blob: 30a88b945a44f58e812a5eaf824a087b51a5dbec [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080049#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070055#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
Nick Piggin0f8053a2006-03-22 00:08:33 -080057#include "internal.h"
58
Mel Gorman33906bc2010-08-09 17:19:16 -070059#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
Linus Torvalds1da177e2005-04-16 15:20:36 -070062struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080063 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040067 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
Johannes Weineree814fe2014-08-06 16:06:19 -070069 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070070 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080071
Johannes Weineree814fe2014-08-06 16:06:19 -070072 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070077
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070078 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080079 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080083
Johannes Weineree814fe2014-08-06 16:06:19 -070084 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
Mel Gormanb2e18752016-07-28 15:45:37 -070087 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
89
Johannes Weineree814fe2014-08-06 16:06:19 -070090 unsigned int may_writepage:1;
91
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
94
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
97
Johannes Weiner241994e2015-02-11 15:26:06 -080098 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
100
Johannes Weineree814fe2014-08-06 16:06:19 -0700101 unsigned int hibernation_mode:1;
102
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
105
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
108
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111};
112
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113#ifdef ARCH_HAS_PREFETCH
114#define prefetch_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127#ifdef ARCH_HAS_PREFETCHW
128#define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
132 \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
135 } \
136 } while (0)
137#else
138#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139#endif
140
141/*
142 * From 0 .. 100. Higher means more swappy.
143 */
144int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700145/*
146 * The total number of pages which are beyond the high watermark within all
147 * zones.
148 */
149unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150
151static LIST_HEAD(shrinker_list);
152static DECLARE_RWSEM(shrinker_rwsem);
153
Andrew Mortonc255a452012-07-31 16:43:02 -0700154#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155static bool global_reclaim(struct scan_control *sc)
156{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800157 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800158}
Tejun Heo97c93412015-05-22 18:23:36 -0400159
160/**
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
163 *
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
169 *
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
172 */
173static bool sane_reclaim(struct scan_control *sc)
174{
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
176
177 if (!memcg)
178 return true;
179#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400181 return true;
182#endif
183 return false;
184}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800185#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800186static bool global_reclaim(struct scan_control *sc)
187{
188 return true;
189}
Tejun Heo97c93412015-05-22 18:23:36 -0400190
191static bool sane_reclaim(struct scan_control *sc)
192{
193 return true;
194}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800195#endif
196
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700197/*
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
201 */
202unsigned long zone_reclaimable_pages(struct zone *zone)
203{
204 unsigned long nr;
205
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
211
212 return nr;
213}
214
Mel Gorman599d0c92016-07-28 15:45:31 -0700215unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
Lisa Du6e543d52013-09-11 14:22:36 -0700216{
Mel Gorman599d0c92016-07-28 15:45:31 -0700217 unsigned long nr;
218
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
222
223 if (get_nr_swap_pages() > 0)
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
227
228 return nr;
229}
230
231bool pgdat_reclaimable(struct pglist_data *pgdat)
232{
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700235}
236
Michal Hocko71053132017-02-22 15:45:58 -0800237/**
238 * lruvec_lru_size - Returns the number of pages on the given LRU list.
239 * @lruvec: lru vector
240 * @lru: lru to use
241 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
242 */
243unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800244{
Michal Hocko71053132017-02-22 15:45:58 -0800245 unsigned long lru_size;
246 int zid;
247
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700248 if (!mem_cgroup_disabled())
Michal Hocko71053132017-02-22 15:45:58 -0800249 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
250 else
251 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800252
Michal Hocko71053132017-02-22 15:45:58 -0800253 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
254 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
255 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800256
Michal Hocko71053132017-02-22 15:45:58 -0800257 if (!managed_zone(zone))
258 continue;
Michal Hocko07fc9572017-01-10 16:58:04 -0800259
Michal Hocko71053132017-02-22 15:45:58 -0800260 if (!mem_cgroup_disabled())
261 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
262 else
263 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
264 NR_ZONE_LRU_BASE + lru);
265 lru_size -= min(size, lru_size);
266 }
267
268 return lru_size;
269
Michal Hocko07fc9572017-01-10 16:58:04 -0800270}
271
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000273 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000275int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000277 size_t size = sizeof(*shrinker->nr_deferred);
278
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279 if (shrinker->flags & SHRINKER_NUMA_AWARE)
280 size *= nr_node_ids;
281
282 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
283 if (!shrinker->nr_deferred)
284 return -ENOMEM;
285
Rusty Russell8e1f9362007-07-17 04:03:17 -0700286 down_write(&shrinker_rwsem);
287 list_add_tail(&shrinker->list, &shrinker_list);
288 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000289 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700291EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
293/*
294 * Remove one
295 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700296void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 down_write(&shrinker_rwsem);
299 list_del(&shrinker->list);
300 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700301 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700303EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
305#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000306
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800307static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
308 struct shrinker *shrinker,
309 unsigned long nr_scanned,
310 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000312 unsigned long freed = 0;
313 unsigned long long delta;
314 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700315 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000316 long nr;
317 long new_nr;
318 int nid = shrinkctl->nid;
319 long batch_size = shrinker->batch ? shrinker->batch
320 : SHRINK_BATCH;
Shaohua Li04597be2016-12-12 16:41:50 -0800321 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000322
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700323 freeable = shrinker->count_objects(shrinker, shrinkctl);
324 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000325 return 0;
326
327 /*
328 * copy the current shrinker scan count into a local variable
329 * and zero it so that other concurrent shrinker invocations
330 * don't also do this scanning work.
331 */
332 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
333
334 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800335 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700336 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800337 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000338 total_scan += delta;
339 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800340 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000341 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700342 total_scan = freeable;
Shaohua Li04597be2016-12-12 16:41:50 -0800343 next_deferred = nr;
344 } else
345 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000346
347 /*
348 * We need to avoid excessive windup on filesystem shrinkers
349 * due to large numbers of GFP_NOFS allocations causing the
350 * shrinkers to return -1 all the time. This results in a large
351 * nr being built up so when a shrink that can do some work
352 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700353 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000354 * memory.
355 *
356 * Hence only allow the shrinker to scan the entire cache when
357 * a large delta change is calculated directly.
358 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700359 if (delta < freeable / 4)
360 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000361
362 /*
363 * Avoid risking looping forever due to too large nr value:
364 * never try to free more than twice the estimate number of
365 * freeable entries.
366 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700367 if (total_scan > freeable * 2)
368 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000369
370 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800371 nr_scanned, nr_eligible,
372 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000373
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800374 /*
375 * Normally, we should not scan less than batch_size objects in one
376 * pass to avoid too frequent shrinker calls, but if the slab has less
377 * than batch_size objects in total and we are really tight on memory,
378 * we will try to reclaim all available objects, otherwise we can end
379 * up failing allocations although there are plenty of reclaimable
380 * objects spread over several slabs with usage less than the
381 * batch_size.
382 *
383 * We detect the "tight on memory" situations by looking at the total
384 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700385 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800386 * scanning at high prio and therefore should try to reclaim as much as
387 * possible.
388 */
389 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700390 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000391 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800392 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000393
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800394 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000395 ret = shrinker->scan_objects(shrinker, shrinkctl);
396 if (ret == SHRINK_STOP)
397 break;
398 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000399
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800400 count_vm_events(SLABS_SCANNED, nr_to_scan);
401 total_scan -= nr_to_scan;
Shaohua Li04597be2016-12-12 16:41:50 -0800402 scanned += nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000403
404 cond_resched();
405 }
406
Shaohua Li04597be2016-12-12 16:41:50 -0800407 if (next_deferred >= scanned)
408 next_deferred -= scanned;
409 else
410 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000411 /*
412 * move the unused scan count back into the shrinker in a
413 * manner that handles concurrent updates. If we exhausted the
414 * scan, there is no need to do an update.
415 */
Shaohua Li04597be2016-12-12 16:41:50 -0800416 if (next_deferred > 0)
417 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000418 &shrinker->nr_deferred[nid]);
419 else
420 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
421
Dave Hansendf9024a2014-06-04 16:08:07 -0700422 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000423 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424}
425
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800426/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800427 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800428 * @gfp_mask: allocation context
429 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800430 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800431 * @nr_scanned: pressure numerator
432 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800434 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800436 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
437 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800439 * @memcg specifies the memory cgroup to target. If it is not NULL,
440 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700441 * objects from the memory cgroup specified. Otherwise, only unaware
442 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800443 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800444 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
445 * the available objects should be scanned. Page reclaim for example
446 * passes the number of pages scanned and the number of pages on the
447 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
448 * when it encountered mapped pages. The ratio is further biased by
449 * the ->seeks setting of the shrink function, which indicates the
450 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800452 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800454static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
455 struct mem_cgroup *memcg,
456 unsigned long nr_scanned,
457 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458{
459 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000460 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700462 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800463 return 0;
464
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800465 if (nr_scanned == 0)
466 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467
Minchan Kimf06590b2011-05-24 17:11:11 -0700468 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000469 /*
470 * If we would return 0, our callers would understand that we
471 * have nothing else to shrink and give up trying. By returning
472 * 1 we keep it going and assume we'll be able to shrink next
473 * time.
474 */
475 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700476 goto out;
477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
479 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800480 struct shrink_control sc = {
481 .gfp_mask = gfp_mask,
482 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800483 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800484 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800485
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700486 /*
487 * If kernel memory accounting is disabled, we ignore
488 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
489 * passing NULL for memcg.
490 */
491 if (memcg_kmem_enabled() &&
492 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800493 continue;
494
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800495 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
496 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000497
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800498 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800500
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700502out:
503 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000504 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505}
506
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800507void drop_slab_node(int nid)
508{
509 unsigned long freed;
510
511 do {
512 struct mem_cgroup *memcg = NULL;
513
514 freed = 0;
515 do {
516 freed += shrink_slab(GFP_KERNEL, nid, memcg,
517 1000, 1000);
518 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
519 } while (freed > 10);
520}
521
522void drop_slab(void)
523{
524 int nid;
525
526 for_each_online_node(nid)
527 drop_slab_node(nid);
528}
529
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530static inline int is_page_cache_freeable(struct page *page)
531{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700532 /*
533 * A freeable page cache page is referenced only by the caller
534 * that isolated the page, the page cache radix tree and
535 * optional buffer heads at page->private.
536 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700537 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538}
539
Tejun Heo703c2702015-05-22 17:13:44 -0400540static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541{
Christoph Lameter930d9152006-01-08 01:00:47 -0800542 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400544 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400546 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 return 1;
548 return 0;
549}
550
551/*
552 * We detected a synchronous write error writing a page out. Probably
553 * -ENOSPC. We need to propagate that into the address_space for a subsequent
554 * fsync(), msync() or close().
555 *
556 * The tricky part is that after writepage we cannot touch the mapping: nothing
557 * prevents it from being freed up. But we have a ref on the page and once
558 * that page is locked, the mapping is pinned.
559 *
560 * We're allowed to run sleeping lock_page() here because we know the caller has
561 * __GFP_FS.
562 */
563static void handle_write_error(struct address_space *mapping,
564 struct page *page, int error)
565{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100566 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700567 if (page_mapping(page) == mapping)
568 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 unlock_page(page);
570}
571
Christoph Lameter04e62a22006-06-23 02:03:38 -0700572/* possible outcome of pageout() */
573typedef enum {
574 /* failed to write page out, page is locked */
575 PAGE_KEEP,
576 /* move page to the active list, page is locked */
577 PAGE_ACTIVATE,
578 /* page has been sent to the disk successfully, page is unlocked */
579 PAGE_SUCCESS,
580 /* page is clean and locked */
581 PAGE_CLEAN,
582} pageout_t;
583
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584/*
Andrew Morton1742f192006-03-22 00:08:21 -0800585 * pageout is called by shrink_page_list() for each dirty page.
586 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700588static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700589 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590{
591 /*
592 * If the page is dirty, only perform writeback if that write
593 * will be non-blocking. To prevent this allocation from being
594 * stalled by pagecache activity. But note that there may be
595 * stalls if we need to run get_block(). We could test
596 * PagePrivate for that.
597 *
Al Viro81742022014-04-03 03:17:43 -0400598 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 * this page's queue, we can perform writeback even if that
600 * will block.
601 *
602 * If the page is swapcache, write it back even if that would
603 * block, for some throttling. This happens by accident, because
604 * swap_backing_dev_info is bust: it doesn't reflect the
605 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 */
607 if (!is_page_cache_freeable(page))
608 return PAGE_KEEP;
609 if (!mapping) {
610 /*
611 * Some data journaling orphaned pages can have
612 * page->mapping == NULL while being dirty with clean buffers.
613 */
David Howells266cf652009-04-03 16:42:36 +0100614 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 if (try_to_free_buffers(page)) {
616 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700617 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 return PAGE_CLEAN;
619 }
620 }
621 return PAGE_KEEP;
622 }
623 if (mapping->a_ops->writepage == NULL)
624 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400625 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 return PAGE_KEEP;
627
628 if (clear_page_dirty_for_io(page)) {
629 int res;
630 struct writeback_control wbc = {
631 .sync_mode = WB_SYNC_NONE,
632 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700633 .range_start = 0,
634 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 .for_reclaim = 1,
636 };
637
638 SetPageReclaim(page);
639 res = mapping->a_ops->writepage(page, &wbc);
640 if (res < 0)
641 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800642 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 ClearPageReclaim(page);
644 return PAGE_ACTIVATE;
645 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 if (!PageWriteback(page)) {
648 /* synchronous write or broken a_ops? */
649 ClearPageReclaim(page);
650 }
yalin wang3aa23852016-01-14 15:18:30 -0800651 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700652 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 return PAGE_SUCCESS;
654 }
655
656 return PAGE_CLEAN;
657}
658
Andrew Mortona649fd92006-10-17 00:09:36 -0700659/*
Nick Piggine2867812008-07-25 19:45:30 -0700660 * Same as remove_mapping, but if the page is removed from the mapping, it
661 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700662 */
Johannes Weinera5289102014-04-03 14:47:51 -0700663static int __remove_mapping(struct address_space *mapping, struct page *page,
664 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800665{
Greg Thelenc4843a72015-05-22 17:13:16 -0400666 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400667
Nick Piggin28e4d962006-09-25 23:31:23 -0700668 BUG_ON(!PageLocked(page));
669 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800670
Greg Thelenc4843a72015-05-22 17:13:16 -0400671 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800672 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700673 * The non racy check for a busy page.
674 *
675 * Must be careful with the order of the tests. When someone has
676 * a ref to the page, it may be possible that they dirty it then
677 * drop the reference. So if PageDirty is tested before page_count
678 * here, then the following race may occur:
679 *
680 * get_user_pages(&page);
681 * [user mapping goes away]
682 * write_to(page);
683 * !PageDirty(page) [good]
684 * SetPageDirty(page);
685 * put_page(page);
686 * !page_count(page) [good, discard it]
687 *
688 * [oops, our write_to data is lost]
689 *
690 * Reversing the order of the tests ensures such a situation cannot
691 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700692 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700693 *
694 * Note that if SetPageDirty is always performed via set_page_dirty,
695 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800696 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700697 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800698 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700699 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
700 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700701 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800702 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700703 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800704
705 if (PageSwapCache(page)) {
706 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700707 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800708 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400709 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700710 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700711 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500712 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700713 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500714
715 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700716 /*
717 * Remember a shadow entry for reclaimed file cache in
718 * order to detect refaults, thus thrashing, later on.
719 *
720 * But don't store shadows in an address space that is
721 * already exiting. This is not just an optizimation,
722 * inode reclaim needs to empty out the radix tree or
723 * the nodes are lost. Don't plant shadows behind its
724 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800725 *
726 * We also don't store shadows for DAX mappings because the
727 * only page cache pages found in these are zero pages
728 * covering holes, and because we don't want to mix DAX
729 * exceptional entries and shadow exceptional entries in the
730 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700731 */
732 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800733 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700734 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700735 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400736 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500737
738 if (freepage != NULL)
739 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800740 }
741
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800742 return 1;
743
744cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400745 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800746 return 0;
747}
748
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749/*
Nick Piggine2867812008-07-25 19:45:30 -0700750 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
751 * someone else has a ref on the page, abort and return 0. If it was
752 * successfully detached, return 1. Assumes the caller has a single ref on
753 * this page.
754 */
755int remove_mapping(struct address_space *mapping, struct page *page)
756{
Johannes Weinera5289102014-04-03 14:47:51 -0700757 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700758 /*
759 * Unfreezing the refcount with 1 rather than 2 effectively
760 * drops the pagecache ref for us without requiring another
761 * atomic operation.
762 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700763 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700764 return 1;
765 }
766 return 0;
767}
768
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700769/**
770 * putback_lru_page - put previously isolated page onto appropriate LRU list
771 * @page: page to be put back to appropriate lru list
772 *
773 * Add previously isolated @page to appropriate LRU list.
774 * Page may still be unevictable for other reasons.
775 *
776 * lru_lock must not be held, interrupts must be enabled.
777 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700778void putback_lru_page(struct page *page)
779{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700780 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700781 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700782
Sasha Levin309381fea2014-01-23 15:52:54 -0800783 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700784
785redo:
786 ClearPageUnevictable(page);
787
Hugh Dickins39b5f292012-10-08 16:33:18 -0700788 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700789 /*
790 * For evictable pages, we can use the cache.
791 * In event of a race, worst case is we end up with an
792 * unevictable page on [in]active list.
793 * We know how to handle that.
794 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700795 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700796 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700797 } else {
798 /*
799 * Put unevictable pages directly on zone's unevictable
800 * list.
801 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700802 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700803 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700804 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700805 * When racing with an mlock or AS_UNEVICTABLE clearing
806 * (page is unlocked) make sure that if the other thread
807 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800808 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700809 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700810 * the page back to the evictable list.
811 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700812 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700813 */
814 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700815 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700816
817 /*
818 * page's status can change while we move it among lru. If an evictable
819 * page is on unevictable list, it never be freed. To avoid that,
820 * check after we added it to the list, again.
821 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700822 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700823 if (!isolate_lru_page(page)) {
824 put_page(page);
825 goto redo;
826 }
827 /* This means someone else dropped this page from LRU
828 * So, it will be freed or putback to LRU again. There is
829 * nothing to do here.
830 */
831 }
832
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700833 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700834 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700835 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700836 count_vm_event(UNEVICTABLE_PGCULLED);
837
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700838 put_page(page); /* drop ref from isolate */
839}
840
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800841enum page_references {
842 PAGEREF_RECLAIM,
843 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800844 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800845 PAGEREF_ACTIVATE,
846};
847
848static enum page_references page_check_references(struct page *page,
849 struct scan_control *sc)
850{
Johannes Weiner645747462010-03-05 13:42:22 -0800851 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800852 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800853
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700854 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
855 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800856 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800857
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800858 /*
859 * Mlock lost the isolation race with us. Let try_to_unmap()
860 * move the page to the unevictable list.
861 */
862 if (vm_flags & VM_LOCKED)
863 return PAGEREF_RECLAIM;
864
Johannes Weiner645747462010-03-05 13:42:22 -0800865 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700866 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800867 return PAGEREF_ACTIVATE;
868 /*
869 * All mapped pages start out with page table
870 * references from the instantiating fault, so we need
871 * to look twice if a mapped file page is used more
872 * than once.
873 *
874 * Mark it and spare it for another trip around the
875 * inactive list. Another page table reference will
876 * lead to its activation.
877 *
878 * Note: the mark is set for activated pages as well
879 * so that recently deactivated but used pages are
880 * quickly recovered.
881 */
882 SetPageReferenced(page);
883
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800884 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800885 return PAGEREF_ACTIVATE;
886
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800887 /*
888 * Activate file-backed executable pages after first usage.
889 */
890 if (vm_flags & VM_EXEC)
891 return PAGEREF_ACTIVATE;
892
Johannes Weiner645747462010-03-05 13:42:22 -0800893 return PAGEREF_KEEP;
894 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800895
896 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700897 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800898 return PAGEREF_RECLAIM_CLEAN;
899
900 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800901}
902
Mel Gormane2be15f2013-07-03 15:01:57 -0700903/* Check if a page is dirty or under writeback */
904static void page_check_dirty_writeback(struct page *page,
905 bool *dirty, bool *writeback)
906{
Mel Gormanb4597222013-07-03 15:02:05 -0700907 struct address_space *mapping;
908
Mel Gormane2be15f2013-07-03 15:01:57 -0700909 /*
910 * Anonymous pages are not handled by flushers and must be written
911 * from reclaim context. Do not stall reclaim based on them
912 */
913 if (!page_is_file_cache(page)) {
914 *dirty = false;
915 *writeback = false;
916 return;
917 }
918
919 /* By default assume that the page flags are accurate */
920 *dirty = PageDirty(page);
921 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700922
923 /* Verify dirty/writeback state if the filesystem supports it */
924 if (!page_has_private(page))
925 return;
926
927 mapping = page_mapping(page);
928 if (mapping && mapping->a_ops->is_dirty_writeback)
929 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700930}
931
Nick Piggine2867812008-07-25 19:45:30 -0700932/*
Andrew Morton1742f192006-03-22 00:08:21 -0800933 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 */
Andrew Morton1742f192006-03-22 00:08:21 -0800935static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700936 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700937 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700938 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700939 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700940 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700941 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700942 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700943 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700944 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945{
946 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700947 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700949 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700950 unsigned long nr_dirty = 0;
951 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800952 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700953 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700954 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
956 cond_resched();
957
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 while (!list_empty(page_list)) {
959 struct address_space *mapping;
960 struct page *page;
961 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700962 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700963 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800964 bool lazyfree = false;
965 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966
967 cond_resched();
968
969 page = lru_to_page(page_list);
970 list_del(&page->lru);
971
Nick Piggin529ae9a2008-08-02 12:01:03 +0200972 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 goto keep;
974
Sasha Levin309381fea2014-01-23 15:52:54 -0800975 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976
977 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800978
Hugh Dickins39b5f292012-10-08 16:33:18 -0700979 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700980 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700981
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700982 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800983 goto keep_locked;
984
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 /* Double the slab pressure for mapped and swapcache pages */
986 if (page_mapped(page) || PageSwapCache(page))
987 sc->nr_scanned++;
988
Andy Whitcroftc661b072007-08-22 14:01:26 -0700989 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
990 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
991
Mel Gorman283aba92013-07-03 15:01:51 -0700992 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700993 * The number of dirty pages determines if a zone is marked
994 * reclaim_congested which affects wait_iff_congested. kswapd
995 * will stall and start writing pages if the tail of the LRU
996 * is all dirty unqueued pages.
997 */
998 page_check_dirty_writeback(page, &dirty, &writeback);
999 if (dirty || writeback)
1000 nr_dirty++;
1001
1002 if (dirty && !writeback)
1003 nr_unqueued_dirty++;
1004
Mel Gormand04e8ac2013-07-03 15:02:03 -07001005 /*
1006 * Treat this page as congested if the underlying BDI is or if
1007 * pages are cycling through the LRU so quickly that the
1008 * pages marked for immediate reclaim are making it to the
1009 * end of the LRU a second time.
1010 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001011 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001012 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001013 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001014 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001015 nr_congested++;
1016
1017 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001018 * If a page at the tail of the LRU is under writeback, there
1019 * are three cases to consider.
1020 *
1021 * 1) If reclaim is encountering an excessive number of pages
1022 * under writeback and this page is both under writeback and
1023 * PageReclaim then it indicates that pages are being queued
1024 * for IO but are being recycled through the LRU before the
1025 * IO can complete. Waiting on the page itself risks an
1026 * indefinite stall if it is impossible to writeback the
1027 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001028 * note that the LRU is being scanned too quickly and the
1029 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001030 *
Tejun Heo97c93412015-05-22 18:23:36 -04001031 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001032 * not marked for immediate reclaim, or the caller does not
1033 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1034 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001035 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001036 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001037 * Require may_enter_fs because we would wait on fs, which
1038 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001039 * enter reclaim, and deadlock if it waits on a page for
1040 * which it is needed to do the write (loop masks off
1041 * __GFP_IO|__GFP_FS for this reason); but more thought
1042 * would probably show more reasons.
1043 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001044 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001045 * PageReclaim. memcg does not have any dirty pages
1046 * throttling so we could easily OOM just because too many
1047 * pages are in writeback and there is nothing else to
1048 * reclaim. Wait for the writeback to complete.
1049 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001050 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001051 /* Case 1 above */
1052 if (current_is_kswapd() &&
1053 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001054 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001055 nr_immediate++;
1056 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001057
1058 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001059 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001060 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001061 /*
1062 * This is slightly racy - end_page_writeback()
1063 * might have just cleared PageReclaim, then
1064 * setting PageReclaim here end up interpreted
1065 * as PageReadahead - but that does not matter
1066 * enough to care. What we do want is for this
1067 * page to have PageReclaim set next time memcg
1068 * reclaim reaches the tests above, so it will
1069 * then wait_on_page_writeback() to avoid OOM;
1070 * and it's also appropriate in global reclaim.
1071 */
1072 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001073 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001074 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001075
1076 /* Case 3 above */
1077 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001078 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001079 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001080 /* then go back and try same page again */
1081 list_add_tail(&page->lru, page_list);
1082 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001083 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001084 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
Minchan Kim02c6de82012-10-08 16:31:55 -07001086 if (!force_reclaim)
1087 references = page_check_references(page, sc);
1088
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001089 switch (references) {
1090 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001092 case PAGEREF_KEEP:
1093 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001094 case PAGEREF_RECLAIM:
1095 case PAGEREF_RECLAIM_CLEAN:
1096 ; /* try to reclaim the page below */
1097 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 /*
1100 * Anonymous process memory has backing store?
1101 * Try to allocate it some swap space here.
1102 */
Nick Pigginb291f002008-10-18 20:26:44 -07001103 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001104 if (!(sc->gfp_mask & __GFP_IO))
1105 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001106 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001108 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001109 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110
Mel Gormane2be15f2013-07-03 15:01:57 -07001111 /* Adding to swap updated mapping */
1112 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001113 } else if (unlikely(PageTransHuge(page))) {
1114 /* Split file THP */
1115 if (split_huge_page_to_list(page, page_list))
1116 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001117 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001119 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 /*
1122 * The page is mapped into the page tables of one or more
1123 * processes. Try to unmap it here.
1124 */
1125 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001126 switch (ret = try_to_unmap(page, lazyfree ?
1127 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1128 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 case SWAP_FAIL:
1130 goto activate_locked;
1131 case SWAP_AGAIN:
1132 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001133 case SWAP_MLOCK:
1134 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001135 case SWAP_LZFREE:
1136 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 case SWAP_SUCCESS:
1138 ; /* try to free the page below */
1139 }
1140 }
1141
1142 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001143 /*
1144 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001145 * avoid risk of stack overflow but only writeback
1146 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001147 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001148 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001149 (!current_is_kswapd() ||
Mel Gorman599d0c92016-07-28 15:45:31 -07001150 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001151 /*
1152 * Immediately reclaim when written back.
1153 * Similar in principal to deactivate_page()
1154 * except we already have the page isolated
1155 * and know it's dirty
1156 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001157 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001158 SetPageReclaim(page);
1159
Mel Gormanee728862011-10-31 17:07:38 -07001160 goto keep_locked;
1161 }
1162
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001163 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001165 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001167 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 goto keep_locked;
1169
Mel Gormand950c942015-09-04 15:47:35 -07001170 /*
1171 * Page is dirty. Flush the TLB if a writable entry
1172 * potentially exists to avoid CPU writes after IO
1173 * starts and then write it out here.
1174 */
1175 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001176 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 case PAGE_KEEP:
1178 goto keep_locked;
1179 case PAGE_ACTIVATE:
1180 goto activate_locked;
1181 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001182 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001183 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001184 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001186
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 /*
1188 * A synchronous write - probably a ramdisk. Go
1189 * ahead and try to reclaim the page.
1190 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001191 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 goto keep;
1193 if (PageDirty(page) || PageWriteback(page))
1194 goto keep_locked;
1195 mapping = page_mapping(page);
1196 case PAGE_CLEAN:
1197 ; /* try to free the page below */
1198 }
1199 }
1200
1201 /*
1202 * If the page has buffers, try to free the buffer mappings
1203 * associated with this page. If we succeed we try to free
1204 * the page as well.
1205 *
1206 * We do this even if the page is PageDirty().
1207 * try_to_release_page() does not perform I/O, but it is
1208 * possible for a page to have PageDirty set, but it is actually
1209 * clean (all its buffers are clean). This happens if the
1210 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001211 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 * try_to_release_page() will discover that cleanness and will
1213 * drop the buffers and mark the page clean - it can be freed.
1214 *
1215 * Rarely, pages can have buffers and no ->mapping. These are
1216 * the pages which were not successfully invalidated in
1217 * truncate_complete_page(). We try to drop those buffers here
1218 * and if that worked, and the page is no longer mapped into
1219 * process address space (page_count == 1) it can be freed.
1220 * Otherwise, leave the page on the LRU so it is swappable.
1221 */
David Howells266cf652009-04-03 16:42:36 +01001222 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 if (!try_to_release_page(page, sc->gfp_mask))
1224 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001225 if (!mapping && page_count(page) == 1) {
1226 unlock_page(page);
1227 if (put_page_testzero(page))
1228 goto free_it;
1229 else {
1230 /*
1231 * rare race with speculative reference.
1232 * the speculative reference will free
1233 * this page shortly, so we may
1234 * increment nr_reclaimed here (and
1235 * leave it off the LRU).
1236 */
1237 nr_reclaimed++;
1238 continue;
1239 }
1240 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 }
1242
Minchan Kim854e9ed2016-01-15 16:54:53 -08001243lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001244 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001245 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246
Nick Piggina978d6f2008-10-18 20:26:58 -07001247 /*
1248 * At this point, we have no other references and there is
1249 * no way to pick any more up (removed from LRU, removed
1250 * from pagecache). Can use non-atomic bitops now (and
1251 * we obviously don't have to worry about waking up a process
1252 * waiting on the page lock, because there are no references.
1253 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001254 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001255free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001256 if (ret == SWAP_LZFREE)
1257 count_vm_event(PGLAZYFREED);
1258
Andrew Morton05ff5132006-03-22 00:08:20 -08001259 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001260
1261 /*
1262 * Is there need to periodically free_page_list? It would
1263 * appear not as the counts should be low
1264 */
1265 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 continue;
1267
Nick Pigginb291f002008-10-18 20:26:44 -07001268cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001269 if (PageSwapCache(page))
1270 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001271 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001272 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001273 continue;
1274
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001276 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001277 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001278 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001279 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 SetPageActive(page);
1281 pgactivate++;
1282keep_locked:
1283 unlock_page(page);
1284keep:
1285 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001286 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001288
Johannes Weiner747db952014-08-08 14:19:24 -07001289 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001290 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001291 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001292
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001294 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001295
Mel Gorman8e950282013-07-03 15:02:02 -07001296 *ret_nr_dirty += nr_dirty;
1297 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001298 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001299 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001300 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001301 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302}
1303
Minchan Kim02c6de82012-10-08 16:31:55 -07001304unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1305 struct list_head *page_list)
1306{
1307 struct scan_control sc = {
1308 .gfp_mask = GFP_KERNEL,
1309 .priority = DEF_PRIORITY,
1310 .may_unmap = 1,
1311 };
Mel Gorman8e950282013-07-03 15:02:02 -07001312 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001313 struct page *page, *next;
1314 LIST_HEAD(clean_pages);
1315
1316 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001317 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001318 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001319 ClearPageActive(page);
1320 list_move(&page->lru, &clean_pages);
1321 }
1322 }
1323
Mel Gorman599d0c92016-07-28 15:45:31 -07001324 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001325 TTU_UNMAP|TTU_IGNORE_ACCESS,
1326 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001327 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001328 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001329 return ret;
1330}
1331
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001332/*
1333 * Attempt to remove the specified page from its LRU. Only take this page
1334 * if it is of the appropriate PageActive status. Pages which are being
1335 * freed elsewhere are also ignored.
1336 *
1337 * page: page to consider
1338 * mode: one of the LRU isolation modes defined above
1339 *
1340 * returns 0 on success, -ve errno on failure.
1341 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001342int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001343{
1344 int ret = -EINVAL;
1345
1346 /* Only take pages on the LRU. */
1347 if (!PageLRU(page))
1348 return ret;
1349
Minchan Kime46a2872012-10-08 16:33:48 -07001350 /* Compaction should not handle unevictable pages but CMA can do so */
1351 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001352 return ret;
1353
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001354 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001355
Mel Gormanc8244932012-01-12 17:19:38 -08001356 /*
1357 * To minimise LRU disruption, the caller can indicate that it only
1358 * wants to isolate pages it will be able to operate on without
1359 * blocking - clean pages for the most part.
1360 *
1361 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1362 * is used by reclaim when it is cannot write to backing storage
1363 *
1364 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1365 * that it is possible to migrate without blocking
1366 */
1367 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1368 /* All the caller can do on PageWriteback is block */
1369 if (PageWriteback(page))
1370 return ret;
1371
1372 if (PageDirty(page)) {
1373 struct address_space *mapping;
1374
1375 /* ISOLATE_CLEAN means only clean pages */
1376 if (mode & ISOLATE_CLEAN)
1377 return ret;
1378
1379 /*
1380 * Only pages without mappings or that have a
1381 * ->migratepage callback are possible to migrate
1382 * without blocking
1383 */
1384 mapping = page_mapping(page);
1385 if (mapping && !mapping->a_ops->migratepage)
1386 return ret;
1387 }
1388 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001389
Minchan Kimf80c0672011-10-31 17:06:55 -07001390 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1391 return ret;
1392
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001393 if (likely(get_page_unless_zero(page))) {
1394 /*
1395 * Be careful not to clear PageLRU until after we're
1396 * sure the page is not being freed elsewhere -- the
1397 * page release code relies on it.
1398 */
1399 ClearPageLRU(page);
1400 ret = 0;
1401 }
1402
1403 return ret;
1404}
1405
Mel Gorman7ee36a12016-07-28 15:47:17 -07001406
1407/*
1408 * Update LRU sizes after isolating pages. The LRU size updates must
1409 * be complete before mem_cgroup_update_lru_size due to a santity check.
1410 */
1411static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hocko07fc9572017-01-10 16:58:04 -08001412 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001413{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001414 int zid;
1415
Mel Gorman7ee36a12016-07-28 15:47:17 -07001416 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1417 if (!nr_zone_taken[zid])
1418 continue;
1419
1420 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hocko07fc9572017-01-10 16:58:04 -08001421#ifdef CONFIG_MEMCG
1422 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1423#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001424 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001425
Mel Gorman7ee36a12016-07-28 15:47:17 -07001426}
1427
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001428/*
Mel Gormana52633d2016-07-28 15:45:28 -07001429 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 * shrink the lists perform better by taking out a batch of pages
1431 * and working on them outside the LRU lock.
1432 *
1433 * For pagecache intensive workloads, this function is the hottest
1434 * spot in the kernel (apart from copy_*_user functions).
1435 *
1436 * Appropriate locks must be held before calling this function.
1437 *
1438 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001439 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001441 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001442 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001443 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001444 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 *
1446 * returns how many pages were moved onto *@dst.
1447 */
Andrew Morton69e05942006-03-22 00:08:19 -08001448static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001449 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001450 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001451 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001453 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001454 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001455 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001456 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Mel Gorman599d0c92016-07-28 15:45:31 -07001457 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001458 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001460 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Mel Gormand7f05522016-07-28 15:47:37 -07001461 !list_empty(src);) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001462 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001463
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 page = lru_to_page(src);
1465 prefetchw_prev_lru_page(page, src, flags);
1466
Sasha Levin309381fea2014-01-23 15:52:54 -08001467 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001468
Mel Gormanb2e18752016-07-28 15:45:37 -07001469 if (page_zonenum(page) > sc->reclaim_idx) {
1470 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001471 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001472 continue;
1473 }
1474
Mel Gormand7f05522016-07-28 15:47:37 -07001475 /*
1476 * Account for scanned and skipped separetly to avoid the pgdat
1477 * being prematurely marked unreclaimable by pgdat_reclaimable.
1478 */
1479 scan++;
1480
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001481 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001482 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001483 nr_pages = hpage_nr_pages(page);
1484 nr_taken += nr_pages;
1485 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001486 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001487 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001488
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001489 case -EBUSY:
1490 /* else it is being freed elsewhere */
1491 list_move(&page->lru, src);
1492 continue;
1493
1494 default:
1495 BUG();
1496 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 }
1498
Mel Gormanb2e18752016-07-28 15:45:37 -07001499 /*
1500 * Splice any skipped pages to the start of the LRU list. Note that
1501 * this disrupts the LRU order when reclaiming for lower zones but
1502 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1503 * scanning would soon rescan the same pages to skip and put the
1504 * system at risk of premature OOM.
1505 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001506 if (!list_empty(&pages_skipped)) {
1507 int zid;
Mel Gormand7f05522016-07-28 15:47:37 -07001508 unsigned long total_skipped = 0;
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001509
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001510 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1511 if (!nr_skipped[zid])
1512 continue;
1513
1514 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Mel Gormand7f05522016-07-28 15:47:37 -07001515 total_skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001516 }
Mel Gormand7f05522016-07-28 15:47:37 -07001517
1518 /*
1519 * Account skipped pages as a partial scan as the pgdat may be
1520 * close to unreclaimable. If the LRU list is empty, account
1521 * skipped pages as a full scan.
1522 */
1523 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1524
1525 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001526 }
Hugh Dickinsf6260122012-01-12 17:20:06 -08001527 *nr_scanned = scan;
Mel Gormane5146b12016-07-28 15:46:47 -07001528 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
Hugh Dickins75b00af2012-05-29 15:07:09 -07001529 nr_taken, mode, is_file_lru(lru));
Michal Hocko07fc9572017-01-10 16:58:04 -08001530 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 return nr_taken;
1532}
1533
Nick Piggin62695a82008-10-18 20:26:09 -07001534/**
1535 * isolate_lru_page - tries to isolate a page from its LRU list
1536 * @page: page to isolate from its LRU list
1537 *
1538 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1539 * vmstat statistic corresponding to whatever LRU list the page was on.
1540 *
1541 * Returns 0 if the page was removed from an LRU list.
1542 * Returns -EBUSY if the page was not on an LRU list.
1543 *
1544 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001545 * the active list, it will have PageActive set. If it was found on
1546 * the unevictable list, it will have the PageUnevictable bit set. That flag
1547 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001548 *
1549 * The vmstat statistic corresponding to the list on which the page was
1550 * found will be decremented.
1551 *
1552 * Restrictions:
1553 * (1) Must be called with an elevated refcount on the page. This is a
1554 * fundamentnal difference from isolate_lru_pages (which is called
1555 * without a stable reference).
1556 * (2) the lru_lock must not be held.
1557 * (3) interrupts must be enabled.
1558 */
1559int isolate_lru_page(struct page *page)
1560{
1561 int ret = -EBUSY;
1562
Sasha Levin309381fea2014-01-23 15:52:54 -08001563 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001564 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001565
Nick Piggin62695a82008-10-18 20:26:09 -07001566 if (PageLRU(page)) {
1567 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001568 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001569
Mel Gormana52633d2016-07-28 15:45:28 -07001570 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001571 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001572 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001573 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001574 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001575 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001576 del_page_from_lru_list(page, lruvec, lru);
1577 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001578 }
Mel Gormana52633d2016-07-28 15:45:28 -07001579 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001580 }
1581 return ret;
1582}
1583
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001584/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001585 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1586 * then get resheduled. When there are massive number of tasks doing page
1587 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1588 * the LRU list will go small and be scanned faster than necessary, leading to
1589 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001590 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001591static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001592 struct scan_control *sc)
1593{
1594 unsigned long inactive, isolated;
1595
1596 if (current_is_kswapd())
1597 return 0;
1598
Tejun Heo97c93412015-05-22 18:23:36 -04001599 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001600 return 0;
1601
1602 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001603 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1604 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001605 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001606 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1607 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001608 }
1609
Fengguang Wu3cf23842012-12-18 14:23:31 -08001610 /*
1611 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1612 * won't get blocked by normal direct-reclaimers, forming a circular
1613 * deadlock.
1614 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001615 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001616 inactive >>= 3;
1617
Rik van Riel35cd7812009-09-21 17:01:38 -07001618 return isolated > inactive;
1619}
1620
Mel Gorman66635622010-08-09 17:19:30 -07001621static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001622putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001623{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001624 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001625 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001626 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001627
Mel Gorman66635622010-08-09 17:19:30 -07001628 /*
1629 * Put back any unfreeable pages.
1630 */
Mel Gorman66635622010-08-09 17:19:30 -07001631 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001632 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001633 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001634
Sasha Levin309381fea2014-01-23 15:52:54 -08001635 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001636 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001637 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001638 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001639 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001640 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001641 continue;
1642 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001643
Mel Gorman599d0c92016-07-28 15:45:31 -07001644 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001645
Linus Torvalds7a608572011-01-17 14:42:19 -08001646 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001647 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001648 add_page_to_lru_list(page, lruvec, lru);
1649
Mel Gorman66635622010-08-09 17:19:30 -07001650 if (is_active_lru(lru)) {
1651 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001652 int numpages = hpage_nr_pages(page);
1653 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001654 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001655 if (put_page_testzero(page)) {
1656 __ClearPageLRU(page);
1657 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001658 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001659
1660 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001661 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001662 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001663 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001664 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001665 } else
1666 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001667 }
1668 }
Mel Gorman66635622010-08-09 17:19:30 -07001669
Hugh Dickins3f797682012-01-12 17:20:07 -08001670 /*
1671 * To save our caller's stack, now use input list for pages to free.
1672 */
1673 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001674}
1675
1676/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001677 * If a kernel thread (such as nfsd for loop-back mounts) services
1678 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1679 * In that case we should only throttle if the backing device it is
1680 * writing to is congested. In other cases it is safe to throttle.
1681 */
1682static int current_may_throttle(void)
1683{
1684 return !(current->flags & PF_LESS_THROTTLE) ||
1685 current->backing_dev_info == NULL ||
1686 bdi_write_congested(current->backing_dev_info);
1687}
1688
Minchan Kim91dcade2016-07-28 15:47:40 -07001689static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1690 struct scan_control *sc, enum lru_list lru)
1691{
1692 int zid;
1693 struct zone *zone;
1694 int file = is_file_lru(lru);
1695 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1696
1697 if (!global_reclaim(sc))
1698 return true;
1699
1700 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1701 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001702 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001703 continue;
1704
1705 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1706 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1707 return true;
1708 }
1709
1710 return false;
1711}
1712
NeilBrown399ba0b2014-06-04 16:07:42 -07001713/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001714 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001715 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 */
Mel Gorman66635622010-08-09 17:19:30 -07001717static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001718shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001719 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720{
1721 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001722 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001723 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001724 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001725 unsigned long nr_dirty = 0;
1726 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001727 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001728 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001729 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001730 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001731 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001732 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001733 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001734
Minchan Kim91dcade2016-07-28 15:47:40 -07001735 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1736 return 0;
1737
Mel Gorman599d0c92016-07-28 15:45:31 -07001738 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001739 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001740
1741 /* We are about to die and free our memory. Return now. */
1742 if (fatal_signal_pending(current))
1743 return SWAP_CLUSTER_MAX;
1744 }
1745
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001747
1748 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001749 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001750 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001751 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001752
Mel Gorman599d0c92016-07-28 15:45:31 -07001753 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001755 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1756 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001757
Mel Gorman599d0c92016-07-28 15:45:31 -07001758 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001759 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001760
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001761 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001762 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001763 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001764 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001765 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001766 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001767 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001768 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001769
Hillf Dantond563c052012-03-21 16:34:02 -07001770 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001771 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001772
Mel Gorman599d0c92016-07-28 15:45:31 -07001773 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001774 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1775 &nr_writeback, &nr_immediate,
1776 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001777
Mel Gorman599d0c92016-07-28 15:45:31 -07001778 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001779
Ying Han904249a2012-04-25 16:01:48 -07001780 if (global_reclaim(sc)) {
1781 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001782 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001783 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001784 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001785 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001786
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001787 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001788
Mel Gorman599d0c92016-07-28 15:45:31 -07001789 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001790
Mel Gorman599d0c92016-07-28 15:45:31 -07001791 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001792
Johannes Weiner747db952014-08-08 14:19:24 -07001793 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001794 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001795
Mel Gorman92df3a72011-10-31 17:07:56 -07001796 /*
1797 * If reclaim is isolating dirty pages under writeback, it implies
1798 * that the long-lived page allocation rate is exceeding the page
1799 * laundering rate. Either the global limits are not being effective
1800 * at throttling processes due to the page distribution throughout
1801 * zones or there is heavy usage of a slow backing device. The
1802 * only option is to throttle from reclaim context which is not ideal
1803 * as there is no guarantee the dirtying process is throttled in the
1804 * same way balance_dirty_pages() manages.
1805 *
Mel Gorman8e950282013-07-03 15:02:02 -07001806 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1807 * of pages under pages flagged for immediate reclaim and stall if any
1808 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001809 */
Mel Gorman918fc712013-07-08 16:00:25 -07001810 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001811 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001812
Mel Gormand43006d2013-07-03 15:01:50 -07001813 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001814 * Legacy memcg will stall in page writeback so avoid forcibly
1815 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001816 */
Tejun Heo97c93412015-05-22 18:23:36 -04001817 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001818 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001819 * Tag a zone as congested if all the dirty pages scanned were
1820 * backed by a congested BDI and wait_iff_congested will stall.
1821 */
1822 if (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001823 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001824
1825 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001826 * If dirty pages are scanned that are not queued for IO, it
1827 * implies that flushers are not keeping up. In this case, flag
Mel Gorman599d0c92016-07-28 15:45:31 -07001828 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
Johannes Weiner57054652014-10-09 15:28:17 -07001829 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001830 */
1831 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001832 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001833
1834 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001835 * If kswapd scans pages marked marked for immediate
1836 * reclaim and under writeback (nr_immediate), it implies
1837 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001838 * they are written so also forcibly stall.
1839 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001840 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001841 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001842 }
Mel Gormand43006d2013-07-03 15:01:50 -07001843
Mel Gorman8e950282013-07-03 15:02:02 -07001844 /*
1845 * Stall direct reclaim for IO completions if underlying BDIs or zone
1846 * is congested. Allow kswapd to continue until it starts encountering
1847 * unqueued dirty pages or cycling through the LRU too quickly.
1848 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001849 if (!sc->hibernation_mode && !current_is_kswapd() &&
1850 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001851 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001852
Mel Gorman599d0c92016-07-28 15:45:31 -07001853 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1854 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001855 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001856 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857}
1858
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001859/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 * This moves pages from the active list to the inactive list.
1861 *
1862 * We move them the other way if the page is referenced by one or more
1863 * processes, from rmap.
1864 *
1865 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001866 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001868 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 * this, so instead we remove the pages from the LRU while processing them.
1870 * It is safe to rely on PG_active against the non-LRU pages in here because
1871 * nobody will play with that bit on a non-LRU page.
1872 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001873 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 * But we had to alter page->flags anyway.
1875 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001876
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001877static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001878 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001879 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001880 enum lru_list lru)
1881{
Mel Gorman599d0c92016-07-28 15:45:31 -07001882 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001883 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001884 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001885 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001886
Wu Fengguang3eb41402009-06-16 15:33:13 -07001887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001889 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001890
Sasha Levin309381fea2014-01-23 15:52:54 -08001891 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001892 SetPageLRU(page);
1893
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001894 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001895 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001896 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001897 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001898
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001899 if (put_page_testzero(page)) {
1900 __ClearPageLRU(page);
1901 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001902 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001903
1904 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001905 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001906 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001907 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001908 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001909 } else
1910 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001911 }
1912 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001913
Wu Fengguang3eb41402009-06-16 15:33:13 -07001914 if (!is_active_lru(lru))
1915 __count_vm_events(PGDEACTIVATE, pgmoved);
1916}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001917
Hugh Dickinsf6260122012-01-12 17:20:06 -08001918static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001919 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001920 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001921 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001923 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001924 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001925 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001927 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001928 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001930 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001931 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001932 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001933 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001934 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935
1936 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001937
1938 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001939 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001940 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001941 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001942
Mel Gorman599d0c92016-07-28 15:45:31 -07001943 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001944
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001945 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1946 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001947
Mel Gorman599d0c92016-07-28 15:45:31 -07001948 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001949 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001950
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001951 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001952 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1953 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001954
Mel Gorman599d0c92016-07-28 15:45:31 -07001955 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 while (!list_empty(&l_hold)) {
1958 cond_resched();
1959 page = lru_to_page(&l_hold);
1960 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001961
Hugh Dickins39b5f292012-10-08 16:33:18 -07001962 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001963 putback_lru_page(page);
1964 continue;
1965 }
1966
Mel Gormancc715d92012-03-21 16:34:00 -07001967 if (unlikely(buffer_heads_over_limit)) {
1968 if (page_has_private(page) && trylock_page(page)) {
1969 if (page_has_private(page))
1970 try_to_release_page(page, 0);
1971 unlock_page(page);
1972 }
1973 }
1974
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001975 if (page_referenced(page, 0, sc->target_mem_cgroup,
1976 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001977 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001978 /*
1979 * Identify referenced, file-backed active pages and
1980 * give them one more trip around the active list. So
1981 * that executable code get better chances to stay in
1982 * memory under moderate memory pressure. Anon pages
1983 * are not likely to be evicted by use-once streaming
1984 * IO, plus JVM can create lots of anon VM_EXEC pages,
1985 * so we ignore them here.
1986 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001987 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001988 list_add(&page->lru, &l_active);
1989 continue;
1990 }
1991 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001992
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001993 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 list_add(&page->lru, &l_inactive);
1995 }
1996
Andrew Mortonb5557492009-01-06 14:40:13 -08001997 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001998 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001999 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002000 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002001 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002002 * Count referenced pages from currently used mappings as rotated,
2003 * even though only some of them are actually re-activated. This
2004 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002005 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002006 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002007 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002008
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002009 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2010 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07002011 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2012 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002013
Johannes Weiner747db952014-08-08 14:19:24 -07002014 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07002015 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016}
2017
Rik van Riel59dc76b2016-05-20 16:56:31 -07002018/*
2019 * The inactive anon list should be small enough that the VM never has
2020 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002021 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002022 * The inactive file list should be small enough to leave most memory
2023 * to the established workingset on the scan-resistant active list,
2024 * but large enough to avoid thrashing the aggregate readahead window.
2025 *
2026 * Both inactive lists should also be large enough that each inactive
2027 * page has a chance to be referenced again before it is reclaimed.
2028 *
2029 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2030 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2031 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2032 *
2033 * total target max
2034 * memory ratio inactive
2035 * -------------------------------------
2036 * 10MB 1 5MB
2037 * 100MB 1 50MB
2038 * 1GB 3 250MB
2039 * 10GB 10 0.9GB
2040 * 100GB 31 3GB
2041 * 1TB 101 10GB
2042 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002043 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002044static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2045 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002046{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002047 unsigned long inactive_ratio;
Michal Hocko71053132017-02-22 15:45:58 -08002048 unsigned long inactive, active;
2049 enum lru_list inactive_lru = file * LRU_FILE;
2050 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002051 unsigned long gb;
2052
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002053 /*
2054 * If we don't have swap space, anonymous page deactivation
2055 * is pointless.
2056 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002057 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002058 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002059
Michal Hocko71053132017-02-22 15:45:58 -08002060 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2061 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002062
Rik van Riel59dc76b2016-05-20 16:56:31 -07002063 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2064 if (gb)
2065 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002066 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002067 inactive_ratio = 1;
2068
2069 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002070}
2071
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002072static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002073 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002074{
Rik van Rielb39415b2009-12-14 17:59:48 -08002075 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002076 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002077 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002078 return 0;
2079 }
2080
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002081 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002082}
2083
Johannes Weiner9a265112013-02-22 16:32:17 -08002084enum scan_balance {
2085 SCAN_EQUAL,
2086 SCAN_FRACT,
2087 SCAN_ANON,
2088 SCAN_FILE,
2089};
2090
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002092 * Determine how aggressively the anon and file LRU lists should be
2093 * scanned. The relative value of each set of LRU lists is determined
2094 * by looking at the fraction of the pages scanned we did rotate back
2095 * onto the active list instead of evict.
2096 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002097 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2098 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002099 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002100static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002101 struct scan_control *sc, unsigned long *nr,
2102 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002103{
Vladimir Davydov33377672016-01-20 15:02:59 -08002104 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002105 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002106 u64 fraction[2];
2107 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002108 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002109 unsigned long anon_prio, file_prio;
2110 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002111 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002112 bool force_scan = false;
2113 unsigned long ap, fp;
2114 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002115 bool some_scanned;
2116 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07002117
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07002118 /*
2119 * If the zone or memcg is small, nr[l] can be 0. This
2120 * results in no scanning on this priority and a potential
2121 * priority drop. Global direct reclaim can go to the next
2122 * zone and tends to have no problems. Global kswapd is for
2123 * zone balancing and it needs to scan a minimum amount. When
2124 * reclaiming for a memcg, a priority drop can cause high
2125 * latencies, so it's better to scan a minimum amount there as
2126 * well.
2127 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002128 if (current_is_kswapd()) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002129 if (!pgdat_reclaimable(pgdat))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002130 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08002131 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002132 force_scan = true;
2133 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002134 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002135 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002136
2137 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002138 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002139 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002140 goto out;
2141 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002142
Johannes Weiner10316b32013-02-22 16:32:14 -08002143 /*
2144 * Global reclaim will swap to prevent OOM even with no
2145 * swappiness, but memcg users want to use this knob to
2146 * disable swapping for individual groups completely when
2147 * using the memory controller's swap limit feature would be
2148 * too expensive.
2149 */
Johannes Weiner02695172014-08-06 16:06:17 -07002150 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002151 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002152 goto out;
2153 }
2154
2155 /*
2156 * Do not apply any pressure balancing cleverness when the
2157 * system is close to OOM, scan both anon and file equally
2158 * (unless the swappiness setting disagrees with swapping).
2159 */
Johannes Weiner02695172014-08-06 16:06:17 -07002160 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002161 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002162 goto out;
2163 }
2164
Johannes Weiner11d16c22013-02-22 16:32:15 -08002165 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002166 * Prevent the reclaimer from falling into the cache trap: as
2167 * cache pages start out inactive, every cache fault will tip
2168 * the scan balance towards the file LRU. And as the file LRU
2169 * shrinks, so does the window for rotation from references.
2170 * This means we have a runaway feedback loop where a tiny
2171 * thrashing file LRU becomes infinitely more attractive than
2172 * anon pages. Try to detect this based on file LRU size.
2173 */
2174 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002175 unsigned long pgdatfile;
2176 unsigned long pgdatfree;
2177 int z;
2178 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002179
Mel Gorman599d0c92016-07-28 15:45:31 -07002180 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2181 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2182 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002183
Mel Gorman599d0c92016-07-28 15:45:31 -07002184 for (z = 0; z < MAX_NR_ZONES; z++) {
2185 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002186 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002187 continue;
2188
2189 total_high_wmark += high_wmark_pages(zone);
2190 }
2191
2192 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002193 scan_balance = SCAN_ANON;
2194 goto out;
2195 }
2196 }
2197
2198 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002199 * If there is enough inactive page cache, i.e. if the size of the
2200 * inactive list is greater than that of the active list *and* the
2201 * inactive list actually has some pages to scan on this priority, we
2202 * do not reclaim anything from the anonymous working set right now.
2203 * Without the second condition we could end up never scanning an
2204 * lruvec even if it has plenty of old anonymous pages unless the
2205 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002206 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002207 if (!inactive_list_is_low(lruvec, true, sc) &&
Michal Hocko521e92b2017-02-22 15:46:01 -08002208 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002209 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002210 goto out;
2211 }
2212
Johannes Weiner9a265112013-02-22 16:32:17 -08002213 scan_balance = SCAN_FRACT;
2214
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002215 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002216 * With swappiness at 100, anonymous and file have the same priority.
2217 * This scanning priority is essentially the inverse of IO cost.
2218 */
Johannes Weiner02695172014-08-06 16:06:17 -07002219 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002220 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002221
2222 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002223 * OK, so we have swap space and a fair amount of page cache
2224 * pages. We use the recently rotated / recently scanned
2225 * ratios to determine how valuable each cache is.
2226 *
2227 * Because workloads change over time (and to avoid overflow)
2228 * we keep these statistics as a floating average, which ends
2229 * up weighing recent references more than old ones.
2230 *
2231 * anon in [0], file in [1]
2232 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002233
Michal Hocko71053132017-02-22 15:45:58 -08002234 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2235 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2236 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2237 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002238
Mel Gorman599d0c92016-07-28 15:45:31 -07002239 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002240 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002241 reclaim_stat->recent_scanned[0] /= 2;
2242 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002243 }
2244
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002245 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002246 reclaim_stat->recent_scanned[1] /= 2;
2247 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002248 }
2249
2250 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002251 * The amount of pressure on anon vs file pages is inversely
2252 * proportional to the fraction of recently scanned pages on
2253 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002254 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002255 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002256 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002257
Satoru Moriyafe350042012-05-29 15:06:47 -07002258 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002259 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002260 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002261
Shaohua Li76a33fc2010-05-24 14:32:36 -07002262 fraction[0] = ap;
2263 fraction[1] = fp;
2264 denominator = ap + fp + 1;
2265out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002266 some_scanned = false;
2267 /* Only use force_scan on second pass. */
2268 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002269 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002270 for_each_evictable_lru(lru) {
2271 int file = is_file_lru(lru);
2272 unsigned long size;
2273 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002274
Michal Hocko521e92b2017-02-22 15:46:01 -08002275 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002276 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002277
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002278 if (!scan && pass && force_scan)
2279 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002280
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002281 switch (scan_balance) {
2282 case SCAN_EQUAL:
2283 /* Scan lists relative to size */
2284 break;
2285 case SCAN_FRACT:
2286 /*
2287 * Scan types proportional to swappiness and
2288 * their relative recent reclaim efficiency.
2289 */
2290 scan = div64_u64(scan * fraction[file],
2291 denominator);
2292 break;
2293 case SCAN_FILE:
2294 case SCAN_ANON:
2295 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002296 if ((scan_balance == SCAN_FILE) != file) {
2297 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002298 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002299 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002300 break;
2301 default:
2302 /* Look ma, no brain */
2303 BUG();
2304 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002305
2306 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002307 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002308
Johannes Weiner9a265112013-02-22 16:32:17 -08002309 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002310 * Skip the second pass and don't force_scan,
2311 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002312 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002313 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002314 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002315 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002316}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002317
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002318/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002319 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002320 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002321static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002322 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002323{
Mel Gormanef8f2322016-07-28 15:46:05 -07002324 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002325 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002326 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002327 unsigned long nr_to_scan;
2328 enum lru_list lru;
2329 unsigned long nr_reclaimed = 0;
2330 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2331 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002332 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002333
Vladimir Davydov33377672016-01-20 15:02:59 -08002334 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002335
Mel Gormane82e0562013-07-03 15:01:44 -07002336 /* Record the original scan target for proportional adjustments later */
2337 memcpy(targets, nr, sizeof(nr));
2338
Mel Gorman1a501902014-06-04 16:10:49 -07002339 /*
2340 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2341 * event that can occur when there is little memory pressure e.g.
2342 * multiple streaming readers/writers. Hence, we do not abort scanning
2343 * when the requested number of pages are reclaimed when scanning at
2344 * DEF_PRIORITY on the assumption that the fact we are direct
2345 * reclaiming implies that kswapd is not keeping up and it is best to
2346 * do a batch of work at once. For memcg reclaim one check is made to
2347 * abort proportional reclaim if either the file or anon lru has already
2348 * dropped to zero at the first pass.
2349 */
2350 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2351 sc->priority == DEF_PRIORITY);
2352
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002353 blk_start_plug(&plug);
2354 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2355 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002356 unsigned long nr_anon, nr_file, percentage;
2357 unsigned long nr_scanned;
2358
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002359 for_each_evictable_lru(lru) {
2360 if (nr[lru]) {
2361 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2362 nr[lru] -= nr_to_scan;
2363
2364 nr_reclaimed += shrink_list(lru, nr_to_scan,
2365 lruvec, sc);
2366 }
2367 }
Mel Gormane82e0562013-07-03 15:01:44 -07002368
Michal Hockobd041732016-12-02 17:26:48 -08002369 cond_resched();
2370
Mel Gormane82e0562013-07-03 15:01:44 -07002371 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2372 continue;
2373
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002374 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002375 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002376 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002377 * proportionally what was requested by get_scan_count(). We
2378 * stop reclaiming one LRU and reduce the amount scanning
2379 * proportional to the original scan target.
2380 */
2381 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2382 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2383
Mel Gorman1a501902014-06-04 16:10:49 -07002384 /*
2385 * It's just vindictive to attack the larger once the smaller
2386 * has gone to zero. And given the way we stop scanning the
2387 * smaller below, this makes sure that we only make one nudge
2388 * towards proportionality once we've got nr_to_reclaim.
2389 */
2390 if (!nr_file || !nr_anon)
2391 break;
2392
Mel Gormane82e0562013-07-03 15:01:44 -07002393 if (nr_file > nr_anon) {
2394 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2395 targets[LRU_ACTIVE_ANON] + 1;
2396 lru = LRU_BASE;
2397 percentage = nr_anon * 100 / scan_target;
2398 } else {
2399 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2400 targets[LRU_ACTIVE_FILE] + 1;
2401 lru = LRU_FILE;
2402 percentage = nr_file * 100 / scan_target;
2403 }
2404
2405 /* Stop scanning the smaller of the LRU */
2406 nr[lru] = 0;
2407 nr[lru + LRU_ACTIVE] = 0;
2408
2409 /*
2410 * Recalculate the other LRU scan count based on its original
2411 * scan target and the percentage scanning already complete
2412 */
2413 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2414 nr_scanned = targets[lru] - nr[lru];
2415 nr[lru] = targets[lru] * (100 - percentage) / 100;
2416 nr[lru] -= min(nr[lru], nr_scanned);
2417
2418 lru += LRU_ACTIVE;
2419 nr_scanned = targets[lru] - nr[lru];
2420 nr[lru] = targets[lru] * (100 - percentage) / 100;
2421 nr[lru] -= min(nr[lru], nr_scanned);
2422
2423 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002424 }
2425 blk_finish_plug(&plug);
2426 sc->nr_reclaimed += nr_reclaimed;
2427
2428 /*
2429 * Even if we did not try to evict anon pages at all, we want to
2430 * rebalance the anon lru active/inactive ratio.
2431 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002432 if (inactive_list_is_low(lruvec, false, sc))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002433 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2434 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002435}
2436
Mel Gorman23b9da52012-05-29 15:06:20 -07002437/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002438static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002439{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002440 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002441 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002442 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002443 return true;
2444
2445 return false;
2446}
2447
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002448/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002449 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2450 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2451 * true if more pages should be reclaimed such that when the page allocator
2452 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2453 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002454 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002455static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002456 unsigned long nr_reclaimed,
2457 unsigned long nr_scanned,
2458 struct scan_control *sc)
2459{
2460 unsigned long pages_for_compaction;
2461 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002462 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002463
2464 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002465 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002466 return false;
2467
Mel Gorman28765922011-02-25 14:44:20 -08002468 /* Consider stopping depending on scan and reclaim activity */
2469 if (sc->gfp_mask & __GFP_REPEAT) {
2470 /*
2471 * For __GFP_REPEAT allocations, stop reclaiming if the
2472 * full LRU list has been scanned and we are still failing
2473 * to reclaim pages. This full LRU scan is potentially
2474 * expensive but a __GFP_REPEAT caller really wants to succeed
2475 */
2476 if (!nr_reclaimed && !nr_scanned)
2477 return false;
2478 } else {
2479 /*
2480 * For non-__GFP_REPEAT allocations which can presumably
2481 * fail without consequence, stop if we failed to reclaim
2482 * any pages from the last SWAP_CLUSTER_MAX number of
2483 * pages that were scanned. This will return to the
2484 * caller faster at the risk reclaim/compaction and
2485 * the resulting allocation attempt fails
2486 */
2487 if (!nr_reclaimed)
2488 return false;
2489 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002490
2491 /*
2492 * If we have not reclaimed enough pages for compaction and the
2493 * inactive lists are large enough, continue reclaiming
2494 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002495 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002496 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002497 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002498 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002499 if (sc->nr_reclaimed < pages_for_compaction &&
2500 inactive_lru_pages > pages_for_compaction)
2501 return true;
2502
2503 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002504 for (z = 0; z <= sc->reclaim_idx; z++) {
2505 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002506 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002507 continue;
2508
2509 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002510 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002511 case COMPACT_CONTINUE:
2512 return false;
2513 default:
2514 /* check next zone */
2515 ;
2516 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002517 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002518 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002519}
2520
Mel Gorman970a39a2016-07-28 15:46:35 -07002521static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002522{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002523 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002524 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002525 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002526
Johannes Weiner56600482012-01-12 17:17:59 -08002527 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002528 struct mem_cgroup *root = sc->target_mem_cgroup;
2529 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002530 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002531 .priority = sc->priority,
2532 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002533 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002534 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002535
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002536 nr_reclaimed = sc->nr_reclaimed;
2537 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002538
Andrew Morton694fbc02013-09-24 15:27:37 -07002539 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2540 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002541 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002542 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002543 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002544
Johannes Weiner241994e2015-02-11 15:26:06 -08002545 if (mem_cgroup_low(root, memcg)) {
2546 if (!sc->may_thrash)
2547 continue;
2548 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2549 }
2550
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002551 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002552 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002553
Mel Gormana9dd0a82016-07-28 15:46:02 -07002554 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2555 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002556
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002557 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002558 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002559 memcg, sc->nr_scanned - scanned,
2560 lru_pages);
2561
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002562 /* Record the group's reclaim efficiency */
2563 vmpressure(sc->gfp_mask, memcg, false,
2564 sc->nr_scanned - scanned,
2565 sc->nr_reclaimed - reclaimed);
2566
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002567 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002568 * Direct reclaim and kswapd have to scan all memory
2569 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002570 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002571 *
2572 * Limit reclaim, on the other hand, only cares about
2573 * nr_to_reclaim pages to be reclaimed and it will
2574 * retry with decreasing priority if one round over the
2575 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002576 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002577 if (!global_reclaim(sc) &&
2578 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002579 mem_cgroup_iter_break(root, memcg);
2580 break;
2581 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002582 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002583
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002584 /*
2585 * Shrink the slab caches in the same proportion that
2586 * the eligible LRU pages were scanned.
2587 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002588 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002589 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002590 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002591 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002592
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002593 if (reclaim_state) {
2594 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2595 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002596 }
2597
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002598 /* Record the subtree's reclaim efficiency */
2599 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002600 sc->nr_scanned - nr_scanned,
2601 sc->nr_reclaimed - nr_reclaimed);
2602
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002603 if (sc->nr_reclaimed - nr_reclaimed)
2604 reclaimable = true;
2605
Mel Gormana9dd0a82016-07-28 15:46:02 -07002606 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002607 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002608
2609 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002610}
2611
Vlastimil Babka53853e22014-10-09 15:27:02 -07002612/*
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002613 * Returns true if compaction should go ahead for a costly-order request, or
2614 * the allocation would already succeed without compaction. Return false if we
2615 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002616 */
Mel Gorman4f588332016-07-28 15:46:38 -07002617static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002618{
Mel Gorman31483b62016-07-28 15:45:46 -07002619 unsigned long watermark;
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002620 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002621
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002622 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2623 if (suitable == COMPACT_SUCCESS)
2624 /* Allocation should succeed already. Don't reclaim. */
2625 return true;
2626 if (suitable == COMPACT_SKIPPED)
2627 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002628 return false;
2629
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002630 /*
2631 * Compaction is already possible, but it takes time to run and there
2632 * are potentially other callers using the pages just freed. So proceed
2633 * with reclaim to make a buffer of free pages available to give
2634 * compaction a reasonable chance of completing and allocating the page.
2635 * Note that we won't actually reclaim the whole buffer in one attempt
2636 * as the target watermark in should_continue_reclaim() is lower. But if
2637 * we are already above the high+gap watermark, don't reclaim at all.
2638 */
2639 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2640
2641 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002642}
2643
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644/*
2645 * This is the direct reclaim path, for page-allocating processes. We only
2646 * try to reclaim pages from zones which will satisfy the caller's allocation
2647 * request.
2648 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 * If a zone is deemed to be full of pinned pages then just give it a light
2650 * scan then give up on it.
2651 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002652static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653{
Mel Gormandd1a2392008-04-28 02:12:17 -07002654 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002655 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002656 unsigned long nr_soft_reclaimed;
2657 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002658 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002659 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002660
Mel Gormancc715d92012-03-21 16:34:00 -07002661 /*
2662 * If the number of buffer_heads in the machine exceeds the maximum
2663 * allowed level, force direct reclaim to scan the highmem zone as
2664 * highmem pages could be pinning lowmem pages storing buffer_heads
2665 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002666 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002667 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002668 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002669 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002670 }
Mel Gormancc715d92012-03-21 16:34:00 -07002671
Mel Gormand4debc62010-08-09 17:19:29 -07002672 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002673 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002674 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002675 * Take care memory controller reclaiming has small influence
2676 * to global LRU.
2677 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002678 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002679 if (!cpuset_zone_allowed(zone,
2680 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002681 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002682
Lisa Du6e543d52013-09-11 14:22:36 -07002683 if (sc->priority != DEF_PRIORITY &&
Mel Gorman599d0c92016-07-28 15:45:31 -07002684 !pgdat_reclaimable(zone->zone_pgdat))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002685 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002686
2687 /*
2688 * If we already have plenty of memory free for
2689 * compaction in this zone, don't free any more.
2690 * Even though compaction is invoked for any
2691 * non-zero order, only frequent costly order
2692 * reclamation is disruptive enough to become a
2693 * noticeable problem, like transparent huge
2694 * page allocations.
2695 */
2696 if (IS_ENABLED(CONFIG_COMPACTION) &&
2697 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002698 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002699 sc->compaction_ready = true;
2700 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002701 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002702
Andrew Morton0608f432013-09-24 15:27:41 -07002703 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002704 * Shrink each node in the zonelist once. If the
2705 * zonelist is ordered by zone (not the default) then a
2706 * node may be shrunk multiple times but in that case
2707 * the user prefers lower zones being preserved.
2708 */
2709 if (zone->zone_pgdat == last_pgdat)
2710 continue;
2711
2712 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002713 * This steals pages from memory cgroups over softlimit
2714 * and returns the number of reclaimed pages and
2715 * scanned pages. This works for global memory pressure
2716 * and balancing, not for a memcg's limit.
2717 */
2718 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002719 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002720 sc->order, sc->gfp_mask,
2721 &nr_soft_scanned);
2722 sc->nr_reclaimed += nr_soft_reclaimed;
2723 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002724 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002725 }
Nick Piggin408d8542006-09-25 23:31:27 -07002726
Mel Gorman79dafcd2016-07-28 15:45:53 -07002727 /* See comment about same check for global reclaim above */
2728 if (zone->zone_pgdat == last_pgdat)
2729 continue;
2730 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002731 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 }
Mel Gormane0c23272011-10-31 17:09:33 -07002733
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002734 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002735 * Restore to original mask to avoid the impact on the caller if we
2736 * promoted it to __GFP_HIGHMEM.
2737 */
2738 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002740
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741/*
2742 * This is the main entry point to direct page reclaim.
2743 *
2744 * If a full scan of the inactive list fails to free enough memory then we
2745 * are "out of memory" and something needs to be killed.
2746 *
2747 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2748 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002749 * caller can't do much about. We kick the writeback threads and take explicit
2750 * naps in the hope that some of these pages can be written. But if the
2751 * allocating task holds filesystem locks which prevent writeout this might not
2752 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002753 *
2754 * returns: 0, if no pages reclaimed
2755 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 */
Mel Gormandac1d272008-04-28 02:12:12 -07002757static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002758 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759{
Johannes Weiner241994e2015-02-11 15:26:06 -08002760 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002761 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002762 unsigned long writeback_threshold;
Johannes Weiner241994e2015-02-11 15:26:06 -08002763retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002764 delayacct_freepages_start();
2765
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002766 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002767 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002769 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002770 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2771 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002772 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002773 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002774
Balbir Singh66e17072008-02-07 00:13:56 -08002775 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002776 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002777 break;
2778
2779 if (sc->compaction_ready)
2780 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781
2782 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002783 * If we're getting trouble reclaiming, start doing
2784 * writepage even in laptop mode.
2785 */
2786 if (sc->priority < DEF_PRIORITY - 2)
2787 sc->may_writepage = 1;
2788
2789 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 * Try to write back as many pages as we just scanned. This
2791 * tends to cause slow streaming writers to write data to the
2792 * disk smoothly, at the dirtying rate, which is nice. But
2793 * that's undesirable in laptop mode, where we *want* lumpy
2794 * writeout. So in laptop mode, write out the whole world.
2795 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002796 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2797 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002798 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2799 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002800 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002802 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002803
Keika Kobayashi873b4772008-07-25 01:48:52 -07002804 delayacct_freepages_end();
2805
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002806 if (sc->nr_reclaimed)
2807 return sc->nr_reclaimed;
2808
Mel Gorman0cee34f2012-01-12 17:19:49 -08002809 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002810 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002811 return 1;
2812
Johannes Weiner241994e2015-02-11 15:26:06 -08002813 /* Untapped cgroup reserves? Don't OOM, retry. */
2814 if (!sc->may_thrash) {
2815 sc->priority = initial_priority;
2816 sc->may_thrash = 1;
2817 goto retry;
2818 }
2819
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002820 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821}
2822
Mel Gorman55150612012-07-31 16:44:35 -07002823static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2824{
2825 struct zone *zone;
2826 unsigned long pfmemalloc_reserve = 0;
2827 unsigned long free_pages = 0;
2828 int i;
2829 bool wmark_ok;
2830
2831 for (i = 0; i <= ZONE_NORMAL; i++) {
2832 zone = &pgdat->node_zones[i];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002833 if (!managed_zone(zone) ||
Mel Gorman599d0c92016-07-28 15:45:31 -07002834 pgdat_reclaimable_pages(pgdat) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002835 continue;
2836
Mel Gorman55150612012-07-31 16:44:35 -07002837 pfmemalloc_reserve += min_wmark_pages(zone);
2838 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2839 }
2840
Mel Gorman675becc2014-06-04 16:07:35 -07002841 /* If there are no reserves (unexpected config) then do not throttle */
2842 if (!pfmemalloc_reserve)
2843 return true;
2844
Mel Gorman55150612012-07-31 16:44:35 -07002845 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2846
2847 /* kswapd must be awake if processes are being throttled */
2848 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002849 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002850 (enum zone_type)ZONE_NORMAL);
2851 wake_up_interruptible(&pgdat->kswapd_wait);
2852 }
2853
2854 return wmark_ok;
2855}
2856
2857/*
2858 * Throttle direct reclaimers if backing storage is backed by the network
2859 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2860 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002861 * when the low watermark is reached.
2862 *
2863 * Returns true if a fatal signal was delivered during throttling. If this
2864 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002865 */
Mel Gorman50694c22012-11-26 16:29:48 -08002866static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002867 nodemask_t *nodemask)
2868{
Mel Gorman675becc2014-06-04 16:07:35 -07002869 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002870 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002871 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002872
2873 /*
2874 * Kernel threads should not be throttled as they may be indirectly
2875 * responsible for cleaning pages necessary for reclaim to make forward
2876 * progress. kjournald for example may enter direct reclaim while
2877 * committing a transaction where throttling it could forcing other
2878 * processes to block on log_wait_commit().
2879 */
2880 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002881 goto out;
2882
2883 /*
2884 * If a fatal signal is pending, this process should not throttle.
2885 * It should return quickly so it can exit and free its memory
2886 */
2887 if (fatal_signal_pending(current))
2888 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002889
Mel Gorman675becc2014-06-04 16:07:35 -07002890 /*
2891 * Check if the pfmemalloc reserves are ok by finding the first node
2892 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2893 * GFP_KERNEL will be required for allocating network buffers when
2894 * swapping over the network so ZONE_HIGHMEM is unusable.
2895 *
2896 * Throttling is based on the first usable node and throttled processes
2897 * wait on a queue until kswapd makes progress and wakes them. There
2898 * is an affinity then between processes waking up and where reclaim
2899 * progress has been made assuming the process wakes on the same node.
2900 * More importantly, processes running on remote nodes will not compete
2901 * for remote pfmemalloc reserves and processes on different nodes
2902 * should make reasonable progress.
2903 */
2904 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002905 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002906 if (zone_idx(zone) > ZONE_NORMAL)
2907 continue;
2908
2909 /* Throttle based on the first usable node */
2910 pgdat = zone->zone_pgdat;
2911 if (pfmemalloc_watermark_ok(pgdat))
2912 goto out;
2913 break;
2914 }
2915
2916 /* If no zone was usable by the allocation flags then do not throttle */
2917 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002918 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002919
Mel Gorman68243e72012-07-31 16:44:39 -07002920 /* Account for the throttling */
2921 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2922
Mel Gorman55150612012-07-31 16:44:35 -07002923 /*
2924 * If the caller cannot enter the filesystem, it's possible that it
2925 * is due to the caller holding an FS lock or performing a journal
2926 * transaction in the case of a filesystem like ext[3|4]. In this case,
2927 * it is not safe to block on pfmemalloc_wait as kswapd could be
2928 * blocked waiting on the same lock. Instead, throttle for up to a
2929 * second before continuing.
2930 */
2931 if (!(gfp_mask & __GFP_FS)) {
2932 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2933 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002934
2935 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002936 }
2937
2938 /* Throttle until kswapd wakes the process */
2939 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2940 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002941
2942check_pending:
2943 if (fatal_signal_pending(current))
2944 return true;
2945
2946out:
2947 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002948}
2949
Mel Gormandac1d272008-04-28 02:12:12 -07002950unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002951 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002952{
Mel Gorman33906bc2010-08-09 17:19:16 -07002953 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002954 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002955 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002956 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07002957 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07002958 .order = order,
2959 .nodemask = nodemask,
2960 .priority = DEF_PRIORITY,
2961 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002962 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002963 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002964 };
2965
Mel Gorman55150612012-07-31 16:44:35 -07002966 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002967 * Do not enter reclaim if fatal signal was delivered while throttled.
2968 * 1 is returned so that the page allocator does not OOM kill at this
2969 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002970 */
Mel Gorman50694c22012-11-26 16:29:48 -08002971 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002972 return 1;
2973
Mel Gorman33906bc2010-08-09 17:19:16 -07002974 trace_mm_vmscan_direct_reclaim_begin(order,
2975 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07002976 gfp_mask,
2977 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002978
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002979 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002980
2981 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2982
2983 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002984}
2985
Andrew Mortonc255a452012-07-31 16:43:02 -07002986#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002987
Mel Gormana9dd0a82016-07-28 15:46:02 -07002988unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002989 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07002990 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07002991 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002992{
2993 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002994 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002995 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002996 .may_writepage = !laptop_mode,
2997 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07002998 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07002999 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003000 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003001 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003002
Balbir Singh4e416952009-09-23 15:56:39 -07003003 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3004 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003005
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003006 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003007 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003008 sc.gfp_mask,
3009 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003010
Balbir Singh4e416952009-09-23 15:56:39 -07003011 /*
3012 * NOTE: Although we can get the priority field, using it
3013 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003014 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003015 * will pick up pages from other mem cgroup's as well. We hack
3016 * the priority and make it zero.
3017 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003018 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003019
3020 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3021
Ying Han0ae5e892011-05-26 16:25:25 -07003022 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003023 return sc.nr_reclaimed;
3024}
3025
Johannes Weiner72835c82012-01-12 17:18:32 -08003026unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003027 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003028 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003029 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003030{
Balbir Singh4e416952009-09-23 15:56:39 -07003031 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003032 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003033 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003034 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003035 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003036 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3037 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003038 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003039 .target_mem_cgroup = memcg,
3040 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003041 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003042 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003043 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003044 };
Balbir Singh66e17072008-02-07 00:13:56 -08003045
Ying Han889976d2011-05-26 16:25:33 -07003046 /*
3047 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3048 * take care of from where we get pages. So the node where we start the
3049 * scan does not need to be the current node.
3050 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003051 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003052
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003053 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003054
3055 trace_mm_vmscan_memcg_reclaim_begin(0,
3056 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003057 sc.gfp_mask,
3058 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003059
Johannes Weiner89a28482016-10-27 17:46:56 -07003060 current->flags |= PF_MEMALLOC;
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003061 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weiner89a28482016-10-27 17:46:56 -07003062 current->flags &= ~PF_MEMALLOC;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003063
3064 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3065
3066 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003067}
3068#endif
3069
Mel Gorman1d82de62016-07-28 15:45:43 -07003070static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003071 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003072{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003073 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003074
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003075 if (!total_swap_pages)
3076 return;
3077
3078 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3079 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003080 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003081
Mel Gormanf8d1a312016-07-28 15:47:34 -07003082 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003083 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003084 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003085
3086 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3087 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003088}
3089
Mel Gorman31483b62016-07-28 15:45:46 -07003090static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003091{
Mel Gorman31483b62016-07-28 15:45:46 -07003092 unsigned long mark = high_wmark_pages(zone);
Johannes Weiner60cefed2012-11-29 13:54:23 -08003093
Mel Gorman6256c6b42016-07-28 15:45:56 -07003094 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3095 return false;
3096
3097 /*
3098 * If any eligible zone is balanced then the node is not considered
3099 * to be congested or dirty
3100 */
3101 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3102 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3103
3104 return true;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003105}
3106
Mel Gorman1741c872011-01-13 15:46:21 -08003107/*
Mel Gorman55150612012-07-31 16:44:35 -07003108 * Prepare kswapd for sleeping. This verifies that there are no processes
3109 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3110 *
3111 * Returns true if kswapd is ready to sleep
3112 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003113static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003114{
Mel Gorman1d82de62016-07-28 15:45:43 -07003115 int i;
3116
Mel Gorman55150612012-07-31 16:44:35 -07003117 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003118 * The throttled processes are normally woken up in balance_pgdat() as
3119 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3120 * race between when kswapd checks the watermarks and a process gets
3121 * throttled. There is also a potential race if processes get
3122 * throttled, kswapd wakes, a large process exits thereby balancing the
3123 * zones, which causes kswapd to exit balance_pgdat() before reaching
3124 * the wake up checks. If kswapd is going to sleep, no process should
3125 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3126 * the wake up is premature, processes will wake kswapd and get
3127 * throttled again. The difference from wake ups in balance_pgdat() is
3128 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003129 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003130 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3131 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003132
Mel Gorman1d82de62016-07-28 15:45:43 -07003133 for (i = 0; i <= classzone_idx; i++) {
3134 struct zone *zone = pgdat->node_zones + i;
3135
Mel Gorman6aa303d2016-09-01 16:14:55 -07003136 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003137 continue;
3138
Mel Gorman38087d92016-07-28 15:45:49 -07003139 if (!zone_balanced(zone, order, classzone_idx))
3140 return false;
Mel Gorman1d82de62016-07-28 15:45:43 -07003141 }
3142
Mel Gorman38087d92016-07-28 15:45:49 -07003143 return true;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003144}
3145
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003147 * kswapd shrinks a node of pages that are at or below the highest usable
3148 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003149 *
3150 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003151 * reclaim or if the lack of progress was due to pages under writeback.
3152 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003153 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003154static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003155 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003156{
Mel Gorman1d82de62016-07-28 15:45:43 -07003157 struct zone *zone;
3158 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003159
Mel Gorman1d82de62016-07-28 15:45:43 -07003160 /* Reclaim a number of pages proportional to the number of zones */
3161 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003162 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003163 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003164 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003165 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003166
Mel Gorman1d82de62016-07-28 15:45:43 -07003167 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003168 }
3169
Mel Gorman1d82de62016-07-28 15:45:43 -07003170 /*
3171 * Historically care was taken to put equal pressure on all zones but
3172 * now pressure is applied based on node LRU order.
3173 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003174 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003175
3176 /*
3177 * Fragmentation may mean that the system cannot be rebalanced for
3178 * high-order allocations. If twice the allocation size has been
3179 * reclaimed then recheck watermarks only at order-0 to prevent
3180 * excessive reclaim. Assume that a process requested a high-order
3181 * can direct reclaim/compact.
3182 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003183 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003184 sc->order = 0;
3185
Mel Gormanb8e83b92013-07-03 15:01:45 -07003186 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003187}
3188
3189/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003190 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3191 * that are eligible for use by the caller until at least one zone is
3192 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003194 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 *
3196 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003197 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003198 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3199 * or lower is eligible for reclaim until at least one usable zone is
3200 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003202static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003205 unsigned long nr_soft_reclaimed;
3206 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003207 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003208 struct scan_control sc = {
3209 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003210 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003211 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003212 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003213 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003214 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003215 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003216 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003218 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003219 bool raise_priority = true;
3220
3221 sc.nr_reclaimed = 0;
Mel Gorman84c7a772016-07-28 15:46:44 -07003222 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223
Mel Gorman86c79f62016-07-28 15:45:59 -07003224 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003225 * If the number of buffer_heads exceeds the maximum allowed
3226 * then consider reclaiming from all zones. This has a dual
3227 * purpose -- on 64-bit systems it is expected that
3228 * buffer_heads are stripped during active rotation. On 32-bit
3229 * systems, highmem pages can pin lowmem memory and shrinking
3230 * buffers can relieve lowmem pressure. Reclaim may still not
3231 * go ahead if all eligible zones for the original allocation
3232 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003233 */
3234 if (buffer_heads_over_limit) {
3235 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3236 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003237 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003238 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239
Mel Gorman970a39a2016-07-28 15:46:35 -07003240 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003241 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003244
Mel Gorman86c79f62016-07-28 15:45:59 -07003245 /*
3246 * Only reclaim if there are no eligible zones. Check from
3247 * high to low zone as allocations prefer higher zones.
3248 * Scanning from low to high zone would allow congestion to be
3249 * cleared during a very small window when a small low
3250 * zone was balanced even under extreme pressure when the
Mel Gorman84c7a772016-07-28 15:46:44 -07003251 * overall node may be congested. Note that sc.reclaim_idx
3252 * is not used as buffer_heads_over_limit may have adjusted
3253 * it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003254 */
Mel Gorman84c7a772016-07-28 15:46:44 -07003255 for (i = classzone_idx; i >= 0; i--) {
Mel Gorman86c79f62016-07-28 15:45:59 -07003256 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003257 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003258 continue;
3259
Mel Gorman84c7a772016-07-28 15:46:44 -07003260 if (zone_balanced(zone, sc.order, classzone_idx))
Mel Gorman86c79f62016-07-28 15:45:59 -07003261 goto out;
3262 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003263
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003265 * Do some background aging of the anon list, to give
3266 * pages a chance to be referenced before reclaiming. All
3267 * pages are rotated regardless of classzone as this is
3268 * about consistent aging.
3269 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003270 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003271
3272 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003273 * If we're getting trouble reclaiming, start doing writepage
3274 * even in laptop mode.
3275 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003276 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003277 sc.may_writepage = 1;
3278
Mel Gorman1d82de62016-07-28 15:45:43 -07003279 /* Call soft limit reclaim before calling shrink_node. */
3280 sc.nr_scanned = 0;
3281 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003282 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003283 sc.gfp_mask, &nr_soft_scanned);
3284 sc.nr_reclaimed += nr_soft_reclaimed;
3285
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003286 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003287 * There should be no need to raise the scanning priority if
3288 * enough pages are already being scanned that that high
3289 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003291 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003292 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003293
3294 /*
3295 * If the low watermark is met there is no need for processes
3296 * to be throttled on pfmemalloc_wait as they should not be
3297 * able to safely make forward progress. Wake them
3298 */
3299 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3300 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003301 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003302
Mel Gormanb8e83b92013-07-03 15:01:45 -07003303 /* Check if kswapd should be suspending */
3304 if (try_to_freeze() || kthread_should_stop())
3305 break;
3306
3307 /*
3308 * Raise priority if scanning rate is too low or there was no
3309 * progress in reclaiming pages
3310 */
3311 if (raise_priority || !sc.nr_reclaimed)
3312 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003313 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
Mel Gormanb8e83b92013-07-03 15:01:45 -07003315out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003316 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003317 * Return the order kswapd stopped reclaiming at as
3318 * prepare_kswapd_sleep() takes it into account. If another caller
3319 * entered the allocator slow path while kswapd was awake, order will
3320 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003321 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003322 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323}
3324
Mel Gorman38087d92016-07-28 15:45:49 -07003325static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3326 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003327{
3328 long remaining = 0;
3329 DEFINE_WAIT(wait);
3330
3331 if (freezing(current) || kthread_should_stop())
3332 return;
3333
3334 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3335
3336 /* Try to sleep for a short interval */
Mel Gormand9f21d42016-07-28 15:46:41 -07003337 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003338 /*
3339 * Compaction records what page blocks it recently failed to
3340 * isolate pages from and skips them in the future scanning.
3341 * When kswapd is going to sleep, it is reasonable to assume
3342 * that pages and compaction may succeed so reset the cache.
3343 */
3344 reset_isolation_suitable(pgdat);
3345
3346 /*
3347 * We have freed the memory, now we should compact it to make
3348 * allocation of the requested order possible.
3349 */
Mel Gorman38087d92016-07-28 15:45:49 -07003350 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003351
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003352 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003353
3354 /*
3355 * If woken prematurely then reset kswapd_classzone_idx and
3356 * order. The values will either be from a wakeup request or
3357 * the previous request that slept prematurely.
3358 */
3359 if (remaining) {
3360 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3361 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3362 }
3363
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003364 finish_wait(&pgdat->kswapd_wait, &wait);
3365 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3366 }
3367
3368 /*
3369 * After a short sleep, check if it was a premature sleep. If not, then
3370 * go fully to sleep until explicitly woken up.
3371 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003372 if (!remaining &&
3373 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003374 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3375
3376 /*
3377 * vmstat counters are not perfectly accurate and the estimated
3378 * value for counters such as NR_FREE_PAGES can deviate from the
3379 * true value by nr_online_cpus * threshold. To avoid the zone
3380 * watermarks being breached while under pressure, we reduce the
3381 * per-cpu vmstat threshold while kswapd is awake and restore
3382 * them before going back to sleep.
3383 */
3384 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003385
3386 if (!kthread_should_stop())
3387 schedule();
3388
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003389 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3390 } else {
3391 if (remaining)
3392 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3393 else
3394 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3395 }
3396 finish_wait(&pgdat->kswapd_wait, &wait);
3397}
3398
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399/*
3400 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003401 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 *
3403 * This basically trickles out pages so that we have _some_
3404 * free memory available even if there is no other activity
3405 * that frees anything up. This is needed for things like routing
3406 * etc, where we otherwise might have all activity going on in
3407 * asynchronous contexts that cannot page things out.
3408 *
3409 * If there are applications that are active memory-allocators
3410 * (most normal use), this basically shouldn't matter.
3411 */
3412static int kswapd(void *p)
3413{
Mel Gorman38087d92016-07-28 15:45:49 -07003414 unsigned int alloc_order, reclaim_order, classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 pg_data_t *pgdat = (pg_data_t*)p;
3416 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003417
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 struct reclaim_state reclaim_state = {
3419 .reclaimed_slab = 0,
3420 };
Rusty Russella70f7302009-03-13 14:49:46 +10303421 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Nick Piggincf40bd12009-01-21 08:12:39 +01003423 lockdep_set_current_reclaim_state(GFP_KERNEL);
3424
Rusty Russell174596a2009-01-01 10:12:29 +10303425 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003426 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 current->reclaim_state = &reclaim_state;
3428
3429 /*
3430 * Tell the memory management that we're a "memory allocator",
3431 * and that if we need more memory we should get access to it
3432 * regardless (see "__alloc_pages()"). "kswapd" should
3433 * never get caught in the normal page freeing logic.
3434 *
3435 * (Kswapd normally doesn't need memory anyway, but sometimes
3436 * you need a small amount of memory in order to be able to
3437 * page out something else, and this flag essentially protects
3438 * us from recursively trying to free more memory as we're
3439 * trying to free the first piece of memory in the first place).
3440 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003441 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003442 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443
Mel Gorman38087d92016-07-28 15:45:49 -07003444 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3445 pgdat->kswapd_classzone_idx = classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003447 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003448
Mel Gorman38087d92016-07-28 15:45:49 -07003449kswapd_try_sleep:
3450 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3451 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003452
Mel Gorman38087d92016-07-28 15:45:49 -07003453 /* Read the new order and classzone_idx */
3454 alloc_order = reclaim_order = pgdat->kswapd_order;
3455 classzone_idx = pgdat->kswapd_classzone_idx;
3456 pgdat->kswapd_order = 0;
3457 pgdat->kswapd_classzone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458
David Rientjes8fe23e02009-12-14 17:58:33 -08003459 ret = try_to_freeze();
3460 if (kthread_should_stop())
3461 break;
3462
3463 /*
3464 * We can speed up thawing tasks if we don't call balance_pgdat
3465 * after returning from the refrigerator
3466 */
Mel Gorman38087d92016-07-28 15:45:49 -07003467 if (ret)
3468 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003469
Mel Gorman38087d92016-07-28 15:45:49 -07003470 /*
3471 * Reclaim begins at the requested order but if a high-order
3472 * reclaim fails then kswapd falls back to reclaiming for
3473 * order-0. If that happens, kswapd will consider sleeping
3474 * for the order it finished reclaiming at (reclaim_order)
3475 * but kcompactd is woken to compact for the original
3476 * request (alloc_order).
3477 */
Mel Gormane5146b12016-07-28 15:46:47 -07003478 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3479 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003480 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3481 if (reclaim_order < alloc_order)
3482 goto kswapd_try_sleep;
3483
3484 alloc_order = reclaim_order = pgdat->kswapd_order;
3485 classzone_idx = pgdat->kswapd_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003487
Johannes Weiner71abdc12014-06-06 14:35:35 -07003488 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003489 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003490 lockdep_clear_current_reclaim_state();
3491
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 return 0;
3493}
3494
3495/*
3496 * A zone is low on free memory, so wake its kswapd task to service it.
3497 */
Mel Gorman99504742011-01-13 15:46:20 -08003498void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499{
3500 pg_data_t *pgdat;
Mel Gormane1a55632016-07-28 15:46:26 -07003501 int z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502
Mel Gorman6aa303d2016-09-01 16:14:55 -07003503 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 return;
3505
Vladimir Davydov344736f2014-10-20 15:50:30 +04003506 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003508 pgdat = zone->zone_pgdat;
Mel Gorman38087d92016-07-28 15:45:49 -07003509 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3510 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003511 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003513
3514 /* Only wake kswapd if all zones are unbalanced */
3515 for (z = 0; z <= classzone_idx; z++) {
3516 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003517 if (!managed_zone(zone))
Mel Gormane1a55632016-07-28 15:46:26 -07003518 continue;
3519
3520 if (zone_balanced(zone, order, classzone_idx))
3521 return;
3522 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08003523
3524 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003525 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526}
3527
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003528#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003530 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003531 * freed pages.
3532 *
3533 * Rather than trying to age LRUs the aim is to preserve the overall
3534 * LRU order by reclaiming preferentially
3535 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003537unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003539 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003540 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003541 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003542 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003543 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003544 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003545 .may_writepage = 1,
3546 .may_unmap = 1,
3547 .may_swap = 1,
3548 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003550 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003551 struct task_struct *p = current;
3552 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003554 p->flags |= PF_MEMALLOC;
3555 lockdep_set_current_reclaim_state(sc.gfp_mask);
3556 reclaim_state.reclaimed_slab = 0;
3557 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003558
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003559 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003560
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003561 p->reclaim_state = NULL;
3562 lockdep_clear_current_reclaim_state();
3563 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003564
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003565 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003567#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569/* It's optimal to keep kswapds on the same CPUs as their memory, but
3570 not required for correctness. So if the last cpu in a node goes
3571 away, we get changed to run anywhere: as the first one comes back,
3572 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003573static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3574 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003576 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003578 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003579 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003580 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303581 const struct cpumask *mask;
3582
3583 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003584
Rusty Russell3e597942009-01-01 10:12:24 +10303585 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003587 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 }
3589 }
3590 return NOTIFY_OK;
3591}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592
Yasunori Goto3218ae12006-06-27 02:53:33 -07003593/*
3594 * This kswapd start function will be called by init and node-hot-add.
3595 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3596 */
3597int kswapd_run(int nid)
3598{
3599 pg_data_t *pgdat = NODE_DATA(nid);
3600 int ret = 0;
3601
3602 if (pgdat->kswapd)
3603 return 0;
3604
3605 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3606 if (IS_ERR(pgdat->kswapd)) {
3607 /* failure at boot is fatal */
3608 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003609 pr_err("Failed to start kswapd on node %d\n", nid);
3610 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003611 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003612 }
3613 return ret;
3614}
3615
David Rientjes8fe23e02009-12-14 17:58:33 -08003616/*
Jiang Liud8adde12012-07-11 14:01:52 -07003617 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003618 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003619 */
3620void kswapd_stop(int nid)
3621{
3622 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3623
Jiang Liud8adde12012-07-11 14:01:52 -07003624 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003625 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003626 NODE_DATA(nid)->kswapd = NULL;
3627 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003628}
3629
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630static int __init kswapd_init(void)
3631{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003632 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003633
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003635 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003636 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 hotcpu_notifier(cpu_callback, 0);
3638 return 0;
3639}
3640
3641module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003642
3643#ifdef CONFIG_NUMA
3644/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003645 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003646 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003647 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003648 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003649 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003650int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003651
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003652#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003653#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003654#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003655#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003656
Christoph Lameter9eeff232006-01-18 17:42:31 -08003657/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003658 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003659 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3660 * a zone.
3661 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003662#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003663
Christoph Lameter9eeff232006-01-18 17:42:31 -08003664/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003665 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003666 * occur.
3667 */
3668int sysctl_min_unmapped_ratio = 1;
3669
3670/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003671 * If the number of slab pages in a zone grows beyond this percentage then
3672 * slab reclaim needs to occur.
3673 */
3674int sysctl_min_slab_ratio = 5;
3675
Mel Gorman11fb9982016-07-28 15:46:20 -07003676static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003677{
Mel Gorman11fb9982016-07-28 15:46:20 -07003678 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3679 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3680 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003681
3682 /*
3683 * It's possible for there to be more file mapped pages than
3684 * accounted for by the pages on the file LRU lists because
3685 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3686 */
3687 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3688}
3689
3690/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003691static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003692{
Alexandru Moised031a152015-11-05 18:48:08 -08003693 unsigned long nr_pagecache_reclaimable;
3694 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003695
3696 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003697 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003698 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003699 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003700 * a better estimate
3701 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003702 if (node_reclaim_mode & RECLAIM_UNMAP)
3703 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003704 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003705 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003706
3707 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003708 if (!(node_reclaim_mode & RECLAIM_WRITE))
3709 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003710
3711 /* Watch for any possible underflows due to delta */
3712 if (unlikely(delta > nr_pagecache_reclaimable))
3713 delta = nr_pagecache_reclaimable;
3714
3715 return nr_pagecache_reclaimable - delta;
3716}
3717
Christoph Lameter0ff38492006-09-25 23:31:52 -07003718/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003719 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003720 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003721static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003722{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003723 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003724 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003725 struct task_struct *p = current;
3726 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003727 int classzone_idx = gfp_zone(gfp_mask);
Andrew Morton179e9632006-03-22 00:08:18 -08003728 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003729 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003730 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003731 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003732 .priority = NODE_RECLAIM_PRIORITY,
3733 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3734 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003735 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003736 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003737 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003738
Christoph Lameter9eeff232006-01-18 17:42:31 -08003739 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003740 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003741 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003742 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003743 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003744 */
3745 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003746 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003747 reclaim_state.reclaimed_slab = 0;
3748 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003749
Mel Gormana5f5f912016-07-28 15:46:32 -07003750 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003751 /*
3752 * Free memory by calling shrink zone with increasing
3753 * priorities until we have enough memory freed.
3754 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003755 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003756 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003757 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003758 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003759
Christoph Lameter9eeff232006-01-18 17:42:31 -08003760 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003761 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003762 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003763 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003764}
Andrew Morton179e9632006-03-22 00:08:18 -08003765
Mel Gormana5f5f912016-07-28 15:46:32 -07003766int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003767{
David Rientjesd773ed62007-10-16 23:26:01 -07003768 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003769
3770 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003771 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003772 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003773 *
Christoph Lameter96146342006-07-03 00:24:13 -07003774 * A small portion of unmapped file backed pages is needed for
3775 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003776 * thrown out if the node is overallocated. So we do not reclaim
3777 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003778 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003779 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003780 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3781 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3782 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003783
Mel Gormana5f5f912016-07-28 15:46:32 -07003784 if (!pgdat_reclaimable(pgdat))
3785 return NODE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003786
Andrew Morton179e9632006-03-22 00:08:18 -08003787 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003788 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003789 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003790 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003791 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003792
3793 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003794 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003795 * have associated processors. This will favor the local processor
3796 * over remote processors and spread off node memory allocations
3797 * as wide as possible.
3798 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003799 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3800 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003801
Mel Gormana5f5f912016-07-28 15:46:32 -07003802 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3803 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003804
Mel Gormana5f5f912016-07-28 15:46:32 -07003805 ret = __node_reclaim(pgdat, gfp_mask, order);
3806 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003807
Mel Gorman24cf725182009-06-16 15:33:23 -07003808 if (!ret)
3809 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3810
David Rientjesd773ed62007-10-16 23:26:01 -07003811 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003812}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003813#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003814
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003815/*
3816 * page_evictable - test whether a page is evictable
3817 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003818 *
3819 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003820 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003821 *
3822 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003823 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003824 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003825 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003826 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003827int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003828{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003829 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003830}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003831
Hugh Dickins850465792012-01-20 14:34:19 -08003832#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003833/**
Hugh Dickins24513262012-01-20 14:34:21 -08003834 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3835 * @pages: array of pages to check
3836 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003837 *
Hugh Dickins24513262012-01-20 14:34:21 -08003838 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003839 *
3840 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003841 */
Hugh Dickins24513262012-01-20 14:34:21 -08003842void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003843{
Johannes Weiner925b7672012-01-12 17:18:15 -08003844 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003845 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003846 int pgscanned = 0;
3847 int pgrescued = 0;
3848 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003849
Hugh Dickins24513262012-01-20 14:34:21 -08003850 for (i = 0; i < nr_pages; i++) {
3851 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003852 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003853
Hugh Dickins24513262012-01-20 14:34:21 -08003854 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003855 if (pagepgdat != pgdat) {
3856 if (pgdat)
3857 spin_unlock_irq(&pgdat->lru_lock);
3858 pgdat = pagepgdat;
3859 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003860 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003861 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003862
Hugh Dickins24513262012-01-20 14:34:21 -08003863 if (!PageLRU(page) || !PageUnevictable(page))
3864 continue;
3865
Hugh Dickins39b5f292012-10-08 16:33:18 -07003866 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003867 enum lru_list lru = page_lru_base_type(page);
3868
Sasha Levin309381fea2014-01-23 15:52:54 -08003869 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003870 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003871 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3872 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003873 pgrescued++;
3874 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003875 }
Hugh Dickins24513262012-01-20 14:34:21 -08003876
Mel Gorman785b99f2016-07-28 15:47:23 -07003877 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003878 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3879 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003880 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003881 }
Hugh Dickins850465792012-01-20 14:34:19 -08003882}
3883#endif /* CONFIG_SHMEM */