blob: c9177202c8ce2ce2f9bfe9045f2a40a8105de8e8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51
Nick Piggin0f8053a2006-03-22 00:08:33 -080052#include "internal.h"
53
Mel Gorman33906bc2010-08-09 17:19:16 -070054#define CREATE_TRACE_POINTS
55#include <trace/events/vmscan.h>
56
Mel Gormanee64fc92011-01-13 15:45:55 -080057/*
Mel Gormanf3a310b2011-01-13 15:46:00 -080058 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
Mel Gormanee64fc92011-01-13 15:45:55 -080063 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
Mel Gormanf3a310b2011-01-13 15:46:00 -080065 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
Mel Gorman3e7d3442011-01-13 15:45:56 -080066 * order-0 pages and then compact the zone
Mel Gormanee64fc92011-01-13 15:45:55 -080067 */
Mel Gormanf3a310b2011-01-13 15:46:00 -080068typedef unsigned __bitwise__ reclaim_mode_t;
69#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070074
Linus Torvalds1da177e2005-04-16 15:20:36 -070075struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
78
Rik van Riela79311c2009-01-06 14:40:01 -080079 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
81
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080082 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
84
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080085 unsigned long hibernation_mode;
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040088 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 int may_writepage;
91
Johannes Weinera6dc60f2009-03-31 15:19:30 -070092 /* Can mapped pages be reclaimed? */
93 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080094
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070095 /* Can pages be swapped as part of reclaim? */
96 int may_swap;
97
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070098 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070099
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700100 int order;
Balbir Singh66e17072008-02-07 00:13:56 -0800101
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700102 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +0530103 * Intend to reclaim enough continuous memory rather than reclaim
104 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700105 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800106 reclaim_mode_t reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700107
Balbir Singh66e17072008-02-07 00:13:56 -0800108 /* Which cgroup do we reclaim from */
109 struct mem_cgroup *mem_cgroup;
110
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -0700111 /*
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 * are scanned.
114 */
115 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120#ifdef ARCH_HAS_PREFETCH
121#define prefetch_prev_lru_page(_page, _base, _field) \
122 do { \
123 if ((_page)->lru.prev != _base) { \
124 struct page *prev; \
125 \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
128 } \
129 } while (0)
130#else
131#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132#endif
133
134#ifdef ARCH_HAS_PREFETCHW
135#define prefetchw_prev_lru_page(_page, _base, _field) \
136 do { \
137 if ((_page)->lru.prev != _base) { \
138 struct page *prev; \
139 \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
142 } \
143 } while (0)
144#else
145#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146#endif
147
148/*
149 * From 0 .. 100. Higher means more swappy.
150 */
151int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700152long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153
154static LIST_HEAD(shrinker_list);
155static DECLARE_RWSEM(shrinker_rwsem);
156
Balbir Singh00f0b822008-03-04 14:28:39 -0800157#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800158#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800160#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800161#endif
162
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800163static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164 struct scan_control *sc)
165{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800166 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800167 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800169 return &zone->reclaim_stat;
170}
171
Vincent Li0b217672009-09-21 17:03:09 -0700172static unsigned long zone_nr_lru_pages(struct zone *zone,
173 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800174{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800175 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800176 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
177
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800178 return zone_page_state(zone, NR_LRU_BASE + lru);
179}
180
181
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182/*
183 * Add a shrinker callback to be called from the vm
184 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700185void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700187 shrinker->nr = 0;
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700192EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194/*
195 * Remove one
196 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700197void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700203EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
205#define SHRINK_BATCH 128
206/*
207 * Call the shrink functions to age shrinkable caches
208 *
209 * Here we assume it costs one seek to replace a lru page and that it also
210 * takes a seek to recreate a cache object. With this in mind we age equal
211 * percentages of the lru and ageable caches. This should balance the seeks
212 * generated by these structures.
213 *
Simon Arlott183ff222007-10-20 01:27:18 +0200214 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 * slab to avoid swapping.
216 *
217 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
218 *
219 * `lru_pages' represents the number of on-LRU pages in all the zones which
220 * are eligible for the caller's allocation attempt. It is used for balancing
221 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700222 *
223 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
Andrew Morton69e05942006-03-22 00:08:19 -0800225unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
226 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227{
228 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800229 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
231 if (scanned == 0)
232 scanned = SWAP_CLUSTER_MAX;
233
234 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700235 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236
237 list_for_each_entry(shrinker, &shrinker_list, list) {
238 unsigned long long delta;
239 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000240 unsigned long max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241
Dave Chinner7f8275d2010-07-19 14:56:17 +1000242 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800244 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 do_div(delta, lru_pages + 1);
246 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800247 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700248 printk(KERN_ERR "shrink_slab: %pF negative objects to "
249 "delete nr=%ld\n",
250 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800251 shrinker->nr = max_pass;
252 }
253
254 /*
255 * Avoid risking looping forever due to too large nr value:
256 * never try to free more than twice the estimate number of
257 * freeable entries.
258 */
259 if (shrinker->nr > max_pass * 2)
260 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
262 total_scan = shrinker->nr;
263 shrinker->nr = 0;
264
265 while (total_scan >= SHRINK_BATCH) {
266 long this_scan = SHRINK_BATCH;
267 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700268 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
Dave Chinner7f8275d2010-07-19 14:56:17 +1000270 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
271 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
272 gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 if (shrink_ret == -1)
274 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700275 if (shrink_ret < nr_before)
276 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700277 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 total_scan -= this_scan;
279
280 cond_resched();
281 }
282
283 shrinker->nr += total_scan;
284 }
285 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700286 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287}
288
Mel Gormanf3a310b2011-01-13 15:46:00 -0800289static void set_reclaim_mode(int priority, struct scan_control *sc,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700290 bool sync)
291{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800292 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700293
294 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800295 * Initially assume we are entering either lumpy reclaim or
296 * reclaim/compaction.Depending on the order, we will either set the
297 * sync mode or just reclaim order-0 pages later.
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700298 */
Mel Gorman3e7d3442011-01-13 15:45:56 -0800299 if (COMPACTION_BUILD)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800300 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800301 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800302 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700303
304 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800305 * Avoid using lumpy reclaim or reclaim/compaction if possible by
306 * restricting when its set to either costly allocations or when
307 * under memory pressure
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700308 */
309 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800310 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700311 else if (sc->order && priority < DEF_PRIORITY - 2)
Mel Gormanf3a310b2011-01-13 15:46:00 -0800312 sc->reclaim_mode |= syncmode;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700313 else
Mel Gormanf3a310b2011-01-13 15:46:00 -0800314 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700315}
316
Mel Gormanf3a310b2011-01-13 15:46:00 -0800317static void reset_reclaim_mode(struct scan_control *sc)
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700318{
Mel Gormanf3a310b2011-01-13 15:46:00 -0800319 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700320}
321
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322static inline int is_page_cache_freeable(struct page *page)
323{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700324 /*
325 * A freeable page cache page is referenced only by the caller
326 * that isolated the page, the page cache radix tree and
327 * optional buffer heads at page->private.
328 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700329 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330}
331
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700332static int may_write_to_queue(struct backing_dev_info *bdi,
333 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334{
Christoph Lameter930d9152006-01-08 01:00:47 -0800335 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 return 1;
337 if (!bdi_write_congested(bdi))
338 return 1;
339 if (bdi == current->backing_dev_info)
340 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700341
342 /* lumpy reclaim for hugepage often need a lot of write */
343 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
344 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 return 0;
346}
347
348/*
349 * We detected a synchronous write error writing a page out. Probably
350 * -ENOSPC. We need to propagate that into the address_space for a subsequent
351 * fsync(), msync() or close().
352 *
353 * The tricky part is that after writepage we cannot touch the mapping: nothing
354 * prevents it from being freed up. But we have a ref on the page and once
355 * that page is locked, the mapping is pinned.
356 *
357 * We're allowed to run sleeping lock_page() here because we know the caller has
358 * __GFP_FS.
359 */
360static void handle_write_error(struct address_space *mapping,
361 struct page *page, int error)
362{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100363 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700364 if (page_mapping(page) == mapping)
365 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 unlock_page(page);
367}
368
Christoph Lameter04e62a22006-06-23 02:03:38 -0700369/* possible outcome of pageout() */
370typedef enum {
371 /* failed to write page out, page is locked */
372 PAGE_KEEP,
373 /* move page to the active list, page is locked */
374 PAGE_ACTIVATE,
375 /* page has been sent to the disk successfully, page is unlocked */
376 PAGE_SUCCESS,
377 /* page is clean and locked */
378 PAGE_CLEAN,
379} pageout_t;
380
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381/*
Andrew Morton1742f192006-03-22 00:08:21 -0800382 * pageout is called by shrink_page_list() for each dirty page.
383 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700385static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700386 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387{
388 /*
389 * If the page is dirty, only perform writeback if that write
390 * will be non-blocking. To prevent this allocation from being
391 * stalled by pagecache activity. But note that there may be
392 * stalls if we need to run get_block(). We could test
393 * PagePrivate for that.
394 *
Vincent Li6aceb532009-12-14 17:58:49 -0800395 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 * this page's queue, we can perform writeback even if that
397 * will block.
398 *
399 * If the page is swapcache, write it back even if that would
400 * block, for some throttling. This happens by accident, because
401 * swap_backing_dev_info is bust: it doesn't reflect the
402 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 */
404 if (!is_page_cache_freeable(page))
405 return PAGE_KEEP;
406 if (!mapping) {
407 /*
408 * Some data journaling orphaned pages can have
409 * page->mapping == NULL while being dirty with clean buffers.
410 */
David Howells266cf652009-04-03 16:42:36 +0100411 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 if (try_to_free_buffers(page)) {
413 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700414 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 return PAGE_CLEAN;
416 }
417 }
418 return PAGE_KEEP;
419 }
420 if (mapping->a_ops->writepage == NULL)
421 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700422 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 return PAGE_KEEP;
424
425 if (clear_page_dirty_for_io(page)) {
426 int res;
427 struct writeback_control wbc = {
428 .sync_mode = WB_SYNC_NONE,
429 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700430 .range_start = 0,
431 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 .for_reclaim = 1,
433 };
434
435 SetPageReclaim(page);
436 res = mapping->a_ops->writepage(page, &wbc);
437 if (res < 0)
438 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800439 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 ClearPageReclaim(page);
441 return PAGE_ACTIVATE;
442 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700443
444 /*
445 * Wait on writeback if requested to. This happens when
446 * direct reclaiming a large contiguous area and the
447 * first attempt to free a range of pages fails.
448 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700449 if (PageWriteback(page) &&
Mel Gormanf3a310b2011-01-13 15:46:00 -0800450 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
Andy Whitcroftc661b072007-08-22 14:01:26 -0700451 wait_on_page_writeback(page);
452
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 if (!PageWriteback(page)) {
454 /* synchronous write or broken a_ops? */
455 ClearPageReclaim(page);
456 }
Mel Gorman755f0222010-08-09 17:19:18 -0700457 trace_mm_vmscan_writepage(page,
Mel Gormanf3a310b2011-01-13 15:46:00 -0800458 trace_reclaim_flags(page, sc->reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700459 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 return PAGE_SUCCESS;
461 }
462
463 return PAGE_CLEAN;
464}
465
Andrew Mortona649fd92006-10-17 00:09:36 -0700466/*
Nick Piggine2867812008-07-25 19:45:30 -0700467 * Same as remove_mapping, but if the page is removed from the mapping, it
468 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700469 */
Nick Piggine2867812008-07-25 19:45:30 -0700470static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800471{
Nick Piggin28e4d962006-09-25 23:31:23 -0700472 BUG_ON(!PageLocked(page));
473 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800474
Nick Piggin19fd6232008-07-25 19:45:32 -0700475 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800476 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700477 * The non racy check for a busy page.
478 *
479 * Must be careful with the order of the tests. When someone has
480 * a ref to the page, it may be possible that they dirty it then
481 * drop the reference. So if PageDirty is tested before page_count
482 * here, then the following race may occur:
483 *
484 * get_user_pages(&page);
485 * [user mapping goes away]
486 * write_to(page);
487 * !PageDirty(page) [good]
488 * SetPageDirty(page);
489 * put_page(page);
490 * !page_count(page) [good, discard it]
491 *
492 * [oops, our write_to data is lost]
493 *
494 * Reversing the order of the tests ensures such a situation cannot
495 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
496 * load is not satisfied before that of page->_count.
497 *
498 * Note that if SetPageDirty is always performed via set_page_dirty,
499 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800500 */
Nick Piggine2867812008-07-25 19:45:30 -0700501 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800502 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700503 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
504 if (unlikely(PageDirty(page))) {
505 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800506 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700507 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800508
509 if (PageSwapCache(page)) {
510 swp_entry_t swap = { .val = page_private(page) };
511 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700512 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700513 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700514 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500515 void (*freepage)(struct page *);
516
517 freepage = mapping->a_ops->freepage;
518
Minchan Kime64a7822011-03-22 16:32:44 -0700519 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700520 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700521 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500522
523 if (freepage != NULL)
524 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800525 }
526
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800527 return 1;
528
529cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700530 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800531 return 0;
532}
533
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534/*
Nick Piggine2867812008-07-25 19:45:30 -0700535 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
536 * someone else has a ref on the page, abort and return 0. If it was
537 * successfully detached, return 1. Assumes the caller has a single ref on
538 * this page.
539 */
540int remove_mapping(struct address_space *mapping, struct page *page)
541{
542 if (__remove_mapping(mapping, page)) {
543 /*
544 * Unfreezing the refcount with 1 rather than 2 effectively
545 * drops the pagecache ref for us without requiring another
546 * atomic operation.
547 */
548 page_unfreeze_refs(page, 1);
549 return 1;
550 }
551 return 0;
552}
553
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700554/**
555 * putback_lru_page - put previously isolated page onto appropriate LRU list
556 * @page: page to be put back to appropriate lru list
557 *
558 * Add previously isolated @page to appropriate LRU list.
559 * Page may still be unevictable for other reasons.
560 *
561 * lru_lock must not be held, interrupts must be enabled.
562 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700563void putback_lru_page(struct page *page)
564{
565 int lru;
566 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700567 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700568
569 VM_BUG_ON(PageLRU(page));
570
571redo:
572 ClearPageUnevictable(page);
573
574 if (page_evictable(page, NULL)) {
575 /*
576 * For evictable pages, we can use the cache.
577 * In event of a race, worst case is we end up with an
578 * unevictable page on [in]active list.
579 * We know how to handle that.
580 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700581 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700582 lru_cache_add_lru(page, lru);
583 } else {
584 /*
585 * Put unevictable pages directly on zone's unevictable
586 * list.
587 */
588 lru = LRU_UNEVICTABLE;
589 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700590 /*
591 * When racing with an mlock clearing (page is
592 * unlocked), make sure that if the other thread does
593 * not observe our setting of PG_lru and fails
594 * isolation, we see PG_mlocked cleared below and move
595 * the page back to the evictable list.
596 *
597 * The other side is TestClearPageMlocked().
598 */
599 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700600 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700601
602 /*
603 * page's status can change while we move it among lru. If an evictable
604 * page is on unevictable list, it never be freed. To avoid that,
605 * check after we added it to the list, again.
606 */
607 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
608 if (!isolate_lru_page(page)) {
609 put_page(page);
610 goto redo;
611 }
612 /* This means someone else dropped this page from LRU
613 * So, it will be freed or putback to LRU again. There is
614 * nothing to do here.
615 */
616 }
617
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700618 if (was_unevictable && lru != LRU_UNEVICTABLE)
619 count_vm_event(UNEVICTABLE_PGRESCUED);
620 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
621 count_vm_event(UNEVICTABLE_PGCULLED);
622
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700623 put_page(page); /* drop ref from isolate */
624}
625
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800626enum page_references {
627 PAGEREF_RECLAIM,
628 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800629 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800630 PAGEREF_ACTIVATE,
631};
632
633static enum page_references page_check_references(struct page *page,
634 struct scan_control *sc)
635{
Johannes Weiner64574742010-03-05 13:42:22 -0800636 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800637 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800638
Johannes Weiner64574742010-03-05 13:42:22 -0800639 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
640 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800641
642 /* Lumpy reclaim - ignore references */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800643 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800644 return PAGEREF_RECLAIM;
645
646 /*
647 * Mlock lost the isolation race with us. Let try_to_unmap()
648 * move the page to the unevictable list.
649 */
650 if (vm_flags & VM_LOCKED)
651 return PAGEREF_RECLAIM;
652
Johannes Weiner64574742010-03-05 13:42:22 -0800653 if (referenced_ptes) {
654 if (PageAnon(page))
655 return PAGEREF_ACTIVATE;
656 /*
657 * All mapped pages start out with page table
658 * references from the instantiating fault, so we need
659 * to look twice if a mapped file page is used more
660 * than once.
661 *
662 * Mark it and spare it for another trip around the
663 * inactive list. Another page table reference will
664 * lead to its activation.
665 *
666 * Note: the mark is set for activated pages as well
667 * so that recently deactivated but used pages are
668 * quickly recovered.
669 */
670 SetPageReferenced(page);
671
672 if (referenced_page)
673 return PAGEREF_ACTIVATE;
674
675 return PAGEREF_KEEP;
676 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800677
678 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700679 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner64574742010-03-05 13:42:22 -0800680 return PAGEREF_RECLAIM_CLEAN;
681
682 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800683}
684
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700685static noinline_for_stack void free_page_list(struct list_head *free_pages)
686{
687 struct pagevec freed_pvec;
688 struct page *page, *tmp;
689
690 pagevec_init(&freed_pvec, 1);
691
692 list_for_each_entry_safe(page, tmp, free_pages, lru) {
693 list_del(&page->lru);
694 if (!pagevec_add(&freed_pvec, page)) {
695 __pagevec_free(&freed_pvec);
696 pagevec_reinit(&freed_pvec);
697 }
698 }
699
700 pagevec_free(&freed_pvec);
701}
702
Nick Piggine2867812008-07-25 19:45:30 -0700703/*
Andrew Morton1742f192006-03-22 00:08:21 -0800704 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 */
Andrew Morton1742f192006-03-22 00:08:21 -0800706static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700707 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700708 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
710 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700711 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700713 unsigned long nr_dirty = 0;
714 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800715 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716
717 cond_resched();
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800720 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 struct address_space *mapping;
722 struct page *page;
723 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
725 cond_resched();
726
727 page = lru_to_page(page_list);
728 list_del(&page->lru);
729
Nick Piggin529ae9a2008-08-02 12:01:03 +0200730 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 goto keep;
732
Nick Piggin725d7042006-09-25 23:30:55 -0700733 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700734 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
736 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800737
Nick Pigginb291f002008-10-18 20:26:44 -0700738 if (unlikely(!page_evictable(page, NULL)))
739 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700740
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700741 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800742 goto keep_locked;
743
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 /* Double the slab pressure for mapped and swapcache pages */
745 if (page_mapped(page) || PageSwapCache(page))
746 sc->nr_scanned++;
747
Andy Whitcroftc661b072007-08-22 14:01:26 -0700748 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
749 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
750
751 if (PageWriteback(page)) {
752 /*
753 * Synchronous reclaim is performed in two passes,
754 * first an asynchronous pass over the list to
755 * start parallel writeback, and a second synchronous
756 * pass to wait for the IO to complete. Wait here
757 * for any page for which writeback has already
758 * started.
759 */
Mel Gormanf3a310b2011-01-13 15:46:00 -0800760 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700761 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700762 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700763 else {
764 unlock_page(page);
765 goto keep_lumpy;
766 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700767 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800769 references = page_check_references(page, sc);
770 switch (references) {
771 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800773 case PAGEREF_KEEP:
774 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800775 case PAGEREF_RECLAIM:
776 case PAGEREF_RECLAIM_CLEAN:
777 ; /* try to reclaim the page below */
778 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 /*
781 * Anonymous process memory has backing store?
782 * Try to allocate it some swap space here.
783 */
Nick Pigginb291f002008-10-18 20:26:44 -0700784 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800785 if (!(sc->gfp_mask & __GFP_IO))
786 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800787 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800789 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700790 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
792 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793
794 /*
795 * The page is mapped into the page tables of one or more
796 * processes. Try to unmap it here.
797 */
798 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200799 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 case SWAP_FAIL:
801 goto activate_locked;
802 case SWAP_AGAIN:
803 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700804 case SWAP_MLOCK:
805 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 case SWAP_SUCCESS:
807 ; /* try to free the page below */
808 }
809 }
810
811 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700812 nr_dirty++;
813
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800814 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700816 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800818 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 goto keep_locked;
820
821 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700822 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700824 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 goto keep_locked;
826 case PAGE_ACTIVATE:
827 goto activate_locked;
828 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700829 if (PageWriteback(page))
830 goto keep_lumpy;
831 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700833
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 /*
835 * A synchronous write - probably a ramdisk. Go
836 * ahead and try to reclaim the page.
837 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200838 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 goto keep;
840 if (PageDirty(page) || PageWriteback(page))
841 goto keep_locked;
842 mapping = page_mapping(page);
843 case PAGE_CLEAN:
844 ; /* try to free the page below */
845 }
846 }
847
848 /*
849 * If the page has buffers, try to free the buffer mappings
850 * associated with this page. If we succeed we try to free
851 * the page as well.
852 *
853 * We do this even if the page is PageDirty().
854 * try_to_release_page() does not perform I/O, but it is
855 * possible for a page to have PageDirty set, but it is actually
856 * clean (all its buffers are clean). This happens if the
857 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700858 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 * try_to_release_page() will discover that cleanness and will
860 * drop the buffers and mark the page clean - it can be freed.
861 *
862 * Rarely, pages can have buffers and no ->mapping. These are
863 * the pages which were not successfully invalidated in
864 * truncate_complete_page(). We try to drop those buffers here
865 * and if that worked, and the page is no longer mapped into
866 * process address space (page_count == 1) it can be freed.
867 * Otherwise, leave the page on the LRU so it is swappable.
868 */
David Howells266cf652009-04-03 16:42:36 +0100869 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 if (!try_to_release_page(page, sc->gfp_mask))
871 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700872 if (!mapping && page_count(page) == 1) {
873 unlock_page(page);
874 if (put_page_testzero(page))
875 goto free_it;
876 else {
877 /*
878 * rare race with speculative reference.
879 * the speculative reference will free
880 * this page shortly, so we may
881 * increment nr_reclaimed here (and
882 * leave it off the LRU).
883 */
884 nr_reclaimed++;
885 continue;
886 }
887 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 }
889
Nick Piggine2867812008-07-25 19:45:30 -0700890 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800891 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892
Nick Piggina978d6f2008-10-18 20:26:58 -0700893 /*
894 * At this point, we have no other references and there is
895 * no way to pick any more up (removed from LRU, removed
896 * from pagecache). Can use non-atomic bitops now (and
897 * we obviously don't have to worry about waking up a process
898 * waiting on the page lock, because there are no references.
899 */
900 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700901free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800902 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700903
904 /*
905 * Is there need to periodically free_page_list? It would
906 * appear not as the counts should be low
907 */
908 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 continue;
910
Nick Pigginb291f002008-10-18 20:26:44 -0700911cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800912 if (PageSwapCache(page))
913 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700914 unlock_page(page);
915 putback_lru_page(page);
Mel Gormanf3a310b2011-01-13 15:46:00 -0800916 reset_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700917 continue;
918
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700920 /* Not a candidate for swapping, so reclaim swap space. */
921 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800922 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700923 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 SetPageActive(page);
925 pgactivate++;
926keep_locked:
927 unlock_page(page);
928keep:
Mel Gormanf3a310b2011-01-13 15:46:00 -0800929 reset_reclaim_mode(sc);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700930keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700932 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700934
Mel Gorman0e093d992010-10-26 14:21:45 -0700935 /*
936 * Tag a zone as congested if all the dirty pages encountered were
937 * backed by a congested BDI. In this case, reclaimers should just
938 * back off and wait for congestion to clear because further reclaim
939 * will encounter the same problem
940 */
KAMEZAWA Hiroyukid6c438b2011-05-17 15:44:10 -0700941 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
Mel Gorman0e093d992010-10-26 14:21:45 -0700942 zone_set_flag(zone, ZONE_CONGESTED);
943
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700944 free_page_list(&free_pages);
945
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700947 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800948 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949}
950
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700951/*
952 * Attempt to remove the specified page from its LRU. Only take this page
953 * if it is of the appropriate PageActive status. Pages which are being
954 * freed elsewhere are also ignored.
955 *
956 * page: page to consider
957 * mode: one of the LRU isolation modes defined above
958 *
959 * returns 0 on success, -ve errno on failure.
960 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700961int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700962{
963 int ret = -EINVAL;
964
965 /* Only take pages on the LRU. */
966 if (!PageLRU(page))
967 return ret;
968
969 /*
970 * When checking the active state, we need to be sure we are
971 * dealing with comparible boolean values. Take the logical not
972 * of each.
973 */
974 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
975 return ret;
976
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700977 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700978 return ret;
979
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700980 /*
981 * When this function is being called for lumpy reclaim, we
982 * initially look into all LRU pages, active, inactive and
983 * unevictable; only give shrink_page_list evictable pages.
984 */
985 if (PageUnevictable(page))
986 return ret;
987
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700988 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800989
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700990 if (likely(get_page_unless_zero(page))) {
991 /*
992 * Be careful not to clear PageLRU until after we're
993 * sure the page is not being freed elsewhere -- the
994 * page release code relies on it.
995 */
996 ClearPageLRU(page);
997 ret = 0;
998 }
999
1000 return ret;
1001}
1002
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001003/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 * zone->lru_lock is heavily contended. Some of the functions that
1005 * shrink the lists perform better by taking out a batch of pages
1006 * and working on them outside the LRU lock.
1007 *
1008 * For pagecache intensive workloads, this function is the hottest
1009 * spot in the kernel (apart from copy_*_user functions).
1010 *
1011 * Appropriate locks must be held before calling this function.
1012 *
1013 * @nr_to_scan: The number of pages to look through on the list.
1014 * @src: The LRU list to pull pages off.
1015 * @dst: The temp list to put pages on to.
1016 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001017 * @order: The caller's attempted allocation order
1018 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001019 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 *
1021 * returns how many pages were moved onto *@dst.
1022 */
Andrew Morton69e05942006-03-22 00:08:19 -08001023static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1024 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001025 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
Andrew Morton69e05942006-03-22 00:08:19 -08001027 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001028 unsigned long nr_lumpy_taken = 0;
1029 unsigned long nr_lumpy_dirty = 0;
1030 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001031 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001033 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001034 struct page *page;
1035 unsigned long pfn;
1036 unsigned long end_pfn;
1037 unsigned long page_pfn;
1038 int zone_id;
1039
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 page = lru_to_page(src);
1041 prefetchw_prev_lru_page(page, src, flags);
1042
Nick Piggin725d7042006-09-25 23:30:55 -07001043 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001044
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001045 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001046 case 0:
1047 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001048 mem_cgroup_del_lru(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001049 nr_taken += hpage_nr_pages(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001050 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001051
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001052 case -EBUSY:
1053 /* else it is being freed elsewhere */
1054 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001055 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001056 continue;
1057
1058 default:
1059 BUG();
1060 }
1061
1062 if (!order)
1063 continue;
1064
1065 /*
1066 * Attempt to take all pages in the order aligned region
1067 * surrounding the tag page. Only take those pages of
1068 * the same active state as that tag page. We may safely
1069 * round the target page pfn down to the requested order
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001070 * as the mem_map is guaranteed valid out to MAX_ORDER,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001071 * where that page is in a different zone we will detect
1072 * it from its zone id and abort this block scan.
1073 */
1074 zone_id = page_zone_id(page);
1075 page_pfn = page_to_pfn(page);
1076 pfn = page_pfn & ~((1 << order) - 1);
1077 end_pfn = pfn + (1 << order);
1078 for (; pfn < end_pfn; pfn++) {
1079 struct page *cursor_page;
1080
1081 /* The target page is in the block, ignore it. */
1082 if (unlikely(pfn == page_pfn))
1083 continue;
1084
1085 /* Avoid holes within the zone. */
1086 if (unlikely(!pfn_valid_within(pfn)))
1087 break;
1088
1089 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001090
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001091 /* Check that we have not crossed a zone boundary. */
1092 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001093 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001094
1095 /*
1096 * If we don't have enough swap space, reclaiming of
1097 * anon page which don't already have a swap slot is
1098 * pointless.
1099 */
1100 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001101 !PageSwapCache(cursor_page))
1102 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001103
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001104 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001105 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001106 mem_cgroup_del_lru(cursor_page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001107 nr_taken += hpage_nr_pages(page);
Mel Gormana8a94d12010-08-09 17:19:17 -07001108 nr_lumpy_taken++;
1109 if (PageDirty(cursor_page))
1110 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001111 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001112 } else {
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001113 /* the page is freed already. */
1114 if (!page_count(cursor_page))
1115 continue;
1116 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001117 }
1118 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001119
1120 /* If we break out of the loop above, lumpy reclaim failed */
1121 if (pfn < end_pfn)
1122 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 }
1124
1125 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001126
1127 trace_mm_vmscan_lru_isolate(order,
1128 nr_to_scan, scan,
1129 nr_taken,
1130 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1131 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 return nr_taken;
1133}
1134
Balbir Singh66e17072008-02-07 00:13:56 -08001135static unsigned long isolate_pages_global(unsigned long nr,
1136 struct list_head *dst,
1137 unsigned long *scanned, int order,
1138 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001139 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001140{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001141 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001142 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001143 lru += LRU_ACTIVE;
1144 if (file)
1145 lru += LRU_FILE;
1146 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001147 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001148}
1149
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001151 * clear_active_flags() is a helper for shrink_active_list(), clearing
1152 * any active bits from the pages in the list.
1153 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001154static unsigned long clear_active_flags(struct list_head *page_list,
1155 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001156{
1157 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001158 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001159 struct page *page;
1160
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001161 list_for_each_entry(page, page_list, lru) {
Rik van Riel2c888cf2011-01-13 15:47:13 -08001162 int numpages = hpage_nr_pages(page);
Johannes Weiner401a8e12009-09-21 17:02:58 -07001163 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001164 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001165 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001166 ClearPageActive(page);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001167 nr_active += numpages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001168 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001169 if (count)
Rik van Riel2c888cf2011-01-13 15:47:13 -08001170 count[lru] += numpages;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001171 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001172
1173 return nr_active;
1174}
1175
Nick Piggin62695a82008-10-18 20:26:09 -07001176/**
1177 * isolate_lru_page - tries to isolate a page from its LRU list
1178 * @page: page to isolate from its LRU list
1179 *
1180 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1181 * vmstat statistic corresponding to whatever LRU list the page was on.
1182 *
1183 * Returns 0 if the page was removed from an LRU list.
1184 * Returns -EBUSY if the page was not on an LRU list.
1185 *
1186 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001187 * the active list, it will have PageActive set. If it was found on
1188 * the unevictable list, it will have the PageUnevictable bit set. That flag
1189 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001190 *
1191 * The vmstat statistic corresponding to the list on which the page was
1192 * found will be decremented.
1193 *
1194 * Restrictions:
1195 * (1) Must be called with an elevated refcount on the page. This is a
1196 * fundamentnal difference from isolate_lru_pages (which is called
1197 * without a stable reference).
1198 * (2) the lru_lock must not be held.
1199 * (3) interrupts must be enabled.
1200 */
1201int isolate_lru_page(struct page *page)
1202{
1203 int ret = -EBUSY;
1204
1205 if (PageLRU(page)) {
1206 struct zone *zone = page_zone(page);
1207
1208 spin_lock_irq(&zone->lru_lock);
1209 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001210 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001211 ret = 0;
1212 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001213
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001214 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001215 }
1216 spin_unlock_irq(&zone->lru_lock);
1217 }
1218 return ret;
1219}
1220
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001221/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001222 * Are there way too many processes in the direct reclaim path already?
1223 */
1224static int too_many_isolated(struct zone *zone, int file,
1225 struct scan_control *sc)
1226{
1227 unsigned long inactive, isolated;
1228
1229 if (current_is_kswapd())
1230 return 0;
1231
1232 if (!scanning_global_lru(sc))
1233 return 0;
1234
1235 if (file) {
1236 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1237 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1238 } else {
1239 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1240 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1241 }
1242
1243 return isolated > inactive;
1244}
1245
1246/*
Mel Gorman66635622010-08-09 17:19:30 -07001247 * TODO: Try merging with migrations version of putback_lru_pages
1248 */
1249static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001250putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001251 unsigned long nr_anon, unsigned long nr_file,
1252 struct list_head *page_list)
1253{
1254 struct page *page;
1255 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001256 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001257
1258 pagevec_init(&pvec, 1);
1259
1260 /*
1261 * Put back any unfreeable pages.
1262 */
1263 spin_lock(&zone->lru_lock);
1264 while (!list_empty(page_list)) {
1265 int lru;
1266 page = lru_to_page(page_list);
1267 VM_BUG_ON(PageLRU(page));
1268 list_del(&page->lru);
1269 if (unlikely(!page_evictable(page, NULL))) {
1270 spin_unlock_irq(&zone->lru_lock);
1271 putback_lru_page(page);
1272 spin_lock_irq(&zone->lru_lock);
1273 continue;
1274 }
Linus Torvalds7a608572011-01-17 14:42:19 -08001275 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001276 lru = page_lru(page);
Linus Torvalds7a608572011-01-17 14:42:19 -08001277 add_page_to_lru_list(zone, page, lru);
Mel Gorman66635622010-08-09 17:19:30 -07001278 if (is_active_lru(lru)) {
1279 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001280 int numpages = hpage_nr_pages(page);
1281 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001282 }
1283 if (!pagevec_add(&pvec, page)) {
1284 spin_unlock_irq(&zone->lru_lock);
1285 __pagevec_release(&pvec);
1286 spin_lock_irq(&zone->lru_lock);
1287 }
1288 }
1289 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1290 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1291
1292 spin_unlock_irq(&zone->lru_lock);
1293 pagevec_release(&pvec);
1294}
1295
Mel Gorman1489fa12010-08-09 17:19:33 -07001296static noinline_for_stack void update_isolated_counts(struct zone *zone,
1297 struct scan_control *sc,
1298 unsigned long *nr_anon,
1299 unsigned long *nr_file,
1300 struct list_head *isolated_list)
1301{
1302 unsigned long nr_active;
1303 unsigned int count[NR_LRU_LISTS] = { 0, };
1304 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1305
1306 nr_active = clear_active_flags(isolated_list, count);
1307 __count_vm_events(PGDEACTIVATE, nr_active);
1308
1309 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1310 -count[LRU_ACTIVE_FILE]);
1311 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1312 -count[LRU_INACTIVE_FILE]);
1313 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1314 -count[LRU_ACTIVE_ANON]);
1315 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1316 -count[LRU_INACTIVE_ANON]);
1317
1318 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1319 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1320 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1321 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1322
1323 reclaim_stat->recent_scanned[0] += *nr_anon;
1324 reclaim_stat->recent_scanned[1] += *nr_file;
1325}
1326
Mel Gorman66635622010-08-09 17:19:30 -07001327/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001328 * Returns true if the caller should wait to clean dirty/writeback pages.
1329 *
1330 * If we are direct reclaiming for contiguous pages and we do not reclaim
1331 * everything in the list, try again and wait for writeback IO to complete.
1332 * This will stall high-order allocations noticeably. Only do that when really
1333 * need to free the pages under high memory pressure.
1334 */
1335static inline bool should_reclaim_stall(unsigned long nr_taken,
1336 unsigned long nr_freed,
1337 int priority,
1338 struct scan_control *sc)
1339{
1340 int lumpy_stall_priority;
1341
1342 /* kswapd should not stall on sync IO */
1343 if (current_is_kswapd())
1344 return false;
1345
1346 /* Only stall on lumpy reclaim */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001347 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001348 return false;
1349
1350 /* If we have relaimed everything on the isolated list, no stall */
1351 if (nr_freed == nr_taken)
1352 return false;
1353
1354 /*
1355 * For high-order allocations, there are two stall thresholds.
1356 * High-cost allocations stall immediately where as lower
1357 * order allocations such as stacks require the scanning
1358 * priority to be much higher before stalling.
1359 */
1360 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1361 lumpy_stall_priority = DEF_PRIORITY;
1362 else
1363 lumpy_stall_priority = DEF_PRIORITY / 3;
1364
1365 return priority <= lumpy_stall_priority;
1366}
1367
1368/*
Andrew Morton1742f192006-03-22 00:08:21 -08001369 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1370 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 */
Mel Gorman66635622010-08-09 17:19:30 -07001372static noinline_for_stack unsigned long
1373shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1374 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375{
1376 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001377 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001378 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001379 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001380 unsigned long nr_anon;
1381 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001382
Rik van Riel35cd7812009-09-21 17:01:38 -07001383 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001384 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001385
1386 /* We are about to die and free our memory. Return now. */
1387 if (fatal_signal_pending(current))
1388 return SWAP_CLUSTER_MAX;
1389 }
1390
Mel Gormanf3a310b2011-01-13 15:46:00 -08001391 set_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 lru_add_drain();
1393 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001395 if (scanning_global_lru(sc)) {
1396 nr_taken = isolate_pages_global(nr_to_scan,
1397 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001398 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001399 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001400 zone, 0, file);
1401 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001402 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001403 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1404 nr_scanned);
1405 else
1406 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1407 nr_scanned);
1408 } else {
1409 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1410 &page_list, &nr_scanned, sc->order,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001411 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
Mel Gorman3e7d3442011-01-13 15:45:56 -08001412 ISOLATE_BOTH : ISOLATE_INACTIVE,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001413 zone, sc->mem_cgroup,
1414 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001416 * mem_cgroup_isolate_pages() keeps track of
1417 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001419 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001420
Mel Gorman66635622010-08-09 17:19:30 -07001421 if (nr_taken == 0) {
1422 spin_unlock_irq(&zone->lru_lock);
1423 return 0;
1424 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001425
Mel Gorman1489fa12010-08-09 17:19:33 -07001426 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001427
1428 spin_unlock_irq(&zone->lru_lock);
1429
Mel Gorman0e093d992010-10-26 14:21:45 -07001430 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001431
Wu Fengguange31f36982010-08-09 17:20:01 -07001432 /* Check if we should syncronously wait for writeback */
1433 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
Mel Gormanf3a310b2011-01-13 15:46:00 -08001434 set_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001435 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001436 }
1437
1438 local_irq_disable();
1439 if (current_is_kswapd())
1440 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1441 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1442
Mel Gorman1489fa12010-08-09 17:19:33 -07001443 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001444
1445 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1446 zone_idx(zone),
1447 nr_scanned, nr_reclaimed,
1448 priority,
Mel Gormanf3a310b2011-01-13 15:46:00 -08001449 trace_shrink_flags(file, sc->reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001450 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451}
1452
Martin Bligh3bb1a852006-10-28 10:38:24 -07001453/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 * This moves pages from the active list to the inactive list.
1455 *
1456 * We move them the other way if the page is referenced by one or more
1457 * processes, from rmap.
1458 *
1459 * If the pages are mostly unmapped, the processing is fast and it is
1460 * appropriate to hold zone->lru_lock across the whole operation. But if
1461 * the pages are mapped, the processing is slow (page_referenced()) so we
1462 * should drop zone->lru_lock around each page. It's impossible to balance
1463 * this, so instead we remove the pages from the LRU while processing them.
1464 * It is safe to rely on PG_active against the non-LRU pages in here because
1465 * nobody will play with that bit on a non-LRU page.
1466 *
1467 * The downside is that we have to touch page->_count against each page.
1468 * But we had to alter page->flags anyway.
1469 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001470
Wu Fengguang3eb41402009-06-16 15:33:13 -07001471static void move_active_pages_to_lru(struct zone *zone,
1472 struct list_head *list,
1473 enum lru_list lru)
1474{
1475 unsigned long pgmoved = 0;
1476 struct pagevec pvec;
1477 struct page *page;
1478
1479 pagevec_init(&pvec, 1);
1480
1481 while (!list_empty(list)) {
1482 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001483
1484 VM_BUG_ON(PageLRU(page));
1485 SetPageLRU(page);
1486
Wu Fengguang3eb41402009-06-16 15:33:13 -07001487 list_move(&page->lru, &zone->lru[lru].list);
1488 mem_cgroup_add_lru_list(page, lru);
Rik van Riel2c888cf2011-01-13 15:47:13 -08001489 pgmoved += hpage_nr_pages(page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001490
1491 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1492 spin_unlock_irq(&zone->lru_lock);
1493 if (buffer_heads_over_limit)
1494 pagevec_strip(&pvec);
1495 __pagevec_release(&pvec);
1496 spin_lock_irq(&zone->lru_lock);
1497 }
1498 }
1499 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1500 if (!is_active_lru(lru))
1501 __count_vm_events(PGDEACTIVATE, pgmoved);
1502}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001503
Andrew Morton1742f192006-03-22 00:08:21 -08001504static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001505 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001507 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001508 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001509 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001511 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001512 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001514 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001515 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516
1517 lru_add_drain();
1518 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001519 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001520 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1521 &pgscanned, sc->order,
1522 ISOLATE_ACTIVE, zone,
1523 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001524 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001525 } else {
1526 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1527 &pgscanned, sc->order,
1528 ISOLATE_ACTIVE, zone,
1529 sc->mem_cgroup, 1, file);
1530 /*
1531 * mem_cgroup_isolate_pages() keeps track of
1532 * scanned pages on its own.
1533 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001534 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001535
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001536 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001537
Wu Fengguang3eb41402009-06-16 15:33:13 -07001538 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001539 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001540 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001541 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001542 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001543 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 spin_unlock_irq(&zone->lru_lock);
1545
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 while (!list_empty(&l_hold)) {
1547 cond_resched();
1548 page = lru_to_page(&l_hold);
1549 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001550
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001551 if (unlikely(!page_evictable(page, NULL))) {
1552 putback_lru_page(page);
1553 continue;
1554 }
1555
Johannes Weiner64574742010-03-05 13:42:22 -08001556 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001557 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001558 /*
1559 * Identify referenced, file-backed active pages and
1560 * give them one more trip around the active list. So
1561 * that executable code get better chances to stay in
1562 * memory under moderate memory pressure. Anon pages
1563 * are not likely to be evicted by use-once streaming
1564 * IO, plus JVM can create lots of anon VM_EXEC pages,
1565 * so we ignore them here.
1566 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001567 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001568 list_add(&page->lru, &l_active);
1569 continue;
1570 }
1571 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001572
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001573 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 list_add(&page->lru, &l_inactive);
1575 }
1576
Andrew Mortonb5557492009-01-06 14:40:13 -08001577 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001578 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001579 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001580 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001581 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001582 * Count referenced pages from currently used mappings as rotated,
1583 * even though only some of them are actually re-activated. This
1584 * helps balance scan pressure between file and anonymous pages in
1585 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001586 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001587 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001588
Wu Fengguang3eb41402009-06-16 15:33:13 -07001589 move_active_pages_to_lru(zone, &l_active,
1590 LRU_ACTIVE + file * LRU_FILE);
1591 move_active_pages_to_lru(zone, &l_inactive,
1592 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001593 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001594 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595}
1596
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001597#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001598static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001599{
1600 unsigned long active, inactive;
1601
1602 active = zone_page_state(zone, NR_ACTIVE_ANON);
1603 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1604
1605 if (inactive * zone->inactive_ratio < active)
1606 return 1;
1607
1608 return 0;
1609}
1610
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001611/**
1612 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1613 * @zone: zone to check
1614 * @sc: scan control of this context
1615 *
1616 * Returns true if the zone does not have enough inactive anon pages,
1617 * meaning some active anon pages need to be deactivated.
1618 */
1619static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1620{
1621 int low;
1622
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001623 /*
1624 * If we don't have swap space, anonymous page deactivation
1625 * is pointless.
1626 */
1627 if (!total_swap_pages)
1628 return 0;
1629
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001630 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001631 low = inactive_anon_is_low_global(zone);
1632 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001633 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001634 return low;
1635}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001636#else
1637static inline int inactive_anon_is_low(struct zone *zone,
1638 struct scan_control *sc)
1639{
1640 return 0;
1641}
1642#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001643
Rik van Riel56e49d22009-06-16 15:32:28 -07001644static int inactive_file_is_low_global(struct zone *zone)
1645{
1646 unsigned long active, inactive;
1647
1648 active = zone_page_state(zone, NR_ACTIVE_FILE);
1649 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1650
1651 return (active > inactive);
1652}
1653
1654/**
1655 * inactive_file_is_low - check if file pages need to be deactivated
1656 * @zone: zone to check
1657 * @sc: scan control of this context
1658 *
1659 * When the system is doing streaming IO, memory pressure here
1660 * ensures that active file pages get deactivated, until more
1661 * than half of the file pages are on the inactive list.
1662 *
1663 * Once we get to that situation, protect the system's working
1664 * set from being evicted by disabling active file page aging.
1665 *
1666 * This uses a different ratio than the anonymous pages, because
1667 * the page cache uses a use-once replacement algorithm.
1668 */
1669static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1670{
1671 int low;
1672
1673 if (scanning_global_lru(sc))
1674 low = inactive_file_is_low_global(zone);
1675 else
1676 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1677 return low;
1678}
1679
Rik van Rielb39415b2009-12-14 17:59:48 -08001680static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1681 int file)
1682{
1683 if (file)
1684 return inactive_file_is_low(zone, sc);
1685 else
1686 return inactive_anon_is_low(zone, sc);
1687}
1688
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001689static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001690 struct zone *zone, struct scan_control *sc, int priority)
1691{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001692 int file = is_file_lru(lru);
1693
Rik van Rielb39415b2009-12-14 17:59:48 -08001694 if (is_active_lru(lru)) {
1695 if (inactive_list_is_low(zone, sc, file))
1696 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001697 return 0;
1698 }
1699
Rik van Riel33c120e2008-10-18 20:26:36 -07001700 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001701}
1702
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001704 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1705 * until we collected @swap_cluster_max pages to scan.
1706 */
1707static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1708 unsigned long *nr_saved_scan)
1709{
1710 unsigned long nr;
1711
1712 *nr_saved_scan += nr_to_scan;
1713 nr = *nr_saved_scan;
1714
1715 if (nr >= SWAP_CLUSTER_MAX)
1716 *nr_saved_scan = 0;
1717 else
1718 nr = 0;
1719
1720 return nr;
1721}
1722
1723/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001724 * Determine how aggressively the anon and file LRU lists should be
1725 * scanned. The relative value of each set of LRU lists is determined
1726 * by looking at the fraction of the pages scanned we did rotate back
1727 * onto the active list instead of evict.
1728 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001729 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001730 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001731static void get_scan_count(struct zone *zone, struct scan_control *sc,
1732 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001733{
1734 unsigned long anon, file, free;
1735 unsigned long anon_prio, file_prio;
1736 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001737 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001738 u64 fraction[2], denominator;
1739 enum lru_list l;
1740 int noswap = 0;
1741
1742 /* If we have no swap space, do not bother scanning anon pages. */
1743 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1744 noswap = 1;
1745 fraction[0] = 0;
1746 fraction[1] = 1;
1747 denominator = 1;
1748 goto out;
1749 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001750
Vincent Li0b217672009-09-21 17:03:09 -07001751 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1752 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1753 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1754 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001755
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001756 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001757 free = zone_page_state(zone, NR_FREE_PAGES);
1758 /* If we have very few page cache pages,
1759 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001760 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001761 fraction[0] = 1;
1762 fraction[1] = 0;
1763 denominator = 1;
1764 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001765 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001766 }
1767
1768 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001769 * With swappiness at 100, anonymous and file have the same priority.
1770 * This scanning priority is essentially the inverse of IO cost.
1771 */
1772 anon_prio = sc->swappiness;
1773 file_prio = 200 - sc->swappiness;
1774
1775 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001776 * OK, so we have swap space and a fair amount of page cache
1777 * pages. We use the recently rotated / recently scanned
1778 * ratios to determine how valuable each cache is.
1779 *
1780 * Because workloads change over time (and to avoid overflow)
1781 * we keep these statistics as a floating average, which ends
1782 * up weighing recent references more than old ones.
1783 *
1784 * anon in [0], file in [1]
1785 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001786 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001787 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001788 reclaim_stat->recent_scanned[0] /= 2;
1789 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001790 }
1791
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001792 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001793 reclaim_stat->recent_scanned[1] /= 2;
1794 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001795 }
1796
1797 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001798 * The amount of pressure on anon vs file pages is inversely
1799 * proportional to the fraction of recently scanned pages on
1800 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001801 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001802 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1803 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001804
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001805 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1806 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001807 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001808
Shaohua Li76a33fc2010-05-24 14:32:36 -07001809 fraction[0] = ap;
1810 fraction[1] = fp;
1811 denominator = ap + fp + 1;
1812out:
1813 for_each_evictable_lru(l) {
1814 int file = is_file_lru(l);
1815 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001816
Shaohua Li76a33fc2010-05-24 14:32:36 -07001817 scan = zone_nr_lru_pages(zone, sc, l);
1818 if (priority || noswap) {
1819 scan >>= priority;
1820 scan = div64_u64(scan * fraction[file], denominator);
1821 }
1822 nr[l] = nr_scan_try_batch(scan,
1823 &reclaim_stat->nr_saved_scan[l]);
1824 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001825}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001826
1827/*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001828 * Reclaim/compaction depends on a number of pages being freed. To avoid
1829 * disruption to the system, a small number of order-0 pages continue to be
1830 * rotated and reclaimed in the normal fashion. However, by the time we get
1831 * back to the allocator and call try_to_compact_zone(), we ensure that
1832 * there are enough free pages for it to be likely successful
1833 */
1834static inline bool should_continue_reclaim(struct zone *zone,
1835 unsigned long nr_reclaimed,
1836 unsigned long nr_scanned,
1837 struct scan_control *sc)
1838{
1839 unsigned long pages_for_compaction;
1840 unsigned long inactive_lru_pages;
1841
1842 /* If not in reclaim/compaction mode, stop */
Mel Gormanf3a310b2011-01-13 15:46:00 -08001843 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001844 return false;
1845
Mel Gorman28765922011-02-25 14:44:20 -08001846 /* Consider stopping depending on scan and reclaim activity */
1847 if (sc->gfp_mask & __GFP_REPEAT) {
1848 /*
1849 * For __GFP_REPEAT allocations, stop reclaiming if the
1850 * full LRU list has been scanned and we are still failing
1851 * to reclaim pages. This full LRU scan is potentially
1852 * expensive but a __GFP_REPEAT caller really wants to succeed
1853 */
1854 if (!nr_reclaimed && !nr_scanned)
1855 return false;
1856 } else {
1857 /*
1858 * For non-__GFP_REPEAT allocations which can presumably
1859 * fail without consequence, stop if we failed to reclaim
1860 * any pages from the last SWAP_CLUSTER_MAX number of
1861 * pages that were scanned. This will return to the
1862 * caller faster at the risk reclaim/compaction and
1863 * the resulting allocation attempt fails
1864 */
1865 if (!nr_reclaimed)
1866 return false;
1867 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001868
1869 /*
1870 * If we have not reclaimed enough pages for compaction and the
1871 * inactive lists are large enough, continue reclaiming
1872 */
1873 pages_for_compaction = (2UL << sc->order);
1874 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1875 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1876 if (sc->nr_reclaimed < pages_for_compaction &&
1877 inactive_lru_pages > pages_for_compaction)
1878 return true;
1879
1880 /* If compaction would go ahead or the allocation would succeed, stop */
1881 switch (compaction_suitable(zone, sc->order)) {
1882 case COMPACT_PARTIAL:
1883 case COMPACT_CONTINUE:
1884 return false;
1885 default:
1886 return true;
1887 }
1888}
1889
1890/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1892 */
Rik van Riela79311c2009-01-06 14:40:01 -08001893static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001894 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001896 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001897 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001898 enum lru_list l;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001899 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001900 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901
Mel Gorman3e7d3442011-01-13 15:45:56 -08001902restart:
1903 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001904 nr_scanned = sc->nr_scanned;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001905 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001906
Rik van Riel556adec2008-10-18 20:26:34 -07001907 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1908 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001909 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001910 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001911 nr_to_scan = min_t(unsigned long,
1912 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001913 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001915 nr_reclaimed += shrink_list(l, nr_to_scan,
1916 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001917 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 }
Rik van Riela79311c2009-01-06 14:40:01 -08001919 /*
1920 * On large memory systems, scan >> priority can become
1921 * really large. This is fine for the starting priority;
1922 * we want to put equal scanning pressure on each zone.
1923 * However, if the VM has a harder time of freeing pages,
1924 * with multiple processes reclaiming pages, the total
1925 * freeing target can get unreasonably large.
1926 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001927 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001928 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001930 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001931
Rik van Riel556adec2008-10-18 20:26:34 -07001932 /*
1933 * Even if we did not try to evict anon pages at all, we want to
1934 * rebalance the anon lru active/inactive ratio.
1935 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001936 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001937 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1938
Mel Gorman3e7d3442011-01-13 15:45:56 -08001939 /* reclaim/compaction might need reclaim to continue */
1940 if (should_continue_reclaim(zone, nr_reclaimed,
1941 sc->nr_scanned - nr_scanned, sc))
1942 goto restart;
1943
Andrew Morton232ea4d2007-02-28 20:13:21 -08001944 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945}
1946
1947/*
1948 * This is the direct reclaim path, for page-allocating processes. We only
1949 * try to reclaim pages from zones which will satisfy the caller's allocation
1950 * request.
1951 *
Mel Gorman41858962009-06-16 15:32:12 -07001952 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1953 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1955 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001956 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1957 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1958 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 * If a zone is deemed to be full of pinned pages then just give it a light
1961 * scan then give up on it.
1962 */
Minchan Kimd1908362010-09-22 13:05:01 -07001963static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001964 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965{
Mel Gormandd1a2392008-04-28 02:12:17 -07001966 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001967 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001968
Mel Gormand4debc62010-08-09 17:19:29 -07001969 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1970 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001971 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001973 /*
1974 * Take care memory controller reclaiming has small influence
1975 * to global LRU.
1976 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001977 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001978 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1979 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001980 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001981 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001982 }
Nick Piggin408d8542006-09-25 23:31:27 -07001983
Rik van Riela79311c2009-01-06 14:40:01 -08001984 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 }
Minchan Kimd1908362010-09-22 13:05:01 -07001986}
1987
1988static bool zone_reclaimable(struct zone *zone)
1989{
1990 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1991}
1992
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07001993/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07001994static bool all_unreclaimable(struct zonelist *zonelist,
1995 struct scan_control *sc)
1996{
1997 struct zoneref *z;
1998 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07001999
2000 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2001 gfp_zone(sc->gfp_mask), sc->nodemask) {
2002 if (!populated_zone(zone))
2003 continue;
2004 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2005 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002006 if (!zone->all_unreclaimable)
2007 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002008 }
2009
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002010 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002012
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013/*
2014 * This is the main entry point to direct page reclaim.
2015 *
2016 * If a full scan of the inactive list fails to free enough memory then we
2017 * are "out of memory" and something needs to be killed.
2018 *
2019 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2020 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002021 * caller can't do much about. We kick the writeback threads and take explicit
2022 * naps in the hope that some of these pages can be written. But if the
2023 * allocating task holds filesystem locks which prevent writeout this might not
2024 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002025 *
2026 * returns: 0, if no pages reclaimed
2027 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 */
Mel Gormandac1d272008-04-28 02:12:12 -07002029static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07002030 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031{
2032 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002033 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002035 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002036 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002037 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
Miao Xiec0ff7452010-05-24 14:32:08 -07002039 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002040 delayacct_freepages_start();
2041
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002042 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002043 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044
2045 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08002046 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002047 if (!priority)
2048 disable_swap_token();
Minchan Kimd1908362010-09-22 13:05:01 -07002049 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08002050 /*
2051 * Don't shrink slabs when reclaiming memory from
2052 * over limit cgroups
2053 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08002054 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002055 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002056 for_each_zone_zonelist(zone, z, zonelist,
2057 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002058 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2059 continue;
2060
2061 lru_pages += zone_reclaimable_pages(zone);
2062 }
2063
Mel Gormandd1a2392008-04-28 02:12:17 -07002064 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002065 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002066 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002067 reclaim_state->reclaimed_slab = 0;
2068 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 }
Balbir Singh66e17072008-02-07 00:13:56 -08002070 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002071 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073
2074 /*
2075 * Try to write back as many pages as we just scanned. This
2076 * tends to cause slow streaming writers to write data to the
2077 * disk smoothly, at the dirtying rate, which is nice. But
2078 * that's undesirable in laptop mode, where we *want* lumpy
2079 * writeout. So in laptop mode, write out the whole world.
2080 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002081 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2082 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002083 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002084 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 }
2086
2087 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002088 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002089 priority < DEF_PRIORITY - 2) {
2090 struct zone *preferred_zone;
2091
2092 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002093 &cpuset_current_mems_allowed,
2094 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002095 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2096 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002098
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002100 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002101 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002102
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002103 if (sc->nr_reclaimed)
2104 return sc->nr_reclaimed;
2105
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002106 /*
2107 * As hibernation is going on, kswapd is freezed so that it can't mark
2108 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2109 * check.
2110 */
2111 if (oom_killer_disabled)
2112 return 0;
2113
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002114 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002115 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002116 return 1;
2117
2118 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119}
2120
Mel Gormandac1d272008-04-28 02:12:12 -07002121unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002122 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002123{
Mel Gorman33906bc2010-08-09 17:19:16 -07002124 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002125 struct scan_control sc = {
2126 .gfp_mask = gfp_mask,
2127 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002128 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002129 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002130 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002131 .swappiness = vm_swappiness,
2132 .order = order,
2133 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002134 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002135 };
2136
Mel Gorman33906bc2010-08-09 17:19:16 -07002137 trace_mm_vmscan_direct_reclaim_begin(order,
2138 sc.may_writepage,
2139 gfp_mask);
2140
2141 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2142
2143 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2144
2145 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002146}
2147
Balbir Singh00f0b822008-03-04 14:28:39 -08002148#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002149
Balbir Singh4e416952009-09-23 15:56:39 -07002150unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2151 gfp_t gfp_mask, bool noswap,
2152 unsigned int swappiness,
KOSAKI Motohiro14fec792010-08-10 18:03:05 -07002153 struct zone *zone)
Balbir Singh4e416952009-09-23 15:56:39 -07002154{
2155 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002156 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002157 .may_writepage = !laptop_mode,
2158 .may_unmap = 1,
2159 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002160 .swappiness = swappiness,
2161 .order = 0,
2162 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002163 };
Balbir Singh4e416952009-09-23 15:56:39 -07002164 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2165 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002166
2167 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2168 sc.may_writepage,
2169 sc.gfp_mask);
2170
Balbir Singh4e416952009-09-23 15:56:39 -07002171 /*
2172 * NOTE: Although we can get the priority field, using it
2173 * here is not a good idea, since it limits the pages we can scan.
2174 * if we don't reclaim here, the shrink_zone from balance_pgdat
2175 * will pick up pages from other mem cgroup's as well. We hack
2176 * the priority and make it zero.
2177 */
2178 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002179
2180 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2181
Balbir Singh4e416952009-09-23 15:56:39 -07002182 return sc.nr_reclaimed;
2183}
2184
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002185unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002186 gfp_t gfp_mask,
2187 bool noswap,
2188 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002189{
Balbir Singh4e416952009-09-23 15:56:39 -07002190 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002191 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002192 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002193 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002194 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002195 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002196 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002197 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002198 .order = 0,
2199 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002200 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08002201 };
Balbir Singh66e17072008-02-07 00:13:56 -08002202
Mel Gormandd1a2392008-04-28 02:12:17 -07002203 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2204 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2205 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002206
2207 trace_mm_vmscan_memcg_reclaim_begin(0,
2208 sc.may_writepage,
2209 sc.gfp_mask);
2210
2211 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2212
2213 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2214
2215 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002216}
2217#endif
2218
Mel Gorman1741c872011-01-13 15:46:21 -08002219/*
2220 * pgdat_balanced is used when checking if a node is balanced for high-order
2221 * allocations. Only zones that meet watermarks and are in a zone allowed
2222 * by the callers classzone_idx are added to balanced_pages. The total of
2223 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2224 * for the node to be considered balanced. Forcing all zones to be balanced
2225 * for high orders can cause excessive reclaim when there are imbalanced zones.
2226 * The choice of 25% is due to
2227 * o a 16M DMA zone that is balanced will not balance a zone on any
2228 * reasonable sized machine
2229 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002230 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002231 * would need to be at least 256M for it to be balance a whole node.
2232 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2233 * to balance a node on its own. These seemed like reasonable ratios.
2234 */
2235static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2236 int classzone_idx)
2237{
2238 unsigned long present_pages = 0;
2239 int i;
2240
2241 for (i = 0; i <= classzone_idx; i++)
2242 present_pages += pgdat->node_zones[i].present_pages;
2243
2244 return balanced_pages > (present_pages >> 2);
2245}
2246
Mel Gormanf50de2d2009-12-14 17:58:53 -08002247/* is kswapd sleeping prematurely? */
Mel Gormandc83edd2011-01-13 15:46:26 -08002248static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2249 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002250{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002251 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002252 unsigned long balanced = 0;
2253 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002254
2255 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2256 if (remaining)
Mel Gormandc83edd2011-01-13 15:46:26 -08002257 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002258
Mel Gorman0abdee22011-01-13 15:46:22 -08002259 /* Check the watermark levels */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002260 for (i = 0; i < pgdat->nr_zones; i++) {
2261 struct zone *zone = pgdat->node_zones + i;
2262
2263 if (!populated_zone(zone))
2264 continue;
2265
Mel Gorman355b09c2011-01-13 15:46:24 -08002266 /*
2267 * balance_pgdat() skips over all_unreclaimable after
2268 * DEF_PRIORITY. Effectively, it considers them balanced so
2269 * they must be considered balanced here as well if kswapd
2270 * is to sleep
2271 */
2272 if (zone->all_unreclaimable) {
2273 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002274 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002275 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002276
Mel Gorman88f5acf2011-01-13 15:45:41 -08002277 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormandc83edd2011-01-13 15:46:26 -08002278 classzone_idx, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002279 all_zones_ok = false;
2280 else
2281 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002282 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002283
Mel Gorman1741c872011-01-13 15:46:21 -08002284 /*
2285 * For high-order requests, the balanced zones must contain at least
2286 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2287 * must be balanced
2288 */
2289 if (order)
Mel Gormandc83edd2011-01-13 15:46:26 -08002290 return pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002291 else
2292 return !all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002293}
2294
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295/*
2296 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002297 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002299 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 *
2301 * There is special handling here for zones which are full of pinned pages.
2302 * This can happen if the pages are all mlocked, or if they are all used by
2303 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2304 * What we do is to detect the case where all pages in the zone have been
2305 * scanned twice and there has been zero successful reclaim. Mark the zone as
2306 * dead and from now on, only perform a short scan. Basically we're polling
2307 * the zone for when the problem goes away.
2308 *
2309 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002310 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2311 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2312 * lower zones regardless of the number of free pages in the lower zones. This
2313 * interoperates with the page allocator fallback scheme to ensure that aging
2314 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 */
Mel Gorman99504742011-01-13 15:46:20 -08002316static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002317 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002320 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 int priority;
2322 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002323 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002324 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002326 struct scan_control sc = {
2327 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002328 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002329 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002330 /*
2331 * kswapd doesn't want to be bailed out while reclaim. because
2332 * we want to put equal scanning pressure on each zone.
2333 */
2334 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002335 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002336 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002337 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002338 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339loop_again:
2340 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002341 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002342 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002343 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002347 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002349 /* The swap token gets in the way of swapout... */
2350 if (!priority)
2351 disable_swap_token();
2352
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002354 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002356 /*
2357 * Scan in the highmem->dma direction for the highest
2358 * zone which needs scanning
2359 */
2360 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2361 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002363 if (!populated_zone(zone))
2364 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002366 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002367 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
Rik van Riel556adec2008-10-18 20:26:34 -07002369 /*
2370 * Do some background aging of the anon list, to give
2371 * pages a chance to be referenced before reclaiming.
2372 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002373 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002374 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2375 &sc, priority, 0);
2376
Mel Gorman88f5acf2011-01-13 15:45:41 -08002377 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002378 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002379 end_zone = i;
Mel Gormandc83edd2011-01-13 15:46:26 -08002380 *classzone_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002381 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002384 if (i < 0)
2385 goto out;
2386
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 for (i = 0; i <= end_zone; i++) {
2388 struct zone *zone = pgdat->node_zones + i;
2389
Wu Fengguangadea02a2009-09-21 17:01:42 -07002390 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 }
2392
2393 /*
2394 * Now scan the zone in the dma->highmem direction, stopping
2395 * at the last zone which needs scanning.
2396 *
2397 * We do this because the page allocator works in the opposite
2398 * direction. This prevents the page allocator from allocating
2399 * pages behind kswapd's direction of progress, which would
2400 * cause too much scanning of the lower zones.
2401 */
2402 for (i = 0; i <= end_zone; i++) {
2403 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002404 int nr_slab;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002405 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
Con Kolivasf3fe6512006-01-06 00:11:15 -08002407 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 continue;
2409
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002410 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 continue;
2412
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002414
Balbir Singh4e416952009-09-23 15:56:39 -07002415 /*
2416 * Call soft limit reclaim before calling shrink_zone.
2417 * For now we ignore the return value
2418 */
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002419 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2420
Rik van Riel32a43302007-10-16 01:24:50 -07002421 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002422 * We put equal pressure on every zone, unless
2423 * one zone has way too many pages free
2424 * already. The "too many pages" is defined
2425 * as the high wmark plus a "gap" where the
2426 * gap is either the low watermark or 1%
2427 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002428 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002429 balance_gap = min(low_wmark_pages(zone),
2430 (zone->present_pages +
2431 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2432 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Mel Gorman88f5acf2011-01-13 15:45:41 -08002433 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002434 high_wmark_pages(zone) + balance_gap,
2435 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002436 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002438 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2439 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002440 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 total_scanned += sc.nr_scanned;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002442
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002443 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 continue;
Andrea Arcangelid527caf2011-03-22 16:30:38 -07002445 if (nr_slab == 0 &&
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002446 !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002447 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 /*
2449 * If we've done a decent amount of scanning and
2450 * the reclaim ratio is low, start doing writepage
2451 * even in laptop mode
2452 */
2453 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002454 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002456
Mel Gorman88f5acf2011-01-13 15:45:41 -08002457 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002458 high_wmark_pages(zone), end_zone, 0)) {
2459 all_zones_ok = 0;
2460 /*
2461 * We are still under min water mark. This
2462 * means that we have a GFP_ATOMIC allocation
2463 * failure risk. Hurry up!
2464 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002465 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002466 min_wmark_pages(zone), end_zone, 0))
2467 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002468 } else {
2469 /*
2470 * If a zone reaches its high watermark,
2471 * consider it to be no longer congested. It's
2472 * possible there are dirty pages backed by
2473 * congested BDIs but as pressure is relieved,
2474 * spectulatively avoid congestion waits
2475 */
2476 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002477 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002478 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002479 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002480
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 }
Mel Gormandc83edd2011-01-13 15:46:26 -08002482 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 break; /* kswapd: all done */
2484 /*
2485 * OK, kswapd is getting into trouble. Take a nap, then take
2486 * another pass across the zones.
2487 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002488 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2489 if (has_under_min_watermark_zone)
2490 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2491 else
2492 congestion_wait(BLK_RW_ASYNC, HZ/10);
2493 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494
2495 /*
2496 * We do this so kswapd doesn't build up large priorities for
2497 * example when it is freeing in parallel with allocators. It
2498 * matches the direct reclaim path behaviour in terms of impact
2499 * on zone->*_priority.
2500 */
Rik van Riela79311c2009-01-06 14:40:01 -08002501 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 break;
2503 }
2504out:
Mel Gorman99504742011-01-13 15:46:20 -08002505
2506 /*
2507 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002508 * high-order: Balanced zones must make up at least 25% of the node
2509 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002510 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002511 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002513
2514 try_to_freeze();
2515
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002516 /*
2517 * Fragmentation may mean that the system cannot be
2518 * rebalanced for high-order allocations in all zones.
2519 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2520 * it means the zones have been fully scanned and are still
2521 * not balanced. For high-order allocations, there is
2522 * little point trying all over again as kswapd may
2523 * infinite loop.
2524 *
2525 * Instead, recheck all watermarks at order-0 as they
2526 * are the most important. If watermarks are ok, kswapd will go
2527 * back to sleep. High-order users can still perform direct
2528 * reclaim if they wish.
2529 */
2530 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2531 order = sc.order = 0;
2532
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 goto loop_again;
2534 }
2535
Mel Gorman99504742011-01-13 15:46:20 -08002536 /*
2537 * If kswapd was reclaiming at a higher order, it has the option of
2538 * sleeping without all zones being balanced. Before it does, it must
2539 * ensure that the watermarks for order-0 on *all* zones are met and
2540 * that the congestion flags are cleared. The congestion flag must
2541 * be cleared as kswapd is the only mechanism that clears the flag
2542 * and it is potentially going to sleep here.
2543 */
2544 if (order) {
2545 for (i = 0; i <= end_zone; i++) {
2546 struct zone *zone = pgdat->node_zones + i;
2547
2548 if (!populated_zone(zone))
2549 continue;
2550
2551 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2552 continue;
2553
2554 /* Confirm the zone is balanced for order-0 */
2555 if (!zone_watermark_ok(zone, 0,
2556 high_wmark_pages(zone), 0, 0)) {
2557 order = sc.order = 0;
2558 goto loop_again;
2559 }
2560
2561 /* If balanced, clear the congested flag */
2562 zone_clear_flag(zone, ZONE_CONGESTED);
2563 }
2564 }
2565
Mel Gorman0abdee22011-01-13 15:46:22 -08002566 /*
2567 * Return the order we were reclaiming at so sleeping_prematurely()
2568 * makes a decision on the order we were last reclaiming at. However,
2569 * if another caller entered the allocator slow path while kswapd
2570 * was awake, order will remain at the higher level
2571 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002572 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002573 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574}
2575
Mel Gormandc83edd2011-01-13 15:46:26 -08002576static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002577{
2578 long remaining = 0;
2579 DEFINE_WAIT(wait);
2580
2581 if (freezing(current) || kthread_should_stop())
2582 return;
2583
2584 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2585
2586 /* Try to sleep for a short interval */
Mel Gormandc83edd2011-01-13 15:46:26 -08002587 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002588 remaining = schedule_timeout(HZ/10);
2589 finish_wait(&pgdat->kswapd_wait, &wait);
2590 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2591 }
2592
2593 /*
2594 * After a short sleep, check if it was a premature sleep. If not, then
2595 * go fully to sleep until explicitly woken up.
2596 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002597 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002598 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2599
2600 /*
2601 * vmstat counters are not perfectly accurate and the estimated
2602 * value for counters such as NR_FREE_PAGES can deviate from the
2603 * true value by nr_online_cpus * threshold. To avoid the zone
2604 * watermarks being breached while under pressure, we reduce the
2605 * per-cpu vmstat threshold while kswapd is awake and restore
2606 * them before going back to sleep.
2607 */
2608 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2609 schedule();
2610 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2611 } else {
2612 if (remaining)
2613 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2614 else
2615 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2616 }
2617 finish_wait(&pgdat->kswapd_wait, &wait);
2618}
2619
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620/*
2621 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002622 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 *
2624 * This basically trickles out pages so that we have _some_
2625 * free memory available even if there is no other activity
2626 * that frees anything up. This is needed for things like routing
2627 * etc, where we otherwise might have all activity going on in
2628 * asynchronous contexts that cannot page things out.
2629 *
2630 * If there are applications that are active memory-allocators
2631 * (most normal use), this basically shouldn't matter.
2632 */
2633static int kswapd(void *p)
2634{
2635 unsigned long order;
Mel Gorman99504742011-01-13 15:46:20 -08002636 int classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 pg_data_t *pgdat = (pg_data_t*)p;
2638 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002639
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 struct reclaim_state reclaim_state = {
2641 .reclaimed_slab = 0,
2642 };
Rusty Russella70f7302009-03-13 14:49:46 +10302643 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644
Nick Piggincf40bd12009-01-21 08:12:39 +01002645 lockdep_set_current_reclaim_state(GFP_KERNEL);
2646
Rusty Russell174596a2009-01-01 10:12:29 +10302647 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002648 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 current->reclaim_state = &reclaim_state;
2650
2651 /*
2652 * Tell the memory management that we're a "memory allocator",
2653 * and that if we need more memory we should get access to it
2654 * regardless (see "__alloc_pages()"). "kswapd" should
2655 * never get caught in the normal page freeing logic.
2656 *
2657 * (Kswapd normally doesn't need memory anyway, but sometimes
2658 * you need a small amount of memory in order to be able to
2659 * page out something else, and this flag essentially protects
2660 * us from recursively trying to free more memory as we're
2661 * trying to free the first piece of memory in the first place).
2662 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002663 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002664 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665
2666 order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002667 classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 for ( ; ; ) {
2669 unsigned long new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002670 int new_classzone_idx;
David Rientjes8fe23e02009-12-14 17:58:33 -08002671 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002672
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 new_order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002674 new_classzone_idx = pgdat->classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 pgdat->kswapd_max_order = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002676 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2677 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 /*
2679 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002680 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 */
2682 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002683 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 } else {
Mel Gormandc83edd2011-01-13 15:46:26 -08002685 kswapd_try_to_sleep(pgdat, order, classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002687 classzone_idx = pgdat->classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002688 pgdat->kswapd_max_order = 0;
2689 pgdat->classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691
David Rientjes8fe23e02009-12-14 17:58:33 -08002692 ret = try_to_freeze();
2693 if (kthread_should_stop())
2694 break;
2695
2696 /*
2697 * We can speed up thawing tasks if we don't call balance_pgdat
2698 * after returning from the refrigerator
2699 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002700 if (!ret) {
2701 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Mel Gormandc83edd2011-01-13 15:46:26 -08002702 order = balance_pgdat(pgdat, order, &classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002703 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 }
2705 return 0;
2706}
2707
2708/*
2709 * A zone is low on free memory, so wake its kswapd task to service it.
2710 */
Mel Gorman99504742011-01-13 15:46:20 -08002711void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712{
2713 pg_data_t *pgdat;
2714
Con Kolivasf3fe6512006-01-06 00:11:15 -08002715 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 return;
2717
Paul Jackson02a0e532006-12-13 00:34:25 -08002718 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002720 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002721 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002722 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002723 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2724 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002725 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002727 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2728 return;
2729
2730 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002731 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733
Wu Fengguangadea02a2009-09-21 17:01:42 -07002734/*
2735 * The reclaimable count would be mostly accurate.
2736 * The less reclaimable pages may be
2737 * - mlocked pages, which will be moved to unevictable list when encountered
2738 * - mapped pages, which may require several travels to be reclaimed
2739 * - dirty pages, which is not "instantly" reclaimable
2740 */
2741unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002742{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002743 int nr;
2744
2745 nr = global_page_state(NR_ACTIVE_FILE) +
2746 global_page_state(NR_INACTIVE_FILE);
2747
2748 if (nr_swap_pages > 0)
2749 nr += global_page_state(NR_ACTIVE_ANON) +
2750 global_page_state(NR_INACTIVE_ANON);
2751
2752 return nr;
2753}
2754
2755unsigned long zone_reclaimable_pages(struct zone *zone)
2756{
2757 int nr;
2758
2759 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2760 zone_page_state(zone, NR_INACTIVE_FILE);
2761
2762 if (nr_swap_pages > 0)
2763 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2764 zone_page_state(zone, NR_INACTIVE_ANON);
2765
2766 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002767}
2768
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002769#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002771 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002772 * freed pages.
2773 *
2774 * Rather than trying to age LRUs the aim is to preserve the overall
2775 * LRU order by reclaiming preferentially
2776 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002778unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002780 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002781 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002782 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2783 .may_swap = 1,
2784 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002785 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002786 .nr_to_reclaim = nr_to_reclaim,
2787 .hibernation_mode = 1,
2788 .swappiness = vm_swappiness,
2789 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002791 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2792 struct task_struct *p = current;
2793 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002795 p->flags |= PF_MEMALLOC;
2796 lockdep_set_current_reclaim_state(sc.gfp_mask);
2797 reclaim_state.reclaimed_slab = 0;
2798 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002799
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002800 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002801
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002802 p->reclaim_state = NULL;
2803 lockdep_clear_current_reclaim_state();
2804 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002805
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002806 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002808#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810/* It's optimal to keep kswapds on the same CPUs as their memory, but
2811 not required for correctness. So if the last cpu in a node goes
2812 away, we get changed to run anywhere: as the first one comes back,
2813 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002814static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002815 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002817 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002819 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002820 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002821 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302822 const struct cpumask *mask;
2823
2824 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002825
Rusty Russell3e597942009-01-01 10:12:24 +10302826 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002828 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 }
2830 }
2831 return NOTIFY_OK;
2832}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833
Yasunori Goto3218ae12006-06-27 02:53:33 -07002834/*
2835 * This kswapd start function will be called by init and node-hot-add.
2836 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2837 */
2838int kswapd_run(int nid)
2839{
2840 pg_data_t *pgdat = NODE_DATA(nid);
2841 int ret = 0;
2842
2843 if (pgdat->kswapd)
2844 return 0;
2845
2846 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2847 if (IS_ERR(pgdat->kswapd)) {
2848 /* failure at boot is fatal */
2849 BUG_ON(system_state == SYSTEM_BOOTING);
2850 printk("Failed to start kswapd on node %d\n",nid);
2851 ret = -1;
2852 }
2853 return ret;
2854}
2855
David Rientjes8fe23e02009-12-14 17:58:33 -08002856/*
2857 * Called by memory hotplug when all memory in a node is offlined.
2858 */
2859void kswapd_stop(int nid)
2860{
2861 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2862
2863 if (kswapd)
2864 kthread_stop(kswapd);
2865}
2866
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867static int __init kswapd_init(void)
2868{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002869 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002870
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002872 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002873 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 hotcpu_notifier(cpu_callback, 0);
2875 return 0;
2876}
2877
2878module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002879
2880#ifdef CONFIG_NUMA
2881/*
2882 * Zone reclaim mode
2883 *
2884 * If non-zero call zone_reclaim when the number of free pages falls below
2885 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002886 */
2887int zone_reclaim_mode __read_mostly;
2888
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002889#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002890#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002891#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2892#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2893
Christoph Lameter9eeff232006-01-18 17:42:31 -08002894/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002895 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2896 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2897 * a zone.
2898 */
2899#define ZONE_RECLAIM_PRIORITY 4
2900
Christoph Lameter9eeff232006-01-18 17:42:31 -08002901/*
Christoph Lameter96146342006-07-03 00:24:13 -07002902 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2903 * occur.
2904 */
2905int sysctl_min_unmapped_ratio = 1;
2906
2907/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002908 * If the number of slab pages in a zone grows beyond this percentage then
2909 * slab reclaim needs to occur.
2910 */
2911int sysctl_min_slab_ratio = 5;
2912
Mel Gorman90afa5d2009-06-16 15:33:20 -07002913static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2914{
2915 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2916 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2917 zone_page_state(zone, NR_ACTIVE_FILE);
2918
2919 /*
2920 * It's possible for there to be more file mapped pages than
2921 * accounted for by the pages on the file LRU lists because
2922 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2923 */
2924 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2925}
2926
2927/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2928static long zone_pagecache_reclaimable(struct zone *zone)
2929{
2930 long nr_pagecache_reclaimable;
2931 long delta = 0;
2932
2933 /*
2934 * If RECLAIM_SWAP is set, then all file pages are considered
2935 * potentially reclaimable. Otherwise, we have to worry about
2936 * pages like swapcache and zone_unmapped_file_pages() provides
2937 * a better estimate
2938 */
2939 if (zone_reclaim_mode & RECLAIM_SWAP)
2940 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2941 else
2942 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2943
2944 /* If we can't clean pages, remove dirty pages from consideration */
2945 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2946 delta += zone_page_state(zone, NR_FILE_DIRTY);
2947
2948 /* Watch for any possible underflows due to delta */
2949 if (unlikely(delta > nr_pagecache_reclaimable))
2950 delta = nr_pagecache_reclaimable;
2951
2952 return nr_pagecache_reclaimable - delta;
2953}
2954
Christoph Lameter0ff38492006-09-25 23:31:52 -07002955/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002956 * Try to free up some pages from this zone through reclaim.
2957 */
Andrew Morton179e9632006-03-22 00:08:18 -08002958static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002959{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002960 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002961 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002962 struct task_struct *p = current;
2963 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002964 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002965 struct scan_control sc = {
2966 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002967 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002968 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002969 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2970 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002971 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002972 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002973 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002974 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002975 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002976
Christoph Lameter9eeff232006-01-18 17:42:31 -08002977 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002978 /*
2979 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2980 * and we also need to be able to write out pages for RECLAIM_WRITE
2981 * and RECLAIM_SWAP.
2982 */
2983 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002984 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002985 reclaim_state.reclaimed_slab = 0;
2986 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002987
Mel Gorman90afa5d2009-06-16 15:33:20 -07002988 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002989 /*
2990 * Free memory by calling shrink zone with increasing
2991 * priorities until we have enough memory freed.
2992 */
2993 priority = ZONE_RECLAIM_PRIORITY;
2994 do {
Rik van Riela79311c2009-01-06 14:40:01 -08002995 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002996 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002997 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002998 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002999
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003000 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3001 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003002 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003003 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003004 * many pages were freed in this zone. So we take the current
3005 * number of slab pages and shake the slab until it is reduced
3006 * by the same nr_pages that we used for reclaiming unmapped
3007 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003008 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003009 * Note that shrink_slab will free memory on all zones and may
3010 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003011 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003012 for (;;) {
3013 unsigned long lru_pages = zone_reclaimable_pages(zone);
3014
3015 /* No reclaimable slab or very low memory pressure */
3016 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3017 break;
3018
3019 /* Freed enough memory */
3020 nr_slab_pages1 = zone_page_state(zone,
3021 NR_SLAB_RECLAIMABLE);
3022 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3023 break;
3024 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003025
3026 /*
3027 * Update nr_reclaimed by the number of slab pages we
3028 * reclaimed from this zone.
3029 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003030 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3031 if (nr_slab_pages1 < nr_slab_pages0)
3032 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003033 }
3034
Christoph Lameter9eeff232006-01-18 17:42:31 -08003035 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003036 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003037 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003038 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003039}
Andrew Morton179e9632006-03-22 00:08:18 -08003040
3041int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3042{
Andrew Morton179e9632006-03-22 00:08:18 -08003043 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003044 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003045
3046 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003047 * Zone reclaim reclaims unmapped file backed pages and
3048 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003049 *
Christoph Lameter96146342006-07-03 00:24:13 -07003050 * A small portion of unmapped file backed pages is needed for
3051 * file I/O otherwise pages read by file I/O will be immediately
3052 * thrown out if the zone is overallocated. So we do not reclaim
3053 * if less than a specified percentage of the zone is used by
3054 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003055 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003056 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3057 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003058 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003059
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003060 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003061 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003062
Andrew Morton179e9632006-03-22 00:08:18 -08003063 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003064 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003065 */
David Rientjesd773ed62007-10-16 23:26:01 -07003066 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003067 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003068
3069 /*
3070 * Only run zone reclaim on the local zone or on zones that do not
3071 * have associated processors. This will favor the local processor
3072 * over remote processors and spread off node memory allocations
3073 * as wide as possible.
3074 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003075 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003076 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003077 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003078
3079 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003080 return ZONE_RECLAIM_NOSCAN;
3081
David Rientjesd773ed62007-10-16 23:26:01 -07003082 ret = __zone_reclaim(zone, gfp_mask, order);
3083 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3084
Mel Gorman24cf725182009-06-16 15:33:23 -07003085 if (!ret)
3086 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3087
David Rientjesd773ed62007-10-16 23:26:01 -07003088 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003089}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003090#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003091
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003092/*
3093 * page_evictable - test whether a page is evictable
3094 * @page: the page to test
3095 * @vma: the VMA in which the page is or will be mapped, may be NULL
3096 *
3097 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003098 * lists vs unevictable list. The vma argument is !NULL when called from the
3099 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003100 *
3101 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003102 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003103 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003104 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003105 */
3106int page_evictable(struct page *page, struct vm_area_struct *vma)
3107{
3108
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003109 if (mapping_unevictable(page_mapping(page)))
3110 return 0;
3111
Nick Pigginb291f002008-10-18 20:26:44 -07003112 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3113 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003114
3115 return 1;
3116}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003117
3118/**
3119 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3120 * @page: page to check evictability and move to appropriate lru list
3121 * @zone: zone page is in
3122 *
3123 * Checks a page for evictability and moves the page to the appropriate
3124 * zone lru list.
3125 *
3126 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3127 * have PageUnevictable set.
3128 */
3129static void check_move_unevictable_page(struct page *page, struct zone *zone)
3130{
3131 VM_BUG_ON(PageActive(page));
3132
3133retry:
3134 ClearPageUnevictable(page);
3135 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07003136 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003137
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003138 __dec_zone_state(zone, NR_UNEVICTABLE);
3139 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003140 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003141 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3142 __count_vm_event(UNEVICTABLE_PGRESCUED);
3143 } else {
3144 /*
3145 * rotate unevictable list
3146 */
3147 SetPageUnevictable(page);
3148 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08003149 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003150 if (page_evictable(page, NULL))
3151 goto retry;
3152 }
3153}
3154
3155/**
3156 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3157 * @mapping: struct address_space to scan for evictable pages
3158 *
3159 * Scan all pages in mapping. Check unevictable pages for
3160 * evictability and move them to the appropriate zone lru list.
3161 */
3162void scan_mapping_unevictable_pages(struct address_space *mapping)
3163{
3164 pgoff_t next = 0;
3165 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3166 PAGE_CACHE_SHIFT;
3167 struct zone *zone;
3168 struct pagevec pvec;
3169
3170 if (mapping->nrpages == 0)
3171 return;
3172
3173 pagevec_init(&pvec, 0);
3174 while (next < end &&
3175 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3176 int i;
3177 int pg_scanned = 0;
3178
3179 zone = NULL;
3180
3181 for (i = 0; i < pagevec_count(&pvec); i++) {
3182 struct page *page = pvec.pages[i];
3183 pgoff_t page_index = page->index;
3184 struct zone *pagezone = page_zone(page);
3185
3186 pg_scanned++;
3187 if (page_index > next)
3188 next = page_index;
3189 next++;
3190
3191 if (pagezone != zone) {
3192 if (zone)
3193 spin_unlock_irq(&zone->lru_lock);
3194 zone = pagezone;
3195 spin_lock_irq(&zone->lru_lock);
3196 }
3197
3198 if (PageLRU(page) && PageUnevictable(page))
3199 check_move_unevictable_page(page, zone);
3200 }
3201 if (zone)
3202 spin_unlock_irq(&zone->lru_lock);
3203 pagevec_release(&pvec);
3204
3205 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3206 }
3207
3208}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003209
3210/**
3211 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3212 * @zone - zone of which to scan the unevictable list
3213 *
3214 * Scan @zone's unevictable LRU lists to check for pages that have become
3215 * evictable. Move those that have to @zone's inactive list where they
3216 * become candidates for reclaim, unless shrink_inactive_zone() decides
3217 * to reactivate them. Pages that are still unevictable are rotated
3218 * back onto @zone's unevictable list.
3219 */
3220#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08003221static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003222{
3223 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3224 unsigned long scan;
3225 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3226
3227 while (nr_to_scan > 0) {
3228 unsigned long batch_size = min(nr_to_scan,
3229 SCAN_UNEVICTABLE_BATCH_SIZE);
3230
3231 spin_lock_irq(&zone->lru_lock);
3232 for (scan = 0; scan < batch_size; scan++) {
3233 struct page *page = lru_to_page(l_unevictable);
3234
3235 if (!trylock_page(page))
3236 continue;
3237
3238 prefetchw_prev_lru_page(page, l_unevictable, flags);
3239
3240 if (likely(PageLRU(page) && PageUnevictable(page)))
3241 check_move_unevictable_page(page, zone);
3242
3243 unlock_page(page);
3244 }
3245 spin_unlock_irq(&zone->lru_lock);
3246
3247 nr_to_scan -= batch_size;
3248 }
3249}
3250
3251
3252/**
3253 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3254 *
3255 * A really big hammer: scan all zones' unevictable LRU lists to check for
3256 * pages that have become evictable. Move those back to the zones'
3257 * inactive list where they become candidates for reclaim.
3258 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3259 * and we add swap to the system. As such, it runs in the context of a task
3260 * that has possibly/probably made some previously unevictable pages
3261 * evictable.
3262 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003263static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003264{
3265 struct zone *zone;
3266
3267 for_each_zone(zone) {
3268 scan_zone_unevictable_pages(zone);
3269 }
3270}
3271
3272/*
3273 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3274 * all nodes' unevictable lists for evictable pages
3275 */
3276unsigned long scan_unevictable_pages;
3277
3278int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003279 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003280 size_t *length, loff_t *ppos)
3281{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003282 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003283
3284 if (write && *(unsigned long *)table->data)
3285 scan_all_zones_unevictable_pages();
3286
3287 scan_unevictable_pages = 0;
3288 return 0;
3289}
3290
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003291#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003292/*
3293 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3294 * a specified node's per zone unevictable lists for evictable pages.
3295 */
3296
3297static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3298 struct sysdev_attribute *attr,
3299 char *buf)
3300{
3301 return sprintf(buf, "0\n"); /* always zero; should fit... */
3302}
3303
3304static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3305 struct sysdev_attribute *attr,
3306 const char *buf, size_t count)
3307{
3308 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3309 struct zone *zone;
3310 unsigned long res;
3311 unsigned long req = strict_strtoul(buf, 10, &res);
3312
3313 if (!req)
3314 return 1; /* zero is no-op */
3315
3316 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3317 if (!populated_zone(zone))
3318 continue;
3319 scan_zone_unevictable_pages(zone);
3320 }
3321 return 1;
3322}
3323
3324
3325static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3326 read_scan_unevictable_node,
3327 write_scan_unevictable_node);
3328
3329int scan_unevictable_register_node(struct node *node)
3330{
3331 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3332}
3333
3334void scan_unevictable_unregister_node(struct node *node)
3335{
3336 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3337}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003338#endif