blob: 2836b5373b2e7623a1143a98bf3997fa11865731 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
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.
akpm@osdl.orgb15e0902005-06-21 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 Piggin0fd0e6b2006-09-27 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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 while (!list_empty(page_list)) {
826 struct address_space *mapping;
827 struct page *page;
828 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700829 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700830 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 cond_resched();
833
834 page = lru_to_page(page_list);
835 list_del(&page->lru);
836
Nick Piggin529ae9a2008-08-02 12:01:03 +0200837 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 goto keep;
839
Sasha Levin309381fea2014-01-23 15:52:54 -0800840 VM_BUG_ON_PAGE(PageActive(page), page);
841 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
843 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800844
Hugh Dickins39b5f292012-10-08 16:33:18 -0700845 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700846 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700847
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700848 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800849 goto keep_locked;
850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 /* Double the slab pressure for mapped and swapcache pages */
852 if (page_mapped(page) || PageSwapCache(page))
853 sc->nr_scanned++;
854
Andy Whitcroftc661b072007-08-22 14:01:26 -0700855 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
856 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
857
Mel Gorman283aba92013-07-03 15:01:51 -0700858 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700859 * The number of dirty pages determines if a zone is marked
860 * reclaim_congested which affects wait_iff_congested. kswapd
861 * will stall and start writing pages if the tail of the LRU
862 * is all dirty unqueued pages.
863 */
864 page_check_dirty_writeback(page, &dirty, &writeback);
865 if (dirty || writeback)
866 nr_dirty++;
867
868 if (dirty && !writeback)
869 nr_unqueued_dirty++;
870
Mel Gormand04e8ac2013-07-03 15:02:03 -0700871 /*
872 * Treat this page as congested if the underlying BDI is or if
873 * pages are cycling through the LRU so quickly that the
874 * pages marked for immediate reclaim are making it to the
875 * end of the LRU a second time.
876 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700877 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700878 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
879 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700880 nr_congested++;
881
882 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700883 * If a page at the tail of the LRU is under writeback, there
884 * are three cases to consider.
885 *
886 * 1) If reclaim is encountering an excessive number of pages
887 * under writeback and this page is both under writeback and
888 * PageReclaim then it indicates that pages are being queued
889 * for IO but are being recycled through the LRU before the
890 * IO can complete. Waiting on the page itself risks an
891 * indefinite stall if it is impossible to writeback the
892 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700893 * note that the LRU is being scanned too quickly and the
894 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700895 *
896 * 2) Global reclaim encounters a page, memcg encounters a
897 * page that is not marked for immediate reclaim or
898 * the caller does not have __GFP_IO. In this case mark
899 * the page for immediate reclaim and continue scanning.
900 *
901 * __GFP_IO is checked because a loop driver thread might
902 * enter reclaim, and deadlock if it waits on a page for
903 * which it is needed to do the write (loop masks off
904 * __GFP_IO|__GFP_FS for this reason); but more thought
905 * would probably show more reasons.
906 *
907 * Don't require __GFP_FS, since we're not going into the
908 * FS, just waiting on its writeback completion. Worryingly,
909 * ext4 gfs2 and xfs allocate pages with
910 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
911 * may_enter_fs here is liable to OOM on them.
912 *
913 * 3) memcg encounters a page that is not already marked
914 * PageReclaim. memcg does not have any dirty pages
915 * throttling so we could easily OOM just because too many
916 * pages are in writeback and there is nothing else to
917 * reclaim. Wait for the writeback to complete.
918 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700919 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700920 /* Case 1 above */
921 if (current_is_kswapd() &&
922 PageReclaim(page) &&
923 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700924 nr_immediate++;
925 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700926
927 /* Case 2 above */
928 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700929 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
930 /*
931 * This is slightly racy - end_page_writeback()
932 * might have just cleared PageReclaim, then
933 * setting PageReclaim here end up interpreted
934 * as PageReadahead - but that does not matter
935 * enough to care. What we do want is for this
936 * page to have PageReclaim set next time memcg
937 * reclaim reaches the tests above, so it will
938 * then wait_on_page_writeback() to avoid OOM;
939 * and it's also appropriate in global reclaim.
940 */
941 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700942 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700943
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700944 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700945
946 /* Case 3 above */
947 } else {
948 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700949 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700950 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
Minchan Kim02c6de82012-10-08 16:31:55 -0700952 if (!force_reclaim)
953 references = page_check_references(page, sc);
954
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800955 switch (references) {
956 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800958 case PAGEREF_KEEP:
959 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800960 case PAGEREF_RECLAIM:
961 case PAGEREF_RECLAIM_CLEAN:
962 ; /* try to reclaim the page below */
963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 /*
966 * Anonymous process memory has backing store?
967 * Try to allocate it some swap space here.
968 */
Nick Pigginb291f002008-10-18 20:26:44 -0700969 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800970 if (!(sc->gfp_mask & __GFP_IO))
971 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700972 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800974 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
Mel Gormane2be15f2013-07-03 15:01:57 -0700976 /* Adding to swap updated mapping */
977 mapping = page_mapping(page);
978 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
980 /*
981 * The page is mapped into the page tables of one or more
982 * processes. Try to unmap it here.
983 */
984 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700985 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 case SWAP_FAIL:
987 goto activate_locked;
988 case SWAP_AGAIN:
989 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700990 case SWAP_MLOCK:
991 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 case SWAP_SUCCESS:
993 ; /* try to free the page below */
994 }
995 }
996
997 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700998 /*
999 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001000 * avoid risk of stack overflow but only writeback
1001 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001002 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001003 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001004 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -07001005 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001006 /*
1007 * Immediately reclaim when written back.
1008 * Similar in principal to deactivate_page()
1009 * except we already have the page isolated
1010 * and know it's dirty
1011 */
1012 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1013 SetPageReclaim(page);
1014
Mel Gormanee728862011-10-31 17:07:38 -07001015 goto keep_locked;
1016 }
1017
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001018 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001020 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001022 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 goto keep_locked;
1024
1025 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001026 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 case PAGE_KEEP:
1028 goto keep_locked;
1029 case PAGE_ACTIVATE:
1030 goto activate_locked;
1031 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001032 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001033 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001034 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001036
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 /*
1038 * A synchronous write - probably a ramdisk. Go
1039 * ahead and try to reclaim the page.
1040 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001041 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 goto keep;
1043 if (PageDirty(page) || PageWriteback(page))
1044 goto keep_locked;
1045 mapping = page_mapping(page);
1046 case PAGE_CLEAN:
1047 ; /* try to free the page below */
1048 }
1049 }
1050
1051 /*
1052 * If the page has buffers, try to free the buffer mappings
1053 * associated with this page. If we succeed we try to free
1054 * the page as well.
1055 *
1056 * We do this even if the page is PageDirty().
1057 * try_to_release_page() does not perform I/O, but it is
1058 * possible for a page to have PageDirty set, but it is actually
1059 * clean (all its buffers are clean). This happens if the
1060 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001061 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 * try_to_release_page() will discover that cleanness and will
1063 * drop the buffers and mark the page clean - it can be freed.
1064 *
1065 * Rarely, pages can have buffers and no ->mapping. These are
1066 * the pages which were not successfully invalidated in
1067 * truncate_complete_page(). We try to drop those buffers here
1068 * and if that worked, and the page is no longer mapped into
1069 * process address space (page_count == 1) it can be freed.
1070 * Otherwise, leave the page on the LRU so it is swappable.
1071 */
David Howells266cf652009-04-03 16:42:36 +01001072 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 if (!try_to_release_page(page, sc->gfp_mask))
1074 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001075 if (!mapping && page_count(page) == 1) {
1076 unlock_page(page);
1077 if (put_page_testzero(page))
1078 goto free_it;
1079 else {
1080 /*
1081 * rare race with speculative reference.
1082 * the speculative reference will free
1083 * this page shortly, so we may
1084 * increment nr_reclaimed here (and
1085 * leave it off the LRU).
1086 */
1087 nr_reclaimed++;
1088 continue;
1089 }
1090 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 }
1092
Johannes Weinera5289102014-04-03 14:47:51 -07001093 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001094 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095
Nick Piggina978d6f2008-10-18 20:26:58 -07001096 /*
1097 * At this point, we have no other references and there is
1098 * no way to pick any more up (removed from LRU, removed
1099 * from pagecache). Can use non-atomic bitops now (and
1100 * we obviously don't have to worry about waking up a process
1101 * waiting on the page lock, because there are no references.
1102 */
1103 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001104free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001105 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001106
1107 /*
1108 * Is there need to periodically free_page_list? It would
1109 * appear not as the counts should be low
1110 */
1111 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 continue;
1113
Nick Pigginb291f002008-10-18 20:26:44 -07001114cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001115 if (PageSwapCache(page))
1116 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001117 unlock_page(page);
1118 putback_lru_page(page);
1119 continue;
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001122 /* Not a candidate for swapping, so reclaim swap space. */
1123 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001124 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001125 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 SetPageActive(page);
1127 pgactivate++;
1128keep_locked:
1129 unlock_page(page);
1130keep:
1131 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001132 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001134
Johannes Weiner747db952014-08-08 14:19:24 -07001135 mem_cgroup_uncharge_list(&free_pages);
Mel Gormanb745bc82014-06-04 16:10:22 -07001136 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001137
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001139 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001140
Mel Gorman8e950282013-07-03 15:02:02 -07001141 *ret_nr_dirty += nr_dirty;
1142 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001143 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001144 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001145 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001146 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147}
1148
Minchan Kim02c6de82012-10-08 16:31:55 -07001149unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1150 struct list_head *page_list)
1151{
1152 struct scan_control sc = {
1153 .gfp_mask = GFP_KERNEL,
1154 .priority = DEF_PRIORITY,
1155 .may_unmap = 1,
1156 };
Mel Gorman8e950282013-07-03 15:02:02 -07001157 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001158 struct page *page, *next;
1159 LIST_HEAD(clean_pages);
1160
1161 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001162 if (page_is_file_cache(page) && !PageDirty(page) &&
1163 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001164 ClearPageActive(page);
1165 list_move(&page->lru, &clean_pages);
1166 }
1167 }
1168
1169 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001170 TTU_UNMAP|TTU_IGNORE_ACCESS,
1171 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001172 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001173 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001174 return ret;
1175}
1176
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001177/*
1178 * Attempt to remove the specified page from its LRU. Only take this page
1179 * if it is of the appropriate PageActive status. Pages which are being
1180 * freed elsewhere are also ignored.
1181 *
1182 * page: page to consider
1183 * mode: one of the LRU isolation modes defined above
1184 *
1185 * returns 0 on success, -ve errno on failure.
1186 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001187int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001188{
1189 int ret = -EINVAL;
1190
1191 /* Only take pages on the LRU. */
1192 if (!PageLRU(page))
1193 return ret;
1194
Minchan Kime46a2872012-10-08 16:33:48 -07001195 /* Compaction should not handle unevictable pages but CMA can do so */
1196 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001197 return ret;
1198
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001199 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001200
Mel Gormanc8244932012-01-12 17:19:38 -08001201 /*
1202 * To minimise LRU disruption, the caller can indicate that it only
1203 * wants to isolate pages it will be able to operate on without
1204 * blocking - clean pages for the most part.
1205 *
1206 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1207 * is used by reclaim when it is cannot write to backing storage
1208 *
1209 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1210 * that it is possible to migrate without blocking
1211 */
1212 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1213 /* All the caller can do on PageWriteback is block */
1214 if (PageWriteback(page))
1215 return ret;
1216
1217 if (PageDirty(page)) {
1218 struct address_space *mapping;
1219
1220 /* ISOLATE_CLEAN means only clean pages */
1221 if (mode & ISOLATE_CLEAN)
1222 return ret;
1223
1224 /*
1225 * Only pages without mappings or that have a
1226 * ->migratepage callback are possible to migrate
1227 * without blocking
1228 */
1229 mapping = page_mapping(page);
1230 if (mapping && !mapping->a_ops->migratepage)
1231 return ret;
1232 }
1233 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001234
Minchan Kimf80c0672011-10-31 17:06:55 -07001235 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1236 return ret;
1237
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001238 if (likely(get_page_unless_zero(page))) {
1239 /*
1240 * Be careful not to clear PageLRU until after we're
1241 * sure the page is not being freed elsewhere -- the
1242 * page release code relies on it.
1243 */
1244 ClearPageLRU(page);
1245 ret = 0;
1246 }
1247
1248 return ret;
1249}
1250
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001251/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 * zone->lru_lock is heavily contended. Some of the functions that
1253 * shrink the lists perform better by taking out a batch of pages
1254 * and working on them outside the LRU lock.
1255 *
1256 * For pagecache intensive workloads, this function is the hottest
1257 * spot in the kernel (apart from copy_*_user functions).
1258 *
1259 * Appropriate locks must be held before calling this function.
1260 *
1261 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001262 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001264 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001265 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001266 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001267 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 *
1269 * returns how many pages were moved onto *@dst.
1270 */
Andrew Morton69e05942006-03-22 00:08:19 -08001271static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001272 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001273 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001274 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001276 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001277 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001278 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001280 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001281 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001282 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001283
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 page = lru_to_page(src);
1285 prefetchw_prev_lru_page(page, src, flags);
1286
Sasha Levin309381fea2014-01-23 15:52:54 -08001287 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001288
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001289 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001290 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001291 nr_pages = hpage_nr_pages(page);
1292 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001293 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001294 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001295 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001296
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001297 case -EBUSY:
1298 /* else it is being freed elsewhere */
1299 list_move(&page->lru, src);
1300 continue;
1301
1302 default:
1303 BUG();
1304 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 }
1306
Hugh Dickinsf6260122012-01-12 17:20:06 -08001307 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001308 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1309 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 return nr_taken;
1311}
1312
Nick Piggin62695a82008-10-18 20:26:09 -07001313/**
1314 * isolate_lru_page - tries to isolate a page from its LRU list
1315 * @page: page to isolate from its LRU list
1316 *
1317 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1318 * vmstat statistic corresponding to whatever LRU list the page was on.
1319 *
1320 * Returns 0 if the page was removed from an LRU list.
1321 * Returns -EBUSY if the page was not on an LRU list.
1322 *
1323 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001324 * the active list, it will have PageActive set. If it was found on
1325 * the unevictable list, it will have the PageUnevictable bit set. That flag
1326 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001327 *
1328 * The vmstat statistic corresponding to the list on which the page was
1329 * found will be decremented.
1330 *
1331 * Restrictions:
1332 * (1) Must be called with an elevated refcount on the page. This is a
1333 * fundamentnal difference from isolate_lru_pages (which is called
1334 * without a stable reference).
1335 * (2) the lru_lock must not be held.
1336 * (3) interrupts must be enabled.
1337 */
1338int isolate_lru_page(struct page *page)
1339{
1340 int ret = -EBUSY;
1341
Sasha Levin309381fea2014-01-23 15:52:54 -08001342 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001343
Nick Piggin62695a82008-10-18 20:26:09 -07001344 if (PageLRU(page)) {
1345 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001346 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001347
1348 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001349 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001350 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001351 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001352 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001353 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001354 del_page_from_lru_list(page, lruvec, lru);
1355 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001356 }
1357 spin_unlock_irq(&zone->lru_lock);
1358 }
1359 return ret;
1360}
1361
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001362/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001363 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1364 * then get resheduled. When there are massive number of tasks doing page
1365 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1366 * the LRU list will go small and be scanned faster than necessary, leading to
1367 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001368 */
1369static int too_many_isolated(struct zone *zone, int file,
1370 struct scan_control *sc)
1371{
1372 unsigned long inactive, isolated;
1373
1374 if (current_is_kswapd())
1375 return 0;
1376
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001377 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001378 return 0;
1379
1380 if (file) {
1381 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1382 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1383 } else {
1384 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1385 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1386 }
1387
Fengguang Wu3cf23842012-12-18 14:23:31 -08001388 /*
1389 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1390 * won't get blocked by normal direct-reclaimers, forming a circular
1391 * deadlock.
1392 */
1393 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1394 inactive >>= 3;
1395
Rik van Riel35cd7812009-09-21 17:01:38 -07001396 return isolated > inactive;
1397}
1398
Mel Gorman66635622010-08-09 17:19:30 -07001399static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001400putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001401{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001402 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1403 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001404 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001405
Mel Gorman66635622010-08-09 17:19:30 -07001406 /*
1407 * Put back any unfreeable pages.
1408 */
Mel Gorman66635622010-08-09 17:19:30 -07001409 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001410 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001411 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001412
Sasha Levin309381fea2014-01-23 15:52:54 -08001413 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001414 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001415 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001416 spin_unlock_irq(&zone->lru_lock);
1417 putback_lru_page(page);
1418 spin_lock_irq(&zone->lru_lock);
1419 continue;
1420 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001421
1422 lruvec = mem_cgroup_page_lruvec(page, zone);
1423
Linus Torvalds7a608572011-01-17 14:42:19 -08001424 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001425 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001426 add_page_to_lru_list(page, lruvec, lru);
1427
Mel Gorman66635622010-08-09 17:19:30 -07001428 if (is_active_lru(lru)) {
1429 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001430 int numpages = hpage_nr_pages(page);
1431 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001432 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001433 if (put_page_testzero(page)) {
1434 __ClearPageLRU(page);
1435 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001436 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001437
1438 if (unlikely(PageCompound(page))) {
1439 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001440 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001441 (*get_compound_page_dtor(page))(page);
1442 spin_lock_irq(&zone->lru_lock);
1443 } else
1444 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001445 }
1446 }
Mel Gorman66635622010-08-09 17:19:30 -07001447
Hugh Dickins3f797682012-01-12 17:20:07 -08001448 /*
1449 * To save our caller's stack, now use input list for pages to free.
1450 */
1451 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001452}
1453
1454/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001455 * If a kernel thread (such as nfsd for loop-back mounts) services
1456 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1457 * In that case we should only throttle if the backing device it is
1458 * writing to is congested. In other cases it is safe to throttle.
1459 */
1460static int current_may_throttle(void)
1461{
1462 return !(current->flags & PF_LESS_THROTTLE) ||
1463 current->backing_dev_info == NULL ||
1464 bdi_write_congested(current->backing_dev_info);
1465}
1466
1467/*
Andrew Morton1742f192006-03-22 00:08:21 -08001468 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1469 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 */
Mel Gorman66635622010-08-09 17:19:30 -07001471static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001472shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001473 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474{
1475 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001476 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001477 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001478 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001479 unsigned long nr_dirty = 0;
1480 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001481 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001482 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001483 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001484 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001485 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001486 struct zone *zone = lruvec_zone(lruvec);
1487 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001488
Rik van Riel35cd7812009-09-21 17:01:38 -07001489 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001490 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001491
1492 /* We are about to die and free our memory. Return now. */
1493 if (fatal_signal_pending(current))
1494 return SWAP_CLUSTER_MAX;
1495 }
1496
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001498
1499 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001500 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001501 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001502 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001503
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001506 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1507 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001508
1509 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1510 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1511
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001512 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001513 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001514 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001515 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001516 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001517 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001518 }
Hillf Dantond563c052012-03-21 16:34:02 -07001519 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001520
Hillf Dantond563c052012-03-21 16:34:02 -07001521 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001522 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001523
Minchan Kim02c6de82012-10-08 16:31:55 -07001524 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001525 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1526 &nr_writeback, &nr_immediate,
1527 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001528
Hugh Dickins3f797682012-01-12 17:20:07 -08001529 spin_lock_irq(&zone->lru_lock);
1530
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001531 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001532
Ying Han904249a2012-04-25 16:01:48 -07001533 if (global_reclaim(sc)) {
1534 if (current_is_kswapd())
1535 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1536 nr_reclaimed);
1537 else
1538 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1539 nr_reclaimed);
1540 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001541
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001542 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001543
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001544 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001545
1546 spin_unlock_irq(&zone->lru_lock);
1547
Johannes Weiner747db952014-08-08 14:19:24 -07001548 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001549 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001550
Mel Gorman92df3a72011-10-31 17:07:56 -07001551 /*
1552 * If reclaim is isolating dirty pages under writeback, it implies
1553 * that the long-lived page allocation rate is exceeding the page
1554 * laundering rate. Either the global limits are not being effective
1555 * at throttling processes due to the page distribution throughout
1556 * zones or there is heavy usage of a slow backing device. The
1557 * only option is to throttle from reclaim context which is not ideal
1558 * as there is no guarantee the dirtying process is throttled in the
1559 * same way balance_dirty_pages() manages.
1560 *
Mel Gorman8e950282013-07-03 15:02:02 -07001561 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1562 * of pages under pages flagged for immediate reclaim and stall if any
1563 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001564 */
Mel Gorman918fc712013-07-08 16:00:25 -07001565 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001566 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001567
Mel Gormand43006d2013-07-03 15:01:50 -07001568 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001569 * memcg will stall in page writeback so only consider forcibly
1570 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001571 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001572 if (global_reclaim(sc)) {
1573 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001574 * Tag a zone as congested if all the dirty pages scanned were
1575 * backed by a congested BDI and wait_iff_congested will stall.
1576 */
1577 if (nr_dirty && nr_dirty == nr_congested)
1578 zone_set_flag(zone, ZONE_CONGESTED);
1579
1580 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001581 * If dirty pages are scanned that are not queued for IO, it
1582 * implies that flushers are not keeping up. In this case, flag
1583 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
Linus Torvaldsb738d762014-06-08 14:17:00 -07001584 * pages from reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001585 */
1586 if (nr_unqueued_dirty == nr_taken)
1587 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1588
1589 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001590 * If kswapd scans pages marked marked for immediate
1591 * reclaim and under writeback (nr_immediate), it implies
1592 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001593 * they are written so also forcibly stall.
1594 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001595 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001596 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001597 }
Mel Gormand43006d2013-07-03 15:01:50 -07001598
Mel Gorman8e950282013-07-03 15:02:02 -07001599 /*
1600 * Stall direct reclaim for IO completions if underlying BDIs or zone
1601 * is congested. Allow kswapd to continue until it starts encountering
1602 * unqueued dirty pages or cycling through the LRU too quickly.
1603 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001604 if (!sc->hibernation_mode && !current_is_kswapd() &&
1605 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001606 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1607
Mel Gormane11da5b2010-10-26 14:21:40 -07001608 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1609 zone_idx(zone),
1610 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001611 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001612 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001613 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614}
1615
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001616/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 * This moves pages from the active list to the inactive list.
1618 *
1619 * We move them the other way if the page is referenced by one or more
1620 * processes, from rmap.
1621 *
1622 * If the pages are mostly unmapped, the processing is fast and it is
1623 * appropriate to hold zone->lru_lock across the whole operation. But if
1624 * the pages are mapped, the processing is slow (page_referenced()) so we
1625 * should drop zone->lru_lock around each page. It's impossible to balance
1626 * this, so instead we remove the pages from the LRU while processing them.
1627 * It is safe to rely on PG_active against the non-LRU pages in here because
1628 * nobody will play with that bit on a non-LRU page.
1629 *
1630 * The downside is that we have to touch page->_count against each page.
1631 * But we had to alter page->flags anyway.
1632 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001633
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001634static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001635 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001636 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001637 enum lru_list lru)
1638{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001639 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001640 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001641 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001642 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001643
Wu Fengguang3eb41402009-06-16 15:33:13 -07001644 while (!list_empty(list)) {
1645 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001646 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001647
Sasha Levin309381fea2014-01-23 15:52:54 -08001648 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001649 SetPageLRU(page);
1650
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001651 nr_pages = hpage_nr_pages(page);
1652 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001653 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001654 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001655
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001656 if (put_page_testzero(page)) {
1657 __ClearPageLRU(page);
1658 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001659 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001660
1661 if (unlikely(PageCompound(page))) {
1662 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001663 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001664 (*get_compound_page_dtor(page))(page);
1665 spin_lock_irq(&zone->lru_lock);
1666 } else
1667 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001668 }
1669 }
1670 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1671 if (!is_active_lru(lru))
1672 __count_vm_events(PGDEACTIVATE, pgmoved);
1673}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001674
Hugh Dickinsf6260122012-01-12 17:20:06 -08001675static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001676 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001677 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001678 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001680 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001681 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001682 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001684 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001685 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001687 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001688 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001689 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001690 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001691 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692
1693 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001694
1695 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001696 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001697 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001698 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001699
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001701
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001702 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1703 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001704 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001705 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001706
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001707 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001708
Hugh Dickinsf6260122012-01-12 17:20:06 -08001709 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001710 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001711 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 spin_unlock_irq(&zone->lru_lock);
1713
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 while (!list_empty(&l_hold)) {
1715 cond_resched();
1716 page = lru_to_page(&l_hold);
1717 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001718
Hugh Dickins39b5f292012-10-08 16:33:18 -07001719 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001720 putback_lru_page(page);
1721 continue;
1722 }
1723
Mel Gormancc715d92012-03-21 16:34:00 -07001724 if (unlikely(buffer_heads_over_limit)) {
1725 if (page_has_private(page) && trylock_page(page)) {
1726 if (page_has_private(page))
1727 try_to_release_page(page, 0);
1728 unlock_page(page);
1729 }
1730 }
1731
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001732 if (page_referenced(page, 0, sc->target_mem_cgroup,
1733 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001734 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001735 /*
1736 * Identify referenced, file-backed active pages and
1737 * give them one more trip around the active list. So
1738 * that executable code get better chances to stay in
1739 * memory under moderate memory pressure. Anon pages
1740 * are not likely to be evicted by use-once streaming
1741 * IO, plus JVM can create lots of anon VM_EXEC pages,
1742 * so we ignore them here.
1743 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001744 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001745 list_add(&page->lru, &l_active);
1746 continue;
1747 }
1748 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001749
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001750 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 list_add(&page->lru, &l_inactive);
1752 }
1753
Andrew Mortonb5557492009-01-06 14:40:13 -08001754 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001755 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001756 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001757 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001758 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001759 * Count referenced pages from currently used mappings as rotated,
1760 * even though only some of them are actually re-activated. This
1761 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001762 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001763 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001764 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001765
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001766 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1767 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001768 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001769 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001770
Johannes Weiner747db952014-08-08 14:19:24 -07001771 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001772 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773}
1774
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001775#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001776static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001777{
1778 unsigned long active, inactive;
1779
1780 active = zone_page_state(zone, NR_ACTIVE_ANON);
1781 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1782
1783 if (inactive * zone->inactive_ratio < active)
1784 return 1;
1785
1786 return 0;
1787}
1788
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001789/**
1790 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001791 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001792 *
1793 * Returns true if the zone does not have enough inactive anon pages,
1794 * meaning some active anon pages need to be deactivated.
1795 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001796static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001797{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001798 /*
1799 * If we don't have swap space, anonymous page deactivation
1800 * is pointless.
1801 */
1802 if (!total_swap_pages)
1803 return 0;
1804
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001805 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001806 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001807
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001808 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001809}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001810#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001811static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001812{
1813 return 0;
1814}
1815#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001816
Rik van Riel56e49d22009-06-16 15:32:28 -07001817/**
1818 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001819 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001820 *
1821 * When the system is doing streaming IO, memory pressure here
1822 * ensures that active file pages get deactivated, until more
1823 * than half of the file pages are on the inactive list.
1824 *
1825 * Once we get to that situation, protect the system's working
1826 * set from being evicted by disabling active file page aging.
1827 *
1828 * This uses a different ratio than the anonymous pages, because
1829 * the page cache uses a use-once replacement algorithm.
1830 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001831static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001832{
Johannes Weinere3790142013-02-22 16:35:19 -08001833 unsigned long inactive;
1834 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001835
Johannes Weinere3790142013-02-22 16:35:19 -08001836 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1837 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1838
1839 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001840}
1841
Hugh Dickins75b00af2012-05-29 15:07:09 -07001842static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001843{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001844 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001845 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001846 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001847 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001848}
1849
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001850static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001851 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001852{
Rik van Rielb39415b2009-12-14 17:59:48 -08001853 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001854 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001855 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001856 return 0;
1857 }
1858
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001859 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001860}
1861
Johannes Weiner9a265112013-02-22 16:32:17 -08001862enum scan_balance {
1863 SCAN_EQUAL,
1864 SCAN_FRACT,
1865 SCAN_ANON,
1866 SCAN_FILE,
1867};
1868
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001870 * Determine how aggressively the anon and file LRU lists should be
1871 * scanned. The relative value of each set of LRU lists is determined
1872 * by looking at the fraction of the pages scanned we did rotate back
1873 * onto the active list instead of evict.
1874 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001875 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1876 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001877 */
Johannes Weiner02695172014-08-06 16:06:17 -07001878static void get_scan_count(struct lruvec *lruvec, int swappiness,
1879 struct scan_control *sc, unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001880{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001881 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001882 u64 fraction[2];
1883 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001884 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001885 unsigned long anon_prio, file_prio;
1886 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001887 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001888 bool force_scan = false;
1889 unsigned long ap, fp;
1890 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001891 bool some_scanned;
1892 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001893
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001894 /*
1895 * If the zone or memcg is small, nr[l] can be 0. This
1896 * results in no scanning on this priority and a potential
1897 * priority drop. Global direct reclaim can go to the next
1898 * zone and tends to have no problems. Global kswapd is for
1899 * zone balancing and it needs to scan a minimum amount. When
1900 * reclaiming for a memcg, a priority drop can cause high
1901 * latencies, so it's better to scan a minimum amount there as
1902 * well.
1903 */
Lisa Du6e543d52013-09-11 14:22:36 -07001904 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001905 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001906 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001907 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001908
1909 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001910 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001911 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001912 goto out;
1913 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001914
Johannes Weiner10316b32013-02-22 16:32:14 -08001915 /*
1916 * Global reclaim will swap to prevent OOM even with no
1917 * swappiness, but memcg users want to use this knob to
1918 * disable swapping for individual groups completely when
1919 * using the memory controller's swap limit feature would be
1920 * too expensive.
1921 */
Johannes Weiner02695172014-08-06 16:06:17 -07001922 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001923 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001924 goto out;
1925 }
1926
1927 /*
1928 * Do not apply any pressure balancing cleverness when the
1929 * system is close to OOM, scan both anon and file equally
1930 * (unless the swappiness setting disagrees with swapping).
1931 */
Johannes Weiner02695172014-08-06 16:06:17 -07001932 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001933 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001934 goto out;
1935 }
1936
Johannes Weiner11d16c22013-02-22 16:32:15 -08001937 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001938 * Prevent the reclaimer from falling into the cache trap: as
1939 * cache pages start out inactive, every cache fault will tip
1940 * the scan balance towards the file LRU. And as the file LRU
1941 * shrinks, so does the window for rotation from references.
1942 * This means we have a runaway feedback loop where a tiny
1943 * thrashing file LRU becomes infinitely more attractive than
1944 * anon pages. Try to detect this based on file LRU size.
1945 */
1946 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001947 unsigned long zonefile;
1948 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07001949
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001950 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1951 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1952 zone_page_state(zone, NR_INACTIVE_FILE);
1953
1954 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07001955 scan_balance = SCAN_ANON;
1956 goto out;
1957 }
1958 }
1959
1960 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001961 * There is enough inactive page cache, do not reclaim
1962 * anything from the anonymous working set right now.
1963 */
1964 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001965 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001966 goto out;
1967 }
1968
Johannes Weiner9a265112013-02-22 16:32:17 -08001969 scan_balance = SCAN_FRACT;
1970
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001971 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001972 * With swappiness at 100, anonymous and file have the same priority.
1973 * This scanning priority is essentially the inverse of IO cost.
1974 */
Johannes Weiner02695172014-08-06 16:06:17 -07001975 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001976 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001977
1978 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001979 * OK, so we have swap space and a fair amount of page cache
1980 * pages. We use the recently rotated / recently scanned
1981 * ratios to determine how valuable each cache is.
1982 *
1983 * Because workloads change over time (and to avoid overflow)
1984 * we keep these statistics as a floating average, which ends
1985 * up weighing recent references more than old ones.
1986 *
1987 * anon in [0], file in [1]
1988 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07001989
1990 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1991 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1992 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1993 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1994
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001995 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001996 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001997 reclaim_stat->recent_scanned[0] /= 2;
1998 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001999 }
2000
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002001 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002002 reclaim_stat->recent_scanned[1] /= 2;
2003 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002004 }
2005
2006 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002007 * The amount of pressure on anon vs file pages is inversely
2008 * proportional to the fraction of recently scanned pages on
2009 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002010 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002011 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002012 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002013
Satoru Moriyafe350042012-05-29 15:06:47 -07002014 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002015 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002016 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002017
Shaohua Li76a33fc2010-05-24 14:32:36 -07002018 fraction[0] = ap;
2019 fraction[1] = fp;
2020 denominator = ap + fp + 1;
2021out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002022 some_scanned = false;
2023 /* Only use force_scan on second pass. */
2024 for (pass = 0; !some_scanned && pass < 2; pass++) {
2025 for_each_evictable_lru(lru) {
2026 int file = is_file_lru(lru);
2027 unsigned long size;
2028 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002029
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002030 size = get_lru_size(lruvec, lru);
2031 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002032
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002033 if (!scan && pass && force_scan)
2034 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002035
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002036 switch (scan_balance) {
2037 case SCAN_EQUAL:
2038 /* Scan lists relative to size */
2039 break;
2040 case SCAN_FRACT:
2041 /*
2042 * Scan types proportional to swappiness and
2043 * their relative recent reclaim efficiency.
2044 */
2045 scan = div64_u64(scan * fraction[file],
2046 denominator);
2047 break;
2048 case SCAN_FILE:
2049 case SCAN_ANON:
2050 /* Scan one type exclusively */
2051 if ((scan_balance == SCAN_FILE) != file)
2052 scan = 0;
2053 break;
2054 default:
2055 /* Look ma, no brain */
2056 BUG();
2057 }
2058 nr[lru] = scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002059 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002060 * Skip the second pass and don't force_scan,
2061 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002062 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002063 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002064 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002065 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002066}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002067
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002068/*
2069 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2070 */
Johannes Weiner02695172014-08-06 16:06:17 -07002071static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2072 struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002073{
2074 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002075 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002076 unsigned long nr_to_scan;
2077 enum lru_list lru;
2078 unsigned long nr_reclaimed = 0;
2079 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2080 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002081 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002082
Johannes Weiner02695172014-08-06 16:06:17 -07002083 get_scan_count(lruvec, swappiness, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002084
Mel Gormane82e0562013-07-03 15:01:44 -07002085 /* Record the original scan target for proportional adjustments later */
2086 memcpy(targets, nr, sizeof(nr));
2087
Mel Gorman1a501902014-06-04 16:10:49 -07002088 /*
2089 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2090 * event that can occur when there is little memory pressure e.g.
2091 * multiple streaming readers/writers. Hence, we do not abort scanning
2092 * when the requested number of pages are reclaimed when scanning at
2093 * DEF_PRIORITY on the assumption that the fact we are direct
2094 * reclaiming implies that kswapd is not keeping up and it is best to
2095 * do a batch of work at once. For memcg reclaim one check is made to
2096 * abort proportional reclaim if either the file or anon lru has already
2097 * dropped to zero at the first pass.
2098 */
2099 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2100 sc->priority == DEF_PRIORITY);
2101
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002102 blk_start_plug(&plug);
2103 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2104 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002105 unsigned long nr_anon, nr_file, percentage;
2106 unsigned long nr_scanned;
2107
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002108 for_each_evictable_lru(lru) {
2109 if (nr[lru]) {
2110 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2111 nr[lru] -= nr_to_scan;
2112
2113 nr_reclaimed += shrink_list(lru, nr_to_scan,
2114 lruvec, sc);
2115 }
2116 }
Mel Gormane82e0562013-07-03 15:01:44 -07002117
2118 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2119 continue;
2120
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002121 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002122 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002123 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002124 * proportionally what was requested by get_scan_count(). We
2125 * stop reclaiming one LRU and reduce the amount scanning
2126 * proportional to the original scan target.
2127 */
2128 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2129 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2130
Mel Gorman1a501902014-06-04 16:10:49 -07002131 /*
2132 * It's just vindictive to attack the larger once the smaller
2133 * has gone to zero. And given the way we stop scanning the
2134 * smaller below, this makes sure that we only make one nudge
2135 * towards proportionality once we've got nr_to_reclaim.
2136 */
2137 if (!nr_file || !nr_anon)
2138 break;
2139
Mel Gormane82e0562013-07-03 15:01:44 -07002140 if (nr_file > nr_anon) {
2141 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2142 targets[LRU_ACTIVE_ANON] + 1;
2143 lru = LRU_BASE;
2144 percentage = nr_anon * 100 / scan_target;
2145 } else {
2146 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2147 targets[LRU_ACTIVE_FILE] + 1;
2148 lru = LRU_FILE;
2149 percentage = nr_file * 100 / scan_target;
2150 }
2151
2152 /* Stop scanning the smaller of the LRU */
2153 nr[lru] = 0;
2154 nr[lru + LRU_ACTIVE] = 0;
2155
2156 /*
2157 * Recalculate the other LRU scan count based on its original
2158 * scan target and the percentage scanning already complete
2159 */
2160 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2161 nr_scanned = targets[lru] - nr[lru];
2162 nr[lru] = targets[lru] * (100 - percentage) / 100;
2163 nr[lru] -= min(nr[lru], nr_scanned);
2164
2165 lru += LRU_ACTIVE;
2166 nr_scanned = targets[lru] - nr[lru];
2167 nr[lru] = targets[lru] * (100 - percentage) / 100;
2168 nr[lru] -= min(nr[lru], nr_scanned);
2169
2170 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002171 }
2172 blk_finish_plug(&plug);
2173 sc->nr_reclaimed += nr_reclaimed;
2174
2175 /*
2176 * Even if we did not try to evict anon pages at all, we want to
2177 * rebalance the anon lru active/inactive ratio.
2178 */
2179 if (inactive_anon_is_low(lruvec))
2180 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2181 sc, LRU_ACTIVE_ANON);
2182
2183 throttle_vm_writeout(sc->gfp_mask);
2184}
2185
Mel Gorman23b9da52012-05-29 15:06:20 -07002186/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002187static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002188{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002189 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002190 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002191 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002192 return true;
2193
2194 return false;
2195}
2196
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002197/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002198 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2199 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2200 * true if more pages should be reclaimed such that when the page allocator
2201 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2202 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002203 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002204static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002205 unsigned long nr_reclaimed,
2206 unsigned long nr_scanned,
2207 struct scan_control *sc)
2208{
2209 unsigned long pages_for_compaction;
2210 unsigned long inactive_lru_pages;
2211
2212 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002213 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002214 return false;
2215
Mel Gorman28765922011-02-25 14:44:20 -08002216 /* Consider stopping depending on scan and reclaim activity */
2217 if (sc->gfp_mask & __GFP_REPEAT) {
2218 /*
2219 * For __GFP_REPEAT allocations, stop reclaiming if the
2220 * full LRU list has been scanned and we are still failing
2221 * to reclaim pages. This full LRU scan is potentially
2222 * expensive but a __GFP_REPEAT caller really wants to succeed
2223 */
2224 if (!nr_reclaimed && !nr_scanned)
2225 return false;
2226 } else {
2227 /*
2228 * For non-__GFP_REPEAT allocations which can presumably
2229 * fail without consequence, stop if we failed to reclaim
2230 * any pages from the last SWAP_CLUSTER_MAX number of
2231 * pages that were scanned. This will return to the
2232 * caller faster at the risk reclaim/compaction and
2233 * the resulting allocation attempt fails
2234 */
2235 if (!nr_reclaimed)
2236 return false;
2237 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002238
2239 /*
2240 * If we have not reclaimed enough pages for compaction and the
2241 * inactive lists are large enough, continue reclaiming
2242 */
2243 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002244 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002245 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002246 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002247 if (sc->nr_reclaimed < pages_for_compaction &&
2248 inactive_lru_pages > pages_for_compaction)
2249 return true;
2250
2251 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002252 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002253 case COMPACT_PARTIAL:
2254 case COMPACT_CONTINUE:
2255 return false;
2256 default:
2257 return true;
2258 }
2259}
2260
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002261static bool shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002262{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002263 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002264 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002265
Johannes Weiner56600482012-01-12 17:17:59 -08002266 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002267 struct mem_cgroup *root = sc->target_mem_cgroup;
2268 struct mem_cgroup_reclaim_cookie reclaim = {
2269 .zone = zone,
2270 .priority = sc->priority,
2271 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002272 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002273
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002274 nr_reclaimed = sc->nr_reclaimed;
2275 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002276
Andrew Morton694fbc02013-09-24 15:27:37 -07002277 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2278 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002279 struct lruvec *lruvec;
Johannes Weiner02695172014-08-06 16:06:17 -07002280 int swappiness;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002281
2282 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002283 swappiness = mem_cgroup_swappiness(memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002284
Johannes Weiner02695172014-08-06 16:06:17 -07002285 shrink_lruvec(lruvec, swappiness, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002286
2287 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002288 * Direct reclaim and kswapd have to scan all memory
2289 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002290 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002291 *
2292 * Limit reclaim, on the other hand, only cares about
2293 * nr_to_reclaim pages to be reclaimed and it will
2294 * retry with decreasing priority if one round over the
2295 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002296 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002297 if (!global_reclaim(sc) &&
2298 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002299 mem_cgroup_iter_break(root, memcg);
2300 break;
2301 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002302 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2303 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002304
2305 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2306 sc->nr_scanned - nr_scanned,
2307 sc->nr_reclaimed - nr_reclaimed);
2308
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002309 if (sc->nr_reclaimed - nr_reclaimed)
2310 reclaimable = true;
2311
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002312 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2313 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002314
2315 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002316}
2317
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002318/* Returns true if compaction should go ahead for a high-order request */
Johannes Weiner0b064962014-08-06 16:06:12 -07002319static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002320{
2321 unsigned long balance_gap, watermark;
2322 bool watermark_ok;
2323
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002324 /*
2325 * Compaction takes time to run and there are potentially other
2326 * callers using the pages just freed. Continue reclaiming until
2327 * there is a buffer of free pages available to give compaction
2328 * a reasonable chance of completing and allocating the page
2329 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002330 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2331 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002332 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002333 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2334
2335 /*
2336 * If compaction is deferred, reclaim up to a point where
2337 * compaction will have a chance of success when re-enabled
2338 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002339 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002340 return watermark_ok;
2341
2342 /* If compaction is not ready to start, keep reclaiming */
Johannes Weiner0b064962014-08-06 16:06:12 -07002343 if (!compaction_suitable(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002344 return false;
2345
2346 return watermark_ok;
2347}
2348
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349/*
2350 * This is the direct reclaim path, for page-allocating processes. We only
2351 * try to reclaim pages from zones which will satisfy the caller's allocation
2352 * request.
2353 *
Mel Gorman41858962009-06-16 15:32:12 -07002354 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2355 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2357 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002358 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2359 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2360 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 * If a zone is deemed to be full of pinned pages then just give it a light
2363 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002364 *
2365 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002367static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368{
Mel Gormandd1a2392008-04-28 02:12:17 -07002369 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002370 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002371 unsigned long nr_soft_reclaimed;
2372 unsigned long nr_soft_scanned;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002373 unsigned long lru_pages = 0;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002374 struct reclaim_state *reclaim_state = current->reclaim_state;
Weijie Yang619d0d762014-04-07 15:36:59 -07002375 gfp_t orig_mask;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002376 struct shrink_control shrink = {
2377 .gfp_mask = sc->gfp_mask,
2378 };
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002379 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002380 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002381
Mel Gormancc715d92012-03-21 16:34:00 -07002382 /*
2383 * If the number of buffer_heads in the machine exceeds the maximum
2384 * allowed level, force direct reclaim to scan the highmem zone as
2385 * highmem pages could be pinning lowmem pages storing buffer_heads
2386 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002387 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002388 if (buffer_heads_over_limit)
2389 sc->gfp_mask |= __GFP_HIGHMEM;
2390
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002391 nodes_clear(shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002392
Mel Gormand4debc62010-08-09 17:19:29 -07002393 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2394 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002395 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002397 /*
2398 * Take care memory controller reclaiming has small influence
2399 * to global LRU.
2400 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002401 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002402 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2403 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002404
2405 lru_pages += zone_reclaimable_pages(zone);
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002406 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002407
Lisa Du6e543d52013-09-11 14:22:36 -07002408 if (sc->priority != DEF_PRIORITY &&
2409 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002410 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002411
2412 /*
2413 * If we already have plenty of memory free for
2414 * compaction in this zone, don't free any more.
2415 * Even though compaction is invoked for any
2416 * non-zero order, only frequent costly order
2417 * reclamation is disruptive enough to become a
2418 * noticeable problem, like transparent huge
2419 * page allocations.
2420 */
2421 if (IS_ENABLED(CONFIG_COMPACTION) &&
2422 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2423 zonelist_zone_idx(z) <= requested_highidx &&
2424 compaction_ready(zone, sc->order)) {
2425 sc->compaction_ready = true;
2426 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002427 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002428
Andrew Morton0608f432013-09-24 15:27:41 -07002429 /*
2430 * This steals pages from memory cgroups over softlimit
2431 * and returns the number of reclaimed pages and
2432 * scanned pages. This works for global memory pressure
2433 * and balancing, not for a memcg's limit.
2434 */
2435 nr_soft_scanned = 0;
2436 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2437 sc->order, sc->gfp_mask,
2438 &nr_soft_scanned);
2439 sc->nr_reclaimed += nr_soft_reclaimed;
2440 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002441 if (nr_soft_reclaimed)
2442 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002443 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002444 }
Nick Piggin408d8542006-09-25 23:31:27 -07002445
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002446 if (shrink_zone(zone, sc))
2447 reclaimable = true;
2448
2449 if (global_reclaim(sc) &&
2450 !reclaimable && zone_reclaimable(zone))
2451 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 }
Mel Gormane0c23272011-10-31 17:09:33 -07002453
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002454 /*
2455 * Don't shrink slabs when reclaiming memory from over limit cgroups
2456 * but do shrink slab at least once when aborting reclaim for
2457 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2458 * pages.
2459 */
2460 if (global_reclaim(sc)) {
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002461 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002462 if (reclaim_state) {
2463 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2464 reclaim_state->reclaimed_slab = 0;
2465 }
2466 }
2467
Weijie Yang619d0d762014-04-07 15:36:59 -07002468 /*
2469 * Restore to original mask to avoid the impact on the caller if we
2470 * promoted it to __GFP_HIGHMEM.
2471 */
2472 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002473
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002474 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002476
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477/*
2478 * This is the main entry point to direct page reclaim.
2479 *
2480 * If a full scan of the inactive list fails to free enough memory then we
2481 * are "out of memory" and something needs to be killed.
2482 *
2483 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2484 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002485 * caller can't do much about. We kick the writeback threads and take explicit
2486 * naps in the hope that some of these pages can be written. But if the
2487 * allocating task holds filesystem locks which prevent writeout this might not
2488 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002489 *
2490 * returns: 0, if no pages reclaimed
2491 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 */
Mel Gormandac1d272008-04-28 02:12:12 -07002493static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002494 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495{
Andrew Morton69e05942006-03-22 00:08:19 -08002496 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002497 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002498 bool zones_reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Keika Kobayashi873b4772008-07-25 01:48:52 -07002500 delayacct_freepages_start();
2501
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002502 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002503 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002505 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002506 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2507 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002508 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002509 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002510
Balbir Singh66e17072008-02-07 00:13:56 -08002511 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002512 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002513 break;
2514
2515 if (sc->compaction_ready)
2516 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517
2518 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002519 * If we're getting trouble reclaiming, start doing
2520 * writepage even in laptop mode.
2521 */
2522 if (sc->priority < DEF_PRIORITY - 2)
2523 sc->may_writepage = 1;
2524
2525 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 * Try to write back as many pages as we just scanned. This
2527 * tends to cause slow streaming writers to write data to the
2528 * disk smoothly, at the dirtying rate, which is nice. But
2529 * that's undesirable in laptop mode, where we *want* lumpy
2530 * writeout. So in laptop mode, write out the whole world.
2531 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002532 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2533 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002534 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2535 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002536 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002538 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002539
Keika Kobayashi873b4772008-07-25 01:48:52 -07002540 delayacct_freepages_end();
2541
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002542 if (sc->nr_reclaimed)
2543 return sc->nr_reclaimed;
2544
Mel Gorman0cee34f2012-01-12 17:19:49 -08002545 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002546 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002547 return 1;
2548
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002549 /* Any of the zones still reclaimable? Don't OOM. */
2550 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002551 return 1;
2552
2553 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554}
2555
Mel Gorman55150612012-07-31 16:44:35 -07002556static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2557{
2558 struct zone *zone;
2559 unsigned long pfmemalloc_reserve = 0;
2560 unsigned long free_pages = 0;
2561 int i;
2562 bool wmark_ok;
2563
2564 for (i = 0; i <= ZONE_NORMAL; i++) {
2565 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002566 if (!populated_zone(zone))
2567 continue;
2568
Mel Gorman55150612012-07-31 16:44:35 -07002569 pfmemalloc_reserve += min_wmark_pages(zone);
2570 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2571 }
2572
Mel Gorman675becc2014-06-04 16:07:35 -07002573 /* If there are no reserves (unexpected config) then do not throttle */
2574 if (!pfmemalloc_reserve)
2575 return true;
2576
Mel Gorman55150612012-07-31 16:44:35 -07002577 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2578
2579 /* kswapd must be awake if processes are being throttled */
2580 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2581 pgdat->classzone_idx = min(pgdat->classzone_idx,
2582 (enum zone_type)ZONE_NORMAL);
2583 wake_up_interruptible(&pgdat->kswapd_wait);
2584 }
2585
2586 return wmark_ok;
2587}
2588
2589/*
2590 * Throttle direct reclaimers if backing storage is backed by the network
2591 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2592 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002593 * when the low watermark is reached.
2594 *
2595 * Returns true if a fatal signal was delivered during throttling. If this
2596 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002597 */
Mel Gorman50694c22012-11-26 16:29:48 -08002598static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002599 nodemask_t *nodemask)
2600{
Mel Gorman675becc2014-06-04 16:07:35 -07002601 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002602 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002603 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002604
2605 /*
2606 * Kernel threads should not be throttled as they may be indirectly
2607 * responsible for cleaning pages necessary for reclaim to make forward
2608 * progress. kjournald for example may enter direct reclaim while
2609 * committing a transaction where throttling it could forcing other
2610 * processes to block on log_wait_commit().
2611 */
2612 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002613 goto out;
2614
2615 /*
2616 * If a fatal signal is pending, this process should not throttle.
2617 * It should return quickly so it can exit and free its memory
2618 */
2619 if (fatal_signal_pending(current))
2620 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002621
Mel Gorman675becc2014-06-04 16:07:35 -07002622 /*
2623 * Check if the pfmemalloc reserves are ok by finding the first node
2624 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2625 * GFP_KERNEL will be required for allocating network buffers when
2626 * swapping over the network so ZONE_HIGHMEM is unusable.
2627 *
2628 * Throttling is based on the first usable node and throttled processes
2629 * wait on a queue until kswapd makes progress and wakes them. There
2630 * is an affinity then between processes waking up and where reclaim
2631 * progress has been made assuming the process wakes on the same node.
2632 * More importantly, processes running on remote nodes will not compete
2633 * for remote pfmemalloc reserves and processes on different nodes
2634 * should make reasonable progress.
2635 */
2636 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2637 gfp_mask, nodemask) {
2638 if (zone_idx(zone) > ZONE_NORMAL)
2639 continue;
2640
2641 /* Throttle based on the first usable node */
2642 pgdat = zone->zone_pgdat;
2643 if (pfmemalloc_watermark_ok(pgdat))
2644 goto out;
2645 break;
2646 }
2647
2648 /* If no zone was usable by the allocation flags then do not throttle */
2649 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002650 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002651
Mel Gorman68243e72012-07-31 16:44:39 -07002652 /* Account for the throttling */
2653 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2654
Mel Gorman55150612012-07-31 16:44:35 -07002655 /*
2656 * If the caller cannot enter the filesystem, it's possible that it
2657 * is due to the caller holding an FS lock or performing a journal
2658 * transaction in the case of a filesystem like ext[3|4]. In this case,
2659 * it is not safe to block on pfmemalloc_wait as kswapd could be
2660 * blocked waiting on the same lock. Instead, throttle for up to a
2661 * second before continuing.
2662 */
2663 if (!(gfp_mask & __GFP_FS)) {
2664 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2665 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002666
2667 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002668 }
2669
2670 /* Throttle until kswapd wakes the process */
2671 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2672 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002673
2674check_pending:
2675 if (fatal_signal_pending(current))
2676 return true;
2677
2678out:
2679 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002680}
2681
Mel Gormandac1d272008-04-28 02:12:12 -07002682unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002683 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002684{
Mel Gorman33906bc2010-08-09 17:19:16 -07002685 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002686 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002687 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002688 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2689 .order = order,
2690 .nodemask = nodemask,
2691 .priority = DEF_PRIORITY,
2692 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002693 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002694 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002695 };
2696
Mel Gorman55150612012-07-31 16:44:35 -07002697 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002698 * Do not enter reclaim if fatal signal was delivered while throttled.
2699 * 1 is returned so that the page allocator does not OOM kill at this
2700 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002701 */
Mel Gorman50694c22012-11-26 16:29:48 -08002702 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002703 return 1;
2704
Mel Gorman33906bc2010-08-09 17:19:16 -07002705 trace_mm_vmscan_direct_reclaim_begin(order,
2706 sc.may_writepage,
2707 gfp_mask);
2708
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002709 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002710
2711 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2712
2713 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002714}
2715
Andrew Mortonc255a452012-07-31 16:43:02 -07002716#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002717
Johannes Weiner72835c82012-01-12 17:18:32 -08002718unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002719 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002720 struct zone *zone,
2721 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002722{
2723 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002724 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002725 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002726 .may_writepage = !laptop_mode,
2727 .may_unmap = 1,
2728 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002729 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002730 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner02695172014-08-06 16:06:17 -07002731 int swappiness = mem_cgroup_swappiness(memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002732
Balbir Singh4e416952009-09-23 15:56:39 -07002733 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2734 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002735
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002736 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002737 sc.may_writepage,
2738 sc.gfp_mask);
2739
Balbir Singh4e416952009-09-23 15:56:39 -07002740 /*
2741 * NOTE: Although we can get the priority field, using it
2742 * here is not a good idea, since it limits the pages we can scan.
2743 * if we don't reclaim here, the shrink_zone from balance_pgdat
2744 * will pick up pages from other mem cgroup's as well. We hack
2745 * the priority and make it zero.
2746 */
Johannes Weiner02695172014-08-06 16:06:17 -07002747 shrink_lruvec(lruvec, swappiness, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002748
2749 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2750
Ying Han0ae5e892011-05-26 16:25:25 -07002751 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002752 return sc.nr_reclaimed;
2753}
2754
Johannes Weiner72835c82012-01-12 17:18:32 -08002755unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002756 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002757 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002758{
Balbir Singh4e416952009-09-23 15:56:39 -07002759 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002760 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002761 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002762 struct scan_control sc = {
Johannes Weineree814fe2014-08-06 16:06:19 -07002763 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2764 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2765 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2766 .target_mem_cgroup = memcg,
2767 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002768 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002769 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002770 .may_swap = !noswap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002771 };
Balbir Singh66e17072008-02-07 00:13:56 -08002772
Ying Han889976d2011-05-26 16:25:33 -07002773 /*
2774 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2775 * take care of from where we get pages. So the node where we start the
2776 * scan does not need to be the current node.
2777 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002778 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002779
2780 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002781
2782 trace_mm_vmscan_memcg_reclaim_begin(0,
2783 sc.may_writepage,
2784 sc.gfp_mask);
2785
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002786 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002787
2788 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2789
2790 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002791}
2792#endif
2793
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002794static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002795{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002796 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002797
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002798 if (!total_swap_pages)
2799 return;
2800
2801 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2802 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002803 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002804
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002805 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002806 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002807 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002808
2809 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2810 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002811}
2812
Johannes Weiner60cefed2012-11-29 13:54:23 -08002813static bool zone_balanced(struct zone *zone, int order,
2814 unsigned long balance_gap, int classzone_idx)
2815{
2816 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2817 balance_gap, classzone_idx, 0))
2818 return false;
2819
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002820 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2821 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002822 return false;
2823
2824 return true;
2825}
2826
Mel Gorman1741c872011-01-13 15:46:21 -08002827/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002828 * pgdat_balanced() is used when checking if a node is balanced.
2829 *
2830 * For order-0, all zones must be balanced!
2831 *
2832 * For high-order allocations only zones that meet watermarks and are in a
2833 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2834 * total of balanced pages must be at least 25% of the zones allowed by
2835 * classzone_idx for the node to be considered balanced. Forcing all zones to
2836 * be balanced for high orders can cause excessive reclaim when there are
2837 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002838 * The choice of 25% is due to
2839 * o a 16M DMA zone that is balanced will not balance a zone on any
2840 * reasonable sized machine
2841 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002842 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002843 * would need to be at least 256M for it to be balance a whole node.
2844 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2845 * to balance a node on its own. These seemed like reasonable ratios.
2846 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002847static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002848{
Jiang Liub40da042013-02-22 16:33:52 -08002849 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002850 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002851 int i;
2852
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002853 /* Check the watermark levels */
2854 for (i = 0; i <= classzone_idx; i++) {
2855 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002856
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002857 if (!populated_zone(zone))
2858 continue;
2859
Jiang Liub40da042013-02-22 16:33:52 -08002860 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002861
2862 /*
2863 * A special case here:
2864 *
2865 * balance_pgdat() skips over all_unreclaimable after
2866 * DEF_PRIORITY. Effectively, it considers them balanced so
2867 * they must be considered balanced here as well!
2868 */
Lisa Du6e543d52013-09-11 14:22:36 -07002869 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002870 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002871 continue;
2872 }
2873
2874 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002875 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002876 else if (!order)
2877 return false;
2878 }
2879
2880 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002881 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002882 else
2883 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002884}
2885
Mel Gorman55150612012-07-31 16:44:35 -07002886/*
2887 * Prepare kswapd for sleeping. This verifies that there are no processes
2888 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2889 *
2890 * Returns true if kswapd is ready to sleep
2891 */
2892static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002893 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002894{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002895 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2896 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002897 return false;
2898
2899 /*
2900 * There is a potential race between when kswapd checks its watermarks
2901 * and a process gets throttled. There is also a potential race if
2902 * processes get throttled, kswapd wakes, a large process exits therby
2903 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2904 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2905 * so wake them now if necessary. If necessary, processes will wake
2906 * kswapd and get throttled again
2907 */
2908 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2909 wake_up(&pgdat->pfmemalloc_wait);
2910 return false;
2911 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002912
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002913 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002914}
2915
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916/*
Mel Gorman75485362013-07-03 15:01:42 -07002917 * kswapd shrinks the zone by the number of pages required to reach
2918 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002919 *
2920 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002921 * reclaim or if the lack of progress was due to pages under writeback.
2922 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002923 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002924static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002925 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002926 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002927 unsigned long lru_pages,
2928 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002929{
Mel Gorman7c954f62013-07-03 15:01:54 -07002930 int testorder = sc->order;
2931 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002932 struct reclaim_state *reclaim_state = current->reclaim_state;
2933 struct shrink_control shrink = {
2934 .gfp_mask = sc->gfp_mask,
2935 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002936 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002937
2938 /* Reclaim above the high watermark. */
2939 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002940
2941 /*
2942 * Kswapd reclaims only single pages with compaction enabled. Trying
2943 * too hard to reclaim until contiguous free pages have become
2944 * available can hurt performance by evicting too much useful data
2945 * from memory. Do not reclaim more than needed for compaction.
2946 */
2947 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2948 compaction_suitable(zone, sc->order) !=
2949 COMPACT_SKIPPED)
2950 testorder = 0;
2951
2952 /*
2953 * We put equal pressure on every zone, unless one zone has way too
2954 * many pages free already. The "too many pages" is defined as the
2955 * high wmark plus a "gap" where the gap is either the low
2956 * watermark or 1% of the zone, whichever is smaller.
2957 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002958 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2959 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07002960
2961 /*
2962 * If there is no low memory pressure or the zone is balanced then no
2963 * reclaim is necessary
2964 */
2965 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2966 if (!lowmem_pressure && zone_balanced(zone, testorder,
2967 balance_gap, classzone_idx))
2968 return true;
2969
Mel Gorman75485362013-07-03 15:01:42 -07002970 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002971 nodes_clear(shrink.nodes_to_scan);
2972 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002973
2974 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002975 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002976 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2977
Mel Gorman2ab44f42013-07-03 15:01:47 -07002978 /* Account for the number of pages attempted to reclaim */
2979 *nr_attempted += sc->nr_to_reclaim;
2980
Mel Gorman283aba92013-07-03 15:01:51 -07002981 zone_clear_flag(zone, ZONE_WRITEBACK);
2982
Mel Gorman7c954f62013-07-03 15:01:54 -07002983 /*
2984 * If a zone reaches its high watermark, consider it to be no longer
2985 * congested. It's possible there are dirty pages backed by congested
2986 * BDIs but as pressure is relieved, speculatively avoid congestion
2987 * waits.
2988 */
Lisa Du6e543d52013-09-11 14:22:36 -07002989 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002990 zone_balanced(zone, testorder, 0, classzone_idx)) {
2991 zone_clear_flag(zone, ZONE_CONGESTED);
2992 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2993 }
2994
Mel Gormanb8e83b92013-07-03 15:01:45 -07002995 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002996}
2997
2998/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003000 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003002 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 *
3004 * There is special handling here for zones which are full of pinned pages.
3005 * This can happen if the pages are all mlocked, or if they are all used by
3006 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3007 * What we do is to detect the case where all pages in the zone have been
3008 * scanned twice and there has been zero successful reclaim. Mark the zone as
3009 * dead and from now on, only perform a short scan. Basically we're polling
3010 * the zone for when the problem goes away.
3011 *
3012 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003013 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3014 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3015 * lower zones regardless of the number of free pages in the lower zones. This
3016 * interoperates with the page allocator fallback scheme to ensure that aging
3017 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 */
Mel Gorman99504742011-01-13 15:46:20 -08003019static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003020 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003023 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003024 unsigned long nr_soft_reclaimed;
3025 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003026 struct scan_control sc = {
3027 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003028 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003029 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003030 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003031 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003032 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003033 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003034 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003036 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003038 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003039 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003040 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003041
3042 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003044 /*
3045 * Scan in the highmem->dma direction for the highest
3046 * zone which needs scanning
3047 */
3048 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3049 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003051 if (!populated_zone(zone))
3052 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053
Lisa Du6e543d52013-09-11 14:22:36 -07003054 if (sc.priority != DEF_PRIORITY &&
3055 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003056 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057
Rik van Riel556adec2008-10-18 20:26:34 -07003058 /*
3059 * Do some background aging of the anon list, to give
3060 * pages a chance to be referenced before reclaiming.
3061 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003062 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003063
Mel Gormancc715d92012-03-21 16:34:00 -07003064 /*
3065 * If the number of buffer_heads in the machine
3066 * exceeds the maximum allowed level and this node
3067 * has a highmem zone, force kswapd to reclaim from
3068 * it to relieve lowmem pressure.
3069 */
3070 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3071 end_zone = i;
3072 break;
3073 }
3074
Johannes Weiner60cefed2012-11-29 13:54:23 -08003075 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003076 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003077 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003078 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003079 /*
3080 * If balanced, clear the dirty and congested
3081 * flags
3082 */
Shaohua Li439423f2011-08-25 15:59:12 -07003083 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003084 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003087
Mel Gormanb8e83b92013-07-03 15:01:45 -07003088 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003089 goto out;
3090
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 for (i = 0; i <= end_zone; i++) {
3092 struct zone *zone = pgdat->node_zones + i;
3093
Mel Gorman2ab44f42013-07-03 15:01:47 -07003094 if (!populated_zone(zone))
3095 continue;
3096
Wu Fengguangadea02a2009-09-21 17:01:42 -07003097 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003098
3099 /*
3100 * If any zone is currently balanced then kswapd will
3101 * not call compaction as it is expected that the
3102 * necessary pages are already available.
3103 */
3104 if (pgdat_needs_compaction &&
3105 zone_watermark_ok(zone, order,
3106 low_wmark_pages(zone),
3107 *classzone_idx, 0))
3108 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 }
3110
3111 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003112 * If we're getting trouble reclaiming, start doing writepage
3113 * even in laptop mode.
3114 */
3115 if (sc.priority < DEF_PRIORITY - 2)
3116 sc.may_writepage = 1;
3117
3118 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 * Now scan the zone in the dma->highmem direction, stopping
3120 * at the last zone which needs scanning.
3121 *
3122 * We do this because the page allocator works in the opposite
3123 * direction. This prevents the page allocator from allocating
3124 * pages behind kswapd's direction of progress, which would
3125 * cause too much scanning of the lower zones.
3126 */
3127 for (i = 0; i <= end_zone; i++) {
3128 struct zone *zone = pgdat->node_zones + i;
3129
Con Kolivasf3fe6512006-01-06 00:11:15 -08003130 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 continue;
3132
Lisa Du6e543d52013-09-11 14:22:36 -07003133 if (sc.priority != DEF_PRIORITY &&
3134 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 continue;
3136
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003138
Andrew Morton0608f432013-09-24 15:27:41 -07003139 nr_soft_scanned = 0;
3140 /*
3141 * Call soft limit reclaim before calling shrink_zone.
3142 */
3143 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3144 order, sc.gfp_mask,
3145 &nr_soft_scanned);
3146 sc.nr_reclaimed += nr_soft_reclaimed;
3147
Rik van Riel32a43302007-10-16 01:24:50 -07003148 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003149 * There should be no need to raise the scanning
3150 * priority if enough pages are already being scanned
3151 * that that high watermark would be met at 100%
3152 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003153 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003154 if (kswapd_shrink_zone(zone, end_zone, &sc,
3155 lru_pages, &nr_attempted))
3156 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 }
Mel Gorman55150612012-07-31 16:44:35 -07003158
3159 /*
3160 * If the low watermark is met there is no need for processes
3161 * to be throttled on pfmemalloc_wait as they should not be
3162 * able to safely make forward progress. Wake them
3163 */
3164 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3165 pfmemalloc_watermark_ok(pgdat))
3166 wake_up(&pgdat->pfmemalloc_wait);
3167
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003169 * Fragmentation may mean that the system cannot be rebalanced
3170 * for high-order allocations in all zones. If twice the
3171 * allocation size has been reclaimed and the zones are still
3172 * not balanced then recheck the watermarks at order-0 to
3173 * prevent kswapd reclaiming excessively. Assume that a
3174 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003176 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003177 order = sc.order = 0;
3178
Mel Gormanb8e83b92013-07-03 15:01:45 -07003179 /* Check if kswapd should be suspending */
3180 if (try_to_freeze() || kthread_should_stop())
3181 break;
3182
3183 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003184 * Compact if necessary and kswapd is reclaiming at least the
3185 * high watermark number of pages as requsted
3186 */
3187 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3188 compact_pgdat(pgdat, order);
3189
3190 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003191 * Raise priority if scanning rate is too low or there was no
3192 * progress in reclaiming pages
3193 */
3194 if (raise_priority || !sc.nr_reclaimed)
3195 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003196 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003197 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198
Mel Gormanb8e83b92013-07-03 15:01:45 -07003199out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003200 /*
Mel Gorman55150612012-07-31 16:44:35 -07003201 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003202 * makes a decision on the order we were last reclaiming at. However,
3203 * if another caller entered the allocator slow path while kswapd
3204 * was awake, order will remain at the higher level
3205 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003206 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003207 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208}
3209
Mel Gormandc83edd2011-01-13 15:46:26 -08003210static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003211{
3212 long remaining = 0;
3213 DEFINE_WAIT(wait);
3214
3215 if (freezing(current) || kthread_should_stop())
3216 return;
3217
3218 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3219
3220 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003221 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003222 remaining = schedule_timeout(HZ/10);
3223 finish_wait(&pgdat->kswapd_wait, &wait);
3224 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3225 }
3226
3227 /*
3228 * After a short sleep, check if it was a premature sleep. If not, then
3229 * go fully to sleep until explicitly woken up.
3230 */
Mel Gorman55150612012-07-31 16:44:35 -07003231 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003232 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3233
3234 /*
3235 * vmstat counters are not perfectly accurate and the estimated
3236 * value for counters such as NR_FREE_PAGES can deviate from the
3237 * true value by nr_online_cpus * threshold. To avoid the zone
3238 * watermarks being breached while under pressure, we reduce the
3239 * per-cpu vmstat threshold while kswapd is awake and restore
3240 * them before going back to sleep.
3241 */
3242 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003243
Mel Gorman62997022012-10-08 16:32:47 -07003244 /*
3245 * Compaction records what page blocks it recently failed to
3246 * isolate pages from and skips them in the future scanning.
3247 * When kswapd is going to sleep, it is reasonable to assume
3248 * that pages and compaction may succeed so reset the cache.
3249 */
3250 reset_isolation_suitable(pgdat);
3251
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003252 if (!kthread_should_stop())
3253 schedule();
3254
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003255 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3256 } else {
3257 if (remaining)
3258 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3259 else
3260 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3261 }
3262 finish_wait(&pgdat->kswapd_wait, &wait);
3263}
3264
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265/*
3266 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003267 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 *
3269 * This basically trickles out pages so that we have _some_
3270 * free memory available even if there is no other activity
3271 * that frees anything up. This is needed for things like routing
3272 * etc, where we otherwise might have all activity going on in
3273 * asynchronous contexts that cannot page things out.
3274 *
3275 * If there are applications that are active memory-allocators
3276 * (most normal use), this basically shouldn't matter.
3277 */
3278static int kswapd(void *p)
3279{
Mel Gorman215ddd62011-07-08 15:39:40 -07003280 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003281 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003282 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003283 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 pg_data_t *pgdat = (pg_data_t*)p;
3285 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003286
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287 struct reclaim_state reclaim_state = {
3288 .reclaimed_slab = 0,
3289 };
Rusty Russella70f7302009-03-13 14:49:46 +10303290 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291
Nick Piggincf40bd12009-01-21 08:12:39 +01003292 lockdep_set_current_reclaim_state(GFP_KERNEL);
3293
Rusty Russell174596a2009-01-01 10:12:29 +10303294 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003295 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 current->reclaim_state = &reclaim_state;
3297
3298 /*
3299 * Tell the memory management that we're a "memory allocator",
3300 * and that if we need more memory we should get access to it
3301 * regardless (see "__alloc_pages()"). "kswapd" should
3302 * never get caught in the normal page freeing logic.
3303 *
3304 * (Kswapd normally doesn't need memory anyway, but sometimes
3305 * you need a small amount of memory in order to be able to
3306 * page out something else, and this flag essentially protects
3307 * us from recursively trying to free more memory as we're
3308 * trying to free the first piece of memory in the first place).
3309 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003310 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003311 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312
Mel Gorman215ddd62011-07-08 15:39:40 -07003313 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003314 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003315 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003316 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003318 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003319
Mel Gorman215ddd62011-07-08 15:39:40 -07003320 /*
3321 * If the last balance_pgdat was unsuccessful it's unlikely a
3322 * new request of a similar or harder type will succeed soon
3323 * so consider going to sleep on the basis we reclaimed at
3324 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003325 if (balanced_classzone_idx >= new_classzone_idx &&
3326 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003327 new_order = pgdat->kswapd_max_order;
3328 new_classzone_idx = pgdat->classzone_idx;
3329 pgdat->kswapd_max_order = 0;
3330 pgdat->classzone_idx = pgdat->nr_zones - 1;
3331 }
3332
Mel Gorman99504742011-01-13 15:46:20 -08003333 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334 /*
3335 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003336 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 */
3338 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003339 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003341 kswapd_try_to_sleep(pgdat, balanced_order,
3342 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003344 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003345 new_order = order;
3346 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003347 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003348 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350
David Rientjes8fe23e02009-12-14 17:58:33 -08003351 ret = try_to_freeze();
3352 if (kthread_should_stop())
3353 break;
3354
3355 /*
3356 * We can speed up thawing tasks if we don't call balance_pgdat
3357 * after returning from the refrigerator
3358 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003359 if (!ret) {
3360 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003361 balanced_classzone_idx = classzone_idx;
3362 balanced_order = balance_pgdat(pgdat, order,
3363 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003364 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003366
Johannes Weiner71abdc12014-06-06 14:35:35 -07003367 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003368 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003369 lockdep_clear_current_reclaim_state();
3370
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 return 0;
3372}
3373
3374/*
3375 * A zone is low on free memory, so wake its kswapd task to service it.
3376 */
Mel Gorman99504742011-01-13 15:46:20 -08003377void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378{
3379 pg_data_t *pgdat;
3380
Con Kolivasf3fe6512006-01-06 00:11:15 -08003381 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 return;
3383
Paul Jackson02a0e532006-12-13 00:34:25 -08003384 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003386 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003387 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003388 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003389 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3390 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003391 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003393 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003394 return;
3395
3396 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003397 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398}
3399
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003400#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003402 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003403 * freed pages.
3404 *
3405 * Rather than trying to age LRUs the aim is to preserve the overall
3406 * LRU order by reclaiming preferentially
3407 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003409unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003411 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003412 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003413 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003414 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003415 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003416 .may_writepage = 1,
3417 .may_unmap = 1,
3418 .may_swap = 1,
3419 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003421 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003422 struct task_struct *p = current;
3423 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003425 p->flags |= PF_MEMALLOC;
3426 lockdep_set_current_reclaim_state(sc.gfp_mask);
3427 reclaim_state.reclaimed_slab = 0;
3428 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003429
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003430 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003431
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003432 p->reclaim_state = NULL;
3433 lockdep_clear_current_reclaim_state();
3434 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003435
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003436 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003438#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440/* It's optimal to keep kswapds on the same CPUs as their memory, but
3441 not required for correctness. So if the last cpu in a node goes
3442 away, we get changed to run anywhere: as the first one comes back,
3443 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003444static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3445 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003447 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003449 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003450 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003451 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303452 const struct cpumask *mask;
3453
3454 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003455
Rusty Russell3e597942009-01-01 10:12:24 +10303456 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003458 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459 }
3460 }
3461 return NOTIFY_OK;
3462}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
Yasunori Goto3218ae12006-06-27 02:53:33 -07003464/*
3465 * This kswapd start function will be called by init and node-hot-add.
3466 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3467 */
3468int kswapd_run(int nid)
3469{
3470 pg_data_t *pgdat = NODE_DATA(nid);
3471 int ret = 0;
3472
3473 if (pgdat->kswapd)
3474 return 0;
3475
3476 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3477 if (IS_ERR(pgdat->kswapd)) {
3478 /* failure at boot is fatal */
3479 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003480 pr_err("Failed to start kswapd on node %d\n", nid);
3481 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003482 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003483 }
3484 return ret;
3485}
3486
David Rientjes8fe23e02009-12-14 17:58:33 -08003487/*
Jiang Liud8adde12012-07-11 14:01:52 -07003488 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003489 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003490 */
3491void kswapd_stop(int nid)
3492{
3493 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3494
Jiang Liud8adde12012-07-11 14:01:52 -07003495 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003496 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003497 NODE_DATA(nid)->kswapd = NULL;
3498 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003499}
3500
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501static int __init kswapd_init(void)
3502{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003503 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003504
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003506 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003507 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 hotcpu_notifier(cpu_callback, 0);
3509 return 0;
3510}
3511
3512module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003513
3514#ifdef CONFIG_NUMA
3515/*
3516 * Zone reclaim mode
3517 *
3518 * If non-zero call zone_reclaim when the number of free pages falls below
3519 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003520 */
3521int zone_reclaim_mode __read_mostly;
3522
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003523#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003524#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003525#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3526#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3527
Christoph Lameter9eeff232006-01-18 17:42:31 -08003528/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003529 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3530 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3531 * a zone.
3532 */
3533#define ZONE_RECLAIM_PRIORITY 4
3534
Christoph Lameter9eeff232006-01-18 17:42:31 -08003535/*
Christoph Lameter96146342006-07-03 00:24:13 -07003536 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3537 * occur.
3538 */
3539int sysctl_min_unmapped_ratio = 1;
3540
3541/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003542 * If the number of slab pages in a zone grows beyond this percentage then
3543 * slab reclaim needs to occur.
3544 */
3545int sysctl_min_slab_ratio = 5;
3546
Mel Gorman90afa5d2009-06-16 15:33:20 -07003547static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3548{
3549 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3550 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3551 zone_page_state(zone, NR_ACTIVE_FILE);
3552
3553 /*
3554 * It's possible for there to be more file mapped pages than
3555 * accounted for by the pages on the file LRU lists because
3556 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3557 */
3558 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3559}
3560
3561/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3562static long zone_pagecache_reclaimable(struct zone *zone)
3563{
3564 long nr_pagecache_reclaimable;
3565 long delta = 0;
3566
3567 /*
3568 * If RECLAIM_SWAP is set, then all file pages are considered
3569 * potentially reclaimable. Otherwise, we have to worry about
3570 * pages like swapcache and zone_unmapped_file_pages() provides
3571 * a better estimate
3572 */
3573 if (zone_reclaim_mode & RECLAIM_SWAP)
3574 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3575 else
3576 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3577
3578 /* If we can't clean pages, remove dirty pages from consideration */
3579 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3580 delta += zone_page_state(zone, NR_FILE_DIRTY);
3581
3582 /* Watch for any possible underflows due to delta */
3583 if (unlikely(delta > nr_pagecache_reclaimable))
3584 delta = nr_pagecache_reclaimable;
3585
3586 return nr_pagecache_reclaimable - delta;
3587}
3588
Christoph Lameter0ff38492006-09-25 23:31:52 -07003589/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003590 * Try to free up some pages from this zone through reclaim.
3591 */
Andrew Morton179e9632006-03-22 00:08:18 -08003592static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003593{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003594 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003595 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003596 struct task_struct *p = current;
3597 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003598 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003599 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003600 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003601 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003602 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003603 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3604 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3605 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003606 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003607 struct shrink_control shrink = {
3608 .gfp_mask = sc.gfp_mask,
3609 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003610 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003611
Christoph Lameter9eeff232006-01-18 17:42:31 -08003612 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003613 /*
3614 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3615 * and we also need to be able to write out pages for RECLAIM_WRITE
3616 * and RECLAIM_SWAP.
3617 */
3618 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003619 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003620 reclaim_state.reclaimed_slab = 0;
3621 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003622
Mel Gorman90afa5d2009-06-16 15:33:20 -07003623 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003624 /*
3625 * Free memory by calling shrink zone with increasing
3626 * priorities until we have enough memory freed.
3627 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003628 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003629 shrink_zone(zone, &sc);
3630 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003631 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003632
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003633 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3634 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003635 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003636 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003637 * many pages were freed in this zone. So we take the current
3638 * number of slab pages and shake the slab until it is reduced
3639 * by the same nr_pages that we used for reclaiming unmapped
3640 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003641 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003642 nodes_clear(shrink.nodes_to_scan);
3643 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003644 for (;;) {
3645 unsigned long lru_pages = zone_reclaimable_pages(zone);
3646
3647 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003648 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003649 break;
3650
3651 /* Freed enough memory */
3652 nr_slab_pages1 = zone_page_state(zone,
3653 NR_SLAB_RECLAIMABLE);
3654 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3655 break;
3656 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003657
3658 /*
3659 * Update nr_reclaimed by the number of slab pages we
3660 * reclaimed from this zone.
3661 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003662 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3663 if (nr_slab_pages1 < nr_slab_pages0)
3664 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003665 }
3666
Christoph Lameter9eeff232006-01-18 17:42:31 -08003667 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003668 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003669 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003670 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003671}
Andrew Morton179e9632006-03-22 00:08:18 -08003672
3673int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3674{
Andrew Morton179e9632006-03-22 00:08:18 -08003675 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003676 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003677
3678 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003679 * Zone reclaim reclaims unmapped file backed pages and
3680 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003681 *
Christoph Lameter96146342006-07-03 00:24:13 -07003682 * A small portion of unmapped file backed pages is needed for
3683 * file I/O otherwise pages read by file I/O will be immediately
3684 * thrown out if the zone is overallocated. So we do not reclaim
3685 * if less than a specified percentage of the zone is used by
3686 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003687 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003688 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3689 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003690 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003691
Lisa Du6e543d52013-09-11 14:22:36 -07003692 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003693 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003694
Andrew Morton179e9632006-03-22 00:08:18 -08003695 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003696 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003697 */
David Rientjesd773ed62007-10-16 23:26:01 -07003698 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003699 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003700
3701 /*
3702 * Only run zone reclaim on the local zone or on zones that do not
3703 * have associated processors. This will favor the local processor
3704 * over remote processors and spread off node memory allocations
3705 * as wide as possible.
3706 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003707 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003708 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003709 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003710
3711 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003712 return ZONE_RECLAIM_NOSCAN;
3713
David Rientjesd773ed62007-10-16 23:26:01 -07003714 ret = __zone_reclaim(zone, gfp_mask, order);
3715 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3716
Mel Gorman24cf725182009-06-16 15:33:23 -07003717 if (!ret)
3718 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3719
David Rientjesd773ed62007-10-16 23:26:01 -07003720 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003721}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003722#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003723
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003724/*
3725 * page_evictable - test whether a page is evictable
3726 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003727 *
3728 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003729 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003730 *
3731 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003732 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003733 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003734 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003735 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003736int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003737{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003738 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003739}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003740
Hugh Dickins850465792012-01-20 14:34:19 -08003741#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003742/**
Hugh Dickins24513262012-01-20 14:34:21 -08003743 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3744 * @pages: array of pages to check
3745 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003746 *
Hugh Dickins24513262012-01-20 14:34:21 -08003747 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003748 *
3749 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003750 */
Hugh Dickins24513262012-01-20 14:34:21 -08003751void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003752{
Johannes Weiner925b7672012-01-12 17:18:15 -08003753 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003754 struct zone *zone = NULL;
3755 int pgscanned = 0;
3756 int pgrescued = 0;
3757 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003758
Hugh Dickins24513262012-01-20 14:34:21 -08003759 for (i = 0; i < nr_pages; i++) {
3760 struct page *page = pages[i];
3761 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003762
Hugh Dickins24513262012-01-20 14:34:21 -08003763 pgscanned++;
3764 pagezone = page_zone(page);
3765 if (pagezone != zone) {
3766 if (zone)
3767 spin_unlock_irq(&zone->lru_lock);
3768 zone = pagezone;
3769 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003770 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003771 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003772
Hugh Dickins24513262012-01-20 14:34:21 -08003773 if (!PageLRU(page) || !PageUnevictable(page))
3774 continue;
3775
Hugh Dickins39b5f292012-10-08 16:33:18 -07003776 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003777 enum lru_list lru = page_lru_base_type(page);
3778
Sasha Levin309381fea2014-01-23 15:52:54 -08003779 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003780 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003781 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3782 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003783 pgrescued++;
3784 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003785 }
Hugh Dickins24513262012-01-20 14:34:21 -08003786
3787 if (zone) {
3788 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3789 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3790 spin_unlock_irq(&zone->lru_lock);
3791 }
Hugh Dickins850465792012-01-20 14:34:19 -08003792}
3793#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003794
Johannes Weiner264e56d2011-10-31 17:09:13 -07003795static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003796{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003797 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003798 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003799 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003800 "one, please send an email to linux-mm@kvack.org.\n",
3801 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003802}
3803
3804/*
3805 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3806 * all nodes' unevictable lists for evictable pages
3807 */
3808unsigned long scan_unevictable_pages;
3809
3810int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003811 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003812 size_t *length, loff_t *ppos)
3813{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003814 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003815 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003816 scan_unevictable_pages = 0;
3817 return 0;
3818}
3819
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003820#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003821/*
3822 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3823 * a specified node's per zone unevictable lists for evictable pages.
3824 */
3825
Kay Sievers10fbcf42011-12-21 14:48:43 -08003826static ssize_t read_scan_unevictable_node(struct device *dev,
3827 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003828 char *buf)
3829{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003830 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003831 return sprintf(buf, "0\n"); /* always zero; should fit... */
3832}
3833
Kay Sievers10fbcf42011-12-21 14:48:43 -08003834static ssize_t write_scan_unevictable_node(struct device *dev,
3835 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003836 const char *buf, size_t count)
3837{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003838 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003839 return 1;
3840}
3841
3842
Kay Sievers10fbcf42011-12-21 14:48:43 -08003843static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003844 read_scan_unevictable_node,
3845 write_scan_unevictable_node);
3846
3847int scan_unevictable_register_node(struct node *node)
3848{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003849 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003850}
3851
3852void scan_unevictable_unregister_node(struct node *node)
3853{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003854 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003855}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003856#endif