blob: fbcf46076c4fb8cf38bfef18f34b4c498c4185a1 [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
327 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
328 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
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001124 free_hot_cold_page_list(&free_pages, 1);
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/*
Andrew Morton1742f192006-03-22 00:08:21 -08001442 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1443 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 */
Mel Gorman66635622010-08-09 17:19:30 -07001445static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001446shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001447 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448{
1449 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001450 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001451 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001452 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001453 unsigned long nr_dirty = 0;
1454 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001455 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001456 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001457 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001458 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001459 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001460 struct zone *zone = lruvec_zone(lruvec);
1461 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001462
Rik van Riel35cd7812009-09-21 17:01:38 -07001463 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001464 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001465
1466 /* We are about to die and free our memory. Return now. */
1467 if (fatal_signal_pending(current))
1468 return SWAP_CLUSTER_MAX;
1469 }
1470
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001472
1473 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001474 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001475 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001476 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001477
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001480 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1481 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001482
1483 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1484 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1485
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001486 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001487 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001488 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001489 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001490 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001491 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001492 }
Hillf Dantond563c052012-03-21 16:34:02 -07001493 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001494
Hillf Dantond563c052012-03-21 16:34:02 -07001495 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001496 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001497
Minchan Kim02c6de82012-10-08 16:31:55 -07001498 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001499 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1500 &nr_writeback, &nr_immediate,
1501 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001502
Hugh Dickins3f797682012-01-12 17:20:07 -08001503 spin_lock_irq(&zone->lru_lock);
1504
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001505 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001506
Ying Han904249a2012-04-25 16:01:48 -07001507 if (global_reclaim(sc)) {
1508 if (current_is_kswapd())
1509 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1510 nr_reclaimed);
1511 else
1512 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1513 nr_reclaimed);
1514 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001515
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001516 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001517
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001518 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001519
1520 spin_unlock_irq(&zone->lru_lock);
1521
1522 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001523
Mel Gorman92df3a72011-10-31 17:07:56 -07001524 /*
1525 * If reclaim is isolating dirty pages under writeback, it implies
1526 * that the long-lived page allocation rate is exceeding the page
1527 * laundering rate. Either the global limits are not being effective
1528 * at throttling processes due to the page distribution throughout
1529 * zones or there is heavy usage of a slow backing device. The
1530 * only option is to throttle from reclaim context which is not ideal
1531 * as there is no guarantee the dirtying process is throttled in the
1532 * same way balance_dirty_pages() manages.
1533 *
Mel Gorman8e950282013-07-03 15:02:02 -07001534 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1535 * of pages under pages flagged for immediate reclaim and stall if any
1536 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001537 */
Mel Gorman918fc712013-07-08 16:00:25 -07001538 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001539 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001540
Mel Gormand43006d2013-07-03 15:01:50 -07001541 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001542 * memcg will stall in page writeback so only consider forcibly
1543 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001544 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001545 if (global_reclaim(sc)) {
1546 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001547 * Tag a zone as congested if all the dirty pages scanned were
1548 * backed by a congested BDI and wait_iff_congested will stall.
1549 */
1550 if (nr_dirty && nr_dirty == nr_congested)
1551 zone_set_flag(zone, ZONE_CONGESTED);
1552
1553 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001554 * If dirty pages are scanned that are not queued for IO, it
1555 * implies that flushers are not keeping up. In this case, flag
1556 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1557 * pages from reclaim context. It will forcibly stall in the
1558 * next check.
1559 */
1560 if (nr_unqueued_dirty == nr_taken)
1561 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1562
1563 /*
1564 * In addition, if kswapd scans pages marked marked for
1565 * immediate reclaim and under writeback (nr_immediate), it
1566 * implies that pages are cycling through the LRU faster than
1567 * they are written so also forcibly stall.
1568 */
1569 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1570 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001571 }
Mel Gormand43006d2013-07-03 15:01:50 -07001572
Mel Gorman8e950282013-07-03 15:02:02 -07001573 /*
1574 * Stall direct reclaim for IO completions if underlying BDIs or zone
1575 * is congested. Allow kswapd to continue until it starts encountering
1576 * unqueued dirty pages or cycling through the LRU too quickly.
1577 */
1578 if (!sc->hibernation_mode && !current_is_kswapd())
1579 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1580
Mel Gormane11da5b2010-10-26 14:21:40 -07001581 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1582 zone_idx(zone),
1583 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001584 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001585 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001586 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587}
1588
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001589/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 * This moves pages from the active list to the inactive list.
1591 *
1592 * We move them the other way if the page is referenced by one or more
1593 * processes, from rmap.
1594 *
1595 * If the pages are mostly unmapped, the processing is fast and it is
1596 * appropriate to hold zone->lru_lock across the whole operation. But if
1597 * the pages are mapped, the processing is slow (page_referenced()) so we
1598 * should drop zone->lru_lock around each page. It's impossible to balance
1599 * this, so instead we remove the pages from the LRU while processing them.
1600 * It is safe to rely on PG_active against the non-LRU pages in here because
1601 * nobody will play with that bit on a non-LRU page.
1602 *
1603 * The downside is that we have to touch page->_count against each page.
1604 * But we had to alter page->flags anyway.
1605 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001606
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001607static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001608 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001609 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001610 enum lru_list lru)
1611{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001612 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001613 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001614 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001615 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001616
Wu Fengguang3eb41402009-06-16 15:33:13 -07001617 while (!list_empty(list)) {
1618 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001619 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001620
Sasha Levin309381fea2014-01-23 15:52:54 -08001621 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001622 SetPageLRU(page);
1623
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001624 nr_pages = hpage_nr_pages(page);
1625 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001626 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001627 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001628
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001629 if (put_page_testzero(page)) {
1630 __ClearPageLRU(page);
1631 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001632 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001633
1634 if (unlikely(PageCompound(page))) {
1635 spin_unlock_irq(&zone->lru_lock);
1636 (*get_compound_page_dtor(page))(page);
1637 spin_lock_irq(&zone->lru_lock);
1638 } else
1639 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001640 }
1641 }
1642 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1643 if (!is_active_lru(lru))
1644 __count_vm_events(PGDEACTIVATE, pgmoved);
1645}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001646
Hugh Dickinsf6260122012-01-12 17:20:06 -08001647static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001648 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001649 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001650 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001652 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001653 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001654 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001656 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001657 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001659 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001660 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001661 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001662 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001663 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664
1665 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001666
1667 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001668 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001669 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001670 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001671
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001673
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001674 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1675 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001676 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001677 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001678
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001679 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001680
Hugh Dickinsf6260122012-01-12 17:20:06 -08001681 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001682 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001683 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 spin_unlock_irq(&zone->lru_lock);
1685
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 while (!list_empty(&l_hold)) {
1687 cond_resched();
1688 page = lru_to_page(&l_hold);
1689 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001690
Hugh Dickins39b5f292012-10-08 16:33:18 -07001691 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001692 putback_lru_page(page);
1693 continue;
1694 }
1695
Mel Gormancc715d92012-03-21 16:34:00 -07001696 if (unlikely(buffer_heads_over_limit)) {
1697 if (page_has_private(page) && trylock_page(page)) {
1698 if (page_has_private(page))
1699 try_to_release_page(page, 0);
1700 unlock_page(page);
1701 }
1702 }
1703
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001704 if (page_referenced(page, 0, sc->target_mem_cgroup,
1705 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001706 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001707 /*
1708 * Identify referenced, file-backed active pages and
1709 * give them one more trip around the active list. So
1710 * that executable code get better chances to stay in
1711 * memory under moderate memory pressure. Anon pages
1712 * are not likely to be evicted by use-once streaming
1713 * IO, plus JVM can create lots of anon VM_EXEC pages,
1714 * so we ignore them here.
1715 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001716 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001717 list_add(&page->lru, &l_active);
1718 continue;
1719 }
1720 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001721
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001722 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 list_add(&page->lru, &l_inactive);
1724 }
1725
Andrew Mortonb5557492009-01-06 14:40:13 -08001726 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001727 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001728 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001729 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001730 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001731 * Count referenced pages from currently used mappings as rotated,
1732 * even though only some of them are actually re-activated. This
1733 * helps balance scan pressure between file and anonymous pages in
1734 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001735 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001736 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001737
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001738 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1739 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001740 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001741 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001742
1743 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744}
1745
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001746#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001747static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001748{
1749 unsigned long active, inactive;
1750
1751 active = zone_page_state(zone, NR_ACTIVE_ANON);
1752 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1753
1754 if (inactive * zone->inactive_ratio < active)
1755 return 1;
1756
1757 return 0;
1758}
1759
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001760/**
1761 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001762 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001763 *
1764 * Returns true if the zone does not have enough inactive anon pages,
1765 * meaning some active anon pages need to be deactivated.
1766 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001767static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001768{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001769 /*
1770 * If we don't have swap space, anonymous page deactivation
1771 * is pointless.
1772 */
1773 if (!total_swap_pages)
1774 return 0;
1775
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001776 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001777 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001778
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001779 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001780}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001781#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001782static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001783{
1784 return 0;
1785}
1786#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001787
Rik van Riel56e49d22009-06-16 15:32:28 -07001788/**
1789 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001790 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001791 *
1792 * When the system is doing streaming IO, memory pressure here
1793 * ensures that active file pages get deactivated, until more
1794 * than half of the file pages are on the inactive list.
1795 *
1796 * Once we get to that situation, protect the system's working
1797 * set from being evicted by disabling active file page aging.
1798 *
1799 * This uses a different ratio than the anonymous pages, because
1800 * the page cache uses a use-once replacement algorithm.
1801 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001802static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001803{
Johannes Weinere3790142013-02-22 16:35:19 -08001804 unsigned long inactive;
1805 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001806
Johannes Weinere3790142013-02-22 16:35:19 -08001807 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1808 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1809
1810 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001811}
1812
Hugh Dickins75b00af2012-05-29 15:07:09 -07001813static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001814{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001815 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001816 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001817 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001818 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001819}
1820
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001821static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001822 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001823{
Rik van Rielb39415b2009-12-14 17:59:48 -08001824 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001825 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001826 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001827 return 0;
1828 }
1829
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001830 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001831}
1832
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001833static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001834{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001835 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001836 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001837 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001838}
1839
Johannes Weiner9a265112013-02-22 16:32:17 -08001840enum scan_balance {
1841 SCAN_EQUAL,
1842 SCAN_FRACT,
1843 SCAN_ANON,
1844 SCAN_FILE,
1845};
1846
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001848 * Determine how aggressively the anon and file LRU lists should be
1849 * scanned. The relative value of each set of LRU lists is determined
1850 * by looking at the fraction of the pages scanned we did rotate back
1851 * onto the active list instead of evict.
1852 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001853 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1854 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001855 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001856static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001857 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001858{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001859 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001860 u64 fraction[2];
1861 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001862 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001863 unsigned long anon_prio, file_prio;
1864 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001865 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001866 bool force_scan = false;
1867 unsigned long ap, fp;
1868 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001869 bool some_scanned;
1870 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001871
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001872 /*
1873 * If the zone or memcg is small, nr[l] can be 0. This
1874 * results in no scanning on this priority and a potential
1875 * priority drop. Global direct reclaim can go to the next
1876 * zone and tends to have no problems. Global kswapd is for
1877 * zone balancing and it needs to scan a minimum amount. When
1878 * reclaiming for a memcg, a priority drop can cause high
1879 * latencies, so it's better to scan a minimum amount there as
1880 * well.
1881 */
Lisa Du6e543d52013-09-11 14:22:36 -07001882 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001883 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001884 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001885 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001886
1887 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001888 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001889 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001890 goto out;
1891 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001892
Johannes Weiner10316b32013-02-22 16:32:14 -08001893 /*
1894 * Global reclaim will swap to prevent OOM even with no
1895 * swappiness, but memcg users want to use this knob to
1896 * disable swapping for individual groups completely when
1897 * using the memory controller's swap limit feature would be
1898 * too expensive.
1899 */
1900 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001901 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001902 goto out;
1903 }
1904
1905 /*
1906 * Do not apply any pressure balancing cleverness when the
1907 * system is close to OOM, scan both anon and file equally
1908 * (unless the swappiness setting disagrees with swapping).
1909 */
1910 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001911 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001912 goto out;
1913 }
1914
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001915 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1916 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1917 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1918 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001919
Johannes Weiner11d16c22013-02-22 16:32:15 -08001920 /*
Johannes Weiner62376252014-05-06 12:50:07 -07001921 * Prevent the reclaimer from falling into the cache trap: as
1922 * cache pages start out inactive, every cache fault will tip
1923 * the scan balance towards the file LRU. And as the file LRU
1924 * shrinks, so does the window for rotation from references.
1925 * This means we have a runaway feedback loop where a tiny
1926 * thrashing file LRU becomes infinitely more attractive than
1927 * anon pages. Try to detect this based on file LRU size.
1928 */
1929 if (global_reclaim(sc)) {
1930 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1931
1932 if (unlikely(file + free <= high_wmark_pages(zone))) {
1933 scan_balance = SCAN_ANON;
1934 goto out;
1935 }
1936 }
1937
1938 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001939 * There is enough inactive page cache, do not reclaim
1940 * anything from the anonymous working set right now.
1941 */
1942 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001943 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001944 goto out;
1945 }
1946
Johannes Weiner9a265112013-02-22 16:32:17 -08001947 scan_balance = SCAN_FRACT;
1948
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001949 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001950 * With swappiness at 100, anonymous and file have the same priority.
1951 * This scanning priority is essentially the inverse of IO cost.
1952 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001953 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001954 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001955
1956 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001957 * OK, so we have swap space and a fair amount of page cache
1958 * pages. We use the recently rotated / recently scanned
1959 * ratios to determine how valuable each cache is.
1960 *
1961 * Because workloads change over time (and to avoid overflow)
1962 * we keep these statistics as a floating average, which ends
1963 * up weighing recent references more than old ones.
1964 *
1965 * anon in [0], file in [1]
1966 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001967 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001968 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001969 reclaim_stat->recent_scanned[0] /= 2;
1970 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001971 }
1972
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001973 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001974 reclaim_stat->recent_scanned[1] /= 2;
1975 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001976 }
1977
1978 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001979 * The amount of pressure on anon vs file pages is inversely
1980 * proportional to the fraction of recently scanned pages on
1981 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001982 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001983 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001984 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001985
Satoru Moriyafe350042012-05-29 15:06:47 -07001986 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001987 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001988 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001989
Shaohua Li76a33fc2010-05-24 14:32:36 -07001990 fraction[0] = ap;
1991 fraction[1] = fp;
1992 denominator = ap + fp + 1;
1993out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001994 some_scanned = false;
1995 /* Only use force_scan on second pass. */
1996 for (pass = 0; !some_scanned && pass < 2; pass++) {
1997 for_each_evictable_lru(lru) {
1998 int file = is_file_lru(lru);
1999 unsigned long size;
2000 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002001
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002002 size = get_lru_size(lruvec, lru);
2003 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002004
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002005 if (!scan && pass && force_scan)
2006 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002007
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002008 switch (scan_balance) {
2009 case SCAN_EQUAL:
2010 /* Scan lists relative to size */
2011 break;
2012 case SCAN_FRACT:
2013 /*
2014 * Scan types proportional to swappiness and
2015 * their relative recent reclaim efficiency.
2016 */
2017 scan = div64_u64(scan * fraction[file],
2018 denominator);
2019 break;
2020 case SCAN_FILE:
2021 case SCAN_ANON:
2022 /* Scan one type exclusively */
2023 if ((scan_balance == SCAN_FILE) != file)
2024 scan = 0;
2025 break;
2026 default:
2027 /* Look ma, no brain */
2028 BUG();
2029 }
2030 nr[lru] = scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002031 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002032 * Skip the second pass and don't force_scan,
2033 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002034 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002035 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002036 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002037 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002038}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002039
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002040/*
2041 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2042 */
2043static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2044{
2045 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002046 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002047 unsigned long nr_to_scan;
2048 enum lru_list lru;
2049 unsigned long nr_reclaimed = 0;
2050 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2051 struct blk_plug plug;
Mel Gormane82e0562013-07-03 15:01:44 -07002052 bool scan_adjusted = false;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002053
2054 get_scan_count(lruvec, sc, nr);
2055
Mel Gormane82e0562013-07-03 15:01:44 -07002056 /* Record the original scan target for proportional adjustments later */
2057 memcpy(targets, nr, sizeof(nr));
2058
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002059 blk_start_plug(&plug);
2060 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2061 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002062 unsigned long nr_anon, nr_file, percentage;
2063 unsigned long nr_scanned;
2064
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002065 for_each_evictable_lru(lru) {
2066 if (nr[lru]) {
2067 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2068 nr[lru] -= nr_to_scan;
2069
2070 nr_reclaimed += shrink_list(lru, nr_to_scan,
2071 lruvec, sc);
2072 }
2073 }
Mel Gormane82e0562013-07-03 15:01:44 -07002074
2075 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2076 continue;
2077
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002078 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002079 * For global direct reclaim, reclaim only the number of pages
2080 * requested. Less care is taken to scan proportionally as it
2081 * is more important to minimise direct reclaim stall latency
2082 * than it is to properly age the LRU lists.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002083 */
Mel Gormane82e0562013-07-03 15:01:44 -07002084 if (global_reclaim(sc) && !current_is_kswapd())
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002085 break;
Mel Gormane82e0562013-07-03 15:01:44 -07002086
2087 /*
2088 * For kswapd and memcg, reclaim at least the number of pages
2089 * requested. Ensure that the anon and file LRUs shrink
2090 * proportionally what was requested by get_scan_count(). We
2091 * stop reclaiming one LRU and reduce the amount scanning
2092 * proportional to the original scan target.
2093 */
2094 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2095 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2096
2097 if (nr_file > nr_anon) {
2098 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2099 targets[LRU_ACTIVE_ANON] + 1;
2100 lru = LRU_BASE;
2101 percentage = nr_anon * 100 / scan_target;
2102 } else {
2103 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2104 targets[LRU_ACTIVE_FILE] + 1;
2105 lru = LRU_FILE;
2106 percentage = nr_file * 100 / scan_target;
2107 }
2108
2109 /* Stop scanning the smaller of the LRU */
2110 nr[lru] = 0;
2111 nr[lru + LRU_ACTIVE] = 0;
2112
2113 /*
2114 * Recalculate the other LRU scan count based on its original
2115 * scan target and the percentage scanning already complete
2116 */
2117 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2118 nr_scanned = targets[lru] - nr[lru];
2119 nr[lru] = targets[lru] * (100 - percentage) / 100;
2120 nr[lru] -= min(nr[lru], nr_scanned);
2121
2122 lru += LRU_ACTIVE;
2123 nr_scanned = targets[lru] - nr[lru];
2124 nr[lru] = targets[lru] * (100 - percentage) / 100;
2125 nr[lru] -= min(nr[lru], nr_scanned);
2126
2127 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002128 }
2129 blk_finish_plug(&plug);
2130 sc->nr_reclaimed += nr_reclaimed;
2131
2132 /*
2133 * Even if we did not try to evict anon pages at all, we want to
2134 * rebalance the anon lru active/inactive ratio.
2135 */
2136 if (inactive_anon_is_low(lruvec))
2137 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2138 sc, LRU_ACTIVE_ANON);
2139
2140 throttle_vm_writeout(sc->gfp_mask);
2141}
2142
Mel Gorman23b9da52012-05-29 15:06:20 -07002143/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002144static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002145{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002146 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002147 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002148 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002149 return true;
2150
2151 return false;
2152}
2153
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002154/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002155 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2156 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2157 * true if more pages should be reclaimed such that when the page allocator
2158 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2159 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002160 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002161static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002162 unsigned long nr_reclaimed,
2163 unsigned long nr_scanned,
2164 struct scan_control *sc)
2165{
2166 unsigned long pages_for_compaction;
2167 unsigned long inactive_lru_pages;
2168
2169 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002170 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002171 return false;
2172
Mel Gorman28765922011-02-25 14:44:20 -08002173 /* Consider stopping depending on scan and reclaim activity */
2174 if (sc->gfp_mask & __GFP_REPEAT) {
2175 /*
2176 * For __GFP_REPEAT allocations, stop reclaiming if the
2177 * full LRU list has been scanned and we are still failing
2178 * to reclaim pages. This full LRU scan is potentially
2179 * expensive but a __GFP_REPEAT caller really wants to succeed
2180 */
2181 if (!nr_reclaimed && !nr_scanned)
2182 return false;
2183 } else {
2184 /*
2185 * For non-__GFP_REPEAT allocations which can presumably
2186 * fail without consequence, stop if we failed to reclaim
2187 * any pages from the last SWAP_CLUSTER_MAX number of
2188 * pages that were scanned. This will return to the
2189 * caller faster at the risk reclaim/compaction and
2190 * the resulting allocation attempt fails
2191 */
2192 if (!nr_reclaimed)
2193 return false;
2194 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002195
2196 /*
2197 * If we have not reclaimed enough pages for compaction and the
2198 * inactive lists are large enough, continue reclaiming
2199 */
2200 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002201 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002202 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002203 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002204 if (sc->nr_reclaimed < pages_for_compaction &&
2205 inactive_lru_pages > pages_for_compaction)
2206 return true;
2207
2208 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002209 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002210 case COMPACT_PARTIAL:
2211 case COMPACT_CONTINUE:
2212 return false;
2213 default:
2214 return true;
2215 }
2216}
2217
Andrew Morton0608f432013-09-24 15:27:41 -07002218static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002219{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002220 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002221
Johannes Weiner56600482012-01-12 17:17:59 -08002222 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002223 struct mem_cgroup *root = sc->target_mem_cgroup;
2224 struct mem_cgroup_reclaim_cookie reclaim = {
2225 .zone = zone,
2226 .priority = sc->priority,
2227 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002228 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002229
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002230 nr_reclaimed = sc->nr_reclaimed;
2231 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002232
Andrew Morton694fbc02013-09-24 15:27:37 -07002233 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2234 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002235 struct lruvec *lruvec;
2236
2237 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2238
2239 shrink_lruvec(lruvec, sc);
2240
2241 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002242 * Direct reclaim and kswapd have to scan all memory
2243 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002244 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002245 *
2246 * Limit reclaim, on the other hand, only cares about
2247 * nr_to_reclaim pages to be reclaimed and it will
2248 * retry with decreasing priority if one round over the
2249 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002250 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002251 if (!global_reclaim(sc) &&
2252 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002253 mem_cgroup_iter_break(root, memcg);
2254 break;
2255 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002256 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2257 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002258
2259 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2260 sc->nr_scanned - nr_scanned,
2261 sc->nr_reclaimed - nr_reclaimed);
2262
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002263 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2264 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002265}
2266
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002267/* Returns true if compaction should go ahead for a high-order request */
2268static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2269{
2270 unsigned long balance_gap, watermark;
2271 bool watermark_ok;
2272
2273 /* Do not consider compaction for orders reclaim is meant to satisfy */
2274 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2275 return false;
2276
2277 /*
2278 * Compaction takes time to run and there are potentially other
2279 * callers using the pages just freed. Continue reclaiming until
2280 * there is a buffer of free pages available to give compaction
2281 * a reasonable chance of completing and allocating the page
2282 */
2283 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002284 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002285 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2286 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2287 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2288
2289 /*
2290 * If compaction is deferred, reclaim up to a point where
2291 * compaction will have a chance of success when re-enabled
2292 */
Rik van Rielaff62242012-03-21 16:33:52 -07002293 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002294 return watermark_ok;
2295
2296 /* If compaction is not ready to start, keep reclaiming */
2297 if (!compaction_suitable(zone, sc->order))
2298 return false;
2299
2300 return watermark_ok;
2301}
2302
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303/*
2304 * This is the direct reclaim path, for page-allocating processes. We only
2305 * try to reclaim pages from zones which will satisfy the caller's allocation
2306 * request.
2307 *
Mel Gorman41858962009-06-16 15:32:12 -07002308 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2309 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2311 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002312 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2313 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2314 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 * If a zone is deemed to be full of pinned pages then just give it a light
2317 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002318 *
2319 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002320 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002321 * the caller that it should consider retrying the allocation instead of
2322 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 */
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002324static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325{
Mel Gormandd1a2392008-04-28 02:12:17 -07002326 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002327 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002328 unsigned long nr_soft_reclaimed;
2329 unsigned long nr_soft_scanned;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002330 unsigned long lru_pages = 0;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002331 bool aborted_reclaim = false;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002332 struct reclaim_state *reclaim_state = current->reclaim_state;
Weijie Yang619d0d762014-04-07 15:36:59 -07002333 gfp_t orig_mask;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002334 struct shrink_control shrink = {
2335 .gfp_mask = sc->gfp_mask,
2336 };
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002337 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002338
Mel Gormancc715d92012-03-21 16:34:00 -07002339 /*
2340 * If the number of buffer_heads in the machine exceeds the maximum
2341 * allowed level, force direct reclaim to scan the highmem zone as
2342 * highmem pages could be pinning lowmem pages storing buffer_heads
2343 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002344 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002345 if (buffer_heads_over_limit)
2346 sc->gfp_mask |= __GFP_HIGHMEM;
2347
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002348 nodes_clear(shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002349
Mel Gormand4debc62010-08-09 17:19:29 -07002350 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2351 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002352 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002354 /*
2355 * Take care memory controller reclaiming has small influence
2356 * to global LRU.
2357 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002358 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002359 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2360 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002361
2362 lru_pages += zone_reclaimable_pages(zone);
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002363 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002364
Lisa Du6e543d52013-09-11 14:22:36 -07002365 if (sc->priority != DEF_PRIORITY &&
2366 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002367 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002368 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002369 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002370 * If we already have plenty of memory free for
2371 * compaction in this zone, don't free any more.
2372 * Even though compaction is invoked for any
2373 * non-zero order, only frequent costly order
2374 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002375 * noticeable problem, like transparent huge
2376 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002377 */
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002378 if ((zonelist_zone_idx(z) <= requested_highidx)
2379 && compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002380 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002381 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002382 }
Rik van Riele0887c12011-10-31 17:09:31 -07002383 }
Andrew Morton0608f432013-09-24 15:27:41 -07002384 /*
2385 * This steals pages from memory cgroups over softlimit
2386 * and returns the number of reclaimed pages and
2387 * scanned pages. This works for global memory pressure
2388 * and balancing, not for a memcg's limit.
2389 */
2390 nr_soft_scanned = 0;
2391 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2392 sc->order, sc->gfp_mask,
2393 &nr_soft_scanned);
2394 sc->nr_reclaimed += nr_soft_reclaimed;
2395 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002396 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002397 }
Nick Piggin408d8542006-09-25 23:31:27 -07002398
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002399 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 }
Mel Gormane0c23272011-10-31 17:09:33 -07002401
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002402 /*
2403 * Don't shrink slabs when reclaiming memory from over limit cgroups
2404 * but do shrink slab at least once when aborting reclaim for
2405 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2406 * pages.
2407 */
2408 if (global_reclaim(sc)) {
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002409 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002410 if (reclaim_state) {
2411 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2412 reclaim_state->reclaimed_slab = 0;
2413 }
2414 }
2415
Weijie Yang619d0d762014-04-07 15:36:59 -07002416 /*
2417 * Restore to original mask to avoid the impact on the caller if we
2418 * promoted it to __GFP_HIGHMEM.
2419 */
2420 sc->gfp_mask = orig_mask;
2421
Mel Gorman0cee34f2012-01-12 17:19:49 -08002422 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002423}
2424
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002425/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002426static bool all_unreclaimable(struct zonelist *zonelist,
2427 struct scan_control *sc)
2428{
2429 struct zoneref *z;
2430 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002431
2432 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2433 gfp_zone(sc->gfp_mask), sc->nodemask) {
2434 if (!populated_zone(zone))
2435 continue;
2436 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2437 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002438 if (zone_reclaimable(zone))
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002439 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002440 }
2441
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002442 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445/*
2446 * This is the main entry point to direct page reclaim.
2447 *
2448 * If a full scan of the inactive list fails to free enough memory then we
2449 * are "out of memory" and something needs to be killed.
2450 *
2451 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2452 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002453 * caller can't do much about. We kick the writeback threads and take explicit
2454 * naps in the hope that some of these pages can be written. But if the
2455 * allocating task holds filesystem locks which prevent writeout this might not
2456 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002457 *
2458 * returns: 0, if no pages reclaimed
2459 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 */
Mel Gormandac1d272008-04-28 02:12:12 -07002461static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002462 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463{
Andrew Morton69e05942006-03-22 00:08:19 -08002464 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002465 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002466 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467
Keika Kobayashi873b4772008-07-25 01:48:52 -07002468 delayacct_freepages_start();
2469
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002470 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002471 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002473 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002474 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2475 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002476 sc->nr_scanned = 0;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002477 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002478
Balbir Singh66e17072008-02-07 00:13:56 -08002479 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002480 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482
2483 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002484 * If we're getting trouble reclaiming, start doing
2485 * writepage even in laptop mode.
2486 */
2487 if (sc->priority < DEF_PRIORITY - 2)
2488 sc->may_writepage = 1;
2489
2490 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 * Try to write back as many pages as we just scanned. This
2492 * tends to cause slow streaming writers to write data to the
2493 * disk smoothly, at the dirtying rate, which is nice. But
2494 * that's undesirable in laptop mode, where we *want* lumpy
2495 * writeout. So in laptop mode, write out the whole world.
2496 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002497 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2498 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002499 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2500 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002501 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 }
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002503 } while (--sc->priority >= 0 && !aborted_reclaim);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002504
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002506 delayacct_freepages_end();
2507
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002508 if (sc->nr_reclaimed)
2509 return sc->nr_reclaimed;
2510
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002511 /*
2512 * As hibernation is going on, kswapd is freezed so that it can't mark
2513 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2514 * check.
2515 */
2516 if (oom_killer_disabled)
2517 return 0;
2518
Mel Gorman0cee34f2012-01-12 17:19:49 -08002519 /* Aborted reclaim to try compaction? don't OOM, then */
2520 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002521 return 1;
2522
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002523 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002524 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002525 return 1;
2526
2527 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528}
2529
Mel Gorman55150612012-07-31 16:44:35 -07002530static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2531{
2532 struct zone *zone;
2533 unsigned long pfmemalloc_reserve = 0;
2534 unsigned long free_pages = 0;
2535 int i;
2536 bool wmark_ok;
2537
2538 for (i = 0; i <= ZONE_NORMAL; i++) {
2539 zone = &pgdat->node_zones[i];
2540 pfmemalloc_reserve += min_wmark_pages(zone);
2541 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2542 }
2543
2544 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2545
2546 /* kswapd must be awake if processes are being throttled */
2547 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2548 pgdat->classzone_idx = min(pgdat->classzone_idx,
2549 (enum zone_type)ZONE_NORMAL);
2550 wake_up_interruptible(&pgdat->kswapd_wait);
2551 }
2552
2553 return wmark_ok;
2554}
2555
2556/*
2557 * Throttle direct reclaimers if backing storage is backed by the network
2558 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2559 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002560 * when the low watermark is reached.
2561 *
2562 * Returns true if a fatal signal was delivered during throttling. If this
2563 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002564 */
Mel Gorman50694c22012-11-26 16:29:48 -08002565static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002566 nodemask_t *nodemask)
2567{
2568 struct zone *zone;
2569 int high_zoneidx = gfp_zone(gfp_mask);
2570 pg_data_t *pgdat;
2571
2572 /*
2573 * Kernel threads should not be throttled as they may be indirectly
2574 * responsible for cleaning pages necessary for reclaim to make forward
2575 * progress. kjournald for example may enter direct reclaim while
2576 * committing a transaction where throttling it could forcing other
2577 * processes to block on log_wait_commit().
2578 */
2579 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002580 goto out;
2581
2582 /*
2583 * If a fatal signal is pending, this process should not throttle.
2584 * It should return quickly so it can exit and free its memory
2585 */
2586 if (fatal_signal_pending(current))
2587 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002588
2589 /* Check if the pfmemalloc reserves are ok */
2590 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2591 pgdat = zone->zone_pgdat;
2592 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002593 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002594
Mel Gorman68243e72012-07-31 16:44:39 -07002595 /* Account for the throttling */
2596 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2597
Mel Gorman55150612012-07-31 16:44:35 -07002598 /*
2599 * If the caller cannot enter the filesystem, it's possible that it
2600 * is due to the caller holding an FS lock or performing a journal
2601 * transaction in the case of a filesystem like ext[3|4]. In this case,
2602 * it is not safe to block on pfmemalloc_wait as kswapd could be
2603 * blocked waiting on the same lock. Instead, throttle for up to a
2604 * second before continuing.
2605 */
2606 if (!(gfp_mask & __GFP_FS)) {
2607 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2608 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002609
2610 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002611 }
2612
2613 /* Throttle until kswapd wakes the process */
2614 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2615 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002616
2617check_pending:
2618 if (fatal_signal_pending(current))
2619 return true;
2620
2621out:
2622 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002623}
2624
Mel Gormandac1d272008-04-28 02:12:12 -07002625unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002626 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002627{
Mel Gorman33906bc2010-08-09 17:19:16 -07002628 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002629 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002630 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002631 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002632 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002633 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002634 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002635 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002636 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002637 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002638 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002639 };
2640
Mel Gorman55150612012-07-31 16:44:35 -07002641 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002642 * Do not enter reclaim if fatal signal was delivered while throttled.
2643 * 1 is returned so that the page allocator does not OOM kill at this
2644 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002645 */
Mel Gorman50694c22012-11-26 16:29:48 -08002646 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002647 return 1;
2648
Mel Gorman33906bc2010-08-09 17:19:16 -07002649 trace_mm_vmscan_direct_reclaim_begin(order,
2650 sc.may_writepage,
2651 gfp_mask);
2652
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002653 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002654
2655 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2656
2657 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002658}
2659
Andrew Mortonc255a452012-07-31 16:43:02 -07002660#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002661
Johannes Weiner72835c82012-01-12 17:18:32 -08002662unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002663 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002664 struct zone *zone,
2665 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002666{
2667 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002668 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002669 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002670 .may_writepage = !laptop_mode,
2671 .may_unmap = 1,
2672 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002673 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002674 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002675 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002676 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002677 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002678
Balbir Singh4e416952009-09-23 15:56:39 -07002679 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2680 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002681
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002682 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002683 sc.may_writepage,
2684 sc.gfp_mask);
2685
Balbir Singh4e416952009-09-23 15:56:39 -07002686 /*
2687 * NOTE: Although we can get the priority field, using it
2688 * here is not a good idea, since it limits the pages we can scan.
2689 * if we don't reclaim here, the shrink_zone from balance_pgdat
2690 * will pick up pages from other mem cgroup's as well. We hack
2691 * the priority and make it zero.
2692 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002693 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002694
2695 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2696
Ying Han0ae5e892011-05-26 16:25:25 -07002697 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002698 return sc.nr_reclaimed;
2699}
2700
Johannes Weiner72835c82012-01-12 17:18:32 -08002701unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002702 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002703 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002704{
Balbir Singh4e416952009-09-23 15:56:39 -07002705 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002706 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002707 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002708 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002709 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002710 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002711 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002712 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002713 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002714 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002715 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002716 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002717 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2718 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2719 };
Balbir Singh66e17072008-02-07 00:13:56 -08002720
Ying Han889976d2011-05-26 16:25:33 -07002721 /*
2722 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2723 * take care of from where we get pages. So the node where we start the
2724 * scan does not need to be the current node.
2725 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002726 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002727
2728 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002729
2730 trace_mm_vmscan_memcg_reclaim_begin(0,
2731 sc.may_writepage,
2732 sc.gfp_mask);
2733
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002734 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002735
2736 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2737
2738 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002739}
2740#endif
2741
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002742static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002743{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002744 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002745
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002746 if (!total_swap_pages)
2747 return;
2748
2749 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2750 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002751 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002752
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002753 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002754 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002755 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002756
2757 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2758 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002759}
2760
Johannes Weiner60cefed2012-11-29 13:54:23 -08002761static bool zone_balanced(struct zone *zone, int order,
2762 unsigned long balance_gap, int classzone_idx)
2763{
2764 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2765 balance_gap, classzone_idx, 0))
2766 return false;
2767
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002768 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2769 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002770 return false;
2771
2772 return true;
2773}
2774
Mel Gorman1741c872011-01-13 15:46:21 -08002775/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002776 * pgdat_balanced() is used when checking if a node is balanced.
2777 *
2778 * For order-0, all zones must be balanced!
2779 *
2780 * For high-order allocations only zones that meet watermarks and are in a
2781 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2782 * total of balanced pages must be at least 25% of the zones allowed by
2783 * classzone_idx for the node to be considered balanced. Forcing all zones to
2784 * be balanced for high orders can cause excessive reclaim when there are
2785 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002786 * The choice of 25% is due to
2787 * o a 16M DMA zone that is balanced will not balance a zone on any
2788 * reasonable sized machine
2789 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002790 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002791 * would need to be at least 256M for it to be balance a whole node.
2792 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2793 * to balance a node on its own. These seemed like reasonable ratios.
2794 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002795static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002796{
Jiang Liub40da042013-02-22 16:33:52 -08002797 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002798 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002799 int i;
2800
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002801 /* Check the watermark levels */
2802 for (i = 0; i <= classzone_idx; i++) {
2803 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002804
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002805 if (!populated_zone(zone))
2806 continue;
2807
Jiang Liub40da042013-02-22 16:33:52 -08002808 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002809
2810 /*
2811 * A special case here:
2812 *
2813 * balance_pgdat() skips over all_unreclaimable after
2814 * DEF_PRIORITY. Effectively, it considers them balanced so
2815 * they must be considered balanced here as well!
2816 */
Lisa Du6e543d52013-09-11 14:22:36 -07002817 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002818 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002819 continue;
2820 }
2821
2822 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002823 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002824 else if (!order)
2825 return false;
2826 }
2827
2828 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002829 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002830 else
2831 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002832}
2833
Mel Gorman55150612012-07-31 16:44:35 -07002834/*
2835 * Prepare kswapd for sleeping. This verifies that there are no processes
2836 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2837 *
2838 * Returns true if kswapd is ready to sleep
2839 */
2840static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002841 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002842{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002843 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2844 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002845 return false;
2846
2847 /*
2848 * There is a potential race between when kswapd checks its watermarks
2849 * and a process gets throttled. There is also a potential race if
2850 * processes get throttled, kswapd wakes, a large process exits therby
2851 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2852 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2853 * so wake them now if necessary. If necessary, processes will wake
2854 * kswapd and get throttled again
2855 */
2856 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2857 wake_up(&pgdat->pfmemalloc_wait);
2858 return false;
2859 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002860
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002861 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002862}
2863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864/*
Mel Gorman75485362013-07-03 15:01:42 -07002865 * kswapd shrinks the zone by the number of pages required to reach
2866 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002867 *
2868 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002869 * reclaim or if the lack of progress was due to pages under writeback.
2870 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002871 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002872static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002873 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002874 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002875 unsigned long lru_pages,
2876 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002877{
Mel Gorman7c954f62013-07-03 15:01:54 -07002878 int testorder = sc->order;
2879 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002880 struct reclaim_state *reclaim_state = current->reclaim_state;
2881 struct shrink_control shrink = {
2882 .gfp_mask = sc->gfp_mask,
2883 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002884 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002885
2886 /* Reclaim above the high watermark. */
2887 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002888
2889 /*
2890 * Kswapd reclaims only single pages with compaction enabled. Trying
2891 * too hard to reclaim until contiguous free pages have become
2892 * available can hurt performance by evicting too much useful data
2893 * from memory. Do not reclaim more than needed for compaction.
2894 */
2895 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2896 compaction_suitable(zone, sc->order) !=
2897 COMPACT_SKIPPED)
2898 testorder = 0;
2899
2900 /*
2901 * We put equal pressure on every zone, unless one zone has way too
2902 * many pages free already. The "too many pages" is defined as the
2903 * high wmark plus a "gap" where the gap is either the low
2904 * watermark or 1% of the zone, whichever is smaller.
2905 */
2906 balance_gap = min(low_wmark_pages(zone),
2907 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2908 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2909
2910 /*
2911 * If there is no low memory pressure or the zone is balanced then no
2912 * reclaim is necessary
2913 */
2914 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2915 if (!lowmem_pressure && zone_balanced(zone, testorder,
2916 balance_gap, classzone_idx))
2917 return true;
2918
Mel Gorman75485362013-07-03 15:01:42 -07002919 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002920 nodes_clear(shrink.nodes_to_scan);
2921 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002922
2923 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002924 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002925 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2926
Mel Gorman2ab44f42013-07-03 15:01:47 -07002927 /* Account for the number of pages attempted to reclaim */
2928 *nr_attempted += sc->nr_to_reclaim;
2929
Mel Gorman283aba92013-07-03 15:01:51 -07002930 zone_clear_flag(zone, ZONE_WRITEBACK);
2931
Mel Gorman7c954f62013-07-03 15:01:54 -07002932 /*
2933 * If a zone reaches its high watermark, consider it to be no longer
2934 * congested. It's possible there are dirty pages backed by congested
2935 * BDIs but as pressure is relieved, speculatively avoid congestion
2936 * waits.
2937 */
Lisa Du6e543d52013-09-11 14:22:36 -07002938 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002939 zone_balanced(zone, testorder, 0, classzone_idx)) {
2940 zone_clear_flag(zone, ZONE_CONGESTED);
2941 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2942 }
2943
Mel Gormanb8e83b92013-07-03 15:01:45 -07002944 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002945}
2946
2947/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002949 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002951 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 *
2953 * There is special handling here for zones which are full of pinned pages.
2954 * This can happen if the pages are all mlocked, or if they are all used by
2955 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2956 * What we do is to detect the case where all pages in the zone have been
2957 * scanned twice and there has been zero successful reclaim. Mark the zone as
2958 * dead and from now on, only perform a short scan. Basically we're polling
2959 * the zone for when the problem goes away.
2960 *
2961 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002962 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2963 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2964 * lower zones regardless of the number of free pages in the lower zones. This
2965 * interoperates with the page allocator fallback scheme to ensure that aging
2966 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 */
Mel Gorman99504742011-01-13 15:46:20 -08002968static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002969 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002972 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07002973 unsigned long nr_soft_reclaimed;
2974 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002975 struct scan_control sc = {
2976 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002977 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002978 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002979 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002980 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002981 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002982 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002983 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07002984 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002986 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002988 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002989 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002990 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07002991
2992 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002994 /*
2995 * Scan in the highmem->dma direction for the highest
2996 * zone which needs scanning
2997 */
2998 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2999 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003001 if (!populated_zone(zone))
3002 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
Lisa Du6e543d52013-09-11 14:22:36 -07003004 if (sc.priority != DEF_PRIORITY &&
3005 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003006 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007
Rik van Riel556adec2008-10-18 20:26:34 -07003008 /*
3009 * Do some background aging of the anon list, to give
3010 * pages a chance to be referenced before reclaiming.
3011 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003012 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003013
Mel Gormancc715d92012-03-21 16:34:00 -07003014 /*
3015 * If the number of buffer_heads in the machine
3016 * exceeds the maximum allowed level and this node
3017 * has a highmem zone, force kswapd to reclaim from
3018 * it to relieve lowmem pressure.
3019 */
3020 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3021 end_zone = i;
3022 break;
3023 }
3024
Johannes Weiner60cefed2012-11-29 13:54:23 -08003025 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003026 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003027 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003028 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003029 /*
3030 * If balanced, clear the dirty and congested
3031 * flags
3032 */
Shaohua Li439423f2011-08-25 15:59:12 -07003033 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003034 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003037
Mel Gormanb8e83b92013-07-03 15:01:45 -07003038 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003039 goto out;
3040
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 for (i = 0; i <= end_zone; i++) {
3042 struct zone *zone = pgdat->node_zones + i;
3043
Mel Gorman2ab44f42013-07-03 15:01:47 -07003044 if (!populated_zone(zone))
3045 continue;
3046
Wu Fengguangadea02a2009-09-21 17:01:42 -07003047 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003048
3049 /*
3050 * If any zone is currently balanced then kswapd will
3051 * not call compaction as it is expected that the
3052 * necessary pages are already available.
3053 */
3054 if (pgdat_needs_compaction &&
3055 zone_watermark_ok(zone, order,
3056 low_wmark_pages(zone),
3057 *classzone_idx, 0))
3058 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059 }
3060
3061 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003062 * If we're getting trouble reclaiming, start doing writepage
3063 * even in laptop mode.
3064 */
3065 if (sc.priority < DEF_PRIORITY - 2)
3066 sc.may_writepage = 1;
3067
3068 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069 * Now scan the zone in the dma->highmem direction, stopping
3070 * at the last zone which needs scanning.
3071 *
3072 * We do this because the page allocator works in the opposite
3073 * direction. This prevents the page allocator from allocating
3074 * pages behind kswapd's direction of progress, which would
3075 * cause too much scanning of the lower zones.
3076 */
3077 for (i = 0; i <= end_zone; i++) {
3078 struct zone *zone = pgdat->node_zones + i;
3079
Con Kolivasf3fe6512006-01-06 00:11:15 -08003080 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 continue;
3082
Lisa Du6e543d52013-09-11 14:22:36 -07003083 if (sc.priority != DEF_PRIORITY &&
3084 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 continue;
3086
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003088
Andrew Morton0608f432013-09-24 15:27:41 -07003089 nr_soft_scanned = 0;
3090 /*
3091 * Call soft limit reclaim before calling shrink_zone.
3092 */
3093 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3094 order, sc.gfp_mask,
3095 &nr_soft_scanned);
3096 sc.nr_reclaimed += nr_soft_reclaimed;
3097
Rik van Riel32a43302007-10-16 01:24:50 -07003098 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003099 * There should be no need to raise the scanning
3100 * priority if enough pages are already being scanned
3101 * that that high watermark would be met at 100%
3102 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003103 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003104 if (kswapd_shrink_zone(zone, end_zone, &sc,
3105 lru_pages, &nr_attempted))
3106 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107 }
Mel Gorman55150612012-07-31 16:44:35 -07003108
3109 /*
3110 * If the low watermark is met there is no need for processes
3111 * to be throttled on pfmemalloc_wait as they should not be
3112 * able to safely make forward progress. Wake them
3113 */
3114 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3115 pfmemalloc_watermark_ok(pgdat))
3116 wake_up(&pgdat->pfmemalloc_wait);
3117
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003119 * Fragmentation may mean that the system cannot be rebalanced
3120 * for high-order allocations in all zones. If twice the
3121 * allocation size has been reclaimed and the zones are still
3122 * not balanced then recheck the watermarks at order-0 to
3123 * prevent kswapd reclaiming excessively. Assume that a
3124 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003126 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003127 order = sc.order = 0;
3128
Mel Gormanb8e83b92013-07-03 15:01:45 -07003129 /* Check if kswapd should be suspending */
3130 if (try_to_freeze() || kthread_should_stop())
3131 break;
3132
3133 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003134 * Compact if necessary and kswapd is reclaiming at least the
3135 * high watermark number of pages as requsted
3136 */
3137 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3138 compact_pgdat(pgdat, order);
3139
3140 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003141 * Raise priority if scanning rate is too low or there was no
3142 * progress in reclaiming pages
3143 */
3144 if (raise_priority || !sc.nr_reclaimed)
3145 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003146 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003147 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148
Mel Gormanb8e83b92013-07-03 15:01:45 -07003149out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003150 /*
Mel Gorman55150612012-07-31 16:44:35 -07003151 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003152 * makes a decision on the order we were last reclaiming at. However,
3153 * if another caller entered the allocator slow path while kswapd
3154 * was awake, order will remain at the higher level
3155 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003156 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003157 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158}
3159
Mel Gormandc83edd2011-01-13 15:46:26 -08003160static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003161{
3162 long remaining = 0;
3163 DEFINE_WAIT(wait);
3164
3165 if (freezing(current) || kthread_should_stop())
3166 return;
3167
3168 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3169
3170 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003171 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003172 remaining = schedule_timeout(HZ/10);
3173 finish_wait(&pgdat->kswapd_wait, &wait);
3174 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3175 }
3176
3177 /*
3178 * After a short sleep, check if it was a premature sleep. If not, then
3179 * go fully to sleep until explicitly woken up.
3180 */
Mel Gorman55150612012-07-31 16:44:35 -07003181 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003182 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3183
3184 /*
3185 * vmstat counters are not perfectly accurate and the estimated
3186 * value for counters such as NR_FREE_PAGES can deviate from the
3187 * true value by nr_online_cpus * threshold. To avoid the zone
3188 * watermarks being breached while under pressure, we reduce the
3189 * per-cpu vmstat threshold while kswapd is awake and restore
3190 * them before going back to sleep.
3191 */
3192 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003193
Mel Gorman62997022012-10-08 16:32:47 -07003194 /*
3195 * Compaction records what page blocks it recently failed to
3196 * isolate pages from and skips them in the future scanning.
3197 * When kswapd is going to sleep, it is reasonable to assume
3198 * that pages and compaction may succeed so reset the cache.
3199 */
3200 reset_isolation_suitable(pgdat);
3201
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003202 if (!kthread_should_stop())
3203 schedule();
3204
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003205 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3206 } else {
3207 if (remaining)
3208 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3209 else
3210 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3211 }
3212 finish_wait(&pgdat->kswapd_wait, &wait);
3213}
3214
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215/*
3216 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003217 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 *
3219 * This basically trickles out pages so that we have _some_
3220 * free memory available even if there is no other activity
3221 * that frees anything up. This is needed for things like routing
3222 * etc, where we otherwise might have all activity going on in
3223 * asynchronous contexts that cannot page things out.
3224 *
3225 * If there are applications that are active memory-allocators
3226 * (most normal use), this basically shouldn't matter.
3227 */
3228static int kswapd(void *p)
3229{
Mel Gorman215ddd62011-07-08 15:39:40 -07003230 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003231 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003232 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003233 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 pg_data_t *pgdat = (pg_data_t*)p;
3235 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003236
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 struct reclaim_state reclaim_state = {
3238 .reclaimed_slab = 0,
3239 };
Rusty Russella70f7302009-03-13 14:49:46 +10303240 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
Nick Piggincf40bd12009-01-21 08:12:39 +01003242 lockdep_set_current_reclaim_state(GFP_KERNEL);
3243
Rusty Russell174596a2009-01-01 10:12:29 +10303244 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003245 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 current->reclaim_state = &reclaim_state;
3247
3248 /*
3249 * Tell the memory management that we're a "memory allocator",
3250 * and that if we need more memory we should get access to it
3251 * regardless (see "__alloc_pages()"). "kswapd" should
3252 * never get caught in the normal page freeing logic.
3253 *
3254 * (Kswapd normally doesn't need memory anyway, but sometimes
3255 * you need a small amount of memory in order to be able to
3256 * page out something else, and this flag essentially protects
3257 * us from recursively trying to free more memory as we're
3258 * trying to free the first piece of memory in the first place).
3259 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003260 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003261 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262
Mel Gorman215ddd62011-07-08 15:39:40 -07003263 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003264 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003265 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003266 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003268 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003269
Mel Gorman215ddd62011-07-08 15:39:40 -07003270 /*
3271 * If the last balance_pgdat was unsuccessful it's unlikely a
3272 * new request of a similar or harder type will succeed soon
3273 * so consider going to sleep on the basis we reclaimed at
3274 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003275 if (balanced_classzone_idx >= new_classzone_idx &&
3276 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003277 new_order = pgdat->kswapd_max_order;
3278 new_classzone_idx = pgdat->classzone_idx;
3279 pgdat->kswapd_max_order = 0;
3280 pgdat->classzone_idx = pgdat->nr_zones - 1;
3281 }
3282
Mel Gorman99504742011-01-13 15:46:20 -08003283 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 /*
3285 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003286 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287 */
3288 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003289 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003291 kswapd_try_to_sleep(pgdat, balanced_order,
3292 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003294 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003295 new_order = order;
3296 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003297 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003298 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300
David Rientjes8fe23e02009-12-14 17:58:33 -08003301 ret = try_to_freeze();
3302 if (kthread_should_stop())
3303 break;
3304
3305 /*
3306 * We can speed up thawing tasks if we don't call balance_pgdat
3307 * after returning from the refrigerator
3308 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003309 if (!ret) {
3310 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003311 balanced_classzone_idx = classzone_idx;
3312 balanced_order = balance_pgdat(pgdat, order,
3313 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003314 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003316
3317 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 return 0;
3319}
3320
3321/*
3322 * A zone is low on free memory, so wake its kswapd task to service it.
3323 */
Mel Gorman99504742011-01-13 15:46:20 -08003324void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325{
3326 pg_data_t *pgdat;
3327
Con Kolivasf3fe6512006-01-06 00:11:15 -08003328 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 return;
3330
Paul Jackson02a0e532006-12-13 00:34:25 -08003331 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003333 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003334 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003335 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003336 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3337 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003338 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003340 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003341 return;
3342
3343 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003344 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345}
3346
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003347#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003349 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003350 * freed pages.
3351 *
3352 * Rather than trying to age LRUs the aim is to preserve the overall
3353 * LRU order by reclaiming preferentially
3354 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003356unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003358 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003359 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003360 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3361 .may_swap = 1,
3362 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003363 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003364 .nr_to_reclaim = nr_to_reclaim,
3365 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003366 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003367 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003369 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003370 struct task_struct *p = current;
3371 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003373 p->flags |= PF_MEMALLOC;
3374 lockdep_set_current_reclaim_state(sc.gfp_mask);
3375 reclaim_state.reclaimed_slab = 0;
3376 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003377
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003378 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003379
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003380 p->reclaim_state = NULL;
3381 lockdep_clear_current_reclaim_state();
3382 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003383
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003384 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003386#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388/* It's optimal to keep kswapds on the same CPUs as their memory, but
3389 not required for correctness. So if the last cpu in a node goes
3390 away, we get changed to run anywhere: as the first one comes back,
3391 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003392static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3393 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003395 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003397 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003398 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003399 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303400 const struct cpumask *mask;
3401
3402 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003403
Rusty Russell3e597942009-01-01 10:12:24 +10303404 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003406 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 }
3408 }
3409 return NOTIFY_OK;
3410}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411
Yasunori Goto3218ae12006-06-27 02:53:33 -07003412/*
3413 * This kswapd start function will be called by init and node-hot-add.
3414 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3415 */
3416int kswapd_run(int nid)
3417{
3418 pg_data_t *pgdat = NODE_DATA(nid);
3419 int ret = 0;
3420
3421 if (pgdat->kswapd)
3422 return 0;
3423
3424 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3425 if (IS_ERR(pgdat->kswapd)) {
3426 /* failure at boot is fatal */
3427 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003428 pr_err("Failed to start kswapd on node %d\n", nid);
3429 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003430 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003431 }
3432 return ret;
3433}
3434
David Rientjes8fe23e02009-12-14 17:58:33 -08003435/*
Jiang Liud8adde12012-07-11 14:01:52 -07003436 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003437 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003438 */
3439void kswapd_stop(int nid)
3440{
3441 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3442
Jiang Liud8adde12012-07-11 14:01:52 -07003443 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003444 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003445 NODE_DATA(nid)->kswapd = NULL;
3446 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003447}
3448
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449static int __init kswapd_init(void)
3450{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003451 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003452
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003454 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003455 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456 hotcpu_notifier(cpu_callback, 0);
3457 return 0;
3458}
3459
3460module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003461
3462#ifdef CONFIG_NUMA
3463/*
3464 * Zone reclaim mode
3465 *
3466 * If non-zero call zone_reclaim when the number of free pages falls below
3467 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003468 */
3469int zone_reclaim_mode __read_mostly;
3470
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003471#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003472#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003473#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3474#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3475
Christoph Lameter9eeff232006-01-18 17:42:31 -08003476/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003477 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3478 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3479 * a zone.
3480 */
3481#define ZONE_RECLAIM_PRIORITY 4
3482
Christoph Lameter9eeff232006-01-18 17:42:31 -08003483/*
Christoph Lameter96146342006-07-03 00:24:13 -07003484 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3485 * occur.
3486 */
3487int sysctl_min_unmapped_ratio = 1;
3488
3489/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003490 * If the number of slab pages in a zone grows beyond this percentage then
3491 * slab reclaim needs to occur.
3492 */
3493int sysctl_min_slab_ratio = 5;
3494
Mel Gorman90afa5d2009-06-16 15:33:20 -07003495static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3496{
3497 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3498 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3499 zone_page_state(zone, NR_ACTIVE_FILE);
3500
3501 /*
3502 * It's possible for there to be more file mapped pages than
3503 * accounted for by the pages on the file LRU lists because
3504 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3505 */
3506 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3507}
3508
3509/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3510static long zone_pagecache_reclaimable(struct zone *zone)
3511{
3512 long nr_pagecache_reclaimable;
3513 long delta = 0;
3514
3515 /*
3516 * If RECLAIM_SWAP is set, then all file pages are considered
3517 * potentially reclaimable. Otherwise, we have to worry about
3518 * pages like swapcache and zone_unmapped_file_pages() provides
3519 * a better estimate
3520 */
3521 if (zone_reclaim_mode & RECLAIM_SWAP)
3522 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3523 else
3524 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3525
3526 /* If we can't clean pages, remove dirty pages from consideration */
3527 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3528 delta += zone_page_state(zone, NR_FILE_DIRTY);
3529
3530 /* Watch for any possible underflows due to delta */
3531 if (unlikely(delta > nr_pagecache_reclaimable))
3532 delta = nr_pagecache_reclaimable;
3533
3534 return nr_pagecache_reclaimable - delta;
3535}
3536
Christoph Lameter0ff38492006-09-25 23:31:52 -07003537/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003538 * Try to free up some pages from this zone through reclaim.
3539 */
Andrew Morton179e9632006-03-22 00:08:18 -08003540static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003541{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003542 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003543 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003544 struct task_struct *p = current;
3545 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003546 struct scan_control sc = {
3547 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003548 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003549 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003550 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003551 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003552 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003553 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003554 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003555 struct shrink_control shrink = {
3556 .gfp_mask = sc.gfp_mask,
3557 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003558 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003559
Christoph Lameter9eeff232006-01-18 17:42:31 -08003560 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003561 /*
3562 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3563 * and we also need to be able to write out pages for RECLAIM_WRITE
3564 * and RECLAIM_SWAP.
3565 */
3566 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003567 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003568 reclaim_state.reclaimed_slab = 0;
3569 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003570
Mel Gorman90afa5d2009-06-16 15:33:20 -07003571 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003572 /*
3573 * Free memory by calling shrink zone with increasing
3574 * priorities until we have enough memory freed.
3575 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003576 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003577 shrink_zone(zone, &sc);
3578 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003579 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003580
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003581 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3582 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003583 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003584 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003585 * many pages were freed in this zone. So we take the current
3586 * number of slab pages and shake the slab until it is reduced
3587 * by the same nr_pages that we used for reclaiming unmapped
3588 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003589 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003590 nodes_clear(shrink.nodes_to_scan);
3591 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003592 for (;;) {
3593 unsigned long lru_pages = zone_reclaimable_pages(zone);
3594
3595 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003596 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003597 break;
3598
3599 /* Freed enough memory */
3600 nr_slab_pages1 = zone_page_state(zone,
3601 NR_SLAB_RECLAIMABLE);
3602 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3603 break;
3604 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003605
3606 /*
3607 * Update nr_reclaimed by the number of slab pages we
3608 * reclaimed from this zone.
3609 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003610 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3611 if (nr_slab_pages1 < nr_slab_pages0)
3612 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003613 }
3614
Christoph Lameter9eeff232006-01-18 17:42:31 -08003615 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003616 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003617 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003618 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003619}
Andrew Morton179e9632006-03-22 00:08:18 -08003620
3621int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3622{
Andrew Morton179e9632006-03-22 00:08:18 -08003623 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003624 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003625
3626 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003627 * Zone reclaim reclaims unmapped file backed pages and
3628 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003629 *
Christoph Lameter96146342006-07-03 00:24:13 -07003630 * A small portion of unmapped file backed pages is needed for
3631 * file I/O otherwise pages read by file I/O will be immediately
3632 * thrown out if the zone is overallocated. So we do not reclaim
3633 * if less than a specified percentage of the zone is used by
3634 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003635 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003636 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3637 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003638 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003639
Lisa Du6e543d52013-09-11 14:22:36 -07003640 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003641 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003642
Andrew Morton179e9632006-03-22 00:08:18 -08003643 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003644 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003645 */
David Rientjesd773ed62007-10-16 23:26:01 -07003646 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003647 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003648
3649 /*
3650 * Only run zone reclaim on the local zone or on zones that do not
3651 * have associated processors. This will favor the local processor
3652 * over remote processors and spread off node memory allocations
3653 * as wide as possible.
3654 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003655 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003656 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003657 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003658
3659 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003660 return ZONE_RECLAIM_NOSCAN;
3661
David Rientjesd773ed62007-10-16 23:26:01 -07003662 ret = __zone_reclaim(zone, gfp_mask, order);
3663 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3664
Mel Gorman24cf725182009-06-16 15:33:23 -07003665 if (!ret)
3666 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3667
David Rientjesd773ed62007-10-16 23:26:01 -07003668 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003669}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003670#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003671
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003672/*
3673 * page_evictable - test whether a page is evictable
3674 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003675 *
3676 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003677 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003678 *
3679 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003680 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003681 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003682 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003683 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003684int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003685{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003686 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003687}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003688
Hugh Dickins850465792012-01-20 14:34:19 -08003689#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003690/**
Hugh Dickins24513262012-01-20 14:34:21 -08003691 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3692 * @pages: array of pages to check
3693 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003694 *
Hugh Dickins24513262012-01-20 14:34:21 -08003695 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003696 *
3697 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003698 */
Hugh Dickins24513262012-01-20 14:34:21 -08003699void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003700{
Johannes Weiner925b7672012-01-12 17:18:15 -08003701 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003702 struct zone *zone = NULL;
3703 int pgscanned = 0;
3704 int pgrescued = 0;
3705 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003706
Hugh Dickins24513262012-01-20 14:34:21 -08003707 for (i = 0; i < nr_pages; i++) {
3708 struct page *page = pages[i];
3709 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003710
Hugh Dickins24513262012-01-20 14:34:21 -08003711 pgscanned++;
3712 pagezone = page_zone(page);
3713 if (pagezone != zone) {
3714 if (zone)
3715 spin_unlock_irq(&zone->lru_lock);
3716 zone = pagezone;
3717 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003718 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003719 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003720
Hugh Dickins24513262012-01-20 14:34:21 -08003721 if (!PageLRU(page) || !PageUnevictable(page))
3722 continue;
3723
Hugh Dickins39b5f292012-10-08 16:33:18 -07003724 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003725 enum lru_list lru = page_lru_base_type(page);
3726
Sasha Levin309381fea2014-01-23 15:52:54 -08003727 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003728 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003729 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3730 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003731 pgrescued++;
3732 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003733 }
Hugh Dickins24513262012-01-20 14:34:21 -08003734
3735 if (zone) {
3736 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3737 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3738 spin_unlock_irq(&zone->lru_lock);
3739 }
Hugh Dickins850465792012-01-20 14:34:19 -08003740}
3741#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003742
Johannes Weiner264e56d2011-10-31 17:09:13 -07003743static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003744{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003745 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003746 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003747 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003748 "one, please send an email to linux-mm@kvack.org.\n",
3749 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003750}
3751
3752/*
3753 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3754 * all nodes' unevictable lists for evictable pages
3755 */
3756unsigned long scan_unevictable_pages;
3757
3758int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003759 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003760 size_t *length, loff_t *ppos)
3761{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003762 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003763 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003764 scan_unevictable_pages = 0;
3765 return 0;
3766}
3767
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003768#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003769/*
3770 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3771 * a specified node's per zone unevictable lists for evictable pages.
3772 */
3773
Kay Sievers10fbcf42011-12-21 14:48:43 -08003774static ssize_t read_scan_unevictable_node(struct device *dev,
3775 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003776 char *buf)
3777{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003778 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003779 return sprintf(buf, "0\n"); /* always zero; should fit... */
3780}
3781
Kay Sievers10fbcf42011-12-21 14:48:43 -08003782static ssize_t write_scan_unevictable_node(struct device *dev,
3783 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003784 const char *buf, size_t count)
3785{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003786 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003787 return 1;
3788}
3789
3790
Kay Sievers10fbcf42011-12-21 14:48:43 -08003791static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003792 read_scan_unevictable_node,
3793 write_scan_unevictable_node);
3794
3795int scan_unevictable_register_node(struct node *node)
3796{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003797 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003798}
3799
3800void scan_unevictable_unregister_node(struct node *node)
3801{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003802 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003803}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003804#endif