blob: 7068e838d22b2fc55f942e2ee5899b8b5cb076f4 [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 ([email protected]).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049
50#include <asm/tlbflush.h>
51#include <asm/div64.h>
52
53#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070054#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Nick Piggin0f8053a2006-03-22 00:08:33 -080056#include "internal.h"
57
Mel Gorman33906bc2010-08-09 17:19:16 -070058#define CREATE_TRACE_POINTS
59#include <trace/events/vmscan.h>
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080062 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
64
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040066 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Johannes Weineree814fe2014-08-06 16:06:19 -070068 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070069 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080070
Johannes Weineree814fe2014-08-06 16:06:19 -070071 /*
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
73 * are scanned.
74 */
75 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070076
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070077 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080078 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
80 */
81 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080082
Johannes Weineree814fe2014-08-06 16:06:19 -070083 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
86 unsigned int may_writepage:1;
87
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
90
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
93
94 unsigned int hibernation_mode:1;
95
96 /* One of the zones is ready for compaction */
97 unsigned int compaction_ready:1;
98
99 /* Incremented by the number of inactive pages that were scanned */
100 unsigned long nr_scanned;
101
102 /* Number of pages freed so far during a call to shrink_zones() */
103 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104};
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
107
108#ifdef ARCH_HAS_PREFETCH
109#define prefetch_prev_lru_page(_page, _base, _field) \
110 do { \
111 if ((_page)->lru.prev != _base) { \
112 struct page *prev; \
113 \
114 prev = lru_to_page(&(_page->lru)); \
115 prefetch(&prev->_field); \
116 } \
117 } while (0)
118#else
119#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
120#endif
121
122#ifdef ARCH_HAS_PREFETCHW
123#define prefetchw_prev_lru_page(_page, _base, _field) \
124 do { \
125 if ((_page)->lru.prev != _base) { \
126 struct page *prev; \
127 \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetchw(&prev->_field); \
130 } \
131 } while (0)
132#else
133#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
134#endif
135
136/*
137 * From 0 .. 100. Higher means more swappy.
138 */
139int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700140/*
141 * The total number of pages which are beyond the high watermark within all
142 * zones.
143 */
144unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146static LIST_HEAD(shrinker_list);
147static DECLARE_RWSEM(shrinker_rwsem);
148
Andrew Mortonc255a452012-07-31 16:43:02 -0700149#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800150static bool global_reclaim(struct scan_control *sc)
151{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800152 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800153}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800154#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155static bool global_reclaim(struct scan_control *sc)
156{
157 return true;
158}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800159#endif
160
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800161static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700162{
163 int nr;
164
165 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
166 zone_page_state(zone, NR_INACTIVE_FILE);
167
168 if (get_nr_swap_pages() > 0)
169 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
170 zone_page_state(zone, NR_INACTIVE_ANON);
171
172 return nr;
173}
174
175bool zone_reclaimable(struct zone *zone)
176{
Mel Gorman0d5d8232014-08-06 16:07:16 -0700177 return zone_page_state(zone, NR_PAGES_SCANNED) <
178 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700179}
180
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700181static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800182{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700183 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700184 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800185
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700186 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800187}
188
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000190 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000192int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000194 size_t size = sizeof(*shrinker->nr_deferred);
195
196 /*
197 * If we only have one possible node in the system anyway, save
198 * ourselves the trouble and disable NUMA aware behavior. This way we
199 * will save memory and some small loop time later.
200 */
201 if (nr_node_ids == 1)
202 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
203
204 if (shrinker->flags & SHRINKER_NUMA_AWARE)
205 size *= nr_node_ids;
206
207 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
208 if (!shrinker->nr_deferred)
209 return -ENOMEM;
210
Rusty Russell8e1f9362007-07-17 04:03:17 -0700211 down_write(&shrinker_rwsem);
212 list_add_tail(&shrinker->list, &shrinker_list);
213 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000214 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218/*
219 * Remove one
220 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700221void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222{
223 down_write(&shrinker_rwsem);
224 list_del(&shrinker->list);
225 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700226 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700228EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229
230#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000231
232static unsigned long
233shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
234 unsigned long nr_pages_scanned, unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000236 unsigned long freed = 0;
237 unsigned long long delta;
238 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700239 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000240 long nr;
241 long new_nr;
242 int nid = shrinkctl->nid;
243 long batch_size = shrinker->batch ? shrinker->batch
244 : SHRINK_BATCH;
245
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700246 freeable = shrinker->count_objects(shrinker, shrinkctl);
247 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000248 return 0;
249
250 /*
251 * copy the current shrinker scan count into a local variable
252 * and zero it so that other concurrent shrinker invocations
253 * don't also do this scanning work.
254 */
255 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
256
257 total_scan = nr;
258 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700259 delta *= freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000260 do_div(delta, lru_pages + 1);
261 total_scan += delta;
262 if (total_scan < 0) {
263 printk(KERN_ERR
264 "shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000265 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700266 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000267 }
268
269 /*
270 * We need to avoid excessive windup on filesystem shrinkers
271 * due to large numbers of GFP_NOFS allocations causing the
272 * shrinkers to return -1 all the time. This results in a large
273 * nr being built up so when a shrink that can do some work
274 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700275 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000276 * memory.
277 *
278 * Hence only allow the shrinker to scan the entire cache when
279 * a large delta change is calculated directly.
280 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700281 if (delta < freeable / 4)
282 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000283
284 /*
285 * Avoid risking looping forever due to too large nr value:
286 * never try to free more than twice the estimate number of
287 * freeable entries.
288 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700289 if (total_scan > freeable * 2)
290 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000291
292 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
293 nr_pages_scanned, lru_pages,
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700294 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000295
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800296 /*
297 * Normally, we should not scan less than batch_size objects in one
298 * pass to avoid too frequent shrinker calls, but if the slab has less
299 * than batch_size objects in total and we are really tight on memory,
300 * we will try to reclaim all available objects, otherwise we can end
301 * up failing allocations although there are plenty of reclaimable
302 * objects spread over several slabs with usage less than the
303 * batch_size.
304 *
305 * We detect the "tight on memory" situations by looking at the total
306 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700307 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800308 * scanning at high prio and therefore should try to reclaim as much as
309 * possible.
310 */
311 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700312 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000313 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800314 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000315
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800316 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000317 ret = shrinker->scan_objects(shrinker, shrinkctl);
318 if (ret == SHRINK_STOP)
319 break;
320 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000321
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800322 count_vm_events(SLABS_SCANNED, nr_to_scan);
323 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000324
325 cond_resched();
326 }
327
328 /*
329 * move the unused scan count back into the shrinker in a
330 * manner that handles concurrent updates. If we exhausted the
331 * scan, there is no need to do an update.
332 */
333 if (total_scan > 0)
334 new_nr = atomic_long_add_return(total_scan,
335 &shrinker->nr_deferred[nid]);
336 else
337 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
338
Dave Hansendf9024a2014-06-04 16:08:07 -0700339 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000340 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343/*
344 * Call the shrink functions to age shrinkable caches
345 *
346 * Here we assume it costs one seek to replace a lru page and that it also
347 * takes a seek to recreate a cache object. With this in mind we age equal
348 * percentages of the lru and ageable caches. This should balance the seeks
349 * generated by these structures.
350 *
Simon Arlott183ff222007-10-20 01:27:18 +0200351 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 * slab to avoid swapping.
353 *
354 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
355 *
356 * `lru_pages' represents the number of on-LRU pages in all the zones which
357 * are eligible for the caller's allocation attempt. It is used for balancing
358 * slab reclaim versus page reclaim.
[email protected]b15e0902005-06-21 17:14:35 -0700359 *
360 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 */
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000362unsigned long shrink_slab(struct shrink_control *shrinkctl,
Ying Han1495f232011-05-24 17:12:27 -0700363 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700364 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365{
366 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000367 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368
Ying Han1495f232011-05-24 17:12:27 -0700369 if (nr_pages_scanned == 0)
370 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
Minchan Kimf06590b2011-05-24 17:11:11 -0700372 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000373 /*
374 * If we would return 0, our callers would understand that we
375 * have nothing else to shrink and give up trying. By returning
376 * 1 we keep it going and assume we'll be able to shrink next
377 * time.
378 */
379 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700380 goto out;
381 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
383 list_for_each_entry(shrinker, &shrinker_list, list) {
Vladimir Davydovec970972014-01-23 15:53:23 -0800384 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
385 shrinkctl->nid = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000386 freed += shrink_slab_node(shrinkctl, shrinker,
Vladimir Davydovec970972014-01-23 15:53:23 -0800387 nr_pages_scanned, lru_pages);
388 continue;
389 }
390
391 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
392 if (node_online(shrinkctl->nid))
393 freed += shrink_slab_node(shrinkctl, shrinker,
394 nr_pages_scanned, lru_pages);
Dave Chinneracf92b42011-07-08 14:14:35 +1000395
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 }
398 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700399out:
400 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000401 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402}
403
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404static inline int is_page_cache_freeable(struct page *page)
405{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700406 /*
407 * A freeable page cache page is referenced only by the caller
408 * that isolated the page, the page cache radix tree and
409 * optional buffer heads at page->private.
410 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700411 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412}
413
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700414static int may_write_to_queue(struct backing_dev_info *bdi,
415 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
Christoph Lameter930d9152006-01-08 01:00:47 -0800417 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 return 1;
419 if (!bdi_write_congested(bdi))
420 return 1;
421 if (bdi == current->backing_dev_info)
422 return 1;
423 return 0;
424}
425
426/*
427 * We detected a synchronous write error writing a page out. Probably
428 * -ENOSPC. We need to propagate that into the address_space for a subsequent
429 * fsync(), msync() or close().
430 *
431 * The tricky part is that after writepage we cannot touch the mapping: nothing
432 * prevents it from being freed up. But we have a ref on the page and once
433 * that page is locked, the mapping is pinned.
434 *
435 * We're allowed to run sleeping lock_page() here because we know the caller has
436 * __GFP_FS.
437 */
438static void handle_write_error(struct address_space *mapping,
439 struct page *page, int error)
440{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100441 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700442 if (page_mapping(page) == mapping)
443 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 unlock_page(page);
445}
446
Christoph Lameter04e62a22006-06-23 02:03:38 -0700447/* possible outcome of pageout() */
448typedef enum {
449 /* failed to write page out, page is locked */
450 PAGE_KEEP,
451 /* move page to the active list, page is locked */
452 PAGE_ACTIVATE,
453 /* page has been sent to the disk successfully, page is unlocked */
454 PAGE_SUCCESS,
455 /* page is clean and locked */
456 PAGE_CLEAN,
457} pageout_t;
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/*
Andrew Morton1742f192006-03-22 00:08:21 -0800460 * pageout is called by shrink_page_list() for each dirty page.
461 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700463static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700464 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465{
466 /*
467 * If the page is dirty, only perform writeback if that write
468 * will be non-blocking. To prevent this allocation from being
469 * stalled by pagecache activity. But note that there may be
470 * stalls if we need to run get_block(). We could test
471 * PagePrivate for that.
472 *
Al Viro81742022014-04-03 03:17:43 -0400473 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 * this page's queue, we can perform writeback even if that
475 * will block.
476 *
477 * If the page is swapcache, write it back even if that would
478 * block, for some throttling. This happens by accident, because
479 * swap_backing_dev_info is bust: it doesn't reflect the
480 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 */
482 if (!is_page_cache_freeable(page))
483 return PAGE_KEEP;
484 if (!mapping) {
485 /*
486 * Some data journaling orphaned pages can have
487 * page->mapping == NULL while being dirty with clean buffers.
488 */
David Howells266cf652009-04-03 16:42:36 +0100489 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 if (try_to_free_buffers(page)) {
491 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700492 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 return PAGE_CLEAN;
494 }
495 }
496 return PAGE_KEEP;
497 }
498 if (mapping->a_ops->writepage == NULL)
499 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700500 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 return PAGE_KEEP;
502
503 if (clear_page_dirty_for_io(page)) {
504 int res;
505 struct writeback_control wbc = {
506 .sync_mode = WB_SYNC_NONE,
507 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700508 .range_start = 0,
509 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 .for_reclaim = 1,
511 };
512
513 SetPageReclaim(page);
514 res = mapping->a_ops->writepage(page, &wbc);
515 if (res < 0)
516 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800517 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 ClearPageReclaim(page);
519 return PAGE_ACTIVATE;
520 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 if (!PageWriteback(page)) {
523 /* synchronous write or broken a_ops? */
524 ClearPageReclaim(page);
525 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700526 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700527 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return PAGE_SUCCESS;
529 }
530
531 return PAGE_CLEAN;
532}
533
Andrew Mortona649fd92006-10-17 00:09:36 -0700534/*
Nick Piggine2867812008-07-25 19:45:30 -0700535 * Same as remove_mapping, but if the page is removed from the mapping, it
536 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700537 */
Johannes Weinera5289102014-04-03 14:47:51 -0700538static int __remove_mapping(struct address_space *mapping, struct page *page,
539 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800540{
Nick Piggin28e4d962006-09-25 23:31:23 -0700541 BUG_ON(!PageLocked(page));
542 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800543
Nick Piggin19fd6232008-07-25 19:45:32 -0700544 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800545 /*
Nick Piggin0fd0e6b02006-09-27 01:50:02 -0700546 * The non racy check for a busy page.
547 *
548 * Must be careful with the order of the tests. When someone has
549 * a ref to the page, it may be possible that they dirty it then
550 * drop the reference. So if PageDirty is tested before page_count
551 * here, then the following race may occur:
552 *
553 * get_user_pages(&page);
554 * [user mapping goes away]
555 * write_to(page);
556 * !PageDirty(page) [good]
557 * SetPageDirty(page);
558 * put_page(page);
559 * !page_count(page) [good, discard it]
560 *
561 * [oops, our write_to data is lost]
562 *
563 * Reversing the order of the tests ensures such a situation cannot
564 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
565 * load is not satisfied before that of page->_count.
566 *
567 * Note that if SetPageDirty is always performed via set_page_dirty,
568 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800569 */
Nick Piggine2867812008-07-25 19:45:30 -0700570 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800571 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700572 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
573 if (unlikely(PageDirty(page))) {
574 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800575 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700576 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800577
578 if (PageSwapCache(page)) {
579 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700580 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800581 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700582 spin_unlock_irq(&mapping->tree_lock);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700583 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700584 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500585 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700586 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500587
588 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700589 /*
590 * Remember a shadow entry for reclaimed file cache in
591 * order to detect refaults, thus thrashing, later on.
592 *
593 * But don't store shadows in an address space that is
594 * already exiting. This is not just an optizimation,
595 * inode reclaim needs to empty out the radix tree or
596 * the nodes are lost. Don't plant shadows behind its
597 * back.
598 */
599 if (reclaimed && page_is_file_cache(page) &&
600 !mapping_exiting(mapping))
601 shadow = workingset_eviction(mapping, page);
602 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700603 spin_unlock_irq(&mapping->tree_lock);
Linus Torvalds6072d132010-12-01 13:35:19 -0500604
605 if (freepage != NULL)
606 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800607 }
608
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800609 return 1;
610
611cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700612 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800613 return 0;
614}
615
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616/*
Nick Piggine2867812008-07-25 19:45:30 -0700617 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
618 * someone else has a ref on the page, abort and return 0. If it was
619 * successfully detached, return 1. Assumes the caller has a single ref on
620 * this page.
621 */
622int remove_mapping(struct address_space *mapping, struct page *page)
623{
Johannes Weinera5289102014-04-03 14:47:51 -0700624 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700625 /*
626 * Unfreezing the refcount with 1 rather than 2 effectively
627 * drops the pagecache ref for us without requiring another
628 * atomic operation.
629 */
630 page_unfreeze_refs(page, 1);
631 return 1;
632 }
633 return 0;
634}
635
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700636/**
637 * putback_lru_page - put previously isolated page onto appropriate LRU list
638 * @page: page to be put back to appropriate lru list
639 *
640 * Add previously isolated @page to appropriate LRU list.
641 * Page may still be unevictable for other reasons.
642 *
643 * lru_lock must not be held, interrupts must be enabled.
644 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700645void putback_lru_page(struct page *page)
646{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700647 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700648 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700649
Sasha Levin309381fea2014-01-23 15:52:54 -0800650 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700651
652redo:
653 ClearPageUnevictable(page);
654
Hugh Dickins39b5f292012-10-08 16:33:18 -0700655 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700656 /*
657 * For evictable pages, we can use the cache.
658 * In event of a race, worst case is we end up with an
659 * unevictable page on [in]active list.
660 * We know how to handle that.
661 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700662 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700663 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700664 } else {
665 /*
666 * Put unevictable pages directly on zone's unevictable
667 * list.
668 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700669 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700670 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700671 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700672 * When racing with an mlock or AS_UNEVICTABLE clearing
673 * (page is unlocked) make sure that if the other thread
674 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800675 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700676 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700677 * the page back to the evictable list.
678 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700679 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700680 */
681 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700682 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700683
684 /*
685 * page's status can change while we move it among lru. If an evictable
686 * page is on unevictable list, it never be freed. To avoid that,
687 * check after we added it to the list, again.
688 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700689 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700690 if (!isolate_lru_page(page)) {
691 put_page(page);
692 goto redo;
693 }
694 /* This means someone else dropped this page from LRU
695 * So, it will be freed or putback to LRU again. There is
696 * nothing to do here.
697 */
698 }
699
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700700 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700701 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700702 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700703 count_vm_event(UNEVICTABLE_PGCULLED);
704
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700705 put_page(page); /* drop ref from isolate */
706}
707
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800708enum page_references {
709 PAGEREF_RECLAIM,
710 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800711 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800712 PAGEREF_ACTIVATE,
713};
714
715static enum page_references page_check_references(struct page *page,
716 struct scan_control *sc)
717{
Johannes Weiner645747462010-03-05 13:42:22 -0800718 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800719 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800720
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700721 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
722 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800723 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800725 /*
726 * Mlock lost the isolation race with us. Let try_to_unmap()
727 * move the page to the unevictable list.
728 */
729 if (vm_flags & VM_LOCKED)
730 return PAGEREF_RECLAIM;
731
Johannes Weiner645747462010-03-05 13:42:22 -0800732 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700733 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800734 return PAGEREF_ACTIVATE;
735 /*
736 * All mapped pages start out with page table
737 * references from the instantiating fault, so we need
738 * to look twice if a mapped file page is used more
739 * than once.
740 *
741 * Mark it and spare it for another trip around the
742 * inactive list. Another page table reference will
743 * lead to its activation.
744 *
745 * Note: the mark is set for activated pages as well
746 * so that recently deactivated but used pages are
747 * quickly recovered.
748 */
749 SetPageReferenced(page);
750
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800751 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800752 return PAGEREF_ACTIVATE;
753
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800754 /*
755 * Activate file-backed executable pages after first usage.
756 */
757 if (vm_flags & VM_EXEC)
758 return PAGEREF_ACTIVATE;
759
Johannes Weiner645747462010-03-05 13:42:22 -0800760 return PAGEREF_KEEP;
761 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800762
763 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700764 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800765 return PAGEREF_RECLAIM_CLEAN;
766
767 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800768}
769
Mel Gormane2be15f2013-07-03 15:01:57 -0700770/* Check if a page is dirty or under writeback */
771static void page_check_dirty_writeback(struct page *page,
772 bool *dirty, bool *writeback)
773{
Mel Gormanb4597222013-07-03 15:02:05 -0700774 struct address_space *mapping;
775
Mel Gormane2be15f2013-07-03 15:01:57 -0700776 /*
777 * Anonymous pages are not handled by flushers and must be written
778 * from reclaim context. Do not stall reclaim based on them
779 */
780 if (!page_is_file_cache(page)) {
781 *dirty = false;
782 *writeback = false;
783 return;
784 }
785
786 /* By default assume that the page flags are accurate */
787 *dirty = PageDirty(page);
788 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700789
790 /* Verify dirty/writeback state if the filesystem supports it */
791 if (!page_has_private(page))
792 return;
793
794 mapping = page_mapping(page);
795 if (mapping && mapping->a_ops->is_dirty_writeback)
796 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700797}
798
Nick Piggine2867812008-07-25 19:45:30 -0700799/*
Andrew Morton1742f192006-03-22 00:08:21 -0800800 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 */
Andrew Morton1742f192006-03-22 00:08:21 -0800802static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700803 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700804 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700805 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700806 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700807 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700808 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700809 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700810 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700811 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812{
813 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700814 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700816 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700817 unsigned long nr_dirty = 0;
818 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800819 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700820 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700821 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
823 cond_resched();
824
Tim Chen69980e32012-07-31 16:46:08 -0700825 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 while (!list_empty(page_list)) {
827 struct address_space *mapping;
828 struct page *page;
829 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700830 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700831 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
833 cond_resched();
834
835 page = lru_to_page(page_list);
836 list_del(&page->lru);
837
Nick Piggin529ae9a2008-08-02 12:01:03 +0200838 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 goto keep;
840
Sasha Levin309381fea2014-01-23 15:52:54 -0800841 VM_BUG_ON_PAGE(PageActive(page), page);
842 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
844 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800845
Hugh Dickins39b5f292012-10-08 16:33:18 -0700846 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700847 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700848
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700849 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800850 goto keep_locked;
851
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 /* Double the slab pressure for mapped and swapcache pages */
853 if (page_mapped(page) || PageSwapCache(page))
854 sc->nr_scanned++;
855
Andy Whitcroftc661b072007-08-22 14:01:26 -0700856 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
857 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
858
Mel Gorman283aba92013-07-03 15:01:51 -0700859 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700860 * The number of dirty pages determines if a zone is marked
861 * reclaim_congested which affects wait_iff_congested. kswapd
862 * will stall and start writing pages if the tail of the LRU
863 * is all dirty unqueued pages.
864 */
865 page_check_dirty_writeback(page, &dirty, &writeback);
866 if (dirty || writeback)
867 nr_dirty++;
868
869 if (dirty && !writeback)
870 nr_unqueued_dirty++;
871
Mel Gormand04e8ac2013-07-03 15:02:03 -0700872 /*
873 * Treat this page as congested if the underlying BDI is or if
874 * pages are cycling through the LRU so quickly that the
875 * pages marked for immediate reclaim are making it to the
876 * end of the LRU a second time.
877 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700878 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700879 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
880 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700881 nr_congested++;
882
883 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700884 * If a page at the tail of the LRU is under writeback, there
885 * are three cases to consider.
886 *
887 * 1) If reclaim is encountering an excessive number of pages
888 * under writeback and this page is both under writeback and
889 * PageReclaim then it indicates that pages are being queued
890 * for IO but are being recycled through the LRU before the
891 * IO can complete. Waiting on the page itself risks an
892 * indefinite stall if it is impossible to writeback the
893 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700894 * note that the LRU is being scanned too quickly and the
895 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700896 *
897 * 2) Global reclaim encounters a page, memcg encounters a
898 * page that is not marked for immediate reclaim or
899 * the caller does not have __GFP_IO. In this case mark
900 * the page for immediate reclaim and continue scanning.
901 *
902 * __GFP_IO is checked because a loop driver thread might
903 * enter reclaim, and deadlock if it waits on a page for
904 * which it is needed to do the write (loop masks off
905 * __GFP_IO|__GFP_FS for this reason); but more thought
906 * would probably show more reasons.
907 *
908 * Don't require __GFP_FS, since we're not going into the
909 * FS, just waiting on its writeback completion. Worryingly,
910 * ext4 gfs2 and xfs allocate pages with
911 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
912 * may_enter_fs here is liable to OOM on them.
913 *
914 * 3) memcg encounters a page that is not already marked
915 * PageReclaim. memcg does not have any dirty pages
916 * throttling so we could easily OOM just because too many
917 * pages are in writeback and there is nothing else to
918 * reclaim. Wait for the writeback to complete.
919 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700920 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700921 /* Case 1 above */
922 if (current_is_kswapd() &&
923 PageReclaim(page) &&
924 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700925 nr_immediate++;
926 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700927
928 /* Case 2 above */
929 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700930 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
931 /*
932 * This is slightly racy - end_page_writeback()
933 * might have just cleared PageReclaim, then
934 * setting PageReclaim here end up interpreted
935 * as PageReadahead - but that does not matter
936 * enough to care. What we do want is for this
937 * page to have PageReclaim set next time memcg
938 * reclaim reaches the tests above, so it will
939 * then wait_on_page_writeback() to avoid OOM;
940 * and it's also appropriate in global reclaim.
941 */
942 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700943 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700944
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700945 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700946
947 /* Case 3 above */
948 } else {
949 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700950 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700951 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
Minchan Kim02c6de82012-10-08 16:31:55 -0700953 if (!force_reclaim)
954 references = page_check_references(page, sc);
955
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800956 switch (references) {
957 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800959 case PAGEREF_KEEP:
960 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800961 case PAGEREF_RECLAIM:
962 case PAGEREF_RECLAIM_CLEAN:
963 ; /* try to reclaim the page below */
964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 /*
967 * Anonymous process memory has backing store?
968 * Try to allocate it some swap space here.
969 */
Nick Pigginb291f002008-10-18 20:26:44 -0700970 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b932008-11-19 15:36:37 -0800971 if (!(sc->gfp_mask & __GFP_IO))
972 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700973 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 goto activate_locked;
Hugh Dickins63eb6b932008-11-19 15:36:37 -0800975 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976
Mel Gormane2be15f2013-07-03 15:01:57 -0700977 /* Adding to swap updated mapping */
978 mapping = page_mapping(page);
979 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
981 /*
982 * The page is mapped into the page tables of one or more
983 * processes. Try to unmap it here.
984 */
985 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700986 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 case SWAP_FAIL:
988 goto activate_locked;
989 case SWAP_AGAIN:
990 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700991 case SWAP_MLOCK:
992 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 case SWAP_SUCCESS:
994 ; /* try to free the page below */
995 }
996 }
997
998 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700999 /*
1000 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001001 * avoid risk of stack overflow but only writeback
1002 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001003 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001004 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001005 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -07001006 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001007 /*
1008 * Immediately reclaim when written back.
1009 * Similar in principal to deactivate_page()
1010 * except we already have the page isolated
1011 * and know it's dirty
1012 */
1013 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1014 SetPageReclaim(page);
1015
Mel Gormanee728862011-10-31 17:07:38 -07001016 goto keep_locked;
1017 }
1018
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001019 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001021 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001023 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 goto keep_locked;
1025
1026 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001027 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 case PAGE_KEEP:
1029 goto keep_locked;
1030 case PAGE_ACTIVATE:
1031 goto activate_locked;
1032 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001033 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001034 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001035 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001037
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 /*
1039 * A synchronous write - probably a ramdisk. Go
1040 * ahead and try to reclaim the page.
1041 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001042 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 goto keep;
1044 if (PageDirty(page) || PageWriteback(page))
1045 goto keep_locked;
1046 mapping = page_mapping(page);
1047 case PAGE_CLEAN:
1048 ; /* try to free the page below */
1049 }
1050 }
1051
1052 /*
1053 * If the page has buffers, try to free the buffer mappings
1054 * associated with this page. If we succeed we try to free
1055 * the page as well.
1056 *
1057 * We do this even if the page is PageDirty().
1058 * try_to_release_page() does not perform I/O, but it is
1059 * possible for a page to have PageDirty set, but it is actually
1060 * clean (all its buffers are clean). This happens if the
1061 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001062 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 * try_to_release_page() will discover that cleanness and will
1064 * drop the buffers and mark the page clean - it can be freed.
1065 *
1066 * Rarely, pages can have buffers and no ->mapping. These are
1067 * the pages which were not successfully invalidated in
1068 * truncate_complete_page(). We try to drop those buffers here
1069 * and if that worked, and the page is no longer mapped into
1070 * process address space (page_count == 1) it can be freed.
1071 * Otherwise, leave the page on the LRU so it is swappable.
1072 */
David Howells266cf652009-04-03 16:42:36 +01001073 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 if (!try_to_release_page(page, sc->gfp_mask))
1075 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001076 if (!mapping && page_count(page) == 1) {
1077 unlock_page(page);
1078 if (put_page_testzero(page))
1079 goto free_it;
1080 else {
1081 /*
1082 * rare race with speculative reference.
1083 * the speculative reference will free
1084 * this page shortly, so we may
1085 * increment nr_reclaimed here (and
1086 * leave it off the LRU).
1087 */
1088 nr_reclaimed++;
1089 continue;
1090 }
1091 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 }
1093
Johannes Weinera5289102014-04-03 14:47:51 -07001094 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001095 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096
Nick Piggina978d6f2008-10-18 20:26:58 -07001097 /*
1098 * At this point, we have no other references and there is
1099 * no way to pick any more up (removed from LRU, removed
1100 * from pagecache). Can use non-atomic bitops now (and
1101 * we obviously don't have to worry about waking up a process
1102 * waiting on the page lock, because there are no references.
1103 */
1104 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001105free_it:
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001106 mem_cgroup_uncharge(page);
Andrew Morton05ff5132006-03-22 00:08:20 -08001107 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001108
1109 /*
1110 * Is there need to periodically free_page_list? It would
1111 * appear not as the counts should be low
1112 */
1113 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 continue;
1115
Nick Pigginb291f002008-10-18 20:26:44 -07001116cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001117 if (PageSwapCache(page))
1118 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001119 unlock_page(page);
1120 putback_lru_page(page);
1121 continue;
1122
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001124 /* Not a candidate for swapping, so reclaim swap space. */
1125 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001126 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001127 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 SetPageActive(page);
1129 pgactivate++;
1130keep_locked:
1131 unlock_page(page);
1132keep:
1133 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001134 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 }
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001136 mem_cgroup_uncharge_end();
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001137
Mel Gormanb745bc82014-06-04 16:10:22 -07001138 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001139
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001141 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001142
Mel Gorman8e950282013-07-03 15:02:02 -07001143 *ret_nr_dirty += nr_dirty;
1144 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001145 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001146 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001147 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001148 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149}
1150
Minchan Kim02c6de82012-10-08 16:31:55 -07001151unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1152 struct list_head *page_list)
1153{
1154 struct scan_control sc = {
1155 .gfp_mask = GFP_KERNEL,
1156 .priority = DEF_PRIORITY,
1157 .may_unmap = 1,
1158 };
Mel Gorman8e950282013-07-03 15:02:02 -07001159 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001160 struct page *page, *next;
1161 LIST_HEAD(clean_pages);
1162
1163 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001164 if (page_is_file_cache(page) && !PageDirty(page) &&
1165 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001166 ClearPageActive(page);
1167 list_move(&page->lru, &clean_pages);
1168 }
1169 }
1170
1171 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001172 TTU_UNMAP|TTU_IGNORE_ACCESS,
1173 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001174 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001175 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001176 return ret;
1177}
1178
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001179/*
1180 * Attempt to remove the specified page from its LRU. Only take this page
1181 * if it is of the appropriate PageActive status. Pages which are being
1182 * freed elsewhere are also ignored.
1183 *
1184 * page: page to consider
1185 * mode: one of the LRU isolation modes defined above
1186 *
1187 * returns 0 on success, -ve errno on failure.
1188 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001189int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001190{
1191 int ret = -EINVAL;
1192
1193 /* Only take pages on the LRU. */
1194 if (!PageLRU(page))
1195 return ret;
1196
Minchan Kime46a2872012-10-08 16:33:48 -07001197 /* Compaction should not handle unevictable pages but CMA can do so */
1198 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001199 return ret;
1200
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001201 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001202
Mel Gormanc8244932012-01-12 17:19:38 -08001203 /*
1204 * To minimise LRU disruption, the caller can indicate that it only
1205 * wants to isolate pages it will be able to operate on without
1206 * blocking - clean pages for the most part.
1207 *
1208 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1209 * is used by reclaim when it is cannot write to backing storage
1210 *
1211 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1212 * that it is possible to migrate without blocking
1213 */
1214 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1215 /* All the caller can do on PageWriteback is block */
1216 if (PageWriteback(page))
1217 return ret;
1218
1219 if (PageDirty(page)) {
1220 struct address_space *mapping;
1221
1222 /* ISOLATE_CLEAN means only clean pages */
1223 if (mode & ISOLATE_CLEAN)
1224 return ret;
1225
1226 /*
1227 * Only pages without mappings or that have a
1228 * ->migratepage callback are possible to migrate
1229 * without blocking
1230 */
1231 mapping = page_mapping(page);
1232 if (mapping && !mapping->a_ops->migratepage)
1233 return ret;
1234 }
1235 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001236
Minchan Kimf80c0672011-10-31 17:06:55 -07001237 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1238 return ret;
1239
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001240 if (likely(get_page_unless_zero(page))) {
1241 /*
1242 * Be careful not to clear PageLRU until after we're
1243 * sure the page is not being freed elsewhere -- the
1244 * page release code relies on it.
1245 */
1246 ClearPageLRU(page);
1247 ret = 0;
1248 }
1249
1250 return ret;
1251}
1252
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001253/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 * zone->lru_lock is heavily contended. Some of the functions that
1255 * shrink the lists perform better by taking out a batch of pages
1256 * and working on them outside the LRU lock.
1257 *
1258 * For pagecache intensive workloads, this function is the hottest
1259 * spot in the kernel (apart from copy_*_user functions).
1260 *
1261 * Appropriate locks must be held before calling this function.
1262 *
1263 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001264 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 * @dst: The temp list to put pages on to.
Hugh Dickinsf626012d2012-01-12 17:20:06 -08001266 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001267 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001268 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001269 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 *
1271 * returns how many pages were moved onto *@dst.
1272 */
Andrew Morton69e05942006-03-22 00:08:19 -08001273static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001274 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001275 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001276 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001278 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001279 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001280 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001282 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001283 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001284 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001285
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 page = lru_to_page(src);
1287 prefetchw_prev_lru_page(page, src, flags);
1288
Sasha Levin309381fea2014-01-23 15:52:54 -08001289 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001290
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001291 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001292 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001293 nr_pages = hpage_nr_pages(page);
1294 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001295 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001296 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001297 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001298
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001299 case -EBUSY:
1300 /* else it is being freed elsewhere */
1301 list_move(&page->lru, src);
1302 continue;
1303
1304 default:
1305 BUG();
1306 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 }
1308
Hugh Dickinsf626012d2012-01-12 17:20:06 -08001309 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001310 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1311 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 return nr_taken;
1313}
1314
Nick Piggin62695a82008-10-18 20:26:09 -07001315/**
1316 * isolate_lru_page - tries to isolate a page from its LRU list
1317 * @page: page to isolate from its LRU list
1318 *
1319 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1320 * vmstat statistic corresponding to whatever LRU list the page was on.
1321 *
1322 * Returns 0 if the page was removed from an LRU list.
1323 * Returns -EBUSY if the page was not on an LRU list.
1324 *
1325 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001326 * the active list, it will have PageActive set. If it was found on
1327 * the unevictable list, it will have the PageUnevictable bit set. That flag
1328 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001329 *
1330 * The vmstat statistic corresponding to the list on which the page was
1331 * found will be decremented.
1332 *
1333 * Restrictions:
1334 * (1) Must be called with an elevated refcount on the page. This is a
1335 * fundamentnal difference from isolate_lru_pages (which is called
1336 * without a stable reference).
1337 * (2) the lru_lock must not be held.
1338 * (3) interrupts must be enabled.
1339 */
1340int isolate_lru_page(struct page *page)
1341{
1342 int ret = -EBUSY;
1343
Sasha Levin309381fea2014-01-23 15:52:54 -08001344 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001345
Nick Piggin62695a82008-10-18 20:26:09 -07001346 if (PageLRU(page)) {
1347 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001348 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001349
1350 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001351 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001352 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001353 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001354 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001355 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001356 del_page_from_lru_list(page, lruvec, lru);
1357 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001358 }
1359 spin_unlock_irq(&zone->lru_lock);
1360 }
1361 return ret;
1362}
1363
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001364/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001365 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1366 * then get resheduled. When there are massive number of tasks doing page
1367 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1368 * the LRU list will go small and be scanned faster than necessary, leading to
1369 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001370 */
1371static int too_many_isolated(struct zone *zone, int file,
1372 struct scan_control *sc)
1373{
1374 unsigned long inactive, isolated;
1375
1376 if (current_is_kswapd())
1377 return 0;
1378
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001379 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001380 return 0;
1381
1382 if (file) {
1383 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1384 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1385 } else {
1386 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1387 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1388 }
1389
Fengguang Wu3cf23842012-12-18 14:23:31 -08001390 /*
1391 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1392 * won't get blocked by normal direct-reclaimers, forming a circular
1393 * deadlock.
1394 */
1395 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1396 inactive >>= 3;
1397
Rik van Riel35cd7812009-09-21 17:01:38 -07001398 return isolated > inactive;
1399}
1400
Mel Gorman66635622010-08-09 17:19:30 -07001401static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001402putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001403{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001404 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1405 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001406 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001407
Mel Gorman66635622010-08-09 17:19:30 -07001408 /*
1409 * Put back any unfreeable pages.
1410 */
Mel Gorman66635622010-08-09 17:19:30 -07001411 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001412 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001413 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001414
Sasha Levin309381fea2014-01-23 15:52:54 -08001415 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001416 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001417 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001418 spin_unlock_irq(&zone->lru_lock);
1419 putback_lru_page(page);
1420 spin_lock_irq(&zone->lru_lock);
1421 continue;
1422 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001423
1424 lruvec = mem_cgroup_page_lruvec(page, zone);
1425
Linus Torvalds7a608572011-01-17 14:42:19 -08001426 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001427 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001428 add_page_to_lru_list(page, lruvec, lru);
1429
Mel Gorman66635622010-08-09 17:19:30 -07001430 if (is_active_lru(lru)) {
1431 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001432 int numpages = hpage_nr_pages(page);
1433 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001434 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001435 if (put_page_testzero(page)) {
1436 __ClearPageLRU(page);
1437 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001438 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001439
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001440 mem_cgroup_uncharge(page);
1441
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001442 if (unlikely(PageCompound(page))) {
1443 spin_unlock_irq(&zone->lru_lock);
1444 (*get_compound_page_dtor(page))(page);
1445 spin_lock_irq(&zone->lru_lock);
1446 } else
1447 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001448 }
1449 }
Mel Gorman66635622010-08-09 17:19:30 -07001450
Hugh Dickins3f797682012-01-12 17:20:07 -08001451 /*
1452 * To save our caller's stack, now use input list for pages to free.
1453 */
1454 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001455}
1456
1457/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001458 * If a kernel thread (such as nfsd for loop-back mounts) services
1459 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1460 * In that case we should only throttle if the backing device it is
1461 * writing to is congested. In other cases it is safe to throttle.
1462 */
1463static int current_may_throttle(void)
1464{
1465 return !(current->flags & PF_LESS_THROTTLE) ||
1466 current->backing_dev_info == NULL ||
1467 bdi_write_congested(current->backing_dev_info);
1468}
1469
1470/*
Andrew Morton1742f192006-03-22 00:08:21 -08001471 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1472 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 */
Mel Gorman66635622010-08-09 17:19:30 -07001474static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001475shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001476 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477{
1478 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001479 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001480 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001481 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001482 unsigned long nr_dirty = 0;
1483 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001484 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001485 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001486 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001487 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001488 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001489 struct zone *zone = lruvec_zone(lruvec);
1490 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001491
Rik van Riel35cd7812009-09-21 17:01:38 -07001492 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001493 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001494
1495 /* We are about to die and free our memory. Return now. */
1496 if (fatal_signal_pending(current))
1497 return SWAP_CLUSTER_MAX;
1498 }
1499
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001501
1502 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001503 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001504 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001505 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001506
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001509 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1510 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001511
1512 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1513 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1514
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001515 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001516 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001517 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001518 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001519 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001520 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001521 }
Hillf Dantond563c052012-03-21 16:34:02 -07001522 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001523
Hillf Dantond563c052012-03-21 16:34:02 -07001524 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001525 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001526
Minchan Kim02c6de82012-10-08 16:31:55 -07001527 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001528 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1529 &nr_writeback, &nr_immediate,
1530 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001531
Hugh Dickins3f797682012-01-12 17:20:07 -08001532 spin_lock_irq(&zone->lru_lock);
1533
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001534 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001535
Ying Han904249a2012-04-25 16:01:48 -07001536 if (global_reclaim(sc)) {
1537 if (current_is_kswapd())
1538 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1539 nr_reclaimed);
1540 else
1541 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1542 nr_reclaimed);
1543 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001544
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001545 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001546
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001547 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001548
1549 spin_unlock_irq(&zone->lru_lock);
1550
Mel Gormanb745bc82014-06-04 16:10:22 -07001551 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001552
Mel Gorman92df3a72011-10-31 17:07:56 -07001553 /*
1554 * If reclaim is isolating dirty pages under writeback, it implies
1555 * that the long-lived page allocation rate is exceeding the page
1556 * laundering rate. Either the global limits are not being effective
1557 * at throttling processes due to the page distribution throughout
1558 * zones or there is heavy usage of a slow backing device. The
1559 * only option is to throttle from reclaim context which is not ideal
1560 * as there is no guarantee the dirtying process is throttled in the
1561 * same way balance_dirty_pages() manages.
1562 *
Mel Gorman8e950282013-07-03 15:02:02 -07001563 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1564 * of pages under pages flagged for immediate reclaim and stall if any
1565 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001566 */
Mel Gorman918fc712013-07-08 16:00:25 -07001567 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001568 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001569
Mel Gormand43006d2013-07-03 15:01:50 -07001570 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001571 * memcg will stall in page writeback so only consider forcibly
1572 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001573 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001574 if (global_reclaim(sc)) {
1575 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001576 * Tag a zone as congested if all the dirty pages scanned were
1577 * backed by a congested BDI and wait_iff_congested will stall.
1578 */
1579 if (nr_dirty && nr_dirty == nr_congested)
1580 zone_set_flag(zone, ZONE_CONGESTED);
1581
1582 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001583 * If dirty pages are scanned that are not queued for IO, it
1584 * implies that flushers are not keeping up. In this case, flag
1585 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
Linus Torvaldsb738d762014-06-08 14:17:00 -07001586 * pages from reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001587 */
1588 if (nr_unqueued_dirty == nr_taken)
1589 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1590
1591 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001592 * If kswapd scans pages marked marked for immediate
1593 * reclaim and under writeback (nr_immediate), it implies
1594 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001595 * they are written so also forcibly stall.
1596 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001597 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001598 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001599 }
Mel Gormand43006d2013-07-03 15:01:50 -07001600
Mel Gorman8e950282013-07-03 15:02:02 -07001601 /*
1602 * Stall direct reclaim for IO completions if underlying BDIs or zone
1603 * is congested. Allow kswapd to continue until it starts encountering
1604 * unqueued dirty pages or cycling through the LRU too quickly.
1605 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001606 if (!sc->hibernation_mode && !current_is_kswapd() &&
1607 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001608 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1609
Mel Gormane11da5b2010-10-26 14:21:40 -07001610 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1611 zone_idx(zone),
1612 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001613 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001614 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001615 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616}
1617
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001618/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 * This moves pages from the active list to the inactive list.
1620 *
1621 * We move them the other way if the page is referenced by one or more
1622 * processes, from rmap.
1623 *
1624 * If the pages are mostly unmapped, the processing is fast and it is
1625 * appropriate to hold zone->lru_lock across the whole operation. But if
1626 * the pages are mapped, the processing is slow (page_referenced()) so we
1627 * should drop zone->lru_lock around each page. It's impossible to balance
1628 * this, so instead we remove the pages from the LRU while processing them.
1629 * It is safe to rely on PG_active against the non-LRU pages in here because
1630 * nobody will play with that bit on a non-LRU page.
1631 *
1632 * The downside is that we have to touch page->_count against each page.
1633 * But we had to alter page->flags anyway.
1634 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001635
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001636static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001637 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001638 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001639 enum lru_list lru)
1640{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001641 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001642 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001643 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001644 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001645
Wu Fengguang3eb41402009-06-16 15:33:13 -07001646 while (!list_empty(list)) {
1647 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001648 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001649
Sasha Levin309381fea2014-01-23 15:52:54 -08001650 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001651 SetPageLRU(page);
1652
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001653 nr_pages = hpage_nr_pages(page);
1654 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001655 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001656 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001657
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001658 if (put_page_testzero(page)) {
1659 __ClearPageLRU(page);
1660 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001661 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001662
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001663 mem_cgroup_uncharge(page);
1664
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001665 if (unlikely(PageCompound(page))) {
1666 spin_unlock_irq(&zone->lru_lock);
1667 (*get_compound_page_dtor(page))(page);
1668 spin_lock_irq(&zone->lru_lock);
1669 } else
1670 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001671 }
1672 }
1673 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1674 if (!is_active_lru(lru))
1675 __count_vm_events(PGDEACTIVATE, pgmoved);
1676}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001677
Hugh Dickinsf626012d2012-01-12 17:20:06 -08001678static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001679 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001680 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001681 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001683 unsigned long nr_taken;
Hugh Dickinsf626012d2012-01-12 17:20:06 -08001684 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001685 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001687 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001688 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001690 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001691 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001692 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001693 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001694 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695
1696 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001697
1698 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001699 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001700 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001701 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001702
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001704
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001705 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1706 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001707 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001708 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001709
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001710 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001711
Hugh Dickinsf626012d2012-01-12 17:20:06 -08001712 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001713 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001714 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 spin_unlock_irq(&zone->lru_lock);
1716
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 while (!list_empty(&l_hold)) {
1718 cond_resched();
1719 page = lru_to_page(&l_hold);
1720 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001721
Hugh Dickins39b5f292012-10-08 16:33:18 -07001722 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001723 putback_lru_page(page);
1724 continue;
1725 }
1726
Mel Gormancc715d92012-03-21 16:34:00 -07001727 if (unlikely(buffer_heads_over_limit)) {
1728 if (page_has_private(page) && trylock_page(page)) {
1729 if (page_has_private(page))
1730 try_to_release_page(page, 0);
1731 unlock_page(page);
1732 }
1733 }
1734
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001735 if (page_referenced(page, 0, sc->target_mem_cgroup,
1736 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001737 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001738 /*
1739 * Identify referenced, file-backed active pages and
1740 * give them one more trip around the active list. So
1741 * that executable code get better chances to stay in
1742 * memory under moderate memory pressure. Anon pages
1743 * are not likely to be evicted by use-once streaming
1744 * IO, plus JVM can create lots of anon VM_EXEC pages,
1745 * so we ignore them here.
1746 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001747 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001748 list_add(&page->lru, &l_active);
1749 continue;
1750 }
1751 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001752
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001753 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 list_add(&page->lru, &l_inactive);
1755 }
1756
Andrew Mortonb5557492009-01-06 14:40:13 -08001757 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001758 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001759 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001760 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001761 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001762 * Count referenced pages from currently used mappings as rotated,
1763 * even though only some of them are actually re-activated. This
1764 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001765 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001766 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001767 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001768
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001769 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1770 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001771 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001772 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001773
Mel Gormanb745bc82014-06-04 16:10:22 -07001774 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775}
1776
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001777#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001778static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001779{
1780 unsigned long active, inactive;
1781
1782 active = zone_page_state(zone, NR_ACTIVE_ANON);
1783 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1784
1785 if (inactive * zone->inactive_ratio < active)
1786 return 1;
1787
1788 return 0;
1789}
1790
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001791/**
1792 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001793 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001794 *
1795 * Returns true if the zone does not have enough inactive anon pages,
1796 * meaning some active anon pages need to be deactivated.
1797 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001798static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001799{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001800 /*
1801 * If we don't have swap space, anonymous page deactivation
1802 * is pointless.
1803 */
1804 if (!total_swap_pages)
1805 return 0;
1806
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001807 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001808 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001809
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001810 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001811}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001812#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001813static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001814{
1815 return 0;
1816}
1817#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001818
Rik van Riel56e49d22009-06-16 15:32:28 -07001819/**
1820 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001821 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001822 *
1823 * When the system is doing streaming IO, memory pressure here
1824 * ensures that active file pages get deactivated, until more
1825 * than half of the file pages are on the inactive list.
1826 *
1827 * Once we get to that situation, protect the system's working
1828 * set from being evicted by disabling active file page aging.
1829 *
1830 * This uses a different ratio than the anonymous pages, because
1831 * the page cache uses a use-once replacement algorithm.
1832 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001833static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001834{
Johannes Weinere3790142013-02-22 16:35:19 -08001835 unsigned long inactive;
1836 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001837
Johannes Weinere3790142013-02-22 16:35:19 -08001838 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1839 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1840
1841 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001842}
1843
Hugh Dickins75b00af2012-05-29 15:07:09 -07001844static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001845{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001846 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001847 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001848 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001849 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001850}
1851
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001852static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001853 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001854{
Rik van Rielb39415b2009-12-14 17:59:48 -08001855 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001856 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001857 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001858 return 0;
1859 }
1860
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001861 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001862}
1863
Johannes Weiner9a265112013-02-22 16:32:17 -08001864enum scan_balance {
1865 SCAN_EQUAL,
1866 SCAN_FRACT,
1867 SCAN_ANON,
1868 SCAN_FILE,
1869};
1870
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001872 * Determine how aggressively the anon and file LRU lists should be
1873 * scanned. The relative value of each set of LRU lists is determined
1874 * by looking at the fraction of the pages scanned we did rotate back
1875 * onto the active list instead of evict.
1876 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001877 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1878 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001879 */
Johannes Weiner02695172014-08-06 16:06:17 -07001880static void get_scan_count(struct lruvec *lruvec, int swappiness,
1881 struct scan_control *sc, unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001882{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001883 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001884 u64 fraction[2];
1885 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001886 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001887 unsigned long anon_prio, file_prio;
1888 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001889 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001890 bool force_scan = false;
1891 unsigned long ap, fp;
1892 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001893 bool some_scanned;
1894 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001895
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001896 /*
1897 * If the zone or memcg is small, nr[l] can be 0. This
1898 * results in no scanning on this priority and a potential
1899 * priority drop. Global direct reclaim can go to the next
1900 * zone and tends to have no problems. Global kswapd is for
1901 * zone balancing and it needs to scan a minimum amount. When
1902 * reclaiming for a memcg, a priority drop can cause high
1903 * latencies, so it's better to scan a minimum amount there as
1904 * well.
1905 */
Lisa Du6e543d52013-09-11 14:22:36 -07001906 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001907 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001908 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001909 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001910
1911 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001912 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001913 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001914 goto out;
1915 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001916
Johannes Weiner10316b32013-02-22 16:32:14 -08001917 /*
1918 * Global reclaim will swap to prevent OOM even with no
1919 * swappiness, but memcg users want to use this knob to
1920 * disable swapping for individual groups completely when
1921 * using the memory controller's swap limit feature would be
1922 * too expensive.
1923 */
Johannes Weiner02695172014-08-06 16:06:17 -07001924 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001925 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001926 goto out;
1927 }
1928
1929 /*
1930 * Do not apply any pressure balancing cleverness when the
1931 * system is close to OOM, scan both anon and file equally
1932 * (unless the swappiness setting disagrees with swapping).
1933 */
Johannes Weiner02695172014-08-06 16:06:17 -07001934 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001935 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001936 goto out;
1937 }
1938
Johannes Weiner11d16c22013-02-22 16:32:15 -08001939 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001940 * Prevent the reclaimer from falling into the cache trap: as
1941 * cache pages start out inactive, every cache fault will tip
1942 * the scan balance towards the file LRU. And as the file LRU
1943 * shrinks, so does the window for rotation from references.
1944 * This means we have a runaway feedback loop where a tiny
1945 * thrashing file LRU becomes infinitely more attractive than
1946 * anon pages. Try to detect this based on file LRU size.
1947 */
1948 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001949 unsigned long zonefile;
1950 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07001951
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001952 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1953 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1954 zone_page_state(zone, NR_INACTIVE_FILE);
1955
1956 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07001957 scan_balance = SCAN_ANON;
1958 goto out;
1959 }
1960 }
1961
1962 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001963 * There is enough inactive page cache, do not reclaim
1964 * anything from the anonymous working set right now.
1965 */
1966 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001967 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001968 goto out;
1969 }
1970
Johannes Weiner9a265112013-02-22 16:32:17 -08001971 scan_balance = SCAN_FRACT;
1972
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001973 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001974 * With swappiness at 100, anonymous and file have the same priority.
1975 * This scanning priority is essentially the inverse of IO cost.
1976 */
Johannes Weiner02695172014-08-06 16:06:17 -07001977 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001978 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001979
1980 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001981 * OK, so we have swap space and a fair amount of page cache
1982 * pages. We use the recently rotated / recently scanned
1983 * ratios to determine how valuable each cache is.
1984 *
1985 * Because workloads change over time (and to avoid overflow)
1986 * we keep these statistics as a floating average, which ends
1987 * up weighing recent references more than old ones.
1988 *
1989 * anon in [0], file in [1]
1990 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001991
1992 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1993 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1994 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1995 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1996
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001997 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001998 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001999 reclaim_stat->recent_scanned[0] /= 2;
2000 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002001 }
2002
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002003 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002004 reclaim_stat->recent_scanned[1] /= 2;
2005 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002006 }
2007
2008 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002009 * The amount of pressure on anon vs file pages is inversely
2010 * proportional to the fraction of recently scanned pages on
2011 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002012 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002013 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002014 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002015
Satoru Moriyafe350042012-05-29 15:06:47 -07002016 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002017 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002018 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002019
Shaohua Li76a33fc2010-05-24 14:32:36 -07002020 fraction[0] = ap;
2021 fraction[1] = fp;
2022 denominator = ap + fp + 1;
2023out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002024 some_scanned = false;
2025 /* Only use force_scan on second pass. */
2026 for (pass = 0; !some_scanned && pass < 2; pass++) {
2027 for_each_evictable_lru(lru) {
2028 int file = is_file_lru(lru);
2029 unsigned long size;
2030 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002031
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002032 size = get_lru_size(lruvec, lru);
2033 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002034
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002035 if (!scan && pass && force_scan)
2036 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002037
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002038 switch (scan_balance) {
2039 case SCAN_EQUAL:
2040 /* Scan lists relative to size */
2041 break;
2042 case SCAN_FRACT:
2043 /*
2044 * Scan types proportional to swappiness and
2045 * their relative recent reclaim efficiency.
2046 */
2047 scan = div64_u64(scan * fraction[file],
2048 denominator);
2049 break;
2050 case SCAN_FILE:
2051 case SCAN_ANON:
2052 /* Scan one type exclusively */
2053 if ((scan_balance == SCAN_FILE) != file)
2054 scan = 0;
2055 break;
2056 default:
2057 /* Look ma, no brain */
2058 BUG();
2059 }
2060 nr[lru] = scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002061 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002062 * Skip the second pass and don't force_scan,
2063 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002064 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002065 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002066 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002067 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002068}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002069
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002070/*
2071 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2072 */
Johannes Weiner02695172014-08-06 16:06:17 -07002073static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2074 struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002075{
2076 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002077 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002078 unsigned long nr_to_scan;
2079 enum lru_list lru;
2080 unsigned long nr_reclaimed = 0;
2081 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2082 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002083 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002084
Johannes Weiner02695172014-08-06 16:06:17 -07002085 get_scan_count(lruvec, swappiness, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002086
Mel Gormane82e0562013-07-03 15:01:44 -07002087 /* Record the original scan target for proportional adjustments later */
2088 memcpy(targets, nr, sizeof(nr));
2089
Mel Gorman1a501902014-06-04 16:10:49 -07002090 /*
2091 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2092 * event that can occur when there is little memory pressure e.g.
2093 * multiple streaming readers/writers. Hence, we do not abort scanning
2094 * when the requested number of pages are reclaimed when scanning at
2095 * DEF_PRIORITY on the assumption that the fact we are direct
2096 * reclaiming implies that kswapd is not keeping up and it is best to
2097 * do a batch of work at once. For memcg reclaim one check is made to
2098 * abort proportional reclaim if either the file or anon lru has already
2099 * dropped to zero at the first pass.
2100 */
2101 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2102 sc->priority == DEF_PRIORITY);
2103
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002104 blk_start_plug(&plug);
2105 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2106 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002107 unsigned long nr_anon, nr_file, percentage;
2108 unsigned long nr_scanned;
2109
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002110 for_each_evictable_lru(lru) {
2111 if (nr[lru]) {
2112 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2113 nr[lru] -= nr_to_scan;
2114
2115 nr_reclaimed += shrink_list(lru, nr_to_scan,
2116 lruvec, sc);
2117 }
2118 }
Mel Gormane82e0562013-07-03 15:01:44 -07002119
2120 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2121 continue;
2122
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002123 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002124 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002125 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002126 * proportionally what was requested by get_scan_count(). We
2127 * stop reclaiming one LRU and reduce the amount scanning
2128 * proportional to the original scan target.
2129 */
2130 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2131 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2132
Mel Gorman1a501902014-06-04 16:10:49 -07002133 /*
2134 * It's just vindictive to attack the larger once the smaller
2135 * has gone to zero. And given the way we stop scanning the
2136 * smaller below, this makes sure that we only make one nudge
2137 * towards proportionality once we've got nr_to_reclaim.
2138 */
2139 if (!nr_file || !nr_anon)
2140 break;
2141
Mel Gormane82e0562013-07-03 15:01:44 -07002142 if (nr_file > nr_anon) {
2143 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2144 targets[LRU_ACTIVE_ANON] + 1;
2145 lru = LRU_BASE;
2146 percentage = nr_anon * 100 / scan_target;
2147 } else {
2148 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2149 targets[LRU_ACTIVE_FILE] + 1;
2150 lru = LRU_FILE;
2151 percentage = nr_file * 100 / scan_target;
2152 }
2153
2154 /* Stop scanning the smaller of the LRU */
2155 nr[lru] = 0;
2156 nr[lru + LRU_ACTIVE] = 0;
2157
2158 /*
2159 * Recalculate the other LRU scan count based on its original
2160 * scan target and the percentage scanning already complete
2161 */
2162 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2163 nr_scanned = targets[lru] - nr[lru];
2164 nr[lru] = targets[lru] * (100 - percentage) / 100;
2165 nr[lru] -= min(nr[lru], nr_scanned);
2166
2167 lru += LRU_ACTIVE;
2168 nr_scanned = targets[lru] - nr[lru];
2169 nr[lru] = targets[lru] * (100 - percentage) / 100;
2170 nr[lru] -= min(nr[lru], nr_scanned);
2171
2172 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002173 }
2174 blk_finish_plug(&plug);
2175 sc->nr_reclaimed += nr_reclaimed;
2176
2177 /*
2178 * Even if we did not try to evict anon pages at all, we want to
2179 * rebalance the anon lru active/inactive ratio.
2180 */
2181 if (inactive_anon_is_low(lruvec))
2182 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2183 sc, LRU_ACTIVE_ANON);
2184
2185 throttle_vm_writeout(sc->gfp_mask);
2186}
2187
Mel Gorman23b9da52012-05-29 15:06:20 -07002188/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002189static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002190{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002191 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002192 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002193 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002194 return true;
2195
2196 return false;
2197}
2198
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002199/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002200 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2201 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2202 * true if more pages should be reclaimed such that when the page allocator
2203 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2204 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002205 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002206static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002207 unsigned long nr_reclaimed,
2208 unsigned long nr_scanned,
2209 struct scan_control *sc)
2210{
2211 unsigned long pages_for_compaction;
2212 unsigned long inactive_lru_pages;
2213
2214 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002215 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002216 return false;
2217
Mel Gorman28765922011-02-25 14:44:20 -08002218 /* Consider stopping depending on scan and reclaim activity */
2219 if (sc->gfp_mask & __GFP_REPEAT) {
2220 /*
2221 * For __GFP_REPEAT allocations, stop reclaiming if the
2222 * full LRU list has been scanned and we are still failing
2223 * to reclaim pages. This full LRU scan is potentially
2224 * expensive but a __GFP_REPEAT caller really wants to succeed
2225 */
2226 if (!nr_reclaimed && !nr_scanned)
2227 return false;
2228 } else {
2229 /*
2230 * For non-__GFP_REPEAT allocations which can presumably
2231 * fail without consequence, stop if we failed to reclaim
2232 * any pages from the last SWAP_CLUSTER_MAX number of
2233 * pages that were scanned. This will return to the
2234 * caller faster at the risk reclaim/compaction and
2235 * the resulting allocation attempt fails
2236 */
2237 if (!nr_reclaimed)
2238 return false;
2239 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002240
2241 /*
2242 * If we have not reclaimed enough pages for compaction and the
2243 * inactive lists are large enough, continue reclaiming
2244 */
2245 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002246 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002247 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002248 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002249 if (sc->nr_reclaimed < pages_for_compaction &&
2250 inactive_lru_pages > pages_for_compaction)
2251 return true;
2252
2253 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002254 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002255 case COMPACT_PARTIAL:
2256 case COMPACT_CONTINUE:
2257 return false;
2258 default:
2259 return true;
2260 }
2261}
2262
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002263static bool shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002264{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002265 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002266 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002267
Johannes Weiner5660048c2012-01-12 17:17:59 -08002268 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002269 struct mem_cgroup *root = sc->target_mem_cgroup;
2270 struct mem_cgroup_reclaim_cookie reclaim = {
2271 .zone = zone,
2272 .priority = sc->priority,
2273 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002274 struct mem_cgroup *memcg;
Johannes Weiner5660048c2012-01-12 17:17:59 -08002275
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002276 nr_reclaimed = sc->nr_reclaimed;
2277 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002278
Andrew Morton694fbc02013-09-24 15:27:37 -07002279 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2280 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002281 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002282 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002283
2284 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002285 swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002286
Johannes Weiner02695172014-08-06 16:06:17 -07002287 shrink_lruvec(lruvec, swappiness, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002288
2289 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002290 * Direct reclaim and kswapd have to scan all memory
2291 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002292 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002293 *
2294 * Limit reclaim, on the other hand, only cares about
2295 * nr_to_reclaim pages to be reclaimed and it will
2296 * retry with decreasing priority if one round over the
2297 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002298 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002299 if (!global_reclaim(sc) &&
2300 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002301 mem_cgroup_iter_break(root, memcg);
2302 break;
2303 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002304 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2305 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002306
2307 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2308 sc->nr_scanned - nr_scanned,
2309 sc->nr_reclaimed - nr_reclaimed);
2310
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002311 if (sc->nr_reclaimed - nr_reclaimed)
2312 reclaimable = true;
2313
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002314 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2315 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002316
2317 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002318}
2319
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002320/* Returns true if compaction should go ahead for a high-order request */
Johannes Weiner0b064962014-08-06 16:06:12 -07002321static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002322{
2323 unsigned long balance_gap, watermark;
2324 bool watermark_ok;
2325
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002326 /*
2327 * Compaction takes time to run and there are potentially other
2328 * callers using the pages just freed. Continue reclaiming until
2329 * there is a buffer of free pages available to give compaction
2330 * a reasonable chance of completing and allocating the page
2331 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002332 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2333 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002334 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002335 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2336
2337 /*
2338 * If compaction is deferred, reclaim up to a point where
2339 * compaction will have a chance of success when re-enabled
2340 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002341 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002342 return watermark_ok;
2343
2344 /* If compaction is not ready to start, keep reclaiming */
Johannes Weiner0b064962014-08-06 16:06:12 -07002345 if (!compaction_suitable(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002346 return false;
2347
2348 return watermark_ok;
2349}
2350
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351/*
2352 * This is the direct reclaim path, for page-allocating processes. We only
2353 * try to reclaim pages from zones which will satisfy the caller's allocation
2354 * request.
2355 *
Mel Gorman41858962009-06-16 15:32:12 -07002356 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2357 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2359 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002360 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2361 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2362 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 * If a zone is deemed to be full of pinned pages then just give it a light
2365 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002366 *
2367 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002369static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
Mel Gormandd1a2392008-04-28 02:12:17 -07002371 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002372 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002373 unsigned long nr_soft_reclaimed;
2374 unsigned long nr_soft_scanned;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002375 unsigned long lru_pages = 0;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002376 struct reclaim_state *reclaim_state = current->reclaim_state;
Weijie Yang619d0d762014-04-07 15:36:59 -07002377 gfp_t orig_mask;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002378 struct shrink_control shrink = {
2379 .gfp_mask = sc->gfp_mask,
2380 };
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002381 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002382 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002383
Mel Gormancc715d92012-03-21 16:34:00 -07002384 /*
2385 * If the number of buffer_heads in the machine exceeds the maximum
2386 * allowed level, force direct reclaim to scan the highmem zone as
2387 * highmem pages could be pinning lowmem pages storing buffer_heads
2388 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002389 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002390 if (buffer_heads_over_limit)
2391 sc->gfp_mask |= __GFP_HIGHMEM;
2392
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002393 nodes_clear(shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002394
Mel Gormand4debc62010-08-09 17:19:29 -07002395 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2396 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002397 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002399 /*
2400 * Take care memory controller reclaiming has small influence
2401 * to global LRU.
2402 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002403 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002404 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2405 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002406
2407 lru_pages += zone_reclaimable_pages(zone);
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002408 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002409
Lisa Du6e543d52013-09-11 14:22:36 -07002410 if (sc->priority != DEF_PRIORITY &&
2411 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002412 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002413
2414 /*
2415 * If we already have plenty of memory free for
2416 * compaction in this zone, don't free any more.
2417 * Even though compaction is invoked for any
2418 * non-zero order, only frequent costly order
2419 * reclamation is disruptive enough to become a
2420 * noticeable problem, like transparent huge
2421 * page allocations.
2422 */
2423 if (IS_ENABLED(CONFIG_COMPACTION) &&
2424 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2425 zonelist_zone_idx(z) <= requested_highidx &&
2426 compaction_ready(zone, sc->order)) {
2427 sc->compaction_ready = true;
2428 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002429 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002430
Andrew Morton0608f432013-09-24 15:27:41 -07002431 /*
2432 * This steals pages from memory cgroups over softlimit
2433 * and returns the number of reclaimed pages and
2434 * scanned pages. This works for global memory pressure
2435 * and balancing, not for a memcg's limit.
2436 */
2437 nr_soft_scanned = 0;
2438 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2439 sc->order, sc->gfp_mask,
2440 &nr_soft_scanned);
2441 sc->nr_reclaimed += nr_soft_reclaimed;
2442 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002443 if (nr_soft_reclaimed)
2444 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002445 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002446 }
Nick Piggin408d8542006-09-25 23:31:27 -07002447
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002448 if (shrink_zone(zone, sc))
2449 reclaimable = true;
2450
2451 if (global_reclaim(sc) &&
2452 !reclaimable && zone_reclaimable(zone))
2453 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 }
Mel Gormane0c23272011-10-31 17:09:33 -07002455
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002456 /*
2457 * Don't shrink slabs when reclaiming memory from over limit cgroups
2458 * but do shrink slab at least once when aborting reclaim for
2459 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2460 * pages.
2461 */
2462 if (global_reclaim(sc)) {
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002463 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002464 if (reclaim_state) {
2465 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2466 reclaim_state->reclaimed_slab = 0;
2467 }
2468 }
2469
Weijie Yang619d0d762014-04-07 15:36:59 -07002470 /*
2471 * Restore to original mask to avoid the impact on the caller if we
2472 * promoted it to __GFP_HIGHMEM.
2473 */
2474 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002475
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002476 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002478
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479/*
2480 * This is the main entry point to direct page reclaim.
2481 *
2482 * If a full scan of the inactive list fails to free enough memory then we
2483 * are "out of memory" and something needs to be killed.
2484 *
2485 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2486 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002487 * caller can't do much about. We kick the writeback threads and take explicit
2488 * naps in the hope that some of these pages can be written. But if the
2489 * allocating task holds filesystem locks which prevent writeout this might not
2490 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002491 *
2492 * returns: 0, if no pages reclaimed
2493 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494 */
Mel Gormandac1d272008-04-28 02:12:12 -07002495static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002496 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497{
Andrew Morton69e05942006-03-22 00:08:19 -08002498 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002499 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002500 bool zones_reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501
Keika Kobayashi873b4772008-07-25 01:48:52 -07002502 delayacct_freepages_start();
2503
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002504 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002505 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002507 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002508 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2509 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002510 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002511 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002512
Balbir Singh66e17072008-02-07 00:13:56 -08002513 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002514 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002515 break;
2516
2517 if (sc->compaction_ready)
2518 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519
2520 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002521 * If we're getting trouble reclaiming, start doing
2522 * writepage even in laptop mode.
2523 */
2524 if (sc->priority < DEF_PRIORITY - 2)
2525 sc->may_writepage = 1;
2526
2527 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 * Try to write back as many pages as we just scanned. This
2529 * tends to cause slow streaming writers to write data to the
2530 * disk smoothly, at the dirtying rate, which is nice. But
2531 * that's undesirable in laptop mode, where we *want* lumpy
2532 * writeout. So in laptop mode, write out the whole world.
2533 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002534 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2535 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002536 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2537 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002538 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002540 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002541
Keika Kobayashi873b4772008-07-25 01:48:52 -07002542 delayacct_freepages_end();
2543
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002544 if (sc->nr_reclaimed)
2545 return sc->nr_reclaimed;
2546
Mel Gorman0cee34f2012-01-12 17:19:49 -08002547 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002548 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002549 return 1;
2550
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002551 /* Any of the zones still reclaimable? Don't OOM. */
2552 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002553 return 1;
2554
2555 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556}
2557
Mel Gorman55150612012-07-31 16:44:35 -07002558static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2559{
2560 struct zone *zone;
2561 unsigned long pfmemalloc_reserve = 0;
2562 unsigned long free_pages = 0;
2563 int i;
2564 bool wmark_ok;
2565
2566 for (i = 0; i <= ZONE_NORMAL; i++) {
2567 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002568 if (!populated_zone(zone))
2569 continue;
2570
Mel Gorman55150612012-07-31 16:44:35 -07002571 pfmemalloc_reserve += min_wmark_pages(zone);
2572 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2573 }
2574
Mel Gorman675becc2014-06-04 16:07:35 -07002575 /* If there are no reserves (unexpected config) then do not throttle */
2576 if (!pfmemalloc_reserve)
2577 return true;
2578
Mel Gorman55150612012-07-31 16:44:35 -07002579 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2580
2581 /* kswapd must be awake if processes are being throttled */
2582 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2583 pgdat->classzone_idx = min(pgdat->classzone_idx,
2584 (enum zone_type)ZONE_NORMAL);
2585 wake_up_interruptible(&pgdat->kswapd_wait);
2586 }
2587
2588 return wmark_ok;
2589}
2590
2591/*
2592 * Throttle direct reclaimers if backing storage is backed by the network
2593 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2594 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002595 * when the low watermark is reached.
2596 *
2597 * Returns true if a fatal signal was delivered during throttling. If this
2598 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002599 */
Mel Gorman50694c22012-11-26 16:29:48 -08002600static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002601 nodemask_t *nodemask)
2602{
Mel Gorman675becc2014-06-04 16:07:35 -07002603 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002604 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002605 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002606
2607 /*
2608 * Kernel threads should not be throttled as they may be indirectly
2609 * responsible for cleaning pages necessary for reclaim to make forward
2610 * progress. kjournald for example may enter direct reclaim while
2611 * committing a transaction where throttling it could forcing other
2612 * processes to block on log_wait_commit().
2613 */
2614 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002615 goto out;
2616
2617 /*
2618 * If a fatal signal is pending, this process should not throttle.
2619 * It should return quickly so it can exit and free its memory
2620 */
2621 if (fatal_signal_pending(current))
2622 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002623
Mel Gorman675becc2014-06-04 16:07:35 -07002624 /*
2625 * Check if the pfmemalloc reserves are ok by finding the first node
2626 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2627 * GFP_KERNEL will be required for allocating network buffers when
2628 * swapping over the network so ZONE_HIGHMEM is unusable.
2629 *
2630 * Throttling is based on the first usable node and throttled processes
2631 * wait on a queue until kswapd makes progress and wakes them. There
2632 * is an affinity then between processes waking up and where reclaim
2633 * progress has been made assuming the process wakes on the same node.
2634 * More importantly, processes running on remote nodes will not compete
2635 * for remote pfmemalloc reserves and processes on different nodes
2636 * should make reasonable progress.
2637 */
2638 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2639 gfp_mask, nodemask) {
2640 if (zone_idx(zone) > ZONE_NORMAL)
2641 continue;
2642
2643 /* Throttle based on the first usable node */
2644 pgdat = zone->zone_pgdat;
2645 if (pfmemalloc_watermark_ok(pgdat))
2646 goto out;
2647 break;
2648 }
2649
2650 /* If no zone was usable by the allocation flags then do not throttle */
2651 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002652 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002653
Mel Gorman68243e72012-07-31 16:44:39 -07002654 /* Account for the throttling */
2655 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2656
Mel Gorman55150612012-07-31 16:44:35 -07002657 /*
2658 * If the caller cannot enter the filesystem, it's possible that it
2659 * is due to the caller holding an FS lock or performing a journal
2660 * transaction in the case of a filesystem like ext[3|4]. In this case,
2661 * it is not safe to block on pfmemalloc_wait as kswapd could be
2662 * blocked waiting on the same lock. Instead, throttle for up to a
2663 * second before continuing.
2664 */
2665 if (!(gfp_mask & __GFP_FS)) {
2666 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2667 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002668
2669 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002670 }
2671
2672 /* Throttle until kswapd wakes the process */
2673 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2674 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002675
2676check_pending:
2677 if (fatal_signal_pending(current))
2678 return true;
2679
2680out:
2681 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002682}
2683
Mel Gormandac1d272008-04-28 02:12:12 -07002684unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002685 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002686{
Mel Gorman33906bc2010-08-09 17:19:16 -07002687 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002688 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002689 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002690 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2691 .order = order,
2692 .nodemask = nodemask,
2693 .priority = DEF_PRIORITY,
2694 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002695 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002696 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002697 };
2698
Mel Gorman55150612012-07-31 16:44:35 -07002699 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002700 * Do not enter reclaim if fatal signal was delivered while throttled.
2701 * 1 is returned so that the page allocator does not OOM kill at this
2702 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002703 */
Mel Gorman50694c22012-11-26 16:29:48 -08002704 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002705 return 1;
2706
Mel Gorman33906bc2010-08-09 17:19:16 -07002707 trace_mm_vmscan_direct_reclaim_begin(order,
2708 sc.may_writepage,
2709 gfp_mask);
2710
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002711 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002712
2713 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2714
2715 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002716}
2717
Andrew Mortonc255a452012-07-31 16:43:02 -07002718#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002719
Johannes Weiner72835c862012-01-12 17:18:32 -08002720unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002721 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002722 struct zone *zone,
2723 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002724{
2725 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002726 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002727 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002728 .may_writepage = !laptop_mode,
2729 .may_unmap = 1,
2730 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002731 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002732 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002733 int swappiness = mem_cgroup_swappiness(memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002734
Balbir Singh4e416952009-09-23 15:56:39 -07002735 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2736 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002737
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002738 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002739 sc.may_writepage,
2740 sc.gfp_mask);
2741
Balbir Singh4e416952009-09-23 15:56:39 -07002742 /*
2743 * NOTE: Although we can get the priority field, using it
2744 * here is not a good idea, since it limits the pages we can scan.
2745 * if we don't reclaim here, the shrink_zone from balance_pgdat
2746 * will pick up pages from other mem cgroup's as well. We hack
2747 * the priority and make it zero.
2748 */
Johannes Weiner02695172014-08-06 16:06:17 -07002749 shrink_lruvec(lruvec, swappiness, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002750
2751 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2752
Ying Han0ae5e892011-05-26 16:25:25 -07002753 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002754 return sc.nr_reclaimed;
2755}
2756
Johannes Weiner72835c862012-01-12 17:18:32 -08002757unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002758 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002759 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002760{
Balbir Singh4e416952009-09-23 15:56:39 -07002761 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002762 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002763 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002764 struct scan_control sc = {
Johannes Weineree814fe2014-08-06 16:06:19 -07002765 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2766 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2767 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2768 .target_mem_cgroup = memcg,
2769 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002770 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002771 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002772 .may_swap = !noswap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002773 };
Balbir Singh66e17072008-02-07 00:13:56 -08002774
Ying Han889976d2011-05-26 16:25:33 -07002775 /*
2776 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2777 * take care of from where we get pages. So the node where we start the
2778 * scan does not need to be the current node.
2779 */
Johannes Weiner72835c862012-01-12 17:18:32 -08002780 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002781
2782 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002783
2784 trace_mm_vmscan_memcg_reclaim_begin(0,
2785 sc.may_writepage,
2786 sc.gfp_mask);
2787
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002788 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002789
2790 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2791
2792 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002793}
2794#endif
2795
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002796static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002797{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002798 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002799
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002800 if (!total_swap_pages)
2801 return;
2802
2803 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2804 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002805 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002806
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002807 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002808 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002809 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002810
2811 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2812 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002813}
2814
Johannes Weiner60cefed2012-11-29 13:54:23 -08002815static bool zone_balanced(struct zone *zone, int order,
2816 unsigned long balance_gap, int classzone_idx)
2817{
2818 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2819 balance_gap, classzone_idx, 0))
2820 return false;
2821
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002822 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2823 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002824 return false;
2825
2826 return true;
2827}
2828
Mel Gorman1741c872011-01-13 15:46:21 -08002829/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002830 * pgdat_balanced() is used when checking if a node is balanced.
2831 *
2832 * For order-0, all zones must be balanced!
2833 *
2834 * For high-order allocations only zones that meet watermarks and are in a
2835 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2836 * total of balanced pages must be at least 25% of the zones allowed by
2837 * classzone_idx for the node to be considered balanced. Forcing all zones to
2838 * be balanced for high orders can cause excessive reclaim when there are
2839 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002840 * The choice of 25% is due to
2841 * o a 16M DMA zone that is balanced will not balance a zone on any
2842 * reasonable sized machine
2843 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002844 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002845 * would need to be at least 256M for it to be balance a whole node.
2846 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2847 * to balance a node on its own. These seemed like reasonable ratios.
2848 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002849static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002850{
Jiang Liub40da042013-02-22 16:33:52 -08002851 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002852 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002853 int i;
2854
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002855 /* Check the watermark levels */
2856 for (i = 0; i <= classzone_idx; i++) {
2857 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002858
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002859 if (!populated_zone(zone))
2860 continue;
2861
Jiang Liub40da042013-02-22 16:33:52 -08002862 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002863
2864 /*
2865 * A special case here:
2866 *
2867 * balance_pgdat() skips over all_unreclaimable after
2868 * DEF_PRIORITY. Effectively, it considers them balanced so
2869 * they must be considered balanced here as well!
2870 */
Lisa Du6e543d52013-09-11 14:22:36 -07002871 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002872 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002873 continue;
2874 }
2875
2876 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002877 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002878 else if (!order)
2879 return false;
2880 }
2881
2882 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002883 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002884 else
2885 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002886}
2887
Mel Gorman55150612012-07-31 16:44:35 -07002888/*
2889 * Prepare kswapd for sleeping. This verifies that there are no processes
2890 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2891 *
2892 * Returns true if kswapd is ready to sleep
2893 */
2894static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002895 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002896{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002897 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2898 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002899 return false;
2900
2901 /*
2902 * There is a potential race between when kswapd checks its watermarks
2903 * and a process gets throttled. There is also a potential race if
2904 * processes get throttled, kswapd wakes, a large process exits therby
2905 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2906 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2907 * so wake them now if necessary. If necessary, processes will wake
2908 * kswapd and get throttled again
2909 */
2910 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2911 wake_up(&pgdat->pfmemalloc_wait);
2912 return false;
2913 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002914
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002915 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002916}
2917
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918/*
Mel Gorman75485362013-07-03 15:01:42 -07002919 * kswapd shrinks the zone by the number of pages required to reach
2920 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002921 *
2922 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002923 * reclaim or if the lack of progress was due to pages under writeback.
2924 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002925 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002926static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002927 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002928 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002929 unsigned long lru_pages,
2930 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002931{
Mel Gorman7c954f62013-07-03 15:01:54 -07002932 int testorder = sc->order;
2933 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002934 struct reclaim_state *reclaim_state = current->reclaim_state;
2935 struct shrink_control shrink = {
2936 .gfp_mask = sc->gfp_mask,
2937 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002938 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002939
2940 /* Reclaim above the high watermark. */
2941 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002942
2943 /*
2944 * Kswapd reclaims only single pages with compaction enabled. Trying
2945 * too hard to reclaim until contiguous free pages have become
2946 * available can hurt performance by evicting too much useful data
2947 * from memory. Do not reclaim more than needed for compaction.
2948 */
2949 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2950 compaction_suitable(zone, sc->order) !=
2951 COMPACT_SKIPPED)
2952 testorder = 0;
2953
2954 /*
2955 * We put equal pressure on every zone, unless one zone has way too
2956 * many pages free already. The "too many pages" is defined as the
2957 * high wmark plus a "gap" where the gap is either the low
2958 * watermark or 1% of the zone, whichever is smaller.
2959 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002960 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2961 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07002962
2963 /*
2964 * If there is no low memory pressure or the zone is balanced then no
2965 * reclaim is necessary
2966 */
2967 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2968 if (!lowmem_pressure && zone_balanced(zone, testorder,
2969 balance_gap, classzone_idx))
2970 return true;
2971
Mel Gorman75485362013-07-03 15:01:42 -07002972 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002973 nodes_clear(shrink.nodes_to_scan);
2974 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002975
2976 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002977 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002978 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2979
Mel Gorman2ab44f42013-07-03 15:01:47 -07002980 /* Account for the number of pages attempted to reclaim */
2981 *nr_attempted += sc->nr_to_reclaim;
2982
Mel Gorman283aba92013-07-03 15:01:51 -07002983 zone_clear_flag(zone, ZONE_WRITEBACK);
2984
Mel Gorman7c954f62013-07-03 15:01:54 -07002985 /*
2986 * If a zone reaches its high watermark, consider it to be no longer
2987 * congested. It's possible there are dirty pages backed by congested
2988 * BDIs but as pressure is relieved, speculatively avoid congestion
2989 * waits.
2990 */
Lisa Du6e543d52013-09-11 14:22:36 -07002991 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002992 zone_balanced(zone, testorder, 0, classzone_idx)) {
2993 zone_clear_flag(zone, ZONE_CONGESTED);
2994 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2995 }
2996
Mel Gormanb8e83b92013-07-03 15:01:45 -07002997 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002998}
2999
3000/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003002 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003004 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 *
3006 * There is special handling here for zones which are full of pinned pages.
3007 * This can happen if the pages are all mlocked, or if they are all used by
3008 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3009 * What we do is to detect the case where all pages in the zone have been
3010 * scanned twice and there has been zero successful reclaim. Mark the zone as
3011 * dead and from now on, only perform a short scan. Basically we're polling
3012 * the zone for when the problem goes away.
3013 *
3014 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003015 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3016 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3017 * lower zones regardless of the number of free pages in the lower zones. This
3018 * interoperates with the page allocator fallback scheme to ensure that aging
3019 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 */
Mel Gorman99504742011-01-13 15:46:20 -08003021static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003022 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003025 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003026 unsigned long nr_soft_reclaimed;
3027 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003028 struct scan_control sc = {
3029 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003030 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003031 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003032 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003033 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003034 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003035 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003036 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003038 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003040 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003041 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003042 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003043
3044 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003046 /*
3047 * Scan in the highmem->dma direction for the highest
3048 * zone which needs scanning
3049 */
3050 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3051 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003053 if (!populated_zone(zone))
3054 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055
Lisa Du6e543d52013-09-11 14:22:36 -07003056 if (sc.priority != DEF_PRIORITY &&
3057 !zone_reclaimable(zone))
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003058 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059
Rik van Riel556adec2008-10-18 20:26:34 -07003060 /*
3061 * Do some background aging of the anon list, to give
3062 * pages a chance to be referenced before reclaiming.
3063 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003064 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003065
Mel Gormancc715d92012-03-21 16:34:00 -07003066 /*
3067 * If the number of buffer_heads in the machine
3068 * exceeds the maximum allowed level and this node
3069 * has a highmem zone, force kswapd to reclaim from
3070 * it to relieve lowmem pressure.
3071 */
3072 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3073 end_zone = i;
3074 break;
3075 }
3076
Johannes Weiner60cefed2012-11-29 13:54:23 -08003077 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003078 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003079 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003080 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003081 /*
3082 * If balanced, clear the dirty and congested
3083 * flags
3084 */
Shaohua Li439423f2011-08-25 15:59:12 -07003085 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003086 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003089
Mel Gormanb8e83b92013-07-03 15:01:45 -07003090 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003091 goto out;
3092
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 for (i = 0; i <= end_zone; i++) {
3094 struct zone *zone = pgdat->node_zones + i;
3095
Mel Gorman2ab44f42013-07-03 15:01:47 -07003096 if (!populated_zone(zone))
3097 continue;
3098
Wu Fengguangadea02a2009-09-21 17:01:42 -07003099 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003100
3101 /*
3102 * If any zone is currently balanced then kswapd will
3103 * not call compaction as it is expected that the
3104 * necessary pages are already available.
3105 */
3106 if (pgdat_needs_compaction &&
3107 zone_watermark_ok(zone, order,
3108 low_wmark_pages(zone),
3109 *classzone_idx, 0))
3110 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 }
3112
3113 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003114 * If we're getting trouble reclaiming, start doing writepage
3115 * even in laptop mode.
3116 */
3117 if (sc.priority < DEF_PRIORITY - 2)
3118 sc.may_writepage = 1;
3119
3120 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121 * Now scan the zone in the dma->highmem direction, stopping
3122 * at the last zone which needs scanning.
3123 *
3124 * We do this because the page allocator works in the opposite
3125 * direction. This prevents the page allocator from allocating
3126 * pages behind kswapd's direction of progress, which would
3127 * cause too much scanning of the lower zones.
3128 */
3129 for (i = 0; i <= end_zone; i++) {
3130 struct zone *zone = pgdat->node_zones + i;
3131
Con Kolivasf3fe6512006-01-06 00:11:15 -08003132 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 continue;
3134
Lisa Du6e543d52013-09-11 14:22:36 -07003135 if (sc.priority != DEF_PRIORITY &&
3136 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 continue;
3138
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003140
Andrew Morton0608f432013-09-24 15:27:41 -07003141 nr_soft_scanned = 0;
3142 /*
3143 * Call soft limit reclaim before calling shrink_zone.
3144 */
3145 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3146 order, sc.gfp_mask,
3147 &nr_soft_scanned);
3148 sc.nr_reclaimed += nr_soft_reclaimed;
3149
Rik van Riel32a43302007-10-16 01:24:50 -07003150 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003151 * There should be no need to raise the scanning
3152 * priority if enough pages are already being scanned
3153 * that that high watermark would be met at 100%
3154 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003155 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003156 if (kswapd_shrink_zone(zone, end_zone, &sc,
3157 lru_pages, &nr_attempted))
3158 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 }
Mel Gorman55150612012-07-31 16:44:35 -07003160
3161 /*
3162 * If the low watermark is met there is no need for processes
3163 * to be throttled on pfmemalloc_wait as they should not be
3164 * able to safely make forward progress. Wake them
3165 */
3166 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3167 pfmemalloc_watermark_ok(pgdat))
3168 wake_up(&pgdat->pfmemalloc_wait);
3169
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003171 * Fragmentation may mean that the system cannot be rebalanced
3172 * for high-order allocations in all zones. If twice the
3173 * allocation size has been reclaimed and the zones are still
3174 * not balanced then recheck the watermarks at order-0 to
3175 * prevent kswapd reclaiming excessively. Assume that a
3176 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003178 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003179 order = sc.order = 0;
3180
Mel Gormanb8e83b92013-07-03 15:01:45 -07003181 /* Check if kswapd should be suspending */
3182 if (try_to_freeze() || kthread_should_stop())
3183 break;
3184
3185 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003186 * Compact if necessary and kswapd is reclaiming at least the
3187 * high watermark number of pages as requsted
3188 */
3189 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3190 compact_pgdat(pgdat, order);
3191
3192 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003193 * Raise priority if scanning rate is too low or there was no
3194 * progress in reclaiming pages
3195 */
3196 if (raise_priority || !sc.nr_reclaimed)
3197 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003198 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003199 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200
Mel Gormanb8e83b92013-07-03 15:01:45 -07003201out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003202 /*
Mel Gorman55150612012-07-31 16:44:35 -07003203 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003204 * makes a decision on the order we were last reclaiming at. However,
3205 * if another caller entered the allocator slow path while kswapd
3206 * was awake, order will remain at the higher level
3207 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003208 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003209 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210}
3211
Mel Gormandc83edd2011-01-13 15:46:26 -08003212static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003213{
3214 long remaining = 0;
3215 DEFINE_WAIT(wait);
3216
3217 if (freezing(current) || kthread_should_stop())
3218 return;
3219
3220 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3221
3222 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003223 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003224 remaining = schedule_timeout(HZ/10);
3225 finish_wait(&pgdat->kswapd_wait, &wait);
3226 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3227 }
3228
3229 /*
3230 * After a short sleep, check if it was a premature sleep. If not, then
3231 * go fully to sleep until explicitly woken up.
3232 */
Mel Gorman55150612012-07-31 16:44:35 -07003233 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003234 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3235
3236 /*
3237 * vmstat counters are not perfectly accurate and the estimated
3238 * value for counters such as NR_FREE_PAGES can deviate from the
3239 * true value by nr_online_cpus * threshold. To avoid the zone
3240 * watermarks being breached while under pressure, we reduce the
3241 * per-cpu vmstat threshold while kswapd is awake and restore
3242 * them before going back to sleep.
3243 */
3244 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003245
Mel Gorman62997022012-10-08 16:32:47 -07003246 /*
3247 * Compaction records what page blocks it recently failed to
3248 * isolate pages from and skips them in the future scanning.
3249 * When kswapd is going to sleep, it is reasonable to assume
3250 * that pages and compaction may succeed so reset the cache.
3251 */
3252 reset_isolation_suitable(pgdat);
3253
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003254 if (!kthread_should_stop())
3255 schedule();
3256
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003257 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3258 } else {
3259 if (remaining)
3260 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3261 else
3262 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3263 }
3264 finish_wait(&pgdat->kswapd_wait, &wait);
3265}
3266
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267/*
3268 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003269 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270 *
3271 * This basically trickles out pages so that we have _some_
3272 * free memory available even if there is no other activity
3273 * that frees anything up. This is needed for things like routing
3274 * etc, where we otherwise might have all activity going on in
3275 * asynchronous contexts that cannot page things out.
3276 *
3277 * If there are applications that are active memory-allocators
3278 * (most normal use), this basically shouldn't matter.
3279 */
3280static int kswapd(void *p)
3281{
Mel Gorman215ddd62011-07-08 15:39:40 -07003282 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003283 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003284 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003285 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 pg_data_t *pgdat = (pg_data_t*)p;
3287 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003288
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289 struct reclaim_state reclaim_state = {
3290 .reclaimed_slab = 0,
3291 };
Rusty Russella70f7302009-03-13 14:49:46 +10303292 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293
Nick Piggincf40bd12009-01-21 08:12:39 +01003294 lockdep_set_current_reclaim_state(GFP_KERNEL);
3295
Rusty Russell174596a2009-01-01 10:12:29 +10303296 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003297 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 current->reclaim_state = &reclaim_state;
3299
3300 /*
3301 * Tell the memory management that we're a "memory allocator",
3302 * and that if we need more memory we should get access to it
3303 * regardless (see "__alloc_pages()"). "kswapd" should
3304 * never get caught in the normal page freeing logic.
3305 *
3306 * (Kswapd normally doesn't need memory anyway, but sometimes
3307 * you need a small amount of memory in order to be able to
3308 * page out something else, and this flag essentially protects
3309 * us from recursively trying to free more memory as we're
3310 * trying to free the first piece of memory in the first place).
3311 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003312 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003313 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
Mel Gorman215ddd62011-07-08 15:39:40 -07003315 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003316 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003317 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003318 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003320 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003321
Mel Gorman215ddd62011-07-08 15:39:40 -07003322 /*
3323 * If the last balance_pgdat was unsuccessful it's unlikely a
3324 * new request of a similar or harder type will succeed soon
3325 * so consider going to sleep on the basis we reclaimed at
3326 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003327 if (balanced_classzone_idx >= new_classzone_idx &&
3328 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003329 new_order = pgdat->kswapd_max_order;
3330 new_classzone_idx = pgdat->classzone_idx;
3331 pgdat->kswapd_max_order = 0;
3332 pgdat->classzone_idx = pgdat->nr_zones - 1;
3333 }
3334
Mel Gorman99504742011-01-13 15:46:20 -08003335 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336 /*
3337 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003338 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339 */
3340 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003341 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003343 kswapd_try_to_sleep(pgdat, balanced_order,
3344 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003346 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003347 new_order = order;
3348 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003349 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003350 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352
David Rientjes8fe23e02009-12-14 17:58:33 -08003353 ret = try_to_freeze();
3354 if (kthread_should_stop())
3355 break;
3356
3357 /*
3358 * We can speed up thawing tasks if we don't call balance_pgdat
3359 * after returning from the refrigerator
3360 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003361 if (!ret) {
3362 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003363 balanced_classzone_idx = classzone_idx;
3364 balanced_order = balance_pgdat(pgdat, order,
3365 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003366 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003368
Johannes Weiner71abdc12014-06-06 14:35:35 -07003369 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003370 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003371 lockdep_clear_current_reclaim_state();
3372
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 return 0;
3374}
3375
3376/*
3377 * A zone is low on free memory, so wake its kswapd task to service it.
3378 */
Mel Gorman99504742011-01-13 15:46:20 -08003379void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380{
3381 pg_data_t *pgdat;
3382
Con Kolivasf3fe6512006-01-06 00:11:15 -08003383 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 return;
3385
Paul Jackson02a0e532006-12-13 00:34:25 -08003386 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003388 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003389 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003390 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003391 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3392 }
Con Kolivas8d0986e22005-09-13 01:25:07 -07003393 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003395 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003396 return;
3397
3398 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e22005-09-13 01:25:07 -07003399 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400}
3401
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003402#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003404 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003405 * freed pages.
3406 *
3407 * Rather than trying to age LRUs the aim is to preserve the overall
3408 * LRU order by reclaiming preferentially
3409 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003411unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412{
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003413 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003414 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003415 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003416 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003417 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003418 .may_writepage = 1,
3419 .may_unmap = 1,
3420 .may_swap = 1,
3421 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003423 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003424 struct task_struct *p = current;
3425 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003427 p->flags |= PF_MEMALLOC;
3428 lockdep_set_current_reclaim_state(sc.gfp_mask);
3429 reclaim_state.reclaimed_slab = 0;
3430 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003431
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003432 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003433
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003434 p->reclaim_state = NULL;
3435 lockdep_clear_current_reclaim_state();
3436 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db42006-06-23 02:03:18 -07003437
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003438 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003440#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442/* It's optimal to keep kswapds on the same CPUs as their memory, but
3443 not required for correctness. So if the last cpu in a node goes
3444 away, we get changed to run anywhere: as the first one comes back,
3445 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003446static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3447 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003449 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003451 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003452 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003453 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303454 const struct cpumask *mask;
3455
3456 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003457
Rusty Russell3e597942009-01-01 10:12:24 +10303458 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003460 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 }
3462 }
3463 return NOTIFY_OK;
3464}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
Yasunori Goto3218ae12006-06-27 02:53:33 -07003466/*
3467 * This kswapd start function will be called by init and node-hot-add.
3468 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3469 */
3470int kswapd_run(int nid)
3471{
3472 pg_data_t *pgdat = NODE_DATA(nid);
3473 int ret = 0;
3474
3475 if (pgdat->kswapd)
3476 return 0;
3477
3478 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3479 if (IS_ERR(pgdat->kswapd)) {
3480 /* failure at boot is fatal */
3481 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003482 pr_err("Failed to start kswapd on node %d\n", nid);
3483 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003484 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003485 }
3486 return ret;
3487}
3488
David Rientjes8fe23e02009-12-14 17:58:33 -08003489/*
Jiang Liud8adde12012-07-11 14:01:52 -07003490 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003491 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003492 */
3493void kswapd_stop(int nid)
3494{
3495 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3496
Jiang Liud8adde12012-07-11 14:01:52 -07003497 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003498 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003499 NODE_DATA(nid)->kswapd = NULL;
3500 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003501}
3502
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503static int __init kswapd_init(void)
3504{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003505 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003506
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003508 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003509 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 hotcpu_notifier(cpu_callback, 0);
3511 return 0;
3512}
3513
3514module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003515
3516#ifdef CONFIG_NUMA
3517/*
3518 * Zone reclaim mode
3519 *
3520 * If non-zero call zone_reclaim when the number of free pages falls below
3521 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003522 */
3523int zone_reclaim_mode __read_mostly;
3524
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003525#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003526#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003527#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3528#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3529
Christoph Lameter9eeff232006-01-18 17:42:31 -08003530/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003531 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3532 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3533 * a zone.
3534 */
3535#define ZONE_RECLAIM_PRIORITY 4
3536
Christoph Lameter9eeff232006-01-18 17:42:31 -08003537/*
Christoph Lameter96146342006-07-03 00:24:13 -07003538 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3539 * occur.
3540 */
3541int sysctl_min_unmapped_ratio = 1;
3542
3543/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003544 * If the number of slab pages in a zone grows beyond this percentage then
3545 * slab reclaim needs to occur.
3546 */
3547int sysctl_min_slab_ratio = 5;
3548
Mel Gorman90afa5de2009-06-16 15:33:20 -07003549static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3550{
3551 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3552 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3553 zone_page_state(zone, NR_ACTIVE_FILE);
3554
3555 /*
3556 * It's possible for there to be more file mapped pages than
3557 * accounted for by the pages on the file LRU lists because
3558 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3559 */
3560 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3561}
3562
3563/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3564static long zone_pagecache_reclaimable(struct zone *zone)
3565{
3566 long nr_pagecache_reclaimable;
3567 long delta = 0;
3568
3569 /*
3570 * If RECLAIM_SWAP is set, then all file pages are considered
3571 * potentially reclaimable. Otherwise, we have to worry about
3572 * pages like swapcache and zone_unmapped_file_pages() provides
3573 * a better estimate
3574 */
3575 if (zone_reclaim_mode & RECLAIM_SWAP)
3576 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3577 else
3578 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3579
3580 /* If we can't clean pages, remove dirty pages from consideration */
3581 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3582 delta += zone_page_state(zone, NR_FILE_DIRTY);
3583
3584 /* Watch for any possible underflows due to delta */
3585 if (unlikely(delta > nr_pagecache_reclaimable))
3586 delta = nr_pagecache_reclaimable;
3587
3588 return nr_pagecache_reclaimable - delta;
3589}
3590
Christoph Lameter0ff38492006-09-25 23:31:52 -07003591/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003592 * Try to free up some pages from this zone through reclaim.
3593 */
Andrew Morton179e9632006-03-22 00:08:18 -08003594static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003595{
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08003596 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003597 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003598 struct task_struct *p = current;
3599 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003600 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003601 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003602 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003603 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003604 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003605 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3606 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3607 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003608 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003609 struct shrink_control shrink = {
3610 .gfp_mask = sc.gfp_mask,
3611 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003612 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003613
Christoph Lameter9eeff232006-01-18 17:42:31 -08003614 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003615 /*
3616 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3617 * and we also need to be able to write out pages for RECLAIM_WRITE
3618 * and RECLAIM_SWAP.
3619 */
3620 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003621 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003622 reclaim_state.reclaimed_slab = 0;
3623 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003624
Mel Gorman90afa5de2009-06-16 15:33:20 -07003625 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003626 /*
3627 * Free memory by calling shrink zone with increasing
3628 * priorities until we have enough memory freed.
3629 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003630 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003631 shrink_zone(zone, &sc);
3632 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003633 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003634
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003635 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3636 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003637 /*
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08003638 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003639 * many pages were freed in this zone. So we take the current
3640 * number of slab pages and shake the slab until it is reduced
3641 * by the same nr_pages that we used for reclaiming unmapped
3642 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003643 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003644 nodes_clear(shrink.nodes_to_scan);
3645 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003646 for (;;) {
3647 unsigned long lru_pages = zone_reclaimable_pages(zone);
3648
3649 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003650 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003651 break;
3652
3653 /* Freed enough memory */
3654 nr_slab_pages1 = zone_page_state(zone,
3655 NR_SLAB_RECLAIMABLE);
3656 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3657 break;
3658 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003659
3660 /*
3661 * Update nr_reclaimed by the number of slab pages we
3662 * reclaimed from this zone.
3663 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003664 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3665 if (nr_slab_pages1 < nr_slab_pages0)
3666 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003667 }
3668
Christoph Lameter9eeff232006-01-18 17:42:31 -08003669 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003670 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003671 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003672 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003673}
Andrew Morton179e9632006-03-22 00:08:18 -08003674
3675int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3676{
Andrew Morton179e9632006-03-22 00:08:18 -08003677 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003678 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003679
3680 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003681 * Zone reclaim reclaims unmapped file backed pages and
3682 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003683 *
Christoph Lameter96146342006-07-03 00:24:13 -07003684 * A small portion of unmapped file backed pages is needed for
3685 * file I/O otherwise pages read by file I/O will be immediately
3686 * thrown out if the zone is overallocated. So we do not reclaim
3687 * if less than a specified percentage of the zone is used by
3688 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003689 */
Mel Gorman90afa5de2009-06-16 15:33:20 -07003690 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3691 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003692 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003693
Lisa Du6e543d52013-09-11 14:22:36 -07003694 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003695 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003696
Andrew Morton179e9632006-03-22 00:08:18 -08003697 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003698 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003699 */
David Rientjesd773ed62007-10-16 23:26:01 -07003700 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003701 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003702
3703 /*
3704 * Only run zone reclaim on the local zone or on zones that do not
3705 * have associated processors. This will favor the local processor
3706 * over remote processors and spread off node memory allocations
3707 * as wide as possible.
3708 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003709 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003710 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003711 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003712
3713 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003714 return ZONE_RECLAIM_NOSCAN;
3715
David Rientjesd773ed62007-10-16 23:26:01 -07003716 ret = __zone_reclaim(zone, gfp_mask, order);
3717 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3718
Mel Gorman24cf725182009-06-16 15:33:23 -07003719 if (!ret)
3720 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3721
David Rientjesd773ed62007-10-16 23:26:01 -07003722 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003723}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003724#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003725
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003726/*
3727 * page_evictable - test whether a page is evictable
3728 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003729 *
3730 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003731 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003732 *
3733 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003734 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003735 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003736 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003737 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003738int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003739{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003740 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003741}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003742
Hugh Dickins850465792012-01-20 14:34:19 -08003743#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003744/**
Hugh Dickins24513262012-01-20 14:34:21 -08003745 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3746 * @pages: array of pages to check
3747 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003748 *
Hugh Dickins24513262012-01-20 14:34:21 -08003749 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003750 *
3751 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003752 */
Hugh Dickins24513262012-01-20 14:34:21 -08003753void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003754{
Johannes Weiner925b7672012-01-12 17:18:15 -08003755 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003756 struct zone *zone = NULL;
3757 int pgscanned = 0;
3758 int pgrescued = 0;
3759 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003760
Hugh Dickins24513262012-01-20 14:34:21 -08003761 for (i = 0; i < nr_pages; i++) {
3762 struct page *page = pages[i];
3763 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003764
Hugh Dickins24513262012-01-20 14:34:21 -08003765 pgscanned++;
3766 pagezone = page_zone(page);
3767 if (pagezone != zone) {
3768 if (zone)
3769 spin_unlock_irq(&zone->lru_lock);
3770 zone = pagezone;
3771 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003772 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003773 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003774
Hugh Dickins24513262012-01-20 14:34:21 -08003775 if (!PageLRU(page) || !PageUnevictable(page))
3776 continue;
3777
Hugh Dickins39b5f292012-10-08 16:33:18 -07003778 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003779 enum lru_list lru = page_lru_base_type(page);
3780
Sasha Levin309381fea2014-01-23 15:52:54 -08003781 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003782 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003783 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3784 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003785 pgrescued++;
3786 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003787 }
Hugh Dickins24513262012-01-20 14:34:21 -08003788
3789 if (zone) {
3790 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3791 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3792 spin_unlock_irq(&zone->lru_lock);
3793 }
Hugh Dickins850465792012-01-20 14:34:19 -08003794}
3795#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003796
Johannes Weiner264e56d2011-10-31 17:09:13 -07003797static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003798{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003799 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003800 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003801 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003802 "one, please send an email to [email protected].\n",
3803 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003804}
3805
3806/*
3807 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3808 * all nodes' unevictable lists for evictable pages
3809 */
3810unsigned long scan_unevictable_pages;
3811
3812int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003813 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003814 size_t *length, loff_t *ppos)
3815{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003816 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003817 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003818 scan_unevictable_pages = 0;
3819 return 0;
3820}
3821
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003822#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003823/*
3824 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3825 * a specified node's per zone unevictable lists for evictable pages.
3826 */
3827
Kay Sievers10fbcf42011-12-21 14:48:43 -08003828static ssize_t read_scan_unevictable_node(struct device *dev,
3829 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003830 char *buf)
3831{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003832 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003833 return sprintf(buf, "0\n"); /* always zero; should fit... */
3834}
3835
Kay Sievers10fbcf42011-12-21 14:48:43 -08003836static ssize_t write_scan_unevictable_node(struct device *dev,
3837 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003838 const char *buf, size_t count)
3839{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003840 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003841 return 1;
3842}
3843
3844
Kay Sievers10fbcf42011-12-21 14:48:43 -08003845static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003846 read_scan_unevictable_node,
3847 write_scan_unevictable_node);
3848
3849int scan_unevictable_register_node(struct node *node)
3850{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003851 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003852}
3853
3854void scan_unevictable_unregister_node(struct node *node)
3855{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003856 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003857}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003858#endif