blob: 05d41c0d7f6c01bbf1e3acf58a2fcc119c13bafe [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070022#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070023#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/file.h>
25#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070051#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
Nick Piggin0f8053a2006-03-22 00:08:33 -080053#include "internal.h"
54
Mel Gorman33906bc2010-08-09 17:19:16 -070055#define CREATE_TRACE_POINTS
56#include <trace/events/vmscan.h>
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 /* Incremented by the number of inactive pages that were scanned */
60 unsigned long nr_scanned;
61
Rik van Riela79311c2009-01-06 14:40:01 -080062 /* Number of pages freed so far during a call to shrink_zones() */
63 unsigned long nr_reclaimed;
64
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080065 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080068 unsigned long hibernation_mode;
69
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040071 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070072
73 int may_writepage;
74
Johannes Weinera6dc60f82009-03-31 15:19:30 -070075 /* Can mapped pages be reclaimed? */
76 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080077
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070078 /* Can pages be swapped as part of reclaim? */
79 int may_swap;
80
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070081 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080082
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070083 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070086 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080087 * The memory cgroup that hit its limit and as a result is the
88 * primary target of this reclaim invocation.
89 */
90 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080091
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070092 /*
93 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 * are scanned.
95 */
96 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097};
98
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100
101#ifdef ARCH_HAS_PREFETCH
102#define prefetch_prev_lru_page(_page, _base, _field) \
103 do { \
104 if ((_page)->lru.prev != _base) { \
105 struct page *prev; \
106 \
107 prev = lru_to_page(&(_page->lru)); \
108 prefetch(&prev->_field); \
109 } \
110 } while (0)
111#else
112#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113#endif
114
115#ifdef ARCH_HAS_PREFETCHW
116#define prefetchw_prev_lru_page(_page, _base, _field) \
117 do { \
118 if ((_page)->lru.prev != _base) { \
119 struct page *prev; \
120 \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetchw(&prev->_field); \
123 } \
124 } while (0)
125#else
126#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127#endif
128
129/*
130 * From 0 .. 100. Higher means more swappy.
131 */
132int vm_swappiness = 60;
Zhang Yanfeib21e0b92013-02-22 16:35:48 -0800133unsigned long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135static LIST_HEAD(shrinker_list);
136static DECLARE_RWSEM(shrinker_rwsem);
137
Andrew Mortonc255a452012-07-31 16:43:02 -0700138#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800139static bool global_reclaim(struct scan_control *sc)
140{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800141 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800142}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800143#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800144static bool global_reclaim(struct scan_control *sc)
145{
146 return true;
147}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800148#endif
149
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800150static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700151{
152 int nr;
153
154 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
155 zone_page_state(zone, NR_INACTIVE_FILE);
156
157 if (get_nr_swap_pages() > 0)
158 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
159 zone_page_state(zone, NR_INACTIVE_ANON);
160
161 return nr;
162}
163
164bool zone_reclaimable(struct zone *zone)
165{
166 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
167}
168
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700169static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800170{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700171 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700172 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800173
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700174 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800175}
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000178 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000180int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000182 size_t size = sizeof(*shrinker->nr_deferred);
183
184 /*
185 * If we only have one possible node in the system anyway, save
186 * ourselves the trouble and disable NUMA aware behavior. This way we
187 * will save memory and some small loop time later.
188 */
189 if (nr_node_ids == 1)
190 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
191
192 if (shrinker->flags & SHRINKER_NUMA_AWARE)
193 size *= nr_node_ids;
194
195 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
196 if (!shrinker->nr_deferred)
197 return -ENOMEM;
198
Rusty Russell8e1f9362007-07-17 04:03:17 -0700199 down_write(&shrinker_rwsem);
200 list_add_tail(&shrinker->list, &shrinker_list);
201 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000202 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700204EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
206/*
207 * Remove one
208 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700209void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
211 down_write(&shrinker_rwsem);
212 list_del(&shrinker->list);
213 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700214 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000219
220static unsigned long
221shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
222 unsigned long nr_pages_scanned, unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000224 unsigned long freed = 0;
225 unsigned long long delta;
226 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700227 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000228 long nr;
229 long new_nr;
230 int nid = shrinkctl->nid;
231 long batch_size = shrinker->batch ? shrinker->batch
232 : SHRINK_BATCH;
233
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700234 freeable = shrinker->count_objects(shrinker, shrinkctl);
235 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000236 return 0;
237
238 /*
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
242 */
243 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
244
245 total_scan = nr;
246 delta = (4 * nr_pages_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700247 delta *= freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000248 do_div(delta, lru_pages + 1);
249 total_scan += delta;
250 if (total_scan < 0) {
251 printk(KERN_ERR
252 "shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000253 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700254 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000255 }
256
257 /*
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700263 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000264 * memory.
265 *
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
268 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700269 if (delta < freeable / 4)
270 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000271
272 /*
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
275 * freeable entries.
276 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700277 if (total_scan > freeable * 2)
278 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279
280 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
281 nr_pages_scanned, lru_pages,
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700282 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000283
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800284 /*
285 * Normally, we should not scan less than batch_size objects in one
286 * pass to avoid too frequent shrinker calls, but if the slab has less
287 * than batch_size objects in total and we are really tight on memory,
288 * we will try to reclaim all available objects, otherwise we can end
289 * up failing allocations although there are plenty of reclaimable
290 * objects spread over several slabs with usage less than the
291 * batch_size.
292 *
293 * We detect the "tight on memory" situations by looking at the total
294 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700295 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800296 * scanning at high prio and therefore should try to reclaim as much as
297 * possible.
298 */
299 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700300 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000301 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800302 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000303
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800304 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000305 ret = shrinker->scan_objects(shrinker, shrinkctl);
306 if (ret == SHRINK_STOP)
307 break;
308 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000309
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800310 count_vm_events(SLABS_SCANNED, nr_to_scan);
311 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000312
313 cond_resched();
314 }
315
316 /*
317 * move the unused scan count back into the shrinker in a
318 * manner that handles concurrent updates. If we exhausted the
319 * scan, there is no need to do an update.
320 */
321 if (total_scan > 0)
322 new_nr = atomic_long_add_return(total_scan,
323 &shrinker->nr_deferred[nid]);
324 else
325 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
326
Dave Hansendf9024a2014-06-04 16:08:07 -0700327 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000328 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329}
330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331/*
332 * Call the shrink functions to age shrinkable caches
333 *
334 * Here we assume it costs one seek to replace a lru page and that it also
335 * takes a seek to recreate a cache object. With this in mind we age equal
336 * percentages of the lru and ageable caches. This should balance the seeks
337 * generated by these structures.
338 *
Simon Arlott183ff222007-10-20 01:27:18 +0200339 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * slab to avoid swapping.
341 *
342 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
343 *
344 * `lru_pages' represents the number of on-LRU pages in all the zones which
345 * are eligible for the caller's allocation attempt. It is used for balancing
346 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700347 *
348 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 */
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000350unsigned long shrink_slab(struct shrink_control *shrinkctl,
Ying Han1495f232011-05-24 17:12:27 -0700351 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700352 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000355 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Ying Han1495f232011-05-24 17:12:27 -0700357 if (nr_pages_scanned == 0)
358 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Minchan Kimf06590b2011-05-24 17:11:11 -0700360 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000361 /*
362 * If we would return 0, our callers would understand that we
363 * have nothing else to shrink and give up trying. By returning
364 * 1 we keep it going and assume we'll be able to shrink next
365 * time.
366 */
367 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700368 goto out;
369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370
371 list_for_each_entry(shrinker, &shrinker_list, list) {
Vladimir Davydovec970972014-01-23 15:53:23 -0800372 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
373 shrinkctl->nid = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000374 freed += shrink_slab_node(shrinkctl, shrinker,
Vladimir Davydovec970972014-01-23 15:53:23 -0800375 nr_pages_scanned, lru_pages);
376 continue;
377 }
378
379 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
380 if (node_online(shrinkctl->nid))
381 freed += shrink_slab_node(shrinkctl, shrinker,
382 nr_pages_scanned, lru_pages);
Dave Chinneracf92b42011-07-08 14:14:35 +1000383
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 }
386 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700387out:
388 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000389 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390}
391
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392static inline int is_page_cache_freeable(struct page *page)
393{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700394 /*
395 * A freeable page cache page is referenced only by the caller
396 * that isolated the page, the page cache radix tree and
397 * optional buffer heads at page->private.
398 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700399 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400}
401
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700402static int may_write_to_queue(struct backing_dev_info *bdi,
403 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404{
Christoph Lameter930d9152006-01-08 01:00:47 -0800405 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 return 1;
407 if (!bdi_write_congested(bdi))
408 return 1;
409 if (bdi == current->backing_dev_info)
410 return 1;
411 return 0;
412}
413
414/*
415 * We detected a synchronous write error writing a page out. Probably
416 * -ENOSPC. We need to propagate that into the address_space for a subsequent
417 * fsync(), msync() or close().
418 *
419 * The tricky part is that after writepage we cannot touch the mapping: nothing
420 * prevents it from being freed up. But we have a ref on the page and once
421 * that page is locked, the mapping is pinned.
422 *
423 * We're allowed to run sleeping lock_page() here because we know the caller has
424 * __GFP_FS.
425 */
426static void handle_write_error(struct address_space *mapping,
427 struct page *page, int error)
428{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100429 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700430 if (page_mapping(page) == mapping)
431 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 unlock_page(page);
433}
434
Christoph Lameter04e62a22006-06-23 02:03:38 -0700435/* possible outcome of pageout() */
436typedef enum {
437 /* failed to write page out, page is locked */
438 PAGE_KEEP,
439 /* move page to the active list, page is locked */
440 PAGE_ACTIVATE,
441 /* page has been sent to the disk successfully, page is unlocked */
442 PAGE_SUCCESS,
443 /* page is clean and locked */
444 PAGE_CLEAN,
445} pageout_t;
446
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447/*
Andrew Morton1742f192006-03-22 00:08:21 -0800448 * pageout is called by shrink_page_list() for each dirty page.
449 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700451static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700452 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 /*
455 * If the page is dirty, only perform writeback if that write
456 * will be non-blocking. To prevent this allocation from being
457 * stalled by pagecache activity. But note that there may be
458 * stalls if we need to run get_block(). We could test
459 * PagePrivate for that.
460 *
Vincent Li6aceb532009-12-14 17:58:49 -0800461 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 * this page's queue, we can perform writeback even if that
463 * will block.
464 *
465 * If the page is swapcache, write it back even if that would
466 * block, for some throttling. This happens by accident, because
467 * swap_backing_dev_info is bust: it doesn't reflect the
468 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 */
470 if (!is_page_cache_freeable(page))
471 return PAGE_KEEP;
472 if (!mapping) {
473 /*
474 * Some data journaling orphaned pages can have
475 * page->mapping == NULL while being dirty with clean buffers.
476 */
David Howells266cf652009-04-03 16:42:36 +0100477 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 if (try_to_free_buffers(page)) {
479 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700480 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 return PAGE_CLEAN;
482 }
483 }
484 return PAGE_KEEP;
485 }
486 if (mapping->a_ops->writepage == NULL)
487 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700488 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return PAGE_KEEP;
490
491 if (clear_page_dirty_for_io(page)) {
492 int res;
493 struct writeback_control wbc = {
494 .sync_mode = WB_SYNC_NONE,
495 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700496 .range_start = 0,
497 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 .for_reclaim = 1,
499 };
500
501 SetPageReclaim(page);
502 res = mapping->a_ops->writepage(page, &wbc);
503 if (res < 0)
504 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800505 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 ClearPageReclaim(page);
507 return PAGE_ACTIVATE;
508 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 if (!PageWriteback(page)) {
511 /* synchronous write or broken a_ops? */
512 ClearPageReclaim(page);
513 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700514 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700515 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 return PAGE_SUCCESS;
517 }
518
519 return PAGE_CLEAN;
520}
521
Andrew Mortona649fd92006-10-17 00:09:36 -0700522/*
Nick Piggine2867812008-07-25 19:45:30 -0700523 * Same as remove_mapping, but if the page is removed from the mapping, it
524 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700525 */
Johannes Weinera5289102014-04-03 14:47:51 -0700526static int __remove_mapping(struct address_space *mapping, struct page *page,
527 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800528{
Nick Piggin28e4d962006-09-25 23:31:23 -0700529 BUG_ON(!PageLocked(page));
530 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800531
Nick Piggin19fd6232008-07-25 19:45:32 -0700532 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800533 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700534 * The non racy check for a busy page.
535 *
536 * Must be careful with the order of the tests. When someone has
537 * a ref to the page, it may be possible that they dirty it then
538 * drop the reference. So if PageDirty is tested before page_count
539 * here, then the following race may occur:
540 *
541 * get_user_pages(&page);
542 * [user mapping goes away]
543 * write_to(page);
544 * !PageDirty(page) [good]
545 * SetPageDirty(page);
546 * put_page(page);
547 * !page_count(page) [good, discard it]
548 *
549 * [oops, our write_to data is lost]
550 *
551 * Reversing the order of the tests ensures such a situation cannot
552 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
553 * load is not satisfied before that of page->_count.
554 *
555 * Note that if SetPageDirty is always performed via set_page_dirty,
556 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800557 */
Nick Piggine2867812008-07-25 19:45:30 -0700558 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800559 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700560 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
561 if (unlikely(PageDirty(page))) {
562 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800563 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700564 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800565
566 if (PageSwapCache(page)) {
567 swp_entry_t swap = { .val = page_private(page) };
568 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700569 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700570 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700571 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500572 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700573 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500574
575 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700576 /*
577 * Remember a shadow entry for reclaimed file cache in
578 * order to detect refaults, thus thrashing, later on.
579 *
580 * But don't store shadows in an address space that is
581 * already exiting. This is not just an optizimation,
582 * inode reclaim needs to empty out the radix tree or
583 * the nodes are lost. Don't plant shadows behind its
584 * back.
585 */
586 if (reclaimed && page_is_file_cache(page) &&
587 !mapping_exiting(mapping))
588 shadow = workingset_eviction(mapping, page);
589 __delete_from_page_cache(page, shadow);
Nick Piggin19fd6232008-07-25 19:45:32 -0700590 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700591 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500592
593 if (freepage != NULL)
594 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800595 }
596
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800597 return 1;
598
599cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700600 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800601 return 0;
602}
603
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604/*
Nick Piggine2867812008-07-25 19:45:30 -0700605 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
606 * someone else has a ref on the page, abort and return 0. If it was
607 * successfully detached, return 1. Assumes the caller has a single ref on
608 * this page.
609 */
610int remove_mapping(struct address_space *mapping, struct page *page)
611{
Johannes Weinera5289102014-04-03 14:47:51 -0700612 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700613 /*
614 * Unfreezing the refcount with 1 rather than 2 effectively
615 * drops the pagecache ref for us without requiring another
616 * atomic operation.
617 */
618 page_unfreeze_refs(page, 1);
619 return 1;
620 }
621 return 0;
622}
623
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700624/**
625 * putback_lru_page - put previously isolated page onto appropriate LRU list
626 * @page: page to be put back to appropriate lru list
627 *
628 * Add previously isolated @page to appropriate LRU list.
629 * Page may still be unevictable for other reasons.
630 *
631 * lru_lock must not be held, interrupts must be enabled.
632 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700633void putback_lru_page(struct page *page)
634{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700635 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700636 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700637
Sasha Levin309381fea2014-01-23 15:52:54 -0800638 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700639
640redo:
641 ClearPageUnevictable(page);
642
Hugh Dickins39b5f292012-10-08 16:33:18 -0700643 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700644 /*
645 * For evictable pages, we can use the cache.
646 * In event of a race, worst case is we end up with an
647 * unevictable page on [in]active list.
648 * We know how to handle that.
649 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700650 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700651 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700652 } else {
653 /*
654 * Put unevictable pages directly on zone's unevictable
655 * list.
656 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700657 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700658 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700659 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700660 * When racing with an mlock or AS_UNEVICTABLE clearing
661 * (page is unlocked) make sure that if the other thread
662 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800663 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700664 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700665 * the page back to the evictable list.
666 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700667 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700668 */
669 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700670 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700671
672 /*
673 * page's status can change while we move it among lru. If an evictable
674 * page is on unevictable list, it never be freed. To avoid that,
675 * check after we added it to the list, again.
676 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700677 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700678 if (!isolate_lru_page(page)) {
679 put_page(page);
680 goto redo;
681 }
682 /* This means someone else dropped this page from LRU
683 * So, it will be freed or putback to LRU again. There is
684 * nothing to do here.
685 */
686 }
687
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700688 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700689 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700690 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700691 count_vm_event(UNEVICTABLE_PGCULLED);
692
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700693 put_page(page); /* drop ref from isolate */
694}
695
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800696enum page_references {
697 PAGEREF_RECLAIM,
698 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800699 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800700 PAGEREF_ACTIVATE,
701};
702
703static enum page_references page_check_references(struct page *page,
704 struct scan_control *sc)
705{
Johannes Weiner645747462010-03-05 13:42:22 -0800706 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800707 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800708
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700709 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
710 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800711 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800712
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800713 /*
714 * Mlock lost the isolation race with us. Let try_to_unmap()
715 * move the page to the unevictable list.
716 */
717 if (vm_flags & VM_LOCKED)
718 return PAGEREF_RECLAIM;
719
Johannes Weiner645747462010-03-05 13:42:22 -0800720 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700721 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800722 return PAGEREF_ACTIVATE;
723 /*
724 * All mapped pages start out with page table
725 * references from the instantiating fault, so we need
726 * to look twice if a mapped file page is used more
727 * than once.
728 *
729 * Mark it and spare it for another trip around the
730 * inactive list. Another page table reference will
731 * lead to its activation.
732 *
733 * Note: the mark is set for activated pages as well
734 * so that recently deactivated but used pages are
735 * quickly recovered.
736 */
737 SetPageReferenced(page);
738
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800739 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800740 return PAGEREF_ACTIVATE;
741
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800742 /*
743 * Activate file-backed executable pages after first usage.
744 */
745 if (vm_flags & VM_EXEC)
746 return PAGEREF_ACTIVATE;
747
Johannes Weiner645747462010-03-05 13:42:22 -0800748 return PAGEREF_KEEP;
749 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800750
751 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700752 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800753 return PAGEREF_RECLAIM_CLEAN;
754
755 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800756}
757
Mel Gormane2be15f2013-07-03 15:01:57 -0700758/* Check if a page is dirty or under writeback */
759static void page_check_dirty_writeback(struct page *page,
760 bool *dirty, bool *writeback)
761{
Mel Gormanb4597222013-07-03 15:02:05 -0700762 struct address_space *mapping;
763
Mel Gormane2be15f2013-07-03 15:01:57 -0700764 /*
765 * Anonymous pages are not handled by flushers and must be written
766 * from reclaim context. Do not stall reclaim based on them
767 */
768 if (!page_is_file_cache(page)) {
769 *dirty = false;
770 *writeback = false;
771 return;
772 }
773
774 /* By default assume that the page flags are accurate */
775 *dirty = PageDirty(page);
776 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700777
778 /* Verify dirty/writeback state if the filesystem supports it */
779 if (!page_has_private(page))
780 return;
781
782 mapping = page_mapping(page);
783 if (mapping && mapping->a_ops->is_dirty_writeback)
784 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700785}
786
Nick Piggine2867812008-07-25 19:45:30 -0700787/*
Andrew Morton1742f192006-03-22 00:08:21 -0800788 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 */
Andrew Morton1742f192006-03-22 00:08:21 -0800790static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700791 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700792 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700793 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700794 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700795 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700796 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700797 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700798 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700799 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
801 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700802 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700804 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700805 unsigned long nr_dirty = 0;
806 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800807 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700808 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700809 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
811 cond_resched();
812
Tim Chen69980e32012-07-31 16:46:08 -0700813 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 while (!list_empty(page_list)) {
815 struct address_space *mapping;
816 struct page *page;
817 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700818 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700819 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
821 cond_resched();
822
823 page = lru_to_page(page_list);
824 list_del(&page->lru);
825
Nick Piggin529ae9a2008-08-02 12:01:03 +0200826 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 goto keep;
828
Sasha Levin309381fea2014-01-23 15:52:54 -0800829 VM_BUG_ON_PAGE(PageActive(page), page);
830 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800833
Hugh Dickins39b5f292012-10-08 16:33:18 -0700834 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700835 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700836
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700837 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800838 goto keep_locked;
839
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 /* Double the slab pressure for mapped and swapcache pages */
841 if (page_mapped(page) || PageSwapCache(page))
842 sc->nr_scanned++;
843
Andy Whitcroftc661b072007-08-22 14:01:26 -0700844 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
845 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
846
Mel Gorman283aba92013-07-03 15:01:51 -0700847 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700848 * The number of dirty pages determines if a zone is marked
849 * reclaim_congested which affects wait_iff_congested. kswapd
850 * will stall and start writing pages if the tail of the LRU
851 * is all dirty unqueued pages.
852 */
853 page_check_dirty_writeback(page, &dirty, &writeback);
854 if (dirty || writeback)
855 nr_dirty++;
856
857 if (dirty && !writeback)
858 nr_unqueued_dirty++;
859
Mel Gormand04e8ac2013-07-03 15:02:03 -0700860 /*
861 * Treat this page as congested if the underlying BDI is or if
862 * pages are cycling through the LRU so quickly that the
863 * pages marked for immediate reclaim are making it to the
864 * end of the LRU a second time.
865 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700866 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700867 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
868 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700869 nr_congested++;
870
871 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700872 * If a page at the tail of the LRU is under writeback, there
873 * are three cases to consider.
874 *
875 * 1) If reclaim is encountering an excessive number of pages
876 * under writeback and this page is both under writeback and
877 * PageReclaim then it indicates that pages are being queued
878 * for IO but are being recycled through the LRU before the
879 * IO can complete. Waiting on the page itself risks an
880 * indefinite stall if it is impossible to writeback the
881 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700882 * note that the LRU is being scanned too quickly and the
883 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700884 *
885 * 2) Global reclaim encounters a page, memcg encounters a
886 * page that is not marked for immediate reclaim or
887 * the caller does not have __GFP_IO. In this case mark
888 * the page for immediate reclaim and continue scanning.
889 *
890 * __GFP_IO is checked because a loop driver thread might
891 * enter reclaim, and deadlock if it waits on a page for
892 * which it is needed to do the write (loop masks off
893 * __GFP_IO|__GFP_FS for this reason); but more thought
894 * would probably show more reasons.
895 *
896 * Don't require __GFP_FS, since we're not going into the
897 * FS, just waiting on its writeback completion. Worryingly,
898 * ext4 gfs2 and xfs allocate pages with
899 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
900 * may_enter_fs here is liable to OOM on them.
901 *
902 * 3) memcg encounters a page that is not already marked
903 * PageReclaim. memcg does not have any dirty pages
904 * throttling so we could easily OOM just because too many
905 * pages are in writeback and there is nothing else to
906 * reclaim. Wait for the writeback to complete.
907 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700908 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700909 /* Case 1 above */
910 if (current_is_kswapd() &&
911 PageReclaim(page) &&
912 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700913 nr_immediate++;
914 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700915
916 /* Case 2 above */
917 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700918 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
919 /*
920 * This is slightly racy - end_page_writeback()
921 * might have just cleared PageReclaim, then
922 * setting PageReclaim here end up interpreted
923 * as PageReadahead - but that does not matter
924 * enough to care. What we do want is for this
925 * page to have PageReclaim set next time memcg
926 * reclaim reaches the tests above, so it will
927 * then wait_on_page_writeback() to avoid OOM;
928 * and it's also appropriate in global reclaim.
929 */
930 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700931 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700932
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700933 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700934
935 /* Case 3 above */
936 } else {
937 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700938 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700939 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
Minchan Kim02c6de82012-10-08 16:31:55 -0700941 if (!force_reclaim)
942 references = page_check_references(page, sc);
943
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800944 switch (references) {
945 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800947 case PAGEREF_KEEP:
948 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800949 case PAGEREF_RECLAIM:
950 case PAGEREF_RECLAIM_CLEAN:
951 ; /* try to reclaim the page below */
952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 /*
955 * Anonymous process memory has backing store?
956 * Try to allocate it some swap space here.
957 */
Nick Pigginb291f002008-10-18 20:26:44 -0700958 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800959 if (!(sc->gfp_mask & __GFP_IO))
960 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700961 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800963 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Mel Gormane2be15f2013-07-03 15:01:57 -0700965 /* Adding to swap updated mapping */
966 mapping = page_mapping(page);
967 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
969 /*
970 * The page is mapped into the page tables of one or more
971 * processes. Try to unmap it here.
972 */
973 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700974 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 case SWAP_FAIL:
976 goto activate_locked;
977 case SWAP_AGAIN:
978 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700979 case SWAP_MLOCK:
980 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 case SWAP_SUCCESS:
982 ; /* try to free the page below */
983 }
984 }
985
986 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700987 /*
988 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -0700989 * avoid risk of stack overflow but only writeback
990 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -0700991 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700992 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700993 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -0700994 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700995 /*
996 * Immediately reclaim when written back.
997 * Similar in principal to deactivate_page()
998 * except we already have the page isolated
999 * and know it's dirty
1000 */
1001 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1002 SetPageReclaim(page);
1003
Mel Gormanee728862011-10-31 17:07:38 -07001004 goto keep_locked;
1005 }
1006
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001007 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001009 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001011 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 goto keep_locked;
1013
1014 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001015 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016 case PAGE_KEEP:
1017 goto keep_locked;
1018 case PAGE_ACTIVATE:
1019 goto activate_locked;
1020 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001021 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001022 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001023 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001025
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 /*
1027 * A synchronous write - probably a ramdisk. Go
1028 * ahead and try to reclaim the page.
1029 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001030 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 goto keep;
1032 if (PageDirty(page) || PageWriteback(page))
1033 goto keep_locked;
1034 mapping = page_mapping(page);
1035 case PAGE_CLEAN:
1036 ; /* try to free the page below */
1037 }
1038 }
1039
1040 /*
1041 * If the page has buffers, try to free the buffer mappings
1042 * associated with this page. If we succeed we try to free
1043 * the page as well.
1044 *
1045 * We do this even if the page is PageDirty().
1046 * try_to_release_page() does not perform I/O, but it is
1047 * possible for a page to have PageDirty set, but it is actually
1048 * clean (all its buffers are clean). This happens if the
1049 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001050 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 * try_to_release_page() will discover that cleanness and will
1052 * drop the buffers and mark the page clean - it can be freed.
1053 *
1054 * Rarely, pages can have buffers and no ->mapping. These are
1055 * the pages which were not successfully invalidated in
1056 * truncate_complete_page(). We try to drop those buffers here
1057 * and if that worked, and the page is no longer mapped into
1058 * process address space (page_count == 1) it can be freed.
1059 * Otherwise, leave the page on the LRU so it is swappable.
1060 */
David Howells266cf652009-04-03 16:42:36 +01001061 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 if (!try_to_release_page(page, sc->gfp_mask))
1063 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001064 if (!mapping && page_count(page) == 1) {
1065 unlock_page(page);
1066 if (put_page_testzero(page))
1067 goto free_it;
1068 else {
1069 /*
1070 * rare race with speculative reference.
1071 * the speculative reference will free
1072 * this page shortly, so we may
1073 * increment nr_reclaimed here (and
1074 * leave it off the LRU).
1075 */
1076 nr_reclaimed++;
1077 continue;
1078 }
1079 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 }
1081
Johannes Weinera5289102014-04-03 14:47:51 -07001082 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001083 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
Nick Piggina978d6f2008-10-18 20:26:58 -07001085 /*
1086 * At this point, we have no other references and there is
1087 * no way to pick any more up (removed from LRU, removed
1088 * from pagecache). Can use non-atomic bitops now (and
1089 * we obviously don't have to worry about waking up a process
1090 * waiting on the page lock, because there are no references.
1091 */
1092 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001093free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001094 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001095
1096 /*
1097 * Is there need to periodically free_page_list? It would
1098 * appear not as the counts should be low
1099 */
1100 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 continue;
1102
Nick Pigginb291f002008-10-18 20:26:44 -07001103cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001104 if (PageSwapCache(page))
1105 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001106 unlock_page(page);
1107 putback_lru_page(page);
1108 continue;
1109
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001111 /* Not a candidate for swapping, so reclaim swap space. */
1112 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001113 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001114 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 SetPageActive(page);
1116 pgactivate++;
1117keep_locked:
1118 unlock_page(page);
1119keep:
1120 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001121 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001123
Mel Gormanb745bc82014-06-04 16:10:22 -07001124 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001127 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001128 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001129 *ret_nr_dirty += nr_dirty;
1130 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001131 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001132 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001133 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001134 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135}
1136
Minchan Kim02c6de82012-10-08 16:31:55 -07001137unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1138 struct list_head *page_list)
1139{
1140 struct scan_control sc = {
1141 .gfp_mask = GFP_KERNEL,
1142 .priority = DEF_PRIORITY,
1143 .may_unmap = 1,
1144 };
Mel Gorman8e950282013-07-03 15:02:02 -07001145 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001146 struct page *page, *next;
1147 LIST_HEAD(clean_pages);
1148
1149 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001150 if (page_is_file_cache(page) && !PageDirty(page) &&
1151 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001152 ClearPageActive(page);
1153 list_move(&page->lru, &clean_pages);
1154 }
1155 }
1156
1157 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001158 TTU_UNMAP|TTU_IGNORE_ACCESS,
1159 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001160 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001161 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001162 return ret;
1163}
1164
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001165/*
1166 * Attempt to remove the specified page from its LRU. Only take this page
1167 * if it is of the appropriate PageActive status. Pages which are being
1168 * freed elsewhere are also ignored.
1169 *
1170 * page: page to consider
1171 * mode: one of the LRU isolation modes defined above
1172 *
1173 * returns 0 on success, -ve errno on failure.
1174 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001175int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001176{
1177 int ret = -EINVAL;
1178
1179 /* Only take pages on the LRU. */
1180 if (!PageLRU(page))
1181 return ret;
1182
Minchan Kime46a2872012-10-08 16:33:48 -07001183 /* Compaction should not handle unevictable pages but CMA can do so */
1184 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001185 return ret;
1186
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001187 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001188
Mel Gormanc8244932012-01-12 17:19:38 -08001189 /*
1190 * To minimise LRU disruption, the caller can indicate that it only
1191 * wants to isolate pages it will be able to operate on without
1192 * blocking - clean pages for the most part.
1193 *
1194 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1195 * is used by reclaim when it is cannot write to backing storage
1196 *
1197 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1198 * that it is possible to migrate without blocking
1199 */
1200 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1201 /* All the caller can do on PageWriteback is block */
1202 if (PageWriteback(page))
1203 return ret;
1204
1205 if (PageDirty(page)) {
1206 struct address_space *mapping;
1207
1208 /* ISOLATE_CLEAN means only clean pages */
1209 if (mode & ISOLATE_CLEAN)
1210 return ret;
1211
1212 /*
1213 * Only pages without mappings or that have a
1214 * ->migratepage callback are possible to migrate
1215 * without blocking
1216 */
1217 mapping = page_mapping(page);
1218 if (mapping && !mapping->a_ops->migratepage)
1219 return ret;
1220 }
1221 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001222
Minchan Kimf80c0672011-10-31 17:06:55 -07001223 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1224 return ret;
1225
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001226 if (likely(get_page_unless_zero(page))) {
1227 /*
1228 * Be careful not to clear PageLRU until after we're
1229 * sure the page is not being freed elsewhere -- the
1230 * page release code relies on it.
1231 */
1232 ClearPageLRU(page);
1233 ret = 0;
1234 }
1235
1236 return ret;
1237}
1238
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001239/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 * zone->lru_lock is heavily contended. Some of the functions that
1241 * shrink the lists perform better by taking out a batch of pages
1242 * and working on them outside the LRU lock.
1243 *
1244 * For pagecache intensive workloads, this function is the hottest
1245 * spot in the kernel (apart from copy_*_user functions).
1246 *
1247 * Appropriate locks must be held before calling this function.
1248 *
1249 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001250 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001252 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001253 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001254 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001255 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 *
1257 * returns how many pages were moved onto *@dst.
1258 */
Andrew Morton69e05942006-03-22 00:08:19 -08001259static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001260 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001261 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001262 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001264 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001265 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001266 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001268 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001269 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001270 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001271
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 page = lru_to_page(src);
1273 prefetchw_prev_lru_page(page, src, flags);
1274
Sasha Levin309381fea2014-01-23 15:52:54 -08001275 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001276
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001277 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001278 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001279 nr_pages = hpage_nr_pages(page);
1280 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001281 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001282 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001283 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001284
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001285 case -EBUSY:
1286 /* else it is being freed elsewhere */
1287 list_move(&page->lru, src);
1288 continue;
1289
1290 default:
1291 BUG();
1292 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 }
1294
Hugh Dickinsf6260122012-01-12 17:20:06 -08001295 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001296 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1297 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 return nr_taken;
1299}
1300
Nick Piggin62695a82008-10-18 20:26:09 -07001301/**
1302 * isolate_lru_page - tries to isolate a page from its LRU list
1303 * @page: page to isolate from its LRU list
1304 *
1305 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1306 * vmstat statistic corresponding to whatever LRU list the page was on.
1307 *
1308 * Returns 0 if the page was removed from an LRU list.
1309 * Returns -EBUSY if the page was not on an LRU list.
1310 *
1311 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001312 * the active list, it will have PageActive set. If it was found on
1313 * the unevictable list, it will have the PageUnevictable bit set. That flag
1314 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001315 *
1316 * The vmstat statistic corresponding to the list on which the page was
1317 * found will be decremented.
1318 *
1319 * Restrictions:
1320 * (1) Must be called with an elevated refcount on the page. This is a
1321 * fundamentnal difference from isolate_lru_pages (which is called
1322 * without a stable reference).
1323 * (2) the lru_lock must not be held.
1324 * (3) interrupts must be enabled.
1325 */
1326int isolate_lru_page(struct page *page)
1327{
1328 int ret = -EBUSY;
1329
Sasha Levin309381fea2014-01-23 15:52:54 -08001330 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001331
Nick Piggin62695a82008-10-18 20:26:09 -07001332 if (PageLRU(page)) {
1333 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001334 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001335
1336 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001337 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001338 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001339 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001340 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001341 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001342 del_page_from_lru_list(page, lruvec, lru);
1343 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001344 }
1345 spin_unlock_irq(&zone->lru_lock);
1346 }
1347 return ret;
1348}
1349
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001350/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001351 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1352 * then get resheduled. When there are massive number of tasks doing page
1353 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1354 * the LRU list will go small and be scanned faster than necessary, leading to
1355 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001356 */
1357static int too_many_isolated(struct zone *zone, int file,
1358 struct scan_control *sc)
1359{
1360 unsigned long inactive, isolated;
1361
1362 if (current_is_kswapd())
1363 return 0;
1364
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001365 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001366 return 0;
1367
1368 if (file) {
1369 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1370 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1371 } else {
1372 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1373 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1374 }
1375
Fengguang Wu3cf23842012-12-18 14:23:31 -08001376 /*
1377 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1378 * won't get blocked by normal direct-reclaimers, forming a circular
1379 * deadlock.
1380 */
1381 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1382 inactive >>= 3;
1383
Rik van Riel35cd7812009-09-21 17:01:38 -07001384 return isolated > inactive;
1385}
1386
Mel Gorman66635622010-08-09 17:19:30 -07001387static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001388putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001389{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001390 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1391 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001392 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001393
Mel Gorman66635622010-08-09 17:19:30 -07001394 /*
1395 * Put back any unfreeable pages.
1396 */
Mel Gorman66635622010-08-09 17:19:30 -07001397 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001398 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001399 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001400
Sasha Levin309381fea2014-01-23 15:52:54 -08001401 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001402 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001403 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001404 spin_unlock_irq(&zone->lru_lock);
1405 putback_lru_page(page);
1406 spin_lock_irq(&zone->lru_lock);
1407 continue;
1408 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001409
1410 lruvec = mem_cgroup_page_lruvec(page, zone);
1411
Linus Torvalds7a608572011-01-17 14:42:19 -08001412 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001413 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001414 add_page_to_lru_list(page, lruvec, lru);
1415
Mel Gorman66635622010-08-09 17:19:30 -07001416 if (is_active_lru(lru)) {
1417 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001418 int numpages = hpage_nr_pages(page);
1419 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001420 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001421 if (put_page_testzero(page)) {
1422 __ClearPageLRU(page);
1423 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001424 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001425
1426 if (unlikely(PageCompound(page))) {
1427 spin_unlock_irq(&zone->lru_lock);
1428 (*get_compound_page_dtor(page))(page);
1429 spin_lock_irq(&zone->lru_lock);
1430 } else
1431 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001432 }
1433 }
Mel Gorman66635622010-08-09 17:19:30 -07001434
Hugh Dickins3f797682012-01-12 17:20:07 -08001435 /*
1436 * To save our caller's stack, now use input list for pages to free.
1437 */
1438 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001439}
1440
1441/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001442 * If a kernel thread (such as nfsd for loop-back mounts) services
1443 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1444 * In that case we should only throttle if the backing device it is
1445 * writing to is congested. In other cases it is safe to throttle.
1446 */
1447static int current_may_throttle(void)
1448{
1449 return !(current->flags & PF_LESS_THROTTLE) ||
1450 current->backing_dev_info == NULL ||
1451 bdi_write_congested(current->backing_dev_info);
1452}
1453
1454/*
Andrew Morton1742f192006-03-22 00:08:21 -08001455 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1456 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 */
Mel Gorman66635622010-08-09 17:19:30 -07001458static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001459shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001460 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461{
1462 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001463 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001464 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001465 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001466 unsigned long nr_dirty = 0;
1467 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001468 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001469 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001470 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001471 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001472 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001473 struct zone *zone = lruvec_zone(lruvec);
1474 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001475
Rik van Riel35cd7812009-09-21 17:01:38 -07001476 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001477 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001478
1479 /* We are about to die and free our memory. Return now. */
1480 if (fatal_signal_pending(current))
1481 return SWAP_CLUSTER_MAX;
1482 }
1483
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001485
1486 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001487 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001488 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001489 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001490
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001493 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1494 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001495
1496 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1497 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1498
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001499 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001500 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001501 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001502 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001503 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001504 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001505 }
Hillf Dantond563c052012-03-21 16:34:02 -07001506 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001507
Hillf Dantond563c052012-03-21 16:34:02 -07001508 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001509 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001510
Minchan Kim02c6de82012-10-08 16:31:55 -07001511 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001512 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1513 &nr_writeback, &nr_immediate,
1514 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001515
Hugh Dickins3f797682012-01-12 17:20:07 -08001516 spin_lock_irq(&zone->lru_lock);
1517
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001518 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001519
Ying Han904249a2012-04-25 16:01:48 -07001520 if (global_reclaim(sc)) {
1521 if (current_is_kswapd())
1522 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1523 nr_reclaimed);
1524 else
1525 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1526 nr_reclaimed);
1527 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001528
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001529 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001530
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001531 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001532
1533 spin_unlock_irq(&zone->lru_lock);
1534
Mel Gormanb745bc82014-06-04 16:10:22 -07001535 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001536
Mel Gorman92df3a72011-10-31 17:07:56 -07001537 /*
1538 * If reclaim is isolating dirty pages under writeback, it implies
1539 * that the long-lived page allocation rate is exceeding the page
1540 * laundering rate. Either the global limits are not being effective
1541 * at throttling processes due to the page distribution throughout
1542 * zones or there is heavy usage of a slow backing device. The
1543 * only option is to throttle from reclaim context which is not ideal
1544 * as there is no guarantee the dirtying process is throttled in the
1545 * same way balance_dirty_pages() manages.
1546 *
Mel Gorman8e950282013-07-03 15:02:02 -07001547 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1548 * of pages under pages flagged for immediate reclaim and stall if any
1549 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001550 */
Mel Gorman918fc712013-07-08 16:00:25 -07001551 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001552 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001553
Mel Gormand43006d2013-07-03 15:01:50 -07001554 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001555 * memcg will stall in page writeback so only consider forcibly
1556 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001557 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001558 if (global_reclaim(sc)) {
1559 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001560 * Tag a zone as congested if all the dirty pages scanned were
1561 * backed by a congested BDI and wait_iff_congested will stall.
1562 */
1563 if (nr_dirty && nr_dirty == nr_congested)
1564 zone_set_flag(zone, ZONE_CONGESTED);
1565
1566 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001567 * If dirty pages are scanned that are not queued for IO, it
1568 * implies that flushers are not keeping up. In this case, flag
1569 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1570 * pages from reclaim context. It will forcibly stall in the
1571 * next check.
1572 */
1573 if (nr_unqueued_dirty == nr_taken)
1574 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1575
1576 /*
1577 * In addition, if kswapd scans pages marked marked for
1578 * immediate reclaim and under writeback (nr_immediate), it
1579 * implies that pages are cycling through the LRU faster than
1580 * they are written so also forcibly stall.
1581 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001582 if ((nr_unqueued_dirty == nr_taken || nr_immediate) &&
1583 current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001584 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001585 }
Mel Gormand43006d2013-07-03 15:01:50 -07001586
Mel Gorman8e950282013-07-03 15:02:02 -07001587 /*
1588 * Stall direct reclaim for IO completions if underlying BDIs or zone
1589 * is congested. Allow kswapd to continue until it starts encountering
1590 * unqueued dirty pages or cycling through the LRU too quickly.
1591 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001592 if (!sc->hibernation_mode && !current_is_kswapd() &&
1593 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001594 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1595
Mel Gormane11da5b2010-10-26 14:21:40 -07001596 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1597 zone_idx(zone),
1598 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001599 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001600 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001601 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602}
1603
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001604/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 * This moves pages from the active list to the inactive list.
1606 *
1607 * We move them the other way if the page is referenced by one or more
1608 * processes, from rmap.
1609 *
1610 * If the pages are mostly unmapped, the processing is fast and it is
1611 * appropriate to hold zone->lru_lock across the whole operation. But if
1612 * the pages are mapped, the processing is slow (page_referenced()) so we
1613 * should drop zone->lru_lock around each page. It's impossible to balance
1614 * this, so instead we remove the pages from the LRU while processing them.
1615 * It is safe to rely on PG_active against the non-LRU pages in here because
1616 * nobody will play with that bit on a non-LRU page.
1617 *
1618 * The downside is that we have to touch page->_count against each page.
1619 * But we had to alter page->flags anyway.
1620 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001621
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001622static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001623 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001624 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001625 enum lru_list lru)
1626{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001627 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001628 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001629 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001630 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001631
Wu Fengguang3eb41402009-06-16 15:33:13 -07001632 while (!list_empty(list)) {
1633 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001634 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001635
Sasha Levin309381fea2014-01-23 15:52:54 -08001636 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001637 SetPageLRU(page);
1638
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001639 nr_pages = hpage_nr_pages(page);
1640 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001641 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001642 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001643
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001644 if (put_page_testzero(page)) {
1645 __ClearPageLRU(page);
1646 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001647 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001648
1649 if (unlikely(PageCompound(page))) {
1650 spin_unlock_irq(&zone->lru_lock);
1651 (*get_compound_page_dtor(page))(page);
1652 spin_lock_irq(&zone->lru_lock);
1653 } else
1654 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001655 }
1656 }
1657 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1658 if (!is_active_lru(lru))
1659 __count_vm_events(PGDEACTIVATE, pgmoved);
1660}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001661
Hugh Dickinsf6260122012-01-12 17:20:06 -08001662static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001663 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001664 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001665 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001667 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001668 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001669 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001671 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001672 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001674 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001675 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001676 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001677 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001678 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
1680 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001681
1682 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001683 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001684 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001685 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001686
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001688
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001689 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1690 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001691 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001692 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001693
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001694 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001695
Hugh Dickinsf6260122012-01-12 17:20:06 -08001696 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001697 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001698 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 spin_unlock_irq(&zone->lru_lock);
1700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 while (!list_empty(&l_hold)) {
1702 cond_resched();
1703 page = lru_to_page(&l_hold);
1704 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001705
Hugh Dickins39b5f292012-10-08 16:33:18 -07001706 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001707 putback_lru_page(page);
1708 continue;
1709 }
1710
Mel Gormancc715d92012-03-21 16:34:00 -07001711 if (unlikely(buffer_heads_over_limit)) {
1712 if (page_has_private(page) && trylock_page(page)) {
1713 if (page_has_private(page))
1714 try_to_release_page(page, 0);
1715 unlock_page(page);
1716 }
1717 }
1718
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001719 if (page_referenced(page, 0, sc->target_mem_cgroup,
1720 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001721 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001722 /*
1723 * Identify referenced, file-backed active pages and
1724 * give them one more trip around the active list. So
1725 * that executable code get better chances to stay in
1726 * memory under moderate memory pressure. Anon pages
1727 * are not likely to be evicted by use-once streaming
1728 * IO, plus JVM can create lots of anon VM_EXEC pages,
1729 * so we ignore them here.
1730 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001731 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001732 list_add(&page->lru, &l_active);
1733 continue;
1734 }
1735 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001736
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001737 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 list_add(&page->lru, &l_inactive);
1739 }
1740
Andrew Mortonb5557492009-01-06 14:40:13 -08001741 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001742 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001743 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001744 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001745 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001746 * Count referenced pages from currently used mappings as rotated,
1747 * even though only some of them are actually re-activated. This
1748 * helps balance scan pressure between file and anonymous pages in
1749 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001750 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001751 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001752
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001753 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1754 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001755 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001756 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001757
Mel Gormanb745bc82014-06-04 16:10:22 -07001758 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759}
1760
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001761#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001762static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001763{
1764 unsigned long active, inactive;
1765
1766 active = zone_page_state(zone, NR_ACTIVE_ANON);
1767 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1768
1769 if (inactive * zone->inactive_ratio < active)
1770 return 1;
1771
1772 return 0;
1773}
1774
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001775/**
1776 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001777 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001778 *
1779 * Returns true if the zone does not have enough inactive anon pages,
1780 * meaning some active anon pages need to be deactivated.
1781 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001782static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001783{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001784 /*
1785 * If we don't have swap space, anonymous page deactivation
1786 * is pointless.
1787 */
1788 if (!total_swap_pages)
1789 return 0;
1790
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001791 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001792 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001793
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001794 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001795}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001796#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001797static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001798{
1799 return 0;
1800}
1801#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001802
Rik van Riel56e49d22009-06-16 15:32:28 -07001803/**
1804 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001805 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001806 *
1807 * When the system is doing streaming IO, memory pressure here
1808 * ensures that active file pages get deactivated, until more
1809 * than half of the file pages are on the inactive list.
1810 *
1811 * Once we get to that situation, protect the system's working
1812 * set from being evicted by disabling active file page aging.
1813 *
1814 * This uses a different ratio than the anonymous pages, because
1815 * the page cache uses a use-once replacement algorithm.
1816 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001817static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001818{
Johannes Weinere3790142013-02-22 16:35:19 -08001819 unsigned long inactive;
1820 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001821
Johannes Weinere3790142013-02-22 16:35:19 -08001822 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1823 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1824
1825 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001826}
1827
Hugh Dickins75b00af2012-05-29 15:07:09 -07001828static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001829{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001830 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001831 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001832 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001833 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001834}
1835
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001836static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001837 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001838{
Rik van Rielb39415b2009-12-14 17:59:48 -08001839 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001840 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001841 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001842 return 0;
1843 }
1844
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001845 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001846}
1847
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001848static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001849{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001850 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001851 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001852 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001853}
1854
Johannes Weiner9a265112013-02-22 16:32:17 -08001855enum scan_balance {
1856 SCAN_EQUAL,
1857 SCAN_FRACT,
1858 SCAN_ANON,
1859 SCAN_FILE,
1860};
1861
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001863 * Determine how aggressively the anon and file LRU lists should be
1864 * scanned. The relative value of each set of LRU lists is determined
1865 * by looking at the fraction of the pages scanned we did rotate back
1866 * onto the active list instead of evict.
1867 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001868 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1869 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001870 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001871static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001872 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001873{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001874 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001875 u64 fraction[2];
1876 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001877 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001878 unsigned long anon_prio, file_prio;
1879 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001880 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001881 bool force_scan = false;
1882 unsigned long ap, fp;
1883 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001884 bool some_scanned;
1885 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001886
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001887 /*
1888 * If the zone or memcg is small, nr[l] can be 0. This
1889 * results in no scanning on this priority and a potential
1890 * priority drop. Global direct reclaim can go to the next
1891 * zone and tends to have no problems. Global kswapd is for
1892 * zone balancing and it needs to scan a minimum amount. When
1893 * reclaiming for a memcg, a priority drop can cause high
1894 * latencies, so it's better to scan a minimum amount there as
1895 * well.
1896 */
Lisa Du6e543d52013-09-11 14:22:36 -07001897 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001898 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001899 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001900 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001901
1902 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001903 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001904 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001905 goto out;
1906 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001907
Johannes Weiner10316b32013-02-22 16:32:14 -08001908 /*
1909 * Global reclaim will swap to prevent OOM even with no
1910 * swappiness, but memcg users want to use this knob to
1911 * disable swapping for individual groups completely when
1912 * using the memory controller's swap limit feature would be
1913 * too expensive.
1914 */
1915 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001916 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001917 goto out;
1918 }
1919
1920 /*
1921 * Do not apply any pressure balancing cleverness when the
1922 * system is close to OOM, scan both anon and file equally
1923 * (unless the swappiness setting disagrees with swapping).
1924 */
1925 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001926 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001927 goto out;
1928 }
1929
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001930 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1931 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1932 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1933 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001934
Johannes Weiner11d16c22013-02-22 16:32:15 -08001935 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001936 * Prevent the reclaimer from falling into the cache trap: as
1937 * cache pages start out inactive, every cache fault will tip
1938 * the scan balance towards the file LRU. And as the file LRU
1939 * shrinks, so does the window for rotation from references.
1940 * This means we have a runaway feedback loop where a tiny
1941 * thrashing file LRU becomes infinitely more attractive than
1942 * anon pages. Try to detect this based on file LRU size.
1943 */
1944 if (global_reclaim(sc)) {
1945 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1946
1947 if (unlikely(file + free <= high_wmark_pages(zone))) {
1948 scan_balance = SCAN_ANON;
1949 goto out;
1950 }
1951 }
1952
1953 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001954 * There is enough inactive page cache, do not reclaim
1955 * anything from the anonymous working set right now.
1956 */
1957 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001958 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001959 goto out;
1960 }
1961
Johannes Weiner9a265112013-02-22 16:32:17 -08001962 scan_balance = SCAN_FRACT;
1963
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001964 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001965 * With swappiness at 100, anonymous and file have the same priority.
1966 * This scanning priority is essentially the inverse of IO cost.
1967 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001968 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001969 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001970
1971 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001972 * OK, so we have swap space and a fair amount of page cache
1973 * pages. We use the recently rotated / recently scanned
1974 * ratios to determine how valuable each cache is.
1975 *
1976 * Because workloads change over time (and to avoid overflow)
1977 * we keep these statistics as a floating average, which ends
1978 * up weighing recent references more than old ones.
1979 *
1980 * anon in [0], file in [1]
1981 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001982 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001983 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001984 reclaim_stat->recent_scanned[0] /= 2;
1985 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001986 }
1987
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001988 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001989 reclaim_stat->recent_scanned[1] /= 2;
1990 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001991 }
1992
1993 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001994 * The amount of pressure on anon vs file pages is inversely
1995 * proportional to the fraction of recently scanned pages on
1996 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001997 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001998 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001999 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002000
Satoru Moriyafe350042012-05-29 15:06:47 -07002001 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002002 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002003 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002004
Shaohua Li76a33fc2010-05-24 14:32:36 -07002005 fraction[0] = ap;
2006 fraction[1] = fp;
2007 denominator = ap + fp + 1;
2008out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002009 some_scanned = false;
2010 /* Only use force_scan on second pass. */
2011 for (pass = 0; !some_scanned && pass < 2; pass++) {
2012 for_each_evictable_lru(lru) {
2013 int file = is_file_lru(lru);
2014 unsigned long size;
2015 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002016
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002017 size = get_lru_size(lruvec, lru);
2018 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002019
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002020 if (!scan && pass && force_scan)
2021 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002022
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002023 switch (scan_balance) {
2024 case SCAN_EQUAL:
2025 /* Scan lists relative to size */
2026 break;
2027 case SCAN_FRACT:
2028 /*
2029 * Scan types proportional to swappiness and
2030 * their relative recent reclaim efficiency.
2031 */
2032 scan = div64_u64(scan * fraction[file],
2033 denominator);
2034 break;
2035 case SCAN_FILE:
2036 case SCAN_ANON:
2037 /* Scan one type exclusively */
2038 if ((scan_balance == SCAN_FILE) != file)
2039 scan = 0;
2040 break;
2041 default:
2042 /* Look ma, no brain */
2043 BUG();
2044 }
2045 nr[lru] = scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002046 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002047 * Skip the second pass and don't force_scan,
2048 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002049 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002050 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002051 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002052 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002053}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002054
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002055/*
2056 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2057 */
2058static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2059{
2060 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002061 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002062 unsigned long nr_to_scan;
2063 enum lru_list lru;
2064 unsigned long nr_reclaimed = 0;
2065 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2066 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002067 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002068
2069 get_scan_count(lruvec, sc, nr);
2070
Mel Gormane82e0562013-07-03 15:01:44 -07002071 /* Record the original scan target for proportional adjustments later */
2072 memcpy(targets, nr, sizeof(nr));
2073
Mel Gorman1a501902014-06-04 16:10:49 -07002074 /*
2075 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2076 * event that can occur when there is little memory pressure e.g.
2077 * multiple streaming readers/writers. Hence, we do not abort scanning
2078 * when the requested number of pages are reclaimed when scanning at
2079 * DEF_PRIORITY on the assumption that the fact we are direct
2080 * reclaiming implies that kswapd is not keeping up and it is best to
2081 * do a batch of work at once. For memcg reclaim one check is made to
2082 * abort proportional reclaim if either the file or anon lru has already
2083 * dropped to zero at the first pass.
2084 */
2085 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2086 sc->priority == DEF_PRIORITY);
2087
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002088 blk_start_plug(&plug);
2089 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2090 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002091 unsigned long nr_anon, nr_file, percentage;
2092 unsigned long nr_scanned;
2093
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002094 for_each_evictable_lru(lru) {
2095 if (nr[lru]) {
2096 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2097 nr[lru] -= nr_to_scan;
2098
2099 nr_reclaimed += shrink_list(lru, nr_to_scan,
2100 lruvec, sc);
2101 }
2102 }
Mel Gormane82e0562013-07-03 15:01:44 -07002103
2104 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2105 continue;
2106
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002107 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002108 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002109 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002110 * proportionally what was requested by get_scan_count(). We
2111 * stop reclaiming one LRU and reduce the amount scanning
2112 * proportional to the original scan target.
2113 */
2114 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2115 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2116
Mel Gorman1a501902014-06-04 16:10:49 -07002117 /*
2118 * It's just vindictive to attack the larger once the smaller
2119 * has gone to zero. And given the way we stop scanning the
2120 * smaller below, this makes sure that we only make one nudge
2121 * towards proportionality once we've got nr_to_reclaim.
2122 */
2123 if (!nr_file || !nr_anon)
2124 break;
2125
Mel Gormane82e0562013-07-03 15:01:44 -07002126 if (nr_file > nr_anon) {
2127 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2128 targets[LRU_ACTIVE_ANON] + 1;
2129 lru = LRU_BASE;
2130 percentage = nr_anon * 100 / scan_target;
2131 } else {
2132 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2133 targets[LRU_ACTIVE_FILE] + 1;
2134 lru = LRU_FILE;
2135 percentage = nr_file * 100 / scan_target;
2136 }
2137
2138 /* Stop scanning the smaller of the LRU */
2139 nr[lru] = 0;
2140 nr[lru + LRU_ACTIVE] = 0;
2141
2142 /*
2143 * Recalculate the other LRU scan count based on its original
2144 * scan target and the percentage scanning already complete
2145 */
2146 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2147 nr_scanned = targets[lru] - nr[lru];
2148 nr[lru] = targets[lru] * (100 - percentage) / 100;
2149 nr[lru] -= min(nr[lru], nr_scanned);
2150
2151 lru += LRU_ACTIVE;
2152 nr_scanned = targets[lru] - nr[lru];
2153 nr[lru] = targets[lru] * (100 - percentage) / 100;
2154 nr[lru] -= min(nr[lru], nr_scanned);
2155
2156 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002157 }
2158 blk_finish_plug(&plug);
2159 sc->nr_reclaimed += nr_reclaimed;
2160
2161 /*
2162 * Even if we did not try to evict anon pages at all, we want to
2163 * rebalance the anon lru active/inactive ratio.
2164 */
2165 if (inactive_anon_is_low(lruvec))
2166 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2167 sc, LRU_ACTIVE_ANON);
2168
2169 throttle_vm_writeout(sc->gfp_mask);
2170}
2171
Mel Gorman23b9da52012-05-29 15:06:20 -07002172/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002173static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002174{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002175 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002176 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002177 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002178 return true;
2179
2180 return false;
2181}
2182
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002183/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002184 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2185 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2186 * true if more pages should be reclaimed such that when the page allocator
2187 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2188 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002189 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002190static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002191 unsigned long nr_reclaimed,
2192 unsigned long nr_scanned,
2193 struct scan_control *sc)
2194{
2195 unsigned long pages_for_compaction;
2196 unsigned long inactive_lru_pages;
2197
2198 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002199 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002200 return false;
2201
Mel Gorman28765922011-02-25 14:44:20 -08002202 /* Consider stopping depending on scan and reclaim activity */
2203 if (sc->gfp_mask & __GFP_REPEAT) {
2204 /*
2205 * For __GFP_REPEAT allocations, stop reclaiming if the
2206 * full LRU list has been scanned and we are still failing
2207 * to reclaim pages. This full LRU scan is potentially
2208 * expensive but a __GFP_REPEAT caller really wants to succeed
2209 */
2210 if (!nr_reclaimed && !nr_scanned)
2211 return false;
2212 } else {
2213 /*
2214 * For non-__GFP_REPEAT allocations which can presumably
2215 * fail without consequence, stop if we failed to reclaim
2216 * any pages from the last SWAP_CLUSTER_MAX number of
2217 * pages that were scanned. This will return to the
2218 * caller faster at the risk reclaim/compaction and
2219 * the resulting allocation attempt fails
2220 */
2221 if (!nr_reclaimed)
2222 return false;
2223 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002224
2225 /*
2226 * If we have not reclaimed enough pages for compaction and the
2227 * inactive lists are large enough, continue reclaiming
2228 */
2229 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002230 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002231 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002232 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002233 if (sc->nr_reclaimed < pages_for_compaction &&
2234 inactive_lru_pages > pages_for_compaction)
2235 return true;
2236
2237 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002238 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002239 case COMPACT_PARTIAL:
2240 case COMPACT_CONTINUE:
2241 return false;
2242 default:
2243 return true;
2244 }
2245}
2246
Andrew Morton0608f432013-09-24 15:27:41 -07002247static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002248{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002249 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002250
Johannes Weiner56600482012-01-12 17:17:59 -08002251 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002252 struct mem_cgroup *root = sc->target_mem_cgroup;
2253 struct mem_cgroup_reclaim_cookie reclaim = {
2254 .zone = zone,
2255 .priority = sc->priority,
2256 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002257 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002258
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002259 nr_reclaimed = sc->nr_reclaimed;
2260 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002261
Andrew Morton694fbc02013-09-24 15:27:37 -07002262 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2263 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002264 struct lruvec *lruvec;
2265
2266 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2267
2268 shrink_lruvec(lruvec, sc);
2269
2270 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002271 * Direct reclaim and kswapd have to scan all memory
2272 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002273 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002274 *
2275 * Limit reclaim, on the other hand, only cares about
2276 * nr_to_reclaim pages to be reclaimed and it will
2277 * retry with decreasing priority if one round over the
2278 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002279 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002280 if (!global_reclaim(sc) &&
2281 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002282 mem_cgroup_iter_break(root, memcg);
2283 break;
2284 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002285 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2286 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002287
2288 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2289 sc->nr_scanned - nr_scanned,
2290 sc->nr_reclaimed - nr_reclaimed);
2291
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002292 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2293 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002294}
2295
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002296/* Returns true if compaction should go ahead for a high-order request */
2297static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2298{
2299 unsigned long balance_gap, watermark;
2300 bool watermark_ok;
2301
2302 /* Do not consider compaction for orders reclaim is meant to satisfy */
2303 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2304 return false;
2305
2306 /*
2307 * Compaction takes time to run and there are potentially other
2308 * callers using the pages just freed. Continue reclaiming until
2309 * there is a buffer of free pages available to give compaction
2310 * a reasonable chance of completing and allocating the page
2311 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002312 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2313 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002314 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2315 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2316
2317 /*
2318 * If compaction is deferred, reclaim up to a point where
2319 * compaction will have a chance of success when re-enabled
2320 */
Rik van Rielaff62242012-03-21 16:33:52 -07002321 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002322 return watermark_ok;
2323
2324 /* If compaction is not ready to start, keep reclaiming */
2325 if (!compaction_suitable(zone, sc->order))
2326 return false;
2327
2328 return watermark_ok;
2329}
2330
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331/*
2332 * This is the direct reclaim path, for page-allocating processes. We only
2333 * try to reclaim pages from zones which will satisfy the caller's allocation
2334 * request.
2335 *
Mel Gorman41858962009-06-16 15:32:12 -07002336 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2337 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2339 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002340 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2341 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2342 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 * If a zone is deemed to be full of pinned pages then just give it a light
2345 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002346 *
2347 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002348 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002349 * the caller that it should consider retrying the allocation instead of
2350 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 */
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002352static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353{
Mel Gormandd1a2392008-04-28 02:12:17 -07002354 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002355 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002356 unsigned long nr_soft_reclaimed;
2357 unsigned long nr_soft_scanned;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002358 unsigned long lru_pages = 0;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002359 bool aborted_reclaim = false;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002360 struct reclaim_state *reclaim_state = current->reclaim_state;
Weijie Yang619d0d762014-04-07 15:36:59 -07002361 gfp_t orig_mask;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002362 struct shrink_control shrink = {
2363 .gfp_mask = sc->gfp_mask,
2364 };
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002365 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002366
Mel Gormancc715d92012-03-21 16:34:00 -07002367 /*
2368 * If the number of buffer_heads in the machine exceeds the maximum
2369 * allowed level, force direct reclaim to scan the highmem zone as
2370 * highmem pages could be pinning lowmem pages storing buffer_heads
2371 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002372 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002373 if (buffer_heads_over_limit)
2374 sc->gfp_mask |= __GFP_HIGHMEM;
2375
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002376 nodes_clear(shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002377
Mel Gormand4debc62010-08-09 17:19:29 -07002378 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2379 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002380 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002382 /*
2383 * Take care memory controller reclaiming has small influence
2384 * to global LRU.
2385 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002386 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002387 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2388 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002389
2390 lru_pages += zone_reclaimable_pages(zone);
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002391 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002392
Lisa Du6e543d52013-09-11 14:22:36 -07002393 if (sc->priority != DEF_PRIORITY &&
2394 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002395 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002396 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002397 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002398 * If we already have plenty of memory free for
2399 * compaction in this zone, don't free any more.
2400 * Even though compaction is invoked for any
2401 * non-zero order, only frequent costly order
2402 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002403 * noticeable problem, like transparent huge
2404 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002405 */
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002406 if ((zonelist_zone_idx(z) <= requested_highidx)
2407 && compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002408 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002409 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002410 }
Rik van Riele0887c12011-10-31 17:09:31 -07002411 }
Andrew Morton0608f432013-09-24 15:27:41 -07002412 /*
2413 * This steals pages from memory cgroups over softlimit
2414 * and returns the number of reclaimed pages and
2415 * scanned pages. This works for global memory pressure
2416 * and balancing, not for a memcg's limit.
2417 */
2418 nr_soft_scanned = 0;
2419 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2420 sc->order, sc->gfp_mask,
2421 &nr_soft_scanned);
2422 sc->nr_reclaimed += nr_soft_reclaimed;
2423 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002424 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002425 }
Nick Piggin408d8542006-09-25 23:31:27 -07002426
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002427 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 }
Mel Gormane0c23272011-10-31 17:09:33 -07002429
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002430 /*
2431 * Don't shrink slabs when reclaiming memory from over limit cgroups
2432 * but do shrink slab at least once when aborting reclaim for
2433 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2434 * pages.
2435 */
2436 if (global_reclaim(sc)) {
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002437 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002438 if (reclaim_state) {
2439 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2440 reclaim_state->reclaimed_slab = 0;
2441 }
2442 }
2443
Weijie Yang619d0d762014-04-07 15:36:59 -07002444 /*
2445 * Restore to original mask to avoid the impact on the caller if we
2446 * promoted it to __GFP_HIGHMEM.
2447 */
2448 sc->gfp_mask = orig_mask;
2449
Mel Gorman0cee34f2012-01-12 17:19:49 -08002450 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002451}
2452
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002453/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002454static bool all_unreclaimable(struct zonelist *zonelist,
2455 struct scan_control *sc)
2456{
2457 struct zoneref *z;
2458 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002459
2460 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2461 gfp_zone(sc->gfp_mask), sc->nodemask) {
2462 if (!populated_zone(zone))
2463 continue;
2464 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2465 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002466 if (zone_reclaimable(zone))
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002467 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002468 }
2469
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002470 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002472
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473/*
2474 * This is the main entry point to direct page reclaim.
2475 *
2476 * If a full scan of the inactive list fails to free enough memory then we
2477 * are "out of memory" and something needs to be killed.
2478 *
2479 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2480 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002481 * caller can't do much about. We kick the writeback threads and take explicit
2482 * naps in the hope that some of these pages can be written. But if the
2483 * allocating task holds filesystem locks which prevent writeout this might not
2484 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002485 *
2486 * returns: 0, if no pages reclaimed
2487 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 */
Mel Gormandac1d272008-04-28 02:12:12 -07002489static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002490 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491{
Andrew Morton69e05942006-03-22 00:08:19 -08002492 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002493 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002494 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Keika Kobayashi873b4772008-07-25 01:48:52 -07002496 delayacct_freepages_start();
2497
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002498 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002499 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002501 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002502 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2503 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002504 sc->nr_scanned = 0;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002505 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002506
Balbir Singh66e17072008-02-07 00:13:56 -08002507 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002508 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510
2511 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002512 * If we're getting trouble reclaiming, start doing
2513 * writepage even in laptop mode.
2514 */
2515 if (sc->priority < DEF_PRIORITY - 2)
2516 sc->may_writepage = 1;
2517
2518 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 * Try to write back as many pages as we just scanned. This
2520 * tends to cause slow streaming writers to write data to the
2521 * disk smoothly, at the dirtying rate, which is nice. But
2522 * that's undesirable in laptop mode, where we *want* lumpy
2523 * writeout. So in laptop mode, write out the whole world.
2524 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002525 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2526 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002527 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2528 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002529 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 }
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002531 } while (--sc->priority >= 0 && !aborted_reclaim);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002532
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002534 delayacct_freepages_end();
2535
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002536 if (sc->nr_reclaimed)
2537 return sc->nr_reclaimed;
2538
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002539 /*
2540 * As hibernation is going on, kswapd is freezed so that it can't mark
2541 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2542 * check.
2543 */
2544 if (oom_killer_disabled)
2545 return 0;
2546
Mel Gorman0cee34f2012-01-12 17:19:49 -08002547 /* Aborted reclaim to try compaction? don't OOM, then */
2548 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002549 return 1;
2550
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002551 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002552 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002553 return 1;
2554
2555 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556}
2557
Mel Gorman55150612012-07-31 16:44:35 -07002558static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2559{
2560 struct zone *zone;
2561 unsigned long pfmemalloc_reserve = 0;
2562 unsigned long free_pages = 0;
2563 int i;
2564 bool wmark_ok;
2565
2566 for (i = 0; i <= ZONE_NORMAL; i++) {
2567 zone = &pgdat->node_zones[i];
Mel Gorman675becc2014-06-04 16:07:35 -07002568 if (!populated_zone(zone))
2569 continue;
2570
Mel Gorman55150612012-07-31 16:44:35 -07002571 pfmemalloc_reserve += min_wmark_pages(zone);
2572 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2573 }
2574
Mel Gorman675becc2014-06-04 16:07:35 -07002575 /* If there are no reserves (unexpected config) then do not throttle */
2576 if (!pfmemalloc_reserve)
2577 return true;
2578
Mel Gorman55150612012-07-31 16:44:35 -07002579 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2580
2581 /* kswapd must be awake if processes are being throttled */
2582 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2583 pgdat->classzone_idx = min(pgdat->classzone_idx,
2584 (enum zone_type)ZONE_NORMAL);
2585 wake_up_interruptible(&pgdat->kswapd_wait);
2586 }
2587
2588 return wmark_ok;
2589}
2590
2591/*
2592 * Throttle direct reclaimers if backing storage is backed by the network
2593 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2594 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002595 * when the low watermark is reached.
2596 *
2597 * Returns true if a fatal signal was delivered during throttling. If this
2598 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002599 */
Mel Gorman50694c22012-11-26 16:29:48 -08002600static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002601 nodemask_t *nodemask)
2602{
Mel Gorman675becc2014-06-04 16:07:35 -07002603 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002604 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002605 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002606
2607 /*
2608 * Kernel threads should not be throttled as they may be indirectly
2609 * responsible for cleaning pages necessary for reclaim to make forward
2610 * progress. kjournald for example may enter direct reclaim while
2611 * committing a transaction where throttling it could forcing other
2612 * processes to block on log_wait_commit().
2613 */
2614 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002615 goto out;
2616
2617 /*
2618 * If a fatal signal is pending, this process should not throttle.
2619 * It should return quickly so it can exit and free its memory
2620 */
2621 if (fatal_signal_pending(current))
2622 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002623
Mel Gorman675becc2014-06-04 16:07:35 -07002624 /*
2625 * Check if the pfmemalloc reserves are ok by finding the first node
2626 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2627 * GFP_KERNEL will be required for allocating network buffers when
2628 * swapping over the network so ZONE_HIGHMEM is unusable.
2629 *
2630 * Throttling is based on the first usable node and throttled processes
2631 * wait on a queue until kswapd makes progress and wakes them. There
2632 * is an affinity then between processes waking up and where reclaim
2633 * progress has been made assuming the process wakes on the same node.
2634 * More importantly, processes running on remote nodes will not compete
2635 * for remote pfmemalloc reserves and processes on different nodes
2636 * should make reasonable progress.
2637 */
2638 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2639 gfp_mask, nodemask) {
2640 if (zone_idx(zone) > ZONE_NORMAL)
2641 continue;
2642
2643 /* Throttle based on the first usable node */
2644 pgdat = zone->zone_pgdat;
2645 if (pfmemalloc_watermark_ok(pgdat))
2646 goto out;
2647 break;
2648 }
2649
2650 /* If no zone was usable by the allocation flags then do not throttle */
2651 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002652 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002653
Mel Gorman68243e72012-07-31 16:44:39 -07002654 /* Account for the throttling */
2655 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2656
Mel Gorman55150612012-07-31 16:44:35 -07002657 /*
2658 * If the caller cannot enter the filesystem, it's possible that it
2659 * is due to the caller holding an FS lock or performing a journal
2660 * transaction in the case of a filesystem like ext[3|4]. In this case,
2661 * it is not safe to block on pfmemalloc_wait as kswapd could be
2662 * blocked waiting on the same lock. Instead, throttle for up to a
2663 * second before continuing.
2664 */
2665 if (!(gfp_mask & __GFP_FS)) {
2666 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2667 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002668
2669 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002670 }
2671
2672 /* Throttle until kswapd wakes the process */
2673 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2674 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002675
2676check_pending:
2677 if (fatal_signal_pending(current))
2678 return true;
2679
2680out:
2681 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002682}
2683
Mel Gormandac1d272008-04-28 02:12:12 -07002684unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002685 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002686{
Mel Gorman33906bc2010-08-09 17:19:16 -07002687 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002688 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002689 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002690 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002691 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002692 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002693 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002694 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002695 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002696 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002697 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002698 };
2699
Mel Gorman55150612012-07-31 16:44:35 -07002700 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002701 * Do not enter reclaim if fatal signal was delivered while throttled.
2702 * 1 is returned so that the page allocator does not OOM kill at this
2703 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002704 */
Mel Gorman50694c22012-11-26 16:29:48 -08002705 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002706 return 1;
2707
Mel Gorman33906bc2010-08-09 17:19:16 -07002708 trace_mm_vmscan_direct_reclaim_begin(order,
2709 sc.may_writepage,
2710 gfp_mask);
2711
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002712 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002713
2714 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2715
2716 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002717}
2718
Andrew Mortonc255a452012-07-31 16:43:02 -07002719#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002720
Johannes Weiner72835c82012-01-12 17:18:32 -08002721unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002722 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002723 struct zone *zone,
2724 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002725{
2726 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002727 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002728 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002729 .may_writepage = !laptop_mode,
2730 .may_unmap = 1,
2731 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002732 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002733 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002734 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002735 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002736 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002737
Balbir Singh4e416952009-09-23 15:56:39 -07002738 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2739 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002740
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002741 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002742 sc.may_writepage,
2743 sc.gfp_mask);
2744
Balbir Singh4e416952009-09-23 15:56:39 -07002745 /*
2746 * NOTE: Although we can get the priority field, using it
2747 * here is not a good idea, since it limits the pages we can scan.
2748 * if we don't reclaim here, the shrink_zone from balance_pgdat
2749 * will pick up pages from other mem cgroup's as well. We hack
2750 * the priority and make it zero.
2751 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002752 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002753
2754 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2755
Ying Han0ae5e892011-05-26 16:25:25 -07002756 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002757 return sc.nr_reclaimed;
2758}
2759
Johannes Weiner72835c82012-01-12 17:18:32 -08002760unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002761 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002762 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002763{
Balbir Singh4e416952009-09-23 15:56:39 -07002764 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002765 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002766 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002767 struct scan_control sc = {
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,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002771 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002772 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002773 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002774 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002775 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002776 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2777 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2778 };
Balbir Singh66e17072008-02-07 00:13:56 -08002779
Ying Han889976d2011-05-26 16:25:33 -07002780 /*
2781 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2782 * take care of from where we get pages. So the node where we start the
2783 * scan does not need to be the current node.
2784 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002785 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002786
2787 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002788
2789 trace_mm_vmscan_memcg_reclaim_begin(0,
2790 sc.may_writepage,
2791 sc.gfp_mask);
2792
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002793 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002794
2795 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2796
2797 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002798}
2799#endif
2800
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002801static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002802{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002803 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002804
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002805 if (!total_swap_pages)
2806 return;
2807
2808 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2809 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002810 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002811
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002812 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002813 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002814 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002815
2816 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2817 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002818}
2819
Johannes Weiner60cefed2012-11-29 13:54:23 -08002820static bool zone_balanced(struct zone *zone, int order,
2821 unsigned long balance_gap, int classzone_idx)
2822{
2823 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2824 balance_gap, classzone_idx, 0))
2825 return false;
2826
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002827 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2828 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002829 return false;
2830
2831 return true;
2832}
2833
Mel Gorman1741c872011-01-13 15:46:21 -08002834/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002835 * pgdat_balanced() is used when checking if a node is balanced.
2836 *
2837 * For order-0, all zones must be balanced!
2838 *
2839 * For high-order allocations only zones that meet watermarks and are in a
2840 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2841 * total of balanced pages must be at least 25% of the zones allowed by
2842 * classzone_idx for the node to be considered balanced. Forcing all zones to
2843 * be balanced for high orders can cause excessive reclaim when there are
2844 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002845 * The choice of 25% is due to
2846 * o a 16M DMA zone that is balanced will not balance a zone on any
2847 * reasonable sized machine
2848 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002849 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002850 * would need to be at least 256M for it to be balance a whole node.
2851 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2852 * to balance a node on its own. These seemed like reasonable ratios.
2853 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002854static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002855{
Jiang Liub40da042013-02-22 16:33:52 -08002856 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002857 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002858 int i;
2859
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002860 /* Check the watermark levels */
2861 for (i = 0; i <= classzone_idx; i++) {
2862 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002863
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002864 if (!populated_zone(zone))
2865 continue;
2866
Jiang Liub40da042013-02-22 16:33:52 -08002867 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002868
2869 /*
2870 * A special case here:
2871 *
2872 * balance_pgdat() skips over all_unreclaimable after
2873 * DEF_PRIORITY. Effectively, it considers them balanced so
2874 * they must be considered balanced here as well!
2875 */
Lisa Du6e543d52013-09-11 14:22:36 -07002876 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002877 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002878 continue;
2879 }
2880
2881 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002882 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002883 else if (!order)
2884 return false;
2885 }
2886
2887 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002888 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002889 else
2890 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002891}
2892
Mel Gorman55150612012-07-31 16:44:35 -07002893/*
2894 * Prepare kswapd for sleeping. This verifies that there are no processes
2895 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2896 *
2897 * Returns true if kswapd is ready to sleep
2898 */
2899static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002900 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002901{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002902 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2903 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002904 return false;
2905
2906 /*
2907 * There is a potential race between when kswapd checks its watermarks
2908 * and a process gets throttled. There is also a potential race if
2909 * processes get throttled, kswapd wakes, a large process exits therby
2910 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2911 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2912 * so wake them now if necessary. If necessary, processes will wake
2913 * kswapd and get throttled again
2914 */
2915 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2916 wake_up(&pgdat->pfmemalloc_wait);
2917 return false;
2918 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002919
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002920 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002921}
2922
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923/*
Mel Gorman75485362013-07-03 15:01:42 -07002924 * kswapd shrinks the zone by the number of pages required to reach
2925 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002926 *
2927 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002928 * reclaim or if the lack of progress was due to pages under writeback.
2929 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002930 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002931static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002932 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002933 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002934 unsigned long lru_pages,
2935 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002936{
Mel Gorman7c954f62013-07-03 15:01:54 -07002937 int testorder = sc->order;
2938 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002939 struct reclaim_state *reclaim_state = current->reclaim_state;
2940 struct shrink_control shrink = {
2941 .gfp_mask = sc->gfp_mask,
2942 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002943 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002944
2945 /* Reclaim above the high watermark. */
2946 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002947
2948 /*
2949 * Kswapd reclaims only single pages with compaction enabled. Trying
2950 * too hard to reclaim until contiguous free pages have become
2951 * available can hurt performance by evicting too much useful data
2952 * from memory. Do not reclaim more than needed for compaction.
2953 */
2954 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2955 compaction_suitable(zone, sc->order) !=
2956 COMPACT_SKIPPED)
2957 testorder = 0;
2958
2959 /*
2960 * We put equal pressure on every zone, unless one zone has way too
2961 * many pages free already. The "too many pages" is defined as the
2962 * high wmark plus a "gap" where the gap is either the low
2963 * watermark or 1% of the zone, whichever is smaller.
2964 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002965 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2966 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07002967
2968 /*
2969 * If there is no low memory pressure or the zone is balanced then no
2970 * reclaim is necessary
2971 */
2972 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2973 if (!lowmem_pressure && zone_balanced(zone, testorder,
2974 balance_gap, classzone_idx))
2975 return true;
2976
Mel Gorman75485362013-07-03 15:01:42 -07002977 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002978 nodes_clear(shrink.nodes_to_scan);
2979 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002980
2981 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002982 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002983 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2984
Mel Gorman2ab44f42013-07-03 15:01:47 -07002985 /* Account for the number of pages attempted to reclaim */
2986 *nr_attempted += sc->nr_to_reclaim;
2987
Mel Gorman283aba92013-07-03 15:01:51 -07002988 zone_clear_flag(zone, ZONE_WRITEBACK);
2989
Mel Gorman7c954f62013-07-03 15:01:54 -07002990 /*
2991 * If a zone reaches its high watermark, consider it to be no longer
2992 * congested. It's possible there are dirty pages backed by congested
2993 * BDIs but as pressure is relieved, speculatively avoid congestion
2994 * waits.
2995 */
Lisa Du6e543d52013-09-11 14:22:36 -07002996 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002997 zone_balanced(zone, testorder, 0, classzone_idx)) {
2998 zone_clear_flag(zone, ZONE_CONGESTED);
2999 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3000 }
3001
Mel Gormanb8e83b92013-07-03 15:01:45 -07003002 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003003}
3004
3005/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003007 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008 *
Mel Gorman0abdee22011-01-13 15:46:22 -08003009 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 *
3011 * There is special handling here for zones which are full of pinned pages.
3012 * This can happen if the pages are all mlocked, or if they are all used by
3013 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3014 * What we do is to detect the case where all pages in the zone have been
3015 * scanned twice and there has been zero successful reclaim. Mark the zone as
3016 * dead and from now on, only perform a short scan. Basically we're polling
3017 * the zone for when the problem goes away.
3018 *
3019 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003020 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3021 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3022 * lower zones regardless of the number of free pages in the lower zones. This
3023 * interoperates with the page allocator fallback scheme to ensure that aging
3024 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 */
Mel Gorman99504742011-01-13 15:46:20 -08003026static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08003027 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003030 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003031 unsigned long nr_soft_reclaimed;
3032 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003033 struct scan_control sc = {
3034 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003035 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003036 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003037 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003038 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07003039 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08003040 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08003041 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003042 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003044 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003046 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003047 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07003048 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07003049
3050 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003052 /*
3053 * Scan in the highmem->dma direction for the highest
3054 * zone which needs scanning
3055 */
3056 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3057 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003059 if (!populated_zone(zone))
3060 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061
Lisa Du6e543d52013-09-11 14:22:36 -07003062 if (sc.priority != DEF_PRIORITY &&
3063 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003064 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065
Rik van Riel556adec2008-10-18 20:26:34 -07003066 /*
3067 * Do some background aging of the anon list, to give
3068 * pages a chance to be referenced before reclaiming.
3069 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003070 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003071
Mel Gormancc715d92012-03-21 16:34:00 -07003072 /*
3073 * If the number of buffer_heads in the machine
3074 * exceeds the maximum allowed level and this node
3075 * has a highmem zone, force kswapd to reclaim from
3076 * it to relieve lowmem pressure.
3077 */
3078 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3079 end_zone = i;
3080 break;
3081 }
3082
Johannes Weiner60cefed2012-11-29 13:54:23 -08003083 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003084 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003085 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003086 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003087 /*
3088 * If balanced, clear the dirty and congested
3089 * flags
3090 */
Shaohua Li439423f2011-08-25 15:59:12 -07003091 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003092 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003095
Mel Gormanb8e83b92013-07-03 15:01:45 -07003096 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003097 goto out;
3098
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 for (i = 0; i <= end_zone; i++) {
3100 struct zone *zone = pgdat->node_zones + i;
3101
Mel Gorman2ab44f42013-07-03 15:01:47 -07003102 if (!populated_zone(zone))
3103 continue;
3104
Wu Fengguangadea02a2009-09-21 17:01:42 -07003105 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003106
3107 /*
3108 * If any zone is currently balanced then kswapd will
3109 * not call compaction as it is expected that the
3110 * necessary pages are already available.
3111 */
3112 if (pgdat_needs_compaction &&
3113 zone_watermark_ok(zone, order,
3114 low_wmark_pages(zone),
3115 *classzone_idx, 0))
3116 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 }
3118
3119 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003120 * If we're getting trouble reclaiming, start doing writepage
3121 * even in laptop mode.
3122 */
3123 if (sc.priority < DEF_PRIORITY - 2)
3124 sc.may_writepage = 1;
3125
3126 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 * Now scan the zone in the dma->highmem direction, stopping
3128 * at the last zone which needs scanning.
3129 *
3130 * We do this because the page allocator works in the opposite
3131 * direction. This prevents the page allocator from allocating
3132 * pages behind kswapd's direction of progress, which would
3133 * cause too much scanning of the lower zones.
3134 */
3135 for (i = 0; i <= end_zone; i++) {
3136 struct zone *zone = pgdat->node_zones + i;
3137
Con Kolivasf3fe6512006-01-06 00:11:15 -08003138 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 continue;
3140
Lisa Du6e543d52013-09-11 14:22:36 -07003141 if (sc.priority != DEF_PRIORITY &&
3142 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 continue;
3144
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003146
Andrew Morton0608f432013-09-24 15:27:41 -07003147 nr_soft_scanned = 0;
3148 /*
3149 * Call soft limit reclaim before calling shrink_zone.
3150 */
3151 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3152 order, sc.gfp_mask,
3153 &nr_soft_scanned);
3154 sc.nr_reclaimed += nr_soft_reclaimed;
3155
Rik van Riel32a43302007-10-16 01:24:50 -07003156 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003157 * There should be no need to raise the scanning
3158 * priority if enough pages are already being scanned
3159 * that that high watermark would be met at 100%
3160 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003161 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003162 if (kswapd_shrink_zone(zone, end_zone, &sc,
3163 lru_pages, &nr_attempted))
3164 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 }
Mel Gorman55150612012-07-31 16:44:35 -07003166
3167 /*
3168 * If the low watermark is met there is no need for processes
3169 * to be throttled on pfmemalloc_wait as they should not be
3170 * able to safely make forward progress. Wake them
3171 */
3172 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3173 pfmemalloc_watermark_ok(pgdat))
3174 wake_up(&pgdat->pfmemalloc_wait);
3175
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003177 * Fragmentation may mean that the system cannot be rebalanced
3178 * for high-order allocations in all zones. If twice the
3179 * allocation size has been reclaimed and the zones are still
3180 * not balanced then recheck the watermarks at order-0 to
3181 * prevent kswapd reclaiming excessively. Assume that a
3182 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003184 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003185 order = sc.order = 0;
3186
Mel Gormanb8e83b92013-07-03 15:01:45 -07003187 /* Check if kswapd should be suspending */
3188 if (try_to_freeze() || kthread_should_stop())
3189 break;
3190
3191 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003192 * Compact if necessary and kswapd is reclaiming at least the
3193 * high watermark number of pages as requsted
3194 */
3195 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3196 compact_pgdat(pgdat, order);
3197
3198 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003199 * Raise priority if scanning rate is too low or there was no
3200 * progress in reclaiming pages
3201 */
3202 if (raise_priority || !sc.nr_reclaimed)
3203 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003204 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003205 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206
Mel Gormanb8e83b92013-07-03 15:01:45 -07003207out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003208 /*
Mel Gorman55150612012-07-31 16:44:35 -07003209 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003210 * makes a decision on the order we were last reclaiming at. However,
3211 * if another caller entered the allocator slow path while kswapd
3212 * was awake, order will remain at the higher level
3213 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003214 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003215 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216}
3217
Mel Gormandc83edd2011-01-13 15:46:26 -08003218static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003219{
3220 long remaining = 0;
3221 DEFINE_WAIT(wait);
3222
3223 if (freezing(current) || kthread_should_stop())
3224 return;
3225
3226 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3227
3228 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003229 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003230 remaining = schedule_timeout(HZ/10);
3231 finish_wait(&pgdat->kswapd_wait, &wait);
3232 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3233 }
3234
3235 /*
3236 * After a short sleep, check if it was a premature sleep. If not, then
3237 * go fully to sleep until explicitly woken up.
3238 */
Mel Gorman55150612012-07-31 16:44:35 -07003239 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003240 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3241
3242 /*
3243 * vmstat counters are not perfectly accurate and the estimated
3244 * value for counters such as NR_FREE_PAGES can deviate from the
3245 * true value by nr_online_cpus * threshold. To avoid the zone
3246 * watermarks being breached while under pressure, we reduce the
3247 * per-cpu vmstat threshold while kswapd is awake and restore
3248 * them before going back to sleep.
3249 */
3250 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003251
Mel Gorman62997022012-10-08 16:32:47 -07003252 /*
3253 * Compaction records what page blocks it recently failed to
3254 * isolate pages from and skips them in the future scanning.
3255 * When kswapd is going to sleep, it is reasonable to assume
3256 * that pages and compaction may succeed so reset the cache.
3257 */
3258 reset_isolation_suitable(pgdat);
3259
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003260 if (!kthread_should_stop())
3261 schedule();
3262
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003263 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3264 } else {
3265 if (remaining)
3266 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3267 else
3268 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3269 }
3270 finish_wait(&pgdat->kswapd_wait, &wait);
3271}
3272
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273/*
3274 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003275 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 *
3277 * This basically trickles out pages so that we have _some_
3278 * free memory available even if there is no other activity
3279 * that frees anything up. This is needed for things like routing
3280 * etc, where we otherwise might have all activity going on in
3281 * asynchronous contexts that cannot page things out.
3282 *
3283 * If there are applications that are active memory-allocators
3284 * (most normal use), this basically shouldn't matter.
3285 */
3286static int kswapd(void *p)
3287{
Mel Gorman215ddd62011-07-08 15:39:40 -07003288 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003289 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003290 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003291 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 pg_data_t *pgdat = (pg_data_t*)p;
3293 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003294
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 struct reclaim_state reclaim_state = {
3296 .reclaimed_slab = 0,
3297 };
Rusty Russella70f7302009-03-13 14:49:46 +10303298 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Nick Piggincf40bd12009-01-21 08:12:39 +01003300 lockdep_set_current_reclaim_state(GFP_KERNEL);
3301
Rusty Russell174596a2009-01-01 10:12:29 +10303302 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003303 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304 current->reclaim_state = &reclaim_state;
3305
3306 /*
3307 * Tell the memory management that we're a "memory allocator",
3308 * and that if we need more memory we should get access to it
3309 * regardless (see "__alloc_pages()"). "kswapd" should
3310 * never get caught in the normal page freeing logic.
3311 *
3312 * (Kswapd normally doesn't need memory anyway, but sometimes
3313 * you need a small amount of memory in order to be able to
3314 * page out something else, and this flag essentially protects
3315 * us from recursively trying to free more memory as we're
3316 * trying to free the first piece of memory in the first place).
3317 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003318 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003319 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320
Mel Gorman215ddd62011-07-08 15:39:40 -07003321 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003322 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003323 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003324 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003326 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003327
Mel Gorman215ddd62011-07-08 15:39:40 -07003328 /*
3329 * If the last balance_pgdat was unsuccessful it's unlikely a
3330 * new request of a similar or harder type will succeed soon
3331 * so consider going to sleep on the basis we reclaimed at
3332 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003333 if (balanced_classzone_idx >= new_classzone_idx &&
3334 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003335 new_order = pgdat->kswapd_max_order;
3336 new_classzone_idx = pgdat->classzone_idx;
3337 pgdat->kswapd_max_order = 0;
3338 pgdat->classzone_idx = pgdat->nr_zones - 1;
3339 }
3340
Mel Gorman99504742011-01-13 15:46:20 -08003341 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342 /*
3343 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003344 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 */
3346 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003347 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003349 kswapd_try_to_sleep(pgdat, balanced_order,
3350 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003352 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003353 new_order = order;
3354 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003355 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003356 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358
David Rientjes8fe23e02009-12-14 17:58:33 -08003359 ret = try_to_freeze();
3360 if (kthread_should_stop())
3361 break;
3362
3363 /*
3364 * We can speed up thawing tasks if we don't call balance_pgdat
3365 * after returning from the refrigerator
3366 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003367 if (!ret) {
3368 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003369 balanced_classzone_idx = classzone_idx;
3370 balanced_order = balance_pgdat(pgdat, order,
3371 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003372 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003374
Johannes Weiner71abdc12014-06-06 14:35:35 -07003375 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003376 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003377 lockdep_clear_current_reclaim_state();
3378
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 return 0;
3380}
3381
3382/*
3383 * A zone is low on free memory, so wake its kswapd task to service it.
3384 */
Mel Gorman99504742011-01-13 15:46:20 -08003385void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386{
3387 pg_data_t *pgdat;
3388
Con Kolivasf3fe6512006-01-06 00:11:15 -08003389 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 return;
3391
Paul Jackson02a0e532006-12-13 00:34:25 -08003392 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003394 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003395 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003396 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003397 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3398 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003399 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003401 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003402 return;
3403
3404 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003405 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406}
3407
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003408#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003410 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003411 * freed pages.
3412 *
3413 * Rather than trying to age LRUs the aim is to preserve the overall
3414 * LRU order by reclaiming preferentially
3415 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003417unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003419 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003420 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003421 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3422 .may_swap = 1,
3423 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003424 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003425 .nr_to_reclaim = nr_to_reclaim,
3426 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003427 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003428 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003430 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003431 struct task_struct *p = current;
3432 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003434 p->flags |= PF_MEMALLOC;
3435 lockdep_set_current_reclaim_state(sc.gfp_mask);
3436 reclaim_state.reclaimed_slab = 0;
3437 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003438
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003439 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003440
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003441 p->reclaim_state = NULL;
3442 lockdep_clear_current_reclaim_state();
3443 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003444
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003445 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003447#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449/* It's optimal to keep kswapds on the same CPUs as their memory, but
3450 not required for correctness. So if the last cpu in a node goes
3451 away, we get changed to run anywhere: as the first one comes back,
3452 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003453static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3454 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003456 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003458 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003459 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003460 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303461 const struct cpumask *mask;
3462
3463 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003464
Rusty Russell3e597942009-01-01 10:12:24 +10303465 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003467 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 }
3469 }
3470 return NOTIFY_OK;
3471}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472
Yasunori Goto3218ae12006-06-27 02:53:33 -07003473/*
3474 * This kswapd start function will be called by init and node-hot-add.
3475 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3476 */
3477int kswapd_run(int nid)
3478{
3479 pg_data_t *pgdat = NODE_DATA(nid);
3480 int ret = 0;
3481
3482 if (pgdat->kswapd)
3483 return 0;
3484
3485 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3486 if (IS_ERR(pgdat->kswapd)) {
3487 /* failure at boot is fatal */
3488 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003489 pr_err("Failed to start kswapd on node %d\n", nid);
3490 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003491 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003492 }
3493 return ret;
3494}
3495
David Rientjes8fe23e02009-12-14 17:58:33 -08003496/*
Jiang Liud8adde12012-07-11 14:01:52 -07003497 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003498 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003499 */
3500void kswapd_stop(int nid)
3501{
3502 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3503
Jiang Liud8adde12012-07-11 14:01:52 -07003504 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003505 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003506 NODE_DATA(nid)->kswapd = NULL;
3507 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003508}
3509
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510static int __init kswapd_init(void)
3511{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003512 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003513
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003515 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003516 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 hotcpu_notifier(cpu_callback, 0);
3518 return 0;
3519}
3520
3521module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003522
3523#ifdef CONFIG_NUMA
3524/*
3525 * Zone reclaim mode
3526 *
3527 * If non-zero call zone_reclaim when the number of free pages falls below
3528 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003529 */
3530int zone_reclaim_mode __read_mostly;
3531
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003532#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003533#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003534#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3535#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3536
Christoph Lameter9eeff232006-01-18 17:42:31 -08003537/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003538 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3539 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3540 * a zone.
3541 */
3542#define ZONE_RECLAIM_PRIORITY 4
3543
Christoph Lameter9eeff232006-01-18 17:42:31 -08003544/*
Christoph Lameter96146342006-07-03 00:24:13 -07003545 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3546 * occur.
3547 */
3548int sysctl_min_unmapped_ratio = 1;
3549
3550/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003551 * If the number of slab pages in a zone grows beyond this percentage then
3552 * slab reclaim needs to occur.
3553 */
3554int sysctl_min_slab_ratio = 5;
3555
Mel Gorman90afa5d2009-06-16 15:33:20 -07003556static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3557{
3558 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3559 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3560 zone_page_state(zone, NR_ACTIVE_FILE);
3561
3562 /*
3563 * It's possible for there to be more file mapped pages than
3564 * accounted for by the pages on the file LRU lists because
3565 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3566 */
3567 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3568}
3569
3570/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3571static long zone_pagecache_reclaimable(struct zone *zone)
3572{
3573 long nr_pagecache_reclaimable;
3574 long delta = 0;
3575
3576 /*
3577 * If RECLAIM_SWAP is set, then all file pages are considered
3578 * potentially reclaimable. Otherwise, we have to worry about
3579 * pages like swapcache and zone_unmapped_file_pages() provides
3580 * a better estimate
3581 */
3582 if (zone_reclaim_mode & RECLAIM_SWAP)
3583 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3584 else
3585 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3586
3587 /* If we can't clean pages, remove dirty pages from consideration */
3588 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3589 delta += zone_page_state(zone, NR_FILE_DIRTY);
3590
3591 /* Watch for any possible underflows due to delta */
3592 if (unlikely(delta > nr_pagecache_reclaimable))
3593 delta = nr_pagecache_reclaimable;
3594
3595 return nr_pagecache_reclaimable - delta;
3596}
3597
Christoph Lameter0ff38492006-09-25 23:31:52 -07003598/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003599 * Try to free up some pages from this zone through reclaim.
3600 */
Andrew Morton179e9632006-03-22 00:08:18 -08003601static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003602{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003603 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003604 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003605 struct task_struct *p = current;
3606 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003607 struct scan_control sc = {
3608 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003609 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003610 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003611 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003612 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003613 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003614 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003615 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003616 struct shrink_control shrink = {
3617 .gfp_mask = sc.gfp_mask,
3618 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003619 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003620
Christoph Lameter9eeff232006-01-18 17:42:31 -08003621 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003622 /*
3623 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3624 * and we also need to be able to write out pages for RECLAIM_WRITE
3625 * and RECLAIM_SWAP.
3626 */
3627 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003628 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003629 reclaim_state.reclaimed_slab = 0;
3630 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003631
Mel Gorman90afa5d2009-06-16 15:33:20 -07003632 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003633 /*
3634 * Free memory by calling shrink zone with increasing
3635 * priorities until we have enough memory freed.
3636 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003637 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003638 shrink_zone(zone, &sc);
3639 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003640 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003641
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003642 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3643 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003644 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003645 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003646 * many pages were freed in this zone. So we take the current
3647 * number of slab pages and shake the slab until it is reduced
3648 * by the same nr_pages that we used for reclaiming unmapped
3649 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003650 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003651 nodes_clear(shrink.nodes_to_scan);
3652 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003653 for (;;) {
3654 unsigned long lru_pages = zone_reclaimable_pages(zone);
3655
3656 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003657 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003658 break;
3659
3660 /* Freed enough memory */
3661 nr_slab_pages1 = zone_page_state(zone,
3662 NR_SLAB_RECLAIMABLE);
3663 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3664 break;
3665 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003666
3667 /*
3668 * Update nr_reclaimed by the number of slab pages we
3669 * reclaimed from this zone.
3670 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003671 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3672 if (nr_slab_pages1 < nr_slab_pages0)
3673 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003674 }
3675
Christoph Lameter9eeff232006-01-18 17:42:31 -08003676 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003677 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003678 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003679 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003680}
Andrew Morton179e9632006-03-22 00:08:18 -08003681
3682int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3683{
Andrew Morton179e9632006-03-22 00:08:18 -08003684 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003685 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003686
3687 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003688 * Zone reclaim reclaims unmapped file backed pages and
3689 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003690 *
Christoph Lameter96146342006-07-03 00:24:13 -07003691 * A small portion of unmapped file backed pages is needed for
3692 * file I/O otherwise pages read by file I/O will be immediately
3693 * thrown out if the zone is overallocated. So we do not reclaim
3694 * if less than a specified percentage of the zone is used by
3695 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003696 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003697 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3698 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003699 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003700
Lisa Du6e543d52013-09-11 14:22:36 -07003701 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003702 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003703
Andrew Morton179e9632006-03-22 00:08:18 -08003704 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003705 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003706 */
David Rientjesd773ed62007-10-16 23:26:01 -07003707 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003708 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003709
3710 /*
3711 * Only run zone reclaim on the local zone or on zones that do not
3712 * have associated processors. This will favor the local processor
3713 * over remote processors and spread off node memory allocations
3714 * as wide as possible.
3715 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003716 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003717 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003718 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003719
3720 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003721 return ZONE_RECLAIM_NOSCAN;
3722
David Rientjesd773ed62007-10-16 23:26:01 -07003723 ret = __zone_reclaim(zone, gfp_mask, order);
3724 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3725
Mel Gorman24cf725182009-06-16 15:33:23 -07003726 if (!ret)
3727 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3728
David Rientjesd773ed62007-10-16 23:26:01 -07003729 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003730}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003731#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003732
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003733/*
3734 * page_evictable - test whether a page is evictable
3735 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003736 *
3737 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003738 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003739 *
3740 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003741 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003742 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003743 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003744 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003745int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003746{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003747 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003748}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003749
Hugh Dickins850465792012-01-20 14:34:19 -08003750#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003751/**
Hugh Dickins24513262012-01-20 14:34:21 -08003752 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3753 * @pages: array of pages to check
3754 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003755 *
Hugh Dickins24513262012-01-20 14:34:21 -08003756 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003757 *
3758 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003759 */
Hugh Dickins24513262012-01-20 14:34:21 -08003760void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003761{
Johannes Weiner925b7672012-01-12 17:18:15 -08003762 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003763 struct zone *zone = NULL;
3764 int pgscanned = 0;
3765 int pgrescued = 0;
3766 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003767
Hugh Dickins24513262012-01-20 14:34:21 -08003768 for (i = 0; i < nr_pages; i++) {
3769 struct page *page = pages[i];
3770 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003771
Hugh Dickins24513262012-01-20 14:34:21 -08003772 pgscanned++;
3773 pagezone = page_zone(page);
3774 if (pagezone != zone) {
3775 if (zone)
3776 spin_unlock_irq(&zone->lru_lock);
3777 zone = pagezone;
3778 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003779 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003780 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003781
Hugh Dickins24513262012-01-20 14:34:21 -08003782 if (!PageLRU(page) || !PageUnevictable(page))
3783 continue;
3784
Hugh Dickins39b5f292012-10-08 16:33:18 -07003785 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003786 enum lru_list lru = page_lru_base_type(page);
3787
Sasha Levin309381fea2014-01-23 15:52:54 -08003788 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003789 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003790 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3791 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003792 pgrescued++;
3793 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003794 }
Hugh Dickins24513262012-01-20 14:34:21 -08003795
3796 if (zone) {
3797 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3798 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3799 spin_unlock_irq(&zone->lru_lock);
3800 }
Hugh Dickins850465792012-01-20 14:34:19 -08003801}
3802#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003803
Johannes Weiner264e56d2011-10-31 17:09:13 -07003804static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003805{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003806 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003807 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003808 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003809 "one, please send an email to linux-mm@kvack.org.\n",
3810 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003811}
3812
3813/*
3814 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3815 * all nodes' unevictable lists for evictable pages
3816 */
3817unsigned long scan_unevictable_pages;
3818
3819int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003820 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003821 size_t *length, loff_t *ppos)
3822{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003823 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003824 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003825 scan_unevictable_pages = 0;
3826 return 0;
3827}
3828
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003829#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003830/*
3831 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3832 * a specified node's per zone unevictable lists for evictable pages.
3833 */
3834
Kay Sievers10fbcf42011-12-21 14:48:43 -08003835static ssize_t read_scan_unevictable_node(struct device *dev,
3836 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003837 char *buf)
3838{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003839 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003840 return sprintf(buf, "0\n"); /* always zero; should fit... */
3841}
3842
Kay Sievers10fbcf42011-12-21 14:48:43 -08003843static ssize_t write_scan_unevictable_node(struct device *dev,
3844 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003845 const char *buf, size_t count)
3846{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003847 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003848 return 1;
3849}
3850
3851
Kay Sievers10fbcf42011-12-21 14:48:43 -08003852static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003853 read_scan_unevictable_node,
3854 write_scan_unevictable_node);
3855
3856int scan_unevictable_register_node(struct node *node)
3857{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003858 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003859}
3860
3861void scan_unevictable_unregister_node(struct node *node)
3862{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003863 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003864}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003865#endif