blob: 8b055e9379bc23105dea4e58d6bebe2600756a4b [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
Hugh Dickins39b5f292012-10-08 16:33:18 -0700556 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700557 /*
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 */
Hugh Dickins39b5f292012-10-08 16:33:18 -0700590 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700591 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,
Minchan Kim02c6de82012-10-08 16:31:55 -0700677 enum ttu_flags ttu_flags,
Mel Gorman92df3a72011-10-31 17:07:56 -0700678 unsigned long *ret_nr_dirty,
Minchan Kim02c6de82012-10-08 16:31:55 -0700679 unsigned long *ret_nr_writeback,
680 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681{
682 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700683 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700685 unsigned long nr_dirty = 0;
686 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800687 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700688 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
690 cond_resched();
691
Tim Chen69980e32012-07-31 16:46:08 -0700692 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 while (!list_empty(page_list)) {
694 struct address_space *mapping;
695 struct page *page;
696 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700697 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699 cond_resched();
700
701 page = lru_to_page(page_list);
702 list_del(&page->lru);
703
Nick Piggin529ae9a2008-08-02 12:01:03 +0200704 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 goto keep;
706
Nick Piggin725d7042006-09-25 23:30:55 -0700707 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700708 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800711
Hugh Dickins39b5f292012-10-08 16:33:18 -0700712 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700713 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700714
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700715 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800716 goto keep_locked;
717
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 /* Double the slab pressure for mapped and swapcache pages */
719 if (page_mapped(page) || PageSwapCache(page))
720 sc->nr_scanned++;
721
Andy Whitcroftc661b072007-08-22 14:01:26 -0700722 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
723 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
724
725 if (PageWriteback(page)) {
Michal Hockoe62e3842012-07-31 16:45:55 -0700726 /*
727 * memcg doesn't have any dirty pages throttling so we
728 * could easily OOM just because too many pages are in
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700729 * writeback and there is nothing else to reclaim.
Michal Hockoe62e3842012-07-31 16:45:55 -0700730 *
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700731 * Check __GFP_IO, certainly because a loop driver
Michal Hockoe62e3842012-07-31 16:45:55 -0700732 * thread might enter reclaim, and deadlock if it waits
733 * on a page for which it is needed to do the write
734 * (loop masks off __GFP_IO|__GFP_FS for this reason);
735 * but more thought would probably show more reasons.
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700736 *
737 * Don't require __GFP_FS, since we're not going into
738 * the FS, just waiting on its writeback completion.
739 * Worryingly, ext4 gfs2 and xfs allocate pages with
740 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
741 * testing may_enter_fs here is liable to OOM on them.
Michal Hockoe62e3842012-07-31 16:45:55 -0700742 */
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700743 if (global_reclaim(sc) ||
744 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
745 /*
746 * This is slightly racy - end_page_writeback()
747 * might have just cleared PageReclaim, then
748 * setting PageReclaim here end up interpreted
749 * as PageReadahead - but that does not matter
750 * enough to care. What we do want is for this
751 * page to have PageReclaim set next time memcg
752 * reclaim reaches the tests above, so it will
753 * then wait_on_page_writeback() to avoid OOM;
754 * and it's also appropriate in global reclaim.
755 */
756 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700757 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700758 goto keep_locked;
Michal Hockoe62e3842012-07-31 16:45:55 -0700759 }
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700760 wait_on_page_writeback(page);
Andy Whitcroftc661b072007-08-22 14:01:26 -0700761 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
Minchan Kim02c6de82012-10-08 16:31:55 -0700763 if (!force_reclaim)
764 references = page_check_references(page, sc);
765
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800766 switch (references) {
767 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800769 case PAGEREF_KEEP:
770 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800771 case PAGEREF_RECLAIM:
772 case PAGEREF_RECLAIM_CLEAN:
773 ; /* try to reclaim the page below */
774 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 /*
777 * Anonymous process memory has backing store?
778 * Try to allocate it some swap space here.
779 */
Nick Pigginb291f002008-10-18 20:26:44 -0700780 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800781 if (!(sc->gfp_mask & __GFP_IO))
782 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800783 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800785 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700786 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787
788 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
790 /*
791 * The page is mapped into the page tables of one or more
792 * processes. Try to unmap it here.
793 */
794 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700795 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 case SWAP_FAIL:
797 goto activate_locked;
798 case SWAP_AGAIN:
799 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700800 case SWAP_MLOCK:
801 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 case SWAP_SUCCESS:
803 ; /* try to free the page below */
804 }
805 }
806
807 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700808 nr_dirty++;
809
Mel Gormanee728862011-10-31 17:07:38 -0700810 /*
811 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700812 * avoid risk of stack overflow but do not writeback
813 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700814 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700815 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700816 (!current_is_kswapd() ||
817 sc->priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700818 /*
819 * Immediately reclaim when written back.
820 * Similar in principal to deactivate_page()
821 * except we already have the page isolated
822 * and know it's dirty
823 */
824 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
825 SetPageReclaim(page);
826
Mel Gormanee728862011-10-31 17:07:38 -0700827 goto keep_locked;
828 }
829
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800830 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700832 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800834 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 goto keep_locked;
836
837 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700838 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700840 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 goto keep_locked;
842 case PAGE_ACTIVATE:
843 goto activate_locked;
844 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700845 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700846 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700847 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700849
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 /*
851 * A synchronous write - probably a ramdisk. Go
852 * ahead and try to reclaim the page.
853 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200854 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 goto keep;
856 if (PageDirty(page) || PageWriteback(page))
857 goto keep_locked;
858 mapping = page_mapping(page);
859 case PAGE_CLEAN:
860 ; /* try to free the page below */
861 }
862 }
863
864 /*
865 * If the page has buffers, try to free the buffer mappings
866 * associated with this page. If we succeed we try to free
867 * the page as well.
868 *
869 * We do this even if the page is PageDirty().
870 * try_to_release_page() does not perform I/O, but it is
871 * possible for a page to have PageDirty set, but it is actually
872 * clean (all its buffers are clean). This happens if the
873 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700874 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 * try_to_release_page() will discover that cleanness and will
876 * drop the buffers and mark the page clean - it can be freed.
877 *
878 * Rarely, pages can have buffers and no ->mapping. These are
879 * the pages which were not successfully invalidated in
880 * truncate_complete_page(). We try to drop those buffers here
881 * and if that worked, and the page is no longer mapped into
882 * process address space (page_count == 1) it can be freed.
883 * Otherwise, leave the page on the LRU so it is swappable.
884 */
David Howells266cf652009-04-03 16:42:36 +0100885 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 if (!try_to_release_page(page, sc->gfp_mask))
887 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700888 if (!mapping && page_count(page) == 1) {
889 unlock_page(page);
890 if (put_page_testzero(page))
891 goto free_it;
892 else {
893 /*
894 * rare race with speculative reference.
895 * the speculative reference will free
896 * this page shortly, so we may
897 * increment nr_reclaimed here (and
898 * leave it off the LRU).
899 */
900 nr_reclaimed++;
901 continue;
902 }
903 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 }
905
Nick Piggine2867812008-07-25 19:45:30 -0700906 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800907 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908
Nick Piggina978d6f2008-10-18 20:26:58 -0700909 /*
910 * At this point, we have no other references and there is
911 * no way to pick any more up (removed from LRU, removed
912 * from pagecache). Can use non-atomic bitops now (and
913 * we obviously don't have to worry about waking up a process
914 * waiting on the page lock, because there are no references.
915 */
916 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700917free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800918 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700919
920 /*
921 * Is there need to periodically free_page_list? It would
922 * appear not as the counts should be low
923 */
924 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 continue;
926
Nick Pigginb291f002008-10-18 20:26:44 -0700927cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800928 if (PageSwapCache(page))
929 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700930 unlock_page(page);
931 putback_lru_page(page);
932 continue;
933
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700935 /* Not a candidate for swapping, so reclaim swap space. */
936 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800937 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700938 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 SetPageActive(page);
940 pgactivate++;
941keep_locked:
942 unlock_page(page);
943keep:
944 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700945 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700947
Mel Gorman0e093d992010-10-26 14:21:45 -0700948 /*
949 * Tag a zone as congested if all the dirty pages encountered were
950 * backed by a congested BDI. In this case, reclaimers should just
951 * back off and wait for congestion to clear because further reclaim
952 * will encounter the same problem
953 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800954 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700955 zone_set_flag(zone, ZONE_CONGESTED);
Mel Gorman0e093d992010-10-26 14:21:45 -0700956
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -0800957 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700958
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700960 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -0700961 mem_cgroup_uncharge_end();
Mel Gorman92df3a72011-10-31 17:07:56 -0700962 *ret_nr_dirty += nr_dirty;
963 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -0800964 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965}
966
Minchan Kim02c6de82012-10-08 16:31:55 -0700967unsigned long reclaim_clean_pages_from_list(struct zone *zone,
968 struct list_head *page_list)
969{
970 struct scan_control sc = {
971 .gfp_mask = GFP_KERNEL,
972 .priority = DEF_PRIORITY,
973 .may_unmap = 1,
974 };
975 unsigned long ret, dummy1, dummy2;
976 struct page *page, *next;
977 LIST_HEAD(clean_pages);
978
979 list_for_each_entry_safe(page, next, page_list, lru) {
980 if (page_is_file_cache(page) && !PageDirty(page)) {
981 ClearPageActive(page);
982 list_move(&page->lru, &clean_pages);
983 }
984 }
985
986 ret = shrink_page_list(&clean_pages, zone, &sc,
987 TTU_UNMAP|TTU_IGNORE_ACCESS,
988 &dummy1, &dummy2, true);
989 list_splice(&clean_pages, page_list);
990 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
991 return ret;
992}
993
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700994/*
995 * Attempt to remove the specified page from its LRU. Only take this page
996 * if it is of the appropriate PageActive status. Pages which are being
997 * freed elsewhere are also ignored.
998 *
999 * page: page to consider
1000 * mode: one of the LRU isolation modes defined above
1001 *
1002 * returns 0 on success, -ve errno on failure.
1003 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001004int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001005{
1006 int ret = -EINVAL;
1007
1008 /* Only take pages on the LRU. */
1009 if (!PageLRU(page))
1010 return ret;
1011
Minchan Kime46a2872012-10-08 16:33:48 -07001012 /* Compaction should not handle unevictable pages but CMA can do so */
1013 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001014 return ret;
1015
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001016 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001017
Mel Gormanc8244932012-01-12 17:19:38 -08001018 /*
1019 * To minimise LRU disruption, the caller can indicate that it only
1020 * wants to isolate pages it will be able to operate on without
1021 * blocking - clean pages for the most part.
1022 *
1023 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1024 * is used by reclaim when it is cannot write to backing storage
1025 *
1026 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1027 * that it is possible to migrate without blocking
1028 */
1029 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1030 /* All the caller can do on PageWriteback is block */
1031 if (PageWriteback(page))
1032 return ret;
1033
1034 if (PageDirty(page)) {
1035 struct address_space *mapping;
1036
1037 /* ISOLATE_CLEAN means only clean pages */
1038 if (mode & ISOLATE_CLEAN)
1039 return ret;
1040
1041 /*
1042 * Only pages without mappings or that have a
1043 * ->migratepage callback are possible to migrate
1044 * without blocking
1045 */
1046 mapping = page_mapping(page);
1047 if (mapping && !mapping->a_ops->migratepage)
1048 return ret;
1049 }
1050 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001051
Minchan Kimf80c0672011-10-31 17:06:55 -07001052 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1053 return ret;
1054
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001055 if (likely(get_page_unless_zero(page))) {
1056 /*
1057 * Be careful not to clear PageLRU until after we're
1058 * sure the page is not being freed elsewhere -- the
1059 * page release code relies on it.
1060 */
1061 ClearPageLRU(page);
1062 ret = 0;
1063 }
1064
1065 return ret;
1066}
1067
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001068/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 * zone->lru_lock is heavily contended. Some of the functions that
1070 * shrink the lists perform better by taking out a batch of pages
1071 * and working on them outside the LRU lock.
1072 *
1073 * For pagecache intensive workloads, this function is the hottest
1074 * spot in the kernel (apart from copy_*_user functions).
1075 *
1076 * Appropriate locks must be held before calling this function.
1077 *
1078 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001079 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001081 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001082 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001083 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001084 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 *
1086 * returns how many pages were moved onto *@dst.
1087 */
Andrew Morton69e05942006-03-22 00:08:19 -08001088static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001089 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001090 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001091 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001093 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001094 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001095 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001097 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001098 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001099 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001100
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 page = lru_to_page(src);
1102 prefetchw_prev_lru_page(page, src, flags);
1103
Nick Piggin725d7042006-09-25 23:30:55 -07001104 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001105
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001106 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001107 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001108 nr_pages = hpage_nr_pages(page);
1109 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001110 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001111 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001112 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001113
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001114 case -EBUSY:
1115 /* else it is being freed elsewhere */
1116 list_move(&page->lru, src);
1117 continue;
1118
1119 default:
1120 BUG();
1121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 }
1123
Hugh Dickinsf6260122012-01-12 17:20:06 -08001124 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001125 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1126 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 return nr_taken;
1128}
1129
Nick Piggin62695a82008-10-18 20:26:09 -07001130/**
1131 * isolate_lru_page - tries to isolate a page from its LRU list
1132 * @page: page to isolate from its LRU list
1133 *
1134 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1135 * vmstat statistic corresponding to whatever LRU list the page was on.
1136 *
1137 * Returns 0 if the page was removed from an LRU list.
1138 * Returns -EBUSY if the page was not on an LRU list.
1139 *
1140 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001141 * the active list, it will have PageActive set. If it was found on
1142 * the unevictable list, it will have the PageUnevictable bit set. That flag
1143 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001144 *
1145 * The vmstat statistic corresponding to the list on which the page was
1146 * found will be decremented.
1147 *
1148 * Restrictions:
1149 * (1) Must be called with an elevated refcount on the page. This is a
1150 * fundamentnal difference from isolate_lru_pages (which is called
1151 * without a stable reference).
1152 * (2) the lru_lock must not be held.
1153 * (3) interrupts must be enabled.
1154 */
1155int isolate_lru_page(struct page *page)
1156{
1157 int ret = -EBUSY;
1158
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001159 VM_BUG_ON(!page_count(page));
1160
Nick Piggin62695a82008-10-18 20:26:09 -07001161 if (PageLRU(page)) {
1162 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001163 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001164
1165 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001166 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001167 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001168 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001169 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001170 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001171 del_page_from_lru_list(page, lruvec, lru);
1172 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001173 }
1174 spin_unlock_irq(&zone->lru_lock);
1175 }
1176 return ret;
1177}
1178
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001179/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001180 * Are there way too many processes in the direct reclaim path already?
1181 */
1182static int too_many_isolated(struct zone *zone, int file,
1183 struct scan_control *sc)
1184{
1185 unsigned long inactive, isolated;
1186
1187 if (current_is_kswapd())
1188 return 0;
1189
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001190 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001191 return 0;
1192
1193 if (file) {
1194 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1195 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1196 } else {
1197 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1198 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1199 }
1200
1201 return isolated > inactive;
1202}
1203
Mel Gorman66635622010-08-09 17:19:30 -07001204static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001205putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001206{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001207 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1208 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001209 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001210
Mel Gorman66635622010-08-09 17:19:30 -07001211 /*
1212 * Put back any unfreeable pages.
1213 */
Mel Gorman66635622010-08-09 17:19:30 -07001214 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001215 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001216 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001217
Mel Gorman66635622010-08-09 17:19:30 -07001218 VM_BUG_ON(PageLRU(page));
1219 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001220 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001221 spin_unlock_irq(&zone->lru_lock);
1222 putback_lru_page(page);
1223 spin_lock_irq(&zone->lru_lock);
1224 continue;
1225 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001226
1227 lruvec = mem_cgroup_page_lruvec(page, zone);
1228
Linus Torvalds7a608572011-01-17 14:42:19 -08001229 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001230 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001231 add_page_to_lru_list(page, lruvec, lru);
1232
Mel Gorman66635622010-08-09 17:19:30 -07001233 if (is_active_lru(lru)) {
1234 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001235 int numpages = hpage_nr_pages(page);
1236 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001237 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001238 if (put_page_testzero(page)) {
1239 __ClearPageLRU(page);
1240 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001241 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001242
1243 if (unlikely(PageCompound(page))) {
1244 spin_unlock_irq(&zone->lru_lock);
1245 (*get_compound_page_dtor(page))(page);
1246 spin_lock_irq(&zone->lru_lock);
1247 } else
1248 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001249 }
1250 }
Mel Gorman66635622010-08-09 17:19:30 -07001251
Hugh Dickins3f797682012-01-12 17:20:07 -08001252 /*
1253 * To save our caller's stack, now use input list for pages to free.
1254 */
1255 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001256}
1257
1258/*
Andrew Morton1742f192006-03-22 00:08:21 -08001259 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1260 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 */
Mel Gorman66635622010-08-09 17:19:30 -07001262static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001263shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001264 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265{
1266 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001267 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001268 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001269 unsigned long nr_taken;
Mel Gorman92df3a72011-10-31 17:07:56 -07001270 unsigned long nr_dirty = 0;
1271 unsigned long nr_writeback = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001272 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001273 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001274 struct zone *zone = lruvec_zone(lruvec);
1275 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001276
Rik van Riel35cd7812009-09-21 17:01:38 -07001277 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001278 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001279
1280 /* We are about to die and free our memory. Return now. */
1281 if (fatal_signal_pending(current))
1282 return SWAP_CLUSTER_MAX;
1283 }
1284
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001286
1287 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001288 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001289 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001290 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001291
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001294 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1295 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001296
1297 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1298 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1299
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001300 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001301 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001302 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001303 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001304 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001305 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001306 }
Hillf Dantond563c052012-03-21 16:34:02 -07001307 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001308
Hillf Dantond563c052012-03-21 16:34:02 -07001309 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001310 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001311
Minchan Kim02c6de82012-10-08 16:31:55 -07001312 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1313 &nr_dirty, &nr_writeback, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001314
Hugh Dickins3f797682012-01-12 17:20:07 -08001315 spin_lock_irq(&zone->lru_lock);
1316
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001317 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001318
Ying Han904249a2012-04-25 16:01:48 -07001319 if (global_reclaim(sc)) {
1320 if (current_is_kswapd())
1321 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1322 nr_reclaimed);
1323 else
1324 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1325 nr_reclaimed);
1326 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001327
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001328 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001329
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001330 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001331
1332 spin_unlock_irq(&zone->lru_lock);
1333
1334 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001335
Mel Gorman92df3a72011-10-31 17:07:56 -07001336 /*
1337 * If reclaim is isolating dirty pages under writeback, it implies
1338 * that the long-lived page allocation rate is exceeding the page
1339 * laundering rate. Either the global limits are not being effective
1340 * at throttling processes due to the page distribution throughout
1341 * zones or there is heavy usage of a slow backing device. The
1342 * only option is to throttle from reclaim context which is not ideal
1343 * as there is no guarantee the dirtying process is throttled in the
1344 * same way balance_dirty_pages() manages.
1345 *
1346 * This scales the number of dirty pages that must be under writeback
1347 * before throttling depending on priority. It is a simple backoff
1348 * function that has the most effect in the range DEF_PRIORITY to
1349 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1350 * in trouble and reclaim is considered to be in trouble.
1351 *
1352 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1353 * DEF_PRIORITY-1 50% must be PageWriteback
1354 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1355 * ...
1356 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1357 * isolated page is PageWriteback
1358 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001359 if (nr_writeback && nr_writeback >=
1360 (nr_taken >> (DEF_PRIORITY - sc->priority)))
Mel Gorman92df3a72011-10-31 17:07:56 -07001361 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1362
Mel Gormane11da5b2010-10-26 14:21:40 -07001363 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1364 zone_idx(zone),
1365 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001366 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001367 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001368 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369}
1370
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001371/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 * This moves pages from the active list to the inactive list.
1373 *
1374 * We move them the other way if the page is referenced by one or more
1375 * processes, from rmap.
1376 *
1377 * If the pages are mostly unmapped, the processing is fast and it is
1378 * appropriate to hold zone->lru_lock across the whole operation. But if
1379 * the pages are mapped, the processing is slow (page_referenced()) so we
1380 * should drop zone->lru_lock around each page. It's impossible to balance
1381 * this, so instead we remove the pages from the LRU while processing them.
1382 * It is safe to rely on PG_active against the non-LRU pages in here because
1383 * nobody will play with that bit on a non-LRU page.
1384 *
1385 * The downside is that we have to touch page->_count against each page.
1386 * But we had to alter page->flags anyway.
1387 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001388
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001389static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001390 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001391 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001392 enum lru_list lru)
1393{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001394 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001395 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001396 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001397 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001398
Wu Fengguang3eb41402009-06-16 15:33:13 -07001399 while (!list_empty(list)) {
1400 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001401 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001402
1403 VM_BUG_ON(PageLRU(page));
1404 SetPageLRU(page);
1405
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001406 nr_pages = hpage_nr_pages(page);
1407 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001408 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001409 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001410
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001411 if (put_page_testzero(page)) {
1412 __ClearPageLRU(page);
1413 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001414 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001415
1416 if (unlikely(PageCompound(page))) {
1417 spin_unlock_irq(&zone->lru_lock);
1418 (*get_compound_page_dtor(page))(page);
1419 spin_lock_irq(&zone->lru_lock);
1420 } else
1421 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001422 }
1423 }
1424 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1425 if (!is_active_lru(lru))
1426 __count_vm_events(PGDEACTIVATE, pgmoved);
1427}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001428
Hugh Dickinsf6260122012-01-12 17:20:06 -08001429static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001430 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001431 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001432 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001434 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001435 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001436 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001438 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001439 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001441 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001442 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001443 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001444 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001445 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
1447 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001448
1449 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001450 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001451 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001452 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001453
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001455
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001456 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1457 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001458 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001459 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001460
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001461 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001462
Hugh Dickinsf6260122012-01-12 17:20:06 -08001463 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001464 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001465 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 spin_unlock_irq(&zone->lru_lock);
1467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 while (!list_empty(&l_hold)) {
1469 cond_resched();
1470 page = lru_to_page(&l_hold);
1471 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001472
Hugh Dickins39b5f292012-10-08 16:33:18 -07001473 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001474 putback_lru_page(page);
1475 continue;
1476 }
1477
Mel Gormancc715d92012-03-21 16:34:00 -07001478 if (unlikely(buffer_heads_over_limit)) {
1479 if (page_has_private(page) && trylock_page(page)) {
1480 if (page_has_private(page))
1481 try_to_release_page(page, 0);
1482 unlock_page(page);
1483 }
1484 }
1485
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001486 if (page_referenced(page, 0, sc->target_mem_cgroup,
1487 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001488 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001489 /*
1490 * Identify referenced, file-backed active pages and
1491 * give them one more trip around the active list. So
1492 * that executable code get better chances to stay in
1493 * memory under moderate memory pressure. Anon pages
1494 * are not likely to be evicted by use-once streaming
1495 * IO, plus JVM can create lots of anon VM_EXEC pages,
1496 * so we ignore them here.
1497 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001498 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001499 list_add(&page->lru, &l_active);
1500 continue;
1501 }
1502 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001503
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001504 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 list_add(&page->lru, &l_inactive);
1506 }
1507
Andrew Mortonb5557492009-01-06 14:40:13 -08001508 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001509 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001510 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001511 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001512 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001513 * Count referenced pages from currently used mappings as rotated,
1514 * even though only some of them are actually re-activated. This
1515 * helps balance scan pressure between file and anonymous pages in
1516 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001517 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001518 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001519
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001520 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1521 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001522 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001523 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001524
1525 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526}
1527
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001528#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001529static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001530{
1531 unsigned long active, inactive;
1532
1533 active = zone_page_state(zone, NR_ACTIVE_ANON);
1534 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1535
1536 if (inactive * zone->inactive_ratio < active)
1537 return 1;
1538
1539 return 0;
1540}
1541
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001542/**
1543 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001544 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001545 *
1546 * Returns true if the zone does not have enough inactive anon pages,
1547 * meaning some active anon pages need to be deactivated.
1548 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001549static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001550{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001551 /*
1552 * If we don't have swap space, anonymous page deactivation
1553 * is pointless.
1554 */
1555 if (!total_swap_pages)
1556 return 0;
1557
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001558 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001559 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001560
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001561 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001562}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001563#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001564static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001565{
1566 return 0;
1567}
1568#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001569
Rik van Riel56e49d22009-06-16 15:32:28 -07001570static int inactive_file_is_low_global(struct zone *zone)
1571{
1572 unsigned long active, inactive;
1573
1574 active = zone_page_state(zone, NR_ACTIVE_FILE);
1575 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1576
1577 return (active > inactive);
1578}
1579
1580/**
1581 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001582 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001583 *
1584 * When the system is doing streaming IO, memory pressure here
1585 * ensures that active file pages get deactivated, until more
1586 * than half of the file pages are on the inactive list.
1587 *
1588 * Once we get to that situation, protect the system's working
1589 * set from being evicted by disabling active file page aging.
1590 *
1591 * This uses a different ratio than the anonymous pages, because
1592 * the page cache uses a use-once replacement algorithm.
1593 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001594static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001595{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001596 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001597 return mem_cgroup_inactive_file_is_low(lruvec);
Rik van Riel56e49d22009-06-16 15:32:28 -07001598
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001599 return inactive_file_is_low_global(lruvec_zone(lruvec));
Rik van Riel56e49d22009-06-16 15:32:28 -07001600}
1601
Hugh Dickins75b00af2012-05-29 15:07:09 -07001602static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001603{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001604 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001605 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001606 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001607 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001608}
1609
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001610static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001611 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001612{
Rik van Rielb39415b2009-12-14 17:59:48 -08001613 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001614 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001615 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001616 return 0;
1617 }
1618
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001619 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001620}
1621
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001622static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001623{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001624 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001625 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001626 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001627}
1628
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001630 * Determine how aggressively the anon and file LRU lists should be
1631 * scanned. The relative value of each set of LRU lists is determined
1632 * by looking at the fraction of the pages scanned we did rotate back
1633 * onto the active list instead of evict.
1634 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001635 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1636 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001637 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001638static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001639 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001640{
1641 unsigned long anon, file, free;
1642 unsigned long anon_prio, file_prio;
1643 unsigned long ap, fp;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001644 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001645 u64 fraction[2], denominator;
Hugh Dickins41113042012-01-12 17:20:01 -08001646 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001647 int noswap = 0;
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001648 bool force_scan = false;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001649 struct zone *zone = lruvec_zone(lruvec);
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001650
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001651 /*
1652 * If the zone or memcg is small, nr[l] can be 0. This
1653 * results in no scanning on this priority and a potential
1654 * priority drop. Global direct reclaim can go to the next
1655 * zone and tends to have no problems. Global kswapd is for
1656 * zone balancing and it needs to scan a minimum amount. When
1657 * reclaiming for a memcg, a priority drop can cause high
1658 * latencies, so it's better to scan a minimum amount there as
1659 * well.
1660 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001661 if (current_is_kswapd() && zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001662 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001663 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001664 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001665
1666 /* If we have no swap space, do not bother scanning anon pages. */
1667 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1668 noswap = 1;
1669 fraction[0] = 0;
1670 fraction[1] = 1;
1671 denominator = 1;
1672 goto out;
1673 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001674
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001675 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1676 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1677 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1678 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001679
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001680 if (global_reclaim(sc)) {
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001681 free = zone_page_state(zone, NR_FREE_PAGES);
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001682 /* If we have very few page cache pages,
1683 force-scan anon pages. */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001684 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001685 fraction[0] = 1;
1686 fraction[1] = 0;
1687 denominator = 1;
1688 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001689 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001690 }
1691
1692 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001693 * With swappiness at 100, anonymous and file have the same priority.
1694 * This scanning priority is essentially the inverse of IO cost.
1695 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001696 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001697 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001698
1699 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001700 * OK, so we have swap space and a fair amount of page cache
1701 * pages. We use the recently rotated / recently scanned
1702 * ratios to determine how valuable each cache is.
1703 *
1704 * Because workloads change over time (and to avoid overflow)
1705 * we keep these statistics as a floating average, which ends
1706 * up weighing recent references more than old ones.
1707 *
1708 * anon in [0], file in [1]
1709 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001710 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001711 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001712 reclaim_stat->recent_scanned[0] /= 2;
1713 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001714 }
1715
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001716 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001717 reclaim_stat->recent_scanned[1] /= 2;
1718 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001719 }
1720
1721 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001722 * The amount of pressure on anon vs file pages is inversely
1723 * proportional to the fraction of recently scanned pages on
1724 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001725 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001726 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001727 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001728
Satoru Moriyafe350042012-05-29 15:06:47 -07001729 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001730 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001731 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001732
Shaohua Li76a33fc2010-05-24 14:32:36 -07001733 fraction[0] = ap;
1734 fraction[1] = fp;
1735 denominator = ap + fp + 1;
1736out:
Hugh Dickins41113042012-01-12 17:20:01 -08001737 for_each_evictable_lru(lru) {
1738 int file = is_file_lru(lru);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001739 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001740
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001741 scan = get_lru_size(lruvec, lru);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001742 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1743 scan >>= sc->priority;
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001744 if (!scan && force_scan)
1745 scan = SWAP_CLUSTER_MAX;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001746 scan = div64_u64(scan * fraction[file], denominator);
1747 }
Hugh Dickins41113042012-01-12 17:20:01 -08001748 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001749 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001750}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001751
Mel Gorman23b9da52012-05-29 15:06:20 -07001752/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001753static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07001754{
1755 if (COMPACTION_BUILD && sc->order &&
1756 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001757 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07001758 return true;
1759
1760 return false;
1761}
1762
Mel Gorman83fde0f2012-10-08 16:29:11 -07001763#ifdef CONFIG_COMPACTION
1764/*
1765 * If compaction is deferred for sc->order then scale the number of pages
1766 * reclaimed based on the number of consecutive allocation failures
1767 */
1768static unsigned long scale_for_compaction(unsigned long pages_for_compaction,
1769 struct lruvec *lruvec, struct scan_control *sc)
1770{
1771 struct zone *zone = lruvec_zone(lruvec);
1772
1773 if (zone->compact_order_failed <= sc->order)
1774 pages_for_compaction <<= zone->compact_defer_shift;
1775 return pages_for_compaction;
1776}
1777#else
1778static unsigned long scale_for_compaction(unsigned long pages_for_compaction,
1779 struct lruvec *lruvec, struct scan_control *sc)
1780{
1781 return pages_for_compaction;
1782}
1783#endif
1784
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001785/*
Mel Gorman23b9da52012-05-29 15:06:20 -07001786 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1787 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1788 * true if more pages should be reclaimed such that when the page allocator
1789 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1790 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08001791 */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001792static inline bool should_continue_reclaim(struct lruvec *lruvec,
Mel Gorman3e7d3442011-01-13 15:45:56 -08001793 unsigned long nr_reclaimed,
1794 unsigned long nr_scanned,
1795 struct scan_control *sc)
1796{
1797 unsigned long pages_for_compaction;
1798 unsigned long inactive_lru_pages;
1799
1800 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001801 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001802 return false;
1803
Mel Gorman28765922011-02-25 14:44:20 -08001804 /* Consider stopping depending on scan and reclaim activity */
1805 if (sc->gfp_mask & __GFP_REPEAT) {
1806 /*
1807 * For __GFP_REPEAT allocations, stop reclaiming if the
1808 * full LRU list has been scanned and we are still failing
1809 * to reclaim pages. This full LRU scan is potentially
1810 * expensive but a __GFP_REPEAT caller really wants to succeed
1811 */
1812 if (!nr_reclaimed && !nr_scanned)
1813 return false;
1814 } else {
1815 /*
1816 * For non-__GFP_REPEAT allocations which can presumably
1817 * fail without consequence, stop if we failed to reclaim
1818 * any pages from the last SWAP_CLUSTER_MAX number of
1819 * pages that were scanned. This will return to the
1820 * caller faster at the risk reclaim/compaction and
1821 * the resulting allocation attempt fails
1822 */
1823 if (!nr_reclaimed)
1824 return false;
1825 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001826
1827 /*
1828 * If we have not reclaimed enough pages for compaction and the
1829 * inactive lists are large enough, continue reclaiming
1830 */
1831 pages_for_compaction = (2UL << sc->order);
Mel Gorman83fde0f2012-10-08 16:29:11 -07001832
1833 pages_for_compaction = scale_for_compaction(pages_for_compaction,
1834 lruvec, sc);
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001835 inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
Minchan Kim86cfd3a2012-01-10 15:08:18 -08001836 if (nr_swap_pages > 0)
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001837 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001838 if (sc->nr_reclaimed < pages_for_compaction &&
1839 inactive_lru_pages > pages_for_compaction)
1840 return true;
1841
1842 /* If compaction would go ahead or the allocation would succeed, stop */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001843 switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08001844 case COMPACT_PARTIAL:
1845 case COMPACT_CONTINUE:
1846 return false;
1847 default:
1848 return true;
1849 }
1850}
1851
1852/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1854 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001855static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001857 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001858 unsigned long nr_to_scan;
Hugh Dickins41113042012-01-12 17:20:01 -08001859 enum lru_list lru;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001860 unsigned long nr_reclaimed, nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001861 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Shaohua Li3da367c2011-10-31 17:07:03 -07001862 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863
Mel Gorman3e7d3442011-01-13 15:45:56 -08001864restart:
1865 nr_reclaimed = 0;
Johannes Weinerf0fdc5e2011-02-10 15:01:34 -08001866 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001867 get_scan_count(lruvec, sc, nr);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001868
Shaohua Li3da367c2011-10-31 17:07:03 -07001869 blk_start_plug(&plug);
Rik van Riel556adec2008-10-18 20:26:34 -07001870 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1871 nr[LRU_INACTIVE_FILE]) {
Hugh Dickins41113042012-01-12 17:20:01 -08001872 for_each_evictable_lru(lru) {
1873 if (nr[lru]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001874 nr_to_scan = min_t(unsigned long,
Hugh Dickins41113042012-01-12 17:20:01 -08001875 nr[lru], SWAP_CLUSTER_MAX);
1876 nr[lru] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877
Hugh Dickins41113042012-01-12 17:20:01 -08001878 nr_reclaimed += shrink_list(lru, nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001879 lruvec, sc);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001880 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 }
Rik van Riela79311c2009-01-06 14:40:01 -08001882 /*
1883 * On large memory systems, scan >> priority can become
1884 * really large. This is fine for the starting priority;
1885 * we want to put equal scanning pressure on each zone.
1886 * However, if the VM has a harder time of freeing pages,
1887 * with multiple processes reclaiming pages, the total
1888 * freeing target can get unreasonably large.
1889 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001890 if (nr_reclaimed >= nr_to_reclaim &&
1891 sc->priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001892 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 }
Shaohua Li3da367c2011-10-31 17:07:03 -07001894 blk_finish_plug(&plug);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001895 sc->nr_reclaimed += nr_reclaimed;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001896
Rik van Riel556adec2008-10-18 20:26:34 -07001897 /*
1898 * Even if we did not try to evict anon pages at all, we want to
1899 * rebalance the anon lru active/inactive ratio.
1900 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001901 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001902 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001903 sc, LRU_ACTIVE_ANON);
Rik van Riel556adec2008-10-18 20:26:34 -07001904
Mel Gorman3e7d3442011-01-13 15:45:56 -08001905 /* reclaim/compaction might need reclaim to continue */
Konstantin Khlebnikov90bdcfa2012-05-29 15:07:02 -07001906 if (should_continue_reclaim(lruvec, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001907 sc->nr_scanned - nr_scanned, sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001908 goto restart;
1909
Andrew Morton232ea4d2007-02-28 20:13:21 -08001910 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911}
1912
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001913static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08001914{
Johannes Weiner56600482012-01-12 17:17:59 -08001915 struct mem_cgroup *root = sc->target_mem_cgroup;
1916 struct mem_cgroup_reclaim_cookie reclaim = {
Johannes Weinerf16015f2012-01-12 17:17:52 -08001917 .zone = zone,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001918 .priority = sc->priority,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001919 };
Johannes Weiner56600482012-01-12 17:17:59 -08001920 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001921
Johannes Weiner56600482012-01-12 17:17:59 -08001922 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1923 do {
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001924 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner56600482012-01-12 17:17:59 -08001925
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001926 shrink_lruvec(lruvec, sc);
1927
Johannes Weiner56600482012-01-12 17:17:59 -08001928 /*
1929 * Limit reclaim has historically picked one memcg and
1930 * scanned it with decreasing priority levels until
1931 * nr_to_reclaim had been reclaimed. This priority
1932 * cycle is thus over after a single memcg.
Johannes Weinerb95a2f22012-01-12 17:18:06 -08001933 *
1934 * Direct reclaim and kswapd, on the other hand, have
1935 * to scan all memory cgroups to fulfill the overall
1936 * scan target for the zone.
Johannes Weiner56600482012-01-12 17:17:59 -08001937 */
1938 if (!global_reclaim(sc)) {
1939 mem_cgroup_iter_break(root, memcg);
1940 break;
1941 }
1942 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1943 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001944}
1945
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001946/* Returns true if compaction should go ahead for a high-order request */
1947static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1948{
1949 unsigned long balance_gap, watermark;
1950 bool watermark_ok;
1951
1952 /* Do not consider compaction for orders reclaim is meant to satisfy */
1953 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1954 return false;
1955
1956 /*
1957 * Compaction takes time to run and there are potentially other
1958 * callers using the pages just freed. Continue reclaiming until
1959 * there is a buffer of free pages available to give compaction
1960 * a reasonable chance of completing and allocating the page
1961 */
1962 balance_gap = min(low_wmark_pages(zone),
1963 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1964 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1965 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1966 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1967
1968 /*
1969 * If compaction is deferred, reclaim up to a point where
1970 * compaction will have a chance of success when re-enabled
1971 */
Rik van Rielaff62242012-03-21 16:33:52 -07001972 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001973 return watermark_ok;
1974
1975 /* If compaction is not ready to start, keep reclaiming */
1976 if (!compaction_suitable(zone, sc->order))
1977 return false;
1978
1979 return watermark_ok;
1980}
1981
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982/*
1983 * This is the direct reclaim path, for page-allocating processes. We only
1984 * try to reclaim pages from zones which will satisfy the caller's allocation
1985 * request.
1986 *
Mel Gorman41858962009-06-16 15:32:12 -07001987 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1988 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1990 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001991 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1992 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1993 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 * If a zone is deemed to be full of pinned pages then just give it a light
1996 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07001997 *
1998 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001999 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002000 * the caller that it should consider retrying the allocation instead of
2001 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002003static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004{
Mel Gormandd1a2392008-04-28 02:12:17 -07002005 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002006 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002007 unsigned long nr_soft_reclaimed;
2008 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002009 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002010
Mel Gormancc715d92012-03-21 16:34:00 -07002011 /*
2012 * If the number of buffer_heads in the machine exceeds the maximum
2013 * allowed level, force direct reclaim to scan the highmem zone as
2014 * highmem pages could be pinning lowmem pages storing buffer_heads
2015 */
2016 if (buffer_heads_over_limit)
2017 sc->gfp_mask |= __GFP_HIGHMEM;
2018
Mel Gormand4debc62010-08-09 17:19:29 -07002019 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2020 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002021 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002023 /*
2024 * Take care memory controller reclaiming has small influence
2025 * to global LRU.
2026 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002027 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002028 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2029 continue;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002030 if (zone->all_unreclaimable &&
2031 sc->priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002032 continue; /* Let kswapd poll it */
Rik van Riele0887c12011-10-31 17:09:31 -07002033 if (COMPACTION_BUILD) {
2034 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002035 * If we already have plenty of memory free for
2036 * compaction in this zone, don't free any more.
2037 * Even though compaction is invoked for any
2038 * non-zero order, only frequent costly order
2039 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002040 * noticeable problem, like transparent huge
2041 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002042 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002043 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002044 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002045 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002046 }
Rik van Riele0887c12011-10-31 17:09:31 -07002047 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002048 /*
2049 * This steals pages from memory cgroups over softlimit
2050 * and returns the number of reclaimed pages and
2051 * scanned pages. This works for global memory pressure
2052 * and balancing, not for a memcg's limit.
2053 */
2054 nr_soft_scanned = 0;
2055 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2056 sc->order, sc->gfp_mask,
2057 &nr_soft_scanned);
2058 sc->nr_reclaimed += nr_soft_reclaimed;
2059 sc->nr_scanned += nr_soft_scanned;
2060 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002061 }
Nick Piggin408d8542006-09-25 23:31:27 -07002062
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002063 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 }
Mel Gormane0c23272011-10-31 17:09:33 -07002065
Mel Gorman0cee34f2012-01-12 17:19:49 -08002066 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002067}
2068
2069static bool zone_reclaimable(struct zone *zone)
2070{
2071 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2072}
2073
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002074/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002075static bool all_unreclaimable(struct zonelist *zonelist,
2076 struct scan_control *sc)
2077{
2078 struct zoneref *z;
2079 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002080
2081 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2082 gfp_zone(sc->gfp_mask), sc->nodemask) {
2083 if (!populated_zone(zone))
2084 continue;
2085 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2086 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002087 if (!zone->all_unreclaimable)
2088 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002089 }
2090
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002091 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094/*
2095 * This is the main entry point to direct page reclaim.
2096 *
2097 * If a full scan of the inactive list fails to free enough memory then we
2098 * are "out of memory" and something needs to be killed.
2099 *
2100 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2101 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002102 * caller can't do much about. We kick the writeback threads and take explicit
2103 * naps in the hope that some of these pages can be written. But if the
2104 * allocating task holds filesystem locks which prevent writeout this might not
2105 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002106 *
2107 * returns: 0, if no pages reclaimed
2108 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 */
Mel Gormandac1d272008-04-28 02:12:12 -07002110static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002111 struct scan_control *sc,
2112 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113{
Andrew Morton69e05942006-03-22 00:08:19 -08002114 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002116 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002117 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002118 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002119 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Keika Kobayashi873b4772008-07-25 01:48:52 -07002121 delayacct_freepages_start();
2122
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002123 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002124 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002126 do {
Balbir Singh66e17072008-02-07 00:13:56 -08002127 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002128 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002129
Balbir Singh66e17072008-02-07 00:13:56 -08002130 /*
2131 * Don't shrink slabs when reclaiming memory from
2132 * over limit cgroups
2133 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002134 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002135 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002136 for_each_zone_zonelist(zone, z, zonelist,
2137 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002138 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2139 continue;
2140
2141 lru_pages += zone_reclaimable_pages(zone);
2142 }
2143
Ying Han1495f232011-05-24 17:12:27 -07002144 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002145 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002146 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002147 reclaim_state->reclaimed_slab = 0;
2148 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 }
Balbir Singh66e17072008-02-07 00:13:56 -08002150 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002151 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
2154 /*
2155 * Try to write back as many pages as we just scanned. This
2156 * tends to cause slow streaming writers to write data to the
2157 * disk smoothly, at the dirtying rate, which is nice. But
2158 * that's undesirable in laptop mode, where we *want* lumpy
2159 * writeout. So in laptop mode, write out the whole world.
2160 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002161 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2162 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002163 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2164 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002165 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 }
2167
2168 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002169 if (!sc->hibernation_mode && sc->nr_scanned &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002170 sc->priority < DEF_PRIORITY - 2) {
Mel Gorman0e093d992010-10-26 14:21:45 -07002171 struct zone *preferred_zone;
2172
2173 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002174 &cpuset_current_mems_allowed,
2175 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002176 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2177 }
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002178 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002181 delayacct_freepages_end();
2182
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002183 if (sc->nr_reclaimed)
2184 return sc->nr_reclaimed;
2185
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002186 /*
2187 * As hibernation is going on, kswapd is freezed so that it can't mark
2188 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2189 * check.
2190 */
2191 if (oom_killer_disabled)
2192 return 0;
2193
Mel Gorman0cee34f2012-01-12 17:19:49 -08002194 /* Aborted reclaim to try compaction? don't OOM, then */
2195 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002196 return 1;
2197
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002198 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002199 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002200 return 1;
2201
2202 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203}
2204
Mel Gorman55150612012-07-31 16:44:35 -07002205static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2206{
2207 struct zone *zone;
2208 unsigned long pfmemalloc_reserve = 0;
2209 unsigned long free_pages = 0;
2210 int i;
2211 bool wmark_ok;
2212
2213 for (i = 0; i <= ZONE_NORMAL; i++) {
2214 zone = &pgdat->node_zones[i];
2215 pfmemalloc_reserve += min_wmark_pages(zone);
2216 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2217 }
2218
2219 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2220
2221 /* kswapd must be awake if processes are being throttled */
2222 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2223 pgdat->classzone_idx = min(pgdat->classzone_idx,
2224 (enum zone_type)ZONE_NORMAL);
2225 wake_up_interruptible(&pgdat->kswapd_wait);
2226 }
2227
2228 return wmark_ok;
2229}
2230
2231/*
2232 * Throttle direct reclaimers if backing storage is backed by the network
2233 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2234 * depleted. kswapd will continue to make progress and wake the processes
2235 * when the low watermark is reached
2236 */
2237static void throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2238 nodemask_t *nodemask)
2239{
2240 struct zone *zone;
2241 int high_zoneidx = gfp_zone(gfp_mask);
2242 pg_data_t *pgdat;
2243
2244 /*
2245 * Kernel threads should not be throttled as they may be indirectly
2246 * responsible for cleaning pages necessary for reclaim to make forward
2247 * progress. kjournald for example may enter direct reclaim while
2248 * committing a transaction where throttling it could forcing other
2249 * processes to block on log_wait_commit().
2250 */
2251 if (current->flags & PF_KTHREAD)
2252 return;
2253
2254 /* Check if the pfmemalloc reserves are ok */
2255 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2256 pgdat = zone->zone_pgdat;
2257 if (pfmemalloc_watermark_ok(pgdat))
2258 return;
2259
Mel Gorman68243e72012-07-31 16:44:39 -07002260 /* Account for the throttling */
2261 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2262
Mel Gorman55150612012-07-31 16:44:35 -07002263 /*
2264 * If the caller cannot enter the filesystem, it's possible that it
2265 * is due to the caller holding an FS lock or performing a journal
2266 * transaction in the case of a filesystem like ext[3|4]. In this case,
2267 * it is not safe to block on pfmemalloc_wait as kswapd could be
2268 * blocked waiting on the same lock. Instead, throttle for up to a
2269 * second before continuing.
2270 */
2271 if (!(gfp_mask & __GFP_FS)) {
2272 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2273 pfmemalloc_watermark_ok(pgdat), HZ);
2274 return;
2275 }
2276
2277 /* Throttle until kswapd wakes the process */
2278 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2279 pfmemalloc_watermark_ok(pgdat));
2280}
2281
Mel Gormandac1d272008-04-28 02:12:12 -07002282unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002283 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002284{
Mel Gorman33906bc2010-08-09 17:19:16 -07002285 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002286 struct scan_control sc = {
2287 .gfp_mask = gfp_mask,
2288 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002289 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002290 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002291 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002292 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002293 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002294 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002295 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002296 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002297 struct shrink_control shrink = {
2298 .gfp_mask = sc.gfp_mask,
2299 };
Balbir Singh66e17072008-02-07 00:13:56 -08002300
Mel Gorman55150612012-07-31 16:44:35 -07002301 throttle_direct_reclaim(gfp_mask, zonelist, nodemask);
2302
2303 /*
2304 * Do not enter reclaim if fatal signal is pending. 1 is returned so
2305 * that the page allocator does not consider triggering OOM
2306 */
2307 if (fatal_signal_pending(current))
2308 return 1;
2309
Mel Gorman33906bc2010-08-09 17:19:16 -07002310 trace_mm_vmscan_direct_reclaim_begin(order,
2311 sc.may_writepage,
2312 gfp_mask);
2313
Ying Hana09ed5e2011-05-24 17:12:26 -07002314 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002315
2316 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2317
2318 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002319}
2320
Andrew Mortonc255a452012-07-31 16:43:02 -07002321#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002322
Johannes Weiner72835c82012-01-12 17:18:32 -08002323unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002324 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002325 struct zone *zone,
2326 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002327{
2328 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002329 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002330 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002331 .may_writepage = !laptop_mode,
2332 .may_unmap = 1,
2333 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002334 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002335 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002336 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002337 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002338 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002339
Balbir Singh4e416952009-09-23 15:56:39 -07002340 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2341 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002342
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002343 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002344 sc.may_writepage,
2345 sc.gfp_mask);
2346
Balbir Singh4e416952009-09-23 15:56:39 -07002347 /*
2348 * NOTE: Although we can get the priority field, using it
2349 * here is not a good idea, since it limits the pages we can scan.
2350 * if we don't reclaim here, the shrink_zone from balance_pgdat
2351 * will pick up pages from other mem cgroup's as well. We hack
2352 * the priority and make it zero.
2353 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002354 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002355
2356 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2357
Ying Han0ae5e892011-05-26 16:25:25 -07002358 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002359 return sc.nr_reclaimed;
2360}
2361
Johannes Weiner72835c82012-01-12 17:18:32 -08002362unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002363 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002364 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002365{
Balbir Singh4e416952009-09-23 15:56:39 -07002366 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002367 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002368 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002369 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002370 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002371 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002372 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002373 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002374 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002375 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002376 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002377 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002378 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2379 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2380 };
2381 struct shrink_control shrink = {
2382 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002383 };
Balbir Singh66e17072008-02-07 00:13:56 -08002384
Ying Han889976d2011-05-26 16:25:33 -07002385 /*
2386 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2387 * take care of from where we get pages. So the node where we start the
2388 * scan does not need to be the current node.
2389 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002390 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002391
2392 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002393
2394 trace_mm_vmscan_memcg_reclaim_begin(0,
2395 sc.may_writepage,
2396 sc.gfp_mask);
2397
Ying Hana09ed5e2011-05-24 17:12:26 -07002398 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002399
2400 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2401
2402 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002403}
2404#endif
2405
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002406static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002407{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002408 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002409
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002410 if (!total_swap_pages)
2411 return;
2412
2413 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2414 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002415 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002416
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002417 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002418 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002419 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002420
2421 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2422 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002423}
2424
Mel Gorman1741c872011-01-13 15:46:21 -08002425/*
2426 * pgdat_balanced is used when checking if a node is balanced for high-order
2427 * allocations. Only zones that meet watermarks and are in a zone allowed
2428 * by the callers classzone_idx are added to balanced_pages. The total of
2429 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2430 * for the node to be considered balanced. Forcing all zones to be balanced
2431 * for high orders can cause excessive reclaim when there are imbalanced zones.
2432 * The choice of 25% is due to
2433 * o a 16M DMA zone that is balanced will not balance a zone on any
2434 * reasonable sized machine
2435 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002436 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002437 * would need to be at least 256M for it to be balance a whole node.
2438 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2439 * to balance a node on its own. These seemed like reasonable ratios.
2440 */
2441static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2442 int classzone_idx)
2443{
2444 unsigned long present_pages = 0;
2445 int i;
2446
2447 for (i = 0; i <= classzone_idx; i++)
2448 present_pages += pgdat->node_zones[i].present_pages;
2449
Shaohua Li4746efd2011-07-19 08:49:26 -07002450 /* A special case here: if zone has no page, we think it's balanced */
2451 return balanced_pages >= (present_pages >> 2);
Mel Gorman1741c872011-01-13 15:46:21 -08002452}
2453
Mel Gorman55150612012-07-31 16:44:35 -07002454/*
2455 * Prepare kswapd for sleeping. This verifies that there are no processes
2456 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2457 *
2458 * Returns true if kswapd is ready to sleep
2459 */
2460static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002461 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002462{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002463 int i;
Mel Gorman1741c872011-01-13 15:46:21 -08002464 unsigned long balanced = 0;
2465 bool all_zones_ok = true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002466
2467 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2468 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002469 return false;
2470
2471 /*
2472 * There is a potential race between when kswapd checks its watermarks
2473 * and a process gets throttled. There is also a potential race if
2474 * processes get throttled, kswapd wakes, a large process exits therby
2475 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2476 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2477 * so wake them now if necessary. If necessary, processes will wake
2478 * kswapd and get throttled again
2479 */
2480 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2481 wake_up(&pgdat->pfmemalloc_wait);
2482 return false;
2483 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002484
Mel Gorman0abdee22011-01-13 15:46:22 -08002485 /* Check the watermark levels */
Mel Gorman08951e52011-07-08 15:39:36 -07002486 for (i = 0; i <= classzone_idx; i++) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002487 struct zone *zone = pgdat->node_zones + i;
2488
2489 if (!populated_zone(zone))
2490 continue;
2491
Mel Gorman355b09c2011-01-13 15:46:24 -08002492 /*
2493 * balance_pgdat() skips over all_unreclaimable after
2494 * DEF_PRIORITY. Effectively, it considers them balanced so
2495 * they must be considered balanced here as well if kswapd
2496 * is to sleep
2497 */
2498 if (zone->all_unreclaimable) {
2499 balanced += zone->present_pages;
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002500 continue;
Mel Gorman355b09c2011-01-13 15:46:24 -08002501 }
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002502
Mel Gorman88f5acf2011-01-13 15:45:41 -08002503 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
Mel Gormanda175d02011-07-08 15:39:39 -07002504 i, 0))
Mel Gorman1741c872011-01-13 15:46:21 -08002505 all_zones_ok = false;
2506 else
2507 balanced += zone->present_pages;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002508 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002509
Mel Gorman1741c872011-01-13 15:46:21 -08002510 /*
2511 * For high-order requests, the balanced zones must contain at least
2512 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2513 * must be balanced
2514 */
2515 if (order)
Mel Gorman55150612012-07-31 16:44:35 -07002516 return pgdat_balanced(pgdat, balanced, classzone_idx);
Mel Gorman1741c872011-01-13 15:46:21 -08002517 else
Mel Gorman55150612012-07-31 16:44:35 -07002518 return all_zones_ok;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002519}
2520
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521/*
2522 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002523 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002525 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 *
2527 * There is special handling here for zones which are full of pinned pages.
2528 * This can happen if the pages are all mlocked, or if they are all used by
2529 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2530 * What we do is to detect the case where all pages in the zone have been
2531 * scanned twice and there has been zero successful reclaim. Mark the zone as
2532 * dead and from now on, only perform a short scan. Basically we're polling
2533 * the zone for when the problem goes away.
2534 *
2535 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002536 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2537 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2538 * lower zones regardless of the number of free pages in the lower zones. This
2539 * interoperates with the page allocator fallback scheme to ensure that aging
2540 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 */
Mel Gorman99504742011-01-13 15:46:20 -08002542static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002543 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 int all_zones_ok;
Mel Gorman1741c872011-01-13 15:46:21 -08002546 unsigned long balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002548 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton69e05942006-03-22 00:08:19 -08002549 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002551 unsigned long nr_soft_reclaimed;
2552 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002553 struct scan_control sc = {
2554 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002555 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002556 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002557 /*
2558 * kswapd doesn't want to be bailed out while reclaim. because
2559 * we want to put equal scanning pressure on each zone.
2560 */
2561 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002562 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002563 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002564 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002565 struct shrink_control shrink = {
2566 .gfp_mask = sc.gfp_mask,
2567 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568loop_again:
2569 total_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002570 sc.priority = DEF_PRIORITY;
Rik van Riela79311c2009-01-06 14:40:01 -08002571 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002572 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002573 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002575 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002577 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578
2579 all_zones_ok = 1;
Mel Gorman1741c872011-01-13 15:46:21 -08002580 balanced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002582 /*
2583 * Scan in the highmem->dma direction for the highest
2584 * zone which needs scanning
2585 */
2586 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2587 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002589 if (!populated_zone(zone))
2590 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002592 if (zone->all_unreclaimable &&
2593 sc.priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002594 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595
Rik van Riel556adec2008-10-18 20:26:34 -07002596 /*
2597 * Do some background aging of the anon list, to give
2598 * pages a chance to be referenced before reclaiming.
2599 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002600 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002601
Mel Gormancc715d92012-03-21 16:34:00 -07002602 /*
2603 * If the number of buffer_heads in the machine
2604 * exceeds the maximum allowed level and this node
2605 * has a highmem zone, force kswapd to reclaim from
2606 * it to relieve lowmem pressure.
2607 */
2608 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2609 end_zone = i;
2610 break;
2611 }
2612
Mel Gorman88f5acf2011-01-13 15:45:41 -08002613 if (!zone_watermark_ok_safe(zone, order,
Mel Gorman41858962009-06-16 15:32:12 -07002614 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002615 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002616 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002617 } else {
2618 /* If balanced, clear the congested flag */
2619 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002622 if (i < 0)
2623 goto out;
2624
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 for (i = 0; i <= end_zone; i++) {
2626 struct zone *zone = pgdat->node_zones + i;
2627
Wu Fengguangadea02a2009-09-21 17:01:42 -07002628 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 }
2630
2631 /*
2632 * Now scan the zone in the dma->highmem direction, stopping
2633 * at the last zone which needs scanning.
2634 *
2635 * We do this because the page allocator works in the opposite
2636 * direction. This prevents the page allocator from allocating
2637 * pages behind kswapd's direction of progress, which would
2638 * cause too much scanning of the lower zones.
2639 */
2640 for (i = 0; i <= end_zone; i++) {
2641 struct zone *zone = pgdat->node_zones + i;
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002642 int nr_slab, testorder;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002643 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644
Con Kolivasf3fe6512006-01-06 00:11:15 -08002645 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 continue;
2647
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002648 if (zone->all_unreclaimable &&
2649 sc.priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 continue;
2651
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002653
Ying Han0ae5e892011-05-26 16:25:25 -07002654 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002655 /*
2656 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002657 */
Ying Han0ae5e892011-05-26 16:25:25 -07002658 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2659 order, sc.gfp_mask,
2660 &nr_soft_scanned);
2661 sc.nr_reclaimed += nr_soft_reclaimed;
2662 total_scanned += nr_soft_scanned;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002663
Rik van Riel32a43302007-10-16 01:24:50 -07002664 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002665 * We put equal pressure on every zone, unless
2666 * one zone has way too many pages free
2667 * already. The "too many pages" is defined
2668 * as the high wmark plus a "gap" where the
2669 * gap is either the low watermark or 1%
2670 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002671 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002672 balance_gap = min(low_wmark_pages(zone),
2673 (zone->present_pages +
2674 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2675 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002676 /*
2677 * Kswapd reclaims only single pages with compaction
2678 * enabled. Trying too hard to reclaim until contiguous
2679 * free pages have become available can hurt performance
2680 * by evicting too much useful data from memory.
2681 * Do not reclaim more than needed for compaction.
2682 */
2683 testorder = order;
2684 if (COMPACTION_BUILD && order &&
2685 compaction_suitable(zone, order) !=
2686 COMPACT_SKIPPED)
2687 testorder = 0;
2688
Mel Gormancc715d92012-03-21 16:34:00 -07002689 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
Hugh Dickins643ac9f2012-03-23 02:57:31 -07002690 !zone_watermark_ok_safe(zone, testorder,
Mel Gorman8afdcec2011-03-22 16:33:04 -07002691 high_wmark_pages(zone) + balance_gap,
Mel Gormand7868da2011-07-08 15:39:38 -07002692 end_zone, 0)) {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002693 shrink_zone(zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002694
Mel Gormand7868da2011-07-08 15:39:38 -07002695 reclaim_state->reclaimed_slab = 0;
2696 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2697 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2698 total_scanned += sc.nr_scanned;
2699
2700 if (nr_slab == 0 && !zone_reclaimable(zone))
2701 zone->all_unreclaimable = 1;
2702 }
2703
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 /*
2705 * If we've done a decent amount of scanning and
2706 * the reclaim ratio is low, start doing writepage
2707 * even in laptop mode
2708 */
2709 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002710 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002712
Mel Gorman215ddd62011-07-08 15:39:40 -07002713 if (zone->all_unreclaimable) {
2714 if (end_zone && end_zone == i)
2715 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002716 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002717 }
Mel Gormand7868da2011-07-08 15:39:38 -07002718
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002719 if (!zone_watermark_ok_safe(zone, testorder,
Minchan Kim45973d72010-03-05 13:41:45 -08002720 high_wmark_pages(zone), end_zone, 0)) {
2721 all_zones_ok = 0;
2722 /*
2723 * We are still under min water mark. This
2724 * means that we have a GFP_ATOMIC allocation
2725 * failure risk. Hurry up!
2726 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08002727 if (!zone_watermark_ok_safe(zone, order,
Minchan Kim45973d72010-03-05 13:41:45 -08002728 min_wmark_pages(zone), end_zone, 0))
2729 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002730 } else {
2731 /*
2732 * If a zone reaches its high watermark,
2733 * consider it to be no longer congested. It's
2734 * possible there are dirty pages backed by
2735 * congested BDIs but as pressure is relieved,
Wanpeng Liab8704b2012-06-17 09:27:18 +08002736 * speculatively avoid congestion waits
Mel Gorman0e093d992010-10-26 14:21:45 -07002737 */
2738 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormandc83edd2011-01-13 15:46:26 -08002739 if (i <= *classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002740 balanced += zone->present_pages;
Minchan Kim45973d72010-03-05 13:41:45 -08002741 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002742
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 }
Mel Gorman55150612012-07-31 16:44:35 -07002744
2745 /*
2746 * If the low watermark is met there is no need for processes
2747 * to be throttled on pfmemalloc_wait as they should not be
2748 * able to safely make forward progress. Wake them
2749 */
2750 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2751 pfmemalloc_watermark_ok(pgdat))
2752 wake_up(&pgdat->pfmemalloc_wait);
2753
Mel Gormandc83edd2011-01-13 15:46:26 -08002754 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 break; /* kswapd: all done */
2756 /*
2757 * OK, kswapd is getting into trouble. Take a nap, then take
2758 * another pass across the zones.
2759 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002760 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002761 if (has_under_min_watermark_zone)
2762 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2763 else
2764 congestion_wait(BLK_RW_ASYNC, HZ/10);
2765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767 /*
2768 * We do this so kswapd doesn't build up large priorities for
2769 * example when it is freeing in parallel with allocators. It
2770 * matches the direct reclaim path behaviour in terms of impact
2771 * on zone->*_priority.
2772 */
Rik van Riela79311c2009-01-06 14:40:01 -08002773 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 break;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002775 } while (--sc.priority >= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776out:
Mel Gorman99504742011-01-13 15:46:20 -08002777
2778 /*
2779 * order-0: All zones must meet high watermark for a balanced node
Mel Gorman1741c872011-01-13 15:46:21 -08002780 * high-order: Balanced zones must make up at least 25% of the node
2781 * for the node to be balanced
Mel Gorman99504742011-01-13 15:46:20 -08002782 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002783 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002785
2786 try_to_freeze();
2787
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002788 /*
2789 * Fragmentation may mean that the system cannot be
2790 * rebalanced for high-order allocations in all zones.
2791 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2792 * it means the zones have been fully scanned and are still
2793 * not balanced. For high-order allocations, there is
2794 * little point trying all over again as kswapd may
2795 * infinite loop.
2796 *
2797 * Instead, recheck all watermarks at order-0 as they
2798 * are the most important. If watermarks are ok, kswapd will go
2799 * back to sleep. High-order users can still perform direct
2800 * reclaim if they wish.
2801 */
2802 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2803 order = sc.order = 0;
2804
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 goto loop_again;
2806 }
2807
Mel Gorman99504742011-01-13 15:46:20 -08002808 /*
2809 * If kswapd was reclaiming at a higher order, it has the option of
2810 * sleeping without all zones being balanced. Before it does, it must
2811 * ensure that the watermarks for order-0 on *all* zones are met and
2812 * that the congestion flags are cleared. The congestion flag must
2813 * be cleared as kswapd is the only mechanism that clears the flag
2814 * and it is potentially going to sleep here.
2815 */
2816 if (order) {
Rik van Riel7be62de2012-03-21 16:33:52 -07002817 int zones_need_compaction = 1;
2818
Mel Gorman99504742011-01-13 15:46:20 -08002819 for (i = 0; i <= end_zone; i++) {
2820 struct zone *zone = pgdat->node_zones + i;
2821
2822 if (!populated_zone(zone))
2823 continue;
2824
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002825 if (zone->all_unreclaimable &&
2826 sc.priority != DEF_PRIORITY)
Mel Gorman99504742011-01-13 15:46:20 -08002827 continue;
2828
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002829 /* Would compaction fail due to lack of free memory? */
Rik van Riel496b9192012-03-24 10:26:21 -04002830 if (COMPACTION_BUILD &&
2831 compaction_suitable(zone, order) == COMPACT_SKIPPED)
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002832 goto loop_again;
2833
Mel Gorman99504742011-01-13 15:46:20 -08002834 /* Confirm the zone is balanced for order-0 */
2835 if (!zone_watermark_ok(zone, 0,
2836 high_wmark_pages(zone), 0, 0)) {
2837 order = sc.order = 0;
2838 goto loop_again;
2839 }
2840
Rik van Riel7be62de2012-03-21 16:33:52 -07002841 /* Check if the memory needs to be defragmented. */
2842 if (zone_watermark_ok(zone, order,
2843 low_wmark_pages(zone), *classzone_idx, 0))
2844 zones_need_compaction = 0;
2845
Mel Gorman99504742011-01-13 15:46:20 -08002846 /* If balanced, clear the congested flag */
2847 zone_clear_flag(zone, ZONE_CONGESTED);
2848 }
Rik van Riel7be62de2012-03-21 16:33:52 -07002849
2850 if (zones_need_compaction)
2851 compact_pgdat(pgdat, order);
Mel Gorman99504742011-01-13 15:46:20 -08002852 }
2853
Mel Gorman0abdee22011-01-13 15:46:22 -08002854 /*
Mel Gorman55150612012-07-31 16:44:35 -07002855 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08002856 * makes a decision on the order we were last reclaiming at. However,
2857 * if another caller entered the allocator slow path while kswapd
2858 * was awake, order will remain at the higher level
2859 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002860 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002861 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862}
2863
Mel Gormandc83edd2011-01-13 15:46:26 -08002864static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002865{
2866 long remaining = 0;
2867 DEFINE_WAIT(wait);
2868
2869 if (freezing(current) || kthread_should_stop())
2870 return;
2871
2872 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2873
2874 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07002875 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002876 remaining = schedule_timeout(HZ/10);
2877 finish_wait(&pgdat->kswapd_wait, &wait);
2878 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2879 }
2880
2881 /*
2882 * After a short sleep, check if it was a premature sleep. If not, then
2883 * go fully to sleep until explicitly woken up.
2884 */
Mel Gorman55150612012-07-31 16:44:35 -07002885 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002886 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2887
2888 /*
2889 * vmstat counters are not perfectly accurate and the estimated
2890 * value for counters such as NR_FREE_PAGES can deviate from the
2891 * true value by nr_online_cpus * threshold. To avoid the zone
2892 * watermarks being breached while under pressure, we reduce the
2893 * per-cpu vmstat threshold while kswapd is awake and restore
2894 * them before going back to sleep.
2895 */
2896 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07002897
Mel Gorman62997022012-10-08 16:32:47 -07002898 /*
2899 * Compaction records what page blocks it recently failed to
2900 * isolate pages from and skips them in the future scanning.
2901 * When kswapd is going to sleep, it is reasonable to assume
2902 * that pages and compaction may succeed so reset the cache.
2903 */
2904 reset_isolation_suitable(pgdat);
2905
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07002906 if (!kthread_should_stop())
2907 schedule();
2908
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002909 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2910 } else {
2911 if (remaining)
2912 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2913 else
2914 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2915 }
2916 finish_wait(&pgdat->kswapd_wait, &wait);
2917}
2918
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919/*
2920 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002921 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 *
2923 * This basically trickles out pages so that we have _some_
2924 * free memory available even if there is no other activity
2925 * that frees anything up. This is needed for things like routing
2926 * etc, where we otherwise might have all activity going on in
2927 * asynchronous contexts that cannot page things out.
2928 *
2929 * If there are applications that are active memory-allocators
2930 * (most normal use), this basically shouldn't matter.
2931 */
2932static int kswapd(void *p)
2933{
Mel Gorman215ddd62011-07-08 15:39:40 -07002934 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002935 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07002936 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002937 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 pg_data_t *pgdat = (pg_data_t*)p;
2939 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002940
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 struct reclaim_state reclaim_state = {
2942 .reclaimed_slab = 0,
2943 };
Rusty Russella70f7302009-03-13 14:49:46 +10302944 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
Nick Piggincf40bd12009-01-21 08:12:39 +01002946 lockdep_set_current_reclaim_state(GFP_KERNEL);
2947
Rusty Russell174596a2009-01-01 10:12:29 +10302948 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002949 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 current->reclaim_state = &reclaim_state;
2951
2952 /*
2953 * Tell the memory management that we're a "memory allocator",
2954 * and that if we need more memory we should get access to it
2955 * regardless (see "__alloc_pages()"). "kswapd" should
2956 * never get caught in the normal page freeing logic.
2957 *
2958 * (Kswapd normally doesn't need memory anyway, but sometimes
2959 * you need a small amount of memory in order to be able to
2960 * page out something else, and this flag essentially protects
2961 * us from recursively trying to free more memory as we're
2962 * trying to free the first piece of memory in the first place).
2963 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002964 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002965 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966
Mel Gorman215ddd62011-07-08 15:39:40 -07002967 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002968 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002969 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002970 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 for ( ; ; ) {
David Rientjes8fe23e02009-12-14 17:58:33 -08002972 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002973
Mel Gorman215ddd62011-07-08 15:39:40 -07002974 /*
2975 * If the last balance_pgdat was unsuccessful it's unlikely a
2976 * new request of a similar or harder type will succeed soon
2977 * so consider going to sleep on the basis we reclaimed at
2978 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002979 if (balanced_classzone_idx >= new_classzone_idx &&
2980 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07002981 new_order = pgdat->kswapd_max_order;
2982 new_classzone_idx = pgdat->classzone_idx;
2983 pgdat->kswapd_max_order = 0;
2984 pgdat->classzone_idx = pgdat->nr_zones - 1;
2985 }
2986
Mel Gorman99504742011-01-13 15:46:20 -08002987 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 /*
2989 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08002990 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 */
2992 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08002993 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002995 kswapd_try_to_sleep(pgdat, balanced_order,
2996 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08002998 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07002999 new_order = order;
3000 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003001 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003002 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
David Rientjes8fe23e02009-12-14 17:58:33 -08003005 ret = try_to_freeze();
3006 if (kthread_should_stop())
3007 break;
3008
3009 /*
3010 * We can speed up thawing tasks if we don't call balance_pgdat
3011 * after returning from the refrigerator
3012 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003013 if (!ret) {
3014 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003015 balanced_classzone_idx = classzone_idx;
3016 balanced_order = balance_pgdat(pgdat, order,
3017 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003018 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003020
3021 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 return 0;
3023}
3024
3025/*
3026 * A zone is low on free memory, so wake its kswapd task to service it.
3027 */
Mel Gorman99504742011-01-13 15:46:20 -08003028void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029{
3030 pg_data_t *pgdat;
3031
Con Kolivasf3fe6512006-01-06 00:11:15 -08003032 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 return;
3034
Paul Jackson02a0e532006-12-13 00:34:25 -08003035 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003037 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003038 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003039 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003040 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3041 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003042 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003044 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3045 return;
3046
3047 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003048 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049}
3050
Wu Fengguangadea02a2009-09-21 17:01:42 -07003051/*
3052 * The reclaimable count would be mostly accurate.
3053 * The less reclaimable pages may be
3054 * - mlocked pages, which will be moved to unevictable list when encountered
3055 * - mapped pages, which may require several travels to be reclaimed
3056 * - dirty pages, which is not "instantly" reclaimable
3057 */
3058unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003059{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003060 int nr;
3061
3062 nr = global_page_state(NR_ACTIVE_FILE) +
3063 global_page_state(NR_INACTIVE_FILE);
3064
3065 if (nr_swap_pages > 0)
3066 nr += global_page_state(NR_ACTIVE_ANON) +
3067 global_page_state(NR_INACTIVE_ANON);
3068
3069 return nr;
3070}
3071
3072unsigned long zone_reclaimable_pages(struct zone *zone)
3073{
3074 int nr;
3075
3076 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3077 zone_page_state(zone, NR_INACTIVE_FILE);
3078
3079 if (nr_swap_pages > 0)
3080 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3081 zone_page_state(zone, NR_INACTIVE_ANON);
3082
3083 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003084}
3085
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003086#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003088 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003089 * freed pages.
3090 *
3091 * Rather than trying to age LRUs the aim is to preserve the overall
3092 * LRU order by reclaiming preferentially
3093 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003095unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003097 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003098 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003099 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3100 .may_swap = 1,
3101 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003102 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003103 .nr_to_reclaim = nr_to_reclaim,
3104 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003105 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003106 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003108 struct shrink_control shrink = {
3109 .gfp_mask = sc.gfp_mask,
3110 };
3111 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003112 struct task_struct *p = current;
3113 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003115 p->flags |= PF_MEMALLOC;
3116 lockdep_set_current_reclaim_state(sc.gfp_mask);
3117 reclaim_state.reclaimed_slab = 0;
3118 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003119
Ying Hana09ed5e2011-05-24 17:12:26 -07003120 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003121
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003122 p->reclaim_state = NULL;
3123 lockdep_clear_current_reclaim_state();
3124 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003125
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003126 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003128#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130/* It's optimal to keep kswapds on the same CPUs as their memory, but
3131 not required for correctness. So if the last cpu in a node goes
3132 away, we get changed to run anywhere: as the first one comes back,
3133 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003134static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08003135 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003137 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003139 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003140 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003141 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303142 const struct cpumask *mask;
3143
3144 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003145
Rusty Russell3e597942009-01-01 10:12:24 +10303146 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003148 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149 }
3150 }
3151 return NOTIFY_OK;
3152}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153
Yasunori Goto3218ae12006-06-27 02:53:33 -07003154/*
3155 * This kswapd start function will be called by init and node-hot-add.
3156 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3157 */
3158int kswapd_run(int nid)
3159{
3160 pg_data_t *pgdat = NODE_DATA(nid);
3161 int ret = 0;
3162
3163 if (pgdat->kswapd)
3164 return 0;
3165
3166 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3167 if (IS_ERR(pgdat->kswapd)) {
3168 /* failure at boot is fatal */
3169 BUG_ON(system_state == SYSTEM_BOOTING);
Wen Congyang18b48d52012-09-17 14:08:55 -07003170 pgdat->kswapd = NULL;
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003171 pr_err("Failed to start kswapd on node %d\n", nid);
3172 ret = PTR_ERR(pgdat->kswapd);
Yasunori Goto3218ae12006-06-27 02:53:33 -07003173 }
3174 return ret;
3175}
3176
David Rientjes8fe23e02009-12-14 17:58:33 -08003177/*
Jiang Liud8adde12012-07-11 14:01:52 -07003178 * Called by memory hotplug when all memory in a node is offlined. Caller must
3179 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003180 */
3181void kswapd_stop(int nid)
3182{
3183 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3184
Jiang Liud8adde12012-07-11 14:01:52 -07003185 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003186 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003187 NODE_DATA(nid)->kswapd = NULL;
3188 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003189}
3190
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191static int __init kswapd_init(void)
3192{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003193 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003194
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07003196 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003197 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 hotcpu_notifier(cpu_callback, 0);
3199 return 0;
3200}
3201
3202module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003203
3204#ifdef CONFIG_NUMA
3205/*
3206 * Zone reclaim mode
3207 *
3208 * If non-zero call zone_reclaim when the number of free pages falls below
3209 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003210 */
3211int zone_reclaim_mode __read_mostly;
3212
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003213#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003214#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003215#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3216#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3217
Christoph Lameter9eeff232006-01-18 17:42:31 -08003218/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003219 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3220 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3221 * a zone.
3222 */
3223#define ZONE_RECLAIM_PRIORITY 4
3224
Christoph Lameter9eeff232006-01-18 17:42:31 -08003225/*
Christoph Lameter96146342006-07-03 00:24:13 -07003226 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3227 * occur.
3228 */
3229int sysctl_min_unmapped_ratio = 1;
3230
3231/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003232 * If the number of slab pages in a zone grows beyond this percentage then
3233 * slab reclaim needs to occur.
3234 */
3235int sysctl_min_slab_ratio = 5;
3236
Mel Gorman90afa5d2009-06-16 15:33:20 -07003237static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3238{
3239 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3240 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3241 zone_page_state(zone, NR_ACTIVE_FILE);
3242
3243 /*
3244 * It's possible for there to be more file mapped pages than
3245 * accounted for by the pages on the file LRU lists because
3246 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3247 */
3248 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3249}
3250
3251/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3252static long zone_pagecache_reclaimable(struct zone *zone)
3253{
3254 long nr_pagecache_reclaimable;
3255 long delta = 0;
3256
3257 /*
3258 * If RECLAIM_SWAP is set, then all file pages are considered
3259 * potentially reclaimable. Otherwise, we have to worry about
3260 * pages like swapcache and zone_unmapped_file_pages() provides
3261 * a better estimate
3262 */
3263 if (zone_reclaim_mode & RECLAIM_SWAP)
3264 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3265 else
3266 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3267
3268 /* If we can't clean pages, remove dirty pages from consideration */
3269 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3270 delta += zone_page_state(zone, NR_FILE_DIRTY);
3271
3272 /* Watch for any possible underflows due to delta */
3273 if (unlikely(delta > nr_pagecache_reclaimable))
3274 delta = nr_pagecache_reclaimable;
3275
3276 return nr_pagecache_reclaimable - delta;
3277}
3278
Christoph Lameter0ff38492006-09-25 23:31:52 -07003279/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003280 * Try to free up some pages from this zone through reclaim.
3281 */
Andrew Morton179e9632006-03-22 00:08:18 -08003282static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003283{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003284 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003285 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003286 struct task_struct *p = current;
3287 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003288 struct scan_control sc = {
3289 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003290 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003291 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003292 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3293 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08003294 .gfp_mask = gfp_mask,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003295 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003296 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003297 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003298 struct shrink_control shrink = {
3299 .gfp_mask = sc.gfp_mask,
3300 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003301 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003302
Christoph Lameter9eeff232006-01-18 17:42:31 -08003303 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003304 /*
3305 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3306 * and we also need to be able to write out pages for RECLAIM_WRITE
3307 * and RECLAIM_SWAP.
3308 */
3309 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003310 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003311 reclaim_state.reclaimed_slab = 0;
3312 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003313
Mel Gorman90afa5d2009-06-16 15:33:20 -07003314 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003315 /*
3316 * Free memory by calling shrink zone with increasing
3317 * priorities until we have enough memory freed.
3318 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003319 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003320 shrink_zone(zone, &sc);
3321 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003322 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003323
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003324 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3325 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003326 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003327 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003328 * many pages were freed in this zone. So we take the current
3329 * number of slab pages and shake the slab until it is reduced
3330 * by the same nr_pages that we used for reclaiming unmapped
3331 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003332 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003333 * Note that shrink_slab will free memory on all zones and may
3334 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003335 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003336 for (;;) {
3337 unsigned long lru_pages = zone_reclaimable_pages(zone);
3338
3339 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003340 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003341 break;
3342
3343 /* Freed enough memory */
3344 nr_slab_pages1 = zone_page_state(zone,
3345 NR_SLAB_RECLAIMABLE);
3346 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3347 break;
3348 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003349
3350 /*
3351 * Update nr_reclaimed by the number of slab pages we
3352 * reclaimed from this zone.
3353 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003354 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3355 if (nr_slab_pages1 < nr_slab_pages0)
3356 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003357 }
3358
Christoph Lameter9eeff232006-01-18 17:42:31 -08003359 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003360 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003361 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003362 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003363}
Andrew Morton179e9632006-03-22 00:08:18 -08003364
3365int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3366{
Andrew Morton179e9632006-03-22 00:08:18 -08003367 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003368 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003369
3370 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003371 * Zone reclaim reclaims unmapped file backed pages and
3372 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003373 *
Christoph Lameter96146342006-07-03 00:24:13 -07003374 * A small portion of unmapped file backed pages is needed for
3375 * file I/O otherwise pages read by file I/O will be immediately
3376 * thrown out if the zone is overallocated. So we do not reclaim
3377 * if less than a specified percentage of the zone is used by
3378 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003379 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003380 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3381 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003382 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003383
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003384 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003385 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003386
Andrew Morton179e9632006-03-22 00:08:18 -08003387 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003388 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003389 */
David Rientjesd773ed62007-10-16 23:26:01 -07003390 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003391 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003392
3393 /*
3394 * Only run zone reclaim on the local zone or on zones that do not
3395 * have associated processors. This will favor the local processor
3396 * over remote processors and spread off node memory allocations
3397 * as wide as possible.
3398 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003399 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003400 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003401 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003402
3403 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003404 return ZONE_RECLAIM_NOSCAN;
3405
David Rientjesd773ed62007-10-16 23:26:01 -07003406 ret = __zone_reclaim(zone, gfp_mask, order);
3407 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3408
Mel Gorman24cf725182009-06-16 15:33:23 -07003409 if (!ret)
3410 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3411
David Rientjesd773ed62007-10-16 23:26:01 -07003412 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003413}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003414#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003415
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003416/*
3417 * page_evictable - test whether a page is evictable
3418 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003419 *
3420 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003421 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003422 *
3423 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003424 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003425 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003426 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003427 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003428int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003429{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003430 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003431}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003432
Hugh Dickins850465792012-01-20 14:34:19 -08003433#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003434/**
Hugh Dickins24513262012-01-20 14:34:21 -08003435 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3436 * @pages: array of pages to check
3437 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003438 *
Hugh Dickins24513262012-01-20 14:34:21 -08003439 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003440 *
3441 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003442 */
Hugh Dickins24513262012-01-20 14:34:21 -08003443void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003444{
Johannes Weiner925b7672012-01-12 17:18:15 -08003445 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003446 struct zone *zone = NULL;
3447 int pgscanned = 0;
3448 int pgrescued = 0;
3449 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003450
Hugh Dickins24513262012-01-20 14:34:21 -08003451 for (i = 0; i < nr_pages; i++) {
3452 struct page *page = pages[i];
3453 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003454
Hugh Dickins24513262012-01-20 14:34:21 -08003455 pgscanned++;
3456 pagezone = page_zone(page);
3457 if (pagezone != zone) {
3458 if (zone)
3459 spin_unlock_irq(&zone->lru_lock);
3460 zone = pagezone;
3461 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003462 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003463 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003464
Hugh Dickins24513262012-01-20 14:34:21 -08003465 if (!PageLRU(page) || !PageUnevictable(page))
3466 continue;
3467
Hugh Dickins39b5f292012-10-08 16:33:18 -07003468 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003469 enum lru_list lru = page_lru_base_type(page);
3470
3471 VM_BUG_ON(PageActive(page));
3472 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003473 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3474 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003475 pgrescued++;
3476 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003477 }
Hugh Dickins24513262012-01-20 14:34:21 -08003478
3479 if (zone) {
3480 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3481 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3482 spin_unlock_irq(&zone->lru_lock);
3483 }
Hugh Dickins850465792012-01-20 14:34:19 -08003484}
3485#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003486
Johannes Weiner264e56d2011-10-31 17:09:13 -07003487static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003488{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003489 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003490 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003491 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003492 "one, please send an email to linux-mm@kvack.org.\n",
3493 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003494}
3495
3496/*
3497 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3498 * all nodes' unevictable lists for evictable pages
3499 */
3500unsigned long scan_unevictable_pages;
3501
3502int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003503 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003504 size_t *length, loff_t *ppos)
3505{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003506 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003507 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003508 scan_unevictable_pages = 0;
3509 return 0;
3510}
3511
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003512#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003513/*
3514 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3515 * a specified node's per zone unevictable lists for evictable pages.
3516 */
3517
Kay Sievers10fbcf42011-12-21 14:48:43 -08003518static ssize_t read_scan_unevictable_node(struct device *dev,
3519 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003520 char *buf)
3521{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003522 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003523 return sprintf(buf, "0\n"); /* always zero; should fit... */
3524}
3525
Kay Sievers10fbcf42011-12-21 14:48:43 -08003526static ssize_t write_scan_unevictable_node(struct device *dev,
3527 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003528 const char *buf, size_t count)
3529{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003530 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003531 return 1;
3532}
3533
3534
Kay Sievers10fbcf42011-12-21 14:48:43 -08003535static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003536 read_scan_unevictable_node,
3537 write_scan_unevictable_node);
3538
3539int scan_unevictable_register_node(struct node *node)
3540{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003541 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003542}
3543
3544void scan_unevictable_unregister_node(struct node *node)
3545{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003546 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003547}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003548#endif