blob: 99b434b674c02b4099bcfe58632ad56e183801fd [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/backing-dev.h>
30#include <linux/rmap.h>
31#include <linux/topology.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080034#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070043#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070044#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
46#include <asm/tlbflush.h>
47#include <asm/div64.h>
48
49#include <linux/swapops.h>
50
Nick Piggin0f8053a2006-03-22 00:08:33 -080051#include "internal.h"
52
Mel Gorman33906bc2010-08-09 17:19:16 -070053#define CREATE_TRACE_POINTS
54#include <trace/events/vmscan.h>
55
Linus Torvalds1da177e2005-04-16 15:20:36 -070056struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070057 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned;
59
Rik van Riela79311c2009-01-06 14:40:01 -080060 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed;
62
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080063 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080066 unsigned long hibernation_mode;
67
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040069 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
71 int may_writepage;
72
Johannes Weinera6dc60f82009-03-31 15:19:30 -070073 /* Can mapped pages be reclaimed? */
74 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080075
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070076 /* Can pages be swapped as part of reclaim? */
77 int may_swap;
78
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070079 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080080
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070081 /* Scan (total_size >> priority) pages at once */
82 int priority;
83
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070084 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080085 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
87 */
88 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080089
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070090 /*
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
92 * are scanned.
93 */
94 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070095};
96
Linus Torvalds1da177e2005-04-16 15:20:36 -070097#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99#ifdef ARCH_HAS_PREFETCH
100#define prefetch_prev_lru_page(_page, _base, _field) \
101 do { \
102 if ((_page)->lru.prev != _base) { \
103 struct page *prev; \
104 \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
107 } \
108 } while (0)
109#else
110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111#endif
112
113#ifdef ARCH_HAS_PREFETCHW
114#define prefetchw_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127/*
128 * From 0 .. 100. Higher means more swappy.
129 */
130int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700131long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
133static LIST_HEAD(shrinker_list);
134static DECLARE_RWSEM(shrinker_rwsem);
135
Andrew Mortonc255a452012-07-31 16:43:02 -0700136#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800137static bool global_reclaim(struct scan_control *sc)
138{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800139 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800140}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800141#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800142static bool global_reclaim(struct scan_control *sc)
143{
144 return true;
145}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800146#endif
147
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700148static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800149{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700150 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700151 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800152
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700153 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800154}
155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156/*
157 * Add a shrinker callback to be called from the vm
158 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700159void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160{
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800161 atomic_long_set(&shrinker->nr_in_batch, 0);
Rusty Russell8e1f9362007-07-17 04:03:17 -0700162 down_write(&shrinker_rwsem);
163 list_add_tail(&shrinker->list, &shrinker_list);
164 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700166EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167
168/*
169 * Remove one
170 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700171void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172{
173 down_write(&shrinker_rwsem);
174 list_del(&shrinker->list);
175 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700177EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
Ying Han1495f232011-05-24 17:12:27 -0700179static inline int do_shrinker_shrink(struct shrinker *shrinker,
180 struct shrink_control *sc,
181 unsigned long nr_to_scan)
182{
183 sc->nr_to_scan = nr_to_scan;
184 return (*shrinker->shrink)(shrinker, sc);
185}
186
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#define SHRINK_BATCH 128
188/*
189 * Call the shrink functions to age shrinkable caches
190 *
191 * Here we assume it costs one seek to replace a lru page and that it also
192 * takes a seek to recreate a cache object. With this in mind we age equal
193 * percentages of the lru and ageable caches. This should balance the seeks
194 * generated by these structures.
195 *
Simon Arlott183ff222007-10-20 01:27:18 +0200196 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 * slab to avoid swapping.
198 *
199 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
200 *
201 * `lru_pages' represents the number of on-LRU pages in all the zones which
202 * are eligible for the caller's allocation attempt. It is used for balancing
203 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700204 *
205 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700207unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700208 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700209 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
211 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800212 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Ying Han1495f232011-05-24 17:12:27 -0700214 if (nr_pages_scanned == 0)
215 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
Minchan Kimf06590b2011-05-24 17:11:11 -0700217 if (!down_read_trylock(&shrinker_rwsem)) {
218 /* Assume we'll be able to shrink next time */
219 ret = 1;
220 goto out;
221 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
223 list_for_each_entry(shrinker, &shrinker_list, list) {
224 unsigned long long delta;
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800225 long total_scan;
226 long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000227 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000228 long nr;
229 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000230 long batch_size = shrinker->batch ? shrinker->batch
231 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800233 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
234 if (max_pass <= 0)
235 continue;
236
Dave Chinneracf92b42011-07-08 14:14:35 +1000237 /*
238 * copy the current shrinker scan count into a local variable
239 * and zero it so that other concurrent shrinker invocations
240 * don't also do this scanning work.
241 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800242 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
Dave Chinneracf92b42011-07-08 14:14:35 +1000243
244 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700245 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800246 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000248 total_scan += delta;
249 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700250 printk(KERN_ERR "shrink_slab: %pF negative objects to "
251 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000252 shrinker->shrink, total_scan);
253 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800254 }
255
256 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000257 * We need to avoid excessive windup on filesystem shrinkers
258 * due to large numbers of GFP_NOFS allocations causing the
259 * shrinkers to return -1 all the time. This results in a large
260 * nr being built up so when a shrink that can do some work
261 * comes along it empties the entire cache due to nr >>>
262 * max_pass. This is bad for sustaining a working set in
263 * memory.
264 *
265 * Hence only allow the shrinker to scan the entire cache when
266 * a large delta change is calculated directly.
267 */
268 if (delta < max_pass / 4)
269 total_scan = min(total_scan, max_pass / 2);
270
271 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800272 * Avoid risking looping forever due to too large nr value:
273 * never try to free more than twice the estimate number of
274 * freeable entries.
275 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000276 if (total_scan > max_pass * 2)
277 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
Dave Chinneracf92b42011-07-08 14:14:35 +1000279 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000280 nr_pages_scanned, lru_pages,
281 max_pass, delta, total_scan);
282
Dave Chinnere9299f52011-07-08 14:14:37 +1000283 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700284 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
Ying Han1495f232011-05-24 17:12:27 -0700286 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000288 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 if (shrink_ret == -1)
290 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700291 if (shrink_ret < nr_before)
292 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000293 count_vm_events(SLABS_SCANNED, batch_size);
294 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295
296 cond_resched();
297 }
298
Dave Chinneracf92b42011-07-08 14:14:35 +1000299 /*
300 * move the unused scan count back into the shrinker in a
301 * manner that handles concurrent updates. If we exhausted the
302 * scan, there is no need to do an update.
303 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800304 if (total_scan > 0)
305 new_nr = atomic_long_add_return(total_scan,
306 &shrinker->nr_in_batch);
307 else
308 new_nr = atomic_long_read(&shrinker->nr_in_batch);
Dave Chinneracf92b42011-07-08 14:14:35 +1000309
310 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 }
312 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700313out:
314 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700315 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316}
317
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318static inline int is_page_cache_freeable(struct page *page)
319{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700320 /*
321 * A freeable page cache page is referenced only by the caller
322 * that isolated the page, the page cache radix tree and
323 * optional buffer heads at page->private.
324 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700325 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326}
327
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700328static int may_write_to_queue(struct backing_dev_info *bdi,
329 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330{
Christoph Lameter930d9152006-01-08 01:00:47 -0800331 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 return 1;
333 if (!bdi_write_congested(bdi))
334 return 1;
335 if (bdi == current->backing_dev_info)
336 return 1;
337 return 0;
338}
339
340/*
341 * We detected a synchronous write error writing a page out. Probably
342 * -ENOSPC. We need to propagate that into the address_space for a subsequent
343 * fsync(), msync() or close().
344 *
345 * The tricky part is that after writepage we cannot touch the mapping: nothing
346 * prevents it from being freed up. But we have a ref on the page and once
347 * that page is locked, the mapping is pinned.
348 *
349 * We're allowed to run sleeping lock_page() here because we know the caller has
350 * __GFP_FS.
351 */
352static void handle_write_error(struct address_space *mapping,
353 struct page *page, int error)
354{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100355 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700356 if (page_mapping(page) == mapping)
357 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 unlock_page(page);
359}
360
Christoph Lameter04e62a22006-06-23 02:03:38 -0700361/* possible outcome of pageout() */
362typedef enum {
363 /* failed to write page out, page is locked */
364 PAGE_KEEP,
365 /* move page to the active list, page is locked */
366 PAGE_ACTIVATE,
367 /* page has been sent to the disk successfully, page is unlocked */
368 PAGE_SUCCESS,
369 /* page is clean and locked */
370 PAGE_CLEAN,
371} pageout_t;
372
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373/*
Andrew Morton1742f192006-03-22 00:08:21 -0800374 * pageout is called by shrink_page_list() for each dirty page.
375 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700377static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700378 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379{
380 /*
381 * If the page is dirty, only perform writeback if that write
382 * will be non-blocking. To prevent this allocation from being
383 * stalled by pagecache activity. But note that there may be
384 * stalls if we need to run get_block(). We could test
385 * PagePrivate for that.
386 *
Vincent Li6aceb532009-12-14 17:58:49 -0800387 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 * this page's queue, we can perform writeback even if that
389 * will block.
390 *
391 * If the page is swapcache, write it back even if that would
392 * block, for some throttling. This happens by accident, because
393 * swap_backing_dev_info is bust: it doesn't reflect the
394 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 */
396 if (!is_page_cache_freeable(page))
397 return PAGE_KEEP;
398 if (!mapping) {
399 /*
400 * Some data journaling orphaned pages can have
401 * page->mapping == NULL while being dirty with clean buffers.
402 */
David Howells266cf652009-04-03 16:42:36 +0100403 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 if (try_to_free_buffers(page)) {
405 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700406 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 return PAGE_CLEAN;
408 }
409 }
410 return PAGE_KEEP;
411 }
412 if (mapping->a_ops->writepage == NULL)
413 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700414 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 return PAGE_KEEP;
416
417 if (clear_page_dirty_for_io(page)) {
418 int res;
419 struct writeback_control wbc = {
420 .sync_mode = WB_SYNC_NONE,
421 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700422 .range_start = 0,
423 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 .for_reclaim = 1,
425 };
426
427 SetPageReclaim(page);
428 res = mapping->a_ops->writepage(page, &wbc);
429 if (res < 0)
430 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800431 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 ClearPageReclaim(page);
433 return PAGE_ACTIVATE;
434 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700435
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 if (!PageWriteback(page)) {
437 /* synchronous write or broken a_ops? */
438 ClearPageReclaim(page);
439 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700440 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700441 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 return PAGE_SUCCESS;
443 }
444
445 return PAGE_CLEAN;
446}
447
Andrew Mortona649fd92006-10-17 00:09:36 -0700448/*
Nick Piggine2867812008-07-25 19:45:30 -0700449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700451 */
Nick Piggine2867812008-07-25 19:45:30 -0700452static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800453{
Nick Piggin28e4d962006-09-25 23:31:23 -0700454 BUG_ON(!PageLocked(page));
455 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800456
Nick Piggin19fd6232008-07-25 19:45:32 -0700457 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800458 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700459 * The non racy check for a busy page.
460 *
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
465 *
466 * get_user_pages(&page);
467 * [user mapping goes away]
468 * write_to(page);
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
471 * put_page(page);
472 * !page_count(page) [good, discard it]
473 *
474 * [oops, our write_to data is lost]
475 *
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
479 *
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800482 */
Nick Piggine2867812008-07-25 19:45:30 -0700483 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800484 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page))) {
487 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800488 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700489 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800490
491 if (PageSwapCache(page)) {
492 swp_entry_t swap = { .val = page_private(page) };
493 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700494 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700495 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700496 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500497 void (*freepage)(struct page *);
498
499 freepage = mapping->a_ops->freepage;
500
Minchan Kime64a7822011-03-22 16:32:44 -0700501 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700502 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700503 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500504
505 if (freepage != NULL)
506 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800507 }
508
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800509 return 1;
510
511cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700512 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800513 return 0;
514}
515
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516/*
Nick Piggine2867812008-07-25 19:45:30 -0700517 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
518 * someone else has a ref on the page, abort and return 0. If it was
519 * successfully detached, return 1. Assumes the caller has a single ref on
520 * this page.
521 */
522int remove_mapping(struct address_space *mapping, struct page *page)
523{
524 if (__remove_mapping(mapping, page)) {
525 /*
526 * Unfreezing the refcount with 1 rather than 2 effectively
527 * drops the pagecache ref for us without requiring another
528 * atomic operation.
529 */
530 page_unfreeze_refs(page, 1);
531 return 1;
532 }
533 return 0;
534}
535
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700536/**
537 * putback_lru_page - put previously isolated page onto appropriate LRU list
538 * @page: page to be put back to appropriate lru list
539 *
540 * Add previously isolated @page to appropriate LRU list.
541 * Page may still be unevictable for other reasons.
542 *
543 * lru_lock must not be held, interrupts must be enabled.
544 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700545void putback_lru_page(struct page *page)
546{
547 int lru;
548 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700549 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700550
551 VM_BUG_ON(PageLRU(page));
552
553redo:
554 ClearPageUnevictable(page);
555
556 if (page_evictable(page, NULL)) {
557 /*
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
562 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700563 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700564 lru_cache_add_lru(page, lru);
565 } else {
566 /*
567 * Put unevictable pages directly on zone's unevictable
568 * list.
569 */
570 lru = LRU_UNEVICTABLE;
571 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700572 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800576 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700578 * the page back to the evictable list.
579 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700580 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700581 */
582 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700583 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700584
585 /*
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
589 */
590 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
591 if (!isolate_lru_page(page)) {
592 put_page(page);
593 goto redo;
594 }
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
598 */
599 }
600
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700601 if (was_unevictable && lru != LRU_UNEVICTABLE)
602 count_vm_event(UNEVICTABLE_PGRESCUED);
603 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604 count_vm_event(UNEVICTABLE_PGCULLED);
605
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700606 put_page(page); /* drop ref from isolate */
607}
608
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800609enum page_references {
610 PAGEREF_RECLAIM,
611 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800612 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800613 PAGEREF_ACTIVATE,
614};
615
616static enum page_references page_check_references(struct page *page,
617 struct scan_control *sc)
618{
Johannes Weiner645747462010-03-05 13:42:22 -0800619 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800620 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800621
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700622 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800624 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800625
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800626 /*
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
629 */
630 if (vm_flags & VM_LOCKED)
631 return PAGEREF_RECLAIM;
632
Johannes Weiner645747462010-03-05 13:42:22 -0800633 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700634 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800635 return PAGEREF_ACTIVATE;
636 /*
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
640 * than once.
641 *
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
645 *
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
648 * quickly recovered.
649 */
650 SetPageReferenced(page);
651
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800652 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800653 return PAGEREF_ACTIVATE;
654
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800655 /*
656 * Activate file-backed executable pages after first usage.
657 */
658 if (vm_flags & VM_EXEC)
659 return PAGEREF_ACTIVATE;
660
Johannes Weiner645747462010-03-05 13:42:22 -0800661 return PAGEREF_KEEP;
662 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800663
664 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700665 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800666 return PAGEREF_RECLAIM_CLEAN;
667
668 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800669}
670
Nick Piggine2867812008-07-25 19:45:30 -0700671/*
Andrew Morton1742f192006-03-22 00:08:21 -0800672 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 */
Andrew Morton1742f192006-03-22 00:08:21 -0800674static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700675 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700676 struct scan_control *sc,
Mel Gorman92df3a72011-10-31 17:07:56 -0700677 unsigned long *ret_nr_dirty,
678 unsigned long *ret_nr_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679{
680 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700681 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700683 unsigned long nr_dirty = 0;
684 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800685 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700686 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
688 cond_resched();
689
Tim Chen69980e32012-07-31 16:46:08 -0700690 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800692 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 struct address_space *mapping;
694 struct page *page;
695 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
697 cond_resched();
698
699 page = lru_to_page(page_list);
700 list_del(&page->lru);
701
Nick Piggin529ae9a2008-08-02 12:01:03 +0200702 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 goto keep;
704
Nick Piggin725d7042006-09-25 23:30:55 -0700705 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700706 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707
708 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800709
Nick Pigginb291f002008-10-18 20:26:44 -0700710 if (unlikely(!page_evictable(page, NULL)))
711 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700712
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700713 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800714 goto keep_locked;
715
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 /* Double the slab pressure for mapped and swapcache pages */
717 if (page_mapped(page) || PageSwapCache(page))
718 sc->nr_scanned++;
719
Andy Whitcroftc661b072007-08-22 14:01:26 -0700720 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
721 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
722
723 if (PageWriteback(page)) {
Michal Hockoe62e3842012-07-31 16:45:55 -0700724 /*
725 * memcg doesn't have any dirty pages throttling so we
726 * could easily OOM just because too many pages are in
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700727 * writeback and there is nothing else to reclaim.
Michal Hockoe62e3842012-07-31 16:45:55 -0700728 *
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700729 * Check __GFP_IO, certainly because a loop driver
Michal Hockoe62e3842012-07-31 16:45:55 -0700730 * thread might enter reclaim, and deadlock if it waits
731 * on a page for which it is needed to do the write
732 * (loop masks off __GFP_IO|__GFP_FS for this reason);
733 * but more thought would probably show more reasons.
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700734 *
735 * Don't require __GFP_FS, since we're not going into
736 * the FS, just waiting on its writeback completion.
737 * Worryingly, ext4 gfs2 and xfs allocate pages with
738 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
739 * testing may_enter_fs here is liable to OOM on them.
Michal Hockoe62e3842012-07-31 16:45:55 -0700740 */
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700741 if (global_reclaim(sc) ||
742 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
743 /*
744 * This is slightly racy - end_page_writeback()
745 * might have just cleared PageReclaim, then
746 * setting PageReclaim here end up interpreted
747 * as PageReadahead - but that does not matter
748 * enough to care. What we do want is for this
749 * page to have PageReclaim set next time memcg
750 * reclaim reaches the tests above, so it will
751 * then wait_on_page_writeback() to avoid OOM;
752 * and it's also appropriate in global reclaim.
753 */
754 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700755 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700756 goto keep_locked;
Michal Hockoe62e3842012-07-31 16:45:55 -0700757 }
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700758 wait_on_page_writeback(page);
Andy Whitcroftc661b072007-08-22 14:01:26 -0700759 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700761 references = page_check_references(page, sc);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800762 switch (references) {
763 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800765 case PAGEREF_KEEP:
766 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800767 case PAGEREF_RECLAIM:
768 case PAGEREF_RECLAIM_CLEAN:
769 ; /* try to reclaim the page below */
770 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 /*
773 * Anonymous process memory has backing store?
774 * Try to allocate it some swap space here.
775 */
Nick Pigginb291f002008-10-18 20:26:44 -0700776 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800777 if (!(sc->gfp_mask & __GFP_IO))
778 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800779 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800781 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700782 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
784 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
786 /*
787 * The page is mapped into the page tables of one or more
788 * processes. Try to unmap it here.
789 */
790 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200791 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 case SWAP_FAIL:
793 goto activate_locked;
794 case SWAP_AGAIN:
795 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700796 case SWAP_MLOCK:
797 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 case SWAP_SUCCESS:
799 ; /* try to free the page below */
800 }
801 }
802
803 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700804 nr_dirty++;
805
Mel Gormanee728862011-10-31 17:07:38 -0700806 /*
807 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700808 * avoid risk of stack overflow but do not writeback
809 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700810 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700811 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700812 (!current_is_kswapd() ||
813 sc->priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700814 /*
815 * Immediately reclaim when written back.
816 * Similar in principal to deactivate_page()
817 * except we already have the page isolated
818 * and know it's dirty
819 */
820 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
821 SetPageReclaim(page);
822
Mel Gormanee728862011-10-31 17:07:38 -0700823 goto keep_locked;
824 }
825
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800826 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700828 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800830 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 goto keep_locked;
832
833 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700834 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700836 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 goto keep_locked;
838 case PAGE_ACTIVATE:
839 goto activate_locked;
840 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700841 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700842 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700843 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700845
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 /*
847 * A synchronous write - probably a ramdisk. Go
848 * ahead and try to reclaim the page.
849 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200850 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 goto keep;
852 if (PageDirty(page) || PageWriteback(page))
853 goto keep_locked;
854 mapping = page_mapping(page);
855 case PAGE_CLEAN:
856 ; /* try to free the page below */
857 }
858 }
859
860 /*
861 * If the page has buffers, try to free the buffer mappings
862 * associated with this page. If we succeed we try to free
863 * the page as well.
864 *
865 * We do this even if the page is PageDirty().
866 * try_to_release_page() does not perform I/O, but it is
867 * possible for a page to have PageDirty set, but it is actually
868 * clean (all its buffers are clean). This happens if the
869 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700870 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 * try_to_release_page() will discover that cleanness and will
872 * drop the buffers and mark the page clean - it can be freed.
873 *
874 * Rarely, pages can have buffers and no ->mapping. These are
875 * the pages which were not successfully invalidated in
876 * truncate_complete_page(). We try to drop those buffers here
877 * and if that worked, and the page is no longer mapped into
878 * process address space (page_count == 1) it can be freed.
879 * Otherwise, leave the page on the LRU so it is swappable.
880 */
David Howells266cf652009-04-03 16:42:36 +0100881 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 if (!try_to_release_page(page, sc->gfp_mask))
883 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700884 if (!mapping && page_count(page) == 1) {
885 unlock_page(page);
886 if (put_page_testzero(page))
887 goto free_it;
888 else {
889 /*
890 * rare race with speculative reference.
891 * the speculative reference will free
892 * this page shortly, so we may
893 * increment nr_reclaimed here (and
894 * leave it off the LRU).
895 */
896 nr_reclaimed++;
897 continue;
898 }
899 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 }
901
Nick Piggine2867812008-07-25 19:45:30 -0700902 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800903 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
Nick Piggina978d6f2008-10-18 20:26:58 -0700905 /*
906 * At this point, we have no other references and there is
907 * no way to pick any more up (removed from LRU, removed
908 * from pagecache). Can use non-atomic bitops now (and
909 * we obviously don't have to worry about waking up a process
910 * waiting on the page lock, because there are no references.
911 */
912 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700913free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800914 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700915
916 /*
917 * Is there need to periodically free_page_list? It would
918 * appear not as the counts should be low
919 */
920 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 continue;
922
Nick Pigginb291f002008-10-18 20:26:44 -0700923cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800924 if (PageSwapCache(page))
925 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700926 unlock_page(page);
927 putback_lru_page(page);
928 continue;
929
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700931 /* Not a candidate for swapping, so reclaim swap space. */
932 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800933 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700934 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 SetPageActive(page);
936 pgactivate++;
937keep_locked:
938 unlock_page(page);
939keep:
940 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700941 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700943
Mel Gorman0e093d992010-10-26 14:21:45 -0700944 /*
945 * Tag a zone as congested if all the dirty pages encountered were
946 * backed by a congested BDI. In this case, reclaimers should just
947 * back off and wait for congestion to clear because further reclaim
948 * will encounter the same problem
949 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800950 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700951 zone_set_flag(zone, ZONE_CONGESTED);
Mel Gorman0e093d992010-10-26 14:21:45 -0700952
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -0800953 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700954
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700956 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -0700957 mem_cgroup_uncharge_end();
Mel Gorman92df3a72011-10-31 17:07:56 -0700958 *ret_nr_dirty += nr_dirty;
959 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -0800960 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961}
962
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700963/*
964 * Attempt to remove the specified page from its LRU. Only take this page
965 * if it is of the appropriate PageActive status. Pages which are being
966 * freed elsewhere are also ignored.
967 *
968 * page: page to consider
969 * mode: one of the LRU isolation modes defined above
970 *
971 * returns 0 on success, -ve errno on failure.
972 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700973int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700974{
975 int ret = -EINVAL;
976
977 /* Only take pages on the LRU. */
978 if (!PageLRU(page))
979 return ret;
980
Mel Gormanc53919a2012-05-29 15:06:19 -0700981 /* Do not give back unevictable pages for compaction */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700982 if (PageUnevictable(page))
983 return ret;
984
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700985 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800986
Mel Gormanc8244932012-01-12 17:19:38 -0800987 /*
988 * To minimise LRU disruption, the caller can indicate that it only
989 * wants to isolate pages it will be able to operate on without
990 * blocking - clean pages for the most part.
991 *
992 * ISOLATE_CLEAN means that only clean pages should be isolated. This
993 * is used by reclaim when it is cannot write to backing storage
994 *
995 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
996 * that it is possible to migrate without blocking
997 */
998 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
999 /* All the caller can do on PageWriteback is block */
1000 if (PageWriteback(page))
1001 return ret;
1002
1003 if (PageDirty(page)) {
1004 struct address_space *mapping;
1005
1006 /* ISOLATE_CLEAN means only clean pages */
1007 if (mode & ISOLATE_CLEAN)
1008 return ret;
1009
1010 /*
1011 * Only pages without mappings or that have a
1012 * ->migratepage callback are possible to migrate
1013 * without blocking
1014 */
1015 mapping = page_mapping(page);
1016 if (mapping && !mapping->a_ops->migratepage)
1017 return ret;
1018 }
1019 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001020
Minchan Kimf80c0672011-10-31 17:06:55 -07001021 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1022 return ret;
1023
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001024 if (likely(get_page_unless_zero(page))) {
1025 /*
1026 * Be careful not to clear PageLRU until after we're
1027 * sure the page is not being freed elsewhere -- the
1028 * page release code relies on it.
1029 */
1030 ClearPageLRU(page);
1031 ret = 0;
1032 }
1033
1034 return ret;
1035}
1036
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001037/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 * zone->lru_lock is heavily contended. Some of the functions that
1039 * shrink the lists perform better by taking out a batch of pages
1040 * and working on them outside the LRU lock.
1041 *
1042 * For pagecache intensive workloads, this function is the hottest
1043 * spot in the kernel (apart from copy_*_user functions).
1044 *
1045 * Appropriate locks must be held before calling this function.
1046 *
1047 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001048 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001050 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001051 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001052 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001053 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 *
1055 * returns how many pages were moved onto *@dst.
1056 */
Andrew Morton69e05942006-03-22 00:08:19 -08001057static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001058 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001059 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001060 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001062 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001063 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001064 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001066 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001067 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001068 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001069
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 page = lru_to_page(src);
1071 prefetchw_prev_lru_page(page, src, flags);
1072
Nick Piggin725d7042006-09-25 23:30:55 -07001073 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001074
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001075 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001076 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001077 nr_pages = hpage_nr_pages(page);
1078 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001079 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001080 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001081 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001082
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001083 case -EBUSY:
1084 /* else it is being freed elsewhere */
1085 list_move(&page->lru, src);
1086 continue;
1087
1088 default:
1089 BUG();
1090 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 }
1092
Hugh Dickinsf6260122012-01-12 17:20:06 -08001093 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001094 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1095 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 return nr_taken;
1097}
1098
Nick Piggin62695a82008-10-18 20:26:09 -07001099/**
1100 * isolate_lru_page - tries to isolate a page from its LRU list
1101 * @page: page to isolate from its LRU list
1102 *
1103 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1104 * vmstat statistic corresponding to whatever LRU list the page was on.
1105 *
1106 * Returns 0 if the page was removed from an LRU list.
1107 * Returns -EBUSY if the page was not on an LRU list.
1108 *
1109 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001110 * the active list, it will have PageActive set. If it was found on
1111 * the unevictable list, it will have the PageUnevictable bit set. That flag
1112 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001113 *
1114 * The vmstat statistic corresponding to the list on which the page was
1115 * found will be decremented.
1116 *
1117 * Restrictions:
1118 * (1) Must be called with an elevated refcount on the page. This is a
1119 * fundamentnal difference from isolate_lru_pages (which is called
1120 * without a stable reference).
1121 * (2) the lru_lock must not be held.
1122 * (3) interrupts must be enabled.
1123 */
1124int isolate_lru_page(struct page *page)
1125{
1126 int ret = -EBUSY;
1127
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001128 VM_BUG_ON(!page_count(page));
1129
Nick Piggin62695a82008-10-18 20:26:09 -07001130 if (PageLRU(page)) {
1131 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001132 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001133
1134 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001135 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001136 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001137 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001138 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001139 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001140 del_page_from_lru_list(page, lruvec, lru);
1141 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001142 }
1143 spin_unlock_irq(&zone->lru_lock);
1144 }
1145 return ret;
1146}
1147
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001148/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001149 * Are there way too many processes in the direct reclaim path already?
1150 */
1151static int too_many_isolated(struct zone *zone, int file,
1152 struct scan_control *sc)
1153{
1154 unsigned long inactive, isolated;
1155
1156 if (current_is_kswapd())
1157 return 0;
1158
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001159 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001160 return 0;
1161
1162 if (file) {
1163 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1164 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1165 } else {
1166 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1167 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1168 }
1169
1170 return isolated > inactive;
1171}
1172
Mel Gorman66635622010-08-09 17:19:30 -07001173static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001174putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001175{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001176 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1177 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001178 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001179
Mel Gorman66635622010-08-09 17:19:30 -07001180 /*
1181 * Put back any unfreeable pages.
1182 */
Mel Gorman66635622010-08-09 17:19:30 -07001183 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001184 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001185 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001186
Mel Gorman66635622010-08-09 17:19:30 -07001187 VM_BUG_ON(PageLRU(page));
1188 list_del(&page->lru);
1189 if (unlikely(!page_evictable(page, NULL))) {
1190 spin_unlock_irq(&zone->lru_lock);
1191 putback_lru_page(page);
1192 spin_lock_irq(&zone->lru_lock);
1193 continue;
1194 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001195
1196 lruvec = mem_cgroup_page_lruvec(page, zone);
1197
Linus Torvalds7a608572011-01-17 14:42:19 -08001198 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001199 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001200 add_page_to_lru_list(page, lruvec, lru);
1201
Mel Gorman66635622010-08-09 17:19:30 -07001202 if (is_active_lru(lru)) {
1203 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001204 int numpages = hpage_nr_pages(page);
1205 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001206 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001207 if (put_page_testzero(page)) {
1208 __ClearPageLRU(page);
1209 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001210 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001211
1212 if (unlikely(PageCompound(page))) {
1213 spin_unlock_irq(&zone->lru_lock);
1214 (*get_compound_page_dtor(page))(page);
1215 spin_lock_irq(&zone->lru_lock);
1216 } else
1217 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001218 }
1219 }
Mel Gorman66635622010-08-09 17:19:30 -07001220
Hugh Dickins3f797682012-01-12 17:20:07 -08001221 /*
1222 * To save our caller's stack, now use input list for pages to free.
1223 */
1224 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001225}
1226
1227/*
Andrew Morton1742f192006-03-22 00:08:21 -08001228 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1229 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 */
Mel Gorman66635622010-08-09 17:19:30 -07001231static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001232shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001233 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234{
1235 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001236 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001237 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001238 unsigned long nr_taken;
Mel Gorman92df3a72011-10-31 17:07:56 -07001239 unsigned long nr_dirty = 0;
1240 unsigned long nr_writeback = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001241 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001242 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001243 struct zone *zone = lruvec_zone(lruvec);
1244 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001245
Rik van Riel35cd7812009-09-21 17:01:38 -07001246 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001247 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001248
1249 /* We are about to die and free our memory. Return now. */
1250 if (fatal_signal_pending(current))
1251 return SWAP_CLUSTER_MAX;
1252 }
1253
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001255
1256 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001257 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001258 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001259 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001263 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1264 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001265
1266 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1267 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1268
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001269 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001270 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001271 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001272 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001273 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001274 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001275 }
Hillf Dantond563c052012-03-21 16:34:02 -07001276 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001277
Hillf Dantond563c052012-03-21 16:34:02 -07001278 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001279 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001280
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -07001281 nr_reclaimed = shrink_page_list(&page_list, zone, sc,
Mel Gorman92df3a72011-10-31 17:07:56 -07001282 &nr_dirty, &nr_writeback);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001283
Hugh Dickins3f797682012-01-12 17:20:07 -08001284 spin_lock_irq(&zone->lru_lock);
1285
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001286 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001287
Ying Han904249a2012-04-25 16:01:48 -07001288 if (global_reclaim(sc)) {
1289 if (current_is_kswapd())
1290 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1291 nr_reclaimed);
1292 else
1293 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1294 nr_reclaimed);
1295 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001296
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001297 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001298
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001299 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001300
1301 spin_unlock_irq(&zone->lru_lock);
1302
1303 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001304
Mel Gorman92df3a72011-10-31 17:07:56 -07001305 /*
1306 * If reclaim is isolating dirty pages under writeback, it implies
1307 * that the long-lived page allocation rate is exceeding the page
1308 * laundering rate. Either the global limits are not being effective
1309 * at throttling processes due to the page distribution throughout
1310 * zones or there is heavy usage of a slow backing device. The
1311 * only option is to throttle from reclaim context which is not ideal
1312 * as there is no guarantee the dirtying process is throttled in the
1313 * same way balance_dirty_pages() manages.
1314 *
1315 * This scales the number of dirty pages that must be under writeback
1316 * before throttling depending on priority. It is a simple backoff
1317 * function that has the most effect in the range DEF_PRIORITY to
1318 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1319 * in trouble and reclaim is considered to be in trouble.
1320 *
1321 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1322 * DEF_PRIORITY-1 50% must be PageWriteback
1323 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1324 * ...
1325 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1326 * isolated page is PageWriteback
1327 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001328 if (nr_writeback && nr_writeback >=
1329 (nr_taken >> (DEF_PRIORITY - sc->priority)))
Mel Gorman92df3a72011-10-31 17:07:56 -07001330 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1331
Mel Gormane11da5b2010-10-26 14:21:40 -07001332 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1333 zone_idx(zone),
1334 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001335 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001336 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001337 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338}
1339
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001340/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 * This moves pages from the active list to the inactive list.
1342 *
1343 * We move them the other way if the page is referenced by one or more
1344 * processes, from rmap.
1345 *
1346 * If the pages are mostly unmapped, the processing is fast and it is
1347 * appropriate to hold zone->lru_lock across the whole operation. But if
1348 * the pages are mapped, the processing is slow (page_referenced()) so we
1349 * should drop zone->lru_lock around each page. It's impossible to balance
1350 * this, so instead we remove the pages from the LRU while processing them.
1351 * It is safe to rely on PG_active against the non-LRU pages in here because
1352 * nobody will play with that bit on a non-LRU page.
1353 *
1354 * The downside is that we have to touch page->_count against each page.
1355 * But we had to alter page->flags anyway.
1356 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001357
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001358static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001359 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001360 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001361 enum lru_list lru)
1362{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001363 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001364 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001365 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001366 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001367
Wu Fengguang3eb41402009-06-16 15:33:13 -07001368 while (!list_empty(list)) {
1369 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001370 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001371
1372 VM_BUG_ON(PageLRU(page));
1373 SetPageLRU(page);
1374
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001375 nr_pages = hpage_nr_pages(page);
1376 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001377 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001378 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001379
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001380 if (put_page_testzero(page)) {
1381 __ClearPageLRU(page);
1382 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001383 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001384
1385 if (unlikely(PageCompound(page))) {
1386 spin_unlock_irq(&zone->lru_lock);
1387 (*get_compound_page_dtor(page))(page);
1388 spin_lock_irq(&zone->lru_lock);
1389 } else
1390 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001391 }
1392 }
1393 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1394 if (!is_active_lru(lru))
1395 __count_vm_events(PGDEACTIVATE, pgmoved);
1396}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001397
Hugh Dickinsf6260122012-01-12 17:20:06 -08001398static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001399 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001400 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001401 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001403 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001404 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001405 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001407 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001408 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001410 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001411 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001412 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001413 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001414 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415
1416 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001417
1418 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001419 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001420 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001421 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001422
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001424
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001425 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1426 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001427 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001428 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001429
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001430 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001431
Hugh Dickinsf6260122012-01-12 17:20:06 -08001432 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001433 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001434 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 spin_unlock_irq(&zone->lru_lock);
1436
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 while (!list_empty(&l_hold)) {
1438 cond_resched();
1439 page = lru_to_page(&l_hold);
1440 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001441
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001442 if (unlikely(!page_evictable(page, NULL))) {
1443 putback_lru_page(page);
1444 continue;
1445 }
1446
Mel Gormancc715d92012-03-21 16:34:00 -07001447 if (unlikely(buffer_heads_over_limit)) {
1448 if (page_has_private(page) && trylock_page(page)) {
1449 if (page_has_private(page))
1450 try_to_release_page(page, 0);
1451 unlock_page(page);
1452 }
1453 }
1454
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001455 if (page_referenced(page, 0, sc->target_mem_cgroup,
1456 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001457 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001458 /*
1459 * Identify referenced, file-backed active pages and
1460 * give them one more trip around the active list. So
1461 * that executable code get better chances to stay in
1462 * memory under moderate memory pressure. Anon pages
1463 * are not likely to be evicted by use-once streaming
1464 * IO, plus JVM can create lots of anon VM_EXEC pages,
1465 * so we ignore them here.
1466 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001467 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001468 list_add(&page->lru, &l_active);
1469 continue;
1470 }
1471 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001472
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001473 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 list_add(&page->lru, &l_inactive);
1475 }
1476
Andrew Mortonb5557492009-01-06 14:40:13 -08001477 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001478 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001479 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001480 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001481 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001482 * Count referenced pages from currently used mappings as rotated,
1483 * even though only some of them are actually re-activated. This
1484 * helps balance scan pressure between file and anonymous pages in
1485 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001486 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001487 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001488
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001489 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1490 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001491 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001492 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001493
1494 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495}
1496
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001497#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001498static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001499{
1500 unsigned long active, inactive;
1501
1502 active = zone_page_state(zone, NR_ACTIVE_ANON);
1503 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1504
1505 if (inactive * zone->inactive_ratio < active)
1506 return 1;
1507
1508 return 0;
1509}
1510
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001511/**
1512 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001513 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001514 *
1515 * Returns true if the zone does not have enough inactive anon pages,
1516 * meaning some active anon pages need to be deactivated.
1517 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001518static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001519{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001520 /*
1521 * If we don't have swap space, anonymous page deactivation
1522 * is pointless.
1523 */
1524 if (!total_swap_pages)
1525 return 0;
1526
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001527 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001528 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001529
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001530 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001531}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001532#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001533static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001534{
1535 return 0;
1536}
1537#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001538
Rik van Riel56e49d22009-06-16 15:32:28 -07001539static int inactive_file_is_low_global(struct zone *zone)
1540{
1541 unsigned long active, inactive;
1542
1543 active = zone_page_state(zone, NR_ACTIVE_FILE);
1544 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1545
1546 return (active > inactive);
1547}
1548
1549/**
1550 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001551 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001552 *
1553 * When the system is doing streaming IO, memory pressure here
1554 * ensures that active file pages get deactivated, until more
1555 * than half of the file pages are on the inactive list.
1556 *
1557 * Once we get to that situation, protect the system's working
1558 * set from being evicted by disabling active file page aging.
1559 *
1560 * This uses a different ratio than the anonymous pages, because
1561 * the page cache uses a use-once replacement algorithm.
1562 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001563static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001564{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001565 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001566 return mem_cgroup_inactive_file_is_low(lruvec);
Rik van Riel56e49d22009-06-16 15:32:28 -07001567
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001568 return inactive_file_is_low_global(lruvec_zone(lruvec));
Rik van Riel56e49d22009-06-16 15:32:28 -07001569}
1570
Hugh Dickins75b00af2012-05-29 15:07:09 -07001571static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001572{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001573 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001574 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001575 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001576 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001577}
1578
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001579static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001580 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001581{
Rik van Rielb39415b2009-12-14 17:59:48 -08001582 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001583 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001584 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001585 return 0;
1586 }
1587
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001588 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001589}
1590
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001591static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001592{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001593 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001594 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001595 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001596}
1597
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001599 * Determine how aggressively the anon and file LRU lists should be
1600 * scanned. The relative value of each set of LRU lists is determined
1601 * by looking at the fraction of the pages scanned we did rotate back
1602 * onto the active list instead of evict.
1603 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001604 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1605 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001606 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001607static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001608 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001609{
1610 unsigned long anon, file, free;
1611 unsigned long anon_prio, file_prio;
1612 unsigned long ap, fp;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001613 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001614 u64 fraction[2], denominator;
Hugh Dickins41113042012-01-12 17:20:01 -08001615 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001616 int noswap = 0;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001617 bool force_scan = false;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001618 struct zone *zone = lruvec_zone(lruvec);
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001619
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001620 /*
1621 * If the zone or memcg is small, nr[l] can be 0. This
1622 * results in no scanning on this priority and a potential
1623 * priority drop. Global direct reclaim can go to the next
1624 * zone and tends to have no problems. Global kswapd is for
1625 * zone balancing and it needs to scan a minimum amount. When
1626 * reclaiming for a memcg, a priority drop can cause high
1627 * latencies, so it's better to scan a minimum amount there as
1628 * well.
1629 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001630 if (current_is_kswapd() && zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001631 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001632 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001633 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001634
1635 /* If we have no swap space, do not bother scanning anon pages. */
1636 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1637 noswap = 1;
1638 fraction[0] = 0;
1639 fraction[1] = 1;
1640 denominator = 1;
1641 goto out;
1642 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001643
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001644 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1645 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1646 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1647 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001648
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001649 if (global_reclaim(sc)) {
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001650 free = zone_page_state(zone, NR_FREE_PAGES);
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001651 /* If we have very few page cache pages,
1652 force-scan anon pages. */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001653 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001654 fraction[0] = 1;
1655 fraction[1] = 0;
1656 denominator = 1;
1657 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001658 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001659 }
1660
1661 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001662 * With swappiness at 100, anonymous and file have the same priority.
1663 * This scanning priority is essentially the inverse of IO cost.
1664 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001665 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001666 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001667
1668 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001669 * OK, so we have swap space and a fair amount of page cache
1670 * pages. We use the recently rotated / recently scanned
1671 * ratios to determine how valuable each cache is.
1672 *
1673 * Because workloads change over time (and to avoid overflow)
1674 * we keep these statistics as a floating average, which ends
1675 * up weighing recent references more than old ones.
1676 *
1677 * anon in [0], file in [1]
1678 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001679 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001680 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001681 reclaim_stat->recent_scanned[0] /= 2;
1682 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001683 }
1684
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001685 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001686 reclaim_stat->recent_scanned[1] /= 2;
1687 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001688 }
1689
1690 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001691 * The amount of pressure on anon vs file pages is inversely
1692 * proportional to the fraction of recently scanned pages on
1693 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001694 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001695 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001696 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001697
Satoru Moriyafe350042012-05-29 15:06:47 -07001698 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001699 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001700 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001701
Shaohua Li76a33fc2010-05-24 14:32:36 -07001702 fraction[0] = ap;
1703 fraction[1] = fp;
1704 denominator = ap + fp + 1;
1705out:
Hugh Dickins41113042012-01-12 17:20:01 -08001706 for_each_evictable_lru(lru) {
1707 int file = is_file_lru(lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001708 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001709
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001710 scan = get_lru_size(lruvec, lru);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001711 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1712 scan >>= sc->priority;
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001713 if (!scan && force_scan)
1714 scan = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001715 scan = div64_u64(scan * fraction[file], denominator);
1716 }
Hugh Dickins41113042012-01-12 17:20:01 -08001717 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001718 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001719}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001720
Mel Gorman23b9da52012-05-29 15:06:20 -07001721/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001722static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07001723{
1724 if (COMPACTION_BUILD && sc->order &&
1725 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001726 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07001727 return true;
1728
1729 return false;
1730}
1731
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001732/*
Mel Gorman23b9da52012-05-29 15:06:20 -07001733 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1734 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1735 * true if more pages should be reclaimed such that when the page allocator
1736 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1737 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08001738 */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001739static inline bool should_continue_reclaim(struct lruvec *lruvec,
Mel Gorman3e7d3442011-01-13 15:45:56 -08001740 unsigned long nr_reclaimed,
1741 unsigned long nr_scanned,
1742 struct scan_control *sc)
1743{
1744 unsigned long pages_for_compaction;
1745 unsigned long inactive_lru_pages;
1746
1747 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001748 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001749 return false;
1750
Mel Gorman28765922011-02-25 14:44:20 -08001751 /* Consider stopping depending on scan and reclaim activity */
1752 if (sc->gfp_mask & __GFP_REPEAT) {
1753 /*
1754 * For __GFP_REPEAT allocations, stop reclaiming if the
1755 * full LRU list has been scanned and we are still failing
1756 * to reclaim pages. This full LRU scan is potentially
1757 * expensive but a __GFP_REPEAT caller really wants to succeed
1758 */
1759 if (!nr_reclaimed && !nr_scanned)
1760 return false;
1761 } else {
1762 /*
1763 * For non-__GFP_REPEAT allocations which can presumably
1764 * fail without consequence, stop if we failed to reclaim
1765 * any pages from the last SWAP_CLUSTER_MAX number of
1766 * pages that were scanned. This will return to the
1767 * caller faster at the risk reclaim/compaction and
1768 * the resulting allocation attempt fails
1769 */
1770 if (!nr_reclaimed)
1771 return false;
1772 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001773
1774 /*
1775 * If we have not reclaimed enough pages for compaction and the
1776 * inactive lists are large enough, continue reclaiming
1777 */
1778 pages_for_compaction = (2UL << sc->order);
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001779 inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
Minchan Kim86cfd3a2012-01-10 15:08:18 -08001780 if (nr_swap_pages > 0)
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001781 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001782 if (sc->nr_reclaimed < pages_for_compaction &&
1783 inactive_lru_pages > pages_for_compaction)
1784 return true;
1785
1786 /* If compaction would go ahead or the allocation would succeed, stop */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001787 switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08001788 case COMPACT_PARTIAL:
1789 case COMPACT_CONTINUE:
1790 return false;
1791 default:
1792 return true;
1793 }
1794}
1795
1796/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1798 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001799static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001801 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001802 unsigned long nr_to_scan;
Hugh Dickins41113042012-01-12 17:20:01 -08001803 enum lru_list lru;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001804 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001805 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Shaohua Li3da367c2011-10-31 17:07:03 -07001806 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
Mel Gorman3e7d3442011-01-13 15:45:56 -08001808restart:
1809 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001810 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001811 get_scan_count(lruvec, sc, nr);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001812
Shaohua Li3da367c2011-10-31 17:07:03 -07001813 blk_start_plug(&plug);
Rik van Riel556adec2008-10-18 20:26:34 -07001814 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1815 nr[LRU_INACTIVE_FILE]) {
Hugh Dickins41113042012-01-12 17:20:01 -08001816 for_each_evictable_lru(lru) {
1817 if (nr[lru]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001818 nr_to_scan = min_t(unsigned long,
Hugh Dickins41113042012-01-12 17:20:01 -08001819 nr[lru], SWAP_CLUSTER_MAX);
1820 nr[lru] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821
Hugh Dickins41113042012-01-12 17:20:01 -08001822 nr_reclaimed += shrink_list(lru, nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001823 lruvec, sc);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 }
Rik van Riela79311c2009-01-06 14:40:01 -08001826 /*
1827 * On large memory systems, scan >> priority can become
1828 * really large. This is fine for the starting priority;
1829 * we want to put equal scanning pressure on each zone.
1830 * However, if the VM has a harder time of freeing pages,
1831 * with multiple processes reclaiming pages, the total
1832 * freeing target can get unreasonably large.
1833 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001834 if (nr_reclaimed >= nr_to_reclaim &&
1835 sc->priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001836 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 }
Shaohua Li3da367c2011-10-31 17:07:03 -07001838 blk_finish_plug(&plug);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001839 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001840
Rik van Riel556adec2008-10-18 20:26:34 -07001841 /*
1842 * Even if we did not try to evict anon pages at all, we want to
1843 * rebalance the anon lru active/inactive ratio.
1844 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001845 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001846 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001847 sc, LRU_ACTIVE_ANON);
Rik van Riel556adec2008-10-18 20:26:34 -07001848
Mel Gorman3e7d3442011-01-13 15:45:56 -08001849 /* reclaim/compaction might need reclaim to continue */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001850 if (should_continue_reclaim(lruvec, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001851 sc->nr_scanned - nr_scanned, sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001852 goto restart;
1853
Andrew Morton232ea4d2007-02-28 20:13:21 -08001854 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855}
1856
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001857static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08001858{
Johannes Weiner56600482012-01-12 17:17:59 -08001859 struct mem_cgroup *root = sc->target_mem_cgroup;
1860 struct mem_cgroup_reclaim_cookie reclaim = {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001861 .zone = zone,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001862 .priority = sc->priority,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001863 };
Johannes Weiner56600482012-01-12 17:17:59 -08001864 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001865
Johannes Weiner56600482012-01-12 17:17:59 -08001866 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1867 do {
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001868 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner56600482012-01-12 17:17:59 -08001869
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001870 shrink_lruvec(lruvec, sc);
1871
Johannes Weiner56600482012-01-12 17:17:59 -08001872 /*
1873 * Limit reclaim has historically picked one memcg and
1874 * scanned it with decreasing priority levels until
1875 * nr_to_reclaim had been reclaimed. This priority
1876 * cycle is thus over after a single memcg.
Johannes Weinerb95a2f22012-01-12 17:18:06 -08001877 *
1878 * Direct reclaim and kswapd, on the other hand, have
1879 * to scan all memory cgroups to fulfill the overall
1880 * scan target for the zone.
Johannes Weiner56600482012-01-12 17:17:59 -08001881 */
1882 if (!global_reclaim(sc)) {
1883 mem_cgroup_iter_break(root, memcg);
1884 break;
1885 }
1886 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1887 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001888}
1889
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001890/* Returns true if compaction should go ahead for a high-order request */
1891static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1892{
1893 unsigned long balance_gap, watermark;
1894 bool watermark_ok;
1895
1896 /* Do not consider compaction for orders reclaim is meant to satisfy */
1897 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1898 return false;
1899
1900 /*
1901 * Compaction takes time to run and there are potentially other
1902 * callers using the pages just freed. Continue reclaiming until
1903 * there is a buffer of free pages available to give compaction
1904 * a reasonable chance of completing and allocating the page
1905 */
1906 balance_gap = min(low_wmark_pages(zone),
1907 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1908 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1909 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1910 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1911
1912 /*
1913 * If compaction is deferred, reclaim up to a point where
1914 * compaction will have a chance of success when re-enabled
1915 */
Rik van Rielaff62242012-03-21 16:33:52 -07001916 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001917 return watermark_ok;
1918
1919 /* If compaction is not ready to start, keep reclaiming */
1920 if (!compaction_suitable(zone, sc->order))
1921 return false;
1922
1923 return watermark_ok;
1924}
1925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926/*
1927 * This is the direct reclaim path, for page-allocating processes. We only
1928 * try to reclaim pages from zones which will satisfy the caller's allocation
1929 * request.
1930 *
Mel Gorman41858962009-06-16 15:32:12 -07001931 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1932 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1934 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001935 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1936 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1937 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 * If a zone is deemed to be full of pinned pages then just give it a light
1940 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07001941 *
1942 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001943 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08001944 * the caller that it should consider retrying the allocation instead of
1945 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001947static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948{
Mel Gormandd1a2392008-04-28 02:12:17 -07001949 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001950 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07001951 unsigned long nr_soft_reclaimed;
1952 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08001953 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001954
Mel Gormancc715d92012-03-21 16:34:00 -07001955 /*
1956 * If the number of buffer_heads in the machine exceeds the maximum
1957 * allowed level, force direct reclaim to scan the highmem zone as
1958 * highmem pages could be pinning lowmem pages storing buffer_heads
1959 */
1960 if (buffer_heads_over_limit)
1961 sc->gfp_mask |= __GFP_HIGHMEM;
1962
Mel Gormand4debc62010-08-09 17:19:29 -07001963 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1964 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001965 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001967 /*
1968 * Take care memory controller reclaiming has small influence
1969 * to global LRU.
1970 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001971 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001972 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1973 continue;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001974 if (zone->all_unreclaimable &&
1975 sc->priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001976 continue; /* Let kswapd poll it */
Rik van Riele0887c12011-10-31 17:09:31 -07001977 if (COMPACTION_BUILD) {
1978 /*
Mel Gormane0c23272011-10-31 17:09:33 -07001979 * If we already have plenty of memory free for
1980 * compaction in this zone, don't free any more.
1981 * Even though compaction is invoked for any
1982 * non-zero order, only frequent costly order
1983 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07001984 * noticeable problem, like transparent huge
1985 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07001986 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001987 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08001988 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07001989 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07001990 }
Rik van Riele0887c12011-10-31 17:09:31 -07001991 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07001992 /*
1993 * This steals pages from memory cgroups over softlimit
1994 * and returns the number of reclaimed pages and
1995 * scanned pages. This works for global memory pressure
1996 * and balancing, not for a memcg's limit.
1997 */
1998 nr_soft_scanned = 0;
1999 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2000 sc->order, sc->gfp_mask,
2001 &nr_soft_scanned);
2002 sc->nr_reclaimed += nr_soft_reclaimed;
2003 sc->nr_scanned += nr_soft_scanned;
2004 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002005 }
Nick Piggin408d8542006-09-25 23:31:27 -07002006
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002007 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 }
Mel Gormane0c23272011-10-31 17:09:33 -07002009
Mel Gorman0cee34f2012-01-12 17:19:49 -08002010 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002011}
2012
2013static bool zone_reclaimable(struct zone *zone)
2014{
2015 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2016}
2017
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002018/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002019static bool all_unreclaimable(struct zonelist *zonelist,
2020 struct scan_control *sc)
2021{
2022 struct zoneref *z;
2023 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002024
2025 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2026 gfp_zone(sc->gfp_mask), sc->nodemask) {
2027 if (!populated_zone(zone))
2028 continue;
2029 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2030 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002031 if (!zone->all_unreclaimable)
2032 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002033 }
2034
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002035 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002037
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038/*
2039 * This is the main entry point to direct page reclaim.
2040 *
2041 * If a full scan of the inactive list fails to free enough memory then we
2042 * are "out of memory" and something needs to be killed.
2043 *
2044 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2045 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002046 * caller can't do much about. We kick the writeback threads and take explicit
2047 * naps in the hope that some of these pages can be written. But if the
2048 * allocating task holds filesystem locks which prevent writeout this might not
2049 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002050 *
2051 * returns: 0, if no pages reclaimed
2052 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 */
Mel Gormandac1d272008-04-28 02:12:12 -07002054static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002055 struct scan_control *sc,
2056 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057{
Andrew Morton69e05942006-03-22 00:08:19 -08002058 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002060 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002061 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002062 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002063 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064
Keika Kobayashi873b4772008-07-25 01:48:52 -07002065 delayacct_freepages_start();
2066
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002067 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002068 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002070 do {
Balbir Singh66e17072008-02-07 00:13:56 -08002071 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002072 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002073
Balbir Singh66e17072008-02-07 00:13:56 -08002074 /*
2075 * Don't shrink slabs when reclaiming memory from
2076 * over limit cgroups
2077 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002078 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002079 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002080 for_each_zone_zonelist(zone, z, zonelist,
2081 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002082 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2083 continue;
2084
2085 lru_pages += zone_reclaimable_pages(zone);
2086 }
2087
Ying Han1495f232011-05-24 17:12:27 -07002088 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002089 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002090 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002091 reclaim_state->reclaimed_slab = 0;
2092 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093 }
Balbir Singh66e17072008-02-07 00:13:56 -08002094 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002095 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097
2098 /*
2099 * Try to write back as many pages as we just scanned. This
2100 * tends to cause slow streaming writers to write data to the
2101 * disk smoothly, at the dirtying rate, which is nice. But
2102 * that's undesirable in laptop mode, where we *want* lumpy
2103 * writeout. So in laptop mode, write out the whole world.
2104 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002105 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2106 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002107 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2108 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002109 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 }
2111
2112 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002113 if (!sc->hibernation_mode && sc->nr_scanned &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002114 sc->priority < DEF_PRIORITY - 2) {
Mel Gorman0e093d992010-10-26 14:21:45 -07002115 struct zone *preferred_zone;
2116
2117 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002118 &cpuset_current_mems_allowed,
2119 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002120 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2121 }
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002122 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002123
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002125 delayacct_freepages_end();
2126
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002127 if (sc->nr_reclaimed)
2128 return sc->nr_reclaimed;
2129
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002130 /*
2131 * As hibernation is going on, kswapd is freezed so that it can't mark
2132 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2133 * check.
2134 */
2135 if (oom_killer_disabled)
2136 return 0;
2137
Mel Gorman0cee34f2012-01-12 17:19:49 -08002138 /* Aborted reclaim to try compaction? don't OOM, then */
2139 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002140 return 1;
2141
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002142 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002143 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002144 return 1;
2145
2146 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147}
2148
Mel Gorman55150612012-07-31 16:44:35 -07002149static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2150{
2151 struct zone *zone;
2152 unsigned long pfmemalloc_reserve = 0;
2153 unsigned long free_pages = 0;
2154 int i;
2155 bool wmark_ok;
2156
2157 for (i = 0; i <= ZONE_NORMAL; i++) {
2158 zone = &pgdat->node_zones[i];
2159 pfmemalloc_reserve += min_wmark_pages(zone);
2160 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2161 }
2162
2163 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2164
2165 /* kswapd must be awake if processes are being throttled */
2166 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2167 pgdat->classzone_idx = min(pgdat->classzone_idx,
2168 (enum zone_type)ZONE_NORMAL);
2169 wake_up_interruptible(&pgdat->kswapd_wait);
2170 }
2171
2172 return wmark_ok;
2173}
2174
2175/*
2176 * Throttle direct reclaimers if backing storage is backed by the network
2177 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2178 * depleted. kswapd will continue to make progress and wake the processes
2179 * when the low watermark is reached
2180 */
2181static void throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2182 nodemask_t *nodemask)
2183{
2184 struct zone *zone;
2185 int high_zoneidx = gfp_zone(gfp_mask);
2186 pg_data_t *pgdat;
2187
2188 /*
2189 * Kernel threads should not be throttled as they may be indirectly
2190 * responsible for cleaning pages necessary for reclaim to make forward
2191 * progress. kjournald for example may enter direct reclaim while
2192 * committing a transaction where throttling it could forcing other
2193 * processes to block on log_wait_commit().
2194 */
2195 if (current->flags & PF_KTHREAD)
2196 return;
2197
2198 /* Check if the pfmemalloc reserves are ok */
2199 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2200 pgdat = zone->zone_pgdat;
2201 if (pfmemalloc_watermark_ok(pgdat))
2202 return;
2203
Mel Gorman68243e72012-07-31 16:44:39 -07002204 /* Account for the throttling */
2205 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2206
Mel Gorman55150612012-07-31 16:44:35 -07002207 /*
2208 * If the caller cannot enter the filesystem, it's possible that it
2209 * is due to the caller holding an FS lock or performing a journal
2210 * transaction in the case of a filesystem like ext[3|4]. In this case,
2211 * it is not safe to block on pfmemalloc_wait as kswapd could be
2212 * blocked waiting on the same lock. Instead, throttle for up to a
2213 * second before continuing.
2214 */
2215 if (!(gfp_mask & __GFP_FS)) {
2216 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2217 pfmemalloc_watermark_ok(pgdat), HZ);
2218 return;
2219 }
2220
2221 /* Throttle until kswapd wakes the process */
2222 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2223 pfmemalloc_watermark_ok(pgdat));
2224}
2225
Mel Gormandac1d272008-04-28 02:12:12 -07002226unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002227 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002228{
Mel Gorman33906bc2010-08-09 17:19:16 -07002229 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002230 struct scan_control sc = {
2231 .gfp_mask = gfp_mask,
2232 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002233 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002234 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002235 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002236 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002237 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002238 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002239 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002240 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002241 struct shrink_control shrink = {
2242 .gfp_mask = sc.gfp_mask,
2243 };
Balbir Singh66e17072008-02-07 00:13:56 -08002244
Mel Gorman55150612012-07-31 16:44:35 -07002245 throttle_direct_reclaim(gfp_mask, zonelist, nodemask);
2246
2247 /*
2248 * Do not enter reclaim if fatal signal is pending. 1 is returned so
2249 * that the page allocator does not consider triggering OOM
2250 */
2251 if (fatal_signal_pending(current))
2252 return 1;
2253
Mel Gorman33906bc2010-08-09 17:19:16 -07002254 trace_mm_vmscan_direct_reclaim_begin(order,
2255 sc.may_writepage,
2256 gfp_mask);
2257
Ying Hana09ed5e2011-05-24 17:12:26 -07002258 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002259
2260 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2261
2262 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002263}
2264
Andrew Mortonc255a452012-07-31 16:43:02 -07002265#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002266
Johannes Weiner72835c82012-01-12 17:18:32 -08002267unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002268 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002269 struct zone *zone,
2270 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002271{
2272 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002273 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002274 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002275 .may_writepage = !laptop_mode,
2276 .may_unmap = 1,
2277 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002278 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002279 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002280 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002281 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002282 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002283
Balbir Singh4e416952009-09-23 15:56:39 -07002284 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2285 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002286
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002287 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002288 sc.may_writepage,
2289 sc.gfp_mask);
2290
Balbir Singh4e416952009-09-23 15:56:39 -07002291 /*
2292 * NOTE: Although we can get the priority field, using it
2293 * here is not a good idea, since it limits the pages we can scan.
2294 * if we don't reclaim here, the shrink_zone from balance_pgdat
2295 * will pick up pages from other mem cgroup's as well. We hack
2296 * the priority and make it zero.
2297 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002298 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002299
2300 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2301
Ying Han0ae5e892011-05-26 16:25:25 -07002302 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002303 return sc.nr_reclaimed;
2304}
2305
Johannes Weiner72835c82012-01-12 17:18:32 -08002306unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002307 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002308 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002309{
Balbir Singh4e416952009-09-23 15:56:39 -07002310 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002311 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002312 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002313 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002314 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002315 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002316 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002317 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002318 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002319 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002320 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002321 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002322 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2323 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2324 };
2325 struct shrink_control shrink = {
2326 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002327 };
Balbir Singh66e17072008-02-07 00:13:56 -08002328
Ying Han889976d2011-05-26 16:25:33 -07002329 /*
2330 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2331 * take care of from where we get pages. So the node where we start the
2332 * scan does not need to be the current node.
2333 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002334 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002335
2336 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002337
2338 trace_mm_vmscan_memcg_reclaim_begin(0,
2339 sc.may_writepage,
2340 sc.gfp_mask);
2341
Ying Hana09ed5e2011-05-24 17:12:26 -07002342 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002343
2344 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2345
2346 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002347}
2348#endif
2349
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002350static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002351{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002352 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002353
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002354 if (!total_swap_pages)
2355 return;
2356
2357 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2358 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002359 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002360
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002361 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002362 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002363 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002364
2365 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2366 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002367}
2368
Mel Gorman1741c872011-01-13 15:46:21 -08002369/*
2370 * pgdat_balanced is used when checking if a node is balanced for high-order
2371 * allocations. Only zones that meet watermarks and are in a zone allowed
2372 * by the callers classzone_idx are added to balanced_pages. The total of
2373 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2374 * for the node to be considered balanced. Forcing all zones to be balanced
2375 * for high orders can cause excessive reclaim when there are imbalanced zones.
2376 * The choice of 25% is due to
2377 * o a 16M DMA zone that is balanced will not balance a zone on any
2378 * reasonable sized machine
2379 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002380 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002381 * would need to be at least 256M for it to be balance a whole node.
2382 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2383 * to balance a node on its own. These seemed like reasonable ratios.
2384 */
2385static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2386 int classzone_idx)
2387{
2388 unsigned long present_pages = 0;
2389 int i;
2390
2391 for (i = 0; i <= classzone_idx; i++)
2392 present_pages += pgdat->node_zones[i].present_pages;
2393
Shaohua Li4746efd2011-07-19 08:49:26 -07002394 /* A special case here: if zone has no page, we think it's balanced */
2395 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002396}
2397
Mel Gorman55150612012-07-31 16:44:35 -07002398/*
2399 * Prepare kswapd for sleeping. This verifies that there are no processes
2400 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2401 *
2402 * Returns true if kswapd is ready to sleep
2403 */
2404static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002405 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002406{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002407 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002408 unsigned long balanced = 0;
2409 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002410
2411 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2412 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002413 return false;
2414
2415 /*
2416 * There is a potential race between when kswapd checks its watermarks
2417 * and a process gets throttled. There is also a potential race if
2418 * processes get throttled, kswapd wakes, a large process exits therby
2419 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2420 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2421 * so wake them now if necessary. If necessary, processes will wake
2422 * kswapd and get throttled again
2423 */
2424 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2425 wake_up(&pgdat->pfmemalloc_wait);
2426 return false;
2427 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002428
Mel Gorman0abdee22011-01-13 15:46:22 -08002429 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002430 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002431 struct zone *zone = pgdat->node_zones + i;
2432
2433 if (!populated_zone(zone))
2434 continue;
2435
Mel Gorman355b09c2011-01-13 15:46:24 -08002436 /*
2437 * balance_pgdat() skips over all_unreclaimable after
2438 * DEF_PRIORITY. Effectively, it considers them balanced so
2439 * they must be considered balanced here as well if kswapd
2440 * is to sleep
2441 */
2442 if (zone->all_unreclaimable) {
2443 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002444 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002445 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002446
Mel Gorman88f5acf2011-01-13 15:45:41 -08002447 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002448 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002449 all_zones_ok = false;
2450 else
2451 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002452 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002453
Mel Gorman1741c872011-01-13 15:46:21 -08002454 /*
2455 * For high-order requests, the balanced zones must contain at least
2456 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2457 * must be balanced
2458 */
2459 if (order)
Mel Gorman55150612012-07-31 16:44:35 -07002460 return pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002461 else
Mel Gorman55150612012-07-31 16:44:35 -07002462 return all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002463}
2464
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465/*
2466 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002467 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002469 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 *
2471 * There is special handling here for zones which are full of pinned pages.
2472 * This can happen if the pages are all mlocked, or if they are all used by
2473 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2474 * What we do is to detect the case where all pages in the zone have been
2475 * scanned twice and there has been zero successful reclaim. Mark the zone as
2476 * dead and from now on, only perform a short scan. Basically we're polling
2477 * the zone for when the problem goes away.
2478 *
2479 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002480 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2481 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2482 * lower zones regardless of the number of free pages in the lower zones. This
2483 * interoperates with the page allocator fallback scheme to ensure that aging
2484 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 */
Mel Gorman99504742011-01-13 15:46:20 -08002486static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002487 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002490 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002492 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002493 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002495 unsigned long nr_soft_reclaimed;
2496 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002497 struct scan_control sc = {
2498 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002499 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002500 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002501 /*
2502 * kswapd doesn't want to be bailed out while reclaim. because
2503 * we want to put equal scanning pressure on each zone.
2504 */
2505 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002506 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002507 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002508 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002509 struct shrink_control shrink = {
2510 .gfp_mask = sc.gfp_mask,
2511 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512loop_again:
2513 total_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002514 sc.priority = DEF_PRIORITY;
Rik van Riela79311c2009-01-06 14:40:01 -08002515 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002516 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002517 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002519 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002521 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
2523 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002524 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002526 /*
2527 * Scan in the highmem->dma direction for the highest
2528 * zone which needs scanning
2529 */
2530 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2531 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002533 if (!populated_zone(zone))
2534 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002536 if (zone->all_unreclaimable &&
2537 sc.priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002538 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539
Rik van Riel556adec2008-10-18 20:26:34 -07002540 /*
2541 * Do some background aging of the anon list, to give
2542 * pages a chance to be referenced before reclaiming.
2543 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002544 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002545
Mel Gormancc715d92012-03-21 16:34:00 -07002546 /*
2547 * If the number of buffer_heads in the machine
2548 * exceeds the maximum allowed level and this node
2549 * has a highmem zone, force kswapd to reclaim from
2550 * it to relieve lowmem pressure.
2551 */
2552 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2553 end_zone = i;
2554 break;
2555 }
2556
Mel Gorman88f5acf2011-01-13 15:45:41 -08002557 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002558 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002559 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002560 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002561 } else {
2562 /* If balanced, clear the congested flag */
2563 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002566 if (i < 0)
2567 goto out;
2568
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 for (i = 0; i <= end_zone; i++) {
2570 struct zone *zone = pgdat->node_zones + i;
2571
Wu Fengguangadea02a2009-09-21 17:01:42 -07002572 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 }
2574
2575 /*
2576 * Now scan the zone in the dma->highmem direction, stopping
2577 * at the last zone which needs scanning.
2578 *
2579 * We do this because the page allocator works in the opposite
2580 * direction. This prevents the page allocator from allocating
2581 * pages behind kswapd's direction of progress, which would
2582 * cause too much scanning of the lower zones.
2583 */
2584 for (i = 0; i <= end_zone; i++) {
2585 struct zone *zone = pgdat->node_zones + i;
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002586 int nr_slab, testorder;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002587 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Con Kolivasf3fe6512006-01-06 00:11:15 -08002589 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 continue;
2591
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002592 if (zone->all_unreclaimable &&
2593 sc.priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 continue;
2595
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002597
Ying Han0ae5e892011-05-26 16:25:25 -07002598 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002599 /*
2600 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002601 */
Ying Han0ae5e892011-05-26 16:25:25 -07002602 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2603 order, sc.gfp_mask,
2604 &nr_soft_scanned);
2605 sc.nr_reclaimed += nr_soft_reclaimed;
2606 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002607
Rik van Riel32a43302007-10-16 01:24:50 -07002608 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002609 * We put equal pressure on every zone, unless
2610 * one zone has way too many pages free
2611 * already. The "too many pages" is defined
2612 * as the high wmark plus a "gap" where the
2613 * gap is either the low watermark or 1%
2614 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002615 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002616 balance_gap = min(low_wmark_pages(zone),
2617 (zone->present_pages +
2618 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2619 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002620 /*
2621 * Kswapd reclaims only single pages with compaction
2622 * enabled. Trying too hard to reclaim until contiguous
2623 * free pages have become available can hurt performance
2624 * by evicting too much useful data from memory.
2625 * Do not reclaim more than needed for compaction.
2626 */
2627 testorder = order;
2628 if (COMPACTION_BUILD && order &&
2629 compaction_suitable(zone, order) !=
2630 COMPACT_SKIPPED)
2631 testorder = 0;
2632
Mel Gormancc715d92012-03-21 16:34:00 -07002633 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
Hugh Dickins643ac9f2012-03-23 02:57:31 -07002634 !zone_watermark_ok_safe(zone, testorder,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002635 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002636 end_zone, 0)) {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002637 shrink_zone(zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002638
Mel Gormand7868da2011-07-08 15:39:38 -07002639 reclaim_state->reclaimed_slab = 0;
2640 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2641 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2642 total_scanned += sc.nr_scanned;
2643
2644 if (nr_slab == 0 && !zone_reclaimable(zone))
2645 zone->all_unreclaimable = 1;
2646 }
2647
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 /*
2649 * If we've done a decent amount of scanning and
2650 * the reclaim ratio is low, start doing writepage
2651 * even in laptop mode
2652 */
2653 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002654 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002656
Mel Gorman215ddd62011-07-08 15:39:40 -07002657 if (zone->all_unreclaimable) {
2658 if (end_zone && end_zone == i)
2659 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002660 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002661 }
Mel Gormand7868da2011-07-08 15:39:38 -07002662
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002663 if (!zone_watermark_ok_safe(zone, testorder,
Minchan Kim45973d72010-03-05 13:41:45 -08002664 high_wmark_pages(zone), end_zone, 0)) {
2665 all_zones_ok = 0;
2666 /*
2667 * We are still under min water mark. This
2668 * means that we have a GFP_ATOMIC allocation
2669 * failure risk. Hurry up!
2670 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002671 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002672 min_wmark_pages(zone), end_zone, 0))
2673 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002674 } else {
2675 /*
2676 * If a zone reaches its high watermark,
2677 * consider it to be no longer congested. It's
2678 * possible there are dirty pages backed by
2679 * congested BDIs but as pressure is relieved,
Wanpeng Liab8704b2012-06-17 09:27:18 +08002680 * speculatively avoid congestion waits
Mel Gorman0e093d992010-10-26 14:21:45 -07002681 */
2682 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002683 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002684 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002685 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002686
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 }
Mel Gorman55150612012-07-31 16:44:35 -07002688
2689 /*
2690 * If the low watermark is met there is no need for processes
2691 * to be throttled on pfmemalloc_wait as they should not be
2692 * able to safely make forward progress. Wake them
2693 */
2694 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2695 pfmemalloc_watermark_ok(pgdat))
2696 wake_up(&pgdat->pfmemalloc_wait);
2697
Mel Gormandc83edd2011-01-13 15:46:26 -08002698 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 break; /* kswapd: all done */
2700 /*
2701 * OK, kswapd is getting into trouble. Take a nap, then take
2702 * another pass across the zones.
2703 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002704 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002705 if (has_under_min_watermark_zone)
2706 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2707 else
2708 congestion_wait(BLK_RW_ASYNC, HZ/10);
2709 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710
2711 /*
2712 * We do this so kswapd doesn't build up large priorities for
2713 * example when it is freeing in parallel with allocators. It
2714 * matches the direct reclaim path behaviour in terms of impact
2715 * on zone->*_priority.
2716 */
Rik van Riela79311c2009-01-06 14:40:01 -08002717 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 break;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002719 } while (--sc.priority >= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720out:
Mel Gorman99504742011-01-13 15:46:20 -08002721
2722 /*
2723 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002724 * high-order: Balanced zones must make up at least 25% of the node
2725 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002726 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002727 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002729
2730 try_to_freeze();
2731
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002732 /*
2733 * Fragmentation may mean that the system cannot be
2734 * rebalanced for high-order allocations in all zones.
2735 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2736 * it means the zones have been fully scanned and are still
2737 * not balanced. For high-order allocations, there is
2738 * little point trying all over again as kswapd may
2739 * infinite loop.
2740 *
2741 * Instead, recheck all watermarks at order-0 as they
2742 * are the most important. If watermarks are ok, kswapd will go
2743 * back to sleep. High-order users can still perform direct
2744 * reclaim if they wish.
2745 */
2746 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2747 order = sc.order = 0;
2748
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 goto loop_again;
2750 }
2751
Mel Gorman99504742011-01-13 15:46:20 -08002752 /*
2753 * If kswapd was reclaiming at a higher order, it has the option of
2754 * sleeping without all zones being balanced. Before it does, it must
2755 * ensure that the watermarks for order-0 on *all* zones are met and
2756 * that the congestion flags are cleared. The congestion flag must
2757 * be cleared as kswapd is the only mechanism that clears the flag
2758 * and it is potentially going to sleep here.
2759 */
2760 if (order) {
Rik van Riel7be62de2012-03-21 16:33:52 -07002761 int zones_need_compaction = 1;
2762
Mel Gorman99504742011-01-13 15:46:20 -08002763 for (i = 0; i <= end_zone; i++) {
2764 struct zone *zone = pgdat->node_zones + i;
2765
2766 if (!populated_zone(zone))
2767 continue;
2768
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002769 if (zone->all_unreclaimable &&
2770 sc.priority != DEF_PRIORITY)
Mel Gorman99504742011-01-13 15:46:20 -08002771 continue;
2772
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002773 /* Would compaction fail due to lack of free memory? */
Rik van Riel496b9192012-03-24 10:26:21 -04002774 if (COMPACTION_BUILD &&
2775 compaction_suitable(zone, order) == COMPACT_SKIPPED)
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002776 goto loop_again;
2777
Mel Gorman99504742011-01-13 15:46:20 -08002778 /* Confirm the zone is balanced for order-0 */
2779 if (!zone_watermark_ok(zone, 0,
2780 high_wmark_pages(zone), 0, 0)) {
2781 order = sc.order = 0;
2782 goto loop_again;
2783 }
2784
Rik van Riel7be62de2012-03-21 16:33:52 -07002785 /* Check if the memory needs to be defragmented. */
2786 if (zone_watermark_ok(zone, order,
2787 low_wmark_pages(zone), *classzone_idx, 0))
2788 zones_need_compaction = 0;
2789
Mel Gorman99504742011-01-13 15:46:20 -08002790 /* If balanced, clear the congested flag */
2791 zone_clear_flag(zone, ZONE_CONGESTED);
2792 }
Rik van Riel7be62de2012-03-21 16:33:52 -07002793
2794 if (zones_need_compaction)
2795 compact_pgdat(pgdat, order);
Mel Gorman99504742011-01-13 15:46:20 -08002796 }
2797
Mel Gorman0abdee22011-01-13 15:46:22 -08002798 /*
Mel Gorman55150612012-07-31 16:44:35 -07002799 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08002800 * makes a decision on the order we were last reclaiming at. However,
2801 * if another caller entered the allocator slow path while kswapd
2802 * was awake, order will remain at the higher level
2803 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002804 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002805 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806}
2807
Mel Gormandc83edd2011-01-13 15:46:26 -08002808static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002809{
2810 long remaining = 0;
2811 DEFINE_WAIT(wait);
2812
2813 if (freezing(current) || kthread_should_stop())
2814 return;
2815
2816 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2817
2818 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07002819 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002820 remaining = schedule_timeout(HZ/10);
2821 finish_wait(&pgdat->kswapd_wait, &wait);
2822 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2823 }
2824
2825 /*
2826 * After a short sleep, check if it was a premature sleep. If not, then
2827 * go fully to sleep until explicitly woken up.
2828 */
Mel Gorman55150612012-07-31 16:44:35 -07002829 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002830 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2831
2832 /*
2833 * vmstat counters are not perfectly accurate and the estimated
2834 * value for counters such as NR_FREE_PAGES can deviate from the
2835 * true value by nr_online_cpus * threshold. To avoid the zone
2836 * watermarks being breached while under pressure, we reduce the
2837 * per-cpu vmstat threshold while kswapd is awake and restore
2838 * them before going back to sleep.
2839 */
2840 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07002841
2842 if (!kthread_should_stop())
2843 schedule();
2844
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002845 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2846 } else {
2847 if (remaining)
2848 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2849 else
2850 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2851 }
2852 finish_wait(&pgdat->kswapd_wait, &wait);
2853}
2854
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855/*
2856 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002857 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 *
2859 * This basically trickles out pages so that we have _some_
2860 * free memory available even if there is no other activity
2861 * that frees anything up. This is needed for things like routing
2862 * etc, where we otherwise might have all activity going on in
2863 * asynchronous contexts that cannot page things out.
2864 *
2865 * If there are applications that are active memory-allocators
2866 * (most normal use), this basically shouldn't matter.
2867 */
2868static int kswapd(void *p)
2869{
Mel Gorman215ddd62011-07-08 15:39:40 -07002870 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002871 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07002872 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002873 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 pg_data_t *pgdat = (pg_data_t*)p;
2875 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002876
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 struct reclaim_state reclaim_state = {
2878 .reclaimed_slab = 0,
2879 };
Rusty Russella70f7302009-03-13 14:49:46 +10302880 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881
Nick Piggincf40bd12009-01-21 08:12:39 +01002882 lockdep_set_current_reclaim_state(GFP_KERNEL);
2883
Rusty Russell174596a2009-01-01 10:12:29 +10302884 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002885 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 current->reclaim_state = &reclaim_state;
2887
2888 /*
2889 * Tell the memory management that we're a "memory allocator",
2890 * and that if we need more memory we should get access to it
2891 * regardless (see "__alloc_pages()"). "kswapd" should
2892 * never get caught in the normal page freeing logic.
2893 *
2894 * (Kswapd normally doesn't need memory anyway, but sometimes
2895 * you need a small amount of memory in order to be able to
2896 * page out something else, and this flag essentially protects
2897 * us from recursively trying to free more memory as we're
2898 * trying to free the first piece of memory in the first place).
2899 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002900 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002901 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902
Mel Gorman215ddd62011-07-08 15:39:40 -07002903 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002904 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002905 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002906 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08002908 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002909
Mel Gorman215ddd62011-07-08 15:39:40 -07002910 /*
2911 * If the last balance_pgdat was unsuccessful it's unlikely a
2912 * new request of a similar or harder type will succeed soon
2913 * so consider going to sleep on the basis we reclaimed at
2914 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002915 if (balanced_classzone_idx >= new_classzone_idx &&
2916 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07002917 new_order = pgdat->kswapd_max_order;
2918 new_classzone_idx = pgdat->classzone_idx;
2919 pgdat->kswapd_max_order = 0;
2920 pgdat->classzone_idx = pgdat->nr_zones - 1;
2921 }
2922
Mel Gorman99504742011-01-13 15:46:20 -08002923 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 /*
2925 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002926 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 */
2928 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002929 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002931 kswapd_try_to_sleep(pgdat, balanced_order,
2932 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002934 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07002935 new_order = order;
2936 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08002937 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002938 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940
David Rientjes8fe23e02009-12-14 17:58:33 -08002941 ret = try_to_freeze();
2942 if (kthread_should_stop())
2943 break;
2944
2945 /*
2946 * We can speed up thawing tasks if we don't call balance_pgdat
2947 * after returning from the refrigerator
2948 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002949 if (!ret) {
2950 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002951 balanced_classzone_idx = classzone_idx;
2952 balanced_order = balance_pgdat(pgdat, order,
2953 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 }
2956 return 0;
2957}
2958
2959/*
2960 * A zone is low on free memory, so wake its kswapd task to service it.
2961 */
Mel Gorman99504742011-01-13 15:46:20 -08002962void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963{
2964 pg_data_t *pgdat;
2965
Con Kolivasf3fe6512006-01-06 00:11:15 -08002966 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 return;
2968
Paul Jackson02a0e532006-12-13 00:34:25 -08002969 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002971 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08002972 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08002973 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08002974 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2975 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07002976 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08002978 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2979 return;
2980
2981 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07002982 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983}
2984
Wu Fengguangadea02a2009-09-21 17:01:42 -07002985/*
2986 * The reclaimable count would be mostly accurate.
2987 * The less reclaimable pages may be
2988 * - mlocked pages, which will be moved to unevictable list when encountered
2989 * - mapped pages, which may require several travels to be reclaimed
2990 * - dirty pages, which is not "instantly" reclaimable
2991 */
2992unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002993{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002994 int nr;
2995
2996 nr = global_page_state(NR_ACTIVE_FILE) +
2997 global_page_state(NR_INACTIVE_FILE);
2998
2999 if (nr_swap_pages > 0)
3000 nr += global_page_state(NR_ACTIVE_ANON) +
3001 global_page_state(NR_INACTIVE_ANON);
3002
3003 return nr;
3004}
3005
3006unsigned long zone_reclaimable_pages(struct zone *zone)
3007{
3008 int nr;
3009
3010 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3011 zone_page_state(zone, NR_INACTIVE_FILE);
3012
3013 if (nr_swap_pages > 0)
3014 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3015 zone_page_state(zone, NR_INACTIVE_ANON);
3016
3017 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003018}
3019
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003020#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003022 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003023 * freed pages.
3024 *
3025 * Rather than trying to age LRUs the aim is to preserve the overall
3026 * LRU order by reclaiming preferentially
3027 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003029unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003031 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003032 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003033 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3034 .may_swap = 1,
3035 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003036 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003037 .nr_to_reclaim = nr_to_reclaim,
3038 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003039 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003040 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003042 struct shrink_control shrink = {
3043 .gfp_mask = sc.gfp_mask,
3044 };
3045 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003046 struct task_struct *p = current;
3047 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003049 p->flags |= PF_MEMALLOC;
3050 lockdep_set_current_reclaim_state(sc.gfp_mask);
3051 reclaim_state.reclaimed_slab = 0;
3052 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003053
Ying Hana09ed5e2011-05-24 17:12:26 -07003054 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003055
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003056 p->reclaim_state = NULL;
3057 lockdep_clear_current_reclaim_state();
3058 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003059
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003060 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003062#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064/* It's optimal to keep kswapds on the same CPUs as their memory, but
3065 not required for correctness. So if the last cpu in a node goes
3066 away, we get changed to run anywhere: as the first one comes back,
3067 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003068static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08003069 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003071 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003073 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003074 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003075 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303076 const struct cpumask *mask;
3077
3078 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003079
Rusty Russell3e597942009-01-01 10:12:24 +10303080 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003082 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 }
3084 }
3085 return NOTIFY_OK;
3086}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087
Yasunori Goto3218ae12006-06-27 02:53:33 -07003088/*
3089 * This kswapd start function will be called by init and node-hot-add.
3090 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3091 */
3092int kswapd_run(int nid)
3093{
3094 pg_data_t *pgdat = NODE_DATA(nid);
3095 int ret = 0;
3096
3097 if (pgdat->kswapd)
3098 return 0;
3099
3100 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3101 if (IS_ERR(pgdat->kswapd)) {
3102 /* failure at boot is fatal */
3103 BUG_ON(system_state == SYSTEM_BOOTING);
3104 printk("Failed to start kswapd on node %d\n",nid);
Wen Congyang18b48d52012-09-17 14:08:55 -07003105 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003106 ret = -1;
3107 }
3108 return ret;
3109}
3110
David Rientjes8fe23e02009-12-14 17:58:33 -08003111/*
Jiang Liud8adde12012-07-11 14:01:52 -07003112 * Called by memory hotplug when all memory in a node is offlined. Caller must
3113 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003114 */
3115void kswapd_stop(int nid)
3116{
3117 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3118
Jiang Liud8adde12012-07-11 14:01:52 -07003119 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003120 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003121 NODE_DATA(nid)->kswapd = NULL;
3122 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003123}
3124
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125static int __init kswapd_init(void)
3126{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003127 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003128
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003130 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003131 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 hotcpu_notifier(cpu_callback, 0);
3133 return 0;
3134}
3135
3136module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003137
3138#ifdef CONFIG_NUMA
3139/*
3140 * Zone reclaim mode
3141 *
3142 * If non-zero call zone_reclaim when the number of free pages falls below
3143 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003144 */
3145int zone_reclaim_mode __read_mostly;
3146
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003147#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003148#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003149#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3150#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3151
Christoph Lameter9eeff232006-01-18 17:42:31 -08003152/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003153 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3154 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3155 * a zone.
3156 */
3157#define ZONE_RECLAIM_PRIORITY 4
3158
Christoph Lameter9eeff232006-01-18 17:42:31 -08003159/*
Christoph Lameter96146342006-07-03 00:24:13 -07003160 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3161 * occur.
3162 */
3163int sysctl_min_unmapped_ratio = 1;
3164
3165/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003166 * If the number of slab pages in a zone grows beyond this percentage then
3167 * slab reclaim needs to occur.
3168 */
3169int sysctl_min_slab_ratio = 5;
3170
Mel Gorman90afa5d2009-06-16 15:33:20 -07003171static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3172{
3173 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3174 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3175 zone_page_state(zone, NR_ACTIVE_FILE);
3176
3177 /*
3178 * It's possible for there to be more file mapped pages than
3179 * accounted for by the pages on the file LRU lists because
3180 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3181 */
3182 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3183}
3184
3185/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3186static long zone_pagecache_reclaimable(struct zone *zone)
3187{
3188 long nr_pagecache_reclaimable;
3189 long delta = 0;
3190
3191 /*
3192 * If RECLAIM_SWAP is set, then all file pages are considered
3193 * potentially reclaimable. Otherwise, we have to worry about
3194 * pages like swapcache and zone_unmapped_file_pages() provides
3195 * a better estimate
3196 */
3197 if (zone_reclaim_mode & RECLAIM_SWAP)
3198 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3199 else
3200 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3201
3202 /* If we can't clean pages, remove dirty pages from consideration */
3203 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3204 delta += zone_page_state(zone, NR_FILE_DIRTY);
3205
3206 /* Watch for any possible underflows due to delta */
3207 if (unlikely(delta > nr_pagecache_reclaimable))
3208 delta = nr_pagecache_reclaimable;
3209
3210 return nr_pagecache_reclaimable - delta;
3211}
3212
Christoph Lameter0ff38492006-09-25 23:31:52 -07003213/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003214 * Try to free up some pages from this zone through reclaim.
3215 */
Andrew Morton179e9632006-03-22 00:08:18 -08003216static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003217{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003218 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003219 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003220 struct task_struct *p = current;
3221 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003222 struct scan_control sc = {
3223 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003224 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003225 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003226 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3227 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003228 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003229 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003230 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003231 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003232 struct shrink_control shrink = {
3233 .gfp_mask = sc.gfp_mask,
3234 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003235 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003236
Christoph Lameter9eeff232006-01-18 17:42:31 -08003237 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003238 /*
3239 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3240 * and we also need to be able to write out pages for RECLAIM_WRITE
3241 * and RECLAIM_SWAP.
3242 */
3243 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003244 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003245 reclaim_state.reclaimed_slab = 0;
3246 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003247
Mel Gorman90afa5d2009-06-16 15:33:20 -07003248 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003249 /*
3250 * Free memory by calling shrink zone with increasing
3251 * priorities until we have enough memory freed.
3252 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003253 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003254 shrink_zone(zone, &sc);
3255 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003256 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003257
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003258 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3259 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003260 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003261 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003262 * many pages were freed in this zone. So we take the current
3263 * number of slab pages and shake the slab until it is reduced
3264 * by the same nr_pages that we used for reclaiming unmapped
3265 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003266 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003267 * Note that shrink_slab will free memory on all zones and may
3268 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003269 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003270 for (;;) {
3271 unsigned long lru_pages = zone_reclaimable_pages(zone);
3272
3273 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003274 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003275 break;
3276
3277 /* Freed enough memory */
3278 nr_slab_pages1 = zone_page_state(zone,
3279 NR_SLAB_RECLAIMABLE);
3280 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3281 break;
3282 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003283
3284 /*
3285 * Update nr_reclaimed by the number of slab pages we
3286 * reclaimed from this zone.
3287 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003288 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3289 if (nr_slab_pages1 < nr_slab_pages0)
3290 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003291 }
3292
Christoph Lameter9eeff232006-01-18 17:42:31 -08003293 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003294 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003295 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003296 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003297}
Andrew Morton179e9632006-03-22 00:08:18 -08003298
3299int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3300{
Andrew Morton179e9632006-03-22 00:08:18 -08003301 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003302 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003303
3304 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003305 * Zone reclaim reclaims unmapped file backed pages and
3306 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003307 *
Christoph Lameter96146342006-07-03 00:24:13 -07003308 * A small portion of unmapped file backed pages is needed for
3309 * file I/O otherwise pages read by file I/O will be immediately
3310 * thrown out if the zone is overallocated. So we do not reclaim
3311 * if less than a specified percentage of the zone is used by
3312 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003313 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003314 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3315 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003316 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003317
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003318 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003319 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003320
Andrew Morton179e9632006-03-22 00:08:18 -08003321 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003322 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003323 */
David Rientjesd773ed62007-10-16 23:26:01 -07003324 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003325 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003326
3327 /*
3328 * Only run zone reclaim on the local zone or on zones that do not
3329 * have associated processors. This will favor the local processor
3330 * over remote processors and spread off node memory allocations
3331 * as wide as possible.
3332 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003333 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003334 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003335 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003336
3337 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003338 return ZONE_RECLAIM_NOSCAN;
3339
David Rientjesd773ed62007-10-16 23:26:01 -07003340 ret = __zone_reclaim(zone, gfp_mask, order);
3341 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3342
Mel Gorman24cf725182009-06-16 15:33:23 -07003343 if (!ret)
3344 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3345
David Rientjesd773ed62007-10-16 23:26:01 -07003346 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003347}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003348#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003349
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003350/*
3351 * page_evictable - test whether a page is evictable
3352 * @page: the page to test
3353 * @vma: the VMA in which the page is or will be mapped, may be NULL
3354 *
3355 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07003356 * lists vs unevictable list. The vma argument is !NULL when called from the
3357 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003358 *
3359 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003360 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003361 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003362 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003363 */
3364int page_evictable(struct page *page, struct vm_area_struct *vma)
3365{
3366
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003367 if (mapping_unevictable(page_mapping(page)))
3368 return 0;
3369
Ying Han096a7cf2012-05-29 15:06:25 -07003370 if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
Nick Pigginb291f002008-10-18 20:26:44 -07003371 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003372
3373 return 1;
3374}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003375
Hugh Dickins850465792012-01-20 14:34:19 -08003376#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003377/**
Hugh Dickins24513262012-01-20 14:34:21 -08003378 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3379 * @pages: array of pages to check
3380 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003381 *
Hugh Dickins24513262012-01-20 14:34:21 -08003382 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003383 *
3384 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003385 */
Hugh Dickins24513262012-01-20 14:34:21 -08003386void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003387{
Johannes Weiner925b7672012-01-12 17:18:15 -08003388 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003389 struct zone *zone = NULL;
3390 int pgscanned = 0;
3391 int pgrescued = 0;
3392 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003393
Hugh Dickins24513262012-01-20 14:34:21 -08003394 for (i = 0; i < nr_pages; i++) {
3395 struct page *page = pages[i];
3396 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003397
Hugh Dickins24513262012-01-20 14:34:21 -08003398 pgscanned++;
3399 pagezone = page_zone(page);
3400 if (pagezone != zone) {
3401 if (zone)
3402 spin_unlock_irq(&zone->lru_lock);
3403 zone = pagezone;
3404 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003405 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003406 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003407
Hugh Dickins24513262012-01-20 14:34:21 -08003408 if (!PageLRU(page) || !PageUnevictable(page))
3409 continue;
3410
3411 if (page_evictable(page, NULL)) {
3412 enum lru_list lru = page_lru_base_type(page);
3413
3414 VM_BUG_ON(PageActive(page));
3415 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003416 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3417 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003418 pgrescued++;
3419 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003420 }
Hugh Dickins24513262012-01-20 14:34:21 -08003421
3422 if (zone) {
3423 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3424 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3425 spin_unlock_irq(&zone->lru_lock);
3426 }
Hugh Dickins850465792012-01-20 14:34:19 -08003427}
3428#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003429
Johannes Weiner264e56d2011-10-31 17:09:13 -07003430static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003431{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003432 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003433 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003434 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003435 "one, please send an email to linux-mm@kvack.org.\n",
3436 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003437}
3438
3439/*
3440 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3441 * all nodes' unevictable lists for evictable pages
3442 */
3443unsigned long scan_unevictable_pages;
3444
3445int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003446 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003447 size_t *length, loff_t *ppos)
3448{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003449 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003450 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003451 scan_unevictable_pages = 0;
3452 return 0;
3453}
3454
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003455#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003456/*
3457 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3458 * a specified node's per zone unevictable lists for evictable pages.
3459 */
3460
Kay Sievers10fbcf42011-12-21 14:48:43 -08003461static ssize_t read_scan_unevictable_node(struct device *dev,
3462 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003463 char *buf)
3464{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003465 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003466 return sprintf(buf, "0\n"); /* always zero; should fit... */
3467}
3468
Kay Sievers10fbcf42011-12-21 14:48:43 -08003469static ssize_t write_scan_unevictable_node(struct device *dev,
3470 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003471 const char *buf, size_t count)
3472{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003473 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003474 return 1;
3475}
3476
3477
Kay Sievers10fbcf42011-12-21 14:48:43 -08003478static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003479 read_scan_unevictable_node,
3480 write_scan_unevictable_node);
3481
3482int scan_unevictable_register_node(struct node *node)
3483{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003484 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003485}
3486
3487void scan_unevictable_unregister_node(struct node *node)
3488{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003489 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003490}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003491#endif