blob: 2cff0d491c6dca84391edd100e1726696c1475d5 [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;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700549 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700550
551 VM_BUG_ON(PageLRU(page));
552
553redo:
554 ClearPageUnevictable(page);
555
Hugh Dickins39b5f292012-10-08 16:33:18 -0700556 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700557 /*
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
562 */
Mel Gormanc53954a2013-07-03 15:02:34 -0700563 lru = page_lru_base_type(page);
564 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700565 } else {
566 /*
567 * Put unevictable pages directly on zone's unevictable
568 * list.
569 */
570 lru = LRU_UNEVICTABLE;
571 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700572 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800576 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700578 * the page back to the evictable list.
579 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700580 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700581 */
582 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700583 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700584
585 /*
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
589 */
Hugh Dickins39b5f292012-10-08 16:33:18 -0700590 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700591 if (!isolate_lru_page(page)) {
592 put_page(page);
593 goto redo;
594 }
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
598 */
599 }
600
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700601 if (was_unevictable && lru != LRU_UNEVICTABLE)
602 count_vm_event(UNEVICTABLE_PGRESCUED);
603 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604 count_vm_event(UNEVICTABLE_PGCULLED);
605
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700606 put_page(page); /* drop ref from isolate */
607}
608
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800609enum page_references {
610 PAGEREF_RECLAIM,
611 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800612 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800613 PAGEREF_ACTIVATE,
614};
615
616static enum page_references page_check_references(struct page *page,
617 struct scan_control *sc)
618{
Johannes Weiner645747462010-03-05 13:42:22 -0800619 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800620 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800621
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700622 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800624 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800625
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800626 /*
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
629 */
630 if (vm_flags & VM_LOCKED)
631 return PAGEREF_RECLAIM;
632
Johannes Weiner645747462010-03-05 13:42:22 -0800633 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700634 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800635 return PAGEREF_ACTIVATE;
636 /*
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
640 * than once.
641 *
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
645 *
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
648 * quickly recovered.
649 */
650 SetPageReferenced(page);
651
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800652 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800653 return PAGEREF_ACTIVATE;
654
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800655 /*
656 * Activate file-backed executable pages after first usage.
657 */
658 if (vm_flags & VM_EXEC)
659 return PAGEREF_ACTIVATE;
660
Johannes Weiner645747462010-03-05 13:42:22 -0800661 return PAGEREF_KEEP;
662 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800663
664 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700665 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800666 return PAGEREF_RECLAIM_CLEAN;
667
668 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800669}
670
Mel Gormane2be15f2013-07-03 15:01:57 -0700671/* Check if a page is dirty or under writeback */
672static void page_check_dirty_writeback(struct page *page,
673 bool *dirty, bool *writeback)
674{
Mel Gormanb4597222013-07-03 15:02:05 -0700675 struct address_space *mapping;
676
Mel Gormane2be15f2013-07-03 15:01:57 -0700677 /*
678 * Anonymous pages are not handled by flushers and must be written
679 * from reclaim context. Do not stall reclaim based on them
680 */
681 if (!page_is_file_cache(page)) {
682 *dirty = false;
683 *writeback = false;
684 return;
685 }
686
687 /* By default assume that the page flags are accurate */
688 *dirty = PageDirty(page);
689 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700690
691 /* Verify dirty/writeback state if the filesystem supports it */
692 if (!page_has_private(page))
693 return;
694
695 mapping = page_mapping(page);
696 if (mapping && mapping->a_ops->is_dirty_writeback)
697 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700698}
699
Nick Piggine2867812008-07-25 19:45:30 -0700700/*
Andrew Morton1742f192006-03-22 00:08:21 -0800701 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 */
Andrew Morton1742f192006-03-22 00:08:21 -0800703static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700704 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700705 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700706 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700707 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700708 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700709 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700710 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700711 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700712 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
714 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700715 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700717 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700718 unsigned long nr_dirty = 0;
719 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800720 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700721 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700722 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
724 cond_resched();
725
Tim Chen69980e32012-07-31 16:46:08 -0700726 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 while (!list_empty(page_list)) {
728 struct address_space *mapping;
729 struct page *page;
730 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700731 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700732 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
734 cond_resched();
735
736 page = lru_to_page(page_list);
737 list_del(&page->lru);
738
Nick Piggin529ae9a2008-08-02 12:01:03 +0200739 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 goto keep;
741
Nick Piggin725d7042006-09-25 23:30:55 -0700742 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700743 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744
745 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800746
Hugh Dickins39b5f292012-10-08 16:33:18 -0700747 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700748 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700749
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700750 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800751 goto keep_locked;
752
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 /* Double the slab pressure for mapped and swapcache pages */
754 if (page_mapped(page) || PageSwapCache(page))
755 sc->nr_scanned++;
756
Andy Whitcroftc661b072007-08-22 14:01:26 -0700757 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
758 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
759
Mel Gorman283aba92013-07-03 15:01:51 -0700760 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700761 * The number of dirty pages determines if a zone is marked
762 * reclaim_congested which affects wait_iff_congested. kswapd
763 * will stall and start writing pages if the tail of the LRU
764 * is all dirty unqueued pages.
765 */
766 page_check_dirty_writeback(page, &dirty, &writeback);
767 if (dirty || writeback)
768 nr_dirty++;
769
770 if (dirty && !writeback)
771 nr_unqueued_dirty++;
772
Mel Gormand04e8ac2013-07-03 15:02:03 -0700773 /*
774 * Treat this page as congested if the underlying BDI is or if
775 * pages are cycling through the LRU so quickly that the
776 * pages marked for immediate reclaim are making it to the
777 * end of the LRU a second time.
778 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700779 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700780 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
781 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700782 nr_congested++;
783
784 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700785 * If a page at the tail of the LRU is under writeback, there
786 * are three cases to consider.
787 *
788 * 1) If reclaim is encountering an excessive number of pages
789 * under writeback and this page is both under writeback and
790 * PageReclaim then it indicates that pages are being queued
791 * for IO but are being recycled through the LRU before the
792 * IO can complete. Waiting on the page itself risks an
793 * indefinite stall if it is impossible to writeback the
794 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700795 * note that the LRU is being scanned too quickly and the
796 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700797 *
798 * 2) Global reclaim encounters a page, memcg encounters a
799 * page that is not marked for immediate reclaim or
800 * the caller does not have __GFP_IO. In this case mark
801 * the page for immediate reclaim and continue scanning.
802 *
803 * __GFP_IO is checked because a loop driver thread might
804 * enter reclaim, and deadlock if it waits on a page for
805 * which it is needed to do the write (loop masks off
806 * __GFP_IO|__GFP_FS for this reason); but more thought
807 * would probably show more reasons.
808 *
809 * Don't require __GFP_FS, since we're not going into the
810 * FS, just waiting on its writeback completion. Worryingly,
811 * ext4 gfs2 and xfs allocate pages with
812 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
813 * may_enter_fs here is liable to OOM on them.
814 *
815 * 3) memcg encounters a page that is not already marked
816 * PageReclaim. memcg does not have any dirty pages
817 * throttling so we could easily OOM just because too many
818 * pages are in writeback and there is nothing else to
819 * reclaim. Wait for the writeback to complete.
820 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700821 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700822 /* Case 1 above */
823 if (current_is_kswapd() &&
824 PageReclaim(page) &&
825 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700826 nr_immediate++;
827 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700828
829 /* Case 2 above */
830 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700831 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
832 /*
833 * This is slightly racy - end_page_writeback()
834 * might have just cleared PageReclaim, then
835 * setting PageReclaim here end up interpreted
836 * as PageReadahead - but that does not matter
837 * enough to care. What we do want is for this
838 * page to have PageReclaim set next time memcg
839 * reclaim reaches the tests above, so it will
840 * then wait_on_page_writeback() to avoid OOM;
841 * and it's also appropriate in global reclaim.
842 */
843 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700844 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700845
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700846 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700847
848 /* Case 3 above */
849 } else {
850 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700851 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700852 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853
Minchan Kim02c6de82012-10-08 16:31:55 -0700854 if (!force_reclaim)
855 references = page_check_references(page, sc);
856
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800857 switch (references) {
858 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800860 case PAGEREF_KEEP:
861 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800862 case PAGEREF_RECLAIM:
863 case PAGEREF_RECLAIM_CLEAN:
864 ; /* try to reclaim the page below */
865 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 /*
868 * Anonymous process memory has backing store?
869 * Try to allocate it some swap space here.
870 */
Nick Pigginb291f002008-10-18 20:26:44 -0700871 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800872 if (!(sc->gfp_mask & __GFP_IO))
873 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700874 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800876 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877
Mel Gormane2be15f2013-07-03 15:01:57 -0700878 /* Adding to swap updated mapping */
879 mapping = page_mapping(page);
880 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
882 /*
883 * The page is mapped into the page tables of one or more
884 * processes. Try to unmap it here.
885 */
886 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700887 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 case SWAP_FAIL:
889 goto activate_locked;
890 case SWAP_AGAIN:
891 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700892 case SWAP_MLOCK:
893 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 case SWAP_SUCCESS:
895 ; /* try to free the page below */
896 }
897 }
898
899 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700900 /*
901 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -0700902 * avoid risk of stack overflow but only writeback
903 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -0700904 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700905 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700906 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -0700907 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700908 /*
909 * Immediately reclaim when written back.
910 * Similar in principal to deactivate_page()
911 * except we already have the page isolated
912 * and know it's dirty
913 */
914 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
915 SetPageReclaim(page);
916
Mel Gormanee728862011-10-31 17:07:38 -0700917 goto keep_locked;
918 }
919
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800920 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700922 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800924 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 goto keep_locked;
926
927 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700928 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 case PAGE_KEEP:
930 goto keep_locked;
931 case PAGE_ACTIVATE:
932 goto activate_locked;
933 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700934 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700935 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700936 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700938
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 /*
940 * A synchronous write - probably a ramdisk. Go
941 * ahead and try to reclaim the page.
942 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200943 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 goto keep;
945 if (PageDirty(page) || PageWriteback(page))
946 goto keep_locked;
947 mapping = page_mapping(page);
948 case PAGE_CLEAN:
949 ; /* try to free the page below */
950 }
951 }
952
953 /*
954 * If the page has buffers, try to free the buffer mappings
955 * associated with this page. If we succeed we try to free
956 * the page as well.
957 *
958 * We do this even if the page is PageDirty().
959 * try_to_release_page() does not perform I/O, but it is
960 * possible for a page to have PageDirty set, but it is actually
961 * clean (all its buffers are clean). This happens if the
962 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700963 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 * try_to_release_page() will discover that cleanness and will
965 * drop the buffers and mark the page clean - it can be freed.
966 *
967 * Rarely, pages can have buffers and no ->mapping. These are
968 * the pages which were not successfully invalidated in
969 * truncate_complete_page(). We try to drop those buffers here
970 * and if that worked, and the page is no longer mapped into
971 * process address space (page_count == 1) it can be freed.
972 * Otherwise, leave the page on the LRU so it is swappable.
973 */
David Howells266cf652009-04-03 16:42:36 +0100974 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 if (!try_to_release_page(page, sc->gfp_mask))
976 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700977 if (!mapping && page_count(page) == 1) {
978 unlock_page(page);
979 if (put_page_testzero(page))
980 goto free_it;
981 else {
982 /*
983 * rare race with speculative reference.
984 * the speculative reference will free
985 * this page shortly, so we may
986 * increment nr_reclaimed here (and
987 * leave it off the LRU).
988 */
989 nr_reclaimed++;
990 continue;
991 }
992 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 }
994
Nick Piggine2867812008-07-25 19:45:30 -0700995 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800996 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
Nick Piggina978d6f2008-10-18 20:26:58 -0700998 /*
999 * At this point, we have no other references and there is
1000 * no way to pick any more up (removed from LRU, removed
1001 * from pagecache). Can use non-atomic bitops now (and
1002 * we obviously don't have to worry about waking up a process
1003 * waiting on the page lock, because there are no references.
1004 */
1005 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001006free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001007 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001008
1009 /*
1010 * Is there need to periodically free_page_list? It would
1011 * appear not as the counts should be low
1012 */
1013 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 continue;
1015
Nick Pigginb291f002008-10-18 20:26:44 -07001016cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001017 if (PageSwapCache(page))
1018 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001019 unlock_page(page);
1020 putback_lru_page(page);
1021 continue;
1022
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001024 /* Not a candidate for swapping, so reclaim swap space. */
1025 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001026 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001027 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 SetPageActive(page);
1029 pgactivate++;
1030keep_locked:
1031 unlock_page(page);
1032keep:
1033 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001034 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001036
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001037 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001038
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001040 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001041 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001042 *ret_nr_dirty += nr_dirty;
1043 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001044 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001045 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001046 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001047 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048}
1049
Minchan Kim02c6de82012-10-08 16:31:55 -07001050unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1051 struct list_head *page_list)
1052{
1053 struct scan_control sc = {
1054 .gfp_mask = GFP_KERNEL,
1055 .priority = DEF_PRIORITY,
1056 .may_unmap = 1,
1057 };
Mel Gorman8e950282013-07-03 15:02:02 -07001058 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001059 struct page *page, *next;
1060 LIST_HEAD(clean_pages);
1061
1062 list_for_each_entry_safe(page, next, page_list, lru) {
1063 if (page_is_file_cache(page) && !PageDirty(page)) {
1064 ClearPageActive(page);
1065 list_move(&page->lru, &clean_pages);
1066 }
1067 }
1068
1069 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001070 TTU_UNMAP|TTU_IGNORE_ACCESS,
1071 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001072 list_splice(&clean_pages, page_list);
1073 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1074 return ret;
1075}
1076
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001077/*
1078 * Attempt to remove the specified page from its LRU. Only take this page
1079 * if it is of the appropriate PageActive status. Pages which are being
1080 * freed elsewhere are also ignored.
1081 *
1082 * page: page to consider
1083 * mode: one of the LRU isolation modes defined above
1084 *
1085 * returns 0 on success, -ve errno on failure.
1086 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001087int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001088{
1089 int ret = -EINVAL;
1090
1091 /* Only take pages on the LRU. */
1092 if (!PageLRU(page))
1093 return ret;
1094
Minchan Kime46a2872012-10-08 16:33:48 -07001095 /* Compaction should not handle unevictable pages but CMA can do so */
1096 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001097 return ret;
1098
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001099 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001100
Mel Gormanc8244932012-01-12 17:19:38 -08001101 /*
1102 * To minimise LRU disruption, the caller can indicate that it only
1103 * wants to isolate pages it will be able to operate on without
1104 * blocking - clean pages for the most part.
1105 *
1106 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1107 * is used by reclaim when it is cannot write to backing storage
1108 *
1109 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1110 * that it is possible to migrate without blocking
1111 */
1112 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1113 /* All the caller can do on PageWriteback is block */
1114 if (PageWriteback(page))
1115 return ret;
1116
1117 if (PageDirty(page)) {
1118 struct address_space *mapping;
1119
1120 /* ISOLATE_CLEAN means only clean pages */
1121 if (mode & ISOLATE_CLEAN)
1122 return ret;
1123
1124 /*
1125 * Only pages without mappings or that have a
1126 * ->migratepage callback are possible to migrate
1127 * without blocking
1128 */
1129 mapping = page_mapping(page);
1130 if (mapping && !mapping->a_ops->migratepage)
1131 return ret;
1132 }
1133 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001134
Minchan Kimf80c0672011-10-31 17:06:55 -07001135 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1136 return ret;
1137
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001138 if (likely(get_page_unless_zero(page))) {
1139 /*
1140 * Be careful not to clear PageLRU until after we're
1141 * sure the page is not being freed elsewhere -- the
1142 * page release code relies on it.
1143 */
1144 ClearPageLRU(page);
1145 ret = 0;
1146 }
1147
1148 return ret;
1149}
1150
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001151/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 * zone->lru_lock is heavily contended. Some of the functions that
1153 * shrink the lists perform better by taking out a batch of pages
1154 * and working on them outside the LRU lock.
1155 *
1156 * For pagecache intensive workloads, this function is the hottest
1157 * spot in the kernel (apart from copy_*_user functions).
1158 *
1159 * Appropriate locks must be held before calling this function.
1160 *
1161 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001162 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001164 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001165 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001166 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001167 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 *
1169 * returns how many pages were moved onto *@dst.
1170 */
Andrew Morton69e05942006-03-22 00:08:19 -08001171static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001172 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001173 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001174 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001176 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001177 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001178 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001180 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001181 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001182 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001183
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 page = lru_to_page(src);
1185 prefetchw_prev_lru_page(page, src, flags);
1186
Nick Piggin725d7042006-09-25 23:30:55 -07001187 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001188
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001189 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001190 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001191 nr_pages = hpage_nr_pages(page);
1192 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001193 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001194 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001195 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001196
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001197 case -EBUSY:
1198 /* else it is being freed elsewhere */
1199 list_move(&page->lru, src);
1200 continue;
1201
1202 default:
1203 BUG();
1204 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 }
1206
Hugh Dickinsf6260122012-01-12 17:20:06 -08001207 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001208 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1209 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 return nr_taken;
1211}
1212
Nick Piggin62695a82008-10-18 20:26:09 -07001213/**
1214 * isolate_lru_page - tries to isolate a page from its LRU list
1215 * @page: page to isolate from its LRU list
1216 *
1217 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1218 * vmstat statistic corresponding to whatever LRU list the page was on.
1219 *
1220 * Returns 0 if the page was removed from an LRU list.
1221 * Returns -EBUSY if the page was not on an LRU list.
1222 *
1223 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001224 * the active list, it will have PageActive set. If it was found on
1225 * the unevictable list, it will have the PageUnevictable bit set. That flag
1226 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001227 *
1228 * The vmstat statistic corresponding to the list on which the page was
1229 * found will be decremented.
1230 *
1231 * Restrictions:
1232 * (1) Must be called with an elevated refcount on the page. This is a
1233 * fundamentnal difference from isolate_lru_pages (which is called
1234 * without a stable reference).
1235 * (2) the lru_lock must not be held.
1236 * (3) interrupts must be enabled.
1237 */
1238int isolate_lru_page(struct page *page)
1239{
1240 int ret = -EBUSY;
1241
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001242 VM_BUG_ON(!page_count(page));
1243
Nick Piggin62695a82008-10-18 20:26:09 -07001244 if (PageLRU(page)) {
1245 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001246 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001247
1248 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001249 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001250 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001251 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001252 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001253 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001254 del_page_from_lru_list(page, lruvec, lru);
1255 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001256 }
1257 spin_unlock_irq(&zone->lru_lock);
1258 }
1259 return ret;
1260}
1261
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001262/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001263 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1264 * then get resheduled. When there are massive number of tasks doing page
1265 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1266 * the LRU list will go small and be scanned faster than necessary, leading to
1267 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001268 */
1269static int too_many_isolated(struct zone *zone, int file,
1270 struct scan_control *sc)
1271{
1272 unsigned long inactive, isolated;
1273
1274 if (current_is_kswapd())
1275 return 0;
1276
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001277 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001278 return 0;
1279
1280 if (file) {
1281 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1282 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1283 } else {
1284 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1285 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1286 }
1287
Fengguang Wu3cf23842012-12-18 14:23:31 -08001288 /*
1289 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1290 * won't get blocked by normal direct-reclaimers, forming a circular
1291 * deadlock.
1292 */
1293 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1294 inactive >>= 3;
1295
Rik van Riel35cd7812009-09-21 17:01:38 -07001296 return isolated > inactive;
1297}
1298
Mel Gorman66635622010-08-09 17:19:30 -07001299static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001300putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001301{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001302 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1303 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001304 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001305
Mel Gorman66635622010-08-09 17:19:30 -07001306 /*
1307 * Put back any unfreeable pages.
1308 */
Mel Gorman66635622010-08-09 17:19:30 -07001309 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001310 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001311 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001312
Mel Gorman66635622010-08-09 17:19:30 -07001313 VM_BUG_ON(PageLRU(page));
1314 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001315 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001316 spin_unlock_irq(&zone->lru_lock);
1317 putback_lru_page(page);
1318 spin_lock_irq(&zone->lru_lock);
1319 continue;
1320 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001321
1322 lruvec = mem_cgroup_page_lruvec(page, zone);
1323
Linus Torvalds7a608572011-01-17 14:42:19 -08001324 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001325 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001326 add_page_to_lru_list(page, lruvec, lru);
1327
Mel Gorman66635622010-08-09 17:19:30 -07001328 if (is_active_lru(lru)) {
1329 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001330 int numpages = hpage_nr_pages(page);
1331 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001332 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001333 if (put_page_testzero(page)) {
1334 __ClearPageLRU(page);
1335 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001336 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001337
1338 if (unlikely(PageCompound(page))) {
1339 spin_unlock_irq(&zone->lru_lock);
1340 (*get_compound_page_dtor(page))(page);
1341 spin_lock_irq(&zone->lru_lock);
1342 } else
1343 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001344 }
1345 }
Mel Gorman66635622010-08-09 17:19:30 -07001346
Hugh Dickins3f797682012-01-12 17:20:07 -08001347 /*
1348 * To save our caller's stack, now use input list for pages to free.
1349 */
1350 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001351}
1352
1353/*
Andrew Morton1742f192006-03-22 00:08:21 -08001354 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1355 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 */
Mel Gorman66635622010-08-09 17:19:30 -07001357static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001358shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001359 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360{
1361 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001362 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001363 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001364 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001365 unsigned long nr_dirty = 0;
1366 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001367 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001368 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001369 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001370 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001371 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001372 struct zone *zone = lruvec_zone(lruvec);
1373 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001374
Rik van Riel35cd7812009-09-21 17:01:38 -07001375 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001376 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001377
1378 /* We are about to die and free our memory. Return now. */
1379 if (fatal_signal_pending(current))
1380 return SWAP_CLUSTER_MAX;
1381 }
1382
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001384
1385 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001386 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001387 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001388 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001389
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001392 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1393 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001394
1395 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1396 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1397
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001398 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001399 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001400 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001401 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001402 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001403 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001404 }
Hillf Dantond563c052012-03-21 16:34:02 -07001405 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001406
Hillf Dantond563c052012-03-21 16:34:02 -07001407 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001408 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001409
Minchan Kim02c6de82012-10-08 16:31:55 -07001410 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001411 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1412 &nr_writeback, &nr_immediate,
1413 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001414
Hugh Dickins3f797682012-01-12 17:20:07 -08001415 spin_lock_irq(&zone->lru_lock);
1416
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001417 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001418
Ying Han904249a2012-04-25 16:01:48 -07001419 if (global_reclaim(sc)) {
1420 if (current_is_kswapd())
1421 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1422 nr_reclaimed);
1423 else
1424 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1425 nr_reclaimed);
1426 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001427
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001428 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001429
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001430 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001431
1432 spin_unlock_irq(&zone->lru_lock);
1433
1434 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001435
Mel Gorman92df3a72011-10-31 17:07:56 -07001436 /*
1437 * If reclaim is isolating dirty pages under writeback, it implies
1438 * that the long-lived page allocation rate is exceeding the page
1439 * laundering rate. Either the global limits are not being effective
1440 * at throttling processes due to the page distribution throughout
1441 * zones or there is heavy usage of a slow backing device. The
1442 * only option is to throttle from reclaim context which is not ideal
1443 * as there is no guarantee the dirtying process is throttled in the
1444 * same way balance_dirty_pages() manages.
1445 *
Mel Gorman8e950282013-07-03 15:02:02 -07001446 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1447 * of pages under pages flagged for immediate reclaim and stall if any
1448 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001449 */
Mel Gorman918fc712013-07-08 16:00:25 -07001450 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001451 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001452
Mel Gormand43006d2013-07-03 15:01:50 -07001453 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001454 * memcg will stall in page writeback so only consider forcibly
1455 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001456 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001457 if (global_reclaim(sc)) {
1458 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001459 * Tag a zone as congested if all the dirty pages scanned were
1460 * backed by a congested BDI and wait_iff_congested will stall.
1461 */
1462 if (nr_dirty && nr_dirty == nr_congested)
1463 zone_set_flag(zone, ZONE_CONGESTED);
1464
1465 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001466 * If dirty pages are scanned that are not queued for IO, it
1467 * implies that flushers are not keeping up. In this case, flag
1468 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1469 * pages from reclaim context. It will forcibly stall in the
1470 * next check.
1471 */
1472 if (nr_unqueued_dirty == nr_taken)
1473 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1474
1475 /*
1476 * In addition, if kswapd scans pages marked marked for
1477 * immediate reclaim and under writeback (nr_immediate), it
1478 * implies that pages are cycling through the LRU faster than
1479 * they are written so also forcibly stall.
1480 */
1481 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1482 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001483 }
Mel Gormand43006d2013-07-03 15:01:50 -07001484
Mel Gorman8e950282013-07-03 15:02:02 -07001485 /*
1486 * Stall direct reclaim for IO completions if underlying BDIs or zone
1487 * is congested. Allow kswapd to continue until it starts encountering
1488 * unqueued dirty pages or cycling through the LRU too quickly.
1489 */
1490 if (!sc->hibernation_mode && !current_is_kswapd())
1491 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1492
Mel Gormane11da5b2010-10-26 14:21:40 -07001493 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1494 zone_idx(zone),
1495 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001496 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001497 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001498 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499}
1500
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001501/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 * This moves pages from the active list to the inactive list.
1503 *
1504 * We move them the other way if the page is referenced by one or more
1505 * processes, from rmap.
1506 *
1507 * If the pages are mostly unmapped, the processing is fast and it is
1508 * appropriate to hold zone->lru_lock across the whole operation. But if
1509 * the pages are mapped, the processing is slow (page_referenced()) so we
1510 * should drop zone->lru_lock around each page. It's impossible to balance
1511 * this, so instead we remove the pages from the LRU while processing them.
1512 * It is safe to rely on PG_active against the non-LRU pages in here because
1513 * nobody will play with that bit on a non-LRU page.
1514 *
1515 * The downside is that we have to touch page->_count against each page.
1516 * But we had to alter page->flags anyway.
1517 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001518
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001519static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001520 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001521 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001522 enum lru_list lru)
1523{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001524 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001525 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001526 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001527 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001528
Wu Fengguang3eb41402009-06-16 15:33:13 -07001529 while (!list_empty(list)) {
1530 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001531 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001532
1533 VM_BUG_ON(PageLRU(page));
1534 SetPageLRU(page);
1535
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001536 nr_pages = hpage_nr_pages(page);
1537 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001538 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001539 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001540
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001541 if (put_page_testzero(page)) {
1542 __ClearPageLRU(page);
1543 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001544 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001545
1546 if (unlikely(PageCompound(page))) {
1547 spin_unlock_irq(&zone->lru_lock);
1548 (*get_compound_page_dtor(page))(page);
1549 spin_lock_irq(&zone->lru_lock);
1550 } else
1551 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001552 }
1553 }
1554 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1555 if (!is_active_lru(lru))
1556 __count_vm_events(PGDEACTIVATE, pgmoved);
1557}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001558
Hugh Dickinsf6260122012-01-12 17:20:06 -08001559static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001560 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001561 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001562 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001564 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001565 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001566 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001568 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001569 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001571 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001572 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001573 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001574 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001575 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576
1577 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001578
1579 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001580 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001581 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001582 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001583
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001585
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001586 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1587 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001588 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001589 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001590
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001591 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001592
Hugh Dickinsf6260122012-01-12 17:20:06 -08001593 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001594 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001595 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 spin_unlock_irq(&zone->lru_lock);
1597
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 while (!list_empty(&l_hold)) {
1599 cond_resched();
1600 page = lru_to_page(&l_hold);
1601 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001602
Hugh Dickins39b5f292012-10-08 16:33:18 -07001603 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001604 putback_lru_page(page);
1605 continue;
1606 }
1607
Mel Gormancc715d92012-03-21 16:34:00 -07001608 if (unlikely(buffer_heads_over_limit)) {
1609 if (page_has_private(page) && trylock_page(page)) {
1610 if (page_has_private(page))
1611 try_to_release_page(page, 0);
1612 unlock_page(page);
1613 }
1614 }
1615
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001616 if (page_referenced(page, 0, sc->target_mem_cgroup,
1617 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001618 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001619 /*
1620 * Identify referenced, file-backed active pages and
1621 * give them one more trip around the active list. So
1622 * that executable code get better chances to stay in
1623 * memory under moderate memory pressure. Anon pages
1624 * are not likely to be evicted by use-once streaming
1625 * IO, plus JVM can create lots of anon VM_EXEC pages,
1626 * so we ignore them here.
1627 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001628 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001629 list_add(&page->lru, &l_active);
1630 continue;
1631 }
1632 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001633
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001634 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 list_add(&page->lru, &l_inactive);
1636 }
1637
Andrew Mortonb5557492009-01-06 14:40:13 -08001638 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001639 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001640 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001641 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001642 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001643 * Count referenced pages from currently used mappings as rotated,
1644 * even though only some of them are actually re-activated. This
1645 * helps balance scan pressure between file and anonymous pages in
1646 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001647 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001648 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001649
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001650 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1651 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001652 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001653 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001654
1655 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656}
1657
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001658#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001659static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001660{
1661 unsigned long active, inactive;
1662
1663 active = zone_page_state(zone, NR_ACTIVE_ANON);
1664 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1665
1666 if (inactive * zone->inactive_ratio < active)
1667 return 1;
1668
1669 return 0;
1670}
1671
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001672/**
1673 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001674 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001675 *
1676 * Returns true if the zone does not have enough inactive anon pages,
1677 * meaning some active anon pages need to be deactivated.
1678 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001679static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001680{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001681 /*
1682 * If we don't have swap space, anonymous page deactivation
1683 * is pointless.
1684 */
1685 if (!total_swap_pages)
1686 return 0;
1687
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001688 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001689 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001690
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001691 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001692}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001693#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001694static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001695{
1696 return 0;
1697}
1698#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001699
Rik van Riel56e49d22009-06-16 15:32:28 -07001700/**
1701 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001702 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001703 *
1704 * When the system is doing streaming IO, memory pressure here
1705 * ensures that active file pages get deactivated, until more
1706 * than half of the file pages are on the inactive list.
1707 *
1708 * Once we get to that situation, protect the system's working
1709 * set from being evicted by disabling active file page aging.
1710 *
1711 * This uses a different ratio than the anonymous pages, because
1712 * the page cache uses a use-once replacement algorithm.
1713 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001714static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001715{
Johannes Weinere3790142013-02-22 16:35:19 -08001716 unsigned long inactive;
1717 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001718
Johannes Weinere3790142013-02-22 16:35:19 -08001719 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1720 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1721
1722 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001723}
1724
Hugh Dickins75b00af2012-05-29 15:07:09 -07001725static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001726{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001727 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001728 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001729 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001730 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001731}
1732
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001733static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001734 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001735{
Rik van Rielb39415b2009-12-14 17:59:48 -08001736 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001737 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001738 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001739 return 0;
1740 }
1741
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001742 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001743}
1744
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001745static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001746{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001747 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001748 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001749 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001750}
1751
Johannes Weiner9a265112013-02-22 16:32:17 -08001752enum scan_balance {
1753 SCAN_EQUAL,
1754 SCAN_FRACT,
1755 SCAN_ANON,
1756 SCAN_FILE,
1757};
1758
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001760 * Determine how aggressively the anon and file LRU lists should be
1761 * scanned. The relative value of each set of LRU lists is determined
1762 * by looking at the fraction of the pages scanned we did rotate back
1763 * onto the active list instead of evict.
1764 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001765 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1766 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001767 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001768static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001769 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001770{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001771 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001772 u64 fraction[2];
1773 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001774 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001775 unsigned long anon_prio, file_prio;
1776 enum scan_balance scan_balance;
1777 unsigned long anon, file, free;
1778 bool force_scan = false;
1779 unsigned long ap, fp;
1780 enum lru_list lru;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001781
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001782 /*
1783 * If the zone or memcg is small, nr[l] can be 0. This
1784 * results in no scanning on this priority and a potential
1785 * priority drop. Global direct reclaim can go to the next
1786 * zone and tends to have no problems. Global kswapd is for
1787 * zone balancing and it needs to scan a minimum amount. When
1788 * reclaiming for a memcg, a priority drop can cause high
1789 * latencies, so it's better to scan a minimum amount there as
1790 * well.
1791 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001792 if (current_is_kswapd() && zone->all_unreclaimable)
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001793 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001794 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001795 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001796
1797 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001798 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001799 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001800 goto out;
1801 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001802
Johannes Weiner10316b32013-02-22 16:32:14 -08001803 /*
1804 * Global reclaim will swap to prevent OOM even with no
1805 * swappiness, but memcg users want to use this knob to
1806 * disable swapping for individual groups completely when
1807 * using the memory controller's swap limit feature would be
1808 * too expensive.
1809 */
1810 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001811 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001812 goto out;
1813 }
1814
1815 /*
1816 * Do not apply any pressure balancing cleverness when the
1817 * system is close to OOM, scan both anon and file equally
1818 * (unless the swappiness setting disagrees with swapping).
1819 */
1820 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001821 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001822 goto out;
1823 }
1824
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001825 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1826 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1827 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1828 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001829
Johannes Weiner11d16c22013-02-22 16:32:15 -08001830 /*
1831 * If it's foreseeable that reclaiming the file cache won't be
1832 * enough to get the zone back into a desirable shape, we have
1833 * to swap. Better start now and leave the - probably heavily
1834 * thrashing - remaining file pages alone.
1835 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001836 if (global_reclaim(sc)) {
Johannes Weiner11d16c22013-02-22 16:32:15 -08001837 free = zone_page_state(zone, NR_FREE_PAGES);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001838 if (unlikely(file + free <= high_wmark_pages(zone))) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001839 scan_balance = SCAN_ANON;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001840 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001841 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001842 }
1843
1844 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001845 * There is enough inactive page cache, do not reclaim
1846 * anything from the anonymous working set right now.
1847 */
1848 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001849 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001850 goto out;
1851 }
1852
Johannes Weiner9a265112013-02-22 16:32:17 -08001853 scan_balance = SCAN_FRACT;
1854
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001855 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001856 * With swappiness at 100, anonymous and file have the same priority.
1857 * This scanning priority is essentially the inverse of IO cost.
1858 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001859 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001860 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001861
1862 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001863 * OK, so we have swap space and a fair amount of page cache
1864 * pages. We use the recently rotated / recently scanned
1865 * ratios to determine how valuable each cache is.
1866 *
1867 * Because workloads change over time (and to avoid overflow)
1868 * we keep these statistics as a floating average, which ends
1869 * up weighing recent references more than old ones.
1870 *
1871 * anon in [0], file in [1]
1872 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001873 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001874 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001875 reclaim_stat->recent_scanned[0] /= 2;
1876 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001877 }
1878
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001879 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001880 reclaim_stat->recent_scanned[1] /= 2;
1881 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001882 }
1883
1884 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001885 * The amount of pressure on anon vs file pages is inversely
1886 * proportional to the fraction of recently scanned pages on
1887 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001888 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001889 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001890 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001891
Satoru Moriyafe350042012-05-29 15:06:47 -07001892 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001893 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001894 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001895
Shaohua Li76a33fc2010-05-24 14:32:36 -07001896 fraction[0] = ap;
1897 fraction[1] = fp;
1898 denominator = ap + fp + 1;
1899out:
Hugh Dickins41113042012-01-12 17:20:01 -08001900 for_each_evictable_lru(lru) {
1901 int file = is_file_lru(lru);
Johannes Weinerd778df52013-02-22 16:32:12 -08001902 unsigned long size;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001903 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001904
Johannes Weinerd778df52013-02-22 16:32:12 -08001905 size = get_lru_size(lruvec, lru);
Johannes Weiner10316b32013-02-22 16:32:14 -08001906 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08001907
Johannes Weiner10316b32013-02-22 16:32:14 -08001908 if (!scan && force_scan)
1909 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08001910
1911 switch (scan_balance) {
1912 case SCAN_EQUAL:
1913 /* Scan lists relative to size */
1914 break;
1915 case SCAN_FRACT:
1916 /*
1917 * Scan types proportional to swappiness and
1918 * their relative recent reclaim efficiency.
1919 */
1920 scan = div64_u64(scan * fraction[file], denominator);
1921 break;
1922 case SCAN_FILE:
1923 case SCAN_ANON:
1924 /* Scan one type exclusively */
1925 if ((scan_balance == SCAN_FILE) != file)
1926 scan = 0;
1927 break;
1928 default:
1929 /* Look ma, no brain */
1930 BUG();
1931 }
Hugh Dickins41113042012-01-12 17:20:01 -08001932 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001933 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001934}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001935
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001936/*
1937 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1938 */
1939static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1940{
1941 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07001942 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001943 unsigned long nr_to_scan;
1944 enum lru_list lru;
1945 unsigned long nr_reclaimed = 0;
1946 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1947 struct blk_plug plug;
Mel Gormane82e0562013-07-03 15:01:44 -07001948 bool scan_adjusted = false;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001949
1950 get_scan_count(lruvec, sc, nr);
1951
Mel Gormane82e0562013-07-03 15:01:44 -07001952 /* Record the original scan target for proportional adjustments later */
1953 memcpy(targets, nr, sizeof(nr));
1954
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001955 blk_start_plug(&plug);
1956 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1957 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07001958 unsigned long nr_anon, nr_file, percentage;
1959 unsigned long nr_scanned;
1960
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001961 for_each_evictable_lru(lru) {
1962 if (nr[lru]) {
1963 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1964 nr[lru] -= nr_to_scan;
1965
1966 nr_reclaimed += shrink_list(lru, nr_to_scan,
1967 lruvec, sc);
1968 }
1969 }
Mel Gormane82e0562013-07-03 15:01:44 -07001970
1971 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
1972 continue;
1973
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001974 /*
Mel Gormane82e0562013-07-03 15:01:44 -07001975 * For global direct reclaim, reclaim only the number of pages
1976 * requested. Less care is taken to scan proportionally as it
1977 * is more important to minimise direct reclaim stall latency
1978 * than it is to properly age the LRU lists.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001979 */
Mel Gormane82e0562013-07-03 15:01:44 -07001980 if (global_reclaim(sc) && !current_is_kswapd())
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001981 break;
Mel Gormane82e0562013-07-03 15:01:44 -07001982
1983 /*
1984 * For kswapd and memcg, reclaim at least the number of pages
1985 * requested. Ensure that the anon and file LRUs shrink
1986 * proportionally what was requested by get_scan_count(). We
1987 * stop reclaiming one LRU and reduce the amount scanning
1988 * proportional to the original scan target.
1989 */
1990 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
1991 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
1992
1993 if (nr_file > nr_anon) {
1994 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
1995 targets[LRU_ACTIVE_ANON] + 1;
1996 lru = LRU_BASE;
1997 percentage = nr_anon * 100 / scan_target;
1998 } else {
1999 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2000 targets[LRU_ACTIVE_FILE] + 1;
2001 lru = LRU_FILE;
2002 percentage = nr_file * 100 / scan_target;
2003 }
2004
2005 /* Stop scanning the smaller of the LRU */
2006 nr[lru] = 0;
2007 nr[lru + LRU_ACTIVE] = 0;
2008
2009 /*
2010 * Recalculate the other LRU scan count based on its original
2011 * scan target and the percentage scanning already complete
2012 */
2013 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2014 nr_scanned = targets[lru] - nr[lru];
2015 nr[lru] = targets[lru] * (100 - percentage) / 100;
2016 nr[lru] -= min(nr[lru], nr_scanned);
2017
2018 lru += LRU_ACTIVE;
2019 nr_scanned = targets[lru] - nr[lru];
2020 nr[lru] = targets[lru] * (100 - percentage) / 100;
2021 nr[lru] -= min(nr[lru], nr_scanned);
2022
2023 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002024 }
2025 blk_finish_plug(&plug);
2026 sc->nr_reclaimed += nr_reclaimed;
2027
2028 /*
2029 * Even if we did not try to evict anon pages at all, we want to
2030 * rebalance the anon lru active/inactive ratio.
2031 */
2032 if (inactive_anon_is_low(lruvec))
2033 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2034 sc, LRU_ACTIVE_ANON);
2035
2036 throttle_vm_writeout(sc->gfp_mask);
2037}
2038
Mel Gorman23b9da52012-05-29 15:06:20 -07002039/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002040static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002041{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002042 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002043 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002044 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002045 return true;
2046
2047 return false;
2048}
2049
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002050/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002051 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2052 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2053 * true if more pages should be reclaimed such that when the page allocator
2054 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2055 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002056 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002057static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002058 unsigned long nr_reclaimed,
2059 unsigned long nr_scanned,
2060 struct scan_control *sc)
2061{
2062 unsigned long pages_for_compaction;
2063 unsigned long inactive_lru_pages;
2064
2065 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002066 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002067 return false;
2068
Mel Gorman28765922011-02-25 14:44:20 -08002069 /* Consider stopping depending on scan and reclaim activity */
2070 if (sc->gfp_mask & __GFP_REPEAT) {
2071 /*
2072 * For __GFP_REPEAT allocations, stop reclaiming if the
2073 * full LRU list has been scanned and we are still failing
2074 * to reclaim pages. This full LRU scan is potentially
2075 * expensive but a __GFP_REPEAT caller really wants to succeed
2076 */
2077 if (!nr_reclaimed && !nr_scanned)
2078 return false;
2079 } else {
2080 /*
2081 * For non-__GFP_REPEAT allocations which can presumably
2082 * fail without consequence, stop if we failed to reclaim
2083 * any pages from the last SWAP_CLUSTER_MAX number of
2084 * pages that were scanned. This will return to the
2085 * caller faster at the risk reclaim/compaction and
2086 * the resulting allocation attempt fails
2087 */
2088 if (!nr_reclaimed)
2089 return false;
2090 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002091
2092 /*
2093 * If we have not reclaimed enough pages for compaction and the
2094 * inactive lists are large enough, continue reclaiming
2095 */
2096 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002097 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002098 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002099 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002100 if (sc->nr_reclaimed < pages_for_compaction &&
2101 inactive_lru_pages > pages_for_compaction)
2102 return true;
2103
2104 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002105 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002106 case COMPACT_PARTIAL:
2107 case COMPACT_CONTINUE:
2108 return false;
2109 default:
2110 return true;
2111 }
2112}
2113
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002114static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002115{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002116 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002117
Johannes Weiner56600482012-01-12 17:17:59 -08002118 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002119 struct mem_cgroup *root = sc->target_mem_cgroup;
2120 struct mem_cgroup_reclaim_cookie reclaim = {
2121 .zone = zone,
2122 .priority = sc->priority,
2123 };
2124 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002125
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002126 nr_reclaimed = sc->nr_reclaimed;
2127 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002128
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002129 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2130 do {
2131 struct lruvec *lruvec;
2132
2133 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2134
2135 shrink_lruvec(lruvec, sc);
2136
2137 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002138 * Direct reclaim and kswapd have to scan all memory
2139 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002140 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002141 *
2142 * Limit reclaim, on the other hand, only cares about
2143 * nr_to_reclaim pages to be reclaimed and it will
2144 * retry with decreasing priority if one round over the
2145 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002146 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002147 if (!global_reclaim(sc) &&
2148 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002149 mem_cgroup_iter_break(root, memcg);
2150 break;
2151 }
2152 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2153 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002154
2155 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2156 sc->nr_scanned - nr_scanned,
2157 sc->nr_reclaimed - nr_reclaimed);
2158
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002159 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2160 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002161}
2162
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002163/* Returns true if compaction should go ahead for a high-order request */
2164static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2165{
2166 unsigned long balance_gap, watermark;
2167 bool watermark_ok;
2168
2169 /* Do not consider compaction for orders reclaim is meant to satisfy */
2170 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2171 return false;
2172
2173 /*
2174 * Compaction takes time to run and there are potentially other
2175 * callers using the pages just freed. Continue reclaiming until
2176 * there is a buffer of free pages available to give compaction
2177 * a reasonable chance of completing and allocating the page
2178 */
2179 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002180 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002181 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2182 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2183 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2184
2185 /*
2186 * If compaction is deferred, reclaim up to a point where
2187 * compaction will have a chance of success when re-enabled
2188 */
Rik van Rielaff62242012-03-21 16:33:52 -07002189 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002190 return watermark_ok;
2191
2192 /* If compaction is not ready to start, keep reclaiming */
2193 if (!compaction_suitable(zone, sc->order))
2194 return false;
2195
2196 return watermark_ok;
2197}
2198
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199/*
2200 * This is the direct reclaim path, for page-allocating processes. We only
2201 * try to reclaim pages from zones which will satisfy the caller's allocation
2202 * request.
2203 *
Mel Gorman41858962009-06-16 15:32:12 -07002204 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2205 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2207 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002208 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2209 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2210 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 * If a zone is deemed to be full of pinned pages then just give it a light
2213 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002214 *
2215 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002216 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002217 * the caller that it should consider retrying the allocation instead of
2218 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002220static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221{
Mel Gormandd1a2392008-04-28 02:12:17 -07002222 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002223 struct zone *zone;
Ying Hand149e3b2011-05-26 16:25:27 -07002224 unsigned long nr_soft_reclaimed;
2225 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002226 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002227
Mel Gormancc715d92012-03-21 16:34:00 -07002228 /*
2229 * If the number of buffer_heads in the machine exceeds the maximum
2230 * allowed level, force direct reclaim to scan the highmem zone as
2231 * highmem pages could be pinning lowmem pages storing buffer_heads
2232 */
2233 if (buffer_heads_over_limit)
2234 sc->gfp_mask |= __GFP_HIGHMEM;
2235
Mel Gormand4debc62010-08-09 17:19:29 -07002236 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2237 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002238 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002240 /*
2241 * Take care memory controller reclaiming has small influence
2242 * to global LRU.
2243 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002244 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002245 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2246 continue;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002247 if (zone->all_unreclaimable &&
2248 sc->priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002249 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002250 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002251 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002252 * If we already have plenty of memory free for
2253 * compaction in this zone, don't free any more.
2254 * Even though compaction is invoked for any
2255 * non-zero order, only frequent costly order
2256 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002257 * noticeable problem, like transparent huge
2258 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002259 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002260 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002261 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002262 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002263 }
Rik van Riele0887c12011-10-31 17:09:31 -07002264 }
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002265 /*
2266 * This steals pages from memory cgroups over softlimit
2267 * and returns the number of reclaimed pages and
2268 * scanned pages. This works for global memory pressure
2269 * and balancing, not for a memcg's limit.
2270 */
2271 nr_soft_scanned = 0;
2272 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2273 sc->order, sc->gfp_mask,
2274 &nr_soft_scanned);
2275 sc->nr_reclaimed += nr_soft_reclaimed;
2276 sc->nr_scanned += nr_soft_scanned;
2277 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002278 }
Nick Piggin408d8542006-09-25 23:31:27 -07002279
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002280 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 }
Mel Gormane0c23272011-10-31 17:09:33 -07002282
Mel Gorman0cee34f2012-01-12 17:19:49 -08002283 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002284}
2285
2286static bool zone_reclaimable(struct zone *zone)
2287{
2288 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2289}
2290
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002291/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002292static bool all_unreclaimable(struct zonelist *zonelist,
2293 struct scan_control *sc)
2294{
2295 struct zoneref *z;
2296 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002297
2298 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2299 gfp_zone(sc->gfp_mask), sc->nodemask) {
2300 if (!populated_zone(zone))
2301 continue;
2302 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2303 continue;
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002304 if (!zone->all_unreclaimable)
2305 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002306 }
2307
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002308 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002310
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311/*
2312 * This is the main entry point to direct page reclaim.
2313 *
2314 * If a full scan of the inactive list fails to free enough memory then we
2315 * are "out of memory" and something needs to be killed.
2316 *
2317 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2318 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002319 * caller can't do much about. We kick the writeback threads and take explicit
2320 * naps in the hope that some of these pages can be written. But if the
2321 * allocating task holds filesystem locks which prevent writeout this might not
2322 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002323 *
2324 * returns: 0, if no pages reclaimed
2325 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 */
Mel Gormandac1d272008-04-28 02:12:12 -07002327static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002328 struct scan_control *sc,
2329 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330{
Andrew Morton69e05942006-03-22 00:08:19 -08002331 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002333 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002334 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002335 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002336 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337
Keika Kobayashi873b4772008-07-25 01:48:52 -07002338 delayacct_freepages_start();
2339
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002340 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002341 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002343 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002344 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2345 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002346 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002347 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002348
Balbir Singh66e17072008-02-07 00:13:56 -08002349 /*
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002350 * Don't shrink slabs when reclaiming memory from over limit
2351 * cgroups but do shrink slab at least once when aborting
2352 * reclaim for compaction to avoid unevenly scanning file/anon
2353 * LRU pages over slab pages.
Balbir Singh66e17072008-02-07 00:13:56 -08002354 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002355 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002356 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07002357 for_each_zone_zonelist(zone, z, zonelist,
2358 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002359 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2360 continue;
2361
2362 lru_pages += zone_reclaimable_pages(zone);
2363 }
2364
Ying Han1495f232011-05-24 17:12:27 -07002365 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002366 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002367 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002368 reclaim_state->reclaimed_slab = 0;
2369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 }
Balbir Singh66e17072008-02-07 00:13:56 -08002371 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002372 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
2375 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002376 * If we're getting trouble reclaiming, start doing
2377 * writepage even in laptop mode.
2378 */
2379 if (sc->priority < DEF_PRIORITY - 2)
2380 sc->may_writepage = 1;
2381
2382 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 * Try to write back as many pages as we just scanned. This
2384 * tends to cause slow streaming writers to write data to the
2385 * disk smoothly, at the dirtying rate, which is nice. But
2386 * that's undesirable in laptop mode, where we *want* lumpy
2387 * writeout. So in laptop mode, write out the whole world.
2388 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002389 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2390 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002391 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2392 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002393 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002395 } while (--sc->priority >= 0 && !aborted_reclaim);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002398 delayacct_freepages_end();
2399
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002400 if (sc->nr_reclaimed)
2401 return sc->nr_reclaimed;
2402
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002403 /*
2404 * As hibernation is going on, kswapd is freezed so that it can't mark
2405 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2406 * check.
2407 */
2408 if (oom_killer_disabled)
2409 return 0;
2410
Mel Gorman0cee34f2012-01-12 17:19:49 -08002411 /* Aborted reclaim to try compaction? don't OOM, then */
2412 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002413 return 1;
2414
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002415 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002416 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002417 return 1;
2418
2419 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420}
2421
Mel Gorman55150612012-07-31 16:44:35 -07002422static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2423{
2424 struct zone *zone;
2425 unsigned long pfmemalloc_reserve = 0;
2426 unsigned long free_pages = 0;
2427 int i;
2428 bool wmark_ok;
2429
2430 for (i = 0; i <= ZONE_NORMAL; i++) {
2431 zone = &pgdat->node_zones[i];
2432 pfmemalloc_reserve += min_wmark_pages(zone);
2433 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2434 }
2435
2436 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2437
2438 /* kswapd must be awake if processes are being throttled */
2439 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2440 pgdat->classzone_idx = min(pgdat->classzone_idx,
2441 (enum zone_type)ZONE_NORMAL);
2442 wake_up_interruptible(&pgdat->kswapd_wait);
2443 }
2444
2445 return wmark_ok;
2446}
2447
2448/*
2449 * Throttle direct reclaimers if backing storage is backed by the network
2450 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2451 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002452 * when the low watermark is reached.
2453 *
2454 * Returns true if a fatal signal was delivered during throttling. If this
2455 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002456 */
Mel Gorman50694c22012-11-26 16:29:48 -08002457static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002458 nodemask_t *nodemask)
2459{
2460 struct zone *zone;
2461 int high_zoneidx = gfp_zone(gfp_mask);
2462 pg_data_t *pgdat;
2463
2464 /*
2465 * Kernel threads should not be throttled as they may be indirectly
2466 * responsible for cleaning pages necessary for reclaim to make forward
2467 * progress. kjournald for example may enter direct reclaim while
2468 * committing a transaction where throttling it could forcing other
2469 * processes to block on log_wait_commit().
2470 */
2471 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002472 goto out;
2473
2474 /*
2475 * If a fatal signal is pending, this process should not throttle.
2476 * It should return quickly so it can exit and free its memory
2477 */
2478 if (fatal_signal_pending(current))
2479 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002480
2481 /* Check if the pfmemalloc reserves are ok */
2482 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2483 pgdat = zone->zone_pgdat;
2484 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002485 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002486
Mel Gorman68243e72012-07-31 16:44:39 -07002487 /* Account for the throttling */
2488 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2489
Mel Gorman55150612012-07-31 16:44:35 -07002490 /*
2491 * If the caller cannot enter the filesystem, it's possible that it
2492 * is due to the caller holding an FS lock or performing a journal
2493 * transaction in the case of a filesystem like ext[3|4]. In this case,
2494 * it is not safe to block on pfmemalloc_wait as kswapd could be
2495 * blocked waiting on the same lock. Instead, throttle for up to a
2496 * second before continuing.
2497 */
2498 if (!(gfp_mask & __GFP_FS)) {
2499 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2500 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002501
2502 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002503 }
2504
2505 /* Throttle until kswapd wakes the process */
2506 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2507 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002508
2509check_pending:
2510 if (fatal_signal_pending(current))
2511 return true;
2512
2513out:
2514 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002515}
2516
Mel Gormandac1d272008-04-28 02:12:12 -07002517unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002518 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002519{
Mel Gorman33906bc2010-08-09 17:19:16 -07002520 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002521 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002522 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002523 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002524 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002525 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002526 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002527 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002528 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002529 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002530 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002531 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002532 struct shrink_control shrink = {
2533 .gfp_mask = sc.gfp_mask,
2534 };
Balbir Singh66e17072008-02-07 00:13:56 -08002535
Mel Gorman55150612012-07-31 16:44:35 -07002536 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002537 * Do not enter reclaim if fatal signal was delivered while throttled.
2538 * 1 is returned so that the page allocator does not OOM kill at this
2539 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002540 */
Mel Gorman50694c22012-11-26 16:29:48 -08002541 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002542 return 1;
2543
Mel Gorman33906bc2010-08-09 17:19:16 -07002544 trace_mm_vmscan_direct_reclaim_begin(order,
2545 sc.may_writepage,
2546 gfp_mask);
2547
Ying Hana09ed5e2011-05-24 17:12:26 -07002548 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002549
2550 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2551
2552 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002553}
2554
Andrew Mortonc255a452012-07-31 16:43:02 -07002555#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002556
Johannes Weiner72835c82012-01-12 17:18:32 -08002557unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002558 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002559 struct zone *zone,
2560 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002561{
2562 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002563 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002564 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002565 .may_writepage = !laptop_mode,
2566 .may_unmap = 1,
2567 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002568 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002569 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002570 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002571 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002572 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002573
Balbir Singh4e416952009-09-23 15:56:39 -07002574 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2575 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002576
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002577 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002578 sc.may_writepage,
2579 sc.gfp_mask);
2580
Balbir Singh4e416952009-09-23 15:56:39 -07002581 /*
2582 * NOTE: Although we can get the priority field, using it
2583 * here is not a good idea, since it limits the pages we can scan.
2584 * if we don't reclaim here, the shrink_zone from balance_pgdat
2585 * will pick up pages from other mem cgroup's as well. We hack
2586 * the priority and make it zero.
2587 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002588 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002589
2590 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2591
Ying Han0ae5e892011-05-26 16:25:25 -07002592 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002593 return sc.nr_reclaimed;
2594}
2595
Johannes Weiner72835c82012-01-12 17:18:32 -08002596unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002597 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002598 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002599{
Balbir Singh4e416952009-09-23 15:56:39 -07002600 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002601 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002602 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002603 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002604 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002605 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002606 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002607 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002608 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002609 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002610 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002611 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002612 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2613 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2614 };
2615 struct shrink_control shrink = {
2616 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002617 };
Balbir Singh66e17072008-02-07 00:13:56 -08002618
Ying Han889976d2011-05-26 16:25:33 -07002619 /*
2620 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2621 * take care of from where we get pages. So the node where we start the
2622 * scan does not need to be the current node.
2623 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002624 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002625
2626 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002627
2628 trace_mm_vmscan_memcg_reclaim_begin(0,
2629 sc.may_writepage,
2630 sc.gfp_mask);
2631
Ying Hana09ed5e2011-05-24 17:12:26 -07002632 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002633
2634 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2635
2636 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002637}
2638#endif
2639
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002640static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002641{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002642 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002643
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002644 if (!total_swap_pages)
2645 return;
2646
2647 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2648 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002649 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002650
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002651 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002652 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002653 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002654
2655 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2656 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002657}
2658
Johannes Weiner60cefed2012-11-29 13:54:23 -08002659static bool zone_balanced(struct zone *zone, int order,
2660 unsigned long balance_gap, int classzone_idx)
2661{
2662 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2663 balance_gap, classzone_idx, 0))
2664 return false;
2665
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002666 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2667 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002668 return false;
2669
2670 return true;
2671}
2672
Mel Gorman1741c872011-01-13 15:46:21 -08002673/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002674 * pgdat_balanced() is used when checking if a node is balanced.
2675 *
2676 * For order-0, all zones must be balanced!
2677 *
2678 * For high-order allocations only zones that meet watermarks and are in a
2679 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2680 * total of balanced pages must be at least 25% of the zones allowed by
2681 * classzone_idx for the node to be considered balanced. Forcing all zones to
2682 * be balanced for high orders can cause excessive reclaim when there are
2683 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002684 * The choice of 25% is due to
2685 * o a 16M DMA zone that is balanced will not balance a zone on any
2686 * reasonable sized machine
2687 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002688 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002689 * would need to be at least 256M for it to be balance a whole node.
2690 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2691 * to balance a node on its own. These seemed like reasonable ratios.
2692 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002693static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002694{
Jiang Liub40da042013-02-22 16:33:52 -08002695 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002696 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002697 int i;
2698
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002699 /* Check the watermark levels */
2700 for (i = 0; i <= classzone_idx; i++) {
2701 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002702
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002703 if (!populated_zone(zone))
2704 continue;
2705
Jiang Liub40da042013-02-22 16:33:52 -08002706 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002707
2708 /*
2709 * A special case here:
2710 *
2711 * balance_pgdat() skips over all_unreclaimable after
2712 * DEF_PRIORITY. Effectively, it considers them balanced so
2713 * they must be considered balanced here as well!
2714 */
2715 if (zone->all_unreclaimable) {
Jiang Liub40da042013-02-22 16:33:52 -08002716 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002717 continue;
2718 }
2719
2720 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002721 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002722 else if (!order)
2723 return false;
2724 }
2725
2726 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002727 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002728 else
2729 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002730}
2731
Mel Gorman55150612012-07-31 16:44:35 -07002732/*
2733 * Prepare kswapd for sleeping. This verifies that there are no processes
2734 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2735 *
2736 * Returns true if kswapd is ready to sleep
2737 */
2738static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002739 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002740{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002741 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2742 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002743 return false;
2744
2745 /*
2746 * There is a potential race between when kswapd checks its watermarks
2747 * and a process gets throttled. There is also a potential race if
2748 * processes get throttled, kswapd wakes, a large process exits therby
2749 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2750 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2751 * so wake them now if necessary. If necessary, processes will wake
2752 * kswapd and get throttled again
2753 */
2754 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2755 wake_up(&pgdat->pfmemalloc_wait);
2756 return false;
2757 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002758
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002759 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002760}
2761
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762/*
Mel Gorman75485362013-07-03 15:01:42 -07002763 * kswapd shrinks the zone by the number of pages required to reach
2764 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002765 *
2766 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002767 * reclaim or if the lack of progress was due to pages under writeback.
2768 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002769 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002770static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002771 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002772 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002773 unsigned long lru_pages,
2774 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002775{
2776 unsigned long nr_slab;
Mel Gorman7c954f62013-07-03 15:01:54 -07002777 int testorder = sc->order;
2778 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002779 struct reclaim_state *reclaim_state = current->reclaim_state;
2780 struct shrink_control shrink = {
2781 .gfp_mask = sc->gfp_mask,
2782 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002783 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002784
2785 /* Reclaim above the high watermark. */
2786 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002787
2788 /*
2789 * Kswapd reclaims only single pages with compaction enabled. Trying
2790 * too hard to reclaim until contiguous free pages have become
2791 * available can hurt performance by evicting too much useful data
2792 * from memory. Do not reclaim more than needed for compaction.
2793 */
2794 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2795 compaction_suitable(zone, sc->order) !=
2796 COMPACT_SKIPPED)
2797 testorder = 0;
2798
2799 /*
2800 * We put equal pressure on every zone, unless one zone has way too
2801 * many pages free already. The "too many pages" is defined as the
2802 * high wmark plus a "gap" where the gap is either the low
2803 * watermark or 1% of the zone, whichever is smaller.
2804 */
2805 balance_gap = min(low_wmark_pages(zone),
2806 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2807 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2808
2809 /*
2810 * If there is no low memory pressure or the zone is balanced then no
2811 * reclaim is necessary
2812 */
2813 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2814 if (!lowmem_pressure && zone_balanced(zone, testorder,
2815 balance_gap, classzone_idx))
2816 return true;
2817
Mel Gorman75485362013-07-03 15:01:42 -07002818 shrink_zone(zone, sc);
2819
2820 reclaim_state->reclaimed_slab = 0;
2821 nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2822 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823
Mel Gorman2ab44f42013-07-03 15:01:47 -07002824 /* Account for the number of pages attempted to reclaim */
2825 *nr_attempted += sc->nr_to_reclaim;
2826
Mel Gorman75485362013-07-03 15:01:42 -07002827 if (nr_slab == 0 && !zone_reclaimable(zone))
2828 zone->all_unreclaimable = 1;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002829
Mel Gorman283aba92013-07-03 15:01:51 -07002830 zone_clear_flag(zone, ZONE_WRITEBACK);
2831
Mel Gorman7c954f62013-07-03 15:01:54 -07002832 /*
2833 * If a zone reaches its high watermark, consider it to be no longer
2834 * congested. It's possible there are dirty pages backed by congested
2835 * BDIs but as pressure is relieved, speculatively avoid congestion
2836 * waits.
2837 */
2838 if (!zone->all_unreclaimable &&
2839 zone_balanced(zone, testorder, 0, classzone_idx)) {
2840 zone_clear_flag(zone, ZONE_CONGESTED);
2841 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2842 }
2843
Mel Gormanb8e83b92013-07-03 15:01:45 -07002844 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002845}
2846
2847/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002849 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002851 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 *
2853 * There is special handling here for zones which are full of pinned pages.
2854 * This can happen if the pages are all mlocked, or if they are all used by
2855 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2856 * What we do is to detect the case where all pages in the zone have been
2857 * scanned twice and there has been zero successful reclaim. Mark the zone as
2858 * dead and from now on, only perform a short scan. Basically we're polling
2859 * the zone for when the problem goes away.
2860 *
2861 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002862 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2863 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2864 * lower zones regardless of the number of free pages in the lower zones. This
2865 * interoperates with the page allocator fallback scheme to ensure that aging
2866 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 */
Mel Gorman99504742011-01-13 15:46:20 -08002868static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002869 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002872 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Ying Han0ae5e892011-05-26 16:25:25 -07002873 unsigned long nr_soft_reclaimed;
2874 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002875 struct scan_control sc = {
2876 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002877 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002878 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002879 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002880 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002881 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002882 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002883 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07002884 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002886 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002888 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002889 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002890 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07002891
2892 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002894 /*
2895 * Scan in the highmem->dma direction for the highest
2896 * zone which needs scanning
2897 */
2898 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2899 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002901 if (!populated_zone(zone))
2902 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002904 if (zone->all_unreclaimable &&
2905 sc.priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002906 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907
Rik van Riel556adec2008-10-18 20:26:34 -07002908 /*
2909 * Do some background aging of the anon list, to give
2910 * pages a chance to be referenced before reclaiming.
2911 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002912 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002913
Mel Gormancc715d92012-03-21 16:34:00 -07002914 /*
2915 * If the number of buffer_heads in the machine
2916 * exceeds the maximum allowed level and this node
2917 * has a highmem zone, force kswapd to reclaim from
2918 * it to relieve lowmem pressure.
2919 */
2920 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2921 end_zone = i;
2922 break;
2923 }
2924
Johannes Weiner60cefed2012-11-29 13:54:23 -08002925 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002926 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002927 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002928 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07002929 /*
2930 * If balanced, clear the dirty and congested
2931 * flags
2932 */
Shaohua Li439423f2011-08-25 15:59:12 -07002933 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07002934 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002937
Mel Gormanb8e83b92013-07-03 15:01:45 -07002938 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002939 goto out;
2940
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 for (i = 0; i <= end_zone; i++) {
2942 struct zone *zone = pgdat->node_zones + i;
2943
Mel Gorman2ab44f42013-07-03 15:01:47 -07002944 if (!populated_zone(zone))
2945 continue;
2946
Wu Fengguangadea02a2009-09-21 17:01:42 -07002947 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07002948
2949 /*
2950 * If any zone is currently balanced then kswapd will
2951 * not call compaction as it is expected that the
2952 * necessary pages are already available.
2953 */
2954 if (pgdat_needs_compaction &&
2955 zone_watermark_ok(zone, order,
2956 low_wmark_pages(zone),
2957 *classzone_idx, 0))
2958 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 }
2960
2961 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07002962 * If we're getting trouble reclaiming, start doing writepage
2963 * even in laptop mode.
2964 */
2965 if (sc.priority < DEF_PRIORITY - 2)
2966 sc.may_writepage = 1;
2967
2968 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 * Now scan the zone in the dma->highmem direction, stopping
2970 * at the last zone which needs scanning.
2971 *
2972 * We do this because the page allocator works in the opposite
2973 * direction. This prevents the page allocator from allocating
2974 * pages behind kswapd's direction of progress, which would
2975 * cause too much scanning of the lower zones.
2976 */
2977 for (i = 0; i <= end_zone; i++) {
2978 struct zone *zone = pgdat->node_zones + i;
2979
Con Kolivasf3fe6512006-01-06 00:11:15 -08002980 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 continue;
2982
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002983 if (zone->all_unreclaimable &&
2984 sc.priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 continue;
2986
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002988
Ying Han0ae5e892011-05-26 16:25:25 -07002989 nr_soft_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002990 /*
2991 * Call soft limit reclaim before calling shrink_zone.
Balbir Singh4e416952009-09-23 15:56:39 -07002992 */
Ying Han0ae5e892011-05-26 16:25:25 -07002993 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2994 order, sc.gfp_mask,
2995 &nr_soft_scanned);
2996 sc.nr_reclaimed += nr_soft_reclaimed;
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002997
Rik van Riel32a43302007-10-16 01:24:50 -07002998 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07002999 * There should be no need to raise the scanning
3000 * priority if enough pages are already being scanned
3001 * that that high watermark would be met at 100%
3002 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003003 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003004 if (kswapd_shrink_zone(zone, end_zone, &sc,
3005 lru_pages, &nr_attempted))
3006 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 }
Mel Gorman55150612012-07-31 16:44:35 -07003008
3009 /*
3010 * If the low watermark is met there is no need for processes
3011 * to be throttled on pfmemalloc_wait as they should not be
3012 * able to safely make forward progress. Wake them
3013 */
3014 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3015 pfmemalloc_watermark_ok(pgdat))
3016 wake_up(&pgdat->pfmemalloc_wait);
3017
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003019 * Fragmentation may mean that the system cannot be rebalanced
3020 * for high-order allocations in all zones. If twice the
3021 * allocation size has been reclaimed and the zones are still
3022 * not balanced then recheck the watermarks at order-0 to
3023 * prevent kswapd reclaiming excessively. Assume that a
3024 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003026 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003027 order = sc.order = 0;
3028
Mel Gormanb8e83b92013-07-03 15:01:45 -07003029 /* Check if kswapd should be suspending */
3030 if (try_to_freeze() || kthread_should_stop())
3031 break;
3032
3033 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003034 * Compact if necessary and kswapd is reclaiming at least the
3035 * high watermark number of pages as requsted
3036 */
3037 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3038 compact_pgdat(pgdat, order);
3039
3040 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003041 * Raise priority if scanning rate is too low or there was no
3042 * progress in reclaiming pages
3043 */
3044 if (raise_priority || !sc.nr_reclaimed)
3045 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003046 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003047 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048
Mel Gormanb8e83b92013-07-03 15:01:45 -07003049out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003050 /*
Mel Gorman55150612012-07-31 16:44:35 -07003051 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003052 * makes a decision on the order we were last reclaiming at. However,
3053 * if another caller entered the allocator slow path while kswapd
3054 * was awake, order will remain at the higher level
3055 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003056 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003057 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058}
3059
Mel Gormandc83edd2011-01-13 15:46:26 -08003060static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003061{
3062 long remaining = 0;
3063 DEFINE_WAIT(wait);
3064
3065 if (freezing(current) || kthread_should_stop())
3066 return;
3067
3068 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3069
3070 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003071 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003072 remaining = schedule_timeout(HZ/10);
3073 finish_wait(&pgdat->kswapd_wait, &wait);
3074 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3075 }
3076
3077 /*
3078 * After a short sleep, check if it was a premature sleep. If not, then
3079 * go fully to sleep until explicitly woken up.
3080 */
Mel Gorman55150612012-07-31 16:44:35 -07003081 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003082 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3083
3084 /*
3085 * vmstat counters are not perfectly accurate and the estimated
3086 * value for counters such as NR_FREE_PAGES can deviate from the
3087 * true value by nr_online_cpus * threshold. To avoid the zone
3088 * watermarks being breached while under pressure, we reduce the
3089 * per-cpu vmstat threshold while kswapd is awake and restore
3090 * them before going back to sleep.
3091 */
3092 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003093
Mel Gorman62997022012-10-08 16:32:47 -07003094 /*
3095 * Compaction records what page blocks it recently failed to
3096 * isolate pages from and skips them in the future scanning.
3097 * When kswapd is going to sleep, it is reasonable to assume
3098 * that pages and compaction may succeed so reset the cache.
3099 */
3100 reset_isolation_suitable(pgdat);
3101
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003102 if (!kthread_should_stop())
3103 schedule();
3104
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003105 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3106 } else {
3107 if (remaining)
3108 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3109 else
3110 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3111 }
3112 finish_wait(&pgdat->kswapd_wait, &wait);
3113}
3114
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115/*
3116 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003117 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 *
3119 * This basically trickles out pages so that we have _some_
3120 * free memory available even if there is no other activity
3121 * that frees anything up. This is needed for things like routing
3122 * etc, where we otherwise might have all activity going on in
3123 * asynchronous contexts that cannot page things out.
3124 *
3125 * If there are applications that are active memory-allocators
3126 * (most normal use), this basically shouldn't matter.
3127 */
3128static int kswapd(void *p)
3129{
Mel Gorman215ddd62011-07-08 15:39:40 -07003130 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003131 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003132 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003133 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 pg_data_t *pgdat = (pg_data_t*)p;
3135 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003136
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 struct reclaim_state reclaim_state = {
3138 .reclaimed_slab = 0,
3139 };
Rusty Russella70f7302009-03-13 14:49:46 +10303140 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141
Nick Piggincf40bd12009-01-21 08:12:39 +01003142 lockdep_set_current_reclaim_state(GFP_KERNEL);
3143
Rusty Russell174596a2009-01-01 10:12:29 +10303144 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003145 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 current->reclaim_state = &reclaim_state;
3147
3148 /*
3149 * Tell the memory management that we're a "memory allocator",
3150 * and that if we need more memory we should get access to it
3151 * regardless (see "__alloc_pages()"). "kswapd" should
3152 * never get caught in the normal page freeing logic.
3153 *
3154 * (Kswapd normally doesn't need memory anyway, but sometimes
3155 * you need a small amount of memory in order to be able to
3156 * page out something else, and this flag essentially protects
3157 * us from recursively trying to free more memory as we're
3158 * trying to free the first piece of memory in the first place).
3159 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003160 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003161 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162
Mel Gorman215ddd62011-07-08 15:39:40 -07003163 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003164 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003165 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003166 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003168 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003169
Mel Gorman215ddd62011-07-08 15:39:40 -07003170 /*
3171 * If the last balance_pgdat was unsuccessful it's unlikely a
3172 * new request of a similar or harder type will succeed soon
3173 * so consider going to sleep on the basis we reclaimed at
3174 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003175 if (balanced_classzone_idx >= new_classzone_idx &&
3176 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003177 new_order = pgdat->kswapd_max_order;
3178 new_classzone_idx = pgdat->classzone_idx;
3179 pgdat->kswapd_max_order = 0;
3180 pgdat->classzone_idx = pgdat->nr_zones - 1;
3181 }
3182
Mel Gorman99504742011-01-13 15:46:20 -08003183 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 /*
3185 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003186 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 */
3188 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003189 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003191 kswapd_try_to_sleep(pgdat, balanced_order,
3192 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003194 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003195 new_order = order;
3196 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003197 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003198 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200
David Rientjes8fe23e02009-12-14 17:58:33 -08003201 ret = try_to_freeze();
3202 if (kthread_should_stop())
3203 break;
3204
3205 /*
3206 * We can speed up thawing tasks if we don't call balance_pgdat
3207 * after returning from the refrigerator
3208 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003209 if (!ret) {
3210 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003211 balanced_classzone_idx = classzone_idx;
3212 balanced_order = balance_pgdat(pgdat, order,
3213 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003216
3217 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 return 0;
3219}
3220
3221/*
3222 * A zone is low on free memory, so wake its kswapd task to service it.
3223 */
Mel Gorman99504742011-01-13 15:46:20 -08003224void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225{
3226 pg_data_t *pgdat;
3227
Con Kolivasf3fe6512006-01-06 00:11:15 -08003228 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229 return;
3230
Paul Jackson02a0e532006-12-13 00:34:25 -08003231 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003233 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003234 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003235 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003236 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3237 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003238 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003240 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3241 return;
3242
3243 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003244 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245}
3246
Wu Fengguangadea02a2009-09-21 17:01:42 -07003247/*
3248 * The reclaimable count would be mostly accurate.
3249 * The less reclaimable pages may be
3250 * - mlocked pages, which will be moved to unevictable list when encountered
3251 * - mapped pages, which may require several travels to be reclaimed
3252 * - dirty pages, which is not "instantly" reclaimable
3253 */
3254unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003255{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003256 int nr;
3257
3258 nr = global_page_state(NR_ACTIVE_FILE) +
3259 global_page_state(NR_INACTIVE_FILE);
3260
Shaohua Liec8acf22013-02-22 16:34:38 -08003261 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003262 nr += global_page_state(NR_ACTIVE_ANON) +
3263 global_page_state(NR_INACTIVE_ANON);
3264
3265 return nr;
3266}
3267
3268unsigned long zone_reclaimable_pages(struct zone *zone)
3269{
3270 int nr;
3271
3272 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3273 zone_page_state(zone, NR_INACTIVE_FILE);
3274
Shaohua Liec8acf22013-02-22 16:34:38 -08003275 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003276 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3277 zone_page_state(zone, NR_INACTIVE_ANON);
3278
3279 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003280}
3281
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003282#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003284 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003285 * freed pages.
3286 *
3287 * Rather than trying to age LRUs the aim is to preserve the overall
3288 * LRU order by reclaiming preferentially
3289 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003291unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003293 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003294 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003295 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3296 .may_swap = 1,
3297 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003298 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003299 .nr_to_reclaim = nr_to_reclaim,
3300 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003301 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003302 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003304 struct shrink_control shrink = {
3305 .gfp_mask = sc.gfp_mask,
3306 };
3307 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003308 struct task_struct *p = current;
3309 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003311 p->flags |= PF_MEMALLOC;
3312 lockdep_set_current_reclaim_state(sc.gfp_mask);
3313 reclaim_state.reclaimed_slab = 0;
3314 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003315
Ying Hana09ed5e2011-05-24 17:12:26 -07003316 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003317
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003318 p->reclaim_state = NULL;
3319 lockdep_clear_current_reclaim_state();
3320 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003321
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003322 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003324#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326/* It's optimal to keep kswapds on the same CPUs as their memory, but
3327 not required for correctness. So if the last cpu in a node goes
3328 away, we get changed to run anywhere: as the first one comes back,
3329 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003330static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3331 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003333 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003335 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003336 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003337 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303338 const struct cpumask *mask;
3339
3340 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003341
Rusty Russell3e597942009-01-01 10:12:24 +10303342 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003344 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 }
3346 }
3347 return NOTIFY_OK;
3348}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349
Yasunori Goto3218ae12006-06-27 02:53:33 -07003350/*
3351 * This kswapd start function will be called by init and node-hot-add.
3352 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3353 */
3354int kswapd_run(int nid)
3355{
3356 pg_data_t *pgdat = NODE_DATA(nid);
3357 int ret = 0;
3358
3359 if (pgdat->kswapd)
3360 return 0;
3361
3362 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3363 if (IS_ERR(pgdat->kswapd)) {
3364 /* failure at boot is fatal */
3365 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003366 pr_err("Failed to start kswapd on node %d\n", nid);
3367 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003368 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003369 }
3370 return ret;
3371}
3372
David Rientjes8fe23e02009-12-14 17:58:33 -08003373/*
Jiang Liud8adde12012-07-11 14:01:52 -07003374 * Called by memory hotplug when all memory in a node is offlined. Caller must
3375 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003376 */
3377void kswapd_stop(int nid)
3378{
3379 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3380
Jiang Liud8adde12012-07-11 14:01:52 -07003381 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003382 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003383 NODE_DATA(nid)->kswapd = NULL;
3384 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003385}
3386
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387static int __init kswapd_init(void)
3388{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003389 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003390
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003392 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003393 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 hotcpu_notifier(cpu_callback, 0);
3395 return 0;
3396}
3397
3398module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003399
3400#ifdef CONFIG_NUMA
3401/*
3402 * Zone reclaim mode
3403 *
3404 * If non-zero call zone_reclaim when the number of free pages falls below
3405 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003406 */
3407int zone_reclaim_mode __read_mostly;
3408
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003409#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003410#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003411#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3412#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3413
Christoph Lameter9eeff232006-01-18 17:42:31 -08003414/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003415 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3416 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3417 * a zone.
3418 */
3419#define ZONE_RECLAIM_PRIORITY 4
3420
Christoph Lameter9eeff232006-01-18 17:42:31 -08003421/*
Christoph Lameter96146342006-07-03 00:24:13 -07003422 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3423 * occur.
3424 */
3425int sysctl_min_unmapped_ratio = 1;
3426
3427/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003428 * If the number of slab pages in a zone grows beyond this percentage then
3429 * slab reclaim needs to occur.
3430 */
3431int sysctl_min_slab_ratio = 5;
3432
Mel Gorman90afa5d2009-06-16 15:33:20 -07003433static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3434{
3435 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3436 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3437 zone_page_state(zone, NR_ACTIVE_FILE);
3438
3439 /*
3440 * It's possible for there to be more file mapped pages than
3441 * accounted for by the pages on the file LRU lists because
3442 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3443 */
3444 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3445}
3446
3447/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3448static long zone_pagecache_reclaimable(struct zone *zone)
3449{
3450 long nr_pagecache_reclaimable;
3451 long delta = 0;
3452
3453 /*
3454 * If RECLAIM_SWAP is set, then all file pages are considered
3455 * potentially reclaimable. Otherwise, we have to worry about
3456 * pages like swapcache and zone_unmapped_file_pages() provides
3457 * a better estimate
3458 */
3459 if (zone_reclaim_mode & RECLAIM_SWAP)
3460 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3461 else
3462 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3463
3464 /* If we can't clean pages, remove dirty pages from consideration */
3465 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3466 delta += zone_page_state(zone, NR_FILE_DIRTY);
3467
3468 /* Watch for any possible underflows due to delta */
3469 if (unlikely(delta > nr_pagecache_reclaimable))
3470 delta = nr_pagecache_reclaimable;
3471
3472 return nr_pagecache_reclaimable - delta;
3473}
3474
Christoph Lameter0ff38492006-09-25 23:31:52 -07003475/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003476 * Try to free up some pages from this zone through reclaim.
3477 */
Andrew Morton179e9632006-03-22 00:08:18 -08003478static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003479{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003480 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003481 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003482 struct task_struct *p = current;
3483 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003484 struct scan_control sc = {
3485 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003486 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003487 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003488 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003489 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003490 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003491 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003492 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003493 struct shrink_control shrink = {
3494 .gfp_mask = sc.gfp_mask,
3495 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003496 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003497
Christoph Lameter9eeff232006-01-18 17:42:31 -08003498 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003499 /*
3500 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3501 * and we also need to be able to write out pages for RECLAIM_WRITE
3502 * and RECLAIM_SWAP.
3503 */
3504 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003505 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003506 reclaim_state.reclaimed_slab = 0;
3507 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003508
Mel Gorman90afa5d2009-06-16 15:33:20 -07003509 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003510 /*
3511 * Free memory by calling shrink zone with increasing
3512 * priorities until we have enough memory freed.
3513 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003514 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003515 shrink_zone(zone, &sc);
3516 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003517 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003518
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003519 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3520 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003521 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003522 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003523 * many pages were freed in this zone. So we take the current
3524 * number of slab pages and shake the slab until it is reduced
3525 * by the same nr_pages that we used for reclaiming unmapped
3526 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003527 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07003528 * Note that shrink_slab will free memory on all zones and may
3529 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003530 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003531 for (;;) {
3532 unsigned long lru_pages = zone_reclaimable_pages(zone);
3533
3534 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003535 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003536 break;
3537
3538 /* Freed enough memory */
3539 nr_slab_pages1 = zone_page_state(zone,
3540 NR_SLAB_RECLAIMABLE);
3541 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3542 break;
3543 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003544
3545 /*
3546 * Update nr_reclaimed by the number of slab pages we
3547 * reclaimed from this zone.
3548 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003549 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3550 if (nr_slab_pages1 < nr_slab_pages0)
3551 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003552 }
3553
Christoph Lameter9eeff232006-01-18 17:42:31 -08003554 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003555 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003556 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003557 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003558}
Andrew Morton179e9632006-03-22 00:08:18 -08003559
3560int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3561{
Andrew Morton179e9632006-03-22 00:08:18 -08003562 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003563 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003564
3565 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003566 * Zone reclaim reclaims unmapped file backed pages and
3567 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003568 *
Christoph Lameter96146342006-07-03 00:24:13 -07003569 * A small portion of unmapped file backed pages is needed for
3570 * file I/O otherwise pages read by file I/O will be immediately
3571 * thrown out if the zone is overallocated. So we do not reclaim
3572 * if less than a specified percentage of the zone is used by
3573 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003574 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003575 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3576 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003577 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003578
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08003579 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003580 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003581
Andrew Morton179e9632006-03-22 00:08:18 -08003582 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003583 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003584 */
David Rientjesd773ed62007-10-16 23:26:01 -07003585 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003586 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003587
3588 /*
3589 * Only run zone reclaim on the local zone or on zones that do not
3590 * have associated processors. This will favor the local processor
3591 * over remote processors and spread off node memory allocations
3592 * as wide as possible.
3593 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003594 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003595 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003596 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003597
3598 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003599 return ZONE_RECLAIM_NOSCAN;
3600
David Rientjesd773ed62007-10-16 23:26:01 -07003601 ret = __zone_reclaim(zone, gfp_mask, order);
3602 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3603
Mel Gorman24cf725182009-06-16 15:33:23 -07003604 if (!ret)
3605 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3606
David Rientjesd773ed62007-10-16 23:26:01 -07003607 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003608}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003609#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003610
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003611/*
3612 * page_evictable - test whether a page is evictable
3613 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003614 *
3615 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003616 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003617 *
3618 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003619 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003620 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003621 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003622 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003623int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003624{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003625 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003626}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003627
Hugh Dickins850465792012-01-20 14:34:19 -08003628#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003629/**
Hugh Dickins24513262012-01-20 14:34:21 -08003630 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3631 * @pages: array of pages to check
3632 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003633 *
Hugh Dickins24513262012-01-20 14:34:21 -08003634 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003635 *
3636 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003637 */
Hugh Dickins24513262012-01-20 14:34:21 -08003638void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003639{
Johannes Weiner925b7672012-01-12 17:18:15 -08003640 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003641 struct zone *zone = NULL;
3642 int pgscanned = 0;
3643 int pgrescued = 0;
3644 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003645
Hugh Dickins24513262012-01-20 14:34:21 -08003646 for (i = 0; i < nr_pages; i++) {
3647 struct page *page = pages[i];
3648 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003649
Hugh Dickins24513262012-01-20 14:34:21 -08003650 pgscanned++;
3651 pagezone = page_zone(page);
3652 if (pagezone != zone) {
3653 if (zone)
3654 spin_unlock_irq(&zone->lru_lock);
3655 zone = pagezone;
3656 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003657 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003658 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003659
Hugh Dickins24513262012-01-20 14:34:21 -08003660 if (!PageLRU(page) || !PageUnevictable(page))
3661 continue;
3662
Hugh Dickins39b5f292012-10-08 16:33:18 -07003663 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003664 enum lru_list lru = page_lru_base_type(page);
3665
3666 VM_BUG_ON(PageActive(page));
3667 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003668 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3669 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003670 pgrescued++;
3671 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003672 }
Hugh Dickins24513262012-01-20 14:34:21 -08003673
3674 if (zone) {
3675 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3676 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3677 spin_unlock_irq(&zone->lru_lock);
3678 }
Hugh Dickins850465792012-01-20 14:34:19 -08003679}
3680#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003681
Johannes Weiner264e56d2011-10-31 17:09:13 -07003682static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003683{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003684 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003685 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003686 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003687 "one, please send an email to linux-mm@kvack.org.\n",
3688 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003689}
3690
3691/*
3692 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3693 * all nodes' unevictable lists for evictable pages
3694 */
3695unsigned long scan_unevictable_pages;
3696
3697int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003698 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003699 size_t *length, loff_t *ppos)
3700{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003701 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003702 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003703 scan_unevictable_pages = 0;
3704 return 0;
3705}
3706
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003707#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003708/*
3709 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3710 * a specified node's per zone unevictable lists for evictable pages.
3711 */
3712
Kay Sievers10fbcf42011-12-21 14:48:43 -08003713static ssize_t read_scan_unevictable_node(struct device *dev,
3714 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003715 char *buf)
3716{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003717 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003718 return sprintf(buf, "0\n"); /* always zero; should fit... */
3719}
3720
Kay Sievers10fbcf42011-12-21 14:48:43 -08003721static ssize_t write_scan_unevictable_node(struct device *dev,
3722 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003723 const char *buf, size_t count)
3724{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003725 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003726 return 1;
3727}
3728
3729
Kay Sievers10fbcf42011-12-21 14:48:43 -08003730static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003731 read_scan_unevictable_node,
3732 write_scan_unevictable_node);
3733
3734int scan_unevictable_register_node(struct node *node)
3735{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003736 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003737}
3738
3739void scan_unevictable_unregister_node(struct node *node)
3740{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003741 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003742}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003743#endif