blob: c85794399848df32b43194644422306bc262a71f [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
Mel Gormane2be15f2013-07-03 15:01:57 -0700672/* Check if a page is dirty or under writeback */
673static void page_check_dirty_writeback(struct page *page,
674 bool *dirty, bool *writeback)
675{
Mel Gormanb4597222013-07-03 15:02:05 -0700676 struct address_space *mapping;
677
Mel Gormane2be15f2013-07-03 15:01:57 -0700678 /*
679 * Anonymous pages are not handled by flushers and must be written
680 * from reclaim context. Do not stall reclaim based on them
681 */
682 if (!page_is_file_cache(page)) {
683 *dirty = false;
684 *writeback = false;
685 return;
686 }
687
688 /* By default assume that the page flags are accurate */
689 *dirty = PageDirty(page);
690 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700691
692 /* Verify dirty/writeback state if the filesystem supports it */
693 if (!page_has_private(page))
694 return;
695
696 mapping = page_mapping(page);
697 if (mapping && mapping->a_ops->is_dirty_writeback)
698 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700699}
700
Nick Piggine2867812008-07-25 19:45:30 -0700701/*
Andrew Morton1742f192006-03-22 00:08:21 -0800702 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 */
Andrew Morton1742f192006-03-22 00:08:21 -0800704static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700705 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700706 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700707 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700708 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700709 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700710 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700711 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700712 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700713 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714{
715 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700716 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700718 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700719 unsigned long nr_dirty = 0;
720 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800721 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700722 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700723 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
725 cond_resched();
726
Tim Chen69980e32012-07-31 16:46:08 -0700727 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 while (!list_empty(page_list)) {
729 struct address_space *mapping;
730 struct page *page;
731 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700732 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700733 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734
735 cond_resched();
736
737 page = lru_to_page(page_list);
738 list_del(&page->lru);
739
Nick Piggin529ae9a2008-08-02 12:01:03 +0200740 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 goto keep;
742
Nick Piggin725d7042006-09-25 23:30:55 -0700743 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700744 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745
746 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800747
Hugh Dickins39b5f292012-10-08 16:33:18 -0700748 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700749 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700750
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700751 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800752 goto keep_locked;
753
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 /* Double the slab pressure for mapped and swapcache pages */
755 if (page_mapped(page) || PageSwapCache(page))
756 sc->nr_scanned++;
757
Andy Whitcroftc661b072007-08-22 14:01:26 -0700758 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
759 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
760
Mel Gorman283aba92013-07-03 15:01:51 -0700761 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700762 * The number of dirty pages determines if a zone is marked
763 * reclaim_congested which affects wait_iff_congested. kswapd
764 * will stall and start writing pages if the tail of the LRU
765 * is all dirty unqueued pages.
766 */
767 page_check_dirty_writeback(page, &dirty, &writeback);
768 if (dirty || writeback)
769 nr_dirty++;
770
771 if (dirty && !writeback)
772 nr_unqueued_dirty++;
773
Mel Gormand04e8ac2013-07-03 15:02:03 -0700774 /*
775 * Treat this page as congested if the underlying BDI is or if
776 * pages are cycling through the LRU so quickly that the
777 * pages marked for immediate reclaim are making it to the
778 * end of the LRU a second time.
779 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700780 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700781 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
782 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700783 nr_congested++;
784
785 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700786 * If a page at the tail of the LRU is under writeback, there
787 * are three cases to consider.
788 *
789 * 1) If reclaim is encountering an excessive number of pages
790 * under writeback and this page is both under writeback and
791 * PageReclaim then it indicates that pages are being queued
792 * for IO but are being recycled through the LRU before the
793 * IO can complete. Waiting on the page itself risks an
794 * indefinite stall if it is impossible to writeback the
795 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700796 * note that the LRU is being scanned too quickly and the
797 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700798 *
799 * 2) Global reclaim encounters a page, memcg encounters a
800 * page that is not marked for immediate reclaim or
801 * the caller does not have __GFP_IO. In this case mark
802 * the page for immediate reclaim and continue scanning.
803 *
804 * __GFP_IO is checked because a loop driver thread might
805 * enter reclaim, and deadlock if it waits on a page for
806 * which it is needed to do the write (loop masks off
807 * __GFP_IO|__GFP_FS for this reason); but more thought
808 * would probably show more reasons.
809 *
810 * Don't require __GFP_FS, since we're not going into the
811 * FS, just waiting on its writeback completion. Worryingly,
812 * ext4 gfs2 and xfs allocate pages with
813 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
814 * may_enter_fs here is liable to OOM on them.
815 *
816 * 3) memcg encounters a page that is not already marked
817 * PageReclaim. memcg does not have any dirty pages
818 * throttling so we could easily OOM just because too many
819 * pages are in writeback and there is nothing else to
820 * reclaim. Wait for the writeback to complete.
821 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700822 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700823 /* Case 1 above */
824 if (current_is_kswapd() &&
825 PageReclaim(page) &&
826 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700827 nr_immediate++;
828 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700829
830 /* Case 2 above */
831 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700832 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
833 /*
834 * This is slightly racy - end_page_writeback()
835 * might have just cleared PageReclaim, then
836 * setting PageReclaim here end up interpreted
837 * as PageReadahead - but that does not matter
838 * enough to care. What we do want is for this
839 * page to have PageReclaim set next time memcg
840 * reclaim reaches the tests above, so it will
841 * then wait_on_page_writeback() to avoid OOM;
842 * and it's also appropriate in global reclaim.
843 */
844 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700845 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700846
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700847 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700848
849 /* Case 3 above */
850 } else {
851 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700852 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700853 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854
Minchan Kim02c6de82012-10-08 16:31:55 -0700855 if (!force_reclaim)
856 references = page_check_references(page, sc);
857
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800858 switch (references) {
859 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800861 case PAGEREF_KEEP:
862 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800863 case PAGEREF_RECLAIM:
864 case PAGEREF_RECLAIM_CLEAN:
865 ; /* try to reclaim the page below */
866 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 /*
869 * Anonymous process memory has backing store?
870 * Try to allocate it some swap space here.
871 */
Nick Pigginb291f002008-10-18 20:26:44 -0700872 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800873 if (!(sc->gfp_mask & __GFP_IO))
874 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700875 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800877 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878
Mel Gormane2be15f2013-07-03 15:01:57 -0700879 /* Adding to swap updated mapping */
880 mapping = page_mapping(page);
881 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882
883 /*
884 * The page is mapped into the page tables of one or more
885 * processes. Try to unmap it here.
886 */
887 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700888 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 case SWAP_FAIL:
890 goto activate_locked;
891 case SWAP_AGAIN:
892 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700893 case SWAP_MLOCK:
894 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 case SWAP_SUCCESS:
896 ; /* try to free the page below */
897 }
898 }
899
900 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700901 /*
902 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -0700903 * avoid risk of stack overflow but only writeback
904 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -0700905 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700906 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700907 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -0700908 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700909 /*
910 * Immediately reclaim when written back.
911 * Similar in principal to deactivate_page()
912 * except we already have the page isolated
913 * and know it's dirty
914 */
915 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
916 SetPageReclaim(page);
917
Mel Gormanee728862011-10-31 17:07:38 -0700918 goto keep_locked;
919 }
920
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800921 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700923 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800925 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 goto keep_locked;
927
928 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700929 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 case PAGE_KEEP:
931 goto keep_locked;
932 case PAGE_ACTIVATE:
933 goto activate_locked;
934 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700935 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700936 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700937 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700939
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 /*
941 * A synchronous write - probably a ramdisk. Go
942 * ahead and try to reclaim the page.
943 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200944 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 goto keep;
946 if (PageDirty(page) || PageWriteback(page))
947 goto keep_locked;
948 mapping = page_mapping(page);
949 case PAGE_CLEAN:
950 ; /* try to free the page below */
951 }
952 }
953
954 /*
955 * If the page has buffers, try to free the buffer mappings
956 * associated with this page. If we succeed we try to free
957 * the page as well.
958 *
959 * We do this even if the page is PageDirty().
960 * try_to_release_page() does not perform I/O, but it is
961 * possible for a page to have PageDirty set, but it is actually
962 * clean (all its buffers are clean). This happens if the
963 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700964 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 * try_to_release_page() will discover that cleanness and will
966 * drop the buffers and mark the page clean - it can be freed.
967 *
968 * Rarely, pages can have buffers and no ->mapping. These are
969 * the pages which were not successfully invalidated in
970 * truncate_complete_page(). We try to drop those buffers here
971 * and if that worked, and the page is no longer mapped into
972 * process address space (page_count == 1) it can be freed.
973 * Otherwise, leave the page on the LRU so it is swappable.
974 */
David Howells266cf652009-04-03 16:42:36 +0100975 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 if (!try_to_release_page(page, sc->gfp_mask))
977 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700978 if (!mapping && page_count(page) == 1) {
979 unlock_page(page);
980 if (put_page_testzero(page))
981 goto free_it;
982 else {
983 /*
984 * rare race with speculative reference.
985 * the speculative reference will free
986 * this page shortly, so we may
987 * increment nr_reclaimed here (and
988 * leave it off the LRU).
989 */
990 nr_reclaimed++;
991 continue;
992 }
993 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 }
995
Nick Piggine2867812008-07-25 19:45:30 -0700996 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800997 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998
Nick Piggina978d6f2008-10-18 20:26:58 -0700999 /*
1000 * At this point, we have no other references and there is
1001 * no way to pick any more up (removed from LRU, removed
1002 * from pagecache). Can use non-atomic bitops now (and
1003 * we obviously don't have to worry about waking up a process
1004 * waiting on the page lock, because there are no references.
1005 */
1006 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001007free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001008 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001009
1010 /*
1011 * Is there need to periodically free_page_list? It would
1012 * appear not as the counts should be low
1013 */
1014 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 continue;
1016
Nick Pigginb291f002008-10-18 20:26:44 -07001017cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001018 if (PageSwapCache(page))
1019 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001020 unlock_page(page);
1021 putback_lru_page(page);
1022 continue;
1023
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001025 /* Not a candidate for swapping, so reclaim swap space. */
1026 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001027 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001028 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 SetPageActive(page);
1030 pgactivate++;
1031keep_locked:
1032 unlock_page(page);
1033keep:
1034 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001035 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001037
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001038 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001039
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001041 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001042 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001043 *ret_nr_dirty += nr_dirty;
1044 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001045 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001046 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001047 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001048 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049}
1050
Minchan Kim02c6de82012-10-08 16:31:55 -07001051unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1052 struct list_head *page_list)
1053{
1054 struct scan_control sc = {
1055 .gfp_mask = GFP_KERNEL,
1056 .priority = DEF_PRIORITY,
1057 .may_unmap = 1,
1058 };
Mel Gorman8e950282013-07-03 15:02:02 -07001059 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001060 struct page *page, *next;
1061 LIST_HEAD(clean_pages);
1062
1063 list_for_each_entry_safe(page, next, page_list, lru) {
1064 if (page_is_file_cache(page) && !PageDirty(page)) {
1065 ClearPageActive(page);
1066 list_move(&page->lru, &clean_pages);
1067 }
1068 }
1069
1070 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001071 TTU_UNMAP|TTU_IGNORE_ACCESS,
1072 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001073 list_splice(&clean_pages, page_list);
1074 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1075 return ret;
1076}
1077
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001078/*
1079 * Attempt to remove the specified page from its LRU. Only take this page
1080 * if it is of the appropriate PageActive status. Pages which are being
1081 * freed elsewhere are also ignored.
1082 *
1083 * page: page to consider
1084 * mode: one of the LRU isolation modes defined above
1085 *
1086 * returns 0 on success, -ve errno on failure.
1087 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001088int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001089{
1090 int ret = -EINVAL;
1091
1092 /* Only take pages on the LRU. */
1093 if (!PageLRU(page))
1094 return ret;
1095
Minchan Kime46a2872012-10-08 16:33:48 -07001096 /* Compaction should not handle unevictable pages but CMA can do so */
1097 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001098 return ret;
1099
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001100 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001101
Mel Gormanc8244932012-01-12 17:19:38 -08001102 /*
1103 * To minimise LRU disruption, the caller can indicate that it only
1104 * wants to isolate pages it will be able to operate on without
1105 * blocking - clean pages for the most part.
1106 *
1107 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1108 * is used by reclaim when it is cannot write to backing storage
1109 *
1110 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1111 * that it is possible to migrate without blocking
1112 */
1113 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1114 /* All the caller can do on PageWriteback is block */
1115 if (PageWriteback(page))
1116 return ret;
1117
1118 if (PageDirty(page)) {
1119 struct address_space *mapping;
1120
1121 /* ISOLATE_CLEAN means only clean pages */
1122 if (mode & ISOLATE_CLEAN)
1123 return ret;
1124
1125 /*
1126 * Only pages without mappings or that have a
1127 * ->migratepage callback are possible to migrate
1128 * without blocking
1129 */
1130 mapping = page_mapping(page);
1131 if (mapping && !mapping->a_ops->migratepage)
1132 return ret;
1133 }
1134 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001135
Minchan Kimf80c0672011-10-31 17:06:55 -07001136 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1137 return ret;
1138
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001139 if (likely(get_page_unless_zero(page))) {
1140 /*
1141 * Be careful not to clear PageLRU until after we're
1142 * sure the page is not being freed elsewhere -- the
1143 * page release code relies on it.
1144 */
1145 ClearPageLRU(page);
1146 ret = 0;
1147 }
1148
1149 return ret;
1150}
1151
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001152/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 * zone->lru_lock is heavily contended. Some of the functions that
1154 * shrink the lists perform better by taking out a batch of pages
1155 * and working on them outside the LRU lock.
1156 *
1157 * For pagecache intensive workloads, this function is the hottest
1158 * spot in the kernel (apart from copy_*_user functions).
1159 *
1160 * Appropriate locks must be held before calling this function.
1161 *
1162 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001163 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001165 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001166 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001167 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001168 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 *
1170 * returns how many pages were moved onto *@dst.
1171 */
Andrew Morton69e05942006-03-22 00:08:19 -08001172static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001173 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001174 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001175 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001177 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001178 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001179 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001181 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001182 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001183 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001184
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 page = lru_to_page(src);
1186 prefetchw_prev_lru_page(page, src, flags);
1187
Nick Piggin725d7042006-09-25 23:30:55 -07001188 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001189
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001190 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001191 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001192 nr_pages = hpage_nr_pages(page);
1193 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001194 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001195 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001196 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001197
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001198 case -EBUSY:
1199 /* else it is being freed elsewhere */
1200 list_move(&page->lru, src);
1201 continue;
1202
1203 default:
1204 BUG();
1205 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 }
1207
Hugh Dickinsf6260122012-01-12 17:20:06 -08001208 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001209 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1210 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 return nr_taken;
1212}
1213
Nick Piggin62695a82008-10-18 20:26:09 -07001214/**
1215 * isolate_lru_page - tries to isolate a page from its LRU list
1216 * @page: page to isolate from its LRU list
1217 *
1218 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1219 * vmstat statistic corresponding to whatever LRU list the page was on.
1220 *
1221 * Returns 0 if the page was removed from an LRU list.
1222 * Returns -EBUSY if the page was not on an LRU list.
1223 *
1224 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001225 * the active list, it will have PageActive set. If it was found on
1226 * the unevictable list, it will have the PageUnevictable bit set. That flag
1227 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001228 *
1229 * The vmstat statistic corresponding to the list on which the page was
1230 * found will be decremented.
1231 *
1232 * Restrictions:
1233 * (1) Must be called with an elevated refcount on the page. This is a
1234 * fundamentnal difference from isolate_lru_pages (which is called
1235 * without a stable reference).
1236 * (2) the lru_lock must not be held.
1237 * (3) interrupts must be enabled.
1238 */
1239int isolate_lru_page(struct page *page)
1240{
1241 int ret = -EBUSY;
1242
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001243 VM_BUG_ON(!page_count(page));
1244
Nick Piggin62695a82008-10-18 20:26:09 -07001245 if (PageLRU(page)) {
1246 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001247 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001248
1249 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001250 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001251 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001252 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001253 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001254 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001255 del_page_from_lru_list(page, lruvec, lru);
1256 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001257 }
1258 spin_unlock_irq(&zone->lru_lock);
1259 }
1260 return ret;
1261}
1262
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001263/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001264 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1265 * then get resheduled. When there are massive number of tasks doing page
1266 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1267 * the LRU list will go small and be scanned faster than necessary, leading to
1268 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001269 */
1270static int too_many_isolated(struct zone *zone, int file,
1271 struct scan_control *sc)
1272{
1273 unsigned long inactive, isolated;
1274
1275 if (current_is_kswapd())
1276 return 0;
1277
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001278 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001279 return 0;
1280
1281 if (file) {
1282 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1283 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1284 } else {
1285 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1286 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1287 }
1288
Fengguang Wu3cf23842012-12-18 14:23:31 -08001289 /*
1290 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1291 * won't get blocked by normal direct-reclaimers, forming a circular
1292 * deadlock.
1293 */
1294 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1295 inactive >>= 3;
1296
Rik van Riel35cd7812009-09-21 17:01:38 -07001297 return isolated > inactive;
1298}
1299
Mel Gorman66635622010-08-09 17:19:30 -07001300static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001301putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001302{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001303 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1304 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001305 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001306
Mel Gorman66635622010-08-09 17:19:30 -07001307 /*
1308 * Put back any unfreeable pages.
1309 */
Mel Gorman66635622010-08-09 17:19:30 -07001310 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001311 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001312 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001313
Mel Gorman66635622010-08-09 17:19:30 -07001314 VM_BUG_ON(PageLRU(page));
1315 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001316 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001317 spin_unlock_irq(&zone->lru_lock);
1318 putback_lru_page(page);
1319 spin_lock_irq(&zone->lru_lock);
1320 continue;
1321 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001322
1323 lruvec = mem_cgroup_page_lruvec(page, zone);
1324
Linus Torvalds7a608572011-01-17 14:42:19 -08001325 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001326 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001327 add_page_to_lru_list(page, lruvec, lru);
1328
Mel Gorman66635622010-08-09 17:19:30 -07001329 if (is_active_lru(lru)) {
1330 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001331 int numpages = hpage_nr_pages(page);
1332 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001333 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001334 if (put_page_testzero(page)) {
1335 __ClearPageLRU(page);
1336 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001337 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001338
1339 if (unlikely(PageCompound(page))) {
1340 spin_unlock_irq(&zone->lru_lock);
1341 (*get_compound_page_dtor(page))(page);
1342 spin_lock_irq(&zone->lru_lock);
1343 } else
1344 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001345 }
1346 }
Mel Gorman66635622010-08-09 17:19:30 -07001347
Hugh Dickins3f797682012-01-12 17:20:07 -08001348 /*
1349 * To save our caller's stack, now use input list for pages to free.
1350 */
1351 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001352}
1353
1354/*
Andrew Morton1742f192006-03-22 00:08:21 -08001355 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1356 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 */
Mel Gorman66635622010-08-09 17:19:30 -07001358static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001359shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001360 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
1362 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001363 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001364 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001365 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001366 unsigned long nr_dirty = 0;
1367 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001368 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001369 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001370 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001371 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001372 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001373 struct zone *zone = lruvec_zone(lruvec);
1374 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001375
Rik van Riel35cd7812009-09-21 17:01:38 -07001376 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001377 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001378
1379 /* We are about to die and free our memory. Return now. */
1380 if (fatal_signal_pending(current))
1381 return SWAP_CLUSTER_MAX;
1382 }
1383
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001385
1386 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001387 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001388 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001389 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001390
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001393 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1394 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001395
1396 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1397 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1398
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001399 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001400 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001401 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001402 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001403 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001404 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001405 }
Hillf Dantond563c052012-03-21 16:34:02 -07001406 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001407
Hillf Dantond563c052012-03-21 16:34:02 -07001408 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001409 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001410
Minchan Kim02c6de82012-10-08 16:31:55 -07001411 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001412 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1413 &nr_writeback, &nr_immediate,
1414 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001415
Hugh Dickins3f797682012-01-12 17:20:07 -08001416 spin_lock_irq(&zone->lru_lock);
1417
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001418 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001419
Ying Han904249a2012-04-25 16:01:48 -07001420 if (global_reclaim(sc)) {
1421 if (current_is_kswapd())
1422 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1423 nr_reclaimed);
1424 else
1425 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1426 nr_reclaimed);
1427 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001428
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001429 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001430
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001431 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001432
1433 spin_unlock_irq(&zone->lru_lock);
1434
1435 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001436
Mel Gorman92df3a72011-10-31 17:07:56 -07001437 /*
1438 * If reclaim is isolating dirty pages under writeback, it implies
1439 * that the long-lived page allocation rate is exceeding the page
1440 * laundering rate. Either the global limits are not being effective
1441 * at throttling processes due to the page distribution throughout
1442 * zones or there is heavy usage of a slow backing device. The
1443 * only option is to throttle from reclaim context which is not ideal
1444 * as there is no guarantee the dirtying process is throttled in the
1445 * same way balance_dirty_pages() manages.
1446 *
1447 * This scales the number of dirty pages that must be under writeback
Mel Gorman8e950282013-07-03 15:02:02 -07001448 * before a zone gets flagged ZONE_WRITEBACK. It is a simple backoff
Mel Gorman92df3a72011-10-31 17:07:56 -07001449 * function that has the most effect in the range DEF_PRIORITY to
1450 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1451 * in trouble and reclaim is considered to be in trouble.
1452 *
1453 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1454 * DEF_PRIORITY-1 50% must be PageWriteback
1455 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1456 * ...
1457 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1458 * isolated page is PageWriteback
Mel Gorman8e950282013-07-03 15:02:02 -07001459 *
1460 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1461 * of pages under pages flagged for immediate reclaim and stall if any
1462 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001463 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001464 if (nr_writeback && nr_writeback >=
Mel Gorman8e950282013-07-03 15:02:02 -07001465 (nr_taken >> (DEF_PRIORITY - sc->priority)))
Mel Gorman283aba92013-07-03 15:01:51 -07001466 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001467
Mel Gormand43006d2013-07-03 15:01:50 -07001468 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001469 * memcg will stall in page writeback so only consider forcibly
1470 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001471 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001472 if (global_reclaim(sc)) {
1473 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001474 * Tag a zone as congested if all the dirty pages scanned were
1475 * backed by a congested BDI and wait_iff_congested will stall.
1476 */
1477 if (nr_dirty && nr_dirty == nr_congested)
1478 zone_set_flag(zone, ZONE_CONGESTED);
1479
1480 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001481 * If dirty pages are scanned that are not queued for IO, it
1482 * implies that flushers are not keeping up. In this case, flag
1483 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1484 * pages from reclaim context. It will forcibly stall in the
1485 * next check.
1486 */
1487 if (nr_unqueued_dirty == nr_taken)
1488 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1489
1490 /*
1491 * In addition, if kswapd scans pages marked marked for
1492 * immediate reclaim and under writeback (nr_immediate), it
1493 * implies that pages are cycling through the LRU faster than
1494 * they are written so also forcibly stall.
1495 */
1496 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1497 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001498 }
Mel Gormand43006d2013-07-03 15:01:50 -07001499
Mel Gorman8e950282013-07-03 15:02:02 -07001500 /*
1501 * Stall direct reclaim for IO completions if underlying BDIs or zone
1502 * is congested. Allow kswapd to continue until it starts encountering
1503 * unqueued dirty pages or cycling through the LRU too quickly.
1504 */
1505 if (!sc->hibernation_mode && !current_is_kswapd())
1506 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1507
Mel Gormane11da5b2010-10-26 14:21:40 -07001508 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1509 zone_idx(zone),
1510 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001511 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001512 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001513 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514}
1515
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001516/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 * This moves pages from the active list to the inactive list.
1518 *
1519 * We move them the other way if the page is referenced by one or more
1520 * processes, from rmap.
1521 *
1522 * If the pages are mostly unmapped, the processing is fast and it is
1523 * appropriate to hold zone->lru_lock across the whole operation. But if
1524 * the pages are mapped, the processing is slow (page_referenced()) so we
1525 * should drop zone->lru_lock around each page. It's impossible to balance
1526 * this, so instead we remove the pages from the LRU while processing them.
1527 * It is safe to rely on PG_active against the non-LRU pages in here because
1528 * nobody will play with that bit on a non-LRU page.
1529 *
1530 * The downside is that we have to touch page->_count against each page.
1531 * But we had to alter page->flags anyway.
1532 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001533
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001534static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001535 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001536 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001537 enum lru_list lru)
1538{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001539 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001540 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001541 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001542 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001543
Wu Fengguang3eb41402009-06-16 15:33:13 -07001544 while (!list_empty(list)) {
1545 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001546 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001547
1548 VM_BUG_ON(PageLRU(page));
1549 SetPageLRU(page);
1550
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001551 nr_pages = hpage_nr_pages(page);
1552 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001553 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001554 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001555
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001556 if (put_page_testzero(page)) {
1557 __ClearPageLRU(page);
1558 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001559 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001560
1561 if (unlikely(PageCompound(page))) {
1562 spin_unlock_irq(&zone->lru_lock);
1563 (*get_compound_page_dtor(page))(page);
1564 spin_lock_irq(&zone->lru_lock);
1565 } else
1566 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001567 }
1568 }
1569 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1570 if (!is_active_lru(lru))
1571 __count_vm_events(PGDEACTIVATE, pgmoved);
1572}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001573
Hugh Dickinsf6260122012-01-12 17:20:06 -08001574static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001575 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001576 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001577 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001579 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001580 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001581 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001583 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001584 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001586 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001587 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001588 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001589 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001590 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591
1592 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001593
1594 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001595 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001596 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001597 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001598
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001600
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001601 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1602 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001603 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001604 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001605
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001606 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001607
Hugh Dickinsf6260122012-01-12 17:20:06 -08001608 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001609 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001610 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 spin_unlock_irq(&zone->lru_lock);
1612
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 while (!list_empty(&l_hold)) {
1614 cond_resched();
1615 page = lru_to_page(&l_hold);
1616 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001617
Hugh Dickins39b5f292012-10-08 16:33:18 -07001618 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001619 putback_lru_page(page);
1620 continue;
1621 }
1622
Mel Gormancc715d92012-03-21 16:34:00 -07001623 if (unlikely(buffer_heads_over_limit)) {
1624 if (page_has_private(page) && trylock_page(page)) {
1625 if (page_has_private(page))
1626 try_to_release_page(page, 0);
1627 unlock_page(page);
1628 }
1629 }
1630
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001631 if (page_referenced(page, 0, sc->target_mem_cgroup,
1632 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001633 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001634 /*
1635 * Identify referenced, file-backed active pages and
1636 * give them one more trip around the active list. So
1637 * that executable code get better chances to stay in
1638 * memory under moderate memory pressure. Anon pages
1639 * are not likely to be evicted by use-once streaming
1640 * IO, plus JVM can create lots of anon VM_EXEC pages,
1641 * so we ignore them here.
1642 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001643 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001644 list_add(&page->lru, &l_active);
1645 continue;
1646 }
1647 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001648
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001649 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 list_add(&page->lru, &l_inactive);
1651 }
1652
Andrew Mortonb5557492009-01-06 14:40:13 -08001653 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001654 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001655 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001656 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001657 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001658 * Count referenced pages from currently used mappings as rotated,
1659 * even though only some of them are actually re-activated. This
1660 * helps balance scan pressure between file and anonymous pages in
1661 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001662 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001663 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001664
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001665 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1666 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001667 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001668 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001669
1670 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671}
1672
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001673#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001674static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001675{
1676 unsigned long active, inactive;
1677
1678 active = zone_page_state(zone, NR_ACTIVE_ANON);
1679 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1680
1681 if (inactive * zone->inactive_ratio < active)
1682 return 1;
1683
1684 return 0;
1685}
1686
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001687/**
1688 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001689 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001690 *
1691 * Returns true if the zone does not have enough inactive anon pages,
1692 * meaning some active anon pages need to be deactivated.
1693 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001694static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001695{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001696 /*
1697 * If we don't have swap space, anonymous page deactivation
1698 * is pointless.
1699 */
1700 if (!total_swap_pages)
1701 return 0;
1702
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001703 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001704 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001705
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001706 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001707}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001708#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001709static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001710{
1711 return 0;
1712}
1713#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001714
Rik van Riel56e49d22009-06-16 15:32:28 -07001715/**
1716 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001717 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001718 *
1719 * When the system is doing streaming IO, memory pressure here
1720 * ensures that active file pages get deactivated, until more
1721 * than half of the file pages are on the inactive list.
1722 *
1723 * Once we get to that situation, protect the system's working
1724 * set from being evicted by disabling active file page aging.
1725 *
1726 * This uses a different ratio than the anonymous pages, because
1727 * the page cache uses a use-once replacement algorithm.
1728 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001729static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001730{
Johannes Weinere3790142013-02-22 16:35:19 -08001731 unsigned long inactive;
1732 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001733
Johannes Weinere3790142013-02-22 16:35:19 -08001734 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1735 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1736
1737 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001738}
1739
Hugh Dickins75b00af2012-05-29 15:07:09 -07001740static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001741{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001742 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001743 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001744 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001745 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001746}
1747
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001748static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001749 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001750{
Rik van Rielb39415b2009-12-14 17:59:48 -08001751 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001752 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001753 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001754 return 0;
1755 }
1756
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001757 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001758}
1759
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001760static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001761{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001762 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001763 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001764 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001765}
1766
Johannes Weiner9a265112013-02-22 16:32:17 -08001767enum scan_balance {
1768 SCAN_EQUAL,
1769 SCAN_FRACT,
1770 SCAN_ANON,
1771 SCAN_FILE,
1772};
1773
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001775 * Determine how aggressively the anon and file LRU lists should be
1776 * scanned. The relative value of each set of LRU lists is determined
1777 * by looking at the fraction of the pages scanned we did rotate back
1778 * onto the active list instead of evict.
1779 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001780 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1781 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001782 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001783static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001784 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001785{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001786 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001787 u64 fraction[2];
1788 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001789 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001790 unsigned long anon_prio, file_prio;
1791 enum scan_balance scan_balance;
1792 unsigned long anon, file, free;
1793 bool force_scan = false;
1794 unsigned long ap, fp;
1795 enum lru_list lru;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001796
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001797 /*
1798 * If the zone or memcg is small, nr[l] can be 0. This
1799 * results in no scanning on this priority and a potential
1800 * priority drop. Global direct reclaim can go to the next
1801 * zone and tends to have no problems. Global kswapd is for
1802 * zone balancing and it needs to scan a minimum amount. When
1803 * reclaiming for a memcg, a priority drop can cause high
1804 * latencies, so it's better to scan a minimum amount there as
1805 * well.
1806 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001807 if (current_is_kswapd() && zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001808 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001809 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001810 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001811
1812 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001813 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001814 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001815 goto out;
1816 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001817
Johannes Weiner10316b32013-02-22 16:32:14 -08001818 /*
1819 * Global reclaim will swap to prevent OOM even with no
1820 * swappiness, but memcg users want to use this knob to
1821 * disable swapping for individual groups completely when
1822 * using the memory controller's swap limit feature would be
1823 * too expensive.
1824 */
1825 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001826 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001827 goto out;
1828 }
1829
1830 /*
1831 * Do not apply any pressure balancing cleverness when the
1832 * system is close to OOM, scan both anon and file equally
1833 * (unless the swappiness setting disagrees with swapping).
1834 */
1835 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001836 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001837 goto out;
1838 }
1839
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001840 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1841 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1842 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1843 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001844
Johannes Weiner11d16c22013-02-22 16:32:15 -08001845 /*
1846 * If it's foreseeable that reclaiming the file cache won't be
1847 * enough to get the zone back into a desirable shape, we have
1848 * to swap. Better start now and leave the - probably heavily
1849 * thrashing - remaining file pages alone.
1850 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001851 if (global_reclaim(sc)) {
Johannes Weiner11d16c22013-02-22 16:32:15 -08001852 free = zone_page_state(zone, NR_FREE_PAGES);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001853 if (unlikely(file + free <= high_wmark_pages(zone))) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001854 scan_balance = SCAN_ANON;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001855 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001856 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001857 }
1858
1859 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001860 * There is enough inactive page cache, do not reclaim
1861 * anything from the anonymous working set right now.
1862 */
1863 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001864 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001865 goto out;
1866 }
1867
Johannes Weiner9a265112013-02-22 16:32:17 -08001868 scan_balance = SCAN_FRACT;
1869
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001870 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001871 * With swappiness at 100, anonymous and file have the same priority.
1872 * This scanning priority is essentially the inverse of IO cost.
1873 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001874 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001875 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001876
1877 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001878 * OK, so we have swap space and a fair amount of page cache
1879 * pages. We use the recently rotated / recently scanned
1880 * ratios to determine how valuable each cache is.
1881 *
1882 * Because workloads change over time (and to avoid overflow)
1883 * we keep these statistics as a floating average, which ends
1884 * up weighing recent references more than old ones.
1885 *
1886 * anon in [0], file in [1]
1887 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001888 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001889 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001890 reclaim_stat->recent_scanned[0] /= 2;
1891 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001892 }
1893
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001894 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001895 reclaim_stat->recent_scanned[1] /= 2;
1896 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001897 }
1898
1899 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001900 * The amount of pressure on anon vs file pages is inversely
1901 * proportional to the fraction of recently scanned pages on
1902 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001903 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001904 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001905 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001906
Satoru Moriyafe350042012-05-29 15:06:47 -07001907 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001908 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001909 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001910
Shaohua Li76a33fc2010-05-24 14:32:36 -07001911 fraction[0] = ap;
1912 fraction[1] = fp;
1913 denominator = ap + fp + 1;
1914out:
Hugh Dickins41113042012-01-12 17:20:01 -08001915 for_each_evictable_lru(lru) {
1916 int file = is_file_lru(lru);
Johannes Weinerd778df52013-02-22 16:32:12 -08001917 unsigned long size;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001918 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001919
Johannes Weinerd778df52013-02-22 16:32:12 -08001920 size = get_lru_size(lruvec, lru);
Johannes Weiner10316b32013-02-22 16:32:14 -08001921 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08001922
Johannes Weiner10316b32013-02-22 16:32:14 -08001923 if (!scan && force_scan)
1924 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08001925
1926 switch (scan_balance) {
1927 case SCAN_EQUAL:
1928 /* Scan lists relative to size */
1929 break;
1930 case SCAN_FRACT:
1931 /*
1932 * Scan types proportional to swappiness and
1933 * their relative recent reclaim efficiency.
1934 */
1935 scan = div64_u64(scan * fraction[file], denominator);
1936 break;
1937 case SCAN_FILE:
1938 case SCAN_ANON:
1939 /* Scan one type exclusively */
1940 if ((scan_balance == SCAN_FILE) != file)
1941 scan = 0;
1942 break;
1943 default:
1944 /* Look ma, no brain */
1945 BUG();
1946 }
Hugh Dickins41113042012-01-12 17:20:01 -08001947 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001948 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001949}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001950
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001951/*
1952 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1953 */
1954static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1955{
1956 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07001957 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001958 unsigned long nr_to_scan;
1959 enum lru_list lru;
1960 unsigned long nr_reclaimed = 0;
1961 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1962 struct blk_plug plug;
Mel Gormane82e0562013-07-03 15:01:44 -07001963 bool scan_adjusted = false;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001964
1965 get_scan_count(lruvec, sc, nr);
1966
Mel Gormane82e0562013-07-03 15:01:44 -07001967 /* Record the original scan target for proportional adjustments later */
1968 memcpy(targets, nr, sizeof(nr));
1969
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001970 blk_start_plug(&plug);
1971 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1972 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07001973 unsigned long nr_anon, nr_file, percentage;
1974 unsigned long nr_scanned;
1975
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001976 for_each_evictable_lru(lru) {
1977 if (nr[lru]) {
1978 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1979 nr[lru] -= nr_to_scan;
1980
1981 nr_reclaimed += shrink_list(lru, nr_to_scan,
1982 lruvec, sc);
1983 }
1984 }
Mel Gormane82e0562013-07-03 15:01:44 -07001985
1986 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
1987 continue;
1988
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001989 /*
Mel Gormane82e0562013-07-03 15:01:44 -07001990 * For global direct reclaim, reclaim only the number of pages
1991 * requested. Less care is taken to scan proportionally as it
1992 * is more important to minimise direct reclaim stall latency
1993 * than it is to properly age the LRU lists.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001994 */
Mel Gormane82e0562013-07-03 15:01:44 -07001995 if (global_reclaim(sc) && !current_is_kswapd())
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001996 break;
Mel Gormane82e0562013-07-03 15:01:44 -07001997
1998 /*
1999 * For kswapd and memcg, reclaim at least the number of pages
2000 * requested. Ensure that the anon and file LRUs shrink
2001 * proportionally what was requested by get_scan_count(). We
2002 * stop reclaiming one LRU and reduce the amount scanning
2003 * proportional to the original scan target.
2004 */
2005 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2006 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2007
2008 if (nr_file > nr_anon) {
2009 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2010 targets[LRU_ACTIVE_ANON] + 1;
2011 lru = LRU_BASE;
2012 percentage = nr_anon * 100 / scan_target;
2013 } else {
2014 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2015 targets[LRU_ACTIVE_FILE] + 1;
2016 lru = LRU_FILE;
2017 percentage = nr_file * 100 / scan_target;
2018 }
2019
2020 /* Stop scanning the smaller of the LRU */
2021 nr[lru] = 0;
2022 nr[lru + LRU_ACTIVE] = 0;
2023
2024 /*
2025 * Recalculate the other LRU scan count based on its original
2026 * scan target and the percentage scanning already complete
2027 */
2028 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2029 nr_scanned = targets[lru] - nr[lru];
2030 nr[lru] = targets[lru] * (100 - percentage) / 100;
2031 nr[lru] -= min(nr[lru], nr_scanned);
2032
2033 lru += LRU_ACTIVE;
2034 nr_scanned = targets[lru] - nr[lru];
2035 nr[lru] = targets[lru] * (100 - percentage) / 100;
2036 nr[lru] -= min(nr[lru], nr_scanned);
2037
2038 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002039 }
2040 blk_finish_plug(&plug);
2041 sc->nr_reclaimed += nr_reclaimed;
2042
2043 /*
2044 * Even if we did not try to evict anon pages at all, we want to
2045 * rebalance the anon lru active/inactive ratio.
2046 */
2047 if (inactive_anon_is_low(lruvec))
2048 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2049 sc, LRU_ACTIVE_ANON);
2050
2051 throttle_vm_writeout(sc->gfp_mask);
2052}
2053
Mel Gorman23b9da52012-05-29 15:06:20 -07002054/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002055static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002056{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002057 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002058 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002059 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002060 return true;
2061
2062 return false;
2063}
2064
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002065/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002066 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2067 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2068 * true if more pages should be reclaimed such that when the page allocator
2069 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2070 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002071 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002072static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002073 unsigned long nr_reclaimed,
2074 unsigned long nr_scanned,
2075 struct scan_control *sc)
2076{
2077 unsigned long pages_for_compaction;
2078 unsigned long inactive_lru_pages;
2079
2080 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002081 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002082 return false;
2083
Mel Gorman28765922011-02-25 14:44:20 -08002084 /* Consider stopping depending on scan and reclaim activity */
2085 if (sc->gfp_mask & __GFP_REPEAT) {
2086 /*
2087 * For __GFP_REPEAT allocations, stop reclaiming if the
2088 * full LRU list has been scanned and we are still failing
2089 * to reclaim pages. This full LRU scan is potentially
2090 * expensive but a __GFP_REPEAT caller really wants to succeed
2091 */
2092 if (!nr_reclaimed && !nr_scanned)
2093 return false;
2094 } else {
2095 /*
2096 * For non-__GFP_REPEAT allocations which can presumably
2097 * fail without consequence, stop if we failed to reclaim
2098 * any pages from the last SWAP_CLUSTER_MAX number of
2099 * pages that were scanned. This will return to the
2100 * caller faster at the risk reclaim/compaction and
2101 * the resulting allocation attempt fails
2102 */
2103 if (!nr_reclaimed)
2104 return false;
2105 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002106
2107 /*
2108 * If we have not reclaimed enough pages for compaction and the
2109 * inactive lists are large enough, continue reclaiming
2110 */
2111 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002112 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002113 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002114 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002115 if (sc->nr_reclaimed < pages_for_compaction &&
2116 inactive_lru_pages > pages_for_compaction)
2117 return true;
2118
2119 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002120 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002121 case COMPACT_PARTIAL:
2122 case COMPACT_CONTINUE:
2123 return false;
2124 default:
2125 return true;
2126 }
2127}
2128
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002129static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002130{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002131 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002132
Johannes Weiner56600482012-01-12 17:17:59 -08002133 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002134 struct mem_cgroup *root = sc->target_mem_cgroup;
2135 struct mem_cgroup_reclaim_cookie reclaim = {
2136 .zone = zone,
2137 .priority = sc->priority,
2138 };
2139 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002140
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002141 nr_reclaimed = sc->nr_reclaimed;
2142 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002143
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002144 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2145 do {
2146 struct lruvec *lruvec;
2147
2148 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2149
2150 shrink_lruvec(lruvec, sc);
2151
2152 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002153 * Direct reclaim and kswapd have to scan all memory
2154 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002155 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002156 *
2157 * Limit reclaim, on the other hand, only cares about
2158 * nr_to_reclaim pages to be reclaimed and it will
2159 * retry with decreasing priority if one round over the
2160 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002161 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002162 if (!global_reclaim(sc) &&
2163 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002164 mem_cgroup_iter_break(root, memcg);
2165 break;
2166 }
2167 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2168 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002169
2170 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2171 sc->nr_scanned - nr_scanned,
2172 sc->nr_reclaimed - nr_reclaimed);
2173
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002174 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2175 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002176}
2177
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002178/* Returns true if compaction should go ahead for a high-order request */
2179static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2180{
2181 unsigned long balance_gap, watermark;
2182 bool watermark_ok;
2183
2184 /* Do not consider compaction for orders reclaim is meant to satisfy */
2185 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2186 return false;
2187
2188 /*
2189 * Compaction takes time to run and there are potentially other
2190 * callers using the pages just freed. Continue reclaiming until
2191 * there is a buffer of free pages available to give compaction
2192 * a reasonable chance of completing and allocating the page
2193 */
2194 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002195 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002196 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2197 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2198 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2199
2200 /*
2201 * If compaction is deferred, reclaim up to a point where
2202 * compaction will have a chance of success when re-enabled
2203 */
Rik van Rielaff62242012-03-21 16:33:52 -07002204 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002205 return watermark_ok;
2206
2207 /* If compaction is not ready to start, keep reclaiming */
2208 if (!compaction_suitable(zone, sc->order))
2209 return false;
2210
2211 return watermark_ok;
2212}
2213
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214/*
2215 * This is the direct reclaim path, for page-allocating processes. We only
2216 * try to reclaim pages from zones which will satisfy the caller's allocation
2217 * request.
2218 *
Mel Gorman41858962009-06-16 15:32:12 -07002219 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2220 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2222 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002223 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2224 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2225 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 * If a zone is deemed to be full of pinned pages then just give it a light
2228 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002229 *
2230 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002231 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002232 * the caller that it should consider retrying the allocation instead of
2233 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002235static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236{
Mel Gormandd1a2392008-04-28 02:12:17 -07002237 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002238 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002239 unsigned long nr_soft_reclaimed;
2240 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002241 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002242
Mel Gormancc715d92012-03-21 16:34:00 -07002243 /*
2244 * If the number of buffer_heads in the machine exceeds the maximum
2245 * allowed level, force direct reclaim to scan the highmem zone as
2246 * highmem pages could be pinning lowmem pages storing buffer_heads
2247 */
2248 if (buffer_heads_over_limit)
2249 sc->gfp_mask |= __GFP_HIGHMEM;
2250
Mel Gormand4debc62010-08-09 17:19:29 -07002251 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2252 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002253 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002255 /*
2256 * Take care memory controller reclaiming has small influence
2257 * to global LRU.
2258 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002259 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002260 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2261 continue;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002262 if (zone->all_unreclaimable &&
2263 sc->priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002264 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002265 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002266 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002267 * If we already have plenty of memory free for
2268 * compaction in this zone, don't free any more.
2269 * Even though compaction is invoked for any
2270 * non-zero order, only frequent costly order
2271 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002272 * noticeable problem, like transparent huge
2273 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002274 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002275 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002276 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002277 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002278 }
Rik van Riele0887c12011-10-31 17:09:31 -07002279 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002280 /*
2281 * This steals pages from memory cgroups over softlimit
2282 * and returns the number of reclaimed pages and
2283 * scanned pages. This works for global memory pressure
2284 * and balancing, not for a memcg's limit.
2285 */
2286 nr_soft_scanned = 0;
2287 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2288 sc->order, sc->gfp_mask,
2289 &nr_soft_scanned);
2290 sc->nr_reclaimed += nr_soft_reclaimed;
2291 sc->nr_scanned += nr_soft_scanned;
2292 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002293 }
Nick Piggin408d8542006-09-25 23:31:27 -07002294
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002295 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 }
Mel Gormane0c23272011-10-31 17:09:33 -07002297
Mel Gorman0cee34f2012-01-12 17:19:49 -08002298 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002299}
2300
2301static bool zone_reclaimable(struct zone *zone)
2302{
2303 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2304}
2305
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002306/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002307static bool all_unreclaimable(struct zonelist *zonelist,
2308 struct scan_control *sc)
2309{
2310 struct zoneref *z;
2311 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002312
2313 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2314 gfp_zone(sc->gfp_mask), sc->nodemask) {
2315 if (!populated_zone(zone))
2316 continue;
2317 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2318 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002319 if (!zone->all_unreclaimable)
2320 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002321 }
2322
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002323 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002325
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326/*
2327 * This is the main entry point to direct page reclaim.
2328 *
2329 * If a full scan of the inactive list fails to free enough memory then we
2330 * are "out of memory" and something needs to be killed.
2331 *
2332 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2333 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002334 * caller can't do much about. We kick the writeback threads and take explicit
2335 * naps in the hope that some of these pages can be written. But if the
2336 * allocating task holds filesystem locks which prevent writeout this might not
2337 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002338 *
2339 * returns: 0, if no pages reclaimed
2340 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 */
Mel Gormandac1d272008-04-28 02:12:12 -07002342static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002343 struct scan_control *sc,
2344 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345{
Andrew Morton69e05942006-03-22 00:08:19 -08002346 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002348 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002349 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002350 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002351 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
Keika Kobayashi873b4772008-07-25 01:48:52 -07002353 delayacct_freepages_start();
2354
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002355 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002356 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002358 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002359 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2360 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002361 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002362 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002363
Balbir Singh66e17072008-02-07 00:13:56 -08002364 /*
2365 * Don't shrink slabs when reclaiming memory from
2366 * over limit cgroups
2367 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002368 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002369 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002370 for_each_zone_zonelist(zone, z, zonelist,
2371 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002372 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2373 continue;
2374
2375 lru_pages += zone_reclaimable_pages(zone);
2376 }
2377
Ying Han1495f232011-05-24 17:12:27 -07002378 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002379 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002380 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002381 reclaim_state->reclaimed_slab = 0;
2382 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 }
Balbir Singh66e17072008-02-07 00:13:56 -08002384 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002385 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
2388 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002389 * If we're getting trouble reclaiming, start doing
2390 * writepage even in laptop mode.
2391 */
2392 if (sc->priority < DEF_PRIORITY - 2)
2393 sc->may_writepage = 1;
2394
2395 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 * Try to write back as many pages as we just scanned. This
2397 * tends to cause slow streaming writers to write data to the
2398 * disk smoothly, at the dirtying rate, which is nice. But
2399 * that's undesirable in laptop mode, where we *want* lumpy
2400 * writeout. So in laptop mode, write out the whole world.
2401 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002402 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2403 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002404 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2405 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002406 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 }
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002408 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002409
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002411 delayacct_freepages_end();
2412
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002413 if (sc->nr_reclaimed)
2414 return sc->nr_reclaimed;
2415
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002416 /*
2417 * As hibernation is going on, kswapd is freezed so that it can't mark
2418 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2419 * check.
2420 */
2421 if (oom_killer_disabled)
2422 return 0;
2423
Mel Gorman0cee34f2012-01-12 17:19:49 -08002424 /* Aborted reclaim to try compaction? don't OOM, then */
2425 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002426 return 1;
2427
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002428 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002429 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002430 return 1;
2431
2432 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433}
2434
Mel Gorman55150612012-07-31 16:44:35 -07002435static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2436{
2437 struct zone *zone;
2438 unsigned long pfmemalloc_reserve = 0;
2439 unsigned long free_pages = 0;
2440 int i;
2441 bool wmark_ok;
2442
2443 for (i = 0; i <= ZONE_NORMAL; i++) {
2444 zone = &pgdat->node_zones[i];
2445 pfmemalloc_reserve += min_wmark_pages(zone);
2446 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2447 }
2448
2449 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2450
2451 /* kswapd must be awake if processes are being throttled */
2452 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2453 pgdat->classzone_idx = min(pgdat->classzone_idx,
2454 (enum zone_type)ZONE_NORMAL);
2455 wake_up_interruptible(&pgdat->kswapd_wait);
2456 }
2457
2458 return wmark_ok;
2459}
2460
2461/*
2462 * Throttle direct reclaimers if backing storage is backed by the network
2463 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2464 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002465 * when the low watermark is reached.
2466 *
2467 * Returns true if a fatal signal was delivered during throttling. If this
2468 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002469 */
Mel Gorman50694c22012-11-26 16:29:48 -08002470static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002471 nodemask_t *nodemask)
2472{
2473 struct zone *zone;
2474 int high_zoneidx = gfp_zone(gfp_mask);
2475 pg_data_t *pgdat;
2476
2477 /*
2478 * Kernel threads should not be throttled as they may be indirectly
2479 * responsible for cleaning pages necessary for reclaim to make forward
2480 * progress. kjournald for example may enter direct reclaim while
2481 * committing a transaction where throttling it could forcing other
2482 * processes to block on log_wait_commit().
2483 */
2484 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002485 goto out;
2486
2487 /*
2488 * If a fatal signal is pending, this process should not throttle.
2489 * It should return quickly so it can exit and free its memory
2490 */
2491 if (fatal_signal_pending(current))
2492 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002493
2494 /* Check if the pfmemalloc reserves are ok */
2495 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2496 pgdat = zone->zone_pgdat;
2497 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002498 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002499
Mel Gorman68243e72012-07-31 16:44:39 -07002500 /* Account for the throttling */
2501 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2502
Mel Gorman55150612012-07-31 16:44:35 -07002503 /*
2504 * If the caller cannot enter the filesystem, it's possible that it
2505 * is due to the caller holding an FS lock or performing a journal
2506 * transaction in the case of a filesystem like ext[3|4]. In this case,
2507 * it is not safe to block on pfmemalloc_wait as kswapd could be
2508 * blocked waiting on the same lock. Instead, throttle for up to a
2509 * second before continuing.
2510 */
2511 if (!(gfp_mask & __GFP_FS)) {
2512 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2513 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002514
2515 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002516 }
2517
2518 /* Throttle until kswapd wakes the process */
2519 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2520 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002521
2522check_pending:
2523 if (fatal_signal_pending(current))
2524 return true;
2525
2526out:
2527 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002528}
2529
Mel Gormandac1d272008-04-28 02:12:12 -07002530unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002531 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002532{
Mel Gorman33906bc2010-08-09 17:19:16 -07002533 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002534 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002535 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002536 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002537 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002538 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002539 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002540 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002541 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002542 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002543 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002544 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002545 struct shrink_control shrink = {
2546 .gfp_mask = sc.gfp_mask,
2547 };
Balbir Singh66e17072008-02-07 00:13:56 -08002548
Mel Gorman55150612012-07-31 16:44:35 -07002549 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002550 * Do not enter reclaim if fatal signal was delivered while throttled.
2551 * 1 is returned so that the page allocator does not OOM kill at this
2552 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002553 */
Mel Gorman50694c22012-11-26 16:29:48 -08002554 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002555 return 1;
2556
Mel Gorman33906bc2010-08-09 17:19:16 -07002557 trace_mm_vmscan_direct_reclaim_begin(order,
2558 sc.may_writepage,
2559 gfp_mask);
2560
Ying Hana09ed5e2011-05-24 17:12:26 -07002561 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002562
2563 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2564
2565 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002566}
2567
Andrew Mortonc255a452012-07-31 16:43:02 -07002568#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002569
Johannes Weiner72835c82012-01-12 17:18:32 -08002570unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002571 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002572 struct zone *zone,
2573 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002574{
2575 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002576 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002577 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002578 .may_writepage = !laptop_mode,
2579 .may_unmap = 1,
2580 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002581 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002582 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002583 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002584 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002585 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002586
Balbir Singh4e416952009-09-23 15:56:39 -07002587 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2588 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002589
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002590 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002591 sc.may_writepage,
2592 sc.gfp_mask);
2593
Balbir Singh4e416952009-09-23 15:56:39 -07002594 /*
2595 * NOTE: Although we can get the priority field, using it
2596 * here is not a good idea, since it limits the pages we can scan.
2597 * if we don't reclaim here, the shrink_zone from balance_pgdat
2598 * will pick up pages from other mem cgroup's as well. We hack
2599 * the priority and make it zero.
2600 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002601 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002602
2603 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2604
Ying Han0ae5e892011-05-26 16:25:25 -07002605 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002606 return sc.nr_reclaimed;
2607}
2608
Johannes Weiner72835c82012-01-12 17:18:32 -08002609unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002610 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002611 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002612{
Balbir Singh4e416952009-09-23 15:56:39 -07002613 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002614 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002615 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002616 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002617 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002618 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002619 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002620 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002621 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002622 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002623 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002624 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002625 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2626 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2627 };
2628 struct shrink_control shrink = {
2629 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002630 };
Balbir Singh66e17072008-02-07 00:13:56 -08002631
Ying Han889976d2011-05-26 16:25:33 -07002632 /*
2633 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2634 * take care of from where we get pages. So the node where we start the
2635 * scan does not need to be the current node.
2636 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002637 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002638
2639 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002640
2641 trace_mm_vmscan_memcg_reclaim_begin(0,
2642 sc.may_writepage,
2643 sc.gfp_mask);
2644
Ying Hana09ed5e2011-05-24 17:12:26 -07002645 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002646
2647 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2648
2649 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002650}
2651#endif
2652
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002653static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002654{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002655 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002656
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002657 if (!total_swap_pages)
2658 return;
2659
2660 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2661 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002662 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002663
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002664 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002665 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002666 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002667
2668 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2669 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002670}
2671
Johannes Weiner60cefed2012-11-29 13:54:23 -08002672static bool zone_balanced(struct zone *zone, int order,
2673 unsigned long balance_gap, int classzone_idx)
2674{
2675 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2676 balance_gap, classzone_idx, 0))
2677 return false;
2678
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002679 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2680 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002681 return false;
2682
2683 return true;
2684}
2685
Mel Gorman1741c872011-01-13 15:46:21 -08002686/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002687 * pgdat_balanced() is used when checking if a node is balanced.
2688 *
2689 * For order-0, all zones must be balanced!
2690 *
2691 * For high-order allocations only zones that meet watermarks and are in a
2692 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2693 * total of balanced pages must be at least 25% of the zones allowed by
2694 * classzone_idx for the node to be considered balanced. Forcing all zones to
2695 * be balanced for high orders can cause excessive reclaim when there are
2696 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002697 * The choice of 25% is due to
2698 * o a 16M DMA zone that is balanced will not balance a zone on any
2699 * reasonable sized machine
2700 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002701 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002702 * would need to be at least 256M for it to be balance a whole node.
2703 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2704 * to balance a node on its own. These seemed like reasonable ratios.
2705 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002706static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002707{
Jiang Liub40da042013-02-22 16:33:52 -08002708 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002709 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002710 int i;
2711
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002712 /* Check the watermark levels */
2713 for (i = 0; i <= classzone_idx; i++) {
2714 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002715
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002716 if (!populated_zone(zone))
2717 continue;
2718
Jiang Liub40da042013-02-22 16:33:52 -08002719 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002720
2721 /*
2722 * A special case here:
2723 *
2724 * balance_pgdat() skips over all_unreclaimable after
2725 * DEF_PRIORITY. Effectively, it considers them balanced so
2726 * they must be considered balanced here as well!
2727 */
2728 if (zone->all_unreclaimable) {
Jiang Liub40da042013-02-22 16:33:52 -08002729 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002730 continue;
2731 }
2732
2733 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002734 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002735 else if (!order)
2736 return false;
2737 }
2738
2739 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002740 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002741 else
2742 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002743}
2744
Mel Gorman55150612012-07-31 16:44:35 -07002745/*
2746 * Prepare kswapd for sleeping. This verifies that there are no processes
2747 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2748 *
2749 * Returns true if kswapd is ready to sleep
2750 */
2751static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002752 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002753{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002754 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2755 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002756 return false;
2757
2758 /*
2759 * There is a potential race between when kswapd checks its watermarks
2760 * and a process gets throttled. There is also a potential race if
2761 * processes get throttled, kswapd wakes, a large process exits therby
2762 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2763 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2764 * so wake them now if necessary. If necessary, processes will wake
2765 * kswapd and get throttled again
2766 */
2767 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2768 wake_up(&pgdat->pfmemalloc_wait);
2769 return false;
2770 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002771
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002772 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002773}
2774
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775/*
Mel Gorman75485362013-07-03 15:01:42 -07002776 * kswapd shrinks the zone by the number of pages required to reach
2777 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002778 *
2779 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002780 * reclaim or if the lack of progress was due to pages under writeback.
2781 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002782 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002783static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002784 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002785 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002786 unsigned long lru_pages,
2787 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002788{
2789 unsigned long nr_slab;
Mel Gorman7c954f62013-07-03 15:01:54 -07002790 int testorder = sc->order;
2791 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002792 struct reclaim_state *reclaim_state = current->reclaim_state;
2793 struct shrink_control shrink = {
2794 .gfp_mask = sc->gfp_mask,
2795 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002796 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002797
2798 /* Reclaim above the high watermark. */
2799 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002800
2801 /*
2802 * Kswapd reclaims only single pages with compaction enabled. Trying
2803 * too hard to reclaim until contiguous free pages have become
2804 * available can hurt performance by evicting too much useful data
2805 * from memory. Do not reclaim more than needed for compaction.
2806 */
2807 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2808 compaction_suitable(zone, sc->order) !=
2809 COMPACT_SKIPPED)
2810 testorder = 0;
2811
2812 /*
2813 * We put equal pressure on every zone, unless one zone has way too
2814 * many pages free already. The "too many pages" is defined as the
2815 * high wmark plus a "gap" where the gap is either the low
2816 * watermark or 1% of the zone, whichever is smaller.
2817 */
2818 balance_gap = min(low_wmark_pages(zone),
2819 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2820 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2821
2822 /*
2823 * If there is no low memory pressure or the zone is balanced then no
2824 * reclaim is necessary
2825 */
2826 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2827 if (!lowmem_pressure && zone_balanced(zone, testorder,
2828 balance_gap, classzone_idx))
2829 return true;
2830
Mel Gorman75485362013-07-03 15:01:42 -07002831 shrink_zone(zone, sc);
2832
2833 reclaim_state->reclaimed_slab = 0;
2834 nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2835 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2836
Mel Gorman2ab44f42013-07-03 15:01:47 -07002837 /* Account for the number of pages attempted to reclaim */
2838 *nr_attempted += sc->nr_to_reclaim;
2839
Mel Gorman75485362013-07-03 15:01:42 -07002840 if (nr_slab == 0 && !zone_reclaimable(zone))
2841 zone->all_unreclaimable = 1;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002842
Mel Gorman283aba92013-07-03 15:01:51 -07002843 zone_clear_flag(zone, ZONE_WRITEBACK);
2844
Mel Gorman7c954f62013-07-03 15:01:54 -07002845 /*
2846 * If a zone reaches its high watermark, consider it to be no longer
2847 * congested. It's possible there are dirty pages backed by congested
2848 * BDIs but as pressure is relieved, speculatively avoid congestion
2849 * waits.
2850 */
2851 if (!zone->all_unreclaimable &&
2852 zone_balanced(zone, testorder, 0, classzone_idx)) {
2853 zone_clear_flag(zone, ZONE_CONGESTED);
2854 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2855 }
2856
Mel Gormanb8e83b92013-07-03 15:01:45 -07002857 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002858}
2859
2860/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002862 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002864 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 *
2866 * There is special handling here for zones which are full of pinned pages.
2867 * This can happen if the pages are all mlocked, or if they are all used by
2868 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2869 * What we do is to detect the case where all pages in the zone have been
2870 * scanned twice and there has been zero successful reclaim. Mark the zone as
2871 * dead and from now on, only perform a short scan. Basically we're polling
2872 * the zone for when the problem goes away.
2873 *
2874 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002875 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2876 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2877 * lower zones regardless of the number of free pages in the lower zones. This
2878 * interoperates with the page allocator fallback scheme to ensure that aging
2879 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 */
Mel Gorman99504742011-01-13 15:46:20 -08002881static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002882 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002885 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Ying Han0ae5e892011-05-26 16:25:25 -07002886 unsigned long nr_soft_reclaimed;
2887 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002888 struct scan_control sc = {
2889 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002890 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002891 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002892 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002893 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002894 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002895 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002896 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07002897 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002899 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002901 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002902 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002903 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07002904
2905 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002907 /*
2908 * Scan in the highmem->dma direction for the highest
2909 * zone which needs scanning
2910 */
2911 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2912 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002914 if (!populated_zone(zone))
2915 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002917 if (zone->all_unreclaimable &&
2918 sc.priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002919 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920
Rik van Riel556adec2008-10-18 20:26:34 -07002921 /*
2922 * Do some background aging of the anon list, to give
2923 * pages a chance to be referenced before reclaiming.
2924 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002925 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002926
Mel Gormancc715d92012-03-21 16:34:00 -07002927 /*
2928 * If the number of buffer_heads in the machine
2929 * exceeds the maximum allowed level and this node
2930 * has a highmem zone, force kswapd to reclaim from
2931 * it to relieve lowmem pressure.
2932 */
2933 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2934 end_zone = i;
2935 break;
2936 }
2937
Johannes Weiner60cefed2012-11-29 13:54:23 -08002938 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002939 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002940 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002941 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07002942 /*
2943 * If balanced, clear the dirty and congested
2944 * flags
2945 */
Shaohua Li439423f2011-08-25 15:59:12 -07002946 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07002947 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002950
Mel Gormanb8e83b92013-07-03 15:01:45 -07002951 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002952 goto out;
2953
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 for (i = 0; i <= end_zone; i++) {
2955 struct zone *zone = pgdat->node_zones + i;
2956
Mel Gorman2ab44f42013-07-03 15:01:47 -07002957 if (!populated_zone(zone))
2958 continue;
2959
Wu Fengguangadea02a2009-09-21 17:01:42 -07002960 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07002961
2962 /*
2963 * If any zone is currently balanced then kswapd will
2964 * not call compaction as it is expected that the
2965 * necessary pages are already available.
2966 */
2967 if (pgdat_needs_compaction &&
2968 zone_watermark_ok(zone, order,
2969 low_wmark_pages(zone),
2970 *classzone_idx, 0))
2971 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 }
2973
2974 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07002975 * If we're getting trouble reclaiming, start doing writepage
2976 * even in laptop mode.
2977 */
2978 if (sc.priority < DEF_PRIORITY - 2)
2979 sc.may_writepage = 1;
2980
2981 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 * Now scan the zone in the dma->highmem direction, stopping
2983 * at the last zone which needs scanning.
2984 *
2985 * We do this because the page allocator works in the opposite
2986 * direction. This prevents the page allocator from allocating
2987 * pages behind kswapd's direction of progress, which would
2988 * cause too much scanning of the lower zones.
2989 */
2990 for (i = 0; i <= end_zone; i++) {
2991 struct zone *zone = pgdat->node_zones + i;
2992
Con Kolivasf3fe6512006-01-06 00:11:15 -08002993 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 continue;
2995
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002996 if (zone->all_unreclaimable &&
2997 sc.priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 continue;
2999
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003001
Ying Han0ae5e892011-05-26 16:25:25 -07003002 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003003 /*
3004 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07003005 */
Ying Han0ae5e892011-05-26 16:25:25 -07003006 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3007 order, sc.gfp_mask,
3008 &nr_soft_scanned);
3009 sc.nr_reclaimed += nr_soft_reclaimed;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07003010
Rik van Riel32a43302007-10-16 01:24:50 -07003011 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003012 * There should be no need to raise the scanning
3013 * priority if enough pages are already being scanned
3014 * that that high watermark would be met at 100%
3015 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003016 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003017 if (kswapd_shrink_zone(zone, end_zone, &sc,
3018 lru_pages, &nr_attempted))
3019 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 }
Mel Gorman55150612012-07-31 16:44:35 -07003021
3022 /*
3023 * If the low watermark is met there is no need for processes
3024 * to be throttled on pfmemalloc_wait as they should not be
3025 * able to safely make forward progress. Wake them
3026 */
3027 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3028 pfmemalloc_watermark_ok(pgdat))
3029 wake_up(&pgdat->pfmemalloc_wait);
3030
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003032 * Fragmentation may mean that the system cannot be rebalanced
3033 * for high-order allocations in all zones. If twice the
3034 * allocation size has been reclaimed and the zones are still
3035 * not balanced then recheck the watermarks at order-0 to
3036 * prevent kswapd reclaiming excessively. Assume that a
3037 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003039 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003040 order = sc.order = 0;
3041
Mel Gormanb8e83b92013-07-03 15:01:45 -07003042 /* Check if kswapd should be suspending */
3043 if (try_to_freeze() || kthread_should_stop())
3044 break;
3045
3046 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003047 * Compact if necessary and kswapd is reclaiming at least the
3048 * high watermark number of pages as requsted
3049 */
3050 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3051 compact_pgdat(pgdat, order);
3052
3053 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003054 * Raise priority if scanning rate is too low or there was no
3055 * progress in reclaiming pages
3056 */
3057 if (raise_priority || !sc.nr_reclaimed)
3058 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003059 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003060 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061
Mel Gormanb8e83b92013-07-03 15:01:45 -07003062out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003063 /*
Mel Gorman55150612012-07-31 16:44:35 -07003064 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003065 * makes a decision on the order we were last reclaiming at. However,
3066 * if another caller entered the allocator slow path while kswapd
3067 * was awake, order will remain at the higher level
3068 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003069 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003070 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071}
3072
Mel Gormandc83edd2011-01-13 15:46:26 -08003073static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003074{
3075 long remaining = 0;
3076 DEFINE_WAIT(wait);
3077
3078 if (freezing(current) || kthread_should_stop())
3079 return;
3080
3081 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3082
3083 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003084 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003085 remaining = schedule_timeout(HZ/10);
3086 finish_wait(&pgdat->kswapd_wait, &wait);
3087 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3088 }
3089
3090 /*
3091 * After a short sleep, check if it was a premature sleep. If not, then
3092 * go fully to sleep until explicitly woken up.
3093 */
Mel Gorman55150612012-07-31 16:44:35 -07003094 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003095 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3096
3097 /*
3098 * vmstat counters are not perfectly accurate and the estimated
3099 * value for counters such as NR_FREE_PAGES can deviate from the
3100 * true value by nr_online_cpus * threshold. To avoid the zone
3101 * watermarks being breached while under pressure, we reduce the
3102 * per-cpu vmstat threshold while kswapd is awake and restore
3103 * them before going back to sleep.
3104 */
3105 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003106
Mel Gorman62997022012-10-08 16:32:47 -07003107 /*
3108 * Compaction records what page blocks it recently failed to
3109 * isolate pages from and skips them in the future scanning.
3110 * When kswapd is going to sleep, it is reasonable to assume
3111 * that pages and compaction may succeed so reset the cache.
3112 */
3113 reset_isolation_suitable(pgdat);
3114
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003115 if (!kthread_should_stop())
3116 schedule();
3117
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003118 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3119 } else {
3120 if (remaining)
3121 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3122 else
3123 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3124 }
3125 finish_wait(&pgdat->kswapd_wait, &wait);
3126}
3127
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128/*
3129 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003130 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 *
3132 * This basically trickles out pages so that we have _some_
3133 * free memory available even if there is no other activity
3134 * that frees anything up. This is needed for things like routing
3135 * etc, where we otherwise might have all activity going on in
3136 * asynchronous contexts that cannot page things out.
3137 *
3138 * If there are applications that are active memory-allocators
3139 * (most normal use), this basically shouldn't matter.
3140 */
3141static int kswapd(void *p)
3142{
Mel Gorman215ddd62011-07-08 15:39:40 -07003143 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003144 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003145 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003146 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 pg_data_t *pgdat = (pg_data_t*)p;
3148 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003149
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150 struct reclaim_state reclaim_state = {
3151 .reclaimed_slab = 0,
3152 };
Rusty Russella70f7302009-03-13 14:49:46 +10303153 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154
Nick Piggincf40bd12009-01-21 08:12:39 +01003155 lockdep_set_current_reclaim_state(GFP_KERNEL);
3156
Rusty Russell174596a2009-01-01 10:12:29 +10303157 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003158 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 current->reclaim_state = &reclaim_state;
3160
3161 /*
3162 * Tell the memory management that we're a "memory allocator",
3163 * and that if we need more memory we should get access to it
3164 * regardless (see "__alloc_pages()"). "kswapd" should
3165 * never get caught in the normal page freeing logic.
3166 *
3167 * (Kswapd normally doesn't need memory anyway, but sometimes
3168 * you need a small amount of memory in order to be able to
3169 * page out something else, and this flag essentially protects
3170 * us from recursively trying to free more memory as we're
3171 * trying to free the first piece of memory in the first place).
3172 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003173 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003174 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175
Mel Gorman215ddd62011-07-08 15:39:40 -07003176 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003177 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003178 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003179 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003181 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003182
Mel Gorman215ddd62011-07-08 15:39:40 -07003183 /*
3184 * If the last balance_pgdat was unsuccessful it's unlikely a
3185 * new request of a similar or harder type will succeed soon
3186 * so consider going to sleep on the basis we reclaimed at
3187 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003188 if (balanced_classzone_idx >= new_classzone_idx &&
3189 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003190 new_order = pgdat->kswapd_max_order;
3191 new_classzone_idx = pgdat->classzone_idx;
3192 pgdat->kswapd_max_order = 0;
3193 pgdat->classzone_idx = pgdat->nr_zones - 1;
3194 }
3195
Mel Gorman99504742011-01-13 15:46:20 -08003196 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 /*
3198 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003199 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 */
3201 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003202 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003204 kswapd_try_to_sleep(pgdat, balanced_order,
3205 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003207 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003208 new_order = order;
3209 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003210 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003211 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213
David Rientjes8fe23e02009-12-14 17:58:33 -08003214 ret = try_to_freeze();
3215 if (kthread_should_stop())
3216 break;
3217
3218 /*
3219 * We can speed up thawing tasks if we don't call balance_pgdat
3220 * after returning from the refrigerator
3221 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003222 if (!ret) {
3223 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003224 balanced_classzone_idx = classzone_idx;
3225 balanced_order = balance_pgdat(pgdat, order,
3226 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003227 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003229
3230 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231 return 0;
3232}
3233
3234/*
3235 * A zone is low on free memory, so wake its kswapd task to service it.
3236 */
Mel Gorman99504742011-01-13 15:46:20 -08003237void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238{
3239 pg_data_t *pgdat;
3240
Con Kolivasf3fe6512006-01-06 00:11:15 -08003241 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 return;
3243
Paul Jackson02a0e532006-12-13 00:34:25 -08003244 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003246 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003247 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003248 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003249 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3250 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003251 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003253 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3254 return;
3255
3256 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003257 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258}
3259
Wu Fengguangadea02a2009-09-21 17:01:42 -07003260/*
3261 * The reclaimable count would be mostly accurate.
3262 * The less reclaimable pages may be
3263 * - mlocked pages, which will be moved to unevictable list when encountered
3264 * - mapped pages, which may require several travels to be reclaimed
3265 * - dirty pages, which is not "instantly" reclaimable
3266 */
3267unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003268{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003269 int nr;
3270
3271 nr = global_page_state(NR_ACTIVE_FILE) +
3272 global_page_state(NR_INACTIVE_FILE);
3273
Shaohua Liec8acf22013-02-22 16:34:38 -08003274 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003275 nr += global_page_state(NR_ACTIVE_ANON) +
3276 global_page_state(NR_INACTIVE_ANON);
3277
3278 return nr;
3279}
3280
3281unsigned long zone_reclaimable_pages(struct zone *zone)
3282{
3283 int nr;
3284
3285 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3286 zone_page_state(zone, NR_INACTIVE_FILE);
3287
Shaohua Liec8acf22013-02-22 16:34:38 -08003288 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003289 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3290 zone_page_state(zone, NR_INACTIVE_ANON);
3291
3292 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003293}
3294
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003295#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003297 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003298 * freed pages.
3299 *
3300 * Rather than trying to age LRUs the aim is to preserve the overall
3301 * LRU order by reclaiming preferentially
3302 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003304unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003306 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003307 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003308 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3309 .may_swap = 1,
3310 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003311 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003312 .nr_to_reclaim = nr_to_reclaim,
3313 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003314 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003315 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003317 struct shrink_control shrink = {
3318 .gfp_mask = sc.gfp_mask,
3319 };
3320 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003321 struct task_struct *p = current;
3322 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003324 p->flags |= PF_MEMALLOC;
3325 lockdep_set_current_reclaim_state(sc.gfp_mask);
3326 reclaim_state.reclaimed_slab = 0;
3327 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003328
Ying Hana09ed5e2011-05-24 17:12:26 -07003329 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003330
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003331 p->reclaim_state = NULL;
3332 lockdep_clear_current_reclaim_state();
3333 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003334
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003335 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003337#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339/* It's optimal to keep kswapds on the same CPUs as their memory, but
3340 not required for correctness. So if the last cpu in a node goes
3341 away, we get changed to run anywhere: as the first one comes back,
3342 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003343static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3344 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003346 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003348 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003349 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003350 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303351 const struct cpumask *mask;
3352
3353 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003354
Rusty Russell3e597942009-01-01 10:12:24 +10303355 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003357 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 }
3359 }
3360 return NOTIFY_OK;
3361}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
Yasunori Goto3218ae12006-06-27 02:53:33 -07003363/*
3364 * This kswapd start function will be called by init and node-hot-add.
3365 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3366 */
3367int kswapd_run(int nid)
3368{
3369 pg_data_t *pgdat = NODE_DATA(nid);
3370 int ret = 0;
3371
3372 if (pgdat->kswapd)
3373 return 0;
3374
3375 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3376 if (IS_ERR(pgdat->kswapd)) {
3377 /* failure at boot is fatal */
3378 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003379 pr_err("Failed to start kswapd on node %d\n", nid);
3380 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003381 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003382 }
3383 return ret;
3384}
3385
David Rientjes8fe23e02009-12-14 17:58:33 -08003386/*
Jiang Liud8adde12012-07-11 14:01:52 -07003387 * Called by memory hotplug when all memory in a node is offlined. Caller must
3388 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003389 */
3390void kswapd_stop(int nid)
3391{
3392 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3393
Jiang Liud8adde12012-07-11 14:01:52 -07003394 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003395 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003396 NODE_DATA(nid)->kswapd = NULL;
3397 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003398}
3399
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400static int __init kswapd_init(void)
3401{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003402 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003403
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003405 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003406 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 hotcpu_notifier(cpu_callback, 0);
3408 return 0;
3409}
3410
3411module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003412
3413#ifdef CONFIG_NUMA
3414/*
3415 * Zone reclaim mode
3416 *
3417 * If non-zero call zone_reclaim when the number of free pages falls below
3418 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003419 */
3420int zone_reclaim_mode __read_mostly;
3421
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003422#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003423#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003424#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3425#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3426
Christoph Lameter9eeff232006-01-18 17:42:31 -08003427/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003428 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3429 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3430 * a zone.
3431 */
3432#define ZONE_RECLAIM_PRIORITY 4
3433
Christoph Lameter9eeff232006-01-18 17:42:31 -08003434/*
Christoph Lameter96146342006-07-03 00:24:13 -07003435 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3436 * occur.
3437 */
3438int sysctl_min_unmapped_ratio = 1;
3439
3440/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003441 * If the number of slab pages in a zone grows beyond this percentage then
3442 * slab reclaim needs to occur.
3443 */
3444int sysctl_min_slab_ratio = 5;
3445
Mel Gorman90afa5d2009-06-16 15:33:20 -07003446static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3447{
3448 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3449 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3450 zone_page_state(zone, NR_ACTIVE_FILE);
3451
3452 /*
3453 * It's possible for there to be more file mapped pages than
3454 * accounted for by the pages on the file LRU lists because
3455 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3456 */
3457 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3458}
3459
3460/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3461static long zone_pagecache_reclaimable(struct zone *zone)
3462{
3463 long nr_pagecache_reclaimable;
3464 long delta = 0;
3465
3466 /*
3467 * If RECLAIM_SWAP is set, then all file pages are considered
3468 * potentially reclaimable. Otherwise, we have to worry about
3469 * pages like swapcache and zone_unmapped_file_pages() provides
3470 * a better estimate
3471 */
3472 if (zone_reclaim_mode & RECLAIM_SWAP)
3473 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3474 else
3475 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3476
3477 /* If we can't clean pages, remove dirty pages from consideration */
3478 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3479 delta += zone_page_state(zone, NR_FILE_DIRTY);
3480
3481 /* Watch for any possible underflows due to delta */
3482 if (unlikely(delta > nr_pagecache_reclaimable))
3483 delta = nr_pagecache_reclaimable;
3484
3485 return nr_pagecache_reclaimable - delta;
3486}
3487
Christoph Lameter0ff38492006-09-25 23:31:52 -07003488/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003489 * Try to free up some pages from this zone through reclaim.
3490 */
Andrew Morton179e9632006-03-22 00:08:18 -08003491static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003492{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003493 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003494 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003495 struct task_struct *p = current;
3496 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003497 struct scan_control sc = {
3498 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003499 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003500 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003501 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003502 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003503 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003504 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003505 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003506 struct shrink_control shrink = {
3507 .gfp_mask = sc.gfp_mask,
3508 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003509 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003510
Christoph Lameter9eeff232006-01-18 17:42:31 -08003511 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003512 /*
3513 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3514 * and we also need to be able to write out pages for RECLAIM_WRITE
3515 * and RECLAIM_SWAP.
3516 */
3517 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003518 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003519 reclaim_state.reclaimed_slab = 0;
3520 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003521
Mel Gorman90afa5d2009-06-16 15:33:20 -07003522 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003523 /*
3524 * Free memory by calling shrink zone with increasing
3525 * priorities until we have enough memory freed.
3526 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003527 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003528 shrink_zone(zone, &sc);
3529 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003530 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003531
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003532 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3533 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003534 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003535 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003536 * many pages were freed in this zone. So we take the current
3537 * number of slab pages and shake the slab until it is reduced
3538 * by the same nr_pages that we used for reclaiming unmapped
3539 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003540 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003541 * Note that shrink_slab will free memory on all zones and may
3542 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003543 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003544 for (;;) {
3545 unsigned long lru_pages = zone_reclaimable_pages(zone);
3546
3547 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003548 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003549 break;
3550
3551 /* Freed enough memory */
3552 nr_slab_pages1 = zone_page_state(zone,
3553 NR_SLAB_RECLAIMABLE);
3554 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3555 break;
3556 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003557
3558 /*
3559 * Update nr_reclaimed by the number of slab pages we
3560 * reclaimed from this zone.
3561 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003562 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3563 if (nr_slab_pages1 < nr_slab_pages0)
3564 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003565 }
3566
Christoph Lameter9eeff232006-01-18 17:42:31 -08003567 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003568 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003569 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003570 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003571}
Andrew Morton179e9632006-03-22 00:08:18 -08003572
3573int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3574{
Andrew Morton179e9632006-03-22 00:08:18 -08003575 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003576 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003577
3578 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003579 * Zone reclaim reclaims unmapped file backed pages and
3580 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003581 *
Christoph Lameter96146342006-07-03 00:24:13 -07003582 * A small portion of unmapped file backed pages is needed for
3583 * file I/O otherwise pages read by file I/O will be immediately
3584 * thrown out if the zone is overallocated. So we do not reclaim
3585 * if less than a specified percentage of the zone is used by
3586 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003587 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003588 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3589 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003590 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003591
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003592 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003593 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003594
Andrew Morton179e9632006-03-22 00:08:18 -08003595 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003596 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003597 */
David Rientjesd773ed62007-10-16 23:26:01 -07003598 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003599 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003600
3601 /*
3602 * Only run zone reclaim on the local zone or on zones that do not
3603 * have associated processors. This will favor the local processor
3604 * over remote processors and spread off node memory allocations
3605 * as wide as possible.
3606 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003607 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003608 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003609 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003610
3611 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003612 return ZONE_RECLAIM_NOSCAN;
3613
David Rientjesd773ed62007-10-16 23:26:01 -07003614 ret = __zone_reclaim(zone, gfp_mask, order);
3615 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3616
Mel Gorman24cf725182009-06-16 15:33:23 -07003617 if (!ret)
3618 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3619
David Rientjesd773ed62007-10-16 23:26:01 -07003620 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003621}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003622#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003623
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003624/*
3625 * page_evictable - test whether a page is evictable
3626 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003627 *
3628 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003629 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003630 *
3631 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003632 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003633 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003634 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003635 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003636int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003637{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003638 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003639}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003640
Hugh Dickins850465792012-01-20 14:34:19 -08003641#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003642/**
Hugh Dickins24513262012-01-20 14:34:21 -08003643 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3644 * @pages: array of pages to check
3645 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003646 *
Hugh Dickins24513262012-01-20 14:34:21 -08003647 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003648 *
3649 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003650 */
Hugh Dickins24513262012-01-20 14:34:21 -08003651void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003652{
Johannes Weiner925b7672012-01-12 17:18:15 -08003653 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003654 struct zone *zone = NULL;
3655 int pgscanned = 0;
3656 int pgrescued = 0;
3657 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003658
Hugh Dickins24513262012-01-20 14:34:21 -08003659 for (i = 0; i < nr_pages; i++) {
3660 struct page *page = pages[i];
3661 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003662
Hugh Dickins24513262012-01-20 14:34:21 -08003663 pgscanned++;
3664 pagezone = page_zone(page);
3665 if (pagezone != zone) {
3666 if (zone)
3667 spin_unlock_irq(&zone->lru_lock);
3668 zone = pagezone;
3669 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003670 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003671 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003672
Hugh Dickins24513262012-01-20 14:34:21 -08003673 if (!PageLRU(page) || !PageUnevictable(page))
3674 continue;
3675
Hugh Dickins39b5f292012-10-08 16:33:18 -07003676 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003677 enum lru_list lru = page_lru_base_type(page);
3678
3679 VM_BUG_ON(PageActive(page));
3680 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003681 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3682 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003683 pgrescued++;
3684 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003685 }
Hugh Dickins24513262012-01-20 14:34:21 -08003686
3687 if (zone) {
3688 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3689 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3690 spin_unlock_irq(&zone->lru_lock);
3691 }
Hugh Dickins850465792012-01-20 14:34:19 -08003692}
3693#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003694
Johannes Weiner264e56d2011-10-31 17:09:13 -07003695static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003696{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003697 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003698 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003699 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003700 "one, please send an email to linux-mm@kvack.org.\n",
3701 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003702}
3703
3704/*
3705 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3706 * all nodes' unevictable lists for evictable pages
3707 */
3708unsigned long scan_unevictable_pages;
3709
3710int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003711 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003712 size_t *length, loff_t *ppos)
3713{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003714 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003715 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003716 scan_unevictable_pages = 0;
3717 return 0;
3718}
3719
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003720#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003721/*
3722 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3723 * a specified node's per zone unevictable lists for evictable pages.
3724 */
3725
Kay Sievers10fbcf42011-12-21 14:48:43 -08003726static ssize_t read_scan_unevictable_node(struct device *dev,
3727 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003728 char *buf)
3729{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003730 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003731 return sprintf(buf, "0\n"); /* always zero; should fit... */
3732}
3733
Kay Sievers10fbcf42011-12-21 14:48:43 -08003734static ssize_t write_scan_unevictable_node(struct device *dev,
3735 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003736 const char *buf, size_t count)
3737{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003738 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003739 return 1;
3740}
3741
3742
Kay Sievers10fbcf42011-12-21 14:48:43 -08003743static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003744 read_scan_unevictable_node,
3745 write_scan_unevictable_node);
3746
3747int scan_unevictable_register_node(struct node *node)
3748{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003749 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003750}
3751
3752void scan_unevictable_unregister_node(struct node *node)
3753{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003754 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003755}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003756#endif