blob: fa6a85378ee41400985aca748cd76b59c0f87594 [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>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070022#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070023#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/file.h>
25#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51
Nick Piggin0f8053a2006-03-22 00:08:33 -080052#include "internal.h"
53
Mel Gorman33906bc2010-08-09 17:19:16 -070054#define CREATE_TRACE_POINTS
55#include <trace/events/vmscan.h>
56
Linus Torvalds1da177e2005-04-16 15:20:36 -070057struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned;
60
Rik van Riela79311c2009-01-06 14:40:01 -080061 /* Number of pages freed so far during a call to shrink_zones() */
62 unsigned long nr_reclaimed;
63
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080064 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim;
66
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080067 unsigned long hibernation_mode;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040070 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071
72 int may_writepage;
73
Johannes Weinera6dc60f82009-03-31 15:19:30 -070074 /* Can mapped pages be reclaimed? */
75 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080076
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070077 /* Can pages be swapped as part of reclaim? */
78 int may_swap;
79
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070080 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080081
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070082 /* Scan (total_size >> priority) pages at once */
83 int priority;
84
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070085 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080086 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
88 */
89 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080090
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070091 /*
92 * Nodemask of nodes allowed by the caller. If NULL, all nodes
93 * are scanned.
94 */
95 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096};
97
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99
100#ifdef ARCH_HAS_PREFETCH
101#define prefetch_prev_lru_page(_page, _base, _field) \
102 do { \
103 if ((_page)->lru.prev != _base) { \
104 struct page *prev; \
105 \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
108 } \
109 } while (0)
110#else
111#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
112#endif
113
114#ifdef ARCH_HAS_PREFETCHW
115#define prefetchw_prev_lru_page(_page, _base, _field) \
116 do { \
117 if ((_page)->lru.prev != _base) { \
118 struct page *prev; \
119 \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
122 } \
123 } while (0)
124#else
125#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
126#endif
127
128/*
129 * From 0 .. 100. Higher means more swappy.
130 */
131int vm_swappiness = 60;
Zhang Yanfeib21e0b92013-02-22 16:35:48 -0800132unsigned long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133
134static LIST_HEAD(shrinker_list);
135static DECLARE_RWSEM(shrinker_rwsem);
136
Andrew Mortonc255a452012-07-31 16:43:02 -0700137#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800138static bool global_reclaim(struct scan_control *sc)
139{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800140 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800141}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800142#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800143static bool global_reclaim(struct scan_control *sc)
144{
145 return true;
146}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800147#endif
148
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700149static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800150{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700151 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700152 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800153
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700154 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800155}
156
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157/*
158 * Add a shrinker callback to be called from the vm
159 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700160void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161{
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800162 atomic_long_set(&shrinker->nr_in_batch, 0);
Rusty Russell8e1f9362007-07-17 04:03:17 -0700163 down_write(&shrinker_rwsem);
164 list_add_tail(&shrinker->list, &shrinker_list);
165 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700167EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169/*
170 * Remove one
171 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700172void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173{
174 down_write(&shrinker_rwsem);
175 list_del(&shrinker->list);
176 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700178EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Ying Han1495f232011-05-24 17:12:27 -0700180static inline int do_shrinker_shrink(struct shrinker *shrinker,
181 struct shrink_control *sc,
182 unsigned long nr_to_scan)
183{
184 sc->nr_to_scan = nr_to_scan;
185 return (*shrinker->shrink)(shrinker, sc);
186}
187
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188#define SHRINK_BATCH 128
189/*
190 * Call the shrink functions to age shrinkable caches
191 *
192 * Here we assume it costs one seek to replace a lru page and that it also
193 * takes a seek to recreate a cache object. With this in mind we age equal
194 * percentages of the lru and ageable caches. This should balance the seeks
195 * generated by these structures.
196 *
Simon Arlott183ff222007-10-20 01:27:18 +0200197 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 * slab to avoid swapping.
199 *
200 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
201 *
202 * `lru_pages' represents the number of on-LRU pages in all the zones which
203 * are eligible for the caller's allocation attempt. It is used for balancing
204 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700205 *
206 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207 */
Ying Hana09ed5e2011-05-24 17:12:26 -0700208unsigned long shrink_slab(struct shrink_control *shrink,
Ying Han1495f232011-05-24 17:12:27 -0700209 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700210 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211{
212 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800213 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
Ying Han1495f232011-05-24 17:12:27 -0700215 if (nr_pages_scanned == 0)
216 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
Minchan Kimf06590b2011-05-24 17:11:11 -0700218 if (!down_read_trylock(&shrinker_rwsem)) {
219 /* Assume we'll be able to shrink next time */
220 ret = 1;
221 goto out;
222 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223
224 list_for_each_entry(shrinker, &shrinker_list, list) {
225 unsigned long long delta;
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800226 long total_scan;
227 long max_pass;
Dave Chinner09576072011-07-08 14:14:34 +1000228 int shrink_ret = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000229 long nr;
230 long new_nr;
Dave Chinnere9299f52011-07-08 14:14:37 +1000231 long batch_size = shrinker->batch ? shrinker->batch
232 : SHRINK_BATCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233
Konstantin Khlebnikov635697c2011-12-08 14:33:51 -0800234 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
235 if (max_pass <= 0)
236 continue;
237
Dave Chinneracf92b42011-07-08 14:14:35 +1000238 /*
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
242 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800243 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
Dave Chinneracf92b42011-07-08 14:14:35 +1000244
245 total_scan = nr;
Ying Han1495f232011-05-24 17:12:27 -0700246 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800247 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 do_div(delta, lru_pages + 1);
Dave Chinneracf92b42011-07-08 14:14:35 +1000249 total_scan += delta;
250 if (total_scan < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700251 printk(KERN_ERR "shrink_slab: %pF negative objects to "
252 "delete nr=%ld\n",
Dave Chinneracf92b42011-07-08 14:14:35 +1000253 shrinker->shrink, total_scan);
254 total_scan = max_pass;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800255 }
256
257 /*
Dave Chinner3567b592011-07-08 14:14:36 +1000258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * max_pass. This is bad for sustaining a working set in
264 * memory.
265 *
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
268 */
269 if (delta < max_pass / 4)
270 total_scan = min(total_scan, max_pass / 2);
271
272 /*
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
275 * freeable entries.
276 */
Dave Chinneracf92b42011-07-08 14:14:35 +1000277 if (total_scan > max_pass * 2)
278 total_scan = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Dave Chinneracf92b42011-07-08 14:14:35 +1000280 trace_mm_shrink_slab_start(shrinker, shrink, nr,
Dave Chinner09576072011-07-08 14:14:34 +1000281 nr_pages_scanned, lru_pages,
282 max_pass, delta, total_scan);
283
Dave Chinnere9299f52011-07-08 14:14:37 +1000284 while (total_scan >= batch_size) {
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700285 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
Ying Han1495f232011-05-24 17:12:27 -0700287 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
288 shrink_ret = do_shrinker_shrink(shrinker, shrink,
Dave Chinnere9299f52011-07-08 14:14:37 +1000289 batch_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 if (shrink_ret == -1)
291 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700292 if (shrink_ret < nr_before)
293 ret += nr_before - shrink_ret;
Dave Chinnere9299f52011-07-08 14:14:37 +1000294 count_vm_events(SLABS_SCANNED, batch_size);
295 total_scan -= batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297 cond_resched();
298 }
299
Dave Chinneracf92b42011-07-08 14:14:35 +1000300 /*
301 * move the unused scan count back into the shrinker in a
302 * manner that handles concurrent updates. If we exhausted the
303 * scan, there is no need to do an update.
304 */
Konstantin Khlebnikov83aeead2011-12-08 14:33:54 -0800305 if (total_scan > 0)
306 new_nr = atomic_long_add_return(total_scan,
307 &shrinker->nr_in_batch);
308 else
309 new_nr = atomic_long_read(&shrinker->nr_in_batch);
Dave Chinneracf92b42011-07-08 14:14:35 +1000310
311 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 }
313 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700314out:
315 cond_resched();
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700316 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317}
318
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319static inline int is_page_cache_freeable(struct page *page)
320{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700321 /*
322 * A freeable page cache page is referenced only by the caller
323 * that isolated the page, the page cache radix tree and
324 * optional buffer heads at page->private.
325 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700326 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327}
328
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700329static int may_write_to_queue(struct backing_dev_info *bdi,
330 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
Christoph Lameter930d9152006-01-08 01:00:47 -0800332 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 return 1;
334 if (!bdi_write_congested(bdi))
335 return 1;
336 if (bdi == current->backing_dev_info)
337 return 1;
338 return 0;
339}
340
341/*
342 * We detected a synchronous write error writing a page out. Probably
343 * -ENOSPC. We need to propagate that into the address_space for a subsequent
344 * fsync(), msync() or close().
345 *
346 * The tricky part is that after writepage we cannot touch the mapping: nothing
347 * prevents it from being freed up. But we have a ref on the page and once
348 * that page is locked, the mapping is pinned.
349 *
350 * We're allowed to run sleeping lock_page() here because we know the caller has
351 * __GFP_FS.
352 */
353static void handle_write_error(struct address_space *mapping,
354 struct page *page, int error)
355{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100356 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700357 if (page_mapping(page) == mapping)
358 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 unlock_page(page);
360}
361
Christoph Lameter04e62a22006-06-23 02:03:38 -0700362/* possible outcome of pageout() */
363typedef enum {
364 /* failed to write page out, page is locked */
365 PAGE_KEEP,
366 /* move page to the active list, page is locked */
367 PAGE_ACTIVATE,
368 /* page has been sent to the disk successfully, page is unlocked */
369 PAGE_SUCCESS,
370 /* page is clean and locked */
371 PAGE_CLEAN,
372} pageout_t;
373
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374/*
Andrew Morton1742f192006-03-22 00:08:21 -0800375 * pageout is called by shrink_page_list() for each dirty page.
376 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700378static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700379 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
381 /*
382 * If the page is dirty, only perform writeback if that write
383 * will be non-blocking. To prevent this allocation from being
384 * stalled by pagecache activity. But note that there may be
385 * stalls if we need to run get_block(). We could test
386 * PagePrivate for that.
387 *
Vincent Li6aceb532009-12-14 17:58:49 -0800388 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 * this page's queue, we can perform writeback even if that
390 * will block.
391 *
392 * If the page is swapcache, write it back even if that would
393 * block, for some throttling. This happens by accident, because
394 * swap_backing_dev_info is bust: it doesn't reflect the
395 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 */
397 if (!is_page_cache_freeable(page))
398 return PAGE_KEEP;
399 if (!mapping) {
400 /*
401 * Some data journaling orphaned pages can have
402 * page->mapping == NULL while being dirty with clean buffers.
403 */
David Howells266cf652009-04-03 16:42:36 +0100404 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 if (try_to_free_buffers(page)) {
406 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700407 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 return PAGE_CLEAN;
409 }
410 }
411 return PAGE_KEEP;
412 }
413 if (mapping->a_ops->writepage == NULL)
414 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700415 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 return PAGE_KEEP;
417
418 if (clear_page_dirty_for_io(page)) {
419 int res;
420 struct writeback_control wbc = {
421 .sync_mode = WB_SYNC_NONE,
422 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700423 .range_start = 0,
424 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 .for_reclaim = 1,
426 };
427
428 SetPageReclaim(page);
429 res = mapping->a_ops->writepage(page, &wbc);
430 if (res < 0)
431 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800432 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 ClearPageReclaim(page);
434 return PAGE_ACTIVATE;
435 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700436
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 if (!PageWriteback(page)) {
438 /* synchronous write or broken a_ops? */
439 ClearPageReclaim(page);
440 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700441 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700442 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 return PAGE_SUCCESS;
444 }
445
446 return PAGE_CLEAN;
447}
448
Andrew Mortona649fd92006-10-17 00:09:36 -0700449/*
Nick Piggine2867812008-07-25 19:45:30 -0700450 * Same as remove_mapping, but if the page is removed from the mapping, it
451 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700452 */
Nick Piggine2867812008-07-25 19:45:30 -0700453static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800454{
Nick Piggin28e4d962006-09-25 23:31:23 -0700455 BUG_ON(!PageLocked(page));
456 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800457
Nick Piggin19fd6232008-07-25 19:45:32 -0700458 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800459 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700460 * The non racy check for a busy page.
461 *
462 * Must be careful with the order of the tests. When someone has
463 * a ref to the page, it may be possible that they dirty it then
464 * drop the reference. So if PageDirty is tested before page_count
465 * here, then the following race may occur:
466 *
467 * get_user_pages(&page);
468 * [user mapping goes away]
469 * write_to(page);
470 * !PageDirty(page) [good]
471 * SetPageDirty(page);
472 * put_page(page);
473 * !page_count(page) [good, discard it]
474 *
475 * [oops, our write_to data is lost]
476 *
477 * Reversing the order of the tests ensures such a situation cannot
478 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
479 * load is not satisfied before that of page->_count.
480 *
481 * Note that if SetPageDirty is always performed via set_page_dirty,
482 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800483 */
Nick Piggine2867812008-07-25 19:45:30 -0700484 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800485 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700486 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
487 if (unlikely(PageDirty(page))) {
488 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800489 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700490 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800491
492 if (PageSwapCache(page)) {
493 swp_entry_t swap = { .val = page_private(page) };
494 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700495 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700496 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700497 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500498 void (*freepage)(struct page *);
499
500 freepage = mapping->a_ops->freepage;
501
Minchan Kime64a7822011-03-22 16:32:44 -0700502 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700503 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700504 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500505
506 if (freepage != NULL)
507 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800508 }
509
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800510 return 1;
511
512cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700513 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800514 return 0;
515}
516
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517/*
Nick Piggine2867812008-07-25 19:45:30 -0700518 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
519 * someone else has a ref on the page, abort and return 0. If it was
520 * successfully detached, return 1. Assumes the caller has a single ref on
521 * this page.
522 */
523int remove_mapping(struct address_space *mapping, struct page *page)
524{
525 if (__remove_mapping(mapping, page)) {
526 /*
527 * Unfreezing the refcount with 1 rather than 2 effectively
528 * drops the pagecache ref for us without requiring another
529 * atomic operation.
530 */
531 page_unfreeze_refs(page, 1);
532 return 1;
533 }
534 return 0;
535}
536
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700537/**
538 * putback_lru_page - put previously isolated page onto appropriate LRU list
539 * @page: page to be put back to appropriate lru list
540 *
541 * Add previously isolated @page to appropriate LRU list.
542 * Page may still be unevictable for other reasons.
543 *
544 * lru_lock must not be held, interrupts must be enabled.
545 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700546void putback_lru_page(struct page *page)
547{
548 int lru;
549 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700550 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700551
552 VM_BUG_ON(PageLRU(page));
553
554redo:
555 ClearPageUnevictable(page);
556
Hugh Dickins39b5f292012-10-08 16:33:18 -0700557 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700558 /*
559 * For evictable pages, we can use the cache.
560 * In event of a race, worst case is we end up with an
561 * unevictable page on [in]active list.
562 * We know how to handle that.
563 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700564 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700565 lru_cache_add_lru(page, lru);
566 } else {
567 /*
568 * Put unevictable pages directly on zone's unevictable
569 * list.
570 */
571 lru = LRU_UNEVICTABLE;
572 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700573 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700574 * When racing with an mlock or AS_UNEVICTABLE clearing
575 * (page is unlocked) make sure that if the other thread
576 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800577 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700578 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700579 * the page back to the evictable list.
580 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700581 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700582 */
583 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700584 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700585
586 /*
587 * page's status can change while we move it among lru. If an evictable
588 * page is on unevictable list, it never be freed. To avoid that,
589 * check after we added it to the list, again.
590 */
Hugh Dickins39b5f292012-10-08 16:33:18 -0700591 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700592 if (!isolate_lru_page(page)) {
593 put_page(page);
594 goto redo;
595 }
596 /* This means someone else dropped this page from LRU
597 * So, it will be freed or putback to LRU again. There is
598 * nothing to do here.
599 */
600 }
601
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700602 if (was_unevictable && lru != LRU_UNEVICTABLE)
603 count_vm_event(UNEVICTABLE_PGRESCUED);
604 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
605 count_vm_event(UNEVICTABLE_PGCULLED);
606
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700607 put_page(page); /* drop ref from isolate */
608}
609
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800610enum page_references {
611 PAGEREF_RECLAIM,
612 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800613 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800614 PAGEREF_ACTIVATE,
615};
616
617static enum page_references page_check_references(struct page *page,
618 struct scan_control *sc)
619{
Johannes Weiner645747462010-03-05 13:42:22 -0800620 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800621 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800622
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700623 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
624 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800625 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800626
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800627 /*
628 * Mlock lost the isolation race with us. Let try_to_unmap()
629 * move the page to the unevictable list.
630 */
631 if (vm_flags & VM_LOCKED)
632 return PAGEREF_RECLAIM;
633
Johannes Weiner645747462010-03-05 13:42:22 -0800634 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700635 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800636 return PAGEREF_ACTIVATE;
637 /*
638 * All mapped pages start out with page table
639 * references from the instantiating fault, so we need
640 * to look twice if a mapped file page is used more
641 * than once.
642 *
643 * Mark it and spare it for another trip around the
644 * inactive list. Another page table reference will
645 * lead to its activation.
646 *
647 * Note: the mark is set for activated pages as well
648 * so that recently deactivated but used pages are
649 * quickly recovered.
650 */
651 SetPageReferenced(page);
652
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800653 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800654 return PAGEREF_ACTIVATE;
655
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800656 /*
657 * Activate file-backed executable pages after first usage.
658 */
659 if (vm_flags & VM_EXEC)
660 return PAGEREF_ACTIVATE;
661
Johannes Weiner645747462010-03-05 13:42:22 -0800662 return PAGEREF_KEEP;
663 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800664
665 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700666 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800667 return PAGEREF_RECLAIM_CLEAN;
668
669 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800670}
671
Nick Piggine2867812008-07-25 19:45:30 -0700672/*
Andrew Morton1742f192006-03-22 00:08:21 -0800673 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 */
Andrew Morton1742f192006-03-22 00:08:21 -0800675static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700676 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700677 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700678 enum ttu_flags ttu_flags,
Mel Gorman92df3a72011-10-31 17:07:56 -0700679 unsigned long *ret_nr_dirty,
Minchan Kim02c6de82012-10-08 16:31:55 -0700680 unsigned long *ret_nr_writeback,
681 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682{
683 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700684 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700686 unsigned long nr_dirty = 0;
687 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800688 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700689 unsigned long nr_writeback = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
691 cond_resched();
692
Tim Chen69980e32012-07-31 16:46:08 -0700693 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 while (!list_empty(page_list)) {
695 struct address_space *mapping;
696 struct page *page;
697 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700698 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
700 cond_resched();
701
702 page = lru_to_page(page_list);
703 list_del(&page->lru);
704
Nick Piggin529ae9a2008-08-02 12:01:03 +0200705 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 goto keep;
707
Nick Piggin725d7042006-09-25 23:30:55 -0700708 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700709 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800712
Hugh Dickins39b5f292012-10-08 16:33:18 -0700713 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700714 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700715
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700716 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800717 goto keep_locked;
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 /* Double the slab pressure for mapped and swapcache pages */
720 if (page_mapped(page) || PageSwapCache(page))
721 sc->nr_scanned++;
722
Andy Whitcroftc661b072007-08-22 14:01:26 -0700723 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
724 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
725
726 if (PageWriteback(page)) {
Michal Hockoe62e3842012-07-31 16:45:55 -0700727 /*
728 * memcg doesn't have any dirty pages throttling so we
729 * could easily OOM just because too many pages are in
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700730 * writeback and there is nothing else to reclaim.
Michal Hockoe62e3842012-07-31 16:45:55 -0700731 *
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700732 * Check __GFP_IO, certainly because a loop driver
Michal Hockoe62e3842012-07-31 16:45:55 -0700733 * thread might enter reclaim, and deadlock if it waits
734 * on a page for which it is needed to do the write
735 * (loop masks off __GFP_IO|__GFP_FS for this reason);
736 * but more thought would probably show more reasons.
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700737 *
738 * Don't require __GFP_FS, since we're not going into
739 * the FS, just waiting on its writeback completion.
740 * Worryingly, ext4 gfs2 and xfs allocate pages with
741 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
742 * testing may_enter_fs here is liable to OOM on them.
Michal Hockoe62e3842012-07-31 16:45:55 -0700743 */
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700744 if (global_reclaim(sc) ||
745 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
746 /*
747 * This is slightly racy - end_page_writeback()
748 * might have just cleared PageReclaim, then
749 * setting PageReclaim here end up interpreted
750 * as PageReadahead - but that does not matter
751 * enough to care. What we do want is for this
752 * page to have PageReclaim set next time memcg
753 * reclaim reaches the tests above, so it will
754 * then wait_on_page_writeback() to avoid OOM;
755 * and it's also appropriate in global reclaim.
756 */
757 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700758 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700759 goto keep_locked;
Michal Hockoe62e3842012-07-31 16:45:55 -0700760 }
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700761 wait_on_page_writeback(page);
Andy Whitcroftc661b072007-08-22 14:01:26 -0700762 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
Minchan Kim02c6de82012-10-08 16:31:55 -0700764 if (!force_reclaim)
765 references = page_check_references(page, sc);
766
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800767 switch (references) {
768 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800770 case PAGEREF_KEEP:
771 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800772 case PAGEREF_RECLAIM:
773 case PAGEREF_RECLAIM_CLEAN:
774 ; /* try to reclaim the page below */
775 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 /*
778 * Anonymous process memory has backing store?
779 * Try to allocate it some swap space here.
780 */
Nick Pigginb291f002008-10-18 20:26:44 -0700781 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800782 if (!(sc->gfp_mask & __GFP_IO))
783 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700784 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800786 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700787 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
789 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
791 /*
792 * The page is mapped into the page tables of one or more
793 * processes. Try to unmap it here.
794 */
795 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700796 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 case SWAP_FAIL:
798 goto activate_locked;
799 case SWAP_AGAIN:
800 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700801 case SWAP_MLOCK:
802 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 case SWAP_SUCCESS:
804 ; /* try to free the page below */
805 }
806 }
807
808 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700809 nr_dirty++;
810
Mel Gormanee728862011-10-31 17:07:38 -0700811 /*
812 * Only kswapd can writeback filesystem pages to
Mel Gormanf84f6e22011-10-31 17:07:51 -0700813 * avoid risk of stack overflow but do not writeback
814 * unless under significant pressure.
Mel Gormanee728862011-10-31 17:07:38 -0700815 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700816 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700817 (!current_is_kswapd() ||
818 sc->priority >= DEF_PRIORITY - 2)) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700819 /*
820 * Immediately reclaim when written back.
821 * Similar in principal to deactivate_page()
822 * except we already have the page isolated
823 * and know it's dirty
824 */
825 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
826 SetPageReclaim(page);
827
Mel Gormanee728862011-10-31 17:07:38 -0700828 goto keep_locked;
829 }
830
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800831 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700833 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800835 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 goto keep_locked;
837
838 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700839 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700841 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 goto keep_locked;
843 case PAGE_ACTIVATE:
844 goto activate_locked;
845 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700846 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700847 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700848 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 /*
852 * A synchronous write - probably a ramdisk. Go
853 * ahead and try to reclaim the page.
854 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200855 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 goto keep;
857 if (PageDirty(page) || PageWriteback(page))
858 goto keep_locked;
859 mapping = page_mapping(page);
860 case PAGE_CLEAN:
861 ; /* try to free the page below */
862 }
863 }
864
865 /*
866 * If the page has buffers, try to free the buffer mappings
867 * associated with this page. If we succeed we try to free
868 * the page as well.
869 *
870 * We do this even if the page is PageDirty().
871 * try_to_release_page() does not perform I/O, but it is
872 * possible for a page to have PageDirty set, but it is actually
873 * clean (all its buffers are clean). This happens if the
874 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700875 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 * try_to_release_page() will discover that cleanness and will
877 * drop the buffers and mark the page clean - it can be freed.
878 *
879 * Rarely, pages can have buffers and no ->mapping. These are
880 * the pages which were not successfully invalidated in
881 * truncate_complete_page(). We try to drop those buffers here
882 * and if that worked, and the page is no longer mapped into
883 * process address space (page_count == 1) it can be freed.
884 * Otherwise, leave the page on the LRU so it is swappable.
885 */
David Howells266cf652009-04-03 16:42:36 +0100886 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887 if (!try_to_release_page(page, sc->gfp_mask))
888 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700889 if (!mapping && page_count(page) == 1) {
890 unlock_page(page);
891 if (put_page_testzero(page))
892 goto free_it;
893 else {
894 /*
895 * rare race with speculative reference.
896 * the speculative reference will free
897 * this page shortly, so we may
898 * increment nr_reclaimed here (and
899 * leave it off the LRU).
900 */
901 nr_reclaimed++;
902 continue;
903 }
904 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 }
906
Nick Piggine2867812008-07-25 19:45:30 -0700907 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800908 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909
Nick Piggina978d6f2008-10-18 20:26:58 -0700910 /*
911 * At this point, we have no other references and there is
912 * no way to pick any more up (removed from LRU, removed
913 * from pagecache). Can use non-atomic bitops now (and
914 * we obviously don't have to worry about waking up a process
915 * waiting on the page lock, because there are no references.
916 */
917 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700918free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800919 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700920
921 /*
922 * Is there need to periodically free_page_list? It would
923 * appear not as the counts should be low
924 */
925 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 continue;
927
Nick Pigginb291f002008-10-18 20:26:44 -0700928cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800929 if (PageSwapCache(page))
930 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700931 unlock_page(page);
932 putback_lru_page(page);
933 continue;
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700936 /* Not a candidate for swapping, so reclaim swap space. */
937 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800938 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700939 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 SetPageActive(page);
941 pgactivate++;
942keep_locked:
943 unlock_page(page);
944keep:
945 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700946 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700948
Mel Gorman0e093d992010-10-26 14:21:45 -0700949 /*
950 * Tag a zone as congested if all the dirty pages encountered were
951 * backed by a congested BDI. In this case, reclaimers should just
952 * back off and wait for congestion to clear because further reclaim
953 * will encounter the same problem
954 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800955 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700956 zone_set_flag(zone, ZONE_CONGESTED);
Mel Gorman0e093d992010-10-26 14:21:45 -0700957
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -0800958 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700959
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700961 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -0700962 mem_cgroup_uncharge_end();
Mel Gorman92df3a72011-10-31 17:07:56 -0700963 *ret_nr_dirty += nr_dirty;
964 *ret_nr_writeback += nr_writeback;
Andrew Morton05ff5132006-03-22 00:08:20 -0800965 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966}
967
Minchan Kim02c6de82012-10-08 16:31:55 -0700968unsigned long reclaim_clean_pages_from_list(struct zone *zone,
969 struct list_head *page_list)
970{
971 struct scan_control sc = {
972 .gfp_mask = GFP_KERNEL,
973 .priority = DEF_PRIORITY,
974 .may_unmap = 1,
975 };
976 unsigned long ret, dummy1, dummy2;
977 struct page *page, *next;
978 LIST_HEAD(clean_pages);
979
980 list_for_each_entry_safe(page, next, page_list, lru) {
981 if (page_is_file_cache(page) && !PageDirty(page)) {
982 ClearPageActive(page);
983 list_move(&page->lru, &clean_pages);
984 }
985 }
986
987 ret = shrink_page_list(&clean_pages, zone, &sc,
988 TTU_UNMAP|TTU_IGNORE_ACCESS,
989 &dummy1, &dummy2, true);
990 list_splice(&clean_pages, page_list);
991 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
992 return ret;
993}
994
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700995/*
996 * Attempt to remove the specified page from its LRU. Only take this page
997 * if it is of the appropriate PageActive status. Pages which are being
998 * freed elsewhere are also ignored.
999 *
1000 * page: page to consider
1001 * mode: one of the LRU isolation modes defined above
1002 *
1003 * returns 0 on success, -ve errno on failure.
1004 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001005int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001006{
1007 int ret = -EINVAL;
1008
1009 /* Only take pages on the LRU. */
1010 if (!PageLRU(page))
1011 return ret;
1012
Minchan Kime46a2872012-10-08 16:33:48 -07001013 /* Compaction should not handle unevictable pages but CMA can do so */
1014 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001015 return ret;
1016
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001017 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001018
Mel Gormanc8244932012-01-12 17:19:38 -08001019 /*
1020 * To minimise LRU disruption, the caller can indicate that it only
1021 * wants to isolate pages it will be able to operate on without
1022 * blocking - clean pages for the most part.
1023 *
1024 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1025 * is used by reclaim when it is cannot write to backing storage
1026 *
1027 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1028 * that it is possible to migrate without blocking
1029 */
1030 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1031 /* All the caller can do on PageWriteback is block */
1032 if (PageWriteback(page))
1033 return ret;
1034
1035 if (PageDirty(page)) {
1036 struct address_space *mapping;
1037
1038 /* ISOLATE_CLEAN means only clean pages */
1039 if (mode & ISOLATE_CLEAN)
1040 return ret;
1041
1042 /*
1043 * Only pages without mappings or that have a
1044 * ->migratepage callback are possible to migrate
1045 * without blocking
1046 */
1047 mapping = page_mapping(page);
1048 if (mapping && !mapping->a_ops->migratepage)
1049 return ret;
1050 }
1051 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001052
Minchan Kimf80c0672011-10-31 17:06:55 -07001053 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1054 return ret;
1055
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001056 if (likely(get_page_unless_zero(page))) {
1057 /*
1058 * Be careful not to clear PageLRU until after we're
1059 * sure the page is not being freed elsewhere -- the
1060 * page release code relies on it.
1061 */
1062 ClearPageLRU(page);
1063 ret = 0;
1064 }
1065
1066 return ret;
1067}
1068
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001069/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 * zone->lru_lock is heavily contended. Some of the functions that
1071 * shrink the lists perform better by taking out a batch of pages
1072 * and working on them outside the LRU lock.
1073 *
1074 * For pagecache intensive workloads, this function is the hottest
1075 * spot in the kernel (apart from copy_*_user functions).
1076 *
1077 * Appropriate locks must be held before calling this function.
1078 *
1079 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001080 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001082 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001083 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001084 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001085 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 *
1087 * returns how many pages were moved onto *@dst.
1088 */
Andrew Morton69e05942006-03-22 00:08:19 -08001089static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001090 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001091 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001092 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001094 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001095 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001096 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001098 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001099 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001100 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001101
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 page = lru_to_page(src);
1103 prefetchw_prev_lru_page(page, src, flags);
1104
Nick Piggin725d7042006-09-25 23:30:55 -07001105 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001106
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001107 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001108 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001109 nr_pages = hpage_nr_pages(page);
1110 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001111 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001112 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001113 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001114
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001115 case -EBUSY:
1116 /* else it is being freed elsewhere */
1117 list_move(&page->lru, src);
1118 continue;
1119
1120 default:
1121 BUG();
1122 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 }
1124
Hugh Dickinsf6260122012-01-12 17:20:06 -08001125 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001126 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1127 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 return nr_taken;
1129}
1130
Nick Piggin62695a82008-10-18 20:26:09 -07001131/**
1132 * isolate_lru_page - tries to isolate a page from its LRU list
1133 * @page: page to isolate from its LRU list
1134 *
1135 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1136 * vmstat statistic corresponding to whatever LRU list the page was on.
1137 *
1138 * Returns 0 if the page was removed from an LRU list.
1139 * Returns -EBUSY if the page was not on an LRU list.
1140 *
1141 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001142 * the active list, it will have PageActive set. If it was found on
1143 * the unevictable list, it will have the PageUnevictable bit set. That flag
1144 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001145 *
1146 * The vmstat statistic corresponding to the list on which the page was
1147 * found will be decremented.
1148 *
1149 * Restrictions:
1150 * (1) Must be called with an elevated refcount on the page. This is a
1151 * fundamentnal difference from isolate_lru_pages (which is called
1152 * without a stable reference).
1153 * (2) the lru_lock must not be held.
1154 * (3) interrupts must be enabled.
1155 */
1156int isolate_lru_page(struct page *page)
1157{
1158 int ret = -EBUSY;
1159
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001160 VM_BUG_ON(!page_count(page));
1161
Nick Piggin62695a82008-10-18 20:26:09 -07001162 if (PageLRU(page)) {
1163 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001164 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001165
1166 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001167 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001168 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001169 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001170 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001171 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001172 del_page_from_lru_list(page, lruvec, lru);
1173 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001174 }
1175 spin_unlock_irq(&zone->lru_lock);
1176 }
1177 return ret;
1178}
1179
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001180/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001181 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1182 * then get resheduled. When there are massive number of tasks doing page
1183 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1184 * the LRU list will go small and be scanned faster than necessary, leading to
1185 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001186 */
1187static int too_many_isolated(struct zone *zone, int file,
1188 struct scan_control *sc)
1189{
1190 unsigned long inactive, isolated;
1191
1192 if (current_is_kswapd())
1193 return 0;
1194
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001195 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001196 return 0;
1197
1198 if (file) {
1199 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1200 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1201 } else {
1202 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1203 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1204 }
1205
Fengguang Wu3cf23842012-12-18 14:23:31 -08001206 /*
1207 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1208 * won't get blocked by normal direct-reclaimers, forming a circular
1209 * deadlock.
1210 */
1211 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1212 inactive >>= 3;
1213
Rik van Riel35cd7812009-09-21 17:01:38 -07001214 return isolated > inactive;
1215}
1216
Mel Gorman66635622010-08-09 17:19:30 -07001217static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001218putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001219{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001220 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1221 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001222 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001223
Mel Gorman66635622010-08-09 17:19:30 -07001224 /*
1225 * Put back any unfreeable pages.
1226 */
Mel Gorman66635622010-08-09 17:19:30 -07001227 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001228 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001229 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001230
Mel Gorman66635622010-08-09 17:19:30 -07001231 VM_BUG_ON(PageLRU(page));
1232 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001233 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001234 spin_unlock_irq(&zone->lru_lock);
1235 putback_lru_page(page);
1236 spin_lock_irq(&zone->lru_lock);
1237 continue;
1238 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001239
1240 lruvec = mem_cgroup_page_lruvec(page, zone);
1241
Linus Torvalds7a608572011-01-17 14:42:19 -08001242 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001243 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001244 add_page_to_lru_list(page, lruvec, lru);
1245
Mel Gorman66635622010-08-09 17:19:30 -07001246 if (is_active_lru(lru)) {
1247 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001248 int numpages = hpage_nr_pages(page);
1249 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001250 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001251 if (put_page_testzero(page)) {
1252 __ClearPageLRU(page);
1253 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001254 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001255
1256 if (unlikely(PageCompound(page))) {
1257 spin_unlock_irq(&zone->lru_lock);
1258 (*get_compound_page_dtor(page))(page);
1259 spin_lock_irq(&zone->lru_lock);
1260 } else
1261 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001262 }
1263 }
Mel Gorman66635622010-08-09 17:19:30 -07001264
Hugh Dickins3f797682012-01-12 17:20:07 -08001265 /*
1266 * To save our caller's stack, now use input list for pages to free.
1267 */
1268 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001269}
1270
1271/*
Andrew Morton1742f192006-03-22 00:08:21 -08001272 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1273 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 */
Mel Gorman66635622010-08-09 17:19:30 -07001275static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001276shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001277 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278{
1279 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001280 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001281 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001282 unsigned long nr_taken;
Mel Gorman92df3a72011-10-31 17:07:56 -07001283 unsigned long nr_dirty = 0;
1284 unsigned long nr_writeback = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001285 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001286 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001287 struct zone *zone = lruvec_zone(lruvec);
1288 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001289
Rik van Riel35cd7812009-09-21 17:01:38 -07001290 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001291 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001292
1293 /* We are about to die and free our memory. Return now. */
1294 if (fatal_signal_pending(current))
1295 return SWAP_CLUSTER_MAX;
1296 }
1297
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001299
1300 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001301 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001302 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001303 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001304
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001307 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1308 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001309
1310 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1311 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1312
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001313 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001314 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001315 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001316 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001317 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001318 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001319 }
Hillf Dantond563c052012-03-21 16:34:02 -07001320 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001321
Hillf Dantond563c052012-03-21 16:34:02 -07001322 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001323 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001324
Minchan Kim02c6de82012-10-08 16:31:55 -07001325 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1326 &nr_dirty, &nr_writeback, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001327
Hugh Dickins3f797682012-01-12 17:20:07 -08001328 spin_lock_irq(&zone->lru_lock);
1329
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001330 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001331
Ying Han904249a2012-04-25 16:01:48 -07001332 if (global_reclaim(sc)) {
1333 if (current_is_kswapd())
1334 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1335 nr_reclaimed);
1336 else
1337 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1338 nr_reclaimed);
1339 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001340
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001341 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001342
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001343 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001344
1345 spin_unlock_irq(&zone->lru_lock);
1346
1347 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001348
Mel Gorman92df3a72011-10-31 17:07:56 -07001349 /*
1350 * If reclaim is isolating dirty pages under writeback, it implies
1351 * that the long-lived page allocation rate is exceeding the page
1352 * laundering rate. Either the global limits are not being effective
1353 * at throttling processes due to the page distribution throughout
1354 * zones or there is heavy usage of a slow backing device. The
1355 * only option is to throttle from reclaim context which is not ideal
1356 * as there is no guarantee the dirtying process is throttled in the
1357 * same way balance_dirty_pages() manages.
1358 *
1359 * This scales the number of dirty pages that must be under writeback
1360 * before throttling depending on priority. It is a simple backoff
1361 * function that has the most effect in the range DEF_PRIORITY to
1362 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1363 * in trouble and reclaim is considered to be in trouble.
1364 *
1365 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1366 * DEF_PRIORITY-1 50% must be PageWriteback
1367 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1368 * ...
1369 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1370 * isolated page is PageWriteback
1371 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001372 if (nr_writeback && nr_writeback >=
1373 (nr_taken >> (DEF_PRIORITY - sc->priority)))
Mel Gorman92df3a72011-10-31 17:07:56 -07001374 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1375
Mel Gormane11da5b2010-10-26 14:21:40 -07001376 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1377 zone_idx(zone),
1378 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001379 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001380 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001381 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382}
1383
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001384/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 * This moves pages from the active list to the inactive list.
1386 *
1387 * We move them the other way if the page is referenced by one or more
1388 * processes, from rmap.
1389 *
1390 * If the pages are mostly unmapped, the processing is fast and it is
1391 * appropriate to hold zone->lru_lock across the whole operation. But if
1392 * the pages are mapped, the processing is slow (page_referenced()) so we
1393 * should drop zone->lru_lock around each page. It's impossible to balance
1394 * this, so instead we remove the pages from the LRU while processing them.
1395 * It is safe to rely on PG_active against the non-LRU pages in here because
1396 * nobody will play with that bit on a non-LRU page.
1397 *
1398 * The downside is that we have to touch page->_count against each page.
1399 * But we had to alter page->flags anyway.
1400 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001401
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001402static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001403 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001404 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001405 enum lru_list lru)
1406{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001407 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001408 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001409 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001410 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001411
Wu Fengguang3eb41402009-06-16 15:33:13 -07001412 while (!list_empty(list)) {
1413 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001414 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001415
1416 VM_BUG_ON(PageLRU(page));
1417 SetPageLRU(page);
1418
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001419 nr_pages = hpage_nr_pages(page);
1420 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001421 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001422 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001423
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001424 if (put_page_testzero(page)) {
1425 __ClearPageLRU(page);
1426 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001427 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001428
1429 if (unlikely(PageCompound(page))) {
1430 spin_unlock_irq(&zone->lru_lock);
1431 (*get_compound_page_dtor(page))(page);
1432 spin_lock_irq(&zone->lru_lock);
1433 } else
1434 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001435 }
1436 }
1437 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1438 if (!is_active_lru(lru))
1439 __count_vm_events(PGDEACTIVATE, pgmoved);
1440}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001441
Hugh Dickinsf6260122012-01-12 17:20:06 -08001442static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001443 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001444 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001445 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001447 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001448 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001449 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001451 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001452 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001454 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001455 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001456 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001457 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001458 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459
1460 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001461
1462 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001463 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001464 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001465 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001468
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001469 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1470 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001471 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001472 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001473
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001474 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001475
Hugh Dickinsf6260122012-01-12 17:20:06 -08001476 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001477 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001478 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 spin_unlock_irq(&zone->lru_lock);
1480
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 while (!list_empty(&l_hold)) {
1482 cond_resched();
1483 page = lru_to_page(&l_hold);
1484 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001485
Hugh Dickins39b5f292012-10-08 16:33:18 -07001486 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001487 putback_lru_page(page);
1488 continue;
1489 }
1490
Mel Gormancc715d92012-03-21 16:34:00 -07001491 if (unlikely(buffer_heads_over_limit)) {
1492 if (page_has_private(page) && trylock_page(page)) {
1493 if (page_has_private(page))
1494 try_to_release_page(page, 0);
1495 unlock_page(page);
1496 }
1497 }
1498
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001499 if (page_referenced(page, 0, sc->target_mem_cgroup,
1500 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001501 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001502 /*
1503 * Identify referenced, file-backed active pages and
1504 * give them one more trip around the active list. So
1505 * that executable code get better chances to stay in
1506 * memory under moderate memory pressure. Anon pages
1507 * are not likely to be evicted by use-once streaming
1508 * IO, plus JVM can create lots of anon VM_EXEC pages,
1509 * so we ignore them here.
1510 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001511 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001512 list_add(&page->lru, &l_active);
1513 continue;
1514 }
1515 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001516
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001517 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 list_add(&page->lru, &l_inactive);
1519 }
1520
Andrew Mortonb5557492009-01-06 14:40:13 -08001521 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001522 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001523 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001524 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001525 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001526 * Count referenced pages from currently used mappings as rotated,
1527 * even though only some of them are actually re-activated. This
1528 * helps balance scan pressure between file and anonymous pages in
1529 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001530 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001531 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001532
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001533 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1534 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001535 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001536 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001537
1538 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539}
1540
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001541#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001542static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001543{
1544 unsigned long active, inactive;
1545
1546 active = zone_page_state(zone, NR_ACTIVE_ANON);
1547 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1548
1549 if (inactive * zone->inactive_ratio < active)
1550 return 1;
1551
1552 return 0;
1553}
1554
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001555/**
1556 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001557 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001558 *
1559 * Returns true if the zone does not have enough inactive anon pages,
1560 * meaning some active anon pages need to be deactivated.
1561 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001562static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001563{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001564 /*
1565 * If we don't have swap space, anonymous page deactivation
1566 * is pointless.
1567 */
1568 if (!total_swap_pages)
1569 return 0;
1570
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001571 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001572 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001573
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001574 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001575}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001576#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001577static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001578{
1579 return 0;
1580}
1581#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001582
Rik van Riel56e49d22009-06-16 15:32:28 -07001583/**
1584 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001585 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001586 *
1587 * When the system is doing streaming IO, memory pressure here
1588 * ensures that active file pages get deactivated, until more
1589 * than half of the file pages are on the inactive list.
1590 *
1591 * Once we get to that situation, protect the system's working
1592 * set from being evicted by disabling active file page aging.
1593 *
1594 * This uses a different ratio than the anonymous pages, because
1595 * the page cache uses a use-once replacement algorithm.
1596 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001597static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001598{
Johannes Weinere3790142013-02-22 16:35:19 -08001599 unsigned long inactive;
1600 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001601
Johannes Weinere3790142013-02-22 16:35:19 -08001602 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1603 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1604
1605 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001606}
1607
Hugh Dickins75b00af2012-05-29 15:07:09 -07001608static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001609{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001610 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001611 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001612 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001613 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001614}
1615
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001616static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001617 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001618{
Rik van Rielb39415b2009-12-14 17:59:48 -08001619 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001620 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001621 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001622 return 0;
1623 }
1624
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001625 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001626}
1627
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001628static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001629{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001630 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001631 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001632 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001633}
1634
Johannes Weiner9a265112013-02-22 16:32:17 -08001635enum scan_balance {
1636 SCAN_EQUAL,
1637 SCAN_FRACT,
1638 SCAN_ANON,
1639 SCAN_FILE,
1640};
1641
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001643 * Determine how aggressively the anon and file LRU lists should be
1644 * scanned. The relative value of each set of LRU lists is determined
1645 * by looking at the fraction of the pages scanned we did rotate back
1646 * onto the active list instead of evict.
1647 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001648 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1649 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001650 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001651static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001652 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001653{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001654 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001655 u64 fraction[2];
1656 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001657 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001658 unsigned long anon_prio, file_prio;
1659 enum scan_balance scan_balance;
1660 unsigned long anon, file, free;
1661 bool force_scan = false;
1662 unsigned long ap, fp;
1663 enum lru_list lru;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001664
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001665 /*
1666 * If the zone or memcg is small, nr[l] can be 0. This
1667 * results in no scanning on this priority and a potential
1668 * priority drop. Global direct reclaim can go to the next
1669 * zone and tends to have no problems. Global kswapd is for
1670 * zone balancing and it needs to scan a minimum amount. When
1671 * reclaiming for a memcg, a priority drop can cause high
1672 * latencies, so it's better to scan a minimum amount there as
1673 * well.
1674 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001675 if (current_is_kswapd() && zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001676 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001677 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001678 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001679
1680 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001681 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001682 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001683 goto out;
1684 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001685
Johannes Weiner10316b32013-02-22 16:32:14 -08001686 /*
1687 * Global reclaim will swap to prevent OOM even with no
1688 * swappiness, but memcg users want to use this knob to
1689 * disable swapping for individual groups completely when
1690 * using the memory controller's swap limit feature would be
1691 * too expensive.
1692 */
1693 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001694 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001695 goto out;
1696 }
1697
1698 /*
1699 * Do not apply any pressure balancing cleverness when the
1700 * system is close to OOM, scan both anon and file equally
1701 * (unless the swappiness setting disagrees with swapping).
1702 */
1703 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001704 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001705 goto out;
1706 }
1707
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001708 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1709 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1710 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1711 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001712
Johannes Weiner11d16c22013-02-22 16:32:15 -08001713 /*
1714 * If it's foreseeable that reclaiming the file cache won't be
1715 * enough to get the zone back into a desirable shape, we have
1716 * to swap. Better start now and leave the - probably heavily
1717 * thrashing - remaining file pages alone.
1718 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001719 if (global_reclaim(sc)) {
Johannes Weiner11d16c22013-02-22 16:32:15 -08001720 free = zone_page_state(zone, NR_FREE_PAGES);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001721 if (unlikely(file + free <= high_wmark_pages(zone))) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001722 scan_balance = SCAN_ANON;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001723 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001724 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001725 }
1726
1727 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001728 * There is enough inactive page cache, do not reclaim
1729 * anything from the anonymous working set right now.
1730 */
1731 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001732 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001733 goto out;
1734 }
1735
Johannes Weiner9a265112013-02-22 16:32:17 -08001736 scan_balance = SCAN_FRACT;
1737
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001738 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001739 * With swappiness at 100, anonymous and file have the same priority.
1740 * This scanning priority is essentially the inverse of IO cost.
1741 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001742 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001743 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001744
1745 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001746 * OK, so we have swap space and a fair amount of page cache
1747 * pages. We use the recently rotated / recently scanned
1748 * ratios to determine how valuable each cache is.
1749 *
1750 * Because workloads change over time (and to avoid overflow)
1751 * we keep these statistics as a floating average, which ends
1752 * up weighing recent references more than old ones.
1753 *
1754 * anon in [0], file in [1]
1755 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001756 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001757 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001758 reclaim_stat->recent_scanned[0] /= 2;
1759 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001760 }
1761
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001762 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001763 reclaim_stat->recent_scanned[1] /= 2;
1764 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001765 }
1766
1767 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001768 * The amount of pressure on anon vs file pages is inversely
1769 * proportional to the fraction of recently scanned pages on
1770 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001771 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001772 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001773 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001774
Satoru Moriyafe350042012-05-29 15:06:47 -07001775 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001776 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001777 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001778
Shaohua Li76a33fc2010-05-24 14:32:36 -07001779 fraction[0] = ap;
1780 fraction[1] = fp;
1781 denominator = ap + fp + 1;
1782out:
Hugh Dickins41113042012-01-12 17:20:01 -08001783 for_each_evictable_lru(lru) {
1784 int file = is_file_lru(lru);
Johannes Weinerd778df52013-02-22 16:32:12 -08001785 unsigned long size;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001786 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001787
Johannes Weinerd778df52013-02-22 16:32:12 -08001788 size = get_lru_size(lruvec, lru);
Johannes Weiner10316b32013-02-22 16:32:14 -08001789 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08001790
Johannes Weiner10316b32013-02-22 16:32:14 -08001791 if (!scan && force_scan)
1792 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08001793
1794 switch (scan_balance) {
1795 case SCAN_EQUAL:
1796 /* Scan lists relative to size */
1797 break;
1798 case SCAN_FRACT:
1799 /*
1800 * Scan types proportional to swappiness and
1801 * their relative recent reclaim efficiency.
1802 */
1803 scan = div64_u64(scan * fraction[file], denominator);
1804 break;
1805 case SCAN_FILE:
1806 case SCAN_ANON:
1807 /* Scan one type exclusively */
1808 if ((scan_balance == SCAN_FILE) != file)
1809 scan = 0;
1810 break;
1811 default:
1812 /* Look ma, no brain */
1813 BUG();
1814 }
Hugh Dickins41113042012-01-12 17:20:01 -08001815 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001816 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001817}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001818
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001819/*
1820 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1821 */
1822static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1823{
1824 unsigned long nr[NR_LRU_LISTS];
1825 unsigned long nr_to_scan;
1826 enum lru_list lru;
1827 unsigned long nr_reclaimed = 0;
1828 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1829 struct blk_plug plug;
1830
1831 get_scan_count(lruvec, sc, nr);
1832
1833 blk_start_plug(&plug);
1834 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1835 nr[LRU_INACTIVE_FILE]) {
1836 for_each_evictable_lru(lru) {
1837 if (nr[lru]) {
1838 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1839 nr[lru] -= nr_to_scan;
1840
1841 nr_reclaimed += shrink_list(lru, nr_to_scan,
1842 lruvec, sc);
1843 }
1844 }
1845 /*
1846 * On large memory systems, scan >> priority can become
1847 * really large. This is fine for the starting priority;
1848 * we want to put equal scanning pressure on each zone.
1849 * However, if the VM has a harder time of freeing pages,
1850 * with multiple processes reclaiming pages, the total
1851 * freeing target can get unreasonably large.
1852 */
1853 if (nr_reclaimed >= nr_to_reclaim &&
1854 sc->priority < DEF_PRIORITY)
1855 break;
1856 }
1857 blk_finish_plug(&plug);
1858 sc->nr_reclaimed += nr_reclaimed;
1859
1860 /*
1861 * Even if we did not try to evict anon pages at all, we want to
1862 * rebalance the anon lru active/inactive ratio.
1863 */
1864 if (inactive_anon_is_low(lruvec))
1865 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1866 sc, LRU_ACTIVE_ANON);
1867
1868 throttle_vm_writeout(sc->gfp_mask);
1869}
1870
Mel Gorman23b9da52012-05-29 15:06:20 -07001871/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001872static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07001873{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08001874 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07001875 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001876 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07001877 return true;
1878
1879 return false;
1880}
1881
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001882/*
Mel Gorman23b9da52012-05-29 15:06:20 -07001883 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1884 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1885 * true if more pages should be reclaimed such that when the page allocator
1886 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1887 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08001888 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001889static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08001890 unsigned long nr_reclaimed,
1891 unsigned long nr_scanned,
1892 struct scan_control *sc)
1893{
1894 unsigned long pages_for_compaction;
1895 unsigned long inactive_lru_pages;
1896
1897 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001898 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001899 return false;
1900
Mel Gorman28765922011-02-25 14:44:20 -08001901 /* Consider stopping depending on scan and reclaim activity */
1902 if (sc->gfp_mask & __GFP_REPEAT) {
1903 /*
1904 * For __GFP_REPEAT allocations, stop reclaiming if the
1905 * full LRU list has been scanned and we are still failing
1906 * to reclaim pages. This full LRU scan is potentially
1907 * expensive but a __GFP_REPEAT caller really wants to succeed
1908 */
1909 if (!nr_reclaimed && !nr_scanned)
1910 return false;
1911 } else {
1912 /*
1913 * For non-__GFP_REPEAT allocations which can presumably
1914 * fail without consequence, stop if we failed to reclaim
1915 * any pages from the last SWAP_CLUSTER_MAX number of
1916 * pages that were scanned. This will return to the
1917 * caller faster at the risk reclaim/compaction and
1918 * the resulting allocation attempt fails
1919 */
1920 if (!nr_reclaimed)
1921 return false;
1922 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001923
1924 /*
1925 * If we have not reclaimed enough pages for compaction and the
1926 * inactive lists are large enough, continue reclaiming
1927 */
1928 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001929 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08001930 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001931 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001932 if (sc->nr_reclaimed < pages_for_compaction &&
1933 inactive_lru_pages > pages_for_compaction)
1934 return true;
1935
1936 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001937 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08001938 case COMPACT_PARTIAL:
1939 case COMPACT_CONTINUE:
1940 return false;
1941 default:
1942 return true;
1943 }
1944}
1945
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001946static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08001947{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001948 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08001949
Johannes Weiner56600482012-01-12 17:17:59 -08001950 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001951 struct mem_cgroup *root = sc->target_mem_cgroup;
1952 struct mem_cgroup_reclaim_cookie reclaim = {
1953 .zone = zone,
1954 .priority = sc->priority,
1955 };
1956 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08001957
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001958 nr_reclaimed = sc->nr_reclaimed;
1959 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07001960
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001961 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1962 do {
1963 struct lruvec *lruvec;
1964
1965 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1966
1967 shrink_lruvec(lruvec, sc);
1968
1969 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08001970 * Direct reclaim and kswapd have to scan all memory
1971 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001972 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08001973 *
1974 * Limit reclaim, on the other hand, only cares about
1975 * nr_to_reclaim pages to be reclaimed and it will
1976 * retry with decreasing priority if one round over the
1977 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001978 */
Michal Hockoa394cb82013-02-22 16:32:30 -08001979 if (!global_reclaim(sc) &&
1980 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001981 mem_cgroup_iter_break(root, memcg);
1982 break;
1983 }
1984 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1985 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07001986
1987 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
1988 sc->nr_scanned - nr_scanned,
1989 sc->nr_reclaimed - nr_reclaimed);
1990
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001991 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
1992 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08001993}
1994
Mel Gormanfe4b1b22012-01-12 17:19:45 -08001995/* Returns true if compaction should go ahead for a high-order request */
1996static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1997{
1998 unsigned long balance_gap, watermark;
1999 bool watermark_ok;
2000
2001 /* Do not consider compaction for orders reclaim is meant to satisfy */
2002 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2003 return false;
2004
2005 /*
2006 * Compaction takes time to run and there are potentially other
2007 * callers using the pages just freed. Continue reclaiming until
2008 * there is a buffer of free pages available to give compaction
2009 * a reasonable chance of completing and allocating the page
2010 */
2011 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002012 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002013 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2014 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2015 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2016
2017 /*
2018 * If compaction is deferred, reclaim up to a point where
2019 * compaction will have a chance of success when re-enabled
2020 */
Rik van Rielaff62242012-03-21 16:33:52 -07002021 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002022 return watermark_ok;
2023
2024 /* If compaction is not ready to start, keep reclaiming */
2025 if (!compaction_suitable(zone, sc->order))
2026 return false;
2027
2028 return watermark_ok;
2029}
2030
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031/*
2032 * This is the direct reclaim path, for page-allocating processes. We only
2033 * try to reclaim pages from zones which will satisfy the caller's allocation
2034 * request.
2035 *
Mel Gorman41858962009-06-16 15:32:12 -07002036 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2037 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2039 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002040 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2041 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2042 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 * If a zone is deemed to be full of pinned pages then just give it a light
2045 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002046 *
2047 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002048 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002049 * the caller that it should consider retrying the allocation instead of
2050 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002052static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053{
Mel Gormandd1a2392008-04-28 02:12:17 -07002054 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002055 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002056 unsigned long nr_soft_reclaimed;
2057 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002058 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002059
Mel Gormancc715d92012-03-21 16:34:00 -07002060 /*
2061 * If the number of buffer_heads in the machine exceeds the maximum
2062 * allowed level, force direct reclaim to scan the highmem zone as
2063 * highmem pages could be pinning lowmem pages storing buffer_heads
2064 */
2065 if (buffer_heads_over_limit)
2066 sc->gfp_mask |= __GFP_HIGHMEM;
2067
Mel Gormand4debc62010-08-09 17:19:29 -07002068 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2069 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002070 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002072 /*
2073 * Take care memory controller reclaiming has small influence
2074 * to global LRU.
2075 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002076 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002077 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2078 continue;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002079 if (zone->all_unreclaimable &&
2080 sc->priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002081 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002082 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002083 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002084 * If we already have plenty of memory free for
2085 * compaction in this zone, don't free any more.
2086 * Even though compaction is invoked for any
2087 * non-zero order, only frequent costly order
2088 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002089 * noticeable problem, like transparent huge
2090 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002091 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002092 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002093 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002094 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002095 }
Rik van Riele0887c12011-10-31 17:09:31 -07002096 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002097 /*
2098 * This steals pages from memory cgroups over softlimit
2099 * and returns the number of reclaimed pages and
2100 * scanned pages. This works for global memory pressure
2101 * and balancing, not for a memcg's limit.
2102 */
2103 nr_soft_scanned = 0;
2104 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2105 sc->order, sc->gfp_mask,
2106 &nr_soft_scanned);
2107 sc->nr_reclaimed += nr_soft_reclaimed;
2108 sc->nr_scanned += nr_soft_scanned;
2109 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002110 }
Nick Piggin408d8542006-09-25 23:31:27 -07002111
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002112 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 }
Mel Gormane0c23272011-10-31 17:09:33 -07002114
Mel Gorman0cee34f2012-01-12 17:19:49 -08002115 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002116}
2117
2118static bool zone_reclaimable(struct zone *zone)
2119{
2120 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2121}
2122
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002123/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002124static bool all_unreclaimable(struct zonelist *zonelist,
2125 struct scan_control *sc)
2126{
2127 struct zoneref *z;
2128 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002129
2130 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2131 gfp_zone(sc->gfp_mask), sc->nodemask) {
2132 if (!populated_zone(zone))
2133 continue;
2134 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2135 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002136 if (!zone->all_unreclaimable)
2137 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002138 }
2139
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002140 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002142
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143/*
2144 * This is the main entry point to direct page reclaim.
2145 *
2146 * If a full scan of the inactive list fails to free enough memory then we
2147 * are "out of memory" and something needs to be killed.
2148 *
2149 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2150 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002151 * caller can't do much about. We kick the writeback threads and take explicit
2152 * naps in the hope that some of these pages can be written. But if the
2153 * allocating task holds filesystem locks which prevent writeout this might not
2154 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002155 *
2156 * returns: 0, if no pages reclaimed
2157 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 */
Mel Gormandac1d272008-04-28 02:12:12 -07002159static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002160 struct scan_control *sc,
2161 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
Andrew Morton69e05942006-03-22 00:08:19 -08002163 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002165 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002166 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002167 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002168 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
Keika Kobayashi873b4772008-07-25 01:48:52 -07002170 delayacct_freepages_start();
2171
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002172 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002173 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002175 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002176 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2177 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002178 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002179 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002180
Balbir Singh66e17072008-02-07 00:13:56 -08002181 /*
2182 * Don't shrink slabs when reclaiming memory from
2183 * over limit cgroups
2184 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002185 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002186 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002187 for_each_zone_zonelist(zone, z, zonelist,
2188 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002189 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2190 continue;
2191
2192 lru_pages += zone_reclaimable_pages(zone);
2193 }
2194
Ying Han1495f232011-05-24 17:12:27 -07002195 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002196 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002197 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002198 reclaim_state->reclaimed_slab = 0;
2199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 }
Balbir Singh66e17072008-02-07 00:13:56 -08002201 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002202 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204
2205 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002206 * If we're getting trouble reclaiming, start doing
2207 * writepage even in laptop mode.
2208 */
2209 if (sc->priority < DEF_PRIORITY - 2)
2210 sc->may_writepage = 1;
2211
2212 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 * Try to write back as many pages as we just scanned. This
2214 * tends to cause slow streaming writers to write data to the
2215 * disk smoothly, at the dirtying rate, which is nice. But
2216 * that's undesirable in laptop mode, where we *want* lumpy
2217 * writeout. So in laptop mode, write out the whole world.
2218 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002219 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2220 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002221 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2222 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002223 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 }
2225
2226 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002227 if (!sc->hibernation_mode && sc->nr_scanned &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002228 sc->priority < DEF_PRIORITY - 2) {
Mel Gorman0e093d992010-10-26 14:21:45 -07002229 struct zone *preferred_zone;
2230
2231 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
David Rientjesf33261d2011-01-25 15:07:20 -08002232 &cpuset_current_mems_allowed,
2233 &preferred_zone);
Mel Gorman0e093d992010-10-26 14:21:45 -07002234 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2235 }
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002236 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002237
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002239 delayacct_freepages_end();
2240
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002241 if (sc->nr_reclaimed)
2242 return sc->nr_reclaimed;
2243
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002244 /*
2245 * As hibernation is going on, kswapd is freezed so that it can't mark
2246 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2247 * check.
2248 */
2249 if (oom_killer_disabled)
2250 return 0;
2251
Mel Gorman0cee34f2012-01-12 17:19:49 -08002252 /* Aborted reclaim to try compaction? don't OOM, then */
2253 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002254 return 1;
2255
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002256 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002257 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002258 return 1;
2259
2260 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
2262
Mel Gorman55150612012-07-31 16:44:35 -07002263static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2264{
2265 struct zone *zone;
2266 unsigned long pfmemalloc_reserve = 0;
2267 unsigned long free_pages = 0;
2268 int i;
2269 bool wmark_ok;
2270
2271 for (i = 0; i <= ZONE_NORMAL; i++) {
2272 zone = &pgdat->node_zones[i];
2273 pfmemalloc_reserve += min_wmark_pages(zone);
2274 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2275 }
2276
2277 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2278
2279 /* kswapd must be awake if processes are being throttled */
2280 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2281 pgdat->classzone_idx = min(pgdat->classzone_idx,
2282 (enum zone_type)ZONE_NORMAL);
2283 wake_up_interruptible(&pgdat->kswapd_wait);
2284 }
2285
2286 return wmark_ok;
2287}
2288
2289/*
2290 * Throttle direct reclaimers if backing storage is backed by the network
2291 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2292 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002293 * when the low watermark is reached.
2294 *
2295 * Returns true if a fatal signal was delivered during throttling. If this
2296 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002297 */
Mel Gorman50694c22012-11-26 16:29:48 -08002298static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002299 nodemask_t *nodemask)
2300{
2301 struct zone *zone;
2302 int high_zoneidx = gfp_zone(gfp_mask);
2303 pg_data_t *pgdat;
2304
2305 /*
2306 * Kernel threads should not be throttled as they may be indirectly
2307 * responsible for cleaning pages necessary for reclaim to make forward
2308 * progress. kjournald for example may enter direct reclaim while
2309 * committing a transaction where throttling it could forcing other
2310 * processes to block on log_wait_commit().
2311 */
2312 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002313 goto out;
2314
2315 /*
2316 * If a fatal signal is pending, this process should not throttle.
2317 * It should return quickly so it can exit and free its memory
2318 */
2319 if (fatal_signal_pending(current))
2320 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002321
2322 /* Check if the pfmemalloc reserves are ok */
2323 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2324 pgdat = zone->zone_pgdat;
2325 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002326 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002327
Mel Gorman68243e72012-07-31 16:44:39 -07002328 /* Account for the throttling */
2329 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2330
Mel Gorman55150612012-07-31 16:44:35 -07002331 /*
2332 * If the caller cannot enter the filesystem, it's possible that it
2333 * is due to the caller holding an FS lock or performing a journal
2334 * transaction in the case of a filesystem like ext[3|4]. In this case,
2335 * it is not safe to block on pfmemalloc_wait as kswapd could be
2336 * blocked waiting on the same lock. Instead, throttle for up to a
2337 * second before continuing.
2338 */
2339 if (!(gfp_mask & __GFP_FS)) {
2340 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2341 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002342
2343 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002344 }
2345
2346 /* Throttle until kswapd wakes the process */
2347 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2348 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002349
2350check_pending:
2351 if (fatal_signal_pending(current))
2352 return true;
2353
2354out:
2355 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002356}
2357
Mel Gormandac1d272008-04-28 02:12:12 -07002358unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002359 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002360{
Mel Gorman33906bc2010-08-09 17:19:16 -07002361 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002362 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002363 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002364 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002365 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002366 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002367 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002368 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002369 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002370 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002371 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002372 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002373 struct shrink_control shrink = {
2374 .gfp_mask = sc.gfp_mask,
2375 };
Balbir Singh66e17072008-02-07 00:13:56 -08002376
Mel Gorman55150612012-07-31 16:44:35 -07002377 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002378 * Do not enter reclaim if fatal signal was delivered while throttled.
2379 * 1 is returned so that the page allocator does not OOM kill at this
2380 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002381 */
Mel Gorman50694c22012-11-26 16:29:48 -08002382 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002383 return 1;
2384
Mel Gorman33906bc2010-08-09 17:19:16 -07002385 trace_mm_vmscan_direct_reclaim_begin(order,
2386 sc.may_writepage,
2387 gfp_mask);
2388
Ying Hana09ed5e2011-05-24 17:12:26 -07002389 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002390
2391 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2392
2393 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002394}
2395
Andrew Mortonc255a452012-07-31 16:43:02 -07002396#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002397
Johannes Weiner72835c82012-01-12 17:18:32 -08002398unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002399 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002400 struct zone *zone,
2401 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002402{
2403 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002404 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002405 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002406 .may_writepage = !laptop_mode,
2407 .may_unmap = 1,
2408 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002409 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002410 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002411 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002412 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002413 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002414
Balbir Singh4e416952009-09-23 15:56:39 -07002415 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2416 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002417
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002418 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002419 sc.may_writepage,
2420 sc.gfp_mask);
2421
Balbir Singh4e416952009-09-23 15:56:39 -07002422 /*
2423 * NOTE: Although we can get the priority field, using it
2424 * here is not a good idea, since it limits the pages we can scan.
2425 * if we don't reclaim here, the shrink_zone from balance_pgdat
2426 * will pick up pages from other mem cgroup's as well. We hack
2427 * the priority and make it zero.
2428 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002429 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002430
2431 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2432
Ying Han0ae5e892011-05-26 16:25:25 -07002433 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002434 return sc.nr_reclaimed;
2435}
2436
Johannes Weiner72835c82012-01-12 17:18:32 -08002437unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002438 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002439 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002440{
Balbir Singh4e416952009-09-23 15:56:39 -07002441 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002442 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002443 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002444 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002445 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002446 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002447 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002448 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002449 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002450 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002451 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002452 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002453 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2454 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2455 };
2456 struct shrink_control shrink = {
2457 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002458 };
Balbir Singh66e17072008-02-07 00:13:56 -08002459
Ying Han889976d2011-05-26 16:25:33 -07002460 /*
2461 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2462 * take care of from where we get pages. So the node where we start the
2463 * scan does not need to be the current node.
2464 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002465 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002466
2467 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002468
2469 trace_mm_vmscan_memcg_reclaim_begin(0,
2470 sc.may_writepage,
2471 sc.gfp_mask);
2472
Ying Hana09ed5e2011-05-24 17:12:26 -07002473 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002474
2475 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2476
2477 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002478}
2479#endif
2480
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002481static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002482{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002483 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002484
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002485 if (!total_swap_pages)
2486 return;
2487
2488 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2489 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002490 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002491
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002492 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002493 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002494 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002495
2496 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2497 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002498}
2499
Johannes Weiner60cefed2012-11-29 13:54:23 -08002500static bool zone_balanced(struct zone *zone, int order,
2501 unsigned long balance_gap, int classzone_idx)
2502{
2503 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2504 balance_gap, classzone_idx, 0))
2505 return false;
2506
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002507 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2508 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002509 return false;
2510
2511 return true;
2512}
2513
Mel Gorman1741c872011-01-13 15:46:21 -08002514/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002515 * pgdat_balanced() is used when checking if a node is balanced.
2516 *
2517 * For order-0, all zones must be balanced!
2518 *
2519 * For high-order allocations only zones that meet watermarks and are in a
2520 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2521 * total of balanced pages must be at least 25% of the zones allowed by
2522 * classzone_idx for the node to be considered balanced. Forcing all zones to
2523 * be balanced for high orders can cause excessive reclaim when there are
2524 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002525 * The choice of 25% is due to
2526 * o a 16M DMA zone that is balanced will not balance a zone on any
2527 * reasonable sized machine
2528 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002529 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002530 * would need to be at least 256M for it to be balance a whole node.
2531 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2532 * to balance a node on its own. These seemed like reasonable ratios.
2533 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002534static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002535{
Jiang Liub40da042013-02-22 16:33:52 -08002536 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002537 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002538 int i;
2539
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002540 /* Check the watermark levels */
2541 for (i = 0; i <= classzone_idx; i++) {
2542 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002543
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002544 if (!populated_zone(zone))
2545 continue;
2546
Jiang Liub40da042013-02-22 16:33:52 -08002547 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002548
2549 /*
2550 * A special case here:
2551 *
2552 * balance_pgdat() skips over all_unreclaimable after
2553 * DEF_PRIORITY. Effectively, it considers them balanced so
2554 * they must be considered balanced here as well!
2555 */
2556 if (zone->all_unreclaimable) {
Jiang Liub40da042013-02-22 16:33:52 -08002557 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002558 continue;
2559 }
2560
2561 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002562 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002563 else if (!order)
2564 return false;
2565 }
2566
2567 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002568 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002569 else
2570 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002571}
2572
Mel Gorman55150612012-07-31 16:44:35 -07002573/*
2574 * Prepare kswapd for sleeping. This verifies that there are no processes
2575 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2576 *
2577 * Returns true if kswapd is ready to sleep
2578 */
2579static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002580 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002581{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002582 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2583 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002584 return false;
2585
2586 /*
2587 * There is a potential race between when kswapd checks its watermarks
2588 * and a process gets throttled. There is also a potential race if
2589 * processes get throttled, kswapd wakes, a large process exits therby
2590 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2591 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2592 * so wake them now if necessary. If necessary, processes will wake
2593 * kswapd and get throttled again
2594 */
2595 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2596 wake_up(&pgdat->pfmemalloc_wait);
2597 return false;
2598 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002599
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002600 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002601}
2602
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603/*
2604 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002605 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002607 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 *
2609 * There is special handling here for zones which are full of pinned pages.
2610 * This can happen if the pages are all mlocked, or if they are all used by
2611 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2612 * What we do is to detect the case where all pages in the zone have been
2613 * scanned twice and there has been zero successful reclaim. Mark the zone as
2614 * dead and from now on, only perform a short scan. Basically we're polling
2615 * the zone for when the problem goes away.
2616 *
2617 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002618 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2619 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2620 * lower zones regardless of the number of free pages in the lower zones. This
2621 * interoperates with the page allocator fallback scheme to ensure that aging
2622 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 */
Mel Gorman99504742011-01-13 15:46:20 -08002624static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002625 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626{
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002627 bool pgdat_is_balanced = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002629 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 struct reclaim_state *reclaim_state = current->reclaim_state;
Ying Han0ae5e892011-05-26 16:25:25 -07002631 unsigned long nr_soft_reclaimed;
2632 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002633 struct scan_control sc = {
2634 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002635 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002636 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002637 /*
2638 * kswapd doesn't want to be bailed out while reclaim. because
2639 * we want to put equal scanning pressure on each zone.
2640 */
2641 .nr_to_reclaim = ULONG_MAX,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002642 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002643 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002644 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002645 struct shrink_control shrink = {
2646 .gfp_mask = sc.gfp_mask,
2647 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648loop_again:
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002649 sc.priority = DEF_PRIORITY;
Rik van Riela79311c2009-01-06 14:40:01 -08002650 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002651 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002652 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002654 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 unsigned long lru_pages = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002657 /*
2658 * Scan in the highmem->dma direction for the highest
2659 * zone which needs scanning
2660 */
2661 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2662 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002664 if (!populated_zone(zone))
2665 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002667 if (zone->all_unreclaimable &&
2668 sc.priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002669 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670
Rik van Riel556adec2008-10-18 20:26:34 -07002671 /*
2672 * Do some background aging of the anon list, to give
2673 * pages a chance to be referenced before reclaiming.
2674 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002675 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002676
Mel Gormancc715d92012-03-21 16:34:00 -07002677 /*
2678 * If the number of buffer_heads in the machine
2679 * exceeds the maximum allowed level and this node
2680 * has a highmem zone, force kswapd to reclaim from
2681 * it to relieve lowmem pressure.
2682 */
2683 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2684 end_zone = i;
2685 break;
2686 }
2687
Johannes Weiner60cefed2012-11-29 13:54:23 -08002688 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002689 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002690 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002691 } else {
2692 /* If balanced, clear the congested flag */
2693 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002696
2697 if (i < 0) {
2698 pgdat_is_balanced = true;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002699 goto out;
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002700 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002701
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 for (i = 0; i <= end_zone; i++) {
2703 struct zone *zone = pgdat->node_zones + i;
2704
Wu Fengguangadea02a2009-09-21 17:01:42 -07002705 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 }
2707
2708 /*
2709 * Now scan the zone in the dma->highmem direction, stopping
2710 * at the last zone which needs scanning.
2711 *
2712 * We do this because the page allocator works in the opposite
2713 * direction. This prevents the page allocator from allocating
2714 * pages behind kswapd's direction of progress, which would
2715 * cause too much scanning of the lower zones.
2716 */
2717 for (i = 0; i <= end_zone; i++) {
2718 struct zone *zone = pgdat->node_zones + i;
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002719 int nr_slab, testorder;
Mel Gorman8afdcec2011-03-22 16:33:04 -07002720 unsigned long balance_gap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
Con Kolivasf3fe6512006-01-06 00:11:15 -08002722 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 continue;
2724
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002725 if (zone->all_unreclaimable &&
2726 sc.priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 continue;
2728
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002730
Ying Han0ae5e892011-05-26 16:25:25 -07002731 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002732 /*
2733 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002734 */
Ying Han0ae5e892011-05-26 16:25:25 -07002735 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2736 order, sc.gfp_mask,
2737 &nr_soft_scanned);
2738 sc.nr_reclaimed += nr_soft_reclaimed;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002739
Rik van Riel32a43302007-10-16 01:24:50 -07002740 /*
Mel Gorman8afdcec2011-03-22 16:33:04 -07002741 * We put equal pressure on every zone, unless
2742 * one zone has way too many pages free
2743 * already. The "too many pages" is defined
2744 * as the high wmark plus a "gap" where the
2745 * gap is either the low watermark or 1%
2746 * of the zone, whichever is smaller.
Rik van Riel32a43302007-10-16 01:24:50 -07002747 */
Mel Gorman8afdcec2011-03-22 16:33:04 -07002748 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002749 (zone->managed_pages +
Mel Gorman8afdcec2011-03-22 16:33:04 -07002750 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2751 KSWAPD_ZONE_BALANCE_GAP_RATIO);
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002752 /*
2753 * Kswapd reclaims only single pages with compaction
2754 * enabled. Trying too hard to reclaim until contiguous
2755 * free pages have become available can hurt performance
2756 * by evicting too much useful data from memory.
2757 * Do not reclaim more than needed for compaction.
2758 */
2759 testorder = order;
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002760 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
Rik van Rielfe2c2a12012-03-21 16:33:51 -07002761 compaction_suitable(zone, order) !=
2762 COMPACT_SKIPPED)
2763 testorder = 0;
2764
Mel Gormancc715d92012-03-21 16:34:00 -07002765 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
Johannes Weiner60cefed2012-11-29 13:54:23 -08002766 !zone_balanced(zone, testorder,
2767 balance_gap, end_zone)) {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002768 shrink_zone(zone, &sc);
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08002769
Mel Gormand7868da2011-07-08 15:39:38 -07002770 reclaim_state->reclaimed_slab = 0;
2771 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2772 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Mel Gormand7868da2011-07-08 15:39:38 -07002773
2774 if (nr_slab == 0 && !zone_reclaimable(zone))
2775 zone->all_unreclaimable = 1;
2776 }
2777
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002779 * If we're getting trouble reclaiming, start doing
2780 * writepage even in laptop mode.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 */
Minchan Kim0e50ce32013-02-22 16:35:37 -08002782 if (sc.priority < DEF_PRIORITY - 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002784
Mel Gorman215ddd62011-07-08 15:39:40 -07002785 if (zone->all_unreclaimable) {
2786 if (end_zone && end_zone == i)
2787 end_zone--;
Mel Gormand7868da2011-07-08 15:39:38 -07002788 continue;
Mel Gorman215ddd62011-07-08 15:39:40 -07002789 }
Mel Gormand7868da2011-07-08 15:39:38 -07002790
Zlatko Calusic258401a2013-02-22 16:34:06 -08002791 if (zone_balanced(zone, testorder, 0, end_zone))
Mel Gorman0e093d992010-10-26 14:21:45 -07002792 /*
2793 * If a zone reaches its high watermark,
2794 * consider it to be no longer congested. It's
2795 * possible there are dirty pages backed by
2796 * congested BDIs but as pressure is relieved,
Wanpeng Liab8704b2012-06-17 09:27:18 +08002797 * speculatively avoid congestion waits
Mel Gorman0e093d992010-10-26 14:21:45 -07002798 */
2799 zone_clear_flag(zone, ZONE_CONGESTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 }
Mel Gorman55150612012-07-31 16:44:35 -07002801
2802 /*
2803 * If the low watermark is met there is no need for processes
2804 * to be throttled on pfmemalloc_wait as they should not be
2805 * able to safely make forward progress. Wake them
2806 */
2807 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2808 pfmemalloc_watermark_ok(pgdat))
2809 wake_up(&pgdat->pfmemalloc_wait);
2810
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002811 if (pgdat_balanced(pgdat, order, *classzone_idx)) {
2812 pgdat_is_balanced = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 break; /* kswapd: all done */
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002814 }
2815
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 * We do this so kswapd doesn't build up large priorities for
2818 * example when it is freeing in parallel with allocators. It
2819 * matches the direct reclaim path behaviour in terms of impact
2820 * on zone->*_priority.
2821 */
Rik van Riela79311c2009-01-06 14:40:01 -08002822 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 break;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002824 } while (--sc.priority >= 0);
Mel Gorman99504742011-01-13 15:46:20 -08002825
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002826out:
2827 if (!pgdat_is_balanced) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002829
2830 try_to_freeze();
2831
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002832 /*
2833 * Fragmentation may mean that the system cannot be
2834 * rebalanced for high-order allocations in all zones.
2835 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2836 * it means the zones have been fully scanned and are still
2837 * not balanced. For high-order allocations, there is
2838 * little point trying all over again as kswapd may
2839 * infinite loop.
2840 *
2841 * Instead, recheck all watermarks at order-0 as they
2842 * are the most important. If watermarks are ok, kswapd will go
2843 * back to sleep. High-order users can still perform direct
2844 * reclaim if they wish.
2845 */
2846 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2847 order = sc.order = 0;
2848
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 goto loop_again;
2850 }
2851
Mel Gorman99504742011-01-13 15:46:20 -08002852 /*
2853 * If kswapd was reclaiming at a higher order, it has the option of
2854 * sleeping without all zones being balanced. Before it does, it must
2855 * ensure that the watermarks for order-0 on *all* zones are met and
2856 * that the congestion flags are cleared. The congestion flag must
2857 * be cleared as kswapd is the only mechanism that clears the flag
2858 * and it is potentially going to sleep here.
2859 */
2860 if (order) {
Rik van Riel7be62de2012-03-21 16:33:52 -07002861 int zones_need_compaction = 1;
2862
Mel Gorman99504742011-01-13 15:46:20 -08002863 for (i = 0; i <= end_zone; i++) {
2864 struct zone *zone = pgdat->node_zones + i;
2865
2866 if (!populated_zone(zone))
2867 continue;
2868
Rik van Riel7be62de2012-03-21 16:33:52 -07002869 /* Check if the memory needs to be defragmented. */
2870 if (zone_watermark_ok(zone, order,
2871 low_wmark_pages(zone), *classzone_idx, 0))
2872 zones_need_compaction = 0;
Mel Gorman99504742011-01-13 15:46:20 -08002873 }
Rik van Riel7be62de2012-03-21 16:33:52 -07002874
2875 if (zones_need_compaction)
2876 compact_pgdat(pgdat, order);
Mel Gorman99504742011-01-13 15:46:20 -08002877 }
2878
Mel Gorman0abdee22011-01-13 15:46:22 -08002879 /*
Mel Gorman55150612012-07-31 16:44:35 -07002880 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08002881 * makes a decision on the order we were last reclaiming at. However,
2882 * if another caller entered the allocator slow path while kswapd
2883 * was awake, order will remain at the higher level
2884 */
Mel Gormandc83edd2011-01-13 15:46:26 -08002885 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08002886 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887}
2888
Mel Gormandc83edd2011-01-13 15:46:26 -08002889static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002890{
2891 long remaining = 0;
2892 DEFINE_WAIT(wait);
2893
2894 if (freezing(current) || kthread_should_stop())
2895 return;
2896
2897 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2898
2899 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07002900 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002901 remaining = schedule_timeout(HZ/10);
2902 finish_wait(&pgdat->kswapd_wait, &wait);
2903 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2904 }
2905
2906 /*
2907 * After a short sleep, check if it was a premature sleep. If not, then
2908 * go fully to sleep until explicitly woken up.
2909 */
Mel Gorman55150612012-07-31 16:44:35 -07002910 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002911 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2912
2913 /*
2914 * vmstat counters are not perfectly accurate and the estimated
2915 * value for counters such as NR_FREE_PAGES can deviate from the
2916 * true value by nr_online_cpus * threshold. To avoid the zone
2917 * watermarks being breached while under pressure, we reduce the
2918 * per-cpu vmstat threshold while kswapd is awake and restore
2919 * them before going back to sleep.
2920 */
2921 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07002922
Mel Gorman62997022012-10-08 16:32:47 -07002923 /*
2924 * Compaction records what page blocks it recently failed to
2925 * isolate pages from and skips them in the future scanning.
2926 * When kswapd is going to sleep, it is reasonable to assume
2927 * that pages and compaction may succeed so reset the cache.
2928 */
2929 reset_isolation_suitable(pgdat);
2930
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07002931 if (!kthread_should_stop())
2932 schedule();
2933
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002934 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2935 } else {
2936 if (remaining)
2937 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2938 else
2939 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2940 }
2941 finish_wait(&pgdat->kswapd_wait, &wait);
2942}
2943
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944/*
2945 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002946 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 *
2948 * This basically trickles out pages so that we have _some_
2949 * free memory available even if there is no other activity
2950 * that frees anything up. This is needed for things like routing
2951 * etc, where we otherwise might have all activity going on in
2952 * asynchronous contexts that cannot page things out.
2953 *
2954 * If there are applications that are active memory-allocators
2955 * (most normal use), this basically shouldn't matter.
2956 */
2957static int kswapd(void *p)
2958{
Mel Gorman215ddd62011-07-08 15:39:40 -07002959 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002960 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07002961 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002962 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 pg_data_t *pgdat = (pg_data_t*)p;
2964 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08002965
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 struct reclaim_state reclaim_state = {
2967 .reclaimed_slab = 0,
2968 };
Rusty Russella70f7302009-03-13 14:49:46 +10302969 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970
Nick Piggincf40bd12009-01-21 08:12:39 +01002971 lockdep_set_current_reclaim_state(GFP_KERNEL);
2972
Rusty Russell174596a2009-01-01 10:12:29 +10302973 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002974 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 current->reclaim_state = &reclaim_state;
2976
2977 /*
2978 * Tell the memory management that we're a "memory allocator",
2979 * and that if we need more memory we should get access to it
2980 * regardless (see "__alloc_pages()"). "kswapd" should
2981 * never get caught in the normal page freeing logic.
2982 *
2983 * (Kswapd normally doesn't need memory anyway, but sometimes
2984 * you need a small amount of memory in order to be able to
2985 * page out something else, and this flag essentially protects
2986 * us from recursively trying to free more memory as we're
2987 * trying to free the first piece of memory in the first place).
2988 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002989 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002990 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
Mel Gorman215ddd62011-07-08 15:39:40 -07002992 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002993 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07002994 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07002995 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08002997 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002998
Mel Gorman215ddd62011-07-08 15:39:40 -07002999 /*
3000 * If the last balance_pgdat was unsuccessful it's unlikely a
3001 * new request of a similar or harder type will succeed soon
3002 * so consider going to sleep on the basis we reclaimed at
3003 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003004 if (balanced_classzone_idx >= new_classzone_idx &&
3005 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003006 new_order = pgdat->kswapd_max_order;
3007 new_classzone_idx = pgdat->classzone_idx;
3008 pgdat->kswapd_max_order = 0;
3009 pgdat->classzone_idx = pgdat->nr_zones - 1;
3010 }
3011
Mel Gorman99504742011-01-13 15:46:20 -08003012 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 /*
3014 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003015 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 */
3017 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003018 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003020 kswapd_try_to_sleep(pgdat, balanced_order,
3021 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003023 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003024 new_order = order;
3025 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003026 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003027 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029
David Rientjes8fe23e02009-12-14 17:58:33 -08003030 ret = try_to_freeze();
3031 if (kthread_should_stop())
3032 break;
3033
3034 /*
3035 * We can speed up thawing tasks if we don't call balance_pgdat
3036 * after returning from the refrigerator
3037 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003038 if (!ret) {
3039 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003040 balanced_classzone_idx = classzone_idx;
3041 balanced_order = balance_pgdat(pgdat, order,
3042 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003043 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003045
3046 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 return 0;
3048}
3049
3050/*
3051 * A zone is low on free memory, so wake its kswapd task to service it.
3052 */
Mel Gorman99504742011-01-13 15:46:20 -08003053void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054{
3055 pg_data_t *pgdat;
3056
Con Kolivasf3fe6512006-01-06 00:11:15 -08003057 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058 return;
3059
Paul Jackson02a0e532006-12-13 00:34:25 -08003060 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003062 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003063 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003064 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003065 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3066 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003067 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003069 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3070 return;
3071
3072 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003073 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074}
3075
Wu Fengguangadea02a2009-09-21 17:01:42 -07003076/*
3077 * The reclaimable count would be mostly accurate.
3078 * The less reclaimable pages may be
3079 * - mlocked pages, which will be moved to unevictable list when encountered
3080 * - mapped pages, which may require several travels to be reclaimed
3081 * - dirty pages, which is not "instantly" reclaimable
3082 */
3083unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003084{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003085 int nr;
3086
3087 nr = global_page_state(NR_ACTIVE_FILE) +
3088 global_page_state(NR_INACTIVE_FILE);
3089
Shaohua Liec8acf22013-02-22 16:34:38 -08003090 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003091 nr += global_page_state(NR_ACTIVE_ANON) +
3092 global_page_state(NR_INACTIVE_ANON);
3093
3094 return nr;
3095}
3096
3097unsigned long zone_reclaimable_pages(struct zone *zone)
3098{
3099 int nr;
3100
3101 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3102 zone_page_state(zone, NR_INACTIVE_FILE);
3103
Shaohua Liec8acf22013-02-22 16:34:38 -08003104 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003105 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3106 zone_page_state(zone, NR_INACTIVE_ANON);
3107
3108 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003109}
3110
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003111#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003113 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003114 * freed pages.
3115 *
3116 * Rather than trying to age LRUs the aim is to preserve the overall
3117 * LRU order by reclaiming preferentially
3118 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003120unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003122 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003123 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003124 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3125 .may_swap = 1,
3126 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003127 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003128 .nr_to_reclaim = nr_to_reclaim,
3129 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003130 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003131 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003133 struct shrink_control shrink = {
3134 .gfp_mask = sc.gfp_mask,
3135 };
3136 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003137 struct task_struct *p = current;
3138 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003140 p->flags |= PF_MEMALLOC;
3141 lockdep_set_current_reclaim_state(sc.gfp_mask);
3142 reclaim_state.reclaimed_slab = 0;
3143 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003144
Ying Hana09ed5e2011-05-24 17:12:26 -07003145 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003146
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003147 p->reclaim_state = NULL;
3148 lockdep_clear_current_reclaim_state();
3149 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003150
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003151 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003153#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155/* It's optimal to keep kswapds on the same CPUs as their memory, but
3156 not required for correctness. So if the last cpu in a node goes
3157 away, we get changed to run anywhere: as the first one comes back,
3158 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003159static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3160 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003162 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003164 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003165 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003166 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303167 const struct cpumask *mask;
3168
3169 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003170
Rusty Russell3e597942009-01-01 10:12:24 +10303171 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003173 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 }
3175 }
3176 return NOTIFY_OK;
3177}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178
Yasunori Goto3218ae12006-06-27 02:53:33 -07003179/*
3180 * This kswapd start function will be called by init and node-hot-add.
3181 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3182 */
3183int kswapd_run(int nid)
3184{
3185 pg_data_t *pgdat = NODE_DATA(nid);
3186 int ret = 0;
3187
3188 if (pgdat->kswapd)
3189 return 0;
3190
3191 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3192 if (IS_ERR(pgdat->kswapd)) {
3193 /* failure at boot is fatal */
3194 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003195 pr_err("Failed to start kswapd on node %d\n", nid);
3196 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003197 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003198 }
3199 return ret;
3200}
3201
David Rientjes8fe23e02009-12-14 17:58:33 -08003202/*
Jiang Liud8adde12012-07-11 14:01:52 -07003203 * Called by memory hotplug when all memory in a node is offlined. Caller must
3204 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003205 */
3206void kswapd_stop(int nid)
3207{
3208 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3209
Jiang Liud8adde12012-07-11 14:01:52 -07003210 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003211 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003212 NODE_DATA(nid)->kswapd = NULL;
3213 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003214}
3215
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216static int __init kswapd_init(void)
3217{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003218 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003219
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003221 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003222 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223 hotcpu_notifier(cpu_callback, 0);
3224 return 0;
3225}
3226
3227module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003228
3229#ifdef CONFIG_NUMA
3230/*
3231 * Zone reclaim mode
3232 *
3233 * If non-zero call zone_reclaim when the number of free pages falls below
3234 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003235 */
3236int zone_reclaim_mode __read_mostly;
3237
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003238#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003239#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003240#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3241#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3242
Christoph Lameter9eeff232006-01-18 17:42:31 -08003243/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003244 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3245 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3246 * a zone.
3247 */
3248#define ZONE_RECLAIM_PRIORITY 4
3249
Christoph Lameter9eeff232006-01-18 17:42:31 -08003250/*
Christoph Lameter96146342006-07-03 00:24:13 -07003251 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3252 * occur.
3253 */
3254int sysctl_min_unmapped_ratio = 1;
3255
3256/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003257 * If the number of slab pages in a zone grows beyond this percentage then
3258 * slab reclaim needs to occur.
3259 */
3260int sysctl_min_slab_ratio = 5;
3261
Mel Gorman90afa5d2009-06-16 15:33:20 -07003262static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3263{
3264 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3265 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3266 zone_page_state(zone, NR_ACTIVE_FILE);
3267
3268 /*
3269 * It's possible for there to be more file mapped pages than
3270 * accounted for by the pages on the file LRU lists because
3271 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3272 */
3273 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3274}
3275
3276/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3277static long zone_pagecache_reclaimable(struct zone *zone)
3278{
3279 long nr_pagecache_reclaimable;
3280 long delta = 0;
3281
3282 /*
3283 * If RECLAIM_SWAP is set, then all file pages are considered
3284 * potentially reclaimable. Otherwise, we have to worry about
3285 * pages like swapcache and zone_unmapped_file_pages() provides
3286 * a better estimate
3287 */
3288 if (zone_reclaim_mode & RECLAIM_SWAP)
3289 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3290 else
3291 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3292
3293 /* If we can't clean pages, remove dirty pages from consideration */
3294 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3295 delta += zone_page_state(zone, NR_FILE_DIRTY);
3296
3297 /* Watch for any possible underflows due to delta */
3298 if (unlikely(delta > nr_pagecache_reclaimable))
3299 delta = nr_pagecache_reclaimable;
3300
3301 return nr_pagecache_reclaimable - delta;
3302}
3303
Christoph Lameter0ff38492006-09-25 23:31:52 -07003304/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003305 * Try to free up some pages from this zone through reclaim.
3306 */
Andrew Morton179e9632006-03-22 00:08:18 -08003307static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003308{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003309 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003310 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003311 struct task_struct *p = current;
3312 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003313 struct scan_control sc = {
3314 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003315 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003316 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003317 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003318 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003319 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003320 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003321 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003322 struct shrink_control shrink = {
3323 .gfp_mask = sc.gfp_mask,
3324 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003325 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003326
Christoph Lameter9eeff232006-01-18 17:42:31 -08003327 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003328 /*
3329 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3330 * and we also need to be able to write out pages for RECLAIM_WRITE
3331 * and RECLAIM_SWAP.
3332 */
3333 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003334 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003335 reclaim_state.reclaimed_slab = 0;
3336 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003337
Mel Gorman90afa5d2009-06-16 15:33:20 -07003338 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003339 /*
3340 * Free memory by calling shrink zone with increasing
3341 * priorities until we have enough memory freed.
3342 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003343 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003344 shrink_zone(zone, &sc);
3345 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003346 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003347
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003348 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3349 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003350 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003351 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003352 * many pages were freed in this zone. So we take the current
3353 * number of slab pages and shake the slab until it is reduced
3354 * by the same nr_pages that we used for reclaiming unmapped
3355 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003356 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003357 * Note that shrink_slab will free memory on all zones and may
3358 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003359 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003360 for (;;) {
3361 unsigned long lru_pages = zone_reclaimable_pages(zone);
3362
3363 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003364 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003365 break;
3366
3367 /* Freed enough memory */
3368 nr_slab_pages1 = zone_page_state(zone,
3369 NR_SLAB_RECLAIMABLE);
3370 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3371 break;
3372 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003373
3374 /*
3375 * Update nr_reclaimed by the number of slab pages we
3376 * reclaimed from this zone.
3377 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003378 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3379 if (nr_slab_pages1 < nr_slab_pages0)
3380 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003381 }
3382
Christoph Lameter9eeff232006-01-18 17:42:31 -08003383 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003384 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003385 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003386 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003387}
Andrew Morton179e9632006-03-22 00:08:18 -08003388
3389int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3390{
Andrew Morton179e9632006-03-22 00:08:18 -08003391 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003392 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003393
3394 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003395 * Zone reclaim reclaims unmapped file backed pages and
3396 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003397 *
Christoph Lameter96146342006-07-03 00:24:13 -07003398 * A small portion of unmapped file backed pages is needed for
3399 * file I/O otherwise pages read by file I/O will be immediately
3400 * thrown out if the zone is overallocated. So we do not reclaim
3401 * if less than a specified percentage of the zone is used by
3402 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003403 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003404 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3405 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003406 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003407
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003408 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003409 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003410
Andrew Morton179e9632006-03-22 00:08:18 -08003411 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003412 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003413 */
David Rientjesd773ed62007-10-16 23:26:01 -07003414 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003415 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003416
3417 /*
3418 * Only run zone reclaim on the local zone or on zones that do not
3419 * have associated processors. This will favor the local processor
3420 * over remote processors and spread off node memory allocations
3421 * as wide as possible.
3422 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003423 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003424 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003425 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003426
3427 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003428 return ZONE_RECLAIM_NOSCAN;
3429
David Rientjesd773ed62007-10-16 23:26:01 -07003430 ret = __zone_reclaim(zone, gfp_mask, order);
3431 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3432
Mel Gorman24cf725182009-06-16 15:33:23 -07003433 if (!ret)
3434 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3435
David Rientjesd773ed62007-10-16 23:26:01 -07003436 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003437}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003438#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003439
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003440/*
3441 * page_evictable - test whether a page is evictable
3442 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003443 *
3444 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003445 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003446 *
3447 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003448 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003449 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003450 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003451 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003452int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003453{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003454 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003455}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003456
Hugh Dickins850465792012-01-20 14:34:19 -08003457#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003458/**
Hugh Dickins24513262012-01-20 14:34:21 -08003459 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3460 * @pages: array of pages to check
3461 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003462 *
Hugh Dickins24513262012-01-20 14:34:21 -08003463 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003464 *
3465 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003466 */
Hugh Dickins24513262012-01-20 14:34:21 -08003467void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003468{
Johannes Weiner925b7672012-01-12 17:18:15 -08003469 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003470 struct zone *zone = NULL;
3471 int pgscanned = 0;
3472 int pgrescued = 0;
3473 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003474
Hugh Dickins24513262012-01-20 14:34:21 -08003475 for (i = 0; i < nr_pages; i++) {
3476 struct page *page = pages[i];
3477 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003478
Hugh Dickins24513262012-01-20 14:34:21 -08003479 pgscanned++;
3480 pagezone = page_zone(page);
3481 if (pagezone != zone) {
3482 if (zone)
3483 spin_unlock_irq(&zone->lru_lock);
3484 zone = pagezone;
3485 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003486 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003487 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003488
Hugh Dickins24513262012-01-20 14:34:21 -08003489 if (!PageLRU(page) || !PageUnevictable(page))
3490 continue;
3491
Hugh Dickins39b5f292012-10-08 16:33:18 -07003492 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003493 enum lru_list lru = page_lru_base_type(page);
3494
3495 VM_BUG_ON(PageActive(page));
3496 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003497 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3498 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003499 pgrescued++;
3500 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003501 }
Hugh Dickins24513262012-01-20 14:34:21 -08003502
3503 if (zone) {
3504 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3505 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3506 spin_unlock_irq(&zone->lru_lock);
3507 }
Hugh Dickins850465792012-01-20 14:34:19 -08003508}
3509#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003510
Johannes Weiner264e56d2011-10-31 17:09:13 -07003511static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003512{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003513 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003514 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003515 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003516 "one, please send an email to linux-mm@kvack.org.\n",
3517 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003518}
3519
3520/*
3521 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3522 * all nodes' unevictable lists for evictable pages
3523 */
3524unsigned long scan_unevictable_pages;
3525
3526int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003527 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003528 size_t *length, loff_t *ppos)
3529{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003530 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003531 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003532 scan_unevictable_pages = 0;
3533 return 0;
3534}
3535
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003536#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003537/*
3538 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3539 * a specified node's per zone unevictable lists for evictable pages.
3540 */
3541
Kay Sievers10fbcf42011-12-21 14:48:43 -08003542static ssize_t read_scan_unevictable_node(struct device *dev,
3543 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003544 char *buf)
3545{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003546 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003547 return sprintf(buf, "0\n"); /* always zero; should fit... */
3548}
3549
Kay Sievers10fbcf42011-12-21 14:48:43 -08003550static ssize_t write_scan_unevictable_node(struct device *dev,
3551 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003552 const char *buf, size_t count)
3553{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003554 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003555 return 1;
3556}
3557
3558
Kay Sievers10fbcf42011-12-21 14:48:43 -08003559static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003560 read_scan_unevictable_node,
3561 write_scan_unevictable_node);
3562
3563int scan_unevictable_register_node(struct node *node)
3564{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003565 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003566}
3567
3568void scan_unevictable_unregister_node(struct node *node)
3569{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003570 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003571}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003572#endif