)))
3142 return 0;
3143
Nick Pigginb291f002008-10-18 20:26:44 -07003144 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3145 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003146
3147 return 1;
3148}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003149
3150/**
3151 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3152 * @page: page to check evictability and move to appropriate lru list
3153 * @zone: zone page is in
3154 *
3155 * Checks a page for evictability and moves the page to the appropriate
3156 * zone lru list.
3157 *
3158 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3159 * have PageUnevictable set.
3160 */
3161static void check_move_unevictable_page(struct page *page, struct zone *zone)
3162{
3163 VM_BUG_ON(PageActive(page));
3164
3165retry:
3166 ClearPageUnevictable(page);
3167 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003168 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003169
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003170 __dec_zone_state(zone, NR_UNEVICTABLE);
3171 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003172 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003173 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3174 __count_vm_event(UNEVICTABLE_PGRESCUED);
3175 } else {
3176 /*
3177 * rotate unevictable list
3178 */
3179 SetPageUnevictable(page);
3180 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003181 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003182 if (page_evictable(page, NULL))
3183 goto retry;
3184 }
3185}
3186
3187/**
3188 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3189 * @mapping: struct address_space to scan for evictable pages
3190 *
3191 * Scan all pages in mapping. Check unevictable pages for
3192 * evictability and move them to the appropriate zone lru list.
3193 */
3194void scan_mapping_unevictable_pages(struct address_space *mapping)
3195{
3196 pgoff_t next = 0;
3197 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3198 PAGE_CACHE_SHIFT;
3199 struct zone *zone;
3200 struct pagevec pvec;
3201
3202 if (mapping->nrpages == 0)
3203 return;
3204
3205 pagevec_init(&pvec, 0);
3206 while (next < end &&
3207 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3208 int i;
3209 int pg_scanned = 0;
3210
3211 zone = NULL;
3212
3213 for (i = 0; i < pagevec_count(&pvec); i++) {
3214 struct page *page = pvec.pages[i];
3215 pgoff_t page_index = page->index;
3216 struct zone *pagezone = page_zone(page);
3217
3218 pg_scanned++;
3219 if (page_index > next)
3220 next = page_index;
3221 next++;
3222
3223 if (pagezone != zone) {
3224 if (zone)
3225 spin_unlock_irq(&zone->lru_lock);
3226 zone = pagezone;
3227 spin_lock_irq(&zone->lru_lock);
3228 }
3229
3230 if (PageLRU(page) && PageUnevictable(page))
3231 check_move_unevictable_page(page, zone);
3232 }
3233 if (zone)
3234 spin_unlock_irq(&zone->lru_lock);
3235 pagevec_release(&pvec);
3236
3237 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3238 }
3239
3240}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003241
3242/**
3243 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3244 * @zone - zone of which to scan the unevictable list
3245 *
3246 * Scan @zone's unevictable LRU lists to check for pages that have become
3247 * evictable. Move those that have to @zone's inactive list where they
3248 * become candidates for reclaim, unless shrink_inactive_zone() decides
3249 * to reactivate them. Pages that are still unevictable are rotated
3250 * back onto @zone's unevictable list.
3251 */
3252#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08003253static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003254{
3255 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3256 unsigned long scan;
3257 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3258
3259 while (nr_to_scan > 0) {
3260 unsigned long batch_size = min(nr_to_scan,
3261 SCAN_UNEVICTABLE_BATCH_SIZE);
3262
3263 spin_lock_irq(&zone->lru_lock);
3264 for (scan = 0; scan < batch_size; scan++) {
3265 struct page *page = lru_to_page(l_unevictable);
3266
3267 if (!trylock_page(page))
3268 continue;
3269
3270 prefetchw_prev_lru_page(page, l_unevictable, flags);
3271
3272 if (likely(PageLRU(page) && PageUnevictable(page)))
3273 check_move_unevictable_page(page, zone);
3274
3275 unlock_page(page);
3276 }
3277 spin_unlock_irq(&zone->lru_lock);
3278
3279 nr_to_scan -= batch_size;
3280 }
3281}
3282
3283
3284/**
3285 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3286 *
3287 * A really big hammer: scan all zones' unevictable LRU lists to check for
3288 * pages that have become evictable. Move those back to the zones'
3289 * inactive list where they become candidates for reclaim.
3290 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3291 * and we add swap to the system. As such, it runs in the context of a task
3292 * that has possibly/probably made some previously unevictable pages
3293 * evictable.
3294 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003295static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003296{
3297 struct zone *zone;
3298
3299 for_each_zone(zone) {
3300 scan_zone_unevictable_pages(zone);
3301 }
3302}
3303
3304/*
3305 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3306 * all nodes' unevictable lists for evictable pages
3307 */
3308unsigned long scan_unevictable_pages;
3309
3310int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003311 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003312 size_t *length, loff_t *ppos)
3313{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003314 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003315
3316 if (write && *(unsigned long *)table->data)
3317 scan_all_zones_unevictable_pages();
3318
3319 scan_unevictable_pages = 0;
3320 return 0;
3321}
3322
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003323#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003324/*
3325 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3326 * a specified node's per zone unevictable lists for evictable pages.
3327 */
3328
3329static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3330 struct sysdev_attribute *attr,
3331 char *buf)
3332{
3333 return sprintf(buf, "0\n"); /* always zero; should fit... */
3334}
3335
3336static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3337 struct sysdev_attribute *attr,
3338 const char *buf, size_t count)
3339{
3340 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3341 struct zone *zone;
3342 unsigned long res;
3343 unsigned long req = strict_strtoul(buf, 10, &res);
3344
3345 if (!req)
3346 return 1; /* zero is no-op */
3347
3348 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3349 if (!populated_zone(zone))
3350 continue;
3351 scan_zone_unevictable_pages(zone);
3352 }
3353 return 1;
3354}
3355
3356
3357static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3358 read_scan_unevictable_node,
3359 write_scan_unevictable_node);
3360
3361int scan_unevictable_register_node(struct node *node)
3362{
3363 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3364}
3365
3366void scan_unevictable_unregister_node(struct node *node)
3367{
3368 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3369}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003370#endif