blob: 8ff834e19c2460d33844e43154b058974a6c9849 [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))
Ying Han1bac1802011-05-26 16:25:36 -0700176 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800177
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800178 return zone_page_state(zone, NR_LRU_BASE + lru);
179}
180
181
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182/*
183 * Add a shrinker callback to be called from the vm
184 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700185void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700187 shrinker->nr = 0;
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700192EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194/*
195 * Remove one
196 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700197void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700203EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
Ying Han1495f232011-05-24 17:12:27 -0700205static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
208{
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
211}
212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213#define SHRINK_BATCH 128
214/*
215 * Call the shrink functions to age shrinkable caches
216 *
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
221 *
Simon Arlott183ff222007-10-20 01:27:18 +0200222 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 * slab to avoid swapping.
224 *
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 *
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700230 *
231 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700233unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700234 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700235 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800238 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Ying Han1495f232011-05-24 17:12:27 -0700240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Minchan Kimf06590b2011-05-24 17:11:11 -0700243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
245 ret = 1;
246 goto out;
247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000252 unsigned long max_pass;
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 {
Andrea Arcangelid179e842011-06-15 15:08:51 -07001127 /*
1128 * Check if the page is freed already.
1129 *
1130 * We can't use page_count() as that
1131 * requires compound_head and we don't
1132 * have a pin on the page here. If a
1133 * page is tail, we may or may not
1134 * have isolated the head, so assume
1135 * it's not free, it'd be tricky to
1136 * track the head status without a
1137 * page pin.
1138 */
1139 if (!PageTail(cursor_page) &&
1140 !atomic_read(&cursor_page->_count))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001141 continue;
1142 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001143 }
1144 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001145
1146 /* If we break out of the loop above, lumpy reclaim failed */
1147 if (pfn < end_pfn)
1148 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 }
1150
1151 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001152
1153 trace_mm_vmscan_lru_isolate(order,
1154 nr_to_scan, scan,
1155 nr_taken,
1156 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1157 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 return nr_taken;
1159}
1160
Balbir Singh66e17072008-02-07 00:13:56 -08001161static unsigned long isolate_pages_global(unsigned long nr,
1162 struct list_head *dst,
1163 unsigned long *scanned, int order,
1164 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001165 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001166{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001167 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001168 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001169 lru += LRU_ACTIVE;
1170 if (file)
1171 lru += LRU_FILE;
1172 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001173 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001174}
1175
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001177 * clear_active_flags() is a helper for shrink_active_list(), clearing
1178 * any active bits from the pages in the list.
1179 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001180static unsigned long clear_active_flags(struct list_head *page_list,
1181 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001182{
1183 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001184 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001185 struct page *page;
1186
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001187 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001188 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001189 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001190 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001191 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001192 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001193 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001194 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001195 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001196 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001197 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001198
1199 return nr_active;
1200}
1201
Nick Piggin62695a82008-10-18 20:26:09 -07001202/**
1203 * isolate_lru_page - tries to isolate a page from its LRU list
1204 * @page: page to isolate from its LRU list
1205 *
1206 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1207 * vmstat statistic corresponding to whatever LRU list the page was on.
1208 *
1209 * Returns 0 if the page was removed from an LRU list.
1210 * Returns -EBUSY if the page was not on an LRU list.
1211 *
1212 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001213 * the active list, it will have PageActive set. If it was found on
1214 * the unevictable list, it will have the PageUnevictable bit set. That flag
1215 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001216 *
1217 * The vmstat statistic corresponding to the list on which the page was
1218 * found will be decremented.
1219 *
1220 * Restrictions:
1221 * (1) Must be called with an elevated refcount on the page. This is a
1222 * fundamentnal difference from isolate_lru_pages (which is called
1223 * without a stable reference).
1224 * (2) the lru_lock must not be held.
1225 * (3) interrupts must be enabled.
1226 */
1227int isolate_lru_page(struct page *page)
1228{
1229 int ret = -EBUSY;
1230
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001231 VM_BUG_ON(!page_count(page));
1232
Nick Piggin62695a82008-10-18 20:26:09 -07001233 if (PageLRU(page)) {
1234 struct zone *zone = page_zone(page);
1235
1236 spin_lock_irq(&zone->lru_lock);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001237 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001238 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001239 ret = 0;
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001240 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001241 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001242
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001243 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001244 }
1245 spin_unlock_irq(&zone->lru_lock);
1246 }
1247 return ret;
1248}
1249
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001250/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001251 * Are there way too many processes in the direct reclaim path already?
1252 */
1253static int too_many_isolated(struct zone *zone, int file,
1254 struct scan_control *sc)
1255{
1256 unsigned long inactive, isolated;
1257
1258 if (current_is_kswapd())
1259 return 0;
1260
1261 if (!scanning_global_lru(sc))
1262 return 0;
1263
1264 if (file) {
1265 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1266 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1267 } else {
1268 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1269 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1270 }
1271
1272 return isolated > inactive;
1273}
1274
1275/*
Mel Gorman66635622010-08-09 17:19:30 -07001276 * TODO: Try merging with migrations version of putback_lru_pages
1277 */
1278static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001279putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001280 unsigned long nr_anon, unsigned long nr_file,
1281 struct list_head *page_list)
1282{
1283 struct page *page;
1284 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001285 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001286
1287 pagevec_init(&pvec, 1);
1288
1289 /*
1290 * Put back any unfreeable pages.
1291 */
1292 spin_lock(&zone->lru_lock);
1293 while (!list_empty(page_list)) {
1294 int lru;
1295 page = lru_to_page(page_list);
1296 VM_BUG_ON(PageLRU(page));
1297 list_del(&page->lru);
1298 if (unlikely(!page_evictable(page, NULL))) {
1299 spin_unlock_irq(&zone->lru_lock);
1300 putback_lru_page(page);
1301 spin_lock_irq(&zone->lru_lock);
1302 continue;
1303 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001304 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001305 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001306 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001307 if (is_active_lru(lru)) {
1308 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001309 int numpages = hpage_nr_pages(page);
1310 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001311 }
1312 if (!pagevec_add(&pvec, page)) {
1313 spin_unlock_irq(&zone->lru_lock);
1314 __pagevec_release(&pvec);
1315 spin_lock_irq(&zone->lru_lock);
1316 }
1317 }
1318 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1319 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1320
1321 spin_unlock_irq(&zone->lru_lock);
1322 pagevec_release(&pvec);
1323}
1324
Mel Gorman1489fa12010-08-09 17:19:33 -07001325static noinline_for_stack void update_isolated_counts(struct zone *zone,
1326 struct scan_control *sc,
1327 unsigned long *nr_anon,
1328 unsigned long *nr_file,
1329 struct list_head *isolated_list)
1330{
1331 unsigned long nr_active;
1332 unsigned int count[NR_LRU_LISTS] = { 0, };
1333 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1334
1335 nr_active = clear_active_flags(isolated_list, count);
1336 __count_vm_events(PGDEACTIVATE, nr_active);
1337
1338 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1339 -count[LRU_ACTIVE_FILE]);
1340 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1341 -count[LRU_INACTIVE_FILE]);
1342 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1343 -count[LRU_ACTIVE_ANON]);
1344 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1345 -count[LRU_INACTIVE_ANON]);
1346
1347 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1348 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1349 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1350 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1351
1352 reclaim_stat->recent_scanned[0] += *nr_anon;
1353 reclaim_stat->recent_scanned[1] += *nr_file;
1354}
1355
Mel Gorman66635622010-08-09 17:19:30 -07001356/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001357 * Returns true if the caller should wait to clean dirty/writeback pages.
1358 *
1359 * If we are direct reclaiming for contiguous pages and we do not reclaim
1360 * everything in the list, try again and wait for writeback IO to complete.
1361 * This will stall high-order allocations noticeably. Only do that when really
1362 * need to free the pages under high memory pressure.
1363 */
1364static inline bool should_reclaim_stall(unsigned long nr_taken,
1365 unsigned long nr_freed,
1366 int priority,
1367 struct scan_control *sc)
1368{
1369 int lumpy_stall_priority;
1370
1371 /* kswapd should not stall on sync IO */
1372 if (current_is_kswapd())
1373 return false;
1374
1375 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001376 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001377 return false;
1378
1379 /* If we have relaimed everything on the isolated list, no stall */
1380 if (nr_freed == nr_taken)
1381 return false;
1382
1383 /*
1384 * For high-order allocations, there are two stall thresholds.
1385 * High-cost allocations stall immediately where as lower
1386 * order allocations such as stacks require the scanning
1387 * priority to be much higher before stalling.
1388 */
1389 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1390 lumpy_stall_priority = DEF_PRIORITY;
1391 else
1392 lumpy_stall_priority = DEF_PRIORITY / 3;
1393
1394 return priority <= lumpy_stall_priority;
1395}
1396
1397/*
Andrew Morton1742f192006-03-22 00:08:21 -08001398 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1399 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 */
Mel Gorman66635622010-08-09 17:19:30 -07001401static noinline_for_stack unsigned long
1402shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1403 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404{
1405 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001406 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001407 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001408 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001409 unsigned long nr_anon;
1410 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001411
Rik van Riel35cd7812009-09-21 17:01:38 -07001412 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001413 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001414
1415 /* We are about to die and free our memory. Return now. */
1416 if (fatal_signal_pending(current))
1417 return SWAP_CLUSTER_MAX;
1418 }
1419
Mel Gormanf3a310b2011-01-13 15:46:00 -08001420 set_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 lru_add_drain();
1422 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001424 if (scanning_global_lru(sc)) {
1425 nr_taken = isolate_pages_global(nr_to_scan,
1426 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001427 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001428 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001429 zone, 0, file);
1430 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001431 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001432 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1433 nr_scanned);
1434 else
1435 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1436 nr_scanned);
1437 } else {
1438 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1439 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001440 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001441 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001442 zone, sc->mem_cgroup,
1443 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001445 * mem_cgroup_isolate_pages() keeps track of
1446 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001448 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001449
Mel Gorman66635622010-08-09 17:19:30 -07001450 if (nr_taken == 0) {
1451 spin_unlock_irq(&zone->lru_lock);
1452 return 0;
1453 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001454
Mel Gorman1489fa12010-08-09 17:19:33 -07001455 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001456
1457 spin_unlock_irq(&zone->lru_lock);
1458
Mel Gorman0e093d992010-10-26 14:21:45 -07001459 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001460
Wu Fengguange31f36982010-08-09 17:20:01 -07001461 /* Check if we should syncronously wait for writeback */
1462 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001463 set_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001464 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001465 }
1466
1467 local_irq_disable();
1468 if (current_is_kswapd())
1469 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1470 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1471
Mel Gorman1489fa12010-08-09 17:19:33 -07001472 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001473
1474 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1475 zone_idx(zone),
1476 nr_scanned, nr_reclaimed,
1477 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001478 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001479 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480}
1481
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001482/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 * This moves pages from the active list to the inactive list.
1484 *
1485 * We move them the other way if the page is referenced by one or more
1486 * processes, from rmap.
1487 *
1488 * If the pages are mostly unmapped, the processing is fast and it is
1489 * appropriate to hold zone->lru_lock across the whole operation. But if
1490 * the pages are mapped, the processing is slow (page_referenced()) so we
1491 * should drop zone->lru_lock around each page. It's impossible to balance
1492 * this, so instead we remove the pages from the LRU while processing them.
1493 * It is safe to rely on PG_active against the non-LRU pages in here because
1494 * nobody will play with that bit on a non-LRU page.
1495 *
1496 * The downside is that we have to touch page->_count against each page.
1497 * But we had to alter page->flags anyway.
1498 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001499
Wu Fengguang3eb41402009-06-16 15:33:13 -07001500static void move_active_pages_to_lru(struct zone *zone,
1501 struct list_head *list,
1502 enum lru_list lru)
1503{
1504 unsigned long pgmoved = 0;
1505 struct pagevec pvec;
1506 struct page *page;
1507
1508 pagevec_init(&pvec, 1);
1509
1510 while (!list_empty(list)) {
1511 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001512
1513 VM_BUG_ON(PageLRU(page));
1514 SetPageLRU(page);
1515
Wu Fengguang3eb41402009-06-16 15:33:13 -07001516 list_move(&page->lru, &zone->lru[lru].list);
1517 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001518 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001519
1520 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1521 spin_unlock_irq(&zone->lru_lock);
1522 if (buffer_heads_over_limit)
1523 pagevec_strip(&pvec);
1524 __pagevec_release(&pvec);
1525 spin_lock_irq(&zone->lru_lock);
1526 }
1527 }
1528 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1529 if (!is_active_lru(lru))
1530 __count_vm_events(PGDEACTIVATE, pgmoved);
1531}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001532
Andrew Morton1742f192006-03-22 00:08:21 -08001533static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001534 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001536 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001537 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001538 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001540 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001541 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001543 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001544 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545
1546 lru_add_drain();
1547 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001548 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001549 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1550 &pgscanned, sc->order,
1551 ISOLATE_ACTIVE, zone,
1552 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001553 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001554 } else {
1555 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1556 &pgscanned, sc->order,
1557 ISOLATE_ACTIVE, zone,
1558 sc->mem_cgroup, 1, file);
1559 /*
1560 * mem_cgroup_isolate_pages() keeps track of
1561 * scanned pages on its own.
1562 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001563 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001564
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001565 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001566
Wu Fengguang3eb41402009-06-16 15:33:13 -07001567 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001568 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001569 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001570 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001571 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001572 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 spin_unlock_irq(&zone->lru_lock);
1574
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 while (!list_empty(&l_hold)) {
1576 cond_resched();
1577 page = lru_to_page(&l_hold);
1578 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001579
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001580 if (unlikely(!page_evictable(page, NULL))) {
1581 putback_lru_page(page);
1582 continue;
1583 }
1584
Johannes Weiner645747462010-03-05 13:42:22 -08001585 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001586 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001587 /*
1588 * Identify referenced, file-backed active pages and
1589 * give them one more trip around the active list. So
1590 * that executable code get better chances to stay in
1591 * memory under moderate memory pressure. Anon pages
1592 * are not likely to be evicted by use-once streaming
1593 * IO, plus JVM can create lots of anon VM_EXEC pages,
1594 * so we ignore them here.
1595 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001596 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001597 list_add(&page->lru, &l_active);
1598 continue;
1599 }
1600 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001601
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001602 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 list_add(&page->lru, &l_inactive);
1604 }
1605
Andrew Mortonb5557492009-01-06 14:40:13 -08001606 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001607 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001608 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001609 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001610 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001611 * Count referenced pages from currently used mappings as rotated,
1612 * even though only some of them are actually re-activated. This
1613 * helps balance scan pressure between file and anonymous pages in
1614 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001615 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001616 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001617
Wu Fengguang3eb41402009-06-16 15:33:13 -07001618 move_active_pages_to_lru(zone, &l_active,
1619 LRU_ACTIVE + file * LRU_FILE);
1620 move_active_pages_to_lru(zone, &l_inactive,
1621 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001622 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001623 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624}
1625
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001626#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001627static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001628{
1629 unsigned long active, inactive;
1630
1631 active = zone_page_state(zone, NR_ACTIVE_ANON);
1632 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1633
1634 if (inactive * zone->inactive_ratio < active)
1635 return 1;
1636
1637 return 0;
1638}
1639
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001640/**
1641 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1642 * @zone: zone to check
1643 * @sc: scan control of this context
1644 *
1645 * Returns true if the zone does not have enough inactive anon pages,
1646 * meaning some active anon pages need to be deactivated.
1647 */
1648static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1649{
1650 int low;
1651
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001652 /*
1653 * If we don't have swap space, anonymous page deactivation
1654 * is pointless.
1655 */
1656 if (!total_swap_pages)
1657 return 0;
1658
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001659 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001660 low = inactive_anon_is_low_global(zone);
1661 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001662 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001663 return low;
1664}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001665#else
1666static inline int inactive_anon_is_low(struct zone *zone,
1667 struct scan_control *sc)
1668{
1669 return 0;
1670}
1671#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001672
Rik van Riel56e49d22009-06-16 15:32:28 -07001673static int inactive_file_is_low_global(struct zone *zone)
1674{
1675 unsigned long active, inactive;
1676
1677 active = zone_page_state(zone, NR_ACTIVE_FILE);
1678 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1679
1680 return (active > inactive);
1681}
1682
1683/**
1684 * inactive_file_is_low - check if file pages need to be deactivated
1685 * @zone: zone to check
1686 * @sc: scan control of this context
1687 *
1688 * When the system is doing streaming IO, memory pressure here
1689 * ensures that active file pages get deactivated, until more
1690 * than half of the file pages are on the inactive list.
1691 *
1692 * Once we get to that situation, protect the system's working
1693 * set from being evicted by disabling active file page aging.
1694 *
1695 * This uses a different ratio than the anonymous pages, because
1696 * the page cache uses a use-once replacement algorithm.
1697 */
1698static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1699{
1700 int low;
1701
1702 if (scanning_global_lru(sc))
1703 low = inactive_file_is_low_global(zone);
1704 else
1705 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1706 return low;
1707}
1708
Rik van Rielb39415b2009-12-14 17:59:48 -08001709static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1710 int file)
1711{
1712 if (file)
1713 return inactive_file_is_low(zone, sc);
1714 else
1715 return inactive_anon_is_low(zone, sc);
1716}
1717
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001718static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001719 struct zone *zone, struct scan_control *sc, int priority)
1720{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001721 int file = is_file_lru(lru);
1722
Rik van Rielb39415b2009-12-14 17:59:48 -08001723 if (is_active_lru(lru)) {
1724 if (inactive_list_is_low(zone, sc, file))
1725 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001726 return 0;
1727 }
1728
Rik van Riel33c120e2008-10-18 20:26:36 -07001729 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001730}
1731
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001733 * Determine how aggressively the anon and file LRU lists should be
1734 * scanned. The relative value of each set of LRU lists is determined
1735 * by looking at the fraction of the pages scanned we did rotate back
1736 * onto the active list instead of evict.
1737 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001738 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001739 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001740static void get_scan_count(struct zone *zone, struct scan_control *sc,
1741 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001742{
1743 unsigned long anon, file, free;
1744 unsigned long anon_prio, file_prio;
1745 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001746 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001747 u64 fraction[2], denominator;
1748 enum lru_list l;
1749 int noswap = 0;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001750 int force_scan = 0;
1751
1752
1753 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1754 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1755 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1756 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1757
1758 if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1759 /* kswapd does zone balancing and need to scan this zone */
1760 if (scanning_global_lru(sc) && current_is_kswapd())
1761 force_scan = 1;
1762 /* memcg may have small limit and need to avoid priority drop */
1763 if (!scanning_global_lru(sc))
1764 force_scan = 1;
1765 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07001766
1767 /* If we have no swap space, do not bother scanning anon pages. */
1768 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1769 noswap = 1;
1770 fraction[0] = 0;
1771 fraction[1] = 1;
1772 denominator = 1;
1773 goto out;
1774 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001775
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001776 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001777 free = zone_page_state(zone, NR_FREE_PAGES);
1778 /* If we have very few page cache pages,
1779 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001780 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001781 fraction[0] = 1;
1782 fraction[1] = 0;
1783 denominator = 1;
1784 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001785 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001786 }
1787
1788 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001789 * With swappiness at 100, anonymous and file have the same priority.
1790 * This scanning priority is essentially the inverse of IO cost.
1791 */
1792 anon_prio = sc->swappiness;
1793 file_prio = 200 - sc->swappiness;
1794
1795 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001796 * OK, so we have swap space and a fair amount of page cache
1797 * pages. We use the recently rotated / recently scanned
1798 * ratios to determine how valuable each cache is.
1799 *
1800 * Because workloads change over time (and to avoid overflow)
1801 * we keep these statistics as a floating average, which ends
1802 * up weighing recent references more than old ones.
1803 *
1804 * anon in [0], file in [1]
1805 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001806 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001807 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001808 reclaim_stat->recent_scanned[0] /= 2;
1809 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001810 }
1811
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001812 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001813 reclaim_stat->recent_scanned[1] /= 2;
1814 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001815 }
1816
1817 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001818 * The amount of pressure on anon vs file pages is inversely
1819 * proportional to the fraction of recently scanned pages on
1820 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001821 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001822 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1823 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001824
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001825 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1826 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001827 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001828
Shaohua Li76a33fc2010-05-24 14:32:36 -07001829 fraction[0] = ap;
1830 fraction[1] = fp;
1831 denominator = ap + fp + 1;
1832out:
1833 for_each_evictable_lru(l) {
1834 int file = is_file_lru(l);
1835 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001836
Shaohua Li76a33fc2010-05-24 14:32:36 -07001837 scan = zone_nr_lru_pages(zone, sc, l);
1838 if (priority || noswap) {
1839 scan >>= priority;
1840 scan = div64_u64(scan * fraction[file], denominator);
1841 }
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001842
1843 /*
1844 * If zone is small or memcg is small, nr[l] can be 0.
1845 * This results no-scan on this priority and priority drop down.
1846 * For global direct reclaim, it can visit next zone and tend
1847 * not to have problems. For global kswapd, it's for zone
1848 * balancing and it need to scan a small amounts. When using
1849 * memcg, priority drop can cause big latency. So, it's better
1850 * to scan small amount. See may_noscan above.
1851 */
1852 if (!scan && force_scan) {
1853 if (file)
1854 scan = SWAP_CLUSTER_MAX;
1855 else if (!noswap)
1856 scan = SWAP_CLUSTER_MAX;
1857 }
1858 nr[l] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001859 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001860}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001861
1862/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001863 * Reclaim/compaction depends on a number of pages being freed. To avoid
1864 * disruption to the system, a small number of order-0 pages continue to be
1865 * rotated and reclaimed in the normal fashion. However, by the time we get
1866 * back to the allocator and call try_to_compact_zone(), we ensure that
1867 * there are enough free pages for it to be likely successful
1868 */
1869static inline bool should_continue_reclaim(struct zone *zone,
1870 unsigned long nr_reclaimed,
1871 unsigned long nr_scanned,
1872 struct scan_control *sc)
1873{
1874 unsigned long pages_for_compaction;
1875 unsigned long inactive_lru_pages;
1876
1877 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001878 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001879 return false;
1880
Mel Gorman28765922011-02-25 14:44:20 -08001881 /* Consider stopping depending on scan and reclaim activity */
1882 if (sc->gfp_mask & __GFP_REPEAT) {
1883 /*
1884 * For __GFP_REPEAT allocations, stop reclaiming if the
1885 * full LRU list has been scanned and we are still failing
1886 * to reclaim pages. This full LRU scan is potentially
1887 * expensive but a __GFP_REPEAT caller really wants to succeed
1888 */
1889 if (!nr_reclaimed && !nr_scanned)
1890 return false;
1891 } else {
1892 /*
1893 * For non-__GFP_REPEAT allocations which can presumably
1894 * fail without consequence, stop if we failed to reclaim
1895 * any pages from the last SWAP_CLUSTER_MAX number of
1896 * pages that were scanned. This will return to the
1897 * caller faster at the risk reclaim/compaction and
1898 * the resulting allocation attempt fails
1899 */
1900 if (!nr_reclaimed)
1901 return false;
1902 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001903
1904 /*
1905 * If we have not reclaimed enough pages for compaction and the
1906 * inactive lists are large enough, continue reclaiming
1907 */
1908 pages_for_compaction = (2UL << sc->order);
1909 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1910 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1911 if (sc->nr_reclaimed < pages_for_compaction &&
1912 inactive_lru_pages > pages_for_compaction)
1913 return true;
1914
1915 /* If compaction would go ahead or the allocation would succeed, stop */
1916 switch (compaction_suitable(zone, sc->order)) {
1917 case COMPACT_PARTIAL:
1918 case COMPACT_CONTINUE:
1919 return false;
1920 default:
1921 return true;
1922 }
1923}
1924
1925/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1927 */
Rik van Riela79311c2009-01-06 14:40:01 -08001928static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001929 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001931 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001932 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001933 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001934 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001935 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
Mel Gorman3e7d3442011-01-13 15:45:56 -08001937restart:
1938 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001939 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001940 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001941
Rik van Riel556adec2008-10-18 20:26:34 -07001942 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1943 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001944 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001945 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001946 nr_to_scan = min_t(unsigned long,
1947 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001948 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001950 nr_reclaimed += shrink_list(l, nr_to_scan,
1951 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 }
Rik van Riela79311c2009-01-06 14:40:01 -08001954 /*
1955 * On large memory systems, scan >> priority can become
1956 * really large. This is fine for the starting priority;
1957 * we want to put equal scanning pressure on each zone.
1958 * However, if the VM has a harder time of freeing pages,
1959 * with multiple processes reclaiming pages, the total
1960 * freeing target can get unreasonably large.
1961 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001962 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001963 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001965 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001966
Rik van Riel556adec2008-10-18 20:26:34 -07001967 /*
1968 * Even if we did not try to evict anon pages at all, we want to
1969 * rebalance the anon lru active/inactive ratio.
1970 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001971 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001972 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1973
Mel Gorman3e7d3442011-01-13 15:45:56 -08001974 /* reclaim/compaction might need reclaim to continue */
1975 if (should_continue_reclaim(zone, nr_reclaimed,
1976 sc->nr_scanned - nr_scanned, sc))
1977 goto restart;
1978
Andrew Morton232ea4d2007-02-28 20:13:21 -08001979 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980}
1981
1982/*
1983 * This is the direct reclaim path, for page-allocating processes. We only
1984 * try to reclaim pages from zones which will satisfy the caller's allocation
1985 * request.
1986 *
Mel Gorman41858962009-06-16 15:32:12 -07001987 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1988 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1990 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001991 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1992 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1993 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 * If a zone is deemed to be full of pinned pages then just give it a light
1996 * scan then give up on it.
1997 */
Ying Hand149e3b2011-05-26 16:25:27 -07001998static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001999 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000{
Mel Gormandd1a2392008-04-28 02:12:17 -07002001 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002002 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002003 unsigned long nr_soft_reclaimed;
2004 unsigned long nr_soft_scanned;
2005 unsigned long total_scanned = 0;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002006
Mel Gormand4debc62010-08-09 17:19:29 -07002007 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2008 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002009 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002011 /*
2012 * Take care memory controller reclaiming has small influence
2013 * to global LRU.
2014 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002015 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002016 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2017 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002018 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002019 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002020 }
Nick Piggin408d8542006-09-25 23:31:27 -07002021
Ying Hand149e3b2011-05-26 16:25:27 -07002022 nr_soft_scanned = 0;
2023 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2024 sc->order, sc->gfp_mask,
2025 &nr_soft_scanned);
2026 sc->nr_reclaimed += nr_soft_reclaimed;
2027 total_scanned += nr_soft_scanned;
2028
Rik van Riela79311c2009-01-06 14:40:01 -08002029 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 }
Ying Hand149e3b2011-05-26 16:25:27 -07002031
2032 return total_scanned;
Minchan Kimd1908362010-09-22 13:05:01 -07002033}
2034
2035static bool zone_reclaimable(struct zone *zone)
2036{
2037 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2038}
2039
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002040/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002041static bool all_unreclaimable(struct zonelist *zonelist,
2042 struct scan_control *sc)
2043{
2044 struct zoneref *z;
2045 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002046
2047 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2048 gfp_zone(sc->gfp_mask), sc->nodemask) {
2049 if (!populated_zone(zone))
2050 continue;
2051 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2052 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002053 if (!zone->all_unreclaimable)
2054 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002055 }
2056
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002057 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002059
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060/*
2061 * This is the main entry point to direct page reclaim.
2062 *
2063 * If a full scan of the inactive list fails to free enough memory then we
2064 * are "out of memory" and something needs to be killed.
2065 *
2066 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2067 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002068 * caller can't do much about. We kick the writeback threads and take explicit
2069 * naps in the hope that some of these pages can be written. But if the
2070 * allocating task holds filesystem locks which prevent writeout this might not
2071 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002072 *
2073 * returns: 0, if no pages reclaimed
2074 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 */
Mel Gormandac1d272008-04-28 02:12:12 -07002076static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002077 struct scan_control *sc,
2078 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079{
2080 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002081 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002083 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002084 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002085 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086
Miao Xiec0ff7452010-05-24 14:32:08 -07002087 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002088 delayacct_freepages_start();
2089
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002090 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002091 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
2093 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002094 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002095 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002096 disable_swap_token(sc->mem_cgroup);
Ying Hand149e3b2011-05-26 16:25:27 -07002097 total_scanned += shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08002098 /*
2099 * Don't shrink slabs when reclaiming memory from
2100 * over limit cgroups
2101 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002102 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002103 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002104 for_each_zone_zonelist(zone, z, zonelist,
2105 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002106 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2107 continue;
2108
2109 lru_pages += zone_reclaimable_pages(zone);
2110 }
2111
Ying Han1495f232011-05-24 17:12:27 -07002112 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002113 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002114 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002115 reclaim_state->reclaimed_slab = 0;
2116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 }
Balbir Singh66e17072008-02-07 00:13:56 -08002118 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002119 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121
2122 /*
2123 * Try to write back as many pages as we just scanned. This
2124 * tends to cause slow streaming writers to write data to the
2125 * disk smoothly, at the dirtying rate, which is nice. But
2126 * that's undesirable in laptop mode, where we *want* lumpy
2127 * writeout. So in laptop mode, write out the whole world.
2128 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002129 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2130 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002131 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002132 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 }
2134
2135 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002136 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002137 priority < DEF_PRIORITY - 2) {
2138 struct zone *preferred_zone;
2139
2140 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002141 &cpuset_current_mems_allowed,
2142 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002143 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2144 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002146
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002148 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002149 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002150
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002151 if (sc->nr_reclaimed)
2152 return sc->nr_reclaimed;
2153
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002154 /*
2155 * As hibernation is going on, kswapd is freezed so that it can't mark
2156 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2157 * check.
2158 */
2159 if (oom_killer_disabled)
2160 return 0;
2161
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002162 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002163 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002164 return 1;
2165
2166 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167}
2168
Mel Gormandac1d272008-04-28 02:12:12 -07002169unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002170 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002171{
Mel Gorman33906bc2010-08-09 17:19:16 -07002172 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002173 struct scan_control sc = {
2174 .gfp_mask = gfp_mask,
2175 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002176 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002177 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002178 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002179 .swappiness = vm_swappiness,
2180 .order = order,
2181 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002182 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002183 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002184 struct shrink_control shrink = {
2185 .gfp_mask = sc.gfp_mask,
2186 };
Balbir Singh66e17072008-02-07 00:13:56 -08002187
Mel Gorman33906bc2010-08-09 17:19:16 -07002188 trace_mm_vmscan_direct_reclaim_begin(order,
2189 sc.may_writepage,
2190 gfp_mask);
2191
Ying Hana09ed5e2011-05-24 17:12:26 -07002192 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002193
2194 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2195
2196 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002197}
2198
Balbir Singh00f0b822008-03-04 14:28:39 -08002199#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002200
Balbir Singh4e416952009-09-23 15:56:39 -07002201unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2202 gfp_t gfp_mask, bool noswap,
2203 unsigned int swappiness,
Ying Han0ae5e892011-05-26 16:25:25 -07002204 struct zone *zone,
2205 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002206{
2207 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002208 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002209 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002210 .may_writepage = !laptop_mode,
2211 .may_unmap = 1,
2212 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002213 .swappiness = swappiness,
2214 .order = 0,
2215 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002216 };
Ying Han0ae5e892011-05-26 16:25:25 -07002217
Balbir Singh4e416952009-09-23 15:56:39 -07002218 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2219 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002220
2221 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2222 sc.may_writepage,
2223 sc.gfp_mask);
2224
Balbir Singh4e416952009-09-23 15:56:39 -07002225 /*
2226 * NOTE: Although we can get the priority field, using it
2227 * here is not a good idea, since it limits the pages we can scan.
2228 * if we don't reclaim here, the shrink_zone from balance_pgdat
2229 * will pick up pages from other mem cgroup's as well. We hack
2230 * the priority and make it zero.
2231 */
2232 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002233
2234 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2235
Ying Han0ae5e892011-05-26 16:25:25 -07002236 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002237 return sc.nr_reclaimed;
2238}
2239
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002240unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002241 gfp_t gfp_mask,
2242 bool noswap,
2243 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002244{
Balbir Singh4e416952009-09-23 15:56:39 -07002245 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002246 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002247 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002248 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002249 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002250 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002251 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002252 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002253 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002254 .order = 0,
2255 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002256 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002257 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2258 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2259 };
2260 struct shrink_control shrink = {
2261 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002262 };
Balbir Singh66e17072008-02-07 00:13:56 -08002263
Ying Han889976d2011-05-26 16:25:33 -07002264 /*
2265 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2266 * take care of from where we get pages. So the node where we start the
2267 * scan does not need to be the current node.
2268 */
2269 nid = mem_cgroup_select_victim_node(mem_cont);
2270
2271 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002272
2273 trace_mm_vmscan_memcg_reclaim_begin(0,
2274 sc.may_writepage,
2275 sc.gfp_mask);
2276
Ying Hana09ed5e2011-05-24 17:12:26 -07002277 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002278
2279 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2280
2281 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002282}
2283#endif
2284
Mel Gorman1741c872011-01-13 15:46:21 -08002285/*
2286 * pgdat_balanced is used when checking if a node is balanced for high-order
2287 * allocations. Only zones that meet watermarks and are in a zone allowed
2288 * by the callers classzone_idx are added to balanced_pages. The total of
2289 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2290 * for the node to be considered balanced. Forcing all zones to be balanced
2291 * for high orders can cause excessive reclaim when there are imbalanced zones.
2292 * The choice of 25% is due to
2293 * o a 16M DMA zone that is balanced will not balance a zone on any
2294 * reasonable sized machine
2295 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002296 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002297 * would need to be at least 256M for it to be balance a whole node.
2298 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2299 * to balance a node on its own. These seemed like reasonable ratios.
2300 */
2301static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2302 int classzone_idx)
2303{
2304 unsigned long present_pages = 0;
2305 int i;
2306
2307 for (i = 0; i <= classzone_idx; i++)
2308 present_pages += pgdat->node_zones[i].present_pages;
2309
2310 return balanced_pages > (present_pages >> 2);
2311}
2312
Mel Gormanf50de2d2009-12-14 17:58:53 -08002313/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002314static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2315 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002316{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002317 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002318 unsigned long balanced = 0;
2319 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002320
2321 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2322 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002323 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002324
Mel Gorman0abdee22011-01-13 15:46:22 -08002325 /* Check the watermark levels */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002326 for (i = 0; i < pgdat->nr_zones; i++) {
2327 struct zone *zone = pgdat->node_zones + i;
2328
2329 if (!populated_zone(zone))
2330 continue;
2331
Mel Gorman355b09c2011-01-13 15:46:24 -08002332 /*
2333 * balance_pgdat() skips over all_unreclaimable after
2334 * DEF_PRIORITY. Effectively, it considers them balanced so
2335 * they must be considered balanced here as well if kswapd
2336 * is to sleep
2337 */
2338 if (zone->all_unreclaimable) {
2339 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002340 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002341 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002342
Mel Gorman88f5acf2011-01-13 15:45:41 -08002343 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormandc83edd2011-01-13 15:46:26 -08002344 classzone_idx, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002345 all_zones_ok = false;
2346 else
2347 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002348 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002349
Mel Gorman1741c872011-01-13 15:46:21 -08002350 /*
2351 * For high-order requests, the balanced zones must contain at least
2352 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2353 * must be balanced
2354 */
2355 if (order)
Johannes Weinerafc7e322011-05-24 17:11:09 -07002356 return !pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002357 else
2358 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002359}
2360
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361/*
2362 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002363 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002365 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 *
2367 * There is special handling here for zones which are full of pinned pages.
2368 * This can happen if the pages are all mlocked, or if they are all used by
2369 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2370 * What we do is to detect the case where all pages in the zone have been
2371 * scanned twice and there has been zero successful reclaim. Mark the zone as
2372 * dead and from now on, only perform a short scan. Basically we're polling
2373 * the zone for when the problem goes away.
2374 *
2375 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002376 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2377 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2378 * lower zones regardless of the number of free pages in the lower zones. This
2379 * interoperates with the page allocator fallback scheme to ensure that aging
2380 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 */
Mel Gorman99504742011-01-13 15:46:20 -08002382static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002383 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002386 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 int priority;
2388 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002389 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002390 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002392 unsigned long nr_soft_reclaimed;
2393 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002394 struct scan_control sc = {
2395 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002396 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002397 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002398 /*
2399 * kswapd doesn't want to be bailed out while reclaim. because
2400 * we want to put equal scanning pressure on each zone.
2401 */
2402 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002403 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002404 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002405 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002406 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002407 struct shrink_control shrink = {
2408 .gfp_mask = sc.gfp_mask,
2409 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410loop_again:
2411 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002412 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002413 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002414 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002418 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002420 /* The swap token gets in the way of swapout... */
2421 if (!priority)
KOSAKI Motohiroa4336582011-06-15 15:08:13 -07002422 disable_swap_token(NULL);
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002423
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002425 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002427 /*
2428 * Scan in the highmem->dma direction for the highest
2429 * zone which needs scanning
2430 */
2431 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2432 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002434 if (!populated_zone(zone))
2435 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002437 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002438 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439
Rik van Riel556adec2008-10-18 20:26:34 -07002440 /*
2441 * Do some background aging of the anon list, to give
2442 * pages a chance to be referenced before reclaiming.
2443 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002444 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002445 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2446 &sc, priority, 0);
2447
Mel Gorman88f5acf2011-01-13 15:45:41 -08002448 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002449 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002450 end_zone = i;
Mel Gormandc83edd2011-01-13 15:46:26 -08002451 *classzone_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002452 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002455 if (i < 0)
2456 goto out;
2457
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 for (i = 0; i <= end_zone; i++) {
2459 struct zone *zone = pgdat->node_zones + i;
2460
Wu Fengguangadea02a2009-09-21 17:01:42 -07002461 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 }
2463
2464 /*
2465 * Now scan the zone in the dma->highmem direction, stopping
2466 * at the last zone which needs scanning.
2467 *
2468 * We do this because the page allocator works in the opposite
2469 * direction. This prevents the page allocator from allocating
2470 * pages behind kswapd's direction of progress, which would
2471 * cause too much scanning of the lower zones.
2472 */
2473 for (i = 0; i <= end_zone; i++) {
2474 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002475 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002476 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
Con Kolivasf3fe6512006-01-06 00:11:15 -08002478 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 continue;
2480
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002481 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 continue;
2483
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002485
Ying Han0ae5e892011-05-26 16:25:25 -07002486 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002487 /*
2488 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002489 */
Ying Han0ae5e892011-05-26 16:25:25 -07002490 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2491 order, sc.gfp_mask,
2492 &nr_soft_scanned);
2493 sc.nr_reclaimed += nr_soft_reclaimed;
2494 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002495
Rik van Riel32a43302007-10-16 01:24:50 -07002496 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002497 * We put equal pressure on every zone, unless
2498 * one zone has way too many pages free
2499 * already. The "too many pages" is defined
2500 * as the high wmark plus a "gap" where the
2501 * gap is either the low watermark or 1%
2502 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002503 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002504 balance_gap = min(low_wmark_pages(zone),
2505 (zone->present_pages +
2506 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2507 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002508 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002509 high_wmark_pages(zone) + balance_gap,
2510 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002511 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 reclaim_state->reclaimed_slab = 0;
Ying Han1495f232011-05-24 17:12:27 -07002513 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002514 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 total_scanned += sc.nr_scanned;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002516
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002517 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 continue;
Andrea Arcangelid527caf2011-03-22 16:30:38 -07002519 if (nr_slab == 0 &&
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002520 !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002521 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 /*
2523 * If we've done a decent amount of scanning and
2524 * the reclaim ratio is low, start doing writepage
2525 * even in laptop mode
2526 */
2527 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002528 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002530
Mel Gorman88f5acf2011-01-13 15:45:41 -08002531 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002532 high_wmark_pages(zone), end_zone, 0)) {
2533 all_zones_ok = 0;
2534 /*
2535 * We are still under min water mark. This
2536 * means that we have a GFP_ATOMIC allocation
2537 * failure risk. Hurry up!
2538 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002539 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002540 min_wmark_pages(zone), end_zone, 0))
2541 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002542 } else {
2543 /*
2544 * If a zone reaches its high watermark,
2545 * consider it to be no longer congested. It's
2546 * possible there are dirty pages backed by
2547 * congested BDIs but as pressure is relieved,
2548 * spectulatively avoid congestion waits
2549 */
2550 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002551 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002552 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002553 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002554
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002556 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 break; /* kswapd: all done */
2558 /*
2559 * OK, kswapd is getting into trouble. Take a nap, then take
2560 * another pass across the zones.
2561 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002562 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2563 if (has_under_min_watermark_zone)
2564 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2565 else
2566 congestion_wait(BLK_RW_ASYNC, HZ/10);
2567 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568
2569 /*
2570 * We do this so kswapd doesn't build up large priorities for
2571 * example when it is freeing in parallel with allocators. It
2572 * matches the direct reclaim path behaviour in terms of impact
2573 * on zone->*_priority.
2574 */
Rik van Riela79311c2009-01-06 14:40:01 -08002575 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 break;
2577 }
2578out:
Mel Gorman99504742011-01-13 15:46:20 -08002579
2580 /*
2581 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002582 * high-order: Balanced zones must make up at least 25% of the node
2583 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002584 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002585 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002587
2588 try_to_freeze();
2589
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002590 /*
2591 * Fragmentation may mean that the system cannot be
2592 * rebalanced for high-order allocations in all zones.
2593 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2594 * it means the zones have been fully scanned and are still
2595 * not balanced. For high-order allocations, there is
2596 * little point trying all over again as kswapd may
2597 * infinite loop.
2598 *
2599 * Instead, recheck all watermarks at order-0 as they
2600 * are the most important. If watermarks are ok, kswapd will go
2601 * back to sleep. High-order users can still perform direct
2602 * reclaim if they wish.
2603 */
2604 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2605 order = sc.order = 0;
2606
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 goto loop_again;
2608 }
2609
Mel Gorman99504742011-01-13 15:46:20 -08002610 /*
2611 * If kswapd was reclaiming at a higher order, it has the option of
2612 * sleeping without all zones being balanced. Before it does, it must
2613 * ensure that the watermarks for order-0 on *all* zones are met and
2614 * that the congestion flags are cleared. The congestion flag must
2615 * be cleared as kswapd is the only mechanism that clears the flag
2616 * and it is potentially going to sleep here.
2617 */
2618 if (order) {
2619 for (i = 0; i <= end_zone; i++) {
2620 struct zone *zone = pgdat->node_zones + i;
2621
2622 if (!populated_zone(zone))
2623 continue;
2624
2625 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2626 continue;
2627
2628 /* Confirm the zone is balanced for order-0 */
2629 if (!zone_watermark_ok(zone, 0,
2630 high_wmark_pages(zone), 0, 0)) {
2631 order = sc.order = 0;
2632 goto loop_again;
2633 }
2634
2635 /* If balanced, clear the congested flag */
2636 zone_clear_flag(zone, ZONE_CONGESTED);
2637 }
2638 }
2639
Mel Gorman0abdee22011-01-13 15:46:22 -08002640 /*
2641 * Return the order we were reclaiming at so sleeping_prematurely()
2642 * makes a decision on the order we were last reclaiming at. However,
2643 * if another caller entered the allocator slow path while kswapd
2644 * was awake, order will remain at the higher level
2645 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002646 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002647 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648}
2649
Mel Gormandc83edd2011-01-13 15:46:26 -08002650static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002651{
2652 long remaining = 0;
2653 DEFINE_WAIT(wait);
2654
2655 if (freezing(current) || kthread_should_stop())
2656 return;
2657
2658 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2659
2660 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002661 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002662 remaining = schedule_timeout(HZ/10);
2663 finish_wait(&pgdat->kswapd_wait, &wait);
2664 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2665 }
2666
2667 /*
2668 * After a short sleep, check if it was a premature sleep. If not, then
2669 * go fully to sleep until explicitly woken up.
2670 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002671 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002672 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2673
2674 /*
2675 * vmstat counters are not perfectly accurate and the estimated
2676 * value for counters such as NR_FREE_PAGES can deviate from the
2677 * true value by nr_online_cpus * threshold. To avoid the zone
2678 * watermarks being breached while under pressure, we reduce the
2679 * per-cpu vmstat threshold while kswapd is awake and restore
2680 * them before going back to sleep.
2681 */
2682 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2683 schedule();
2684 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2685 } else {
2686 if (remaining)
2687 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2688 else
2689 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2690 }
2691 finish_wait(&pgdat->kswapd_wait, &wait);
2692}
2693
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694/*
2695 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002696 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 *
2698 * This basically trickles out pages so that we have _some_
2699 * free memory available even if there is no other activity
2700 * that frees anything up. This is needed for things like routing
2701 * etc, where we otherwise might have all activity going on in
2702 * asynchronous contexts that cannot page things out.
2703 *
2704 * If there are applications that are active memory-allocators
2705 * (most normal use), this basically shouldn't matter.
2706 */
2707static int kswapd(void *p)
2708{
2709 unsigned long order;
Mel Gorman99504742011-01-13 15:46:20 -08002710 int classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 pg_data_t *pgdat = (pg_data_t*)p;
2712 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002713
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 struct reclaim_state reclaim_state = {
2715 .reclaimed_slab = 0,
2716 };
Rusty Russella70f7302009-03-13 14:49:46 +10302717 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718
Nick Piggincf40bd12009-01-21 08:12:39 +01002719 lockdep_set_current_reclaim_state(GFP_KERNEL);
2720
Rusty Russell174596a2009-01-01 10:12:29 +10302721 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002722 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 current->reclaim_state = &reclaim_state;
2724
2725 /*
2726 * Tell the memory management that we're a "memory allocator",
2727 * and that if we need more memory we should get access to it
2728 * regardless (see "__alloc_pages()"). "kswapd" should
2729 * never get caught in the normal page freeing logic.
2730 *
2731 * (Kswapd normally doesn't need memory anyway, but sometimes
2732 * you need a small amount of memory in order to be able to
2733 * page out something else, and this flag essentially protects
2734 * us from recursively trying to free more memory as we're
2735 * trying to free the first piece of memory in the first place).
2736 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002737 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002738 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739
2740 order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002741 classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 for ( ; ; ) {
2743 unsigned long new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002744 int new_classzone_idx;
David Rientjes8fe23e02009-12-14 17:58:33 -08002745 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002746
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 new_order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002748 new_classzone_idx = pgdat->classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 pgdat->kswapd_max_order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002750 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2751 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 /*
2753 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002754 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 */
2756 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002757 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 } else {
Mel Gormandc83edd2011-01-13 15:46:26 -08002759 kswapd_try_to_sleep(pgdat, order, classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002761 classzone_idx = pgdat->classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002762 pgdat->kswapd_max_order = 0;
2763 pgdat->classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
David Rientjes8fe23e02009-12-14 17:58:33 -08002766 ret = try_to_freeze();
2767 if (kthread_should_stop())
2768 break;
2769
2770 /*
2771 * We can speed up thawing tasks if we don't call balance_pgdat
2772 * after returning from the refrigerator
2773 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002774 if (!ret) {
2775 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Mel Gormandc83edd2011-01-13 15:46:26 -08002776 order = balance_pgdat(pgdat, order, &classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002777 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 }
2779 return 0;
2780}
2781
2782/*
2783 * A zone is low on free memory, so wake its kswapd task to service it.
2784 */
Mel Gorman99504742011-01-13 15:46:20 -08002785void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786{
2787 pg_data_t *pgdat;
2788
Con Kolivasf3fe6512006-01-06 00:11:15 -08002789 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 return;
2791
Paul Jackson02a0e532006-12-13 00:34:25 -08002792 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002794 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002795 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002796 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002797 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2798 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002799 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002801 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2802 return;
2803
2804 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002805 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806}
2807
Wu Fengguangadea02a2009-09-21 17:01:42 -07002808/*
2809 * The reclaimable count would be mostly accurate.
2810 * The less reclaimable pages may be
2811 * - mlocked pages, which will be moved to unevictable list when encountered
2812 * - mapped pages, which may require several travels to be reclaimed
2813 * - dirty pages, which is not "instantly" reclaimable
2814 */
2815unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002816{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002817 int nr;
2818
2819 nr = global_page_state(NR_ACTIVE_FILE) +
2820 global_page_state(NR_INACTIVE_FILE);
2821
2822 if (nr_swap_pages > 0)
2823 nr += global_page_state(NR_ACTIVE_ANON) +
2824 global_page_state(NR_INACTIVE_ANON);
2825
2826 return nr;
2827}
2828
2829unsigned long zone_reclaimable_pages(struct zone *zone)
2830{
2831 int nr;
2832
2833 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2834 zone_page_state(zone, NR_INACTIVE_FILE);
2835
2836 if (nr_swap_pages > 0)
2837 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2838 zone_page_state(zone, NR_INACTIVE_ANON);
2839
2840 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002841}
2842
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002843#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002845 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002846 * freed pages.
2847 *
2848 * Rather than trying to age LRUs the aim is to preserve the overall
2849 * LRU order by reclaiming preferentially
2850 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002852unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002854 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002855 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002856 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2857 .may_swap = 1,
2858 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002859 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002860 .nr_to_reclaim = nr_to_reclaim,
2861 .hibernation_mode = 1,
2862 .swappiness = vm_swappiness,
2863 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002865 struct shrink_control shrink = {
2866 .gfp_mask = sc.gfp_mask,
2867 };
2868 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002869 struct task_struct *p = current;
2870 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002872 p->flags |= PF_MEMALLOC;
2873 lockdep_set_current_reclaim_state(sc.gfp_mask);
2874 reclaim_state.reclaimed_slab = 0;
2875 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002876
Ying Hana09ed5e2011-05-24 17:12:26 -07002877 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002878
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002879 p->reclaim_state = NULL;
2880 lockdep_clear_current_reclaim_state();
2881 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002882
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002883 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002885#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887/* It's optimal to keep kswapds on the same CPUs as their memory, but
2888 not required for correctness. So if the last cpu in a node goes
2889 away, we get changed to run anywhere: as the first one comes back,
2890 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002891static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002892 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002894 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002896 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002897 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002898 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302899 const struct cpumask *mask;
2900
2901 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002902
Rusty Russell3e597942009-01-01 10:12:24 +10302903 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002905 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 }
2907 }
2908 return NOTIFY_OK;
2909}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910
Yasunori Goto3218ae12006-06-27 02:53:33 -07002911/*
2912 * This kswapd start function will be called by init and node-hot-add.
2913 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2914 */
2915int kswapd_run(int nid)
2916{
2917 pg_data_t *pgdat = NODE_DATA(nid);
2918 int ret = 0;
2919
2920 if (pgdat->kswapd)
2921 return 0;
2922
2923 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2924 if (IS_ERR(pgdat->kswapd)) {
2925 /* failure at boot is fatal */
2926 BUG_ON(system_state == SYSTEM_BOOTING);
2927 printk("Failed to start kswapd on node %d\n",nid);
2928 ret = -1;
2929 }
2930 return ret;
2931}
2932
David Rientjes8fe23e02009-12-14 17:58:33 -08002933/*
2934 * Called by memory hotplug when all memory in a node is offlined.
2935 */
2936void kswapd_stop(int nid)
2937{
2938 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2939
2940 if (kswapd)
2941 kthread_stop(kswapd);
2942}
2943
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944static int __init kswapd_init(void)
2945{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002946 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002947
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002949 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002950 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 hotcpu_notifier(cpu_callback, 0);
2952 return 0;
2953}
2954
2955module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002956
2957#ifdef CONFIG_NUMA
2958/*
2959 * Zone reclaim mode
2960 *
2961 * If non-zero call zone_reclaim when the number of free pages falls below
2962 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002963 */
2964int zone_reclaim_mode __read_mostly;
2965
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002966#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002967#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002968#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2969#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2970
Christoph Lameter9eeff232006-01-18 17:42:31 -08002971/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002972 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2973 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2974 * a zone.
2975 */
2976#define ZONE_RECLAIM_PRIORITY 4
2977
Christoph Lameter9eeff232006-01-18 17:42:31 -08002978/*
Christoph Lameter96146342006-07-03 00:24:13 -07002979 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2980 * occur.
2981 */
2982int sysctl_min_unmapped_ratio = 1;
2983
2984/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002985 * If the number of slab pages in a zone grows beyond this percentage then
2986 * slab reclaim needs to occur.
2987 */
2988int sysctl_min_slab_ratio = 5;
2989
Mel Gorman90afa5d2009-06-16 15:33:20 -07002990static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2991{
2992 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2993 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2994 zone_page_state(zone, NR_ACTIVE_FILE);
2995
2996 /*
2997 * It's possible for there to be more file mapped pages than
2998 * accounted for by the pages on the file LRU lists because
2999 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3000 */
3001 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3002}
3003
3004/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3005static long zone_pagecache_reclaimable(struct zone *zone)
3006{
3007 long nr_pagecache_reclaimable;
3008 long delta = 0;
3009
3010 /*
3011 * If RECLAIM_SWAP is set, then all file pages are considered
3012 * potentially reclaimable. Otherwise, we have to worry about
3013 * pages like swapcache and zone_unmapped_file_pages() provides
3014 * a better estimate
3015 */
3016 if (zone_reclaim_mode & RECLAIM_SWAP)
3017 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3018 else
3019 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3020
3021 /* If we can't clean pages, remove dirty pages from consideration */
3022 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3023 delta += zone_page_state(zone, NR_FILE_DIRTY);
3024
3025 /* Watch for any possible underflows due to delta */
3026 if (unlikely(delta > nr_pagecache_reclaimable))
3027 delta = nr_pagecache_reclaimable;
3028
3029 return nr_pagecache_reclaimable - delta;
3030}
3031
Christoph Lameter0ff38492006-09-25 23:31:52 -07003032/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003033 * Try to free up some pages from this zone through reclaim.
3034 */
Andrew Morton179e9632006-03-22 00:08:18 -08003035static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003036{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003037 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003038 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003039 struct task_struct *p = current;
3040 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08003041 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08003042 struct scan_control sc = {
3043 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003044 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003045 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003046 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3047 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003048 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003049 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003050 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08003051 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003052 struct shrink_control shrink = {
3053 .gfp_mask = sc.gfp_mask,
3054 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003055 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003056
Christoph Lameter9eeff232006-01-18 17:42:31 -08003057 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003058 /*
3059 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3060 * and we also need to be able to write out pages for RECLAIM_WRITE
3061 * and RECLAIM_SWAP.
3062 */
3063 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003064 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003065 reclaim_state.reclaimed_slab = 0;
3066 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003067
Mel Gorman90afa5d2009-06-16 15:33:20 -07003068 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003069 /*
3070 * Free memory by calling shrink zone with increasing
3071 * priorities until we have enough memory freed.
3072 */
3073 priority = ZONE_RECLAIM_PRIORITY;
3074 do {
Rik van Riela79311c2009-01-06 14:40:01 -08003075 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003076 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08003077 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003078 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003079
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003080 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3081 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003082 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003083 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003084 * many pages were freed in this zone. So we take the current
3085 * number of slab pages and shake the slab until it is reduced
3086 * by the same nr_pages that we used for reclaiming unmapped
3087 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003088 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003089 * Note that shrink_slab will free memory on all zones and may
3090 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003091 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003092 for (;;) {
3093 unsigned long lru_pages = zone_reclaimable_pages(zone);
3094
3095 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003096 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003097 break;
3098
3099 /* Freed enough memory */
3100 nr_slab_pages1 = zone_page_state(zone,
3101 NR_SLAB_RECLAIMABLE);
3102 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3103 break;
3104 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003105
3106 /*
3107 * Update nr_reclaimed by the number of slab pages we
3108 * reclaimed from this zone.
3109 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003110 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3111 if (nr_slab_pages1 < nr_slab_pages0)
3112 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003113 }
3114
Christoph Lameter9eeff232006-01-18 17:42:31 -08003115 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003116 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003117 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003118 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003119}
Andrew Morton179e9632006-03-22 00:08:18 -08003120
3121int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3122{
Andrew Morton179e9632006-03-22 00:08:18 -08003123 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003124 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003125
3126 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003127 * Zone reclaim reclaims unmapped file backed pages and
3128 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003129 *
Christoph Lameter96146342006-07-03 00:24:13 -07003130 * A small portion of unmapped file backed pages is needed for
3131 * file I/O otherwise pages read by file I/O will be immediately
3132 * thrown out if the zone is overallocated. So we do not reclaim
3133 * if less than a specified percentage of the zone is used by
3134 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003135 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003136 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3137 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003138 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003139
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003140 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003141 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003142
Andrew Morton179e9632006-03-22 00:08:18 -08003143 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003144 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003145 */
David Rientjesd773ed62007-10-16 23:26:01 -07003146 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003147 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003148
3149 /*
3150 * Only run zone reclaim on the local zone or on zones that do not
3151 * have associated processors. This will favor the local processor
3152 * over remote processors and spread off node memory allocations
3153 * as wide as possible.
3154 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003155 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003156 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003157 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003158
3159 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003160 return ZONE_RECLAIM_NOSCAN;
3161
David Rientjesd773ed62007-10-16 23:26:01 -07003162 ret = __zone_reclaim(zone, gfp_mask, order);
3163 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3164
Mel Gorman24cf725182009-06-16 15:33:23 -07003165 if (!ret)
3166 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3167
David Rientjesd773ed62007-10-16 23:26:01 -07003168 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003169}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003170#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003171
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003172/*
3173 * page_evictable - test whether a page is evictable
3174 * @page: the page to test
3175 * @vma: the VMA in which the page is or will be mapped, may be NULL
3176 *
3177 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003178 * lists vs unevictable list. The vma argument is !NULL when called from the
3179 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003180 *
3181 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003182 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003183 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003184 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003185 */
3186int page_evictable(struct page *page, struct vm_area_struct *vma)
3187{
3188
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003189 if (mapping_unevictable(page_mapping(page)))
3190 return 0;
3191
Nick Pigginb291f002008-10-18 20:26:44 -07003192 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3193 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003194
3195 return 1;
3196}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003197
3198/**
3199 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3200 * @page: page to check evictability and move to appropriate lru list
3201 * @zone: zone page is in
3202 *
3203 * Checks a page for evictability and moves the page to the appropriate
3204 * zone lru list.
3205 *
3206 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3207 * have PageUnevictable set.
3208 */
3209static void check_move_unevictable_page(struct page *page, struct zone *zone)
3210{
3211 VM_BUG_ON(PageActive(page));
3212
3213retry:
3214 ClearPageUnevictable(page);
3215 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003216 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003217
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003218 __dec_zone_state(zone, NR_UNEVICTABLE);
3219 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003220 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003221 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3222 __count_vm_event(UNEVICTABLE_PGRESCUED);
3223 } else {
3224 /*
3225 * rotate unevictable list
3226 */
3227 SetPageUnevictable(page);
3228 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003229 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003230 if (page_evictable(page, NULL))
3231 goto retry;
3232 }
3233}
3234
3235/**
3236 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3237 * @mapping: struct address_space to scan for evictable pages
3238 *
3239 * Scan all pages in mapping. Check unevictable pages for
3240 * evictability and move them to the appropriate zone lru list.
3241 */
3242void scan_mapping_unevictable_pages(struct address_space *mapping)
3243{
3244 pgoff_t next = 0;
3245 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3246 PAGE_CACHE_SHIFT;
3247 struct zone *zone;
3248 struct pagevec pvec;
3249
3250 if (mapping->nrpages == 0)
3251 return;
3252
3253 pagevec_init(&pvec, 0);
3254 while (next < end &&
3255 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3256 int i;
3257 int pg_scanned = 0;
3258
3259 zone = NULL;
3260
3261 for (i = 0; i < pagevec_count(&pvec); i++) {
3262 struct page *page = pvec.pages[i];
3263 pgoff_t page_index = page->index;
3264 struct zone *pagezone = page_zone(page);
3265
3266 pg_scanned++;
3267 if (page_index > next)
3268 next = page_index;
3269 next++;
3270
3271 if (pagezone != zone) {
3272 if (zone)
3273 spin_unlock_irq(&zone->lru_lock);
3274 zone = pagezone;
3275 spin_lock_irq(&zone->lru_lock);
3276 }
3277
3278 if (PageLRU(page) && PageUnevictable(page))
3279 check_move_unevictable_page(page, zone);
3280 }
3281 if (zone)
3282 spin_unlock_irq(&zone->lru_lock);
3283 pagevec_release(&pvec);
3284
3285 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3286 }
3287
3288}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003289
3290/**
3291 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3292 * @zone - zone of which to scan the unevictable list
3293 *
3294 * Scan @zone's unevictable LRU lists to check for pages that have become
3295 * evictable. Move those that have to @zone's inactive list where they
3296 * become candidates for reclaim, unless shrink_inactive_zone() decides
3297 * to reactivate them. Pages that are still unevictable are rotated
3298 * back onto @zone's unevictable list.
3299 */
3300#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08003301static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003302{
3303 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3304 unsigned long scan;
3305 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3306
3307 while (nr_to_scan > 0) {
3308 unsigned long batch_size = min(nr_to_scan,
3309 SCAN_UNEVICTABLE_BATCH_SIZE);
3310
3311 spin_lock_irq(&zone->lru_lock);
3312 for (scan = 0; scan < batch_size; scan++) {
3313 struct page *page = lru_to_page(l_unevictable);
3314
3315 if (!trylock_page(page))
3316 continue;
3317
3318 prefetchw_prev_lru_page(page, l_unevictable, flags);
3319
3320 if (likely(PageLRU(page) && PageUnevictable(page)))
3321 check_move_unevictable_page(page, zone);
3322
3323 unlock_page(page);
3324 }
3325 spin_unlock_irq(&zone->lru_lock);
3326
3327 nr_to_scan -= batch_size;
3328 }
3329}
3330
3331
3332/**
3333 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3334 *
3335 * A really big hammer: scan all zones' unevictable LRU lists to check for
3336 * pages that have become evictable. Move those back to the zones'
3337 * inactive list where they become candidates for reclaim.
3338 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3339 * and we add swap to the system. As such, it runs in the context of a task
3340 * that has possibly/probably made some previously unevictable pages
3341 * evictable.
3342 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003343static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003344{
3345 struct zone *zone;
3346
3347 for_each_zone(zone) {
3348 scan_zone_unevictable_pages(zone);
3349 }
3350}
3351
3352/*
3353 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3354 * all nodes' unevictable lists for evictable pages
3355 */
3356unsigned long scan_unevictable_pages;
3357
3358int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003359 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003360 size_t *length, loff_t *ppos)
3361{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003362 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003363
3364 if (write && *(unsigned long *)table->data)
3365 scan_all_zones_unevictable_pages();
3366
3367 scan_unevictable_pages = 0;
3368 return 0;
3369}
3370
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003371#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003372/*
3373 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3374 * a specified node's per zone unevictable lists for evictable pages.
3375 */
3376
3377static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3378 struct sysdev_attribute *attr,
3379 char *buf)
3380{
3381 return sprintf(buf, "0\n"); /* always zero; should fit... */
3382}
3383
3384static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3385 struct sysdev_attribute *attr,
3386 const char *buf, size_t count)
3387{
3388 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3389 struct zone *zone;
3390 unsigned long res;
3391 unsigned long req = strict_strtoul(buf, 10, &res);
3392
3393 if (!req)
3394 return 1; /* zero is no-op */
3395
3396 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3397 if (!populated_zone(zone))
3398 continue;
3399 scan_zone_unevictable_pages(zone);
3400 }
3401 return 1;
3402}
3403
3404
3405static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3406 read_scan_unevictable_node,
3407 write_scan_unevictable_node);
3408
3409int scan_unevictable_register_node(struct node *node)
3410{
3411 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3412}
3413
3414void scan_unevictable_unregister_node(struct node *node)
3415{
3416 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3417}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003418#endif