blob: c7f5a6d4b75bfff6cc10d7e9b91a79590cd3aea6 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45#include <asm/tlbflush.h>
46#include <asm/div64.h>
47
48#include <linux/swapops.h>
49
Nick Piggin0f8053a2006-03-22 00:08:33 -080050#include "internal.h"
51
Mel Gorman33906bc2010-08-09 17:19:16 -070052#define CREATE_TRACE_POINTS
53#include <trace/events/vmscan.h>
54
Mel Gormanee64fc92011-01-13 15:45:55 -080055/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080056 * reclaim_mode determines how the inactive list is shrunk
57 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
58 * RECLAIM_MODE_ASYNC: Do not block
59 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
60 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080061 * page from the LRU and reclaim all pages within a
62 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080063 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080064 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080065 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080066typedef unsigned __bitwise__ reclaim_mode_t;
67#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
68#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
69#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
70#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
71#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070072
Linus Torvalds1da177e2005-04-16 15:20:36 -070073struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 /* Incremented by the number of inactive pages that were scanned */
75 unsigned long nr_scanned;
76
Rik van Riela79311c2009-01-06 14:40:01 -080077 /* Number of pages freed so far during a call to shrink_zones() */
78 unsigned long nr_reclaimed;
79
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080080 /* How many pages shrink_list() should reclaim */
81 unsigned long nr_to_reclaim;
82
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080083 unsigned long hibernation_mode;
84
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040086 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88 int may_writepage;
89
Johannes Weinera6dc60f82009-03-31 15:19:30 -070090 /* Can mapped pages be reclaimed? */
91 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080092
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070093 /* Can pages be swapped as part of reclaim? */
94 int may_swap;
95
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070096 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070097
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070098 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080099
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700100 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700103 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800104 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700105
Balbir Singh66e17072008-02-07 00:13:56 -0800106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
108
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700109 /*
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
111 * are scanned.
112 */
113 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114};
115
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118#ifdef ARCH_HAS_PREFETCH
119#define prefetch_prev_lru_page(_page, _base, _field) \
120 do { \
121 if ((_page)->lru.prev != _base) { \
122 struct page *prev; \
123 \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
126 } \
127 } while (0)
128#else
129#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130#endif
131
132#ifdef ARCH_HAS_PREFETCHW
133#define prefetchw_prev_lru_page(_page, _base, _field) \
134 do { \
135 if ((_page)->lru.prev != _base) { \
136 struct page *prev; \
137 \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
140 } \
141 } while (0)
142#else
143#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144#endif
145
146/*
147 * From 0 .. 100. Higher means more swappy.
148 */
149int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700150long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
152static LIST_HEAD(shrinker_list);
153static DECLARE_RWSEM(shrinker_rwsem);
154
Balbir Singh00f0b822008-03-04 14:28:39 -0800155#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800156#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800157#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800158#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#endif
160
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800161static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
163{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800164 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800167 return &zone->reclaim_stat;
168}
169
Vincent Li0b217672009-09-21 17:03:09 -0700170static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800172{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800173 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800174 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
175
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800176 return zone_page_state(zone, NR_LRU_BASE + lru);
177}
178
179
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180/*
181 * Add a shrinker callback to be called from the vm
182 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700183void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700185 shrinker->nr = 0;
186 down_write(&shrinker_rwsem);
187 list_add_tail(&shrinker->list, &shrinker_list);
188 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700190EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
192/*
193 * Remove one
194 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700195void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196{
197 down_write(&shrinker_rwsem);
198 list_del(&shrinker->list);
199 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700201EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202
203#define SHRINK_BATCH 128
204/*
205 * Call the shrink functions to age shrinkable caches
206 *
207 * Here we assume it costs one seek to replace a lru page and that it also
208 * takes a seek to recreate a cache object. With this in mind we age equal
209 * percentages of the lru and ageable caches. This should balance the seeks
210 * generated by these structures.
211 *
Simon Arlott183ff222007-10-20 01:27:18 +0200212 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 * slab to avoid swapping.
214 *
215 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
216 *
217 * `lru_pages' represents the number of on-LRU pages in all the zones which
218 * are eligible for the caller's allocation attempt. It is used for balancing
219 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700220 *
221 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 */
Andrew Morton69e05942006-03-22 00:08:19 -0800223unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
224 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225{
226 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800227 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
229 if (scanned == 0)
230 scanned = SWAP_CLUSTER_MAX;
231
232 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700233 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234
235 list_for_each_entry(shrinker, &shrinker_list, list) {
236 unsigned long long delta;
237 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000238 unsigned long max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Dave Chinner7f8275d2010-07-19 14:56:17 +1000240 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800242 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 do_div(delta, lru_pages + 1);
244 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800245 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700246 printk(KERN_ERR "shrink_slab: %pF negative objects to "
247 "delete nr=%ld\n",
248 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800249 shrinker->nr = max_pass;
250 }
251
252 /*
253 * Avoid risking looping forever due to too large nr value:
254 * never try to free more than twice the estimate number of
255 * freeable entries.
256 */
257 if (shrinker->nr > max_pass * 2)
258 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259
260 total_scan = shrinker->nr;
261 shrinker->nr = 0;
262
263 while (total_scan >= SHRINK_BATCH) {
264 long this_scan = SHRINK_BATCH;
265 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700266 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Dave Chinner7f8275d2010-07-19 14:56:17 +1000268 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
269 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
270 gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 if (shrink_ret == -1)
272 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700273 if (shrink_ret < nr_before)
274 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700275 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 total_scan -= this_scan;
277
278 cond_resched();
279 }
280
281 shrinker->nr += total_scan;
282 }
283 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700284 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
286
Mel Gormanf3a310b2011-01-13 15:46:00 -0800287static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700288 bool sync)
289{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800290 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700291
292 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800293 * Initially assume we are entering either lumpy reclaim or
294 * reclaim/compaction.Depending on the order, we will either set the
295 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700296 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800297 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800298 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800299 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800300 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700301
302 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800303 * Avoid using lumpy reclaim or reclaim/compaction if possible by
304 * restricting when its set to either costly allocations or when
305 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700306 */
307 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800308 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700309 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800310 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700311 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800312 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700313}
314
Mel Gormanf3a310b2011-01-13 15:46:00 -0800315static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700316{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800317 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700318}
319
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320static inline int is_page_cache_freeable(struct page *page)
321{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700322 /*
323 * A freeable page cache page is referenced only by the caller
324 * that isolated the page, the page cache radix tree and
325 * optional buffer heads at page->private.
326 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700327 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328}
329
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700330static int may_write_to_queue(struct backing_dev_info *bdi,
331 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332{
Christoph Lameter930d9152006-01-08 01:00:47 -0800333 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 return 1;
335 if (!bdi_write_congested(bdi))
336 return 1;
337 if (bdi == current->backing_dev_info)
338 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700339
340 /* lumpy reclaim for hugepage often need a lot of write */
341 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
342 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 return 0;
344}
345
346/*
347 * We detected a synchronous write error writing a page out. Probably
348 * -ENOSPC. We need to propagate that into the address_space for a subsequent
349 * fsync(), msync() or close().
350 *
351 * The tricky part is that after writepage we cannot touch the mapping: nothing
352 * prevents it from being freed up. But we have a ref on the page and once
353 * that page is locked, the mapping is pinned.
354 *
355 * We're allowed to run sleeping lock_page() here because we know the caller has
356 * __GFP_FS.
357 */
358static void handle_write_error(struct address_space *mapping,
359 struct page *page, int error)
360{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100361 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700362 if (page_mapping(page) == mapping)
363 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 unlock_page(page);
365}
366
Christoph Lameter04e62a22006-06-23 02:03:38 -0700367/* possible outcome of pageout() */
368typedef enum {
369 /* failed to write page out, page is locked */
370 PAGE_KEEP,
371 /* move page to the active list, page is locked */
372 PAGE_ACTIVATE,
373 /* page has been sent to the disk successfully, page is unlocked */
374 PAGE_SUCCESS,
375 /* page is clean and locked */
376 PAGE_CLEAN,
377} pageout_t;
378
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379/*
Andrew Morton1742f192006-03-22 00:08:21 -0800380 * pageout is called by shrink_page_list() for each dirty page.
381 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700383static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700384 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
386 /*
387 * If the page is dirty, only perform writeback if that write
388 * will be non-blocking. To prevent this allocation from being
389 * stalled by pagecache activity. But note that there may be
390 * stalls if we need to run get_block(). We could test
391 * PagePrivate for that.
392 *
Vincent Li6aceb532009-12-14 17:58:49 -0800393 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 * this page's queue, we can perform writeback even if that
395 * will block.
396 *
397 * If the page is swapcache, write it back even if that would
398 * block, for some throttling. This happens by accident, because
399 * swap_backing_dev_info is bust: it doesn't reflect the
400 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 */
402 if (!is_page_cache_freeable(page))
403 return PAGE_KEEP;
404 if (!mapping) {
405 /*
406 * Some data journaling orphaned pages can have
407 * page->mapping == NULL while being dirty with clean buffers.
408 */
David Howells266cf652009-04-03 16:42:36 +0100409 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 if (try_to_free_buffers(page)) {
411 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700412 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 return PAGE_CLEAN;
414 }
415 }
416 return PAGE_KEEP;
417 }
418 if (mapping->a_ops->writepage == NULL)
419 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700420 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 return PAGE_KEEP;
422
423 if (clear_page_dirty_for_io(page)) {
424 int res;
425 struct writeback_control wbc = {
426 .sync_mode = WB_SYNC_NONE,
427 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700428 .range_start = 0,
429 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 .for_reclaim = 1,
431 };
432
433 SetPageReclaim(page);
434 res = mapping->a_ops->writepage(page, &wbc);
435 if (res < 0)
436 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800437 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 ClearPageReclaim(page);
439 return PAGE_ACTIVATE;
440 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700441
442 /*
443 * Wait on writeback if requested to. This happens when
444 * direct reclaiming a large contiguous area and the
445 * first attempt to free a range of pages fails.
446 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700447 if (PageWriteback(page) &&
Mel Gormanf3a310b2011-01-13 15:46:00 -0800448 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
Andy Whitcroftc661b072007-08-22 14:01:26 -0700449 wait_on_page_writeback(page);
450
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 if (!PageWriteback(page)) {
452 /* synchronous write or broken a_ops? */
453 ClearPageReclaim(page);
454 }
Mel Gorman755f0222010-08-09 17:19:18 -0700455 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800456 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700457 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 return PAGE_SUCCESS;
459 }
460
461 return PAGE_CLEAN;
462}
463
Andrew Mortona649fd92006-10-17 00:09:36 -0700464/*
Nick Piggine2867812008-07-25 19:45:30 -0700465 * Same as remove_mapping, but if the page is removed from the mapping, it
466 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700467 */
Nick Piggine2867812008-07-25 19:45:30 -0700468static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800469{
Nick Piggin28e4d962006-09-25 23:31:23 -0700470 BUG_ON(!PageLocked(page));
471 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800472
Nick Piggin19fd6232008-07-25 19:45:32 -0700473 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800474 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700475 * The non racy check for a busy page.
476 *
477 * Must be careful with the order of the tests. When someone has
478 * a ref to the page, it may be possible that they dirty it then
479 * drop the reference. So if PageDirty is tested before page_count
480 * here, then the following race may occur:
481 *
482 * get_user_pages(&page);
483 * [user mapping goes away]
484 * write_to(page);
485 * !PageDirty(page) [good]
486 * SetPageDirty(page);
487 * put_page(page);
488 * !page_count(page) [good, discard it]
489 *
490 * [oops, our write_to data is lost]
491 *
492 * Reversing the order of the tests ensures such a situation cannot
493 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
494 * load is not satisfied before that of page->_count.
495 *
496 * Note that if SetPageDirty is always performed via set_page_dirty,
497 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800498 */
Nick Piggine2867812008-07-25 19:45:30 -0700499 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800500 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700501 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
502 if (unlikely(PageDirty(page))) {
503 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800504 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700505 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800506
507 if (PageSwapCache(page)) {
508 swp_entry_t swap = { .val = page_private(page) };
509 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700510 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700511 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700512 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500513 void (*freepage)(struct page *);
514
515 freepage = mapping->a_ops->freepage;
516
Minchan Kime64a7822011-03-22 16:32:44 -0700517 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700518 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700519 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500520
521 if (freepage != NULL)
522 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800523 }
524
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800525 return 1;
526
527cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700528 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800529 return 0;
530}
531
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532/*
Nick Piggine2867812008-07-25 19:45:30 -0700533 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
534 * someone else has a ref on the page, abort and return 0. If it was
535 * successfully detached, return 1. Assumes the caller has a single ref on
536 * this page.
537 */
538int remove_mapping(struct address_space *mapping, struct page *page)
539{
540 if (__remove_mapping(mapping, page)) {
541 /*
542 * Unfreezing the refcount with 1 rather than 2 effectively
543 * drops the pagecache ref for us without requiring another
544 * atomic operation.
545 */
546 page_unfreeze_refs(page, 1);
547 return 1;
548 }
549 return 0;
550}
551
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700552/**
553 * putback_lru_page - put previously isolated page onto appropriate LRU list
554 * @page: page to be put back to appropriate lru list
555 *
556 * Add previously isolated @page to appropriate LRU list.
557 * Page may still be unevictable for other reasons.
558 *
559 * lru_lock must not be held, interrupts must be enabled.
560 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700561void putback_lru_page(struct page *page)
562{
563 int lru;
564 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700565 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700566
567 VM_BUG_ON(PageLRU(page));
568
569redo:
570 ClearPageUnevictable(page);
571
572 if (page_evictable(page, NULL)) {
573 /*
574 * For evictable pages, we can use the cache.
575 * In event of a race, worst case is we end up with an
576 * unevictable page on [in]active list.
577 * We know how to handle that.
578 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700579 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700580 lru_cache_add_lru(page, lru);
581 } else {
582 /*
583 * Put unevictable pages directly on zone's unevictable
584 * list.
585 */
586 lru = LRU_UNEVICTABLE;
587 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700588 /*
589 * When racing with an mlock clearing (page is
590 * unlocked), make sure that if the other thread does
591 * not observe our setting of PG_lru and fails
592 * isolation, we see PG_mlocked cleared below and move
593 * the page back to the evictable list.
594 *
595 * The other side is TestClearPageMlocked().
596 */
597 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700598 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700599
600 /*
601 * page's status can change while we move it among lru. If an evictable
602 * page is on unevictable list, it never be freed. To avoid that,
603 * check after we added it to the list, again.
604 */
605 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606 if (!isolate_lru_page(page)) {
607 put_page(page);
608 goto redo;
609 }
610 /* This means someone else dropped this page from LRU
611 * So, it will be freed or putback to LRU again. There is
612 * nothing to do here.
613 */
614 }
615
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700616 if (was_unevictable && lru != LRU_UNEVICTABLE)
617 count_vm_event(UNEVICTABLE_PGRESCUED);
618 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619 count_vm_event(UNEVICTABLE_PGCULLED);
620
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700621 put_page(page); /* drop ref from isolate */
622}
623
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800624enum page_references {
625 PAGEREF_RECLAIM,
626 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800627 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800628 PAGEREF_ACTIVATE,
629};
630
631static enum page_references page_check_references(struct page *page,
632 struct scan_control *sc)
633{
Johannes Weiner645747462010-03-05 13:42:22 -0800634 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800635 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800636
Johannes Weiner645747462010-03-05 13:42:22 -0800637 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
638 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800639
640 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800641 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800642 return PAGEREF_RECLAIM;
643
644 /*
645 * Mlock lost the isolation race with us. Let try_to_unmap()
646 * move the page to the unevictable list.
647 */
648 if (vm_flags & VM_LOCKED)
649 return PAGEREF_RECLAIM;
650
Johannes Weiner645747462010-03-05 13:42:22 -0800651 if (referenced_ptes) {
652 if (PageAnon(page))
653 return PAGEREF_ACTIVATE;
654 /*
655 * All mapped pages start out with page table
656 * references from the instantiating fault, so we need
657 * to look twice if a mapped file page is used more
658 * than once.
659 *
660 * Mark it and spare it for another trip around the
661 * inactive list. Another page table reference will
662 * lead to its activation.
663 *
664 * Note: the mark is set for activated pages as well
665 * so that recently deactivated but used pages are
666 * quickly recovered.
667 */
668 SetPageReferenced(page);
669
670 if (referenced_page)
671 return PAGEREF_ACTIVATE;
672
673 return PAGEREF_KEEP;
674 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800675
676 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700677 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800678 return PAGEREF_RECLAIM_CLEAN;
679
680 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800681}
682
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700683static noinline_for_stack void free_page_list(struct list_head *free_pages)
684{
685 struct pagevec freed_pvec;
686 struct page *page, *tmp;
687
688 pagevec_init(&freed_pvec, 1);
689
690 list_for_each_entry_safe(page, tmp, free_pages, lru) {
691 list_del(&page->lru);
692 if (!pagevec_add(&freed_pvec, page)) {
693 __pagevec_free(&freed_pvec);
694 pagevec_reinit(&freed_pvec);
695 }
696 }
697
698 pagevec_free(&freed_pvec);
699}
700
Nick Piggine2867812008-07-25 19:45:30 -0700701/*
Andrew Morton1742f192006-03-22 00:08:21 -0800702 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 */
Andrew Morton1742f192006-03-22 00:08:21 -0800704static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700705 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700706 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707{
708 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700709 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700711 unsigned long nr_dirty = 0;
712 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800713 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714
715 cond_resched();
716
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800718 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 struct address_space *mapping;
720 struct page *page;
721 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
723 cond_resched();
724
725 page = lru_to_page(page_list);
726 list_del(&page->lru);
727
Nick Piggin529ae9a2008-08-02 12:01:03 +0200728 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 goto keep;
730
Nick Piggin725d7042006-09-25 23:30:55 -0700731 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700732 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
734 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800735
Nick Pigginb291f002008-10-18 20:26:44 -0700736 if (unlikely(!page_evictable(page, NULL)))
737 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700738
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700739 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800740 goto keep_locked;
741
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 /* Double the slab pressure for mapped and swapcache pages */
743 if (page_mapped(page) || PageSwapCache(page))
744 sc->nr_scanned++;
745
Andy Whitcroftc661b072007-08-22 14:01:26 -0700746 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
748
749 if (PageWriteback(page)) {
750 /*
751 * Synchronous reclaim is performed in two passes,
752 * first an asynchronous pass over the list to
753 * start parallel writeback, and a second synchronous
754 * pass to wait for the IO to complete. Wait here
755 * for any page for which writeback has already
756 * started.
757 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800758 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700759 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700760 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700761 else {
762 unlock_page(page);
763 goto keep_lumpy;
764 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800767 references = page_check_references(page, sc);
768 switch (references) {
769 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800771 case PAGEREF_KEEP:
772 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800773 case PAGEREF_RECLAIM:
774 case PAGEREF_RECLAIM_CLEAN:
775 ; /* try to reclaim the page below */
776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 /*
779 * Anonymous process memory has backing store?
780 * Try to allocate it some swap space here.
781 */
Nick Pigginb291f002008-10-18 20:26:44 -0700782 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800783 if (!(sc->gfp_mask & __GFP_IO))
784 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800785 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800787 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700788 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
790 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
792 /*
793 * The page is mapped into the page tables of one or more
794 * processes. Try to unmap it here.
795 */
796 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200797 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 case SWAP_FAIL:
799 goto activate_locked;
800 case SWAP_AGAIN:
801 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700802 case SWAP_MLOCK:
803 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 case SWAP_SUCCESS:
805 ; /* try to free the page below */
806 }
807 }
808
809 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700810 nr_dirty++;
811
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800812 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700814 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800816 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 goto keep_locked;
818
819 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700820 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700822 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 goto keep_locked;
824 case PAGE_ACTIVATE:
825 goto activate_locked;
826 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700827 if (PageWriteback(page))
828 goto keep_lumpy;
829 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700831
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 /*
833 * A synchronous write - probably a ramdisk. Go
834 * ahead and try to reclaim the page.
835 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200836 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 goto keep;
838 if (PageDirty(page) || PageWriteback(page))
839 goto keep_locked;
840 mapping = page_mapping(page);
841 case PAGE_CLEAN:
842 ; /* try to free the page below */
843 }
844 }
845
846 /*
847 * If the page has buffers, try to free the buffer mappings
848 * associated with this page. If we succeed we try to free
849 * the page as well.
850 *
851 * We do this even if the page is PageDirty().
852 * try_to_release_page() does not perform I/O, but it is
853 * possible for a page to have PageDirty set, but it is actually
854 * clean (all its buffers are clean). This happens if the
855 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700856 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 * try_to_release_page() will discover that cleanness and will
858 * drop the buffers and mark the page clean - it can be freed.
859 *
860 * Rarely, pages can have buffers and no ->mapping. These are
861 * the pages which were not successfully invalidated in
862 * truncate_complete_page(). We try to drop those buffers here
863 * and if that worked, and the page is no longer mapped into
864 * process address space (page_count == 1) it can be freed.
865 * Otherwise, leave the page on the LRU so it is swappable.
866 */
David Howells266cf652009-04-03 16:42:36 +0100867 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 if (!try_to_release_page(page, sc->gfp_mask))
869 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700870 if (!mapping && page_count(page) == 1) {
871 unlock_page(page);
872 if (put_page_testzero(page))
873 goto free_it;
874 else {
875 /*
876 * rare race with speculative reference.
877 * the speculative reference will free
878 * this page shortly, so we may
879 * increment nr_reclaimed here (and
880 * leave it off the LRU).
881 */
882 nr_reclaimed++;
883 continue;
884 }
885 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 }
887
Nick Piggine2867812008-07-25 19:45:30 -0700888 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800889 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890
Nick Piggina978d6f2008-10-18 20:26:58 -0700891 /*
892 * At this point, we have no other references and there is
893 * no way to pick any more up (removed from LRU, removed
894 * from pagecache). Can use non-atomic bitops now (and
895 * we obviously don't have to worry about waking up a process
896 * waiting on the page lock, because there are no references.
897 */
898 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700899free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800900 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700901
902 /*
903 * Is there need to periodically free_page_list? It would
904 * appear not as the counts should be low
905 */
906 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 continue;
908
Nick Pigginb291f002008-10-18 20:26:44 -0700909cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800910 if (PageSwapCache(page))
911 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700912 unlock_page(page);
913 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800914 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700915 continue;
916
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700918 /* Not a candidate for swapping, so reclaim swap space. */
919 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800920 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700921 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 SetPageActive(page);
923 pgactivate++;
924keep_locked:
925 unlock_page(page);
926keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -0800927 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700928keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700930 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700932
Mel Gorman0e093d992010-10-26 14:21:45 -0700933 /*
934 * Tag a zone as congested if all the dirty pages encountered were
935 * backed by a congested BDI. In this case, reclaimers should just
936 * back off and wait for congestion to clear because further reclaim
937 * will encounter the same problem
938 */
Shaohua Li1dce071e2010-11-11 14:05:17 -0800939 if (nr_dirty == nr_congested && nr_dirty != 0)
Mel Gorman0e093d992010-10-26 14:21:45 -0700940 zone_set_flag(zone, ZONE_CONGESTED);
941
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700942 free_page_list(&free_pages);
943
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700945 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800946 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947}
948
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700949/*
950 * Attempt to remove the specified page from its LRU. Only take this page
951 * if it is of the appropriate PageActive status. Pages which are being
952 * freed elsewhere are also ignored.
953 *
954 * page: page to consider
955 * mode: one of the LRU isolation modes defined above
956 *
957 * returns 0 on success, -ve errno on failure.
958 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700959int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700960{
961 int ret = -EINVAL;
962
963 /* Only take pages on the LRU. */
964 if (!PageLRU(page))
965 return ret;
966
967 /*
968 * When checking the active state, we need to be sure we are
969 * dealing with comparible boolean values. Take the logical not
970 * of each.
971 */
972 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
973 return ret;
974
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700975 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700976 return ret;
977
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700978 /*
979 * When this function is being called for lumpy reclaim, we
980 * initially look into all LRU pages, active, inactive and
981 * unevictable; only give shrink_page_list evictable pages.
982 */
983 if (PageUnevictable(page))
984 return ret;
985
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700986 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800987
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700988 if (likely(get_page_unless_zero(page))) {
989 /*
990 * Be careful not to clear PageLRU until after we're
991 * sure the page is not being freed elsewhere -- the
992 * page release code relies on it.
993 */
994 ClearPageLRU(page);
995 ret = 0;
996 }
997
998 return ret;
999}
1000
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 * zone->lru_lock is heavily contended. Some of the functions that
1003 * shrink the lists perform better by taking out a batch of pages
1004 * and working on them outside the LRU lock.
1005 *
1006 * For pagecache intensive workloads, this function is the hottest
1007 * spot in the kernel (apart from copy_*_user functions).
1008 *
1009 * Appropriate locks must be held before calling this function.
1010 *
1011 * @nr_to_scan: The number of pages to look through on the list.
1012 * @src: The LRU list to pull pages off.
1013 * @dst: The temp list to put pages on to.
1014 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001015 * @order: The caller's attempted allocation order
1016 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001017 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 *
1019 * returns how many pages were moved onto *@dst.
1020 */
Andrew Morton69e05942006-03-22 00:08:19 -08001021static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1022 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001023 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024{
Andrew Morton69e05942006-03-22 00:08:19 -08001025 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001026 unsigned long nr_lumpy_taken = 0;
1027 unsigned long nr_lumpy_dirty = 0;
1028 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001029 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001031 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001032 struct page *page;
1033 unsigned long pfn;
1034 unsigned long end_pfn;
1035 unsigned long page_pfn;
1036 int zone_id;
1037
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 page = lru_to_page(src);
1039 prefetchw_prev_lru_page(page, src, flags);
1040
Nick Piggin725d7042006-09-25 23:30:55 -07001041 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001042
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001043 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001044 case 0:
1045 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001046 mem_cgroup_del_lru(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001047 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001048 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001049
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001050 case -EBUSY:
1051 /* else it is being freed elsewhere */
1052 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001053 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001054 continue;
1055
1056 default:
1057 BUG();
1058 }
1059
1060 if (!order)
1061 continue;
1062
1063 /*
1064 * Attempt to take all pages in the order aligned region
1065 * surrounding the tag page. Only take those pages of
1066 * the same active state as that tag page. We may safely
1067 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001068 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001069 * where that page is in a different zone we will detect
1070 * it from its zone id and abort this block scan.
1071 */
1072 zone_id = page_zone_id(page);
1073 page_pfn = page_to_pfn(page);
1074 pfn = page_pfn & ~((1 << order) - 1);
1075 end_pfn = pfn + (1 << order);
1076 for (; pfn < end_pfn; pfn++) {
1077 struct page *cursor_page;
1078
1079 /* The target page is in the block, ignore it. */
1080 if (unlikely(pfn == page_pfn))
1081 continue;
1082
1083 /* Avoid holes within the zone. */
1084 if (unlikely(!pfn_valid_within(pfn)))
1085 break;
1086
1087 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001088
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001089 /* Check that we have not crossed a zone boundary. */
1090 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001091 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001092
1093 /*
1094 * If we don't have enough swap space, reclaiming of
1095 * anon page which don't already have a swap slot is
1096 * pointless.
1097 */
1098 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001099 !PageSwapCache(cursor_page))
1100 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001101
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001102 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001103 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001104 mem_cgroup_del_lru(cursor_page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001105 nr_taken += hpage_nr_pages(page);
Mel Gormana8a94d12010-08-09 17:19:17 -07001106 nr_lumpy_taken++;
1107 if (PageDirty(cursor_page))
1108 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001109 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001110 } else {
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001111 /* the page is freed already. */
1112 if (!page_count(cursor_page))
1113 continue;
1114 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001115 }
1116 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001117
1118 /* If we break out of the loop above, lumpy reclaim failed */
1119 if (pfn < end_pfn)
1120 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 }
1122
1123 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001124
1125 trace_mm_vmscan_lru_isolate(order,
1126 nr_to_scan, scan,
1127 nr_taken,
1128 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1129 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 return nr_taken;
1131}
1132
Balbir Singh66e17072008-02-07 00:13:56 -08001133static unsigned long isolate_pages_global(unsigned long nr,
1134 struct list_head *dst,
1135 unsigned long *scanned, int order,
1136 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001137 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001138{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001139 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001140 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001141 lru += LRU_ACTIVE;
1142 if (file)
1143 lru += LRU_FILE;
1144 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001145 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001146}
1147
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001149 * clear_active_flags() is a helper for shrink_active_list(), clearing
1150 * any active bits from the pages in the list.
1151 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001152static unsigned long clear_active_flags(struct list_head *page_list,
1153 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001154{
1155 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001156 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001157 struct page *page;
1158
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001159 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001160 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001161 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001162 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001163 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001164 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001165 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001166 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001167 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001168 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001169 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001170
1171 return nr_active;
1172}
1173
Nick Piggin62695a82008-10-18 20:26:09 -07001174/**
1175 * isolate_lru_page - tries to isolate a page from its LRU list
1176 * @page: page to isolate from its LRU list
1177 *
1178 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1179 * vmstat statistic corresponding to whatever LRU list the page was on.
1180 *
1181 * Returns 0 if the page was removed from an LRU list.
1182 * Returns -EBUSY if the page was not on an LRU list.
1183 *
1184 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001185 * the active list, it will have PageActive set. If it was found on
1186 * the unevictable list, it will have the PageUnevictable bit set. That flag
1187 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001188 *
1189 * The vmstat statistic corresponding to the list on which the page was
1190 * found will be decremented.
1191 *
1192 * Restrictions:
1193 * (1) Must be called with an elevated refcount on the page. This is a
1194 * fundamentnal difference from isolate_lru_pages (which is called
1195 * without a stable reference).
1196 * (2) the lru_lock must not be held.
1197 * (3) interrupts must be enabled.
1198 */
1199int isolate_lru_page(struct page *page)
1200{
1201 int ret = -EBUSY;
1202
1203 if (PageLRU(page)) {
1204 struct zone *zone = page_zone(page);
1205
1206 spin_lock_irq(&zone->lru_lock);
1207 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001208 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001209 ret = 0;
1210 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001211
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001212 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001213 }
1214 spin_unlock_irq(&zone->lru_lock);
1215 }
1216 return ret;
1217}
1218
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001219/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001220 * Are there way too many processes in the direct reclaim path already?
1221 */
1222static int too_many_isolated(struct zone *zone, int file,
1223 struct scan_control *sc)
1224{
1225 unsigned long inactive, isolated;
1226
1227 if (current_is_kswapd())
1228 return 0;
1229
1230 if (!scanning_global_lru(sc))
1231 return 0;
1232
1233 if (file) {
1234 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1235 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1236 } else {
1237 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1238 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1239 }
1240
1241 return isolated > inactive;
1242}
1243
1244/*
Mel Gorman66635622010-08-09 17:19:30 -07001245 * TODO: Try merging with migrations version of putback_lru_pages
1246 */
1247static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001248putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001249 unsigned long nr_anon, unsigned long nr_file,
1250 struct list_head *page_list)
1251{
1252 struct page *page;
1253 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001254 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001255
1256 pagevec_init(&pvec, 1);
1257
1258 /*
1259 * Put back any unfreeable pages.
1260 */
1261 spin_lock(&zone->lru_lock);
1262 while (!list_empty(page_list)) {
1263 int lru;
1264 page = lru_to_page(page_list);
1265 VM_BUG_ON(PageLRU(page));
1266 list_del(&page->lru);
1267 if (unlikely(!page_evictable(page, NULL))) {
1268 spin_unlock_irq(&zone->lru_lock);
1269 putback_lru_page(page);
1270 spin_lock_irq(&zone->lru_lock);
1271 continue;
1272 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001273 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001274 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001275 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001276 if (is_active_lru(lru)) {
1277 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001278 int numpages = hpage_nr_pages(page);
1279 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001280 }
1281 if (!pagevec_add(&pvec, page)) {
1282 spin_unlock_irq(&zone->lru_lock);
1283 __pagevec_release(&pvec);
1284 spin_lock_irq(&zone->lru_lock);
1285 }
1286 }
1287 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1288 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1289
1290 spin_unlock_irq(&zone->lru_lock);
1291 pagevec_release(&pvec);
1292}
1293
Mel Gorman1489fa12010-08-09 17:19:33 -07001294static noinline_for_stack void update_isolated_counts(struct zone *zone,
1295 struct scan_control *sc,
1296 unsigned long *nr_anon,
1297 unsigned long *nr_file,
1298 struct list_head *isolated_list)
1299{
1300 unsigned long nr_active;
1301 unsigned int count[NR_LRU_LISTS] = { 0, };
1302 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1303
1304 nr_active = clear_active_flags(isolated_list, count);
1305 __count_vm_events(PGDEACTIVATE, nr_active);
1306
1307 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1308 -count[LRU_ACTIVE_FILE]);
1309 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1310 -count[LRU_INACTIVE_FILE]);
1311 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1312 -count[LRU_ACTIVE_ANON]);
1313 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1314 -count[LRU_INACTIVE_ANON]);
1315
1316 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1317 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1318 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1319 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1320
1321 reclaim_stat->recent_scanned[0] += *nr_anon;
1322 reclaim_stat->recent_scanned[1] += *nr_file;
1323}
1324
Mel Gorman66635622010-08-09 17:19:30 -07001325/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001326 * Returns true if the caller should wait to clean dirty/writeback pages.
1327 *
1328 * If we are direct reclaiming for contiguous pages and we do not reclaim
1329 * everything in the list, try again and wait for writeback IO to complete.
1330 * This will stall high-order allocations noticeably. Only do that when really
1331 * need to free the pages under high memory pressure.
1332 */
1333static inline bool should_reclaim_stall(unsigned long nr_taken,
1334 unsigned long nr_freed,
1335 int priority,
1336 struct scan_control *sc)
1337{
1338 int lumpy_stall_priority;
1339
1340 /* kswapd should not stall on sync IO */
1341 if (current_is_kswapd())
1342 return false;
1343
1344 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001345 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001346 return false;
1347
1348 /* If we have relaimed everything on the isolated list, no stall */
1349 if (nr_freed == nr_taken)
1350 return false;
1351
1352 /*
1353 * For high-order allocations, there are two stall thresholds.
1354 * High-cost allocations stall immediately where as lower
1355 * order allocations such as stacks require the scanning
1356 * priority to be much higher before stalling.
1357 */
1358 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1359 lumpy_stall_priority = DEF_PRIORITY;
1360 else
1361 lumpy_stall_priority = DEF_PRIORITY / 3;
1362
1363 return priority <= lumpy_stall_priority;
1364}
1365
1366/*
Andrew Morton1742f192006-03-22 00:08:21 -08001367 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1368 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 */
Mel Gorman66635622010-08-09 17:19:30 -07001370static noinline_for_stack unsigned long
1371shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1372 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373{
1374 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001375 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001376 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001377 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001378 unsigned long nr_anon;
1379 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001380
Rik van Riel35cd7812009-09-21 17:01:38 -07001381 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001382 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001383
1384 /* We are about to die and free our memory. Return now. */
1385 if (fatal_signal_pending(current))
1386 return SWAP_CLUSTER_MAX;
1387 }
1388
Mel Gormanf3a310b2011-01-13 15:46:00 -08001389 set_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 lru_add_drain();
1391 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001393 if (scanning_global_lru(sc)) {
1394 nr_taken = isolate_pages_global(nr_to_scan,
1395 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001396 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001397 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001398 zone, 0, file);
1399 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001400 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001401 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1402 nr_scanned);
1403 else
1404 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1405 nr_scanned);
1406 } else {
1407 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1408 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001409 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001410 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001411 zone, sc->mem_cgroup,
1412 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001414 * mem_cgroup_isolate_pages() keeps track of
1415 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001417 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001418
Mel Gorman66635622010-08-09 17:19:30 -07001419 if (nr_taken == 0) {
1420 spin_unlock_irq(&zone->lru_lock);
1421 return 0;
1422 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001423
Mel Gorman1489fa12010-08-09 17:19:33 -07001424 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001425
1426 spin_unlock_irq(&zone->lru_lock);
1427
Mel Gorman0e093d992010-10-26 14:21:45 -07001428 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001429
Wu Fengguange31f36982010-08-09 17:20:01 -07001430 /* Check if we should syncronously wait for writeback */
1431 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001432 set_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001433 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001434 }
1435
1436 local_irq_disable();
1437 if (current_is_kswapd())
1438 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1439 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1440
Mel Gorman1489fa12010-08-09 17:19:33 -07001441 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001442
1443 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1444 zone_idx(zone),
1445 nr_scanned, nr_reclaimed,
1446 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001447 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001448 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449}
1450
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001451/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 * This moves pages from the active list to the inactive list.
1453 *
1454 * We move them the other way if the page is referenced by one or more
1455 * processes, from rmap.
1456 *
1457 * If the pages are mostly unmapped, the processing is fast and it is
1458 * appropriate to hold zone->lru_lock across the whole operation. But if
1459 * the pages are mapped, the processing is slow (page_referenced()) so we
1460 * should drop zone->lru_lock around each page. It's impossible to balance
1461 * this, so instead we remove the pages from the LRU while processing them.
1462 * It is safe to rely on PG_active against the non-LRU pages in here because
1463 * nobody will play with that bit on a non-LRU page.
1464 *
1465 * The downside is that we have to touch page->_count against each page.
1466 * But we had to alter page->flags anyway.
1467 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001468
Wu Fengguang3eb41402009-06-16 15:33:13 -07001469static void move_active_pages_to_lru(struct zone *zone,
1470 struct list_head *list,
1471 enum lru_list lru)
1472{
1473 unsigned long pgmoved = 0;
1474 struct pagevec pvec;
1475 struct page *page;
1476
1477 pagevec_init(&pvec, 1);
1478
1479 while (!list_empty(list)) {
1480 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001481
1482 VM_BUG_ON(PageLRU(page));
1483 SetPageLRU(page);
1484
Wu Fengguang3eb41402009-06-16 15:33:13 -07001485 list_move(&page->lru, &zone->lru[lru].list);
1486 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001487 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001488
1489 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1490 spin_unlock_irq(&zone->lru_lock);
1491 if (buffer_heads_over_limit)
1492 pagevec_strip(&pvec);
1493 __pagevec_release(&pvec);
1494 spin_lock_irq(&zone->lru_lock);
1495 }
1496 }
1497 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1498 if (!is_active_lru(lru))
1499 __count_vm_events(PGDEACTIVATE, pgmoved);
1500}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001501
Andrew Morton1742f192006-03-22 00:08:21 -08001502static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001503 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001505 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001506 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001507 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001509 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001510 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001512 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001513 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
1515 lru_add_drain();
1516 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001517 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001518 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1519 &pgscanned, sc->order,
1520 ISOLATE_ACTIVE, zone,
1521 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001522 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001523 } else {
1524 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1525 &pgscanned, sc->order,
1526 ISOLATE_ACTIVE, zone,
1527 sc->mem_cgroup, 1, file);
1528 /*
1529 * mem_cgroup_isolate_pages() keeps track of
1530 * scanned pages on its own.
1531 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001532 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001533
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001534 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001535
Wu Fengguang3eb41402009-06-16 15:33:13 -07001536 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001537 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001538 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001539 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001540 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001541 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 spin_unlock_irq(&zone->lru_lock);
1543
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 while (!list_empty(&l_hold)) {
1545 cond_resched();
1546 page = lru_to_page(&l_hold);
1547 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001548
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001549 if (unlikely(!page_evictable(page, NULL))) {
1550 putback_lru_page(page);
1551 continue;
1552 }
1553
Johannes Weiner645747462010-03-05 13:42:22 -08001554 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001555 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001556 /*
1557 * Identify referenced, file-backed active pages and
1558 * give them one more trip around the active list. So
1559 * that executable code get better chances to stay in
1560 * memory under moderate memory pressure. Anon pages
1561 * are not likely to be evicted by use-once streaming
1562 * IO, plus JVM can create lots of anon VM_EXEC pages,
1563 * so we ignore them here.
1564 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001565 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001566 list_add(&page->lru, &l_active);
1567 continue;
1568 }
1569 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001570
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001571 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 list_add(&page->lru, &l_inactive);
1573 }
1574
Andrew Mortonb5557492009-01-06 14:40:13 -08001575 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001576 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001577 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001578 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001579 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001580 * Count referenced pages from currently used mappings as rotated,
1581 * even though only some of them are actually re-activated. This
1582 * helps balance scan pressure between file and anonymous pages in
1583 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001584 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001585 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001586
Wu Fengguang3eb41402009-06-16 15:33:13 -07001587 move_active_pages_to_lru(zone, &l_active,
1588 LRU_ACTIVE + file * LRU_FILE);
1589 move_active_pages_to_lru(zone, &l_inactive,
1590 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001591 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001592 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593}
1594
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001595#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001596static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001597{
1598 unsigned long active, inactive;
1599
1600 active = zone_page_state(zone, NR_ACTIVE_ANON);
1601 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1602
1603 if (inactive * zone->inactive_ratio < active)
1604 return 1;
1605
1606 return 0;
1607}
1608
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001609/**
1610 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1611 * @zone: zone to check
1612 * @sc: scan control of this context
1613 *
1614 * Returns true if the zone does not have enough inactive anon pages,
1615 * meaning some active anon pages need to be deactivated.
1616 */
1617static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1618{
1619 int low;
1620
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001621 /*
1622 * If we don't have swap space, anonymous page deactivation
1623 * is pointless.
1624 */
1625 if (!total_swap_pages)
1626 return 0;
1627
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001628 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001629 low = inactive_anon_is_low_global(zone);
1630 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001631 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001632 return low;
1633}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001634#else
1635static inline int inactive_anon_is_low(struct zone *zone,
1636 struct scan_control *sc)
1637{
1638 return 0;
1639}
1640#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001641
Rik van Riel56e49d22009-06-16 15:32:28 -07001642static int inactive_file_is_low_global(struct zone *zone)
1643{
1644 unsigned long active, inactive;
1645
1646 active = zone_page_state(zone, NR_ACTIVE_FILE);
1647 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1648
1649 return (active > inactive);
1650}
1651
1652/**
1653 * inactive_file_is_low - check if file pages need to be deactivated
1654 * @zone: zone to check
1655 * @sc: scan control of this context
1656 *
1657 * When the system is doing streaming IO, memory pressure here
1658 * ensures that active file pages get deactivated, until more
1659 * than half of the file pages are on the inactive list.
1660 *
1661 * Once we get to that situation, protect the system's working
1662 * set from being evicted by disabling active file page aging.
1663 *
1664 * This uses a different ratio than the anonymous pages, because
1665 * the page cache uses a use-once replacement algorithm.
1666 */
1667static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1668{
1669 int low;
1670
1671 if (scanning_global_lru(sc))
1672 low = inactive_file_is_low_global(zone);
1673 else
1674 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1675 return low;
1676}
1677
Rik van Rielb39415b2009-12-14 17:59:48 -08001678static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1679 int file)
1680{
1681 if (file)
1682 return inactive_file_is_low(zone, sc);
1683 else
1684 return inactive_anon_is_low(zone, sc);
1685}
1686
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001687static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001688 struct zone *zone, struct scan_control *sc, int priority)
1689{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001690 int file = is_file_lru(lru);
1691
Rik van Rielb39415b2009-12-14 17:59:48 -08001692 if (is_active_lru(lru)) {
1693 if (inactive_list_is_low(zone, sc, file))
1694 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001695 return 0;
1696 }
1697
Rik van Riel33c120e2008-10-18 20:26:36 -07001698 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001699}
1700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001702 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1703 * until we collected @swap_cluster_max pages to scan.
1704 */
1705static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1706 unsigned long *nr_saved_scan)
1707{
1708 unsigned long nr;
1709
1710 *nr_saved_scan += nr_to_scan;
1711 nr = *nr_saved_scan;
1712
1713 if (nr >= SWAP_CLUSTER_MAX)
1714 *nr_saved_scan = 0;
1715 else
1716 nr = 0;
1717
1718 return nr;
1719}
1720
1721/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001722 * Determine how aggressively the anon and file LRU lists should be
1723 * scanned. The relative value of each set of LRU lists is determined
1724 * by looking at the fraction of the pages scanned we did rotate back
1725 * onto the active list instead of evict.
1726 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001727 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001728 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001729static void get_scan_count(struct zone *zone, struct scan_control *sc,
1730 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001731{
1732 unsigned long anon, file, free;
1733 unsigned long anon_prio, file_prio;
1734 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001735 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001736 u64 fraction[2], denominator;
1737 enum lru_list l;
1738 int noswap = 0;
1739
1740 /* If we have no swap space, do not bother scanning anon pages. */
1741 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1742 noswap = 1;
1743 fraction[0] = 0;
1744 fraction[1] = 1;
1745 denominator = 1;
1746 goto out;
1747 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001748
Vincent Li0b217672009-09-21 17:03:09 -07001749 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1750 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1751 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1752 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001753
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001754 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001755 free = zone_page_state(zone, NR_FREE_PAGES);
1756 /* If we have very few page cache pages,
1757 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001758 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001759 fraction[0] = 1;
1760 fraction[1] = 0;
1761 denominator = 1;
1762 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001763 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001764 }
1765
1766 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001767 * With swappiness at 100, anonymous and file have the same priority.
1768 * This scanning priority is essentially the inverse of IO cost.
1769 */
1770 anon_prio = sc->swappiness;
1771 file_prio = 200 - sc->swappiness;
1772
1773 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001774 * OK, so we have swap space and a fair amount of page cache
1775 * pages. We use the recently rotated / recently scanned
1776 * ratios to determine how valuable each cache is.
1777 *
1778 * Because workloads change over time (and to avoid overflow)
1779 * we keep these statistics as a floating average, which ends
1780 * up weighing recent references more than old ones.
1781 *
1782 * anon in [0], file in [1]
1783 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001784 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001785 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001786 reclaim_stat->recent_scanned[0] /= 2;
1787 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001788 }
1789
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001790 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001791 reclaim_stat->recent_scanned[1] /= 2;
1792 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001793 }
1794
1795 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001796 * The amount of pressure on anon vs file pages is inversely
1797 * proportional to the fraction of recently scanned pages on
1798 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001799 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001800 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1801 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001802
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001803 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1804 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001805 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001806
Shaohua Li76a33fc2010-05-24 14:32:36 -07001807 fraction[0] = ap;
1808 fraction[1] = fp;
1809 denominator = ap + fp + 1;
1810out:
1811 for_each_evictable_lru(l) {
1812 int file = is_file_lru(l);
1813 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001814
Shaohua Li76a33fc2010-05-24 14:32:36 -07001815 scan = zone_nr_lru_pages(zone, sc, l);
1816 if (priority || noswap) {
1817 scan >>= priority;
1818 scan = div64_u64(scan * fraction[file], denominator);
1819 }
1820 nr[l] = nr_scan_try_batch(scan,
1821 &reclaim_stat->nr_saved_scan[l]);
1822 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001823}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001824
1825/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001826 * Reclaim/compaction depends on a number of pages being freed. To avoid
1827 * disruption to the system, a small number of order-0 pages continue to be
1828 * rotated and reclaimed in the normal fashion. However, by the time we get
1829 * back to the allocator and call try_to_compact_zone(), we ensure that
1830 * there are enough free pages for it to be likely successful
1831 */
1832static inline bool should_continue_reclaim(struct zone *zone,
1833 unsigned long nr_reclaimed,
1834 unsigned long nr_scanned,
1835 struct scan_control *sc)
1836{
1837 unsigned long pages_for_compaction;
1838 unsigned long inactive_lru_pages;
1839
1840 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001841 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001842 return false;
1843
Mel Gorman28765922011-02-25 14:44:20 -08001844 /* Consider stopping depending on scan and reclaim activity */
1845 if (sc->gfp_mask & __GFP_REPEAT) {
1846 /*
1847 * For __GFP_REPEAT allocations, stop reclaiming if the
1848 * full LRU list has been scanned and we are still failing
1849 * to reclaim pages. This full LRU scan is potentially
1850 * expensive but a __GFP_REPEAT caller really wants to succeed
1851 */
1852 if (!nr_reclaimed && !nr_scanned)
1853 return false;
1854 } else {
1855 /*
1856 * For non-__GFP_REPEAT allocations which can presumably
1857 * fail without consequence, stop if we failed to reclaim
1858 * any pages from the last SWAP_CLUSTER_MAX number of
1859 * pages that were scanned. This will return to the
1860 * caller faster at the risk reclaim/compaction and
1861 * the resulting allocation attempt fails
1862 */
1863 if (!nr_reclaimed)
1864 return false;
1865 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001866
1867 /*
1868 * If we have not reclaimed enough pages for compaction and the
1869 * inactive lists are large enough, continue reclaiming
1870 */
1871 pages_for_compaction = (2UL << sc->order);
1872 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1873 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1874 if (sc->nr_reclaimed < pages_for_compaction &&
1875 inactive_lru_pages > pages_for_compaction)
1876 return true;
1877
1878 /* If compaction would go ahead or the allocation would succeed, stop */
1879 switch (compaction_suitable(zone, sc->order)) {
1880 case COMPACT_PARTIAL:
1881 case COMPACT_CONTINUE:
1882 return false;
1883 default:
1884 return true;
1885 }
1886}
1887
1888/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1890 */
Rik van Riela79311c2009-01-06 14:40:01 -08001891static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001892 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001894 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001895 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001896 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001897 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001898 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899
Mel Gorman3e7d3442011-01-13 15:45:56 -08001900restart:
1901 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001902 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001903 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001904
Rik van Riel556adec2008-10-18 20:26:34 -07001905 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1906 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001907 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001908 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001909 nr_to_scan = min_t(unsigned long,
1910 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001911 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001913 nr_reclaimed += shrink_list(l, nr_to_scan,
1914 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 }
Rik van Riela79311c2009-01-06 14:40:01 -08001917 /*
1918 * On large memory systems, scan >> priority can become
1919 * really large. This is fine for the starting priority;
1920 * we want to put equal scanning pressure on each zone.
1921 * However, if the VM has a harder time of freeing pages,
1922 * with multiple processes reclaiming pages, the total
1923 * freeing target can get unreasonably large.
1924 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001925 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001926 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001928 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001929
Rik van Riel556adec2008-10-18 20:26:34 -07001930 /*
1931 * Even if we did not try to evict anon pages at all, we want to
1932 * rebalance the anon lru active/inactive ratio.
1933 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001934 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001935 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1936
Mel Gorman3e7d3442011-01-13 15:45:56 -08001937 /* reclaim/compaction might need reclaim to continue */
1938 if (should_continue_reclaim(zone, nr_reclaimed,
1939 sc->nr_scanned - nr_scanned, sc))
1940 goto restart;
1941
Andrew Morton232ea4d2007-02-28 20:13:21 -08001942 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943}
1944
1945/*
1946 * This is the direct reclaim path, for page-allocating processes. We only
1947 * try to reclaim pages from zones which will satisfy the caller's allocation
1948 * request.
1949 *
Mel Gorman41858962009-06-16 15:32:12 -07001950 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1951 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1953 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001954 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1955 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1956 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 * If a zone is deemed to be full of pinned pages then just give it a light
1959 * scan then give up on it.
1960 */
Minchan Kimd1908362010-09-22 13:05:01 -07001961static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001962 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963{
Mel Gormandd1a2392008-04-28 02:12:17 -07001964 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001965 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001966
Mel Gormand4debc62010-08-09 17:19:29 -07001967 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1968 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001969 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001971 /*
1972 * Take care memory controller reclaiming has small influence
1973 * to global LRU.
1974 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001975 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001976 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1977 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001978 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001979 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001980 }
Nick Piggin408d8542006-09-25 23:31:27 -07001981
Rik van Riela79311c2009-01-06 14:40:01 -08001982 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 }
Minchan Kimd1908362010-09-22 13:05:01 -07001984}
1985
1986static bool zone_reclaimable(struct zone *zone)
1987{
1988 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1989}
1990
1991/*
1992 * As hibernation is going on, kswapd is freezed so that it can't mark
1993 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1994 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1995 */
1996static bool all_unreclaimable(struct zonelist *zonelist,
1997 struct scan_control *sc)
1998{
1999 struct zoneref *z;
2000 struct zone *zone;
2001 bool all_unreclaimable = true;
2002
2003 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2004 gfp_zone(sc->gfp_mask), sc->nodemask) {
2005 if (!populated_zone(zone))
2006 continue;
2007 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2008 continue;
2009 if (zone_reclaimable(zone)) {
2010 all_unreclaimable = false;
2011 break;
2012 }
2013 }
2014
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002015 return all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002017
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018/*
2019 * This is the main entry point to direct page reclaim.
2020 *
2021 * If a full scan of the inactive list fails to free enough memory then we
2022 * are "out of memory" and something needs to be killed.
2023 *
2024 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2025 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002026 * caller can't do much about. We kick the writeback threads and take explicit
2027 * naps in the hope that some of these pages can be written. But if the
2028 * allocating task holds filesystem locks which prevent writeout this might not
2029 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002030 *
2031 * returns: 0, if no pages reclaimed
2032 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 */
Mel Gormandac1d272008-04-28 02:12:12 -07002034static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07002035 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036{
2037 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002038 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002040 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002041 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002042 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Miao Xiec0ff7452010-05-24 14:32:08 -07002044 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002045 delayacct_freepages_start();
2046
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002047 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002048 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049
2050 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002051 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002052 if (!priority)
2053 disable_swap_token();
Minchan Kimd1908362010-09-22 13:05:01 -07002054 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08002055 /*
2056 * Don't shrink slabs when reclaiming memory from
2057 * over limit cgroups
2058 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002059 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002060 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002061 for_each_zone_zonelist(zone, z, zonelist,
2062 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002063 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2064 continue;
2065
2066 lru_pages += zone_reclaimable_pages(zone);
2067 }
2068
Mel Gormandd1a2392008-04-28 02:12:17 -07002069 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002070 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002071 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002072 reclaim_state->reclaimed_slab = 0;
2073 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 }
Balbir Singh66e17072008-02-07 00:13:56 -08002075 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002076 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078
2079 /*
2080 * Try to write back as many pages as we just scanned. This
2081 * tends to cause slow streaming writers to write data to the
2082 * disk smoothly, at the dirtying rate, which is nice. But
2083 * that's undesirable in laptop mode, where we *want* lumpy
2084 * writeout. So in laptop mode, write out the whole world.
2085 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002086 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2087 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002088 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002089 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 }
2091
2092 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002093 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002094 priority < DEF_PRIORITY - 2) {
2095 struct zone *preferred_zone;
2096
2097 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002098 &cpuset_current_mems_allowed,
2099 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002100 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002103
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002105 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002106 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002107
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002108 if (sc->nr_reclaimed)
2109 return sc->nr_reclaimed;
2110
2111 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002112 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002113 return 1;
2114
2115 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116}
2117
Mel Gormandac1d272008-04-28 02:12:12 -07002118unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002119 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002120{
Mel Gorman33906bc2010-08-09 17:19:16 -07002121 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002122 struct scan_control sc = {
2123 .gfp_mask = gfp_mask,
2124 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002125 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002126 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002127 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002128 .swappiness = vm_swappiness,
2129 .order = order,
2130 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002131 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002132 };
2133
Mel Gorman33906bc2010-08-09 17:19:16 -07002134 trace_mm_vmscan_direct_reclaim_begin(order,
2135 sc.may_writepage,
2136 gfp_mask);
2137
2138 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2139
2140 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2141
2142 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002143}
2144
Balbir Singh00f0b822008-03-04 14:28:39 -08002145#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002146
Balbir Singh4e416952009-09-23 15:56:39 -07002147unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2148 gfp_t gfp_mask, bool noswap,
2149 unsigned int swappiness,
KOSAKI Motohiro14fec792010-08-10 18:03:05 -07002150 struct zone *zone)
Balbir Singh4e416952009-09-23 15:56:39 -07002151{
2152 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002153 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002154 .may_writepage = !laptop_mode,
2155 .may_unmap = 1,
2156 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002157 .swappiness = swappiness,
2158 .order = 0,
2159 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002160 };
Balbir Singh4e416952009-09-23 15:56:39 -07002161 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2162 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002163
2164 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2165 sc.may_writepage,
2166 sc.gfp_mask);
2167
Balbir Singh4e416952009-09-23 15:56:39 -07002168 /*
2169 * NOTE: Although we can get the priority field, using it
2170 * here is not a good idea, since it limits the pages we can scan.
2171 * if we don't reclaim here, the shrink_zone from balance_pgdat
2172 * will pick up pages from other mem cgroup's as well. We hack
2173 * the priority and make it zero.
2174 */
2175 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002176
2177 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2178
Balbir Singh4e416952009-09-23 15:56:39 -07002179 return sc.nr_reclaimed;
2180}
2181
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002182unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002183 gfp_t gfp_mask,
2184 bool noswap,
2185 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002186{
Balbir Singh4e416952009-09-23 15:56:39 -07002187 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002188 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002189 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002190 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002191 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002192 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002193 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002194 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002195 .order = 0,
2196 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002197 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08002198 };
Balbir Singh66e17072008-02-07 00:13:56 -08002199
Mel Gormandd1a2392008-04-28 02:12:17 -07002200 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2201 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2202 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002203
2204 trace_mm_vmscan_memcg_reclaim_begin(0,
2205 sc.may_writepage,
2206 sc.gfp_mask);
2207
2208 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2209
2210 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2211
2212 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002213}
2214#endif
2215
Mel Gorman1741c872011-01-13 15:46:21 -08002216/*
2217 * pgdat_balanced is used when checking if a node is balanced for high-order
2218 * allocations. Only zones that meet watermarks and are in a zone allowed
2219 * by the callers classzone_idx are added to balanced_pages. The total of
2220 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2221 * for the node to be considered balanced. Forcing all zones to be balanced
2222 * for high orders can cause excessive reclaim when there are imbalanced zones.
2223 * The choice of 25% is due to
2224 * o a 16M DMA zone that is balanced will not balance a zone on any
2225 * reasonable sized machine
2226 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002227 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002228 * would need to be at least 256M for it to be balance a whole node.
2229 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2230 * to balance a node on its own. These seemed like reasonable ratios.
2231 */
2232static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2233 int classzone_idx)
2234{
2235 unsigned long present_pages = 0;
2236 int i;
2237
2238 for (i = 0; i <= classzone_idx; i++)
2239 present_pages += pgdat->node_zones[i].present_pages;
2240
2241 return balanced_pages > (present_pages >> 2);
2242}
2243
Mel Gormanf50de2d2009-12-14 17:58:53 -08002244/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002245static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2246 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002247{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002248 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002249 unsigned long balanced = 0;
2250 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002251
2252 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2253 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002254 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002255
Mel Gorman0abdee22011-01-13 15:46:22 -08002256 /* Check the watermark levels */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002257 for (i = 0; i < pgdat->nr_zones; i++) {
2258 struct zone *zone = pgdat->node_zones + i;
2259
2260 if (!populated_zone(zone))
2261 continue;
2262
Mel Gorman355b09c2011-01-13 15:46:24 -08002263 /*
2264 * balance_pgdat() skips over all_unreclaimable after
2265 * DEF_PRIORITY. Effectively, it considers them balanced so
2266 * they must be considered balanced here as well if kswapd
2267 * is to sleep
2268 */
2269 if (zone->all_unreclaimable) {
2270 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002271 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002272 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002273
Mel Gorman88f5acf2011-01-13 15:45:41 -08002274 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormandc83edd2011-01-13 15:46:26 -08002275 classzone_idx, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002276 all_zones_ok = false;
2277 else
2278 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002279 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002280
Mel Gorman1741c872011-01-13 15:46:21 -08002281 /*
2282 * For high-order requests, the balanced zones must contain at least
2283 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2284 * must be balanced
2285 */
2286 if (order)
Mel Gormandc83edd2011-01-13 15:46:26 -08002287 return pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002288 else
2289 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002290}
2291
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292/*
2293 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002294 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002296 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 *
2298 * There is special handling here for zones which are full of pinned pages.
2299 * This can happen if the pages are all mlocked, or if they are all used by
2300 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2301 * What we do is to detect the case where all pages in the zone have been
2302 * scanned twice and there has been zero successful reclaim. Mark the zone as
2303 * dead and from now on, only perform a short scan. Basically we're polling
2304 * the zone for when the problem goes away.
2305 *
2306 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002307 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2308 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2309 * lower zones regardless of the number of free pages in the lower zones. This
2310 * interoperates with the page allocator fallback scheme to ensure that aging
2311 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 */
Mel Gorman99504742011-01-13 15:46:20 -08002313static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002314 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002317 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 int priority;
2319 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002320 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002321 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002323 struct scan_control sc = {
2324 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002325 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002326 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002327 /*
2328 * kswapd doesn't want to be bailed out while reclaim. because
2329 * we want to put equal scanning pressure on each zone.
2330 */
2331 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002332 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002333 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002334 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002335 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336loop_again:
2337 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002338 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002339 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002340 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002344 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002346 /* The swap token gets in the way of swapout... */
2347 if (!priority)
2348 disable_swap_token();
2349
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002351 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002353 /*
2354 * Scan in the highmem->dma direction for the highest
2355 * zone which needs scanning
2356 */
2357 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2358 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002360 if (!populated_zone(zone))
2361 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002363 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002364 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365
Rik van Riel556adec2008-10-18 20:26:34 -07002366 /*
2367 * Do some background aging of the anon list, to give
2368 * pages a chance to be referenced before reclaiming.
2369 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002370 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002371 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2372 &sc, priority, 0);
2373
Mel Gorman88f5acf2011-01-13 15:45:41 -08002374 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002375 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002376 end_zone = i;
Mel Gormandc83edd2011-01-13 15:46:26 -08002377 *classzone_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002378 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002381 if (i < 0)
2382 goto out;
2383
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 for (i = 0; i <= end_zone; i++) {
2385 struct zone *zone = pgdat->node_zones + i;
2386
Wu Fengguangadea02a2009-09-21 17:01:42 -07002387 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 }
2389
2390 /*
2391 * Now scan the zone in the dma->highmem direction, stopping
2392 * at the last zone which needs scanning.
2393 *
2394 * We do this because the page allocator works in the opposite
2395 * direction. This prevents the page allocator from allocating
2396 * pages behind kswapd's direction of progress, which would
2397 * cause too much scanning of the lower zones.
2398 */
2399 for (i = 0; i <= end_zone; i++) {
2400 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002401 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002402 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403
Con Kolivasf3fe6512006-01-06 00:11:15 -08002404 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 continue;
2406
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002407 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 continue;
2409
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002411
Balbir Singh4e416952009-09-23 15:56:39 -07002412 /*
2413 * Call soft limit reclaim before calling shrink_zone.
2414 * For now we ignore the return value
2415 */
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002416 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2417
Rik van Riel32a43302007-10-16 01:24:50 -07002418 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002419 * We put equal pressure on every zone, unless
2420 * one zone has way too many pages free
2421 * already. The "too many pages" is defined
2422 * as the high wmark plus a "gap" where the
2423 * gap is either the low watermark or 1%
2424 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002425 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002426 balance_gap = min(low_wmark_pages(zone),
2427 (zone->present_pages +
2428 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2429 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002430 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002431 high_wmark_pages(zone) + balance_gap,
2432 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002433 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002435 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2436 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002437 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 total_scanned += sc.nr_scanned;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002439
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002440 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 continue;
Andrea Arcangelid527caf2011-03-22 16:30:38 -07002442 if (nr_slab == 0 &&
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002443 !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002444 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 /*
2446 * If we've done a decent amount of scanning and
2447 * the reclaim ratio is low, start doing writepage
2448 * even in laptop mode
2449 */
2450 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002451 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002453
Mel Gorman88f5acf2011-01-13 15:45:41 -08002454 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002455 high_wmark_pages(zone), end_zone, 0)) {
2456 all_zones_ok = 0;
2457 /*
2458 * We are still under min water mark. This
2459 * means that we have a GFP_ATOMIC allocation
2460 * failure risk. Hurry up!
2461 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002462 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002463 min_wmark_pages(zone), end_zone, 0))
2464 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002465 } else {
2466 /*
2467 * If a zone reaches its high watermark,
2468 * consider it to be no longer congested. It's
2469 * possible there are dirty pages backed by
2470 * congested BDIs but as pressure is relieved,
2471 * spectulatively avoid congestion waits
2472 */
2473 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002474 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002475 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002476 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002477
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002479 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 break; /* kswapd: all done */
2481 /*
2482 * OK, kswapd is getting into trouble. Take a nap, then take
2483 * another pass across the zones.
2484 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002485 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2486 if (has_under_min_watermark_zone)
2487 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2488 else
2489 congestion_wait(BLK_RW_ASYNC, HZ/10);
2490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491
2492 /*
2493 * We do this so kswapd doesn't build up large priorities for
2494 * example when it is freeing in parallel with allocators. It
2495 * matches the direct reclaim path behaviour in terms of impact
2496 * on zone->*_priority.
2497 */
Rik van Riela79311c2009-01-06 14:40:01 -08002498 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 break;
2500 }
2501out:
Mel Gorman99504742011-01-13 15:46:20 -08002502
2503 /*
2504 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002505 * high-order: Balanced zones must make up at least 25% of the node
2506 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002507 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002508 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002510
2511 try_to_freeze();
2512
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002513 /*
2514 * Fragmentation may mean that the system cannot be
2515 * rebalanced for high-order allocations in all zones.
2516 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2517 * it means the zones have been fully scanned and are still
2518 * not balanced. For high-order allocations, there is
2519 * little point trying all over again as kswapd may
2520 * infinite loop.
2521 *
2522 * Instead, recheck all watermarks at order-0 as they
2523 * are the most important. If watermarks are ok, kswapd will go
2524 * back to sleep. High-order users can still perform direct
2525 * reclaim if they wish.
2526 */
2527 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2528 order = sc.order = 0;
2529
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 goto loop_again;
2531 }
2532
Mel Gorman99504742011-01-13 15:46:20 -08002533 /*
2534 * If kswapd was reclaiming at a higher order, it has the option of
2535 * sleeping without all zones being balanced. Before it does, it must
2536 * ensure that the watermarks for order-0 on *all* zones are met and
2537 * that the congestion flags are cleared. The congestion flag must
2538 * be cleared as kswapd is the only mechanism that clears the flag
2539 * and it is potentially going to sleep here.
2540 */
2541 if (order) {
2542 for (i = 0; i <= end_zone; i++) {
2543 struct zone *zone = pgdat->node_zones + i;
2544
2545 if (!populated_zone(zone))
2546 continue;
2547
2548 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2549 continue;
2550
2551 /* Confirm the zone is balanced for order-0 */
2552 if (!zone_watermark_ok(zone, 0,
2553 high_wmark_pages(zone), 0, 0)) {
2554 order = sc.order = 0;
2555 goto loop_again;
2556 }
2557
2558 /* If balanced, clear the congested flag */
2559 zone_clear_flag(zone, ZONE_CONGESTED);
2560 }
2561 }
2562
Mel Gorman0abdee22011-01-13 15:46:22 -08002563 /*
2564 * Return the order we were reclaiming at so sleeping_prematurely()
2565 * makes a decision on the order we were last reclaiming at. However,
2566 * if another caller entered the allocator slow path while kswapd
2567 * was awake, order will remain at the higher level
2568 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002569 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002570 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571}
2572
Mel Gormandc83edd2011-01-13 15:46:26 -08002573static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002574{
2575 long remaining = 0;
2576 DEFINE_WAIT(wait);
2577
2578 if (freezing(current) || kthread_should_stop())
2579 return;
2580
2581 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2582
2583 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002584 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002585 remaining = schedule_timeout(HZ/10);
2586 finish_wait(&pgdat->kswapd_wait, &wait);
2587 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2588 }
2589
2590 /*
2591 * After a short sleep, check if it was a premature sleep. If not, then
2592 * go fully to sleep until explicitly woken up.
2593 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002594 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002595 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2596
2597 /*
2598 * vmstat counters are not perfectly accurate and the estimated
2599 * value for counters such as NR_FREE_PAGES can deviate from the
2600 * true value by nr_online_cpus * threshold. To avoid the zone
2601 * watermarks being breached while under pressure, we reduce the
2602 * per-cpu vmstat threshold while kswapd is awake and restore
2603 * them before going back to sleep.
2604 */
2605 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2606 schedule();
2607 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2608 } else {
2609 if (remaining)
2610 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2611 else
2612 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2613 }
2614 finish_wait(&pgdat->kswapd_wait, &wait);
2615}
2616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617/*
2618 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002619 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 *
2621 * This basically trickles out pages so that we have _some_
2622 * free memory available even if there is no other activity
2623 * that frees anything up. This is needed for things like routing
2624 * etc, where we otherwise might have all activity going on in
2625 * asynchronous contexts that cannot page things out.
2626 *
2627 * If there are applications that are active memory-allocators
2628 * (most normal use), this basically shouldn't matter.
2629 */
2630static int kswapd(void *p)
2631{
2632 unsigned long order;
Mel Gorman99504742011-01-13 15:46:20 -08002633 int classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634 pg_data_t *pgdat = (pg_data_t*)p;
2635 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002636
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 struct reclaim_state reclaim_state = {
2638 .reclaimed_slab = 0,
2639 };
Rusty Russella70f7302009-03-13 14:49:46 +10302640 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641
Nick Piggincf40bd12009-01-21 08:12:39 +01002642 lockdep_set_current_reclaim_state(GFP_KERNEL);
2643
Rusty Russell174596a2009-01-01 10:12:29 +10302644 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002645 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 current->reclaim_state = &reclaim_state;
2647
2648 /*
2649 * Tell the memory management that we're a "memory allocator",
2650 * and that if we need more memory we should get access to it
2651 * regardless (see "__alloc_pages()"). "kswapd" should
2652 * never get caught in the normal page freeing logic.
2653 *
2654 * (Kswapd normally doesn't need memory anyway, but sometimes
2655 * you need a small amount of memory in order to be able to
2656 * page out something else, and this flag essentially protects
2657 * us from recursively trying to free more memory as we're
2658 * trying to free the first piece of memory in the first place).
2659 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002660 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002661 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662
2663 order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002664 classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 for ( ; ; ) {
2666 unsigned long new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002667 int new_classzone_idx;
David Rientjes8fe23e02009-12-14 17:58:33 -08002668 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002669
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 new_order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002671 new_classzone_idx = pgdat->classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 pgdat->kswapd_max_order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002673 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2674 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 /*
2676 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002677 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 */
2679 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002680 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 } else {
Mel Gormandc83edd2011-01-13 15:46:26 -08002682 kswapd_try_to_sleep(pgdat, order, classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002684 classzone_idx = pgdat->classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002685 pgdat->kswapd_max_order = 0;
2686 pgdat->classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688
David Rientjes8fe23e02009-12-14 17:58:33 -08002689 ret = try_to_freeze();
2690 if (kthread_should_stop())
2691 break;
2692
2693 /*
2694 * We can speed up thawing tasks if we don't call balance_pgdat
2695 * after returning from the refrigerator
2696 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002697 if (!ret) {
2698 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Mel Gormandc83edd2011-01-13 15:46:26 -08002699 order = balance_pgdat(pgdat, order, &classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002700 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 }
2702 return 0;
2703}
2704
2705/*
2706 * A zone is low on free memory, so wake its kswapd task to service it.
2707 */
Mel Gorman99504742011-01-13 15:46:20 -08002708void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709{
2710 pg_data_t *pgdat;
2711
Con Kolivasf3fe6512006-01-06 00:11:15 -08002712 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 return;
2714
Paul Jackson02a0e532006-12-13 00:34:25 -08002715 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002717 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002718 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002719 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002720 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2721 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002722 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002724 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2725 return;
2726
2727 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002728 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729}
2730
Wu Fengguangadea02a2009-09-21 17:01:42 -07002731/*
2732 * The reclaimable count would be mostly accurate.
2733 * The less reclaimable pages may be
2734 * - mlocked pages, which will be moved to unevictable list when encountered
2735 * - mapped pages, which may require several travels to be reclaimed
2736 * - dirty pages, which is not "instantly" reclaimable
2737 */
2738unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002739{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002740 int nr;
2741
2742 nr = global_page_state(NR_ACTIVE_FILE) +
2743 global_page_state(NR_INACTIVE_FILE);
2744
2745 if (nr_swap_pages > 0)
2746 nr += global_page_state(NR_ACTIVE_ANON) +
2747 global_page_state(NR_INACTIVE_ANON);
2748
2749 return nr;
2750}
2751
2752unsigned long zone_reclaimable_pages(struct zone *zone)
2753{
2754 int nr;
2755
2756 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2757 zone_page_state(zone, NR_INACTIVE_FILE);
2758
2759 if (nr_swap_pages > 0)
2760 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2761 zone_page_state(zone, NR_INACTIVE_ANON);
2762
2763 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002764}
2765
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002766#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002768 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002769 * freed pages.
2770 *
2771 * Rather than trying to age LRUs the aim is to preserve the overall
2772 * LRU order by reclaiming preferentially
2773 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002775unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002777 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002778 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002779 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2780 .may_swap = 1,
2781 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002782 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002783 .nr_to_reclaim = nr_to_reclaim,
2784 .hibernation_mode = 1,
2785 .swappiness = vm_swappiness,
2786 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002788 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2789 struct task_struct *p = current;
2790 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002792 p->flags |= PF_MEMALLOC;
2793 lockdep_set_current_reclaim_state(sc.gfp_mask);
2794 reclaim_state.reclaimed_slab = 0;
2795 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002796
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002797 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002798
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002799 p->reclaim_state = NULL;
2800 lockdep_clear_current_reclaim_state();
2801 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002802
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002803 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002805#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807/* It's optimal to keep kswapds on the same CPUs as their memory, but
2808 not required for correctness. So if the last cpu in a node goes
2809 away, we get changed to run anywhere: as the first one comes back,
2810 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002811static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002812 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002814 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002816 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002817 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002818 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302819 const struct cpumask *mask;
2820
2821 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002822
Rusty Russell3e597942009-01-01 10:12:24 +10302823 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002825 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 }
2827 }
2828 return NOTIFY_OK;
2829}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
Yasunori Goto3218ae12006-06-27 02:53:33 -07002831/*
2832 * This kswapd start function will be called by init and node-hot-add.
2833 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2834 */
2835int kswapd_run(int nid)
2836{
2837 pg_data_t *pgdat = NODE_DATA(nid);
2838 int ret = 0;
2839
2840 if (pgdat->kswapd)
2841 return 0;
2842
2843 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2844 if (IS_ERR(pgdat->kswapd)) {
2845 /* failure at boot is fatal */
2846 BUG_ON(system_state == SYSTEM_BOOTING);
2847 printk("Failed to start kswapd on node %d\n",nid);
2848 ret = -1;
2849 }
2850 return ret;
2851}
2852
David Rientjes8fe23e02009-12-14 17:58:33 -08002853/*
2854 * Called by memory hotplug when all memory in a node is offlined.
2855 */
2856void kswapd_stop(int nid)
2857{
2858 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2859
2860 if (kswapd)
2861 kthread_stop(kswapd);
2862}
2863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864static int __init kswapd_init(void)
2865{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002866 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002867
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002869 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002870 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 hotcpu_notifier(cpu_callback, 0);
2872 return 0;
2873}
2874
2875module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002876
2877#ifdef CONFIG_NUMA
2878/*
2879 * Zone reclaim mode
2880 *
2881 * If non-zero call zone_reclaim when the number of free pages falls below
2882 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002883 */
2884int zone_reclaim_mode __read_mostly;
2885
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002886#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002887#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002888#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2889#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2890
Christoph Lameter9eeff232006-01-18 17:42:31 -08002891/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002892 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2893 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2894 * a zone.
2895 */
2896#define ZONE_RECLAIM_PRIORITY 4
2897
Christoph Lameter9eeff232006-01-18 17:42:31 -08002898/*
Christoph Lameter96146342006-07-03 00:24:13 -07002899 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2900 * occur.
2901 */
2902int sysctl_min_unmapped_ratio = 1;
2903
2904/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002905 * If the number of slab pages in a zone grows beyond this percentage then
2906 * slab reclaim needs to occur.
2907 */
2908int sysctl_min_slab_ratio = 5;
2909
Mel Gorman90afa5d2009-06-16 15:33:20 -07002910static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2911{
2912 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2913 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2914 zone_page_state(zone, NR_ACTIVE_FILE);
2915
2916 /*
2917 * It's possible for there to be more file mapped pages than
2918 * accounted for by the pages on the file LRU lists because
2919 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2920 */
2921 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2922}
2923
2924/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2925static long zone_pagecache_reclaimable(struct zone *zone)
2926{
2927 long nr_pagecache_reclaimable;
2928 long delta = 0;
2929
2930 /*
2931 * If RECLAIM_SWAP is set, then all file pages are considered
2932 * potentially reclaimable. Otherwise, we have to worry about
2933 * pages like swapcache and zone_unmapped_file_pages() provides
2934 * a better estimate
2935 */
2936 if (zone_reclaim_mode & RECLAIM_SWAP)
2937 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2938 else
2939 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2940
2941 /* If we can't clean pages, remove dirty pages from consideration */
2942 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2943 delta += zone_page_state(zone, NR_FILE_DIRTY);
2944
2945 /* Watch for any possible underflows due to delta */
2946 if (unlikely(delta > nr_pagecache_reclaimable))
2947 delta = nr_pagecache_reclaimable;
2948
2949 return nr_pagecache_reclaimable - delta;
2950}
2951
Christoph Lameter0ff38492006-09-25 23:31:52 -07002952/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002953 * Try to free up some pages from this zone through reclaim.
2954 */
Andrew Morton179e9632006-03-22 00:08:18 -08002955static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002956{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002957 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002958 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002959 struct task_struct *p = current;
2960 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002961 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002962 struct scan_control sc = {
2963 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002964 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002965 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002966 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2967 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002968 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002969 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002970 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002971 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002972 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002973
Christoph Lameter9eeff232006-01-18 17:42:31 -08002974 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002975 /*
2976 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2977 * and we also need to be able to write out pages for RECLAIM_WRITE
2978 * and RECLAIM_SWAP.
2979 */
2980 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002981 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002982 reclaim_state.reclaimed_slab = 0;
2983 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002984
Mel Gorman90afa5d2009-06-16 15:33:20 -07002985 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002986 /*
2987 * Free memory by calling shrink zone with increasing
2988 * priorities until we have enough memory freed.
2989 */
2990 priority = ZONE_RECLAIM_PRIORITY;
2991 do {
Rik van Riela79311c2009-01-06 14:40:01 -08002992 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002993 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002994 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002995 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002996
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002997 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2998 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002999 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003000 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003001 * many pages were freed in this zone. So we take the current
3002 * number of slab pages and shake the slab until it is reduced
3003 * by the same nr_pages that we used for reclaiming unmapped
3004 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003005 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003006 * Note that shrink_slab will free memory on all zones and may
3007 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003008 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003009 for (;;) {
3010 unsigned long lru_pages = zone_reclaimable_pages(zone);
3011
3012 /* No reclaimable slab or very low memory pressure */
3013 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3014 break;
3015
3016 /* Freed enough memory */
3017 nr_slab_pages1 = zone_page_state(zone,
3018 NR_SLAB_RECLAIMABLE);
3019 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3020 break;
3021 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003022
3023 /*
3024 * Update nr_reclaimed by the number of slab pages we
3025 * reclaimed from this zone.
3026 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003027 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3028 if (nr_slab_pages1 < nr_slab_pages0)
3029 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003030 }
3031
Christoph Lameter9eeff232006-01-18 17:42:31 -08003032 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003033 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003034 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003035 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003036}
Andrew Morton179e9632006-03-22 00:08:18 -08003037
3038int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3039{
Andrew Morton179e9632006-03-22 00:08:18 -08003040 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003041 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003042
3043 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003044 * Zone reclaim reclaims unmapped file backed pages and
3045 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003046 *
Christoph Lameter96146342006-07-03 00:24:13 -07003047 * A small portion of unmapped file backed pages is needed for
3048 * file I/O otherwise pages read by file I/O will be immediately
3049 * thrown out if the zone is overallocated. So we do not reclaim
3050 * if less than a specified percentage of the zone is used by
3051 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003052 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003053 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3054 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003055 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003056
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003057 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003058 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003059
Andrew Morton179e9632006-03-22 00:08:18 -08003060 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003061 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003062 */
David Rientjesd773ed62007-10-16 23:26:01 -07003063 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003064 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003065
3066 /*
3067 * Only run zone reclaim on the local zone or on zones that do not
3068 * have associated processors. This will favor the local processor
3069 * over remote processors and spread off node memory allocations
3070 * as wide as possible.
3071 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003072 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003073 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003074 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003075
3076 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003077 return ZONE_RECLAIM_NOSCAN;
3078
David Rientjesd773ed62007-10-16 23:26:01 -07003079 ret = __zone_reclaim(zone, gfp_mask, order);
3080 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3081
Mel Gorman24cf725182009-06-16 15:33:23 -07003082 if (!ret)
3083 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3084
David Rientjesd773ed62007-10-16 23:26:01 -07003085 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003086}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003087#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003088
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003089/*
3090 * page_evictable - test whether a page is evictable
3091 * @page: the page to test
3092 * @vma: the VMA in which the page is or will be mapped, may be NULL
3093 *
3094 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003095 * lists vs unevictable list. The vma argument is !NULL when called from the
3096 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003097 *
3098 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003099 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003100 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003101 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003102 */
3103int page_evictable(struct page *page, struct vm_area_struct *vma)
3104{
3105
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003106 if (mapping_unevictable(page_mapping(page)))
3107 return 0;
3108
Nick Pigginb291f002008-10-18 20:26:44 -07003109 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3110 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003111
3112 return 1;
3113}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003114
3115/**
3116 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3117 * @page: page to check evictability and move to appropriate lru list
3118 * @zone: zone page is in
3119 *
3120 * Checks a page for evictability and moves the page to the appropriate
3121 * zone lru list.
3122 *
3123 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3124 * have PageUnevictable set.
3125 */
3126static void check_move_unevictable_page(struct page *page, struct zone *zone)
3127{
3128 VM_BUG_ON(PageActive(page));
3129
3130retry:
3131 ClearPageUnevictable(page);
3132 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003133 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003134
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003135 __dec_zone_state(zone, NR_UNEVICTABLE);
3136 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003137 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003138 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3139 __count_vm_event(UNEVICTABLE_PGRESCUED);
3140 } else {
3141 /*
3142 * rotate unevictable list
3143 */
3144 SetPageUnevictable(page);
3145 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003146 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003147 if (page_evictable(page, NULL))
3148 goto retry;
3149 }
3150}
3151
3152/**
3153 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3154 * @mapping: struct address_space to scan for evictable pages
3155 *
3156 * Scan all pages in mapping. Check unevictable pages for
3157 * evictability and move them to the appropriate zone lru list.
3158 */
3159void scan_mapping_unevictable_pages(struct address_space *mapping)
3160{
3161 pgoff_t next = 0;
3162 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3163 PAGE_CACHE_SHIFT;
3164 struct zone *zone;
3165 struct pagevec pvec;
3166
3167 if (mapping->nrpages == 0)
3168 return;
3169
3170 pagevec_init(&pvec, 0);
3171 while (next < end &&
3172 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3173 int i;
3174 int pg_scanned = 0;
3175
3176 zone = NULL;
3177
3178 for (i = 0; i < pagevec_count(&pvec); i++) {
3179 struct page *page = pvec.pages[i];
3180 pgoff_t page_index = page->index;
3181 struct zone *pagezone = page_zone(page);
3182
3183 pg_scanned++;
3184 if (page_index > next)
3185 next = page_index;
3186 next++;
3187
3188 if (pagezone != zone) {
3189 if (zone)
3190 spin_unlock_irq(&zone->lru_lock);
3191 zone = pagezone;
3192 spin_lock_irq(&zone->lru_lock);
3193 }
3194
3195 if (PageLRU(page) && PageUnevictable(page))
3196 check_move_unevictable_page(page, zone);
3197 }
3198 if (zone)
3199 spin_unlock_irq(&zone->lru_lock);
3200 pagevec_release(&pvec);
3201
3202 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3203 }
3204
3205}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003206
3207/**
3208 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3209 * @zone - zone of which to scan the unevictable list
3210 *
3211 * Scan @zone's unevictable LRU lists to check for pages that have become
3212 * evictable. Move those that have to @zone's inactive list where they
3213 * become candidates for reclaim, unless shrink_inactive_zone() decides
3214 * to reactivate them. Pages that are still unevictable are rotated
3215 * back onto @zone's unevictable list.
3216 */
3217#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08003218static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003219{
3220 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3221 unsigned long scan;
3222 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3223
3224 while (nr_to_scan > 0) {
3225 unsigned long batch_size = min(nr_to_scan,
3226 SCAN_UNEVICTABLE_BATCH_SIZE);
3227
3228 spin_lock_irq(&zone->lru_lock);
3229 for (scan = 0; scan < batch_size; scan++) {
3230 struct page *page = lru_to_page(l_unevictable);
3231
3232 if (!trylock_page(page))
3233 continue;
3234
3235 prefetchw_prev_lru_page(page, l_unevictable, flags);
3236
3237 if (likely(PageLRU(page) && PageUnevictable(page)))
3238 check_move_unevictable_page(page, zone);
3239
3240 unlock_page(page);
3241 }
3242 spin_unlock_irq(&zone->lru_lock);
3243
3244 nr_to_scan -= batch_size;
3245 }
3246}
3247
3248
3249/**
3250 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3251 *
3252 * A really big hammer: scan all zones' unevictable LRU lists to check for
3253 * pages that have become evictable. Move those back to the zones'
3254 * inactive list where they become candidates for reclaim.
3255 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3256 * and we add swap to the system. As such, it runs in the context of a task
3257 * that has possibly/probably made some previously unevictable pages
3258 * evictable.
3259 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003260static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003261{
3262 struct zone *zone;
3263
3264 for_each_zone(zone) {
3265 scan_zone_unevictable_pages(zone);
3266 }
3267}
3268
3269/*
3270 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3271 * all nodes' unevictable lists for evictable pages
3272 */
3273unsigned long scan_unevictable_pages;
3274
3275int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003276 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003277 size_t *length, loff_t *ppos)
3278{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003279 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003280
3281 if (write && *(unsigned long *)table->data)
3282 scan_all_zones_unevictable_pages();
3283
3284 scan_unevictable_pages = 0;
3285 return 0;
3286}
3287
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003288#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003289/*
3290 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3291 * a specified node's per zone unevictable lists for evictable pages.
3292 */
3293
3294static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3295 struct sysdev_attribute *attr,
3296 char *buf)
3297{
3298 return sprintf(buf, "0\n"); /* always zero; should fit... */
3299}
3300
3301static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3302 struct sysdev_attribute *attr,
3303 const char *buf, size_t count)
3304{
3305 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3306 struct zone *zone;
3307 unsigned long res;
3308 unsigned long req = strict_strtoul(buf, 10, &res);
3309
3310 if (!req)
3311 return 1; /* zero is no-op */
3312
3313 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3314 if (!populated_zone(zone))
3315 continue;
3316 scan_zone_unevictable_pages(zone);
3317 }
3318 return 1;
3319}
3320
3321
3322static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3323 read_scan_unevictable_node,
3324 write_scan_unevictable_node);
3325
3326int scan_unevictable_register_node(struct node *node)
3327{
3328 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3329}
3330
3331void scan_unevictable_unregister_node(struct node *node)
3332{
3333 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3334}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003335#endif