blob: 2f45c0520f43fbcb5696063172cd281553302f9d [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>
Ingo Molnar5b3cc152017-02-02 20:43:54 +010017#include <linux/sched/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/kernel_stat.h>
21#include <linux/swap.h>
22#include <linux/pagemap.h>
23#include <linux/init.h>
24#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070025#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070026#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070027#include <linux/file.h>
28#include <linux/writeback.h>
29#include <linux/blkdev.h>
30#include <linux/buffer_head.h> /* for try_to_release_page(),
31 buffer_heads_over_limit */
32#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033#include <linux/backing-dev.h>
34#include <linux/rmap.h>
35#include <linux/topology.h>
36#include <linux/cpu.h>
37#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080038#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/notifier.h>
40#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080041#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070042#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080043#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080044#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070045#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070046#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070047#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070048#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070049#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080050#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051
52#include <asm/tlbflush.h>
53#include <asm/div64.h>
54
55#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070056#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
Nick Piggin0f8053a2006-03-22 00:08:33 -080058#include "internal.h"
59
Mel Gorman33906bc2010-08-09 17:19:16 -070060#define CREATE_TRACE_POINTS
61#include <trace/events/vmscan.h>
62
Linus Torvalds1da177e2005-04-16 15:20:36 -070063struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080064 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim;
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040068 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
Johannes Weineree814fe2014-08-06 16:06:19 -070070 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070071 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080072
Johannes Weineree814fe2014-08-06 16:06:19 -070073 /*
74 * Nodemask of nodes allowed by the caller. If NULL, all nodes
75 * are scanned.
76 */
77 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070078
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070079 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080080 * The memory cgroup that hit its limit and as a result is the
81 * primary target of this reclaim invocation.
82 */
83 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080084
Johannes Weineree814fe2014-08-06 16:06:19 -070085 /* Scan (total_size >> priority) pages at once */
86 int priority;
87
Mel Gormanb2e18752016-07-28 15:45:37 -070088 /* The highest zone to isolate pages for reclaim from */
89 enum zone_type reclaim_idx;
90
Johannes Weiner1276ad62017-02-24 14:56:11 -080091 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -070092 unsigned int may_writepage:1;
93
94 /* Can mapped pages be reclaimed? */
95 unsigned int may_unmap:1;
96
97 /* Can pages be swapped as part of reclaim? */
98 unsigned int may_swap:1;
99
Yisheng Xied6622f62017-05-03 14:53:57 -0700100 /*
101 * Cgroups are not reclaimed below their configured memory.low,
102 * unless we threaten to OOM. If any cgroups are skipped due to
103 * memory.low and nothing was reclaimed, go back for memory.low.
104 */
105 unsigned int memcg_low_reclaim:1;
106 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -0800107
Johannes Weineree814fe2014-08-06 16:06:19 -0700108 unsigned int hibernation_mode:1;
109
110 /* One of the zones is ready for compaction */
111 unsigned int compaction_ready:1;
112
113 /* Incremented by the number of inactive pages that were scanned */
114 unsigned long nr_scanned;
115
116 /* Number of pages freed so far during a call to shrink_zones() */
117 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118};
119
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120#ifdef ARCH_HAS_PREFETCH
121#define prefetch_prev_lru_page(_page, _base, _field) \
122 do { \
123 if ((_page)->lru.prev != _base) { \
124 struct page *prev; \
125 \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
128 } \
129 } while (0)
130#else
131#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132#endif
133
134#ifdef ARCH_HAS_PREFETCHW
135#define prefetchw_prev_lru_page(_page, _base, _field) \
136 do { \
137 if ((_page)->lru.prev != _base) { \
138 struct page *prev; \
139 \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
142 } \
143 } while (0)
144#else
145#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146#endif
147
148/*
149 * From 0 .. 100. Higher means more swappy.
150 */
151int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700152/*
153 * The total number of pages which are beyond the high watermark within all
154 * zones.
155 */
156unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
158static LIST_HEAD(shrinker_list);
159static DECLARE_RWSEM(shrinker_rwsem);
160
Andrew Mortonc255a452012-07-31 16:43:02 -0700161#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800162static bool global_reclaim(struct scan_control *sc)
163{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800164 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800165}
Tejun Heo97c93412015-05-22 18:23:36 -0400166
167/**
168 * sane_reclaim - is the usual dirty throttling mechanism operational?
169 * @sc: scan_control in question
170 *
171 * The normal page dirty throttling mechanism in balance_dirty_pages() is
172 * completely broken with the legacy memcg and direct stalling in
173 * shrink_page_list() is used for throttling instead, which lacks all the
174 * niceties such as fairness, adaptive pausing, bandwidth proportional
175 * allocation and configurability.
176 *
177 * This function tests whether the vmscan currently in progress can assume
178 * that the normal dirty throttling mechanism is operational.
179 */
180static bool sane_reclaim(struct scan_control *sc)
181{
182 struct mem_cgroup *memcg = sc->target_mem_cgroup;
183
184 if (!memcg)
185 return true;
186#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400188 return true;
189#endif
190 return false;
191}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800192#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800193static bool global_reclaim(struct scan_control *sc)
194{
195 return true;
196}
Tejun Heo97c93412015-05-22 18:23:36 -0400197
198static bool sane_reclaim(struct scan_control *sc)
199{
200 return true;
201}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800202#endif
203
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700204/*
205 * This misses isolated pages which are not accounted for to save counters.
206 * As the data only determines if reclaim or compaction continues, it is
207 * not expected that isolated pages will be a dominating factor.
208 */
209unsigned long zone_reclaimable_pages(struct zone *zone)
210{
211 unsigned long nr;
212
213 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
214 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
215 if (get_nr_swap_pages() > 0)
216 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
217 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
218
219 return nr;
220}
221
Mel Gorman599d0c92016-07-28 15:45:31 -0700222unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
Lisa Du6e543d52013-09-11 14:22:36 -0700223{
Mel Gorman599d0c92016-07-28 15:45:31 -0700224 unsigned long nr;
225
226 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
227 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
228 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
229
230 if (get_nr_swap_pages() > 0)
231 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
232 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
233 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
234
235 return nr;
236}
237
Michal Hockofd538802017-02-22 15:45:58 -0800238/**
239 * lruvec_lru_size - Returns the number of pages on the given LRU list.
240 * @lruvec: lru vector
241 * @lru: lru to use
242 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
243 */
244unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800245{
Michal Hockofd538802017-02-22 15:45:58 -0800246 unsigned long lru_size;
247 int zid;
248
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700249 if (!mem_cgroup_disabled())
Michal Hockofd538802017-02-22 15:45:58 -0800250 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
251 else
252 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800253
Michal Hockofd538802017-02-22 15:45:58 -0800254 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
255 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
256 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800257
Michal Hockofd538802017-02-22 15:45:58 -0800258 if (!managed_zone(zone))
259 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800260
Michal Hockofd538802017-02-22 15:45:58 -0800261 if (!mem_cgroup_disabled())
262 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
263 else
264 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
265 NR_ZONE_LRU_BASE + lru);
266 lru_size -= min(size, lru_size);
267 }
268
269 return lru_size;
270
Michal Hockob4536f0c82017-01-10 16:58:04 -0800271}
272
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000274 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000276int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000278 size_t size = sizeof(*shrinker->nr_deferred);
279
Glauber Costa1d3d4432013-08-28 10:18:04 +1000280 if (shrinker->flags & SHRINKER_NUMA_AWARE)
281 size *= nr_node_ids;
282
283 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
284 if (!shrinker->nr_deferred)
285 return -ENOMEM;
286
Rusty Russell8e1f9362007-07-17 04:03:17 -0700287 down_write(&shrinker_rwsem);
288 list_add_tail(&shrinker->list, &shrinker_list);
289 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000290 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700292EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293
294/*
295 * Remove one
296 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700297void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
299 down_write(&shrinker_rwsem);
300 list_del(&shrinker->list);
301 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700302 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700304EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305
306#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000307
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800308static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
309 struct shrinker *shrinker,
310 unsigned long nr_scanned,
311 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000313 unsigned long freed = 0;
314 unsigned long long delta;
315 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700316 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000317 long nr;
318 long new_nr;
319 int nid = shrinkctl->nid;
320 long batch_size = shrinker->batch ? shrinker->batch
321 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800322 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000323
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700324 freeable = shrinker->count_objects(shrinker, shrinkctl);
325 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000326 return 0;
327
328 /*
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
332 */
333 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
334
335 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800336 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700337 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800338 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000339 total_scan += delta;
340 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000342 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700343 total_scan = freeable;
Shaohua Li5f33a082016-12-12 16:41:50 -0800344 next_deferred = nr;
345 } else
346 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000347
348 /*
349 * We need to avoid excessive windup on filesystem shrinkers
350 * due to large numbers of GFP_NOFS allocations causing the
351 * shrinkers to return -1 all the time. This results in a large
352 * nr being built up so when a shrink that can do some work
353 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700354 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000355 * memory.
356 *
357 * Hence only allow the shrinker to scan the entire cache when
358 * a large delta change is calculated directly.
359 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700360 if (delta < freeable / 4)
361 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000362
363 /*
364 * Avoid risking looping forever due to too large nr value:
365 * never try to free more than twice the estimate number of
366 * freeable entries.
367 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700368 if (total_scan > freeable * 2)
369 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000370
371 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800372 nr_scanned, nr_eligible,
373 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000374
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800375 /*
376 * Normally, we should not scan less than batch_size objects in one
377 * pass to avoid too frequent shrinker calls, but if the slab has less
378 * than batch_size objects in total and we are really tight on memory,
379 * we will try to reclaim all available objects, otherwise we can end
380 * up failing allocations although there are plenty of reclaimable
381 * objects spread over several slabs with usage less than the
382 * batch_size.
383 *
384 * We detect the "tight on memory" situations by looking at the total
385 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700386 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800387 * scanning at high prio and therefore should try to reclaim as much as
388 * possible.
389 */
390 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700391 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000392 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800393 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000394
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800395 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000396 ret = shrinker->scan_objects(shrinker, shrinkctl);
397 if (ret == SHRINK_STOP)
398 break;
399 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000400
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800401 count_vm_events(SLABS_SCANNED, nr_to_scan);
402 total_scan -= nr_to_scan;
Shaohua Li5f33a082016-12-12 16:41:50 -0800403 scanned += nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000404
405 cond_resched();
406 }
407
Shaohua Li5f33a082016-12-12 16:41:50 -0800408 if (next_deferred >= scanned)
409 next_deferred -= scanned;
410 else
411 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000412 /*
413 * move the unused scan count back into the shrinker in a
414 * manner that handles concurrent updates. If we exhausted the
415 * scan, there is no need to do an update.
416 */
Shaohua Li5f33a082016-12-12 16:41:50 -0800417 if (next_deferred > 0)
418 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000419 &shrinker->nr_deferred[nid]);
420 else
421 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
422
Dave Hansendf9024a2014-06-04 16:08:07 -0700423 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000424 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425}
426
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800427/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800428 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800429 * @gfp_mask: allocation context
430 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800431 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800432 * @nr_scanned: pressure numerator
433 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800435 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800437 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
438 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800440 * @memcg specifies the memory cgroup to target. If it is not NULL,
441 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700442 * objects from the memory cgroup specified. Otherwise, only unaware
443 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800444 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800445 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
446 * the available objects should be scanned. Page reclaim for example
447 * passes the number of pages scanned and the number of pages on the
448 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
449 * when it encountered mapped pages. The ratio is further biased by
450 * the ->seeks setting of the shrink function, which indicates the
451 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800453 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800455static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
456 struct mem_cgroup *memcg,
457 unsigned long nr_scanned,
458 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459{
460 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000461 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700463 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800464 return 0;
465
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800466 if (nr_scanned == 0)
467 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468
Minchan Kimf06590b2011-05-24 17:11:11 -0700469 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000470 /*
471 * If we would return 0, our callers would understand that we
472 * have nothing else to shrink and give up trying. By returning
473 * 1 we keep it going and assume we'll be able to shrink next
474 * time.
475 */
476 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700477 goto out;
478 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
480 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800481 struct shrink_control sc = {
482 .gfp_mask = gfp_mask,
483 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800484 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800485 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800486
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700487 /*
488 * If kernel memory accounting is disabled, we ignore
489 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
490 * passing NULL for memcg.
491 */
492 if (memcg_kmem_enabled() &&
493 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800494 continue;
495
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800496 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
497 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000498
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800499 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800501
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700503out:
504 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000505 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506}
507
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800508void drop_slab_node(int nid)
509{
510 unsigned long freed;
511
512 do {
513 struct mem_cgroup *memcg = NULL;
514
515 freed = 0;
516 do {
517 freed += shrink_slab(GFP_KERNEL, nid, memcg,
518 1000, 1000);
519 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
520 } while (freed > 10);
521}
522
523void drop_slab(void)
524{
525 int nid;
526
527 for_each_online_node(nid)
528 drop_slab_node(nid);
529}
530
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531static inline int is_page_cache_freeable(struct page *page)
532{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700533 /*
534 * A freeable page cache page is referenced only by the caller
535 * that isolated the page, the page cache radix tree and
536 * optional buffer heads at page->private.
537 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700538 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539}
540
Tejun Heo703c2702015-05-22 17:13:44 -0400541static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542{
Christoph Lameter930d9152006-01-08 01:00:47 -0800543 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400545 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400547 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 return 1;
549 return 0;
550}
551
552/*
553 * We detected a synchronous write error writing a page out. Probably
554 * -ENOSPC. We need to propagate that into the address_space for a subsequent
555 * fsync(), msync() or close().
556 *
557 * The tricky part is that after writepage we cannot touch the mapping: nothing
558 * prevents it from being freed up. But we have a ref on the page and once
559 * that page is locked, the mapping is pinned.
560 *
561 * We're allowed to run sleeping lock_page() here because we know the caller has
562 * __GFP_FS.
563 */
564static void handle_write_error(struct address_space *mapping,
565 struct page *page, int error)
566{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100567 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700568 if (page_mapping(page) == mapping)
569 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 unlock_page(page);
571}
572
Christoph Lameter04e62a22006-06-23 02:03:38 -0700573/* possible outcome of pageout() */
574typedef enum {
575 /* failed to write page out, page is locked */
576 PAGE_KEEP,
577 /* move page to the active list, page is locked */
578 PAGE_ACTIVATE,
579 /* page has been sent to the disk successfully, page is unlocked */
580 PAGE_SUCCESS,
581 /* page is clean and locked */
582 PAGE_CLEAN,
583} pageout_t;
584
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585/*
Andrew Morton1742f192006-03-22 00:08:21 -0800586 * pageout is called by shrink_page_list() for each dirty page.
587 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700589static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700590 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591{
592 /*
593 * If the page is dirty, only perform writeback if that write
594 * will be non-blocking. To prevent this allocation from being
595 * stalled by pagecache activity. But note that there may be
596 * stalls if we need to run get_block(). We could test
597 * PagePrivate for that.
598 *
Al Viro81742022014-04-03 03:17:43 -0400599 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 * this page's queue, we can perform writeback even if that
601 * will block.
602 *
603 * If the page is swapcache, write it back even if that would
604 * block, for some throttling. This happens by accident, because
605 * swap_backing_dev_info is bust: it doesn't reflect the
606 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 */
608 if (!is_page_cache_freeable(page))
609 return PAGE_KEEP;
610 if (!mapping) {
611 /*
612 * Some data journaling orphaned pages can have
613 * page->mapping == NULL while being dirty with clean buffers.
614 */
David Howells266cf652009-04-03 16:42:36 +0100615 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 if (try_to_free_buffers(page)) {
617 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700618 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 return PAGE_CLEAN;
620 }
621 }
622 return PAGE_KEEP;
623 }
624 if (mapping->a_ops->writepage == NULL)
625 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400626 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 return PAGE_KEEP;
628
629 if (clear_page_dirty_for_io(page)) {
630 int res;
631 struct writeback_control wbc = {
632 .sync_mode = WB_SYNC_NONE,
633 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700634 .range_start = 0,
635 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 .for_reclaim = 1,
637 };
638
639 SetPageReclaim(page);
640 res = mapping->a_ops->writepage(page, &wbc);
641 if (res < 0)
642 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800643 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 ClearPageReclaim(page);
645 return PAGE_ACTIVATE;
646 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700647
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 if (!PageWriteback(page)) {
649 /* synchronous write or broken a_ops? */
650 ClearPageReclaim(page);
651 }
yalin wang3aa23852016-01-14 15:18:30 -0800652 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700653 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 return PAGE_SUCCESS;
655 }
656
657 return PAGE_CLEAN;
658}
659
Andrew Mortona649fd92006-10-17 00:09:36 -0700660/*
Nick Piggine2867812008-07-25 19:45:30 -0700661 * Same as remove_mapping, but if the page is removed from the mapping, it
662 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700663 */
Johannes Weinera5289102014-04-03 14:47:51 -0700664static int __remove_mapping(struct address_space *mapping, struct page *page,
665 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800666{
Greg Thelenc4843a72015-05-22 17:13:16 -0400667 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400668
Nick Piggin28e4d962006-09-25 23:31:23 -0700669 BUG_ON(!PageLocked(page));
670 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800671
Greg Thelenc4843a72015-05-22 17:13:16 -0400672 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800673 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700674 * The non racy check for a busy page.
675 *
676 * Must be careful with the order of the tests. When someone has
677 * a ref to the page, it may be possible that they dirty it then
678 * drop the reference. So if PageDirty is tested before page_count
679 * here, then the following race may occur:
680 *
681 * get_user_pages(&page);
682 * [user mapping goes away]
683 * write_to(page);
684 * !PageDirty(page) [good]
685 * SetPageDirty(page);
686 * put_page(page);
687 * !page_count(page) [good, discard it]
688 *
689 * [oops, our write_to data is lost]
690 *
691 * Reversing the order of the tests ensures such a situation cannot
692 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700693 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700694 *
695 * Note that if SetPageDirty is always performed via set_page_dirty,
696 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800697 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700698 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800699 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700700 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
701 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700702 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800703 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700704 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800705
706 if (PageSwapCache(page)) {
707 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700708 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800709 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700711 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700712 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500713 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700714 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500715
716 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700717 /*
718 * Remember a shadow entry for reclaimed file cache in
719 * order to detect refaults, thus thrashing, later on.
720 *
721 * But don't store shadows in an address space that is
722 * already exiting. This is not just an optizimation,
723 * inode reclaim needs to empty out the radix tree or
724 * the nodes are lost. Don't plant shadows behind its
725 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800726 *
727 * We also don't store shadows for DAX mappings because the
728 * only page cache pages found in these are zero pages
729 * covering holes, and because we don't want to mix DAX
730 * exceptional entries and shadow exceptional entries in the
731 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700732 */
733 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800734 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700735 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700736 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400737 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500738
739 if (freepage != NULL)
740 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800741 }
742
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800743 return 1;
744
745cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400746 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800747 return 0;
748}
749
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750/*
Nick Piggine2867812008-07-25 19:45:30 -0700751 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
752 * someone else has a ref on the page, abort and return 0. If it was
753 * successfully detached, return 1. Assumes the caller has a single ref on
754 * this page.
755 */
756int remove_mapping(struct address_space *mapping, struct page *page)
757{
Johannes Weinera5289102014-04-03 14:47:51 -0700758 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700759 /*
760 * Unfreezing the refcount with 1 rather than 2 effectively
761 * drops the pagecache ref for us without requiring another
762 * atomic operation.
763 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700764 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700765 return 1;
766 }
767 return 0;
768}
769
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700770/**
771 * putback_lru_page - put previously isolated page onto appropriate LRU list
772 * @page: page to be put back to appropriate lru list
773 *
774 * Add previously isolated @page to appropriate LRU list.
775 * Page may still be unevictable for other reasons.
776 *
777 * lru_lock must not be held, interrupts must be enabled.
778 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700779void putback_lru_page(struct page *page)
780{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700781 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700782 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700783
Sasha Levin309381fea2014-01-23 15:52:54 -0800784 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700785
786redo:
787 ClearPageUnevictable(page);
788
Hugh Dickins39b5f292012-10-08 16:33:18 -0700789 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700790 /*
791 * For evictable pages, we can use the cache.
792 * In event of a race, worst case is we end up with an
793 * unevictable page on [in]active list.
794 * We know how to handle that.
795 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700796 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700797 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700798 } else {
799 /*
800 * Put unevictable pages directly on zone's unevictable
801 * list.
802 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700803 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700804 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700805 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700806 * When racing with an mlock or AS_UNEVICTABLE clearing
807 * (page is unlocked) make sure that if the other thread
808 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800809 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700810 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700811 * the page back to the evictable list.
812 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700813 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700814 */
815 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700816 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700817
818 /*
819 * page's status can change while we move it among lru. If an evictable
820 * page is on unevictable list, it never be freed. To avoid that,
821 * check after we added it to the list, again.
822 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700823 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700824 if (!isolate_lru_page(page)) {
825 put_page(page);
826 goto redo;
827 }
828 /* This means someone else dropped this page from LRU
829 * So, it will be freed or putback to LRU again. There is
830 * nothing to do here.
831 */
832 }
833
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700834 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700835 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700836 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700837 count_vm_event(UNEVICTABLE_PGCULLED);
838
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700839 put_page(page); /* drop ref from isolate */
840}
841
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800842enum page_references {
843 PAGEREF_RECLAIM,
844 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800845 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800846 PAGEREF_ACTIVATE,
847};
848
849static enum page_references page_check_references(struct page *page,
850 struct scan_control *sc)
851{
Johannes Weiner645747462010-03-05 13:42:22 -0800852 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800853 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800854
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700855 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
856 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800857 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800858
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800859 /*
860 * Mlock lost the isolation race with us. Let try_to_unmap()
861 * move the page to the unevictable list.
862 */
863 if (vm_flags & VM_LOCKED)
864 return PAGEREF_RECLAIM;
865
Johannes Weiner645747462010-03-05 13:42:22 -0800866 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700867 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800868 return PAGEREF_ACTIVATE;
869 /*
870 * All mapped pages start out with page table
871 * references from the instantiating fault, so we need
872 * to look twice if a mapped file page is used more
873 * than once.
874 *
875 * Mark it and spare it for another trip around the
876 * inactive list. Another page table reference will
877 * lead to its activation.
878 *
879 * Note: the mark is set for activated pages as well
880 * so that recently deactivated but used pages are
881 * quickly recovered.
882 */
883 SetPageReferenced(page);
884
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800885 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800886 return PAGEREF_ACTIVATE;
887
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800888 /*
889 * Activate file-backed executable pages after first usage.
890 */
891 if (vm_flags & VM_EXEC)
892 return PAGEREF_ACTIVATE;
893
Johannes Weiner645747462010-03-05 13:42:22 -0800894 return PAGEREF_KEEP;
895 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800896
897 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700898 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800899 return PAGEREF_RECLAIM_CLEAN;
900
901 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800902}
903
Mel Gormane2be15f2013-07-03 15:01:57 -0700904/* Check if a page is dirty or under writeback */
905static void page_check_dirty_writeback(struct page *page,
906 bool *dirty, bool *writeback)
907{
Mel Gormanb4597222013-07-03 15:02:05 -0700908 struct address_space *mapping;
909
Mel Gormane2be15f2013-07-03 15:01:57 -0700910 /*
911 * Anonymous pages are not handled by flushers and must be written
912 * from reclaim context. Do not stall reclaim based on them
913 */
Shaohua Li802a3a92017-05-03 14:52:32 -0700914 if (!page_is_file_cache(page) ||
915 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -0700916 *dirty = false;
917 *writeback = false;
918 return;
919 }
920
921 /* By default assume that the page flags are accurate */
922 *dirty = PageDirty(page);
923 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700924
925 /* Verify dirty/writeback state if the filesystem supports it */
926 if (!page_has_private(page))
927 return;
928
929 mapping = page_mapping(page);
930 if (mapping && mapping->a_ops->is_dirty_writeback)
931 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700932}
933
Michal Hocko3c710c12017-02-22 15:44:27 -0800934struct reclaim_stat {
935 unsigned nr_dirty;
936 unsigned nr_unqueued_dirty;
937 unsigned nr_congested;
938 unsigned nr_writeback;
939 unsigned nr_immediate;
Michal Hocko5bccd162017-02-22 15:44:30 -0800940 unsigned nr_activate;
941 unsigned nr_ref_keep;
942 unsigned nr_unmap_fail;
Michal Hocko3c710c12017-02-22 15:44:27 -0800943};
944
Nick Piggine2867812008-07-25 19:45:30 -0700945/*
Andrew Morton1742f192006-03-22 00:08:21 -0800946 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 */
Andrew Morton1742f192006-03-22 00:08:21 -0800948static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700949 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700950 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700951 enum ttu_flags ttu_flags,
Michal Hocko3c710c12017-02-22 15:44:27 -0800952 struct reclaim_stat *stat,
Minchan Kim02c6de82012-10-08 16:31:55 -0700953 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954{
955 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700956 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 int pgactivate = 0;
Michal Hocko3c710c12017-02-22 15:44:27 -0800958 unsigned nr_unqueued_dirty = 0;
959 unsigned nr_dirty = 0;
960 unsigned nr_congested = 0;
961 unsigned nr_reclaimed = 0;
962 unsigned nr_writeback = 0;
963 unsigned nr_immediate = 0;
Michal Hocko5bccd162017-02-22 15:44:30 -0800964 unsigned nr_ref_keep = 0;
965 unsigned nr_unmap_fail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966
967 cond_resched();
968
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 while (!list_empty(page_list)) {
970 struct address_space *mapping;
971 struct page *page;
972 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700973 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700974 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
976 cond_resched();
977
978 page = lru_to_page(page_list);
979 list_del(&page->lru);
980
Nick Piggin529ae9a2008-08-02 12:01:03 +0200981 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 goto keep;
983
Sasha Levin309381fea2014-01-23 15:52:54 -0800984 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985
986 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800987
Hugh Dickins39b5f292012-10-08 16:33:18 -0700988 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -0700989 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700990
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700991 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800992 goto keep_locked;
993
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 /* Double the slab pressure for mapped and swapcache pages */
Shaohua Li802a3a92017-05-03 14:52:32 -0700995 if ((page_mapped(page) || PageSwapCache(page)) &&
996 !(PageAnon(page) && !PageSwapBacked(page)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 sc->nr_scanned++;
998
Andy Whitcroftc661b072007-08-22 14:01:26 -0700999 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1000 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1001
Mel Gorman283aba92013-07-03 15:01:51 -07001002 /*
Mel Gormane2be15f2013-07-03 15:01:57 -07001003 * The number of dirty pages determines if a zone is marked
1004 * reclaim_congested which affects wait_iff_congested. kswapd
1005 * will stall and start writing pages if the tail of the LRU
1006 * is all dirty unqueued pages.
1007 */
1008 page_check_dirty_writeback(page, &dirty, &writeback);
1009 if (dirty || writeback)
1010 nr_dirty++;
1011
1012 if (dirty && !writeback)
1013 nr_unqueued_dirty++;
1014
Mel Gormand04e8ac2013-07-03 15:02:03 -07001015 /*
1016 * Treat this page as congested if the underlying BDI is or if
1017 * pages are cycling through the LRU so quickly that the
1018 * pages marked for immediate reclaim are making it to the
1019 * end of the LRU a second time.
1020 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001021 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001022 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001023 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001024 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001025 nr_congested++;
1026
1027 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001028 * If a page at the tail of the LRU is under writeback, there
1029 * are three cases to consider.
1030 *
1031 * 1) If reclaim is encountering an excessive number of pages
1032 * under writeback and this page is both under writeback and
1033 * PageReclaim then it indicates that pages are being queued
1034 * for IO but are being recycled through the LRU before the
1035 * IO can complete. Waiting on the page itself risks an
1036 * indefinite stall if it is impossible to writeback the
1037 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001038 * note that the LRU is being scanned too quickly and the
1039 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001040 *
Tejun Heo97c93412015-05-22 18:23:36 -04001041 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001042 * not marked for immediate reclaim, or the caller does not
1043 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1044 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001045 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001046 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001047 * Require may_enter_fs because we would wait on fs, which
1048 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001049 * enter reclaim, and deadlock if it waits on a page for
1050 * which it is needed to do the write (loop masks off
1051 * __GFP_IO|__GFP_FS for this reason); but more thought
1052 * would probably show more reasons.
1053 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001054 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001055 * PageReclaim. memcg does not have any dirty pages
1056 * throttling so we could easily OOM just because too many
1057 * pages are in writeback and there is nothing else to
1058 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001059 *
1060 * In cases 1) and 2) we activate the pages to get them out of
1061 * the way while we continue scanning for clean pages on the
1062 * inactive list and refilling from the active list. The
1063 * observation here is that waiting for disk writes is more
1064 * expensive than potentially causing reloads down the line.
1065 * Since they're marked for immediate reclaim, they won't put
1066 * memory pressure on the cache working set any longer than it
1067 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001068 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001069 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001070 /* Case 1 above */
1071 if (current_is_kswapd() &&
1072 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001073 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001074 nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001075 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001076
1077 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001078 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001079 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001080 /*
1081 * This is slightly racy - end_page_writeback()
1082 * might have just cleared PageReclaim, then
1083 * setting PageReclaim here end up interpreted
1084 * as PageReadahead - but that does not matter
1085 * enough to care. What we do want is for this
1086 * page to have PageReclaim set next time memcg
1087 * reclaim reaches the tests above, so it will
1088 * then wait_on_page_writeback() to avoid OOM;
1089 * and it's also appropriate in global reclaim.
1090 */
1091 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001092 nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001093 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001094
1095 /* Case 3 above */
1096 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001097 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001098 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001099 /* then go back and try same page again */
1100 list_add_tail(&page->lru, page_list);
1101 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001102 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001103 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104
Minchan Kim02c6de82012-10-08 16:31:55 -07001105 if (!force_reclaim)
1106 references = page_check_references(page, sc);
1107
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001108 switch (references) {
1109 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001111 case PAGEREF_KEEP:
Michal Hocko5bccd162017-02-22 15:44:30 -08001112 nr_ref_keep++;
Johannes Weiner645747462010-03-05 13:42:22 -08001113 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001114 case PAGEREF_RECLAIM:
1115 case PAGEREF_RECLAIM_CLEAN:
1116 ; /* try to reclaim the page below */
1117 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 /*
1120 * Anonymous process memory has backing store?
1121 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001122 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001124 if (PageAnon(page) && PageSwapBacked(page) &&
1125 !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001126 if (!(sc->gfp_mask & __GFP_IO))
1127 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001128 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001130 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
Mel Gormane2be15f2013-07-03 15:01:57 -07001132 /* Adding to swap updated mapping */
1133 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001134 } else if (unlikely(PageTransHuge(page))) {
1135 /* Split file THP */
1136 if (split_huge_page_to_list(page, page_list))
1137 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001138 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001140 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1141
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 /*
1143 * The page is mapped into the page tables of one or more
1144 * processes. Try to unmap it here.
1145 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001146 if (page_mapped(page)) {
Minchan Kim666e5a42017-05-03 14:54:20 -07001147 if (!try_to_unmap(page, ttu_flags | TTU_BATCH_FLUSH)) {
Michal Hocko5bccd162017-02-22 15:44:30 -08001148 nr_unmap_fail++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 }
1151 }
1152
1153 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001154 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001155 * Only kswapd can writeback filesystem pages
1156 * to avoid risk of stack overflow. But avoid
1157 * injecting inefficient single-page IO into
1158 * flusher writeback as much as possible: only
1159 * write pages when we've encountered many
1160 * dirty pages, and when we've already scanned
1161 * the rest of the LRU for clean pages and see
1162 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001163 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001164 if (page_is_file_cache(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001165 (!current_is_kswapd() || !PageReclaim(page) ||
1166 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001167 /*
1168 * Immediately reclaim when written back.
1169 * Similar in principal to deactivate_page()
1170 * except we already have the page isolated
1171 * and know it's dirty
1172 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001173 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001174 SetPageReclaim(page);
1175
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001176 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001177 }
1178
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001179 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001181 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001183 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 goto keep_locked;
1185
Mel Gormand950c942015-09-04 15:47:35 -07001186 /*
1187 * Page is dirty. Flush the TLB if a writable entry
1188 * potentially exists to avoid CPU writes after IO
1189 * starts and then write it out here.
1190 */
1191 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001192 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 case PAGE_KEEP:
1194 goto keep_locked;
1195 case PAGE_ACTIVATE:
1196 goto activate_locked;
1197 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001198 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001199 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001200 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001202
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 /*
1204 * A synchronous write - probably a ramdisk. Go
1205 * ahead and try to reclaim the page.
1206 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001207 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 goto keep;
1209 if (PageDirty(page) || PageWriteback(page))
1210 goto keep_locked;
1211 mapping = page_mapping(page);
1212 case PAGE_CLEAN:
1213 ; /* try to free the page below */
1214 }
1215 }
1216
1217 /*
1218 * If the page has buffers, try to free the buffer mappings
1219 * associated with this page. If we succeed we try to free
1220 * the page as well.
1221 *
1222 * We do this even if the page is PageDirty().
1223 * try_to_release_page() does not perform I/O, but it is
1224 * possible for a page to have PageDirty set, but it is actually
1225 * clean (all its buffers are clean). This happens if the
1226 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001227 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 * try_to_release_page() will discover that cleanness and will
1229 * drop the buffers and mark the page clean - it can be freed.
1230 *
1231 * Rarely, pages can have buffers and no ->mapping. These are
1232 * the pages which were not successfully invalidated in
1233 * truncate_complete_page(). We try to drop those buffers here
1234 * and if that worked, and the page is no longer mapped into
1235 * process address space (page_count == 1) it can be freed.
1236 * Otherwise, leave the page on the LRU so it is swappable.
1237 */
David Howells266cf652009-04-03 16:42:36 +01001238 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 if (!try_to_release_page(page, sc->gfp_mask))
1240 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001241 if (!mapping && page_count(page) == 1) {
1242 unlock_page(page);
1243 if (put_page_testzero(page))
1244 goto free_it;
1245 else {
1246 /*
1247 * rare race with speculative reference.
1248 * the speculative reference will free
1249 * this page shortly, so we may
1250 * increment nr_reclaimed here (and
1251 * leave it off the LRU).
1252 */
1253 nr_reclaimed++;
1254 continue;
1255 }
1256 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 }
1258
Shaohua Li802a3a92017-05-03 14:52:32 -07001259 if (PageAnon(page) && !PageSwapBacked(page)) {
1260 /* follow __remove_mapping for reference */
1261 if (!page_ref_freeze(page, 1))
1262 goto keep_locked;
1263 if (PageDirty(page)) {
1264 page_ref_unfreeze(page, 1);
1265 goto keep_locked;
1266 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267
Shaohua Li802a3a92017-05-03 14:52:32 -07001268 count_vm_event(PGLAZYFREED);
1269 } else if (!mapping || !__remove_mapping(mapping, page, true))
1270 goto keep_locked;
Nick Piggina978d6f2008-10-18 20:26:58 -07001271 /*
1272 * At this point, we have no other references and there is
1273 * no way to pick any more up (removed from LRU, removed
1274 * from pagecache). Can use non-atomic bitops now (and
1275 * we obviously don't have to worry about waking up a process
1276 * waiting on the page lock, because there are no references.
1277 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001278 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001279free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001280 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001281
1282 /*
1283 * Is there need to periodically free_page_list? It would
1284 * appear not as the counts should be low
1285 */
1286 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 continue;
1288
1289activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001290 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001291 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1292 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001293 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001294 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001295 if (!PageMlocked(page)) {
1296 SetPageActive(page);
1297 pgactivate++;
1298 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299keep_locked:
1300 unlock_page(page);
1301keep:
1302 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001303 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001305
Johannes Weiner747db952014-08-08 14:19:24 -07001306 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001307 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001308 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001309
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001311 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001312
Michal Hocko3c710c12017-02-22 15:44:27 -08001313 if (stat) {
1314 stat->nr_dirty = nr_dirty;
1315 stat->nr_congested = nr_congested;
1316 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1317 stat->nr_writeback = nr_writeback;
1318 stat->nr_immediate = nr_immediate;
Michal Hocko5bccd162017-02-22 15:44:30 -08001319 stat->nr_activate = pgactivate;
1320 stat->nr_ref_keep = nr_ref_keep;
1321 stat->nr_unmap_fail = nr_unmap_fail;
Michal Hocko3c710c12017-02-22 15:44:27 -08001322 }
Andrew Morton05ff5132006-03-22 00:08:20 -08001323 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324}
1325
Minchan Kim02c6de82012-10-08 16:31:55 -07001326unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1327 struct list_head *page_list)
1328{
1329 struct scan_control sc = {
1330 .gfp_mask = GFP_KERNEL,
1331 .priority = DEF_PRIORITY,
1332 .may_unmap = 1,
1333 };
Michal Hocko3c710c12017-02-22 15:44:27 -08001334 unsigned long ret;
Minchan Kim02c6de82012-10-08 16:31:55 -07001335 struct page *page, *next;
1336 LIST_HEAD(clean_pages);
1337
1338 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001339 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001340 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001341 ClearPageActive(page);
1342 list_move(&page->lru, &clean_pages);
1343 }
1344 }
1345
Mel Gorman599d0c92016-07-28 15:45:31 -07001346 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Shaohua Lia128ca72017-05-03 14:52:22 -07001347 TTU_IGNORE_ACCESS, NULL, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001348 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001349 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001350 return ret;
1351}
1352
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001353/*
1354 * Attempt to remove the specified page from its LRU. Only take this page
1355 * if it is of the appropriate PageActive status. Pages which are being
1356 * freed elsewhere are also ignored.
1357 *
1358 * page: page to consider
1359 * mode: one of the LRU isolation modes defined above
1360 *
1361 * returns 0 on success, -ve errno on failure.
1362 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001363int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001364{
1365 int ret = -EINVAL;
1366
1367 /* Only take pages on the LRU. */
1368 if (!PageLRU(page))
1369 return ret;
1370
Minchan Kime46a2872012-10-08 16:33:48 -07001371 /* Compaction should not handle unevictable pages but CMA can do so */
1372 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001373 return ret;
1374
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001375 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001376
Mel Gormanc8244932012-01-12 17:19:38 -08001377 /*
1378 * To minimise LRU disruption, the caller can indicate that it only
1379 * wants to isolate pages it will be able to operate on without
1380 * blocking - clean pages for the most part.
1381 *
Mel Gormanc8244932012-01-12 17:19:38 -08001382 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1383 * that it is possible to migrate without blocking
1384 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001385 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001386 /* All the caller can do on PageWriteback is block */
1387 if (PageWriteback(page))
1388 return ret;
1389
1390 if (PageDirty(page)) {
1391 struct address_space *mapping;
1392
Mel Gormanc8244932012-01-12 17:19:38 -08001393 /*
1394 * Only pages without mappings or that have a
1395 * ->migratepage callback are possible to migrate
1396 * without blocking
1397 */
1398 mapping = page_mapping(page);
1399 if (mapping && !mapping->a_ops->migratepage)
1400 return ret;
1401 }
1402 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001403
Minchan Kimf80c0672011-10-31 17:06:55 -07001404 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1405 return ret;
1406
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001407 if (likely(get_page_unless_zero(page))) {
1408 /*
1409 * Be careful not to clear PageLRU until after we're
1410 * sure the page is not being freed elsewhere -- the
1411 * page release code relies on it.
1412 */
1413 ClearPageLRU(page);
1414 ret = 0;
1415 }
1416
1417 return ret;
1418}
1419
Mel Gorman7ee36a12016-07-28 15:47:17 -07001420
1421/*
1422 * Update LRU sizes after isolating pages. The LRU size updates must
1423 * be complete before mem_cgroup_update_lru_size due to a santity check.
1424 */
1425static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001426 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001427{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001428 int zid;
1429
Mel Gorman7ee36a12016-07-28 15:47:17 -07001430 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1431 if (!nr_zone_taken[zid])
1432 continue;
1433
1434 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001435#ifdef CONFIG_MEMCG
1436 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1437#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001438 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001439
Mel Gorman7ee36a12016-07-28 15:47:17 -07001440}
1441
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001442/*
Mel Gormana52633d2016-07-28 15:45:28 -07001443 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 * shrink the lists perform better by taking out a batch of pages
1445 * and working on them outside the LRU lock.
1446 *
1447 * For pagecache intensive workloads, this function is the hottest
1448 * spot in the kernel (apart from copy_*_user functions).
1449 *
1450 * Appropriate locks must be held before calling this function.
1451 *
1452 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001453 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001455 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001456 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001457 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001458 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 *
1460 * returns how many pages were moved onto *@dst.
1461 */
Andrew Morton69e05942006-03-22 00:08:19 -08001462static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001463 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001464 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001465 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001467 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001468 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001469 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001470 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07001471 unsigned long skipped = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001472 unsigned long scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001473 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001475 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
Johannes Weiner3db65812017-05-03 14:52:13 -07001476 !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001477 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 page = lru_to_page(src);
1480 prefetchw_prev_lru_page(page, src, flags);
1481
Sasha Levin309381fea2014-01-23 15:52:54 -08001482 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001483
Mel Gormanb2e18752016-07-28 15:45:37 -07001484 if (page_zonenum(page) > sc->reclaim_idx) {
1485 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001486 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001487 continue;
1488 }
1489
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001490 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001491 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001492 nr_pages = hpage_nr_pages(page);
1493 nr_taken += nr_pages;
1494 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001495 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001496 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001497
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001498 case -EBUSY:
1499 /* else it is being freed elsewhere */
1500 list_move(&page->lru, src);
1501 continue;
1502
1503 default:
1504 BUG();
1505 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 }
1507
Mel Gormanb2e18752016-07-28 15:45:37 -07001508 /*
1509 * Splice any skipped pages to the start of the LRU list. Note that
1510 * this disrupts the LRU order when reclaiming for lower zones but
1511 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1512 * scanning would soon rescan the same pages to skip and put the
1513 * system at risk of premature OOM.
1514 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001515 if (!list_empty(&pages_skipped)) {
1516 int zid;
1517
Johannes Weiner3db65812017-05-03 14:52:13 -07001518 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001519 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1520 if (!nr_skipped[zid])
1521 continue;
1522
1523 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08001524 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001525 }
1526 }
Johannes Weiner3db65812017-05-03 14:52:13 -07001527 *nr_scanned = scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08001528 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Michal Hocko32b3f292017-02-22 15:44:24 -08001529 scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 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
1689/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001690 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001691 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 */
Mel Gorman66635622010-08-09 17:19:30 -07001693static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001694shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001695 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696{
1697 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001698 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001699 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001700 unsigned long nr_taken;
Michal Hocko3c710c12017-02-22 15:44:27 -08001701 struct reclaim_stat stat = {};
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001702 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001703 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001704 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001705 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001706
Mel Gorman599d0c92016-07-28 15:45:31 -07001707 while (unlikely(too_many_isolated(pgdat, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001708 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001709
1710 /* We are about to die and free our memory. Return now. */
1711 if (fatal_signal_pending(current))
1712 return SWAP_CLUSTER_MAX;
1713 }
1714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001716
1717 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001718 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001719
Mel Gorman599d0c92016-07-28 15:45:31 -07001720 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001722 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1723 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001724
Mel Gorman599d0c92016-07-28 15:45:31 -07001725 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001726 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001727
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001728 if (global_reclaim(sc)) {
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001729 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001730 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001731 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001732 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001733 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001734 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001735
Hillf Dantond563c052012-03-21 16:34:02 -07001736 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001737 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001738
Shaohua Lia128ca72017-05-03 14:52:22 -07001739 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
Michal Hocko3c710c12017-02-22 15:44:27 -08001740 &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001741
Mel Gorman599d0c92016-07-28 15:45:31 -07001742 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001743
Ying Han904249a2012-04-25 16:01:48 -07001744 if (global_reclaim(sc)) {
1745 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001746 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001747 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001748 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001749 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001750
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001751 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001752
Mel Gorman599d0c92016-07-28 15:45:31 -07001753 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001754
Mel Gorman599d0c92016-07-28 15:45:31 -07001755 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001756
Johannes Weiner747db952014-08-08 14:19:24 -07001757 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001758 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001759
Mel Gorman92df3a72011-10-31 17:07:56 -07001760 /*
1761 * If reclaim is isolating dirty pages under writeback, it implies
1762 * that the long-lived page allocation rate is exceeding the page
1763 * laundering rate. Either the global limits are not being effective
1764 * at throttling processes due to the page distribution throughout
1765 * zones or there is heavy usage of a slow backing device. The
1766 * only option is to throttle from reclaim context which is not ideal
1767 * as there is no guarantee the dirtying process is throttled in the
1768 * same way balance_dirty_pages() manages.
1769 *
Mel Gorman8e950282013-07-03 15:02:02 -07001770 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1771 * of pages under pages flagged for immediate reclaim and stall if any
1772 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001773 */
Michal Hocko3c710c12017-02-22 15:44:27 -08001774 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001775 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001776
Mel Gormand43006d2013-07-03 15:01:50 -07001777 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001778 * Legacy memcg will stall in page writeback so avoid forcibly
1779 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001780 */
Tejun Heo97c93412015-05-22 18:23:36 -04001781 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001782 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001783 * Tag a zone as congested if all the dirty pages scanned were
1784 * backed by a congested BDI and wait_iff_congested will stall.
1785 */
Michal Hocko3c710c12017-02-22 15:44:27 -08001786 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001787 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001788
1789 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001790 * If dirty pages are scanned that are not queued for IO, it
Johannes Weiner726d0612017-02-24 14:56:14 -08001791 * implies that flushers are not doing their job. This can
1792 * happen when memory pressure pushes dirty pages to the end of
1793 * the LRU before the dirty limits are breached and the dirty
1794 * data has expired. It can also happen when the proportion of
1795 * dirty pages grows not through writes but through memory
1796 * pressure reclaiming all the clean cache. And in some cases,
1797 * the flushers simply cannot keep up with the allocation
1798 * rate. Nudge the flusher threads in case they are asleep, but
1799 * also allow kswapd to start writing pages during reclaim.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001800 */
Johannes Weiner726d0612017-02-24 14:56:14 -08001801 if (stat.nr_unqueued_dirty == nr_taken) {
1802 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
Mel Gorman599d0c92016-07-28 15:45:31 -07001803 set_bit(PGDAT_DIRTY, &pgdat->flags);
Johannes Weiner726d0612017-02-24 14:56:14 -08001804 }
Mel Gormanb1a6f212013-07-03 15:01:58 -07001805
1806 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001807 * If kswapd scans pages marked marked for immediate
1808 * reclaim and under writeback (nr_immediate), it implies
1809 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001810 * they are written so also forcibly stall.
1811 */
Michal Hocko3c710c12017-02-22 15:44:27 -08001812 if (stat.nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001813 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001814 }
Mel Gormand43006d2013-07-03 15:01:50 -07001815
Mel Gorman8e950282013-07-03 15:02:02 -07001816 /*
1817 * Stall direct reclaim for IO completions if underlying BDIs or zone
1818 * is congested. Allow kswapd to continue until it starts encountering
1819 * unqueued dirty pages or cycling through the LRU too quickly.
1820 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001821 if (!sc->hibernation_mode && !current_is_kswapd() &&
1822 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001823 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001824
Mel Gorman599d0c92016-07-28 15:45:31 -07001825 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1826 nr_scanned, nr_reclaimed,
Michal Hocko5bccd162017-02-22 15:44:30 -08001827 stat.nr_dirty, stat.nr_writeback,
1828 stat.nr_congested, stat.nr_immediate,
1829 stat.nr_activate, stat.nr_ref_keep,
1830 stat.nr_unmap_fail,
yalin wangba5e9572016-01-14 15:18:48 -08001831 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001832 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833}
1834
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001835/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 * This moves pages from the active list to the inactive list.
1837 *
1838 * We move them the other way if the page is referenced by one or more
1839 * processes, from rmap.
1840 *
1841 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001842 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001844 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 * this, so instead we remove the pages from the LRU while processing them.
1846 * It is safe to rely on PG_active against the non-LRU pages in here because
1847 * nobody will play with that bit on a non-LRU page.
1848 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001849 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 * But we had to alter page->flags anyway.
Michal Hocko9d998b42017-02-22 15:44:18 -08001851 *
1852 * Returns the number of pages moved to the given lru.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001854
Michal Hocko9d998b42017-02-22 15:44:18 -08001855static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001856 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001857 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001858 enum lru_list lru)
1859{
Mel Gorman599d0c92016-07-28 15:45:31 -07001860 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001861 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001862 int nr_pages;
Michal Hocko9d998b42017-02-22 15:44:18 -08001863 int nr_moved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001864
Wu Fengguang3eb41402009-06-16 15:33:13 -07001865 while (!list_empty(list)) {
1866 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001867 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001868
Sasha Levin309381fea2014-01-23 15:52:54 -08001869 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001870 SetPageLRU(page);
1871
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001872 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001873 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001874 list_move(&page->lru, &lruvec->lists[lru]);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001875
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001876 if (put_page_testzero(page)) {
1877 __ClearPageLRU(page);
1878 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001879 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001880
1881 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001882 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001883 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001884 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001885 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001886 } else
1887 list_add(&page->lru, pages_to_free);
Michal Hocko9d998b42017-02-22 15:44:18 -08001888 } else {
1889 nr_moved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001890 }
1891 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001892
Wu Fengguang3eb41402009-06-16 15:33:13 -07001893 if (!is_active_lru(lru))
Michal Hockof0958902017-02-22 15:45:55 -08001894 __count_vm_events(PGDEACTIVATE, nr_moved);
Michal Hocko9d998b42017-02-22 15:44:18 -08001895
1896 return nr_moved;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001897}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001898
Hugh Dickinsf6260122012-01-12 17:20:06 -08001899static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001900 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001901 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001902 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001904 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001905 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001906 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001908 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001909 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001911 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Michal Hocko9d998b42017-02-22 15:44:18 -08001912 unsigned nr_deactivate, nr_activate;
1913 unsigned nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001914 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001915 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001916 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
1918 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001919
1920 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001921 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001922
Mel Gorman599d0c92016-07-28 15:45:31 -07001923 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001924
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001925 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1926 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001927
Mel Gorman599d0c92016-07-28 15:45:31 -07001928 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001929 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001930
Mel Gorman599d0c92016-07-28 15:45:31 -07001931 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001932
Mel Gorman599d0c92016-07-28 15:45:31 -07001933 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 while (!list_empty(&l_hold)) {
1936 cond_resched();
1937 page = lru_to_page(&l_hold);
1938 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001939
Hugh Dickins39b5f292012-10-08 16:33:18 -07001940 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001941 putback_lru_page(page);
1942 continue;
1943 }
1944
Mel Gormancc715d92012-03-21 16:34:00 -07001945 if (unlikely(buffer_heads_over_limit)) {
1946 if (page_has_private(page) && trylock_page(page)) {
1947 if (page_has_private(page))
1948 try_to_release_page(page, 0);
1949 unlock_page(page);
1950 }
1951 }
1952
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001953 if (page_referenced(page, 0, sc->target_mem_cgroup,
1954 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001955 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001956 /*
1957 * Identify referenced, file-backed active pages and
1958 * give them one more trip around the active list. So
1959 * that executable code get better chances to stay in
1960 * memory under moderate memory pressure. Anon pages
1961 * are not likely to be evicted by use-once streaming
1962 * IO, plus JVM can create lots of anon VM_EXEC pages,
1963 * so we ignore them here.
1964 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001965 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001966 list_add(&page->lru, &l_active);
1967 continue;
1968 }
1969 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001970
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001971 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 list_add(&page->lru, &l_inactive);
1973 }
1974
Andrew Mortonb5557492009-01-06 14:40:13 -08001975 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001976 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001977 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001978 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001979 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001980 * Count referenced pages from currently used mappings as rotated,
1981 * even though only some of them are actually re-activated. This
1982 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001983 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001984 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001985 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001986
Michal Hocko9d998b42017-02-22 15:44:18 -08001987 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1988 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07001989 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1990 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001991
Johannes Weiner747db952014-08-08 14:19:24 -07001992 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001993 free_hot_cold_page_list(&l_hold, true);
Michal Hocko9d998b42017-02-22 15:44:18 -08001994 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
1995 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996}
1997
Rik van Riel59dc76b2016-05-20 16:56:31 -07001998/*
1999 * The inactive anon list should be small enough that the VM never has
2000 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002001 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002002 * The inactive file list should be small enough to leave most memory
2003 * to the established workingset on the scan-resistant active list,
2004 * but large enough to avoid thrashing the aggregate readahead window.
2005 *
2006 * Both inactive lists should also be large enough that each inactive
2007 * page has a chance to be referenced again before it is reclaimed.
2008 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002009 * If that fails and refaulting is observed, the inactive list grows.
2010 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002011 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2012 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2013 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2014 *
2015 * total target max
2016 * memory ratio inactive
2017 * -------------------------------------
2018 * 10MB 1 5MB
2019 * 100MB 1 50MB
2020 * 1GB 3 250MB
2021 * 10GB 10 0.9GB
2022 * 100GB 31 3GB
2023 * 1TB 101 10GB
2024 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002025 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002026static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002027 struct mem_cgroup *memcg,
2028 struct scan_control *sc, bool actual_reclaim)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002029{
Michal Hockofd538802017-02-22 15:45:58 -08002030 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002031 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2032 enum lru_list inactive_lru = file * LRU_FILE;
2033 unsigned long inactive, active;
2034 unsigned long inactive_ratio;
2035 unsigned long refaults;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002036 unsigned long gb;
2037
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002038 /*
2039 * If we don't have swap space, anonymous page deactivation
2040 * is pointless.
2041 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002042 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002043 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002044
Michal Hockofd538802017-02-22 15:45:58 -08002045 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2046 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002047
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002048 if (memcg)
Johannes Weinerccda7f42017-05-03 14:55:16 -07002049 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
Rik van Rielb39415b2009-12-14 17:59:48 -08002050 else
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002051 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
Rik van Riel59dc76b2016-05-20 16:56:31 -07002052
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002053 /*
2054 * When refaults are being observed, it means a new workingset
2055 * is being established. Disable active list protection to get
2056 * rid of the stale workingset quickly.
2057 */
2058 if (file && actual_reclaim && lruvec->refaults != refaults) {
2059 inactive_ratio = 0;
2060 } else {
2061 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2062 if (gb)
2063 inactive_ratio = int_sqrt(10 * gb);
2064 else
2065 inactive_ratio = 1;
2066 }
2067
2068 if (actual_reclaim)
2069 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2070 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2071 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2072 inactive_ratio, file);
Michal Hockofd538802017-02-22 15:45:58 -08002073
Rik van Riel59dc76b2016-05-20 16:56:31 -07002074 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002075}
2076
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002077static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002078 struct lruvec *lruvec, struct mem_cgroup *memcg,
2079 struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002080{
Rik van Rielb39415b2009-12-14 17:59:48 -08002081 if (is_active_lru(lru)) {
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002082 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2083 memcg, sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002084 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002085 return 0;
2086 }
2087
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002088 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002089}
2090
Johannes Weiner9a265112013-02-22 16:32:17 -08002091enum scan_balance {
2092 SCAN_EQUAL,
2093 SCAN_FRACT,
2094 SCAN_ANON,
2095 SCAN_FILE,
2096};
2097
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002099 * Determine how aggressively the anon and file LRU lists should be
2100 * scanned. The relative value of each set of LRU lists is determined
2101 * by looking at the fraction of the pages scanned we did rotate back
2102 * onto the active list instead of evict.
2103 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002104 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2105 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002106 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002107static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002108 struct scan_control *sc, unsigned long *nr,
2109 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002110{
Vladimir Davydov33377672016-01-20 15:02:59 -08002111 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002112 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002113 u64 fraction[2];
2114 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002115 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002116 unsigned long anon_prio, file_prio;
2117 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002118 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002119 unsigned long ap, fp;
2120 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002121
2122 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002123 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002124 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002125 goto out;
2126 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002127
Johannes Weiner10316b32013-02-22 16:32:14 -08002128 /*
2129 * Global reclaim will swap to prevent OOM even with no
2130 * swappiness, but memcg users want to use this knob to
2131 * disable swapping for individual groups completely when
2132 * using the memory controller's swap limit feature would be
2133 * too expensive.
2134 */
Johannes Weiner02695172014-08-06 16:06:17 -07002135 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002136 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002137 goto out;
2138 }
2139
2140 /*
2141 * Do not apply any pressure balancing cleverness when the
2142 * system is close to OOM, scan both anon and file equally
2143 * (unless the swappiness setting disagrees with swapping).
2144 */
Johannes Weiner02695172014-08-06 16:06:17 -07002145 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002146 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002147 goto out;
2148 }
2149
Johannes Weiner11d16c22013-02-22 16:32:15 -08002150 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002151 * Prevent the reclaimer from falling into the cache trap: as
2152 * cache pages start out inactive, every cache fault will tip
2153 * the scan balance towards the file LRU. And as the file LRU
2154 * shrinks, so does the window for rotation from references.
2155 * This means we have a runaway feedback loop where a tiny
2156 * thrashing file LRU becomes infinitely more attractive than
2157 * anon pages. Try to detect this based on file LRU size.
2158 */
2159 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002160 unsigned long pgdatfile;
2161 unsigned long pgdatfree;
2162 int z;
2163 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002164
Mel Gorman599d0c92016-07-28 15:45:31 -07002165 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2166 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2167 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002168
Mel Gorman599d0c92016-07-28 15:45:31 -07002169 for (z = 0; z < MAX_NR_ZONES; z++) {
2170 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002171 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002172 continue;
2173
2174 total_high_wmark += high_wmark_pages(zone);
2175 }
2176
2177 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002178 scan_balance = SCAN_ANON;
2179 goto out;
2180 }
2181 }
2182
2183 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002184 * If there is enough inactive page cache, i.e. if the size of the
2185 * inactive list is greater than that of the active list *and* the
2186 * inactive list actually has some pages to scan on this priority, we
2187 * do not reclaim anything from the anonymous working set right now.
2188 * Without the second condition we could end up never scanning an
2189 * lruvec even if it has plenty of old anonymous pages unless the
2190 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002191 */
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002192 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
Michal Hocko71ab6cf2017-02-22 15:46:01 -08002193 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002194 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002195 goto out;
2196 }
2197
Johannes Weiner9a265112013-02-22 16:32:17 -08002198 scan_balance = SCAN_FRACT;
2199
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002200 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002201 * With swappiness at 100, anonymous and file have the same priority.
2202 * This scanning priority is essentially the inverse of IO cost.
2203 */
Johannes Weiner02695172014-08-06 16:06:17 -07002204 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002205 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002206
2207 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002208 * OK, so we have swap space and a fair amount of page cache
2209 * pages. We use the recently rotated / recently scanned
2210 * ratios to determine how valuable each cache is.
2211 *
2212 * Because workloads change over time (and to avoid overflow)
2213 * we keep these statistics as a floating average, which ends
2214 * up weighing recent references more than old ones.
2215 *
2216 * anon in [0], file in [1]
2217 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002218
Michal Hockofd538802017-02-22 15:45:58 -08002219 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2220 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2221 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2222 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002223
Mel Gorman599d0c92016-07-28 15:45:31 -07002224 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002225 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002226 reclaim_stat->recent_scanned[0] /= 2;
2227 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002228 }
2229
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002230 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002231 reclaim_stat->recent_scanned[1] /= 2;
2232 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002233 }
2234
2235 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002236 * The amount of pressure on anon vs file pages is inversely
2237 * proportional to the fraction of recently scanned pages on
2238 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002239 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002240 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002241 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002242
Satoru Moriyafe350042012-05-29 15:06:47 -07002243 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002244 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002245 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002246
Shaohua Li76a33fc2010-05-24 14:32:36 -07002247 fraction[0] = ap;
2248 fraction[1] = fp;
2249 denominator = ap + fp + 1;
2250out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002251 *lru_pages = 0;
2252 for_each_evictable_lru(lru) {
2253 int file = is_file_lru(lru);
2254 unsigned long size;
2255 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002256
Johannes Weiner688035f2017-05-03 14:52:07 -07002257 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2258 scan = size >> sc->priority;
2259 /*
2260 * If the cgroup's already been deleted, make sure to
2261 * scrape out the remaining cache.
2262 */
2263 if (!scan && !mem_cgroup_online(memcg))
2264 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002265
Johannes Weiner688035f2017-05-03 14:52:07 -07002266 switch (scan_balance) {
2267 case SCAN_EQUAL:
2268 /* Scan lists relative to size */
2269 break;
2270 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002271 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002272 * Scan types proportional to swappiness and
2273 * their relative recent reclaim efficiency.
Johannes Weiner9a265112013-02-22 16:32:17 -08002274 */
Johannes Weiner688035f2017-05-03 14:52:07 -07002275 scan = div64_u64(scan * fraction[file],
2276 denominator);
2277 break;
2278 case SCAN_FILE:
2279 case SCAN_ANON:
2280 /* Scan one type exclusively */
2281 if ((scan_balance == SCAN_FILE) != file) {
2282 size = 0;
2283 scan = 0;
2284 }
2285 break;
2286 default:
2287 /* Look ma, no brain */
2288 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002289 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002290
2291 *lru_pages += size;
2292 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002293 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002294}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002295
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002296/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002297 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002298 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002299static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002300 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002301{
Mel Gormanef8f2322016-07-28 15:46:05 -07002302 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002303 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002304 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002305 unsigned long nr_to_scan;
2306 enum lru_list lru;
2307 unsigned long nr_reclaimed = 0;
2308 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2309 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002310 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002311
Vladimir Davydov33377672016-01-20 15:02:59 -08002312 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002313
Mel Gormane82e0562013-07-03 15:01:44 -07002314 /* Record the original scan target for proportional adjustments later */
2315 memcpy(targets, nr, sizeof(nr));
2316
Mel Gorman1a501902014-06-04 16:10:49 -07002317 /*
2318 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2319 * event that can occur when there is little memory pressure e.g.
2320 * multiple streaming readers/writers. Hence, we do not abort scanning
2321 * when the requested number of pages are reclaimed when scanning at
2322 * DEF_PRIORITY on the assumption that the fact we are direct
2323 * reclaiming implies that kswapd is not keeping up and it is best to
2324 * do a batch of work at once. For memcg reclaim one check is made to
2325 * abort proportional reclaim if either the file or anon lru has already
2326 * dropped to zero at the first pass.
2327 */
2328 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2329 sc->priority == DEF_PRIORITY);
2330
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002331 blk_start_plug(&plug);
2332 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2333 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002334 unsigned long nr_anon, nr_file, percentage;
2335 unsigned long nr_scanned;
2336
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002337 for_each_evictable_lru(lru) {
2338 if (nr[lru]) {
2339 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2340 nr[lru] -= nr_to_scan;
2341
2342 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002343 lruvec, memcg, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002344 }
2345 }
Mel Gormane82e0562013-07-03 15:01:44 -07002346
Michal Hockobd041732016-12-02 17:26:48 -08002347 cond_resched();
2348
Mel Gormane82e0562013-07-03 15:01:44 -07002349 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2350 continue;
2351
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002352 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002353 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002354 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002355 * proportionally what was requested by get_scan_count(). We
2356 * stop reclaiming one LRU and reduce the amount scanning
2357 * proportional to the original scan target.
2358 */
2359 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2360 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2361
Mel Gorman1a501902014-06-04 16:10:49 -07002362 /*
2363 * It's just vindictive to attack the larger once the smaller
2364 * has gone to zero. And given the way we stop scanning the
2365 * smaller below, this makes sure that we only make one nudge
2366 * towards proportionality once we've got nr_to_reclaim.
2367 */
2368 if (!nr_file || !nr_anon)
2369 break;
2370
Mel Gormane82e0562013-07-03 15:01:44 -07002371 if (nr_file > nr_anon) {
2372 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2373 targets[LRU_ACTIVE_ANON] + 1;
2374 lru = LRU_BASE;
2375 percentage = nr_anon * 100 / scan_target;
2376 } else {
2377 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2378 targets[LRU_ACTIVE_FILE] + 1;
2379 lru = LRU_FILE;
2380 percentage = nr_file * 100 / scan_target;
2381 }
2382
2383 /* Stop scanning the smaller of the LRU */
2384 nr[lru] = 0;
2385 nr[lru + LRU_ACTIVE] = 0;
2386
2387 /*
2388 * Recalculate the other LRU scan count based on its original
2389 * scan target and the percentage scanning already complete
2390 */
2391 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2392 nr_scanned = targets[lru] - nr[lru];
2393 nr[lru] = targets[lru] * (100 - percentage) / 100;
2394 nr[lru] -= min(nr[lru], nr_scanned);
2395
2396 lru += LRU_ACTIVE;
2397 nr_scanned = targets[lru] - nr[lru];
2398 nr[lru] = targets[lru] * (100 - percentage) / 100;
2399 nr[lru] -= min(nr[lru], nr_scanned);
2400
2401 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002402 }
2403 blk_finish_plug(&plug);
2404 sc->nr_reclaimed += nr_reclaimed;
2405
2406 /*
2407 * Even if we did not try to evict anon pages at all, we want to
2408 * rebalance the anon lru active/inactive ratio.
2409 */
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002410 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002411 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2412 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002413}
2414
Mel Gorman23b9da52012-05-29 15:06:20 -07002415/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002416static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002417{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002418 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002419 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002420 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002421 return true;
2422
2423 return false;
2424}
2425
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002426/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002427 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2428 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2429 * true if more pages should be reclaimed such that when the page allocator
2430 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2431 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002432 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002433static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002434 unsigned long nr_reclaimed,
2435 unsigned long nr_scanned,
2436 struct scan_control *sc)
2437{
2438 unsigned long pages_for_compaction;
2439 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002440 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002441
2442 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002443 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002444 return false;
2445
Mel Gorman28765922011-02-25 14:44:20 -08002446 /* Consider stopping depending on scan and reclaim activity */
2447 if (sc->gfp_mask & __GFP_REPEAT) {
2448 /*
2449 * For __GFP_REPEAT allocations, stop reclaiming if the
2450 * full LRU list has been scanned and we are still failing
2451 * to reclaim pages. This full LRU scan is potentially
2452 * expensive but a __GFP_REPEAT caller really wants to succeed
2453 */
2454 if (!nr_reclaimed && !nr_scanned)
2455 return false;
2456 } else {
2457 /*
2458 * For non-__GFP_REPEAT allocations which can presumably
2459 * fail without consequence, stop if we failed to reclaim
2460 * any pages from the last SWAP_CLUSTER_MAX number of
2461 * pages that were scanned. This will return to the
2462 * caller faster at the risk reclaim/compaction and
2463 * the resulting allocation attempt fails
2464 */
2465 if (!nr_reclaimed)
2466 return false;
2467 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002468
2469 /*
2470 * If we have not reclaimed enough pages for compaction and the
2471 * inactive lists are large enough, continue reclaiming
2472 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002473 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002474 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002475 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002476 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002477 if (sc->nr_reclaimed < pages_for_compaction &&
2478 inactive_lru_pages > pages_for_compaction)
2479 return true;
2480
2481 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002482 for (z = 0; z <= sc->reclaim_idx; z++) {
2483 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002484 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002485 continue;
2486
2487 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002488 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002489 case COMPACT_CONTINUE:
2490 return false;
2491 default:
2492 /* check next zone */
2493 ;
2494 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002495 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002496 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002497}
2498
Mel Gorman970a39a2016-07-28 15:46:35 -07002499static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002500{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002501 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002502 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002503 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002504
Johannes Weiner56600482012-01-12 17:17:59 -08002505 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002506 struct mem_cgroup *root = sc->target_mem_cgroup;
2507 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002508 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002509 .priority = sc->priority,
2510 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002511 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002512 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002513
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002514 nr_reclaimed = sc->nr_reclaimed;
2515 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002516
Andrew Morton694fbc02013-09-24 15:27:37 -07002517 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2518 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002519 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002520 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002521 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002522
Johannes Weiner241994ed2015-02-11 15:26:06 -08002523 if (mem_cgroup_low(root, memcg)) {
Yisheng Xied6622f62017-05-03 14:53:57 -07002524 if (!sc->memcg_low_reclaim) {
2525 sc->memcg_low_skipped = 1;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002526 continue;
Yisheng Xied6622f62017-05-03 14:53:57 -07002527 }
Johannes Weiner31176c72017-05-03 14:55:07 -07002528 mem_cgroup_event(memcg, MEMCG_LOW);
Johannes Weiner241994ed2015-02-11 15:26:06 -08002529 }
2530
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002531 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002532 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002533
Mel Gormana9dd0a82016-07-28 15:46:02 -07002534 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2535 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002536
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002537 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002538 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002539 memcg, sc->nr_scanned - scanned,
2540 lru_pages);
2541
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002542 /* Record the group's reclaim efficiency */
2543 vmpressure(sc->gfp_mask, memcg, false,
2544 sc->nr_scanned - scanned,
2545 sc->nr_reclaimed - reclaimed);
2546
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002547 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002548 * Direct reclaim and kswapd have to scan all memory
2549 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002550 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002551 *
2552 * Limit reclaim, on the other hand, only cares about
2553 * nr_to_reclaim pages to be reclaimed and it will
2554 * retry with decreasing priority if one round over the
2555 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002556 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002557 if (!global_reclaim(sc) &&
2558 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002559 mem_cgroup_iter_break(root, memcg);
2560 break;
2561 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002562 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002563
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002564 /*
2565 * Shrink the slab caches in the same proportion that
2566 * the eligible LRU pages were scanned.
2567 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002568 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002569 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002570 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002571 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002572
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002573 if (reclaim_state) {
2574 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2575 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002576 }
2577
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002578 /* Record the subtree's reclaim efficiency */
2579 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002580 sc->nr_scanned - nr_scanned,
2581 sc->nr_reclaimed - nr_reclaimed);
2582
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002583 if (sc->nr_reclaimed - nr_reclaimed)
2584 reclaimable = true;
2585
Mel Gormana9dd0a82016-07-28 15:46:02 -07002586 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002587 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002588
Johannes Weinerc73322d2017-05-03 14:51:51 -07002589 /*
2590 * Kswapd gives up on balancing particular nodes after too
2591 * many failures to reclaim anything from them and goes to
2592 * sleep. On reclaim progress, reset the failure counter. A
2593 * successful direct reclaim run will revive a dormant kswapd.
2594 */
2595 if (reclaimable)
2596 pgdat->kswapd_failures = 0;
2597
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002598 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002599}
2600
Vlastimil Babka53853e22014-10-09 15:27:02 -07002601/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002602 * Returns true if compaction should go ahead for a costly-order request, or
2603 * the allocation would already succeed without compaction. Return false if we
2604 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002605 */
Mel Gorman4f588332016-07-28 15:46:38 -07002606static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002607{
Mel Gorman31483b62016-07-28 15:45:46 -07002608 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002609 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002610
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002611 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2612 if (suitable == COMPACT_SUCCESS)
2613 /* Allocation should succeed already. Don't reclaim. */
2614 return true;
2615 if (suitable == COMPACT_SKIPPED)
2616 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002617 return false;
2618
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002619 /*
2620 * Compaction is already possible, but it takes time to run and there
2621 * are potentially other callers using the pages just freed. So proceed
2622 * with reclaim to make a buffer of free pages available to give
2623 * compaction a reasonable chance of completing and allocating the page.
2624 * Note that we won't actually reclaim the whole buffer in one attempt
2625 * as the target watermark in should_continue_reclaim() is lower. But if
2626 * we are already above the high+gap watermark, don't reclaim at all.
2627 */
2628 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2629
2630 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002631}
2632
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633/*
2634 * This is the direct reclaim path, for page-allocating processes. We only
2635 * try to reclaim pages from zones which will satisfy the caller's allocation
2636 * request.
2637 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638 * If a zone is deemed to be full of pinned pages then just give it a light
2639 * scan then give up on it.
2640 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002641static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642{
Mel Gormandd1a2392008-04-28 02:12:17 -07002643 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002644 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002645 unsigned long nr_soft_reclaimed;
2646 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002647 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002648 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002649
Mel Gormancc715d92012-03-21 16:34:00 -07002650 /*
2651 * If the number of buffer_heads in the machine exceeds the maximum
2652 * allowed level, force direct reclaim to scan the highmem zone as
2653 * highmem pages could be pinning lowmem pages storing buffer_heads
2654 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002655 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002656 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002657 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002658 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002659 }
Mel Gormancc715d92012-03-21 16:34:00 -07002660
Mel Gormand4debc62010-08-09 17:19:29 -07002661 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002662 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002663 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002664 * Take care memory controller reclaiming has small influence
2665 * to global LRU.
2666 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002667 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002668 if (!cpuset_zone_allowed(zone,
2669 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002670 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002671
Johannes Weiner0b064962014-08-06 16:06:12 -07002672 /*
2673 * If we already have plenty of memory free for
2674 * compaction in this zone, don't free any more.
2675 * Even though compaction is invoked for any
2676 * non-zero order, only frequent costly order
2677 * reclamation is disruptive enough to become a
2678 * noticeable problem, like transparent huge
2679 * page allocations.
2680 */
2681 if (IS_ENABLED(CONFIG_COMPACTION) &&
2682 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002683 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002684 sc->compaction_ready = true;
2685 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002686 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002687
Andrew Morton0608f432013-09-24 15:27:41 -07002688 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002689 * Shrink each node in the zonelist once. If the
2690 * zonelist is ordered by zone (not the default) then a
2691 * node may be shrunk multiple times but in that case
2692 * the user prefers lower zones being preserved.
2693 */
2694 if (zone->zone_pgdat == last_pgdat)
2695 continue;
2696
2697 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002698 * This steals pages from memory cgroups over softlimit
2699 * and returns the number of reclaimed pages and
2700 * scanned pages. This works for global memory pressure
2701 * and balancing, not for a memcg's limit.
2702 */
2703 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002704 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002705 sc->order, sc->gfp_mask,
2706 &nr_soft_scanned);
2707 sc->nr_reclaimed += nr_soft_reclaimed;
2708 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002709 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002710 }
Nick Piggin408d8542006-09-25 23:31:27 -07002711
Mel Gorman79dafcd2016-07-28 15:45:53 -07002712 /* See comment about same check for global reclaim above */
2713 if (zone->zone_pgdat == last_pgdat)
2714 continue;
2715 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002716 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 }
Mel Gormane0c23272011-10-31 17:09:33 -07002718
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002719 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002720 * Restore to original mask to avoid the impact on the caller if we
2721 * promoted it to __GFP_HIGHMEM.
2722 */
2723 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002725
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002726static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2727{
2728 struct mem_cgroup *memcg;
2729
2730 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2731 do {
2732 unsigned long refaults;
2733 struct lruvec *lruvec;
2734
2735 if (memcg)
Johannes Weinerccda7f42017-05-03 14:55:16 -07002736 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002737 else
2738 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2739
2740 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2741 lruvec->refaults = refaults;
2742 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2743}
2744
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745/*
2746 * This is the main entry point to direct page reclaim.
2747 *
2748 * If a full scan of the inactive list fails to free enough memory then we
2749 * are "out of memory" and something needs to be killed.
2750 *
2751 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2752 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002753 * caller can't do much about. We kick the writeback threads and take explicit
2754 * naps in the hope that some of these pages can be written. But if the
2755 * allocating task holds filesystem locks which prevent writeout this might not
2756 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002757 *
2758 * returns: 0, if no pages reclaimed
2759 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 */
Mel Gormandac1d272008-04-28 02:12:12 -07002761static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002762 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002764 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002765 pg_data_t *last_pgdat;
2766 struct zoneref *z;
2767 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002768retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002769 delayacct_freepages_start();
2770
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002771 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002772 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002774 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002775 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2776 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002777 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002778 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002779
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002780 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002781 break;
2782
2783 if (sc->compaction_ready)
2784 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
2786 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002787 * If we're getting trouble reclaiming, start doing
2788 * writepage even in laptop mode.
2789 */
2790 if (sc->priority < DEF_PRIORITY - 2)
2791 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07002792 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002793
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002794 last_pgdat = NULL;
2795 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2796 sc->nodemask) {
2797 if (zone->zone_pgdat == last_pgdat)
2798 continue;
2799 last_pgdat = zone->zone_pgdat;
2800 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2801 }
2802
Keika Kobayashi873b4772008-07-25 01:48:52 -07002803 delayacct_freepages_end();
2804
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002805 if (sc->nr_reclaimed)
2806 return sc->nr_reclaimed;
2807
Mel Gorman0cee34f2012-01-12 17:19:49 -08002808 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002809 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002810 return 1;
2811
Johannes Weiner241994ed2015-02-11 15:26:06 -08002812 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07002813 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08002814 sc->priority = initial_priority;
Yisheng Xied6622f62017-05-03 14:53:57 -07002815 sc->memcg_low_reclaim = 1;
2816 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002817 goto retry;
2818 }
2819
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002820 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821}
2822
Johannes Weinerc73322d2017-05-03 14:51:51 -07002823static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07002824{
2825 struct zone *zone;
2826 unsigned long pfmemalloc_reserve = 0;
2827 unsigned long free_pages = 0;
2828 int i;
2829 bool wmark_ok;
2830
Johannes Weinerc73322d2017-05-03 14:51:51 -07002831 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2832 return true;
2833
Mel Gorman55150612012-07-31 16:44:35 -07002834 for (i = 0; i <= ZONE_NORMAL; i++) {
2835 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd2017-05-03 14:51:54 -07002836 if (!managed_zone(zone))
2837 continue;
2838
2839 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07002840 continue;
2841
Mel Gorman55150612012-07-31 16:44:35 -07002842 pfmemalloc_reserve += min_wmark_pages(zone);
2843 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2844 }
2845
Mel Gorman675becc2014-06-04 16:07:35 -07002846 /* If there are no reserves (unexpected config) then do not throttle */
2847 if (!pfmemalloc_reserve)
2848 return true;
2849
Mel Gorman55150612012-07-31 16:44:35 -07002850 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2851
2852 /* kswapd must be awake if processes are being throttled */
2853 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002854 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002855 (enum zone_type)ZONE_NORMAL);
2856 wake_up_interruptible(&pgdat->kswapd_wait);
2857 }
2858
2859 return wmark_ok;
2860}
2861
2862/*
2863 * Throttle direct reclaimers if backing storage is backed by the network
2864 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2865 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002866 * when the low watermark is reached.
2867 *
2868 * Returns true if a fatal signal was delivered during throttling. If this
2869 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002870 */
Mel Gorman50694c22012-11-26 16:29:48 -08002871static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002872 nodemask_t *nodemask)
2873{
Mel Gorman675becc2014-06-04 16:07:35 -07002874 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002875 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002876 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002877
2878 /*
2879 * Kernel threads should not be throttled as they may be indirectly
2880 * responsible for cleaning pages necessary for reclaim to make forward
2881 * progress. kjournald for example may enter direct reclaim while
2882 * committing a transaction where throttling it could forcing other
2883 * processes to block on log_wait_commit().
2884 */
2885 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002886 goto out;
2887
2888 /*
2889 * If a fatal signal is pending, this process should not throttle.
2890 * It should return quickly so it can exit and free its memory
2891 */
2892 if (fatal_signal_pending(current))
2893 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002894
Mel Gorman675becc2014-06-04 16:07:35 -07002895 /*
2896 * Check if the pfmemalloc reserves are ok by finding the first node
2897 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2898 * GFP_KERNEL will be required for allocating network buffers when
2899 * swapping over the network so ZONE_HIGHMEM is unusable.
2900 *
2901 * Throttling is based on the first usable node and throttled processes
2902 * wait on a queue until kswapd makes progress and wakes them. There
2903 * is an affinity then between processes waking up and where reclaim
2904 * progress has been made assuming the process wakes on the same node.
2905 * More importantly, processes running on remote nodes will not compete
2906 * for remote pfmemalloc reserves and processes on different nodes
2907 * should make reasonable progress.
2908 */
2909 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002910 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002911 if (zone_idx(zone) > ZONE_NORMAL)
2912 continue;
2913
2914 /* Throttle based on the first usable node */
2915 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07002916 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07002917 goto out;
2918 break;
2919 }
2920
2921 /* If no zone was usable by the allocation flags then do not throttle */
2922 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002923 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002924
Mel Gorman68243e72012-07-31 16:44:39 -07002925 /* Account for the throttling */
2926 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2927
Mel Gorman55150612012-07-31 16:44:35 -07002928 /*
2929 * If the caller cannot enter the filesystem, it's possible that it
2930 * is due to the caller holding an FS lock or performing a journal
2931 * transaction in the case of a filesystem like ext[3|4]. In this case,
2932 * it is not safe to block on pfmemalloc_wait as kswapd could be
2933 * blocked waiting on the same lock. Instead, throttle for up to a
2934 * second before continuing.
2935 */
2936 if (!(gfp_mask & __GFP_FS)) {
2937 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07002938 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002939
2940 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002941 }
2942
2943 /* Throttle until kswapd wakes the process */
2944 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07002945 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002946
2947check_pending:
2948 if (fatal_signal_pending(current))
2949 return true;
2950
2951out:
2952 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002953}
2954
Mel Gormandac1d272008-04-28 02:12:12 -07002955unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002956 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002957{
Mel Gorman33906bc2010-08-09 17:19:16 -07002958 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002959 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002960 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Michal Hocko7dea19f2017-05-03 14:53:15 -07002961 .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)),
Mel Gormanb2e18752016-07-28 15:45:37 -07002962 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07002963 .order = order,
2964 .nodemask = nodemask,
2965 .priority = DEF_PRIORITY,
2966 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002967 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002968 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002969 };
2970
Mel Gorman55150612012-07-31 16:44:35 -07002971 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002972 * Do not enter reclaim if fatal signal was delivered while throttled.
2973 * 1 is returned so that the page allocator does not OOM kill at this
2974 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002975 */
Mel Gorman50694c22012-11-26 16:29:48 -08002976 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002977 return 1;
2978
Mel Gorman33906bc2010-08-09 17:19:16 -07002979 trace_mm_vmscan_direct_reclaim_begin(order,
2980 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07002981 gfp_mask,
2982 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07002983
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002984 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002985
2986 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2987
2988 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002989}
2990
Andrew Mortonc255a452012-07-31 16:43:02 -07002991#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002992
Mel Gormana9dd0a82016-07-28 15:46:02 -07002993unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002994 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07002995 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07002996 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002997{
2998 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002999 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003000 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003001 .may_writepage = !laptop_mode,
3002 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003003 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003004 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003005 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003006 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003007
Balbir Singh4e416952009-09-23 15:56:39 -07003008 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3009 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003010
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003011 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003012 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003013 sc.gfp_mask,
3014 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003015
Balbir Singh4e416952009-09-23 15:56:39 -07003016 /*
3017 * NOTE: Although we can get the priority field, using it
3018 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003019 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003020 * will pick up pages from other mem cgroup's as well. We hack
3021 * the priority and make it zero.
3022 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003023 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003024
3025 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3026
Ying Han0ae5e892011-05-26 16:25:25 -07003027 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003028 return sc.nr_reclaimed;
3029}
3030
Johannes Weiner72835c82012-01-12 17:18:32 -08003031unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003032 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003033 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003034 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003035{
Balbir Singh4e416952009-09-23 15:56:39 -07003036 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003037 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003038 int nid;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003039 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08003040 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003041 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003042 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07003043 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003044 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003045 .target_mem_cgroup = memcg,
3046 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003047 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003048 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003049 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003050 };
Balbir Singh66e17072008-02-07 00:13:56 -08003051
Ying Han889976d2011-05-26 16:25:33 -07003052 /*
3053 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3054 * take care of from where we get pages. So the node where we start the
3055 * scan does not need to be the current node.
3056 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003057 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003058
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003059 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003060
3061 trace_mm_vmscan_memcg_reclaim_begin(0,
3062 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003063 sc.gfp_mask,
3064 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003065
Vlastimil Babka499118e2017-05-08 15:59:50 -07003066 noreclaim_flag = memalloc_noreclaim_save();
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003067 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003068 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003069
3070 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3071
3072 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003073}
3074#endif
3075
Mel Gorman1d82de62016-07-28 15:45:43 -07003076static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003077 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003078{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003079 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003080
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003081 if (!total_swap_pages)
3082 return;
3083
3084 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3085 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003086 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003087
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003088 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003089 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003090 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003091
3092 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3093 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003094}
3095
Mel Gormane716f2e2017-05-03 14:53:45 -07003096/*
3097 * Returns true if there is an eligible zone balanced for the request order
3098 * and classzone_idx
3099 */
3100static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003101{
Mel Gormane716f2e2017-05-03 14:53:45 -07003102 int i;
3103 unsigned long mark = -1;
3104 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003105
Mel Gormane716f2e2017-05-03 14:53:45 -07003106 for (i = 0; i <= classzone_idx; i++) {
3107 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07003108
Mel Gormane716f2e2017-05-03 14:53:45 -07003109 if (!managed_zone(zone))
3110 continue;
3111
3112 mark = high_wmark_pages(zone);
3113 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3114 return true;
3115 }
3116
3117 /*
3118 * If a node has no populated zone within classzone_idx, it does not
3119 * need balancing by definition. This can happen if a zone-restricted
3120 * allocation tries to wake a remote kswapd.
3121 */
3122 if (mark == -1)
3123 return true;
3124
3125 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003126}
3127
Mel Gorman631b6e02017-05-03 14:53:41 -07003128/* Clear pgdat state for congested, dirty or under writeback. */
3129static void clear_pgdat_congested(pg_data_t *pgdat)
3130{
3131 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3132 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3133 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3134}
3135
Mel Gorman1741c872011-01-13 15:46:21 -08003136/*
Mel Gorman55150612012-07-31 16:44:35 -07003137 * Prepare kswapd for sleeping. This verifies that there are no processes
3138 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3139 *
3140 * Returns true if kswapd is ready to sleep
3141 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003142static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003143{
Mel Gorman55150612012-07-31 16:44:35 -07003144 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003145 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07003146 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003147 * race between when kswapd checks the watermarks and a process gets
3148 * throttled. There is also a potential race if processes get
3149 * throttled, kswapd wakes, a large process exits thereby balancing the
3150 * zones, which causes kswapd to exit balance_pgdat() before reaching
3151 * the wake up checks. If kswapd is going to sleep, no process should
3152 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3153 * the wake up is premature, processes will wake kswapd and get
3154 * throttled again. The difference from wake ups in balance_pgdat() is
3155 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003156 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003157 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3158 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003159
Johannes Weinerc73322d2017-05-03 14:51:51 -07003160 /* Hopeless node, leave it to direct reclaim */
3161 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3162 return true;
3163
Mel Gormane716f2e2017-05-03 14:53:45 -07003164 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3165 clear_pgdat_congested(pgdat);
3166 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003167 }
3168
Shantanu Goel333b0a42017-05-03 14:53:38 -07003169 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003170}
3171
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003173 * kswapd shrinks a node of pages that are at or below the highest usable
3174 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003175 *
3176 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003177 * reclaim or if the lack of progress was due to pages under writeback.
3178 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003179 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003180static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003181 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003182{
Mel Gorman1d82de62016-07-28 15:45:43 -07003183 struct zone *zone;
3184 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003185
Mel Gorman1d82de62016-07-28 15:45:43 -07003186 /* Reclaim a number of pages proportional to the number of zones */
3187 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003188 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003189 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003190 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003191 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003192
Mel Gorman1d82de62016-07-28 15:45:43 -07003193 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003194 }
3195
Mel Gorman1d82de62016-07-28 15:45:43 -07003196 /*
3197 * Historically care was taken to put equal pressure on all zones but
3198 * now pressure is applied based on node LRU order.
3199 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003200 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003201
3202 /*
3203 * Fragmentation may mean that the system cannot be rebalanced for
3204 * high-order allocations. If twice the allocation size has been
3205 * reclaimed then recheck watermarks only at order-0 to prevent
3206 * excessive reclaim. Assume that a process requested a high-order
3207 * can direct reclaim/compact.
3208 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003209 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003210 sc->order = 0;
3211
Mel Gormanb8e83b92013-07-03 15:01:45 -07003212 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003213}
3214
3215/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003216 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3217 * that are eligible for use by the caller until at least one zone is
3218 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003220 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 *
3222 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003223 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003224 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3225 * or lower is eligible for reclaim until at least one usable zone is
3226 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003228static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003231 unsigned long nr_soft_reclaimed;
3232 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003233 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003234 struct scan_control sc = {
3235 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003236 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003237 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003238 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003239 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003240 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003241 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003242 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003244 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07003245 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003246 bool raise_priority = true;
3247
Mel Gorman84c7a772016-07-28 15:46:44 -07003248 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249
Mel Gorman86c79f62016-07-28 15:45:59 -07003250 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003251 * If the number of buffer_heads exceeds the maximum allowed
3252 * then consider reclaiming from all zones. This has a dual
3253 * purpose -- on 64-bit systems it is expected that
3254 * buffer_heads are stripped during active rotation. On 32-bit
3255 * systems, highmem pages can pin lowmem memory and shrinking
3256 * buffers can relieve lowmem pressure. Reclaim may still not
3257 * go ahead if all eligible zones for the original allocation
3258 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003259 */
3260 if (buffer_heads_over_limit) {
3261 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3262 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003263 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003264 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265
Mel Gorman970a39a2016-07-28 15:46:35 -07003266 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003267 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003270
Mel Gorman86c79f62016-07-28 15:45:59 -07003271 /*
Mel Gormane716f2e2017-05-03 14:53:45 -07003272 * Only reclaim if there are no eligible zones. Note that
3273 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3274 * have adjusted it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003275 */
Mel Gormane716f2e2017-05-03 14:53:45 -07003276 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3277 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003278
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003280 * Do some background aging of the anon list, to give
3281 * pages a chance to be referenced before reclaiming. All
3282 * pages are rotated regardless of classzone as this is
3283 * about consistent aging.
3284 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003285 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003286
3287 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003288 * If we're getting trouble reclaiming, start doing writepage
3289 * even in laptop mode.
3290 */
Johannes Weiner047d72c2017-05-03 14:51:57 -07003291 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003292 sc.may_writepage = 1;
3293
Mel Gorman1d82de62016-07-28 15:45:43 -07003294 /* Call soft limit reclaim before calling shrink_node. */
3295 sc.nr_scanned = 0;
3296 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003297 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003298 sc.gfp_mask, &nr_soft_scanned);
3299 sc.nr_reclaimed += nr_soft_reclaimed;
3300
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003301 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003302 * There should be no need to raise the scanning priority if
3303 * enough pages are already being scanned that that high
3304 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003306 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003307 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003308
3309 /*
3310 * If the low watermark is met there is no need for processes
3311 * to be throttled on pfmemalloc_wait as they should not be
3312 * able to safely make forward progress. Wake them
3313 */
3314 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc73322d2017-05-03 14:51:51 -07003315 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003316 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003317
Mel Gormanb8e83b92013-07-03 15:01:45 -07003318 /* Check if kswapd should be suspending */
3319 if (try_to_freeze() || kthread_should_stop())
3320 break;
3321
3322 /*
3323 * Raise priority if scanning rate is too low or there was no
3324 * progress in reclaiming pages
3325 */
Johannes Weinerc73322d2017-05-03 14:51:51 -07003326 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3327 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003328 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003329 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330
Johannes Weinerc73322d2017-05-03 14:51:51 -07003331 if (!sc.nr_reclaimed)
3332 pgdat->kswapd_failures++;
3333
Mel Gormanb8e83b92013-07-03 15:01:45 -07003334out:
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003335 snapshot_refaults(NULL, pgdat);
Mel Gorman0abdee22011-01-13 15:46:22 -08003336 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003337 * Return the order kswapd stopped reclaiming at as
3338 * prepare_kswapd_sleep() takes it into account. If another caller
3339 * entered the allocator slow path while kswapd was awake, order will
3340 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003341 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003342 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343}
3344
Mel Gormane716f2e2017-05-03 14:53:45 -07003345/*
3346 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3347 * allocation request woke kswapd for. When kswapd has not woken recently,
3348 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3349 * given classzone and returns it or the highest classzone index kswapd
3350 * was recently woke for.
3351 */
3352static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3353 enum zone_type classzone_idx)
3354{
3355 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3356 return classzone_idx;
3357
3358 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3359}
3360
Mel Gorman38087d92016-07-28 15:45:49 -07003361static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3362 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003363{
3364 long remaining = 0;
3365 DEFINE_WAIT(wait);
3366
3367 if (freezing(current) || kthread_should_stop())
3368 return;
3369
3370 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3371
Shantanu Goel333b0a42017-05-03 14:53:38 -07003372 /*
3373 * Try to sleep for a short interval. Note that kcompactd will only be
3374 * woken if it is possible to sleep for a short interval. This is
3375 * deliberate on the assumption that if reclaim cannot keep an
3376 * eligible zone balanced that it's also unlikely that compaction will
3377 * succeed.
3378 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003379 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003380 /*
3381 * Compaction records what page blocks it recently failed to
3382 * isolate pages from and skips them in the future scanning.
3383 * When kswapd is going to sleep, it is reasonable to assume
3384 * that pages and compaction may succeed so reset the cache.
3385 */
3386 reset_isolation_suitable(pgdat);
3387
3388 /*
3389 * We have freed the memory, now we should compact it to make
3390 * allocation of the requested order possible.
3391 */
Mel Gorman38087d92016-07-28 15:45:49 -07003392 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003393
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003394 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003395
3396 /*
3397 * If woken prematurely then reset kswapd_classzone_idx and
3398 * order. The values will either be from a wakeup request or
3399 * the previous request that slept prematurely.
3400 */
3401 if (remaining) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003402 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003403 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3404 }
3405
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003406 finish_wait(&pgdat->kswapd_wait, &wait);
3407 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3408 }
3409
3410 /*
3411 * After a short sleep, check if it was a premature sleep. If not, then
3412 * go fully to sleep until explicitly woken up.
3413 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003414 if (!remaining &&
3415 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003416 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3417
3418 /*
3419 * vmstat counters are not perfectly accurate and the estimated
3420 * value for counters such as NR_FREE_PAGES can deviate from the
3421 * true value by nr_online_cpus * threshold. To avoid the zone
3422 * watermarks being breached while under pressure, we reduce the
3423 * per-cpu vmstat threshold while kswapd is awake and restore
3424 * them before going back to sleep.
3425 */
3426 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003427
3428 if (!kthread_should_stop())
3429 schedule();
3430
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003431 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3432 } else {
3433 if (remaining)
3434 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3435 else
3436 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3437 }
3438 finish_wait(&pgdat->kswapd_wait, &wait);
3439}
3440
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441/*
3442 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003443 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 *
3445 * This basically trickles out pages so that we have _some_
3446 * free memory available even if there is no other activity
3447 * that frees anything up. This is needed for things like routing
3448 * etc, where we otherwise might have all activity going on in
3449 * asynchronous contexts that cannot page things out.
3450 *
3451 * If there are applications that are active memory-allocators
3452 * (most normal use), this basically shouldn't matter.
3453 */
3454static int kswapd(void *p)
3455{
Mel Gormane716f2e2017-05-03 14:53:45 -07003456 unsigned int alloc_order, reclaim_order;
3457 unsigned int classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 pg_data_t *pgdat = (pg_data_t*)p;
3459 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003460
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 struct reclaim_state reclaim_state = {
3462 .reclaimed_slab = 0,
3463 };
Rusty Russella70f7302009-03-13 14:49:46 +10303464 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
Nick Piggincf40bd12009-01-21 08:12:39 +01003466 lockdep_set_current_reclaim_state(GFP_KERNEL);
3467
Rusty Russell174596a2009-01-01 10:12:29 +10303468 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003469 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 current->reclaim_state = &reclaim_state;
3471
3472 /*
3473 * Tell the memory management that we're a "memory allocator",
3474 * and that if we need more memory we should get access to it
3475 * regardless (see "__alloc_pages()"). "kswapd" should
3476 * never get caught in the normal page freeing logic.
3477 *
3478 * (Kswapd normally doesn't need memory anyway, but sometimes
3479 * you need a small amount of memory in order to be able to
3480 * page out something else, and this flag essentially protects
3481 * us from recursively trying to free more memory as we're
3482 * trying to free the first piece of memory in the first place).
3483 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003484 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003485 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486
Mel Gormane716f2e2017-05-03 14:53:45 -07003487 pgdat->kswapd_order = 0;
3488 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003490 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003491
Mel Gormane716f2e2017-05-03 14:53:45 -07003492 alloc_order = reclaim_order = pgdat->kswapd_order;
3493 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3494
Mel Gorman38087d92016-07-28 15:45:49 -07003495kswapd_try_sleep:
3496 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3497 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003498
Mel Gorman38087d92016-07-28 15:45:49 -07003499 /* Read the new order and classzone_idx */
3500 alloc_order = reclaim_order = pgdat->kswapd_order;
Mel Gormane716f2e2017-05-03 14:53:45 -07003501 classzone_idx = kswapd_classzone_idx(pgdat, 0);
Mel Gorman38087d92016-07-28 15:45:49 -07003502 pgdat->kswapd_order = 0;
Mel Gormane716f2e2017-05-03 14:53:45 -07003503 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504
David Rientjes8fe23e02009-12-14 17:58:33 -08003505 ret = try_to_freeze();
3506 if (kthread_should_stop())
3507 break;
3508
3509 /*
3510 * We can speed up thawing tasks if we don't call balance_pgdat
3511 * after returning from the refrigerator
3512 */
Mel Gorman38087d92016-07-28 15:45:49 -07003513 if (ret)
3514 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003515
Mel Gorman38087d92016-07-28 15:45:49 -07003516 /*
3517 * Reclaim begins at the requested order but if a high-order
3518 * reclaim fails then kswapd falls back to reclaiming for
3519 * order-0. If that happens, kswapd will consider sleeping
3520 * for the order it finished reclaiming at (reclaim_order)
3521 * but kcompactd is woken to compact for the original
3522 * request (alloc_order).
3523 */
Mel Gormane5146b12016-07-28 15:46:47 -07003524 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3525 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003526 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3527 if (reclaim_order < alloc_order)
3528 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003530
Johannes Weiner71abdc12014-06-06 14:35:35 -07003531 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003532 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003533 lockdep_clear_current_reclaim_state();
3534
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 return 0;
3536}
3537
3538/*
3539 * A zone is low on free memory, so wake its kswapd task to service it.
3540 */
Mel Gorman99504742011-01-13 15:46:20 -08003541void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542{
3543 pg_data_t *pgdat;
3544
Mel Gorman6aa303d2016-09-01 16:14:55 -07003545 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546 return;
3547
Vladimir Davydov344736f2014-10-20 15:50:30 +04003548 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003550 pgdat = zone->zone_pgdat;
Mel Gormane716f2e2017-05-03 14:53:45 -07003551 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3552 classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003553 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003554 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003556
Johannes Weinerc73322d2017-05-03 14:51:51 -07003557 /* Hopeless node, leave it to direct reclaim */
3558 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3559 return;
3560
Mel Gormane716f2e2017-05-03 14:53:45 -07003561 if (pgdat_balanced(pgdat, order, classzone_idx))
3562 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003563
Mel Gormane716f2e2017-05-03 14:53:45 -07003564 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003565 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566}
3567
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003568#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003570 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003571 * freed pages.
3572 *
3573 * Rather than trying to age LRUs the aim is to preserve the overall
3574 * LRU order by reclaiming preferentially
3575 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003577unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003579 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003580 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003581 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003582 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003583 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003584 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003585 .may_writepage = 1,
3586 .may_unmap = 1,
3587 .may_swap = 1,
3588 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003590 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003591 struct task_struct *p = current;
3592 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003593 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594
Vlastimil Babka499118e2017-05-08 15:59:50 -07003595 noreclaim_flag = memalloc_noreclaim_save();
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003596 lockdep_set_current_reclaim_state(sc.gfp_mask);
3597 reclaim_state.reclaimed_slab = 0;
3598 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003599
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003600 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003601
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003602 p->reclaim_state = NULL;
3603 lockdep_clear_current_reclaim_state();
Vlastimil Babka499118e2017-05-08 15:59:50 -07003604 memalloc_noreclaim_restore(noreclaim_flag);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003605
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003606 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003608#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610/* It's optimal to keep kswapds on the same CPUs as their memory, but
3611 not required for correctness. So if the last cpu in a node goes
3612 away, we get changed to run anywhere: as the first one comes back,
3613 restore their cpu bindings. */
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003614static int kswapd_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003616 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003618 for_each_node_state(nid, N_MEMORY) {
3619 pg_data_t *pgdat = NODE_DATA(nid);
3620 const struct cpumask *mask;
Rusty Russella70f7302009-03-13 14:49:46 +10303621
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003622 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003623
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003624 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3625 /* One of our CPUs online: restore mask */
3626 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 }
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003628 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630
Yasunori Goto3218ae12006-06-27 02:53:33 -07003631/*
3632 * This kswapd start function will be called by init and node-hot-add.
3633 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3634 */
3635int kswapd_run(int nid)
3636{
3637 pg_data_t *pgdat = NODE_DATA(nid);
3638 int ret = 0;
3639
3640 if (pgdat->kswapd)
3641 return 0;
3642
3643 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3644 if (IS_ERR(pgdat->kswapd)) {
3645 /* failure at boot is fatal */
3646 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003647 pr_err("Failed to start kswapd on node %d\n", nid);
3648 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003649 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003650 }
3651 return ret;
3652}
3653
David Rientjes8fe23e02009-12-14 17:58:33 -08003654/*
Jiang Liud8adde12012-07-11 14:01:52 -07003655 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003656 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003657 */
3658void kswapd_stop(int nid)
3659{
3660 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3661
Jiang Liud8adde12012-07-11 14:01:52 -07003662 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003663 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003664 NODE_DATA(nid)->kswapd = NULL;
3665 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003666}
3667
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668static int __init kswapd_init(void)
3669{
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003670 int nid, ret;
Andrew Morton69e05942006-03-22 00:08:19 -08003671
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003673 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003674 kswapd_run(nid);
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003675 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3676 "mm/vmscan:online", kswapd_cpu_online,
3677 NULL);
3678 WARN_ON(ret < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 return 0;
3680}
3681
3682module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003683
3684#ifdef CONFIG_NUMA
3685/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003686 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003687 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003688 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003689 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003690 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003691int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003692
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003693#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003694#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003695#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003696#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003697
Christoph Lameter9eeff232006-01-18 17:42:31 -08003698/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003699 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003700 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3701 * a zone.
3702 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003703#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003704
Christoph Lameter9eeff232006-01-18 17:42:31 -08003705/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003706 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003707 * occur.
3708 */
3709int sysctl_min_unmapped_ratio = 1;
3710
3711/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003712 * If the number of slab pages in a zone grows beyond this percentage then
3713 * slab reclaim needs to occur.
3714 */
3715int sysctl_min_slab_ratio = 5;
3716
Mel Gorman11fb9982016-07-28 15:46:20 -07003717static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003718{
Mel Gorman11fb9982016-07-28 15:46:20 -07003719 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3720 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3721 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003722
3723 /*
3724 * It's possible for there to be more file mapped pages than
3725 * accounted for by the pages on the file LRU lists because
3726 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3727 */
3728 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3729}
3730
3731/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003732static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003733{
Alexandru Moised031a152015-11-05 18:48:08 -08003734 unsigned long nr_pagecache_reclaimable;
3735 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003736
3737 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003738 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003739 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003740 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003741 * a better estimate
3742 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003743 if (node_reclaim_mode & RECLAIM_UNMAP)
3744 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003745 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003746 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003747
3748 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003749 if (!(node_reclaim_mode & RECLAIM_WRITE))
3750 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003751
3752 /* Watch for any possible underflows due to delta */
3753 if (unlikely(delta > nr_pagecache_reclaimable))
3754 delta = nr_pagecache_reclaimable;
3755
3756 return nr_pagecache_reclaimable - delta;
3757}
3758
Christoph Lameter0ff38492006-09-25 23:31:52 -07003759/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003760 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003761 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003762static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003763{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003764 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003765 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003766 struct task_struct *p = current;
3767 struct reclaim_state reclaim_state;
Mel Gormana5f5f912016-07-28 15:46:32 -07003768 int classzone_idx = gfp_zone(gfp_mask);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003769 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08003770 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003771 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003772 .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003773 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003774 .priority = NODE_RECLAIM_PRIORITY,
3775 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3776 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003777 .may_swap = 1,
Mel Gormana5f5f912016-07-28 15:46:32 -07003778 .reclaim_idx = classzone_idx,
Andrew Morton179e9632006-03-22 00:08:18 -08003779 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003780
Christoph Lameter9eeff232006-01-18 17:42:31 -08003781 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003782 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003783 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003784 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003785 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003786 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07003787 noreclaim_flag = memalloc_noreclaim_save();
3788 p->flags |= PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003789 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003790 reclaim_state.reclaimed_slab = 0;
3791 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003792
Mel Gormana5f5f912016-07-28 15:46:32 -07003793 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003794 /*
3795 * Free memory by calling shrink zone with increasing
3796 * priorities until we have enough memory freed.
3797 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003798 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003799 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003800 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003801 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003802
Christoph Lameter9eeff232006-01-18 17:42:31 -08003803 p->reclaim_state = NULL;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003804 current->flags &= ~PF_SWAPWRITE;
3805 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003806 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003807 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003808}
Andrew Morton179e9632006-03-22 00:08:18 -08003809
Mel Gormana5f5f912016-07-28 15:46:32 -07003810int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003811{
David Rientjesd773ed62007-10-16 23:26:01 -07003812 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003813
3814 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003815 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003816 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003817 *
Christoph Lameter96146342006-07-03 00:24:13 -07003818 * A small portion of unmapped file backed pages is needed for
3819 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003820 * thrown out if the node is overallocated. So we do not reclaim
3821 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003822 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003823 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003824 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3825 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3826 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003827
3828 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003829 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003830 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003831 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003832 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003833
3834 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003835 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003836 * have associated processors. This will favor the local processor
3837 * over remote processors and spread off node memory allocations
3838 * as wide as possible.
3839 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003840 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3841 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003842
Mel Gormana5f5f912016-07-28 15:46:32 -07003843 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3844 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003845
Mel Gormana5f5f912016-07-28 15:46:32 -07003846 ret = __node_reclaim(pgdat, gfp_mask, order);
3847 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003848
Mel Gorman24cf725182009-06-16 15:33:23 -07003849 if (!ret)
3850 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3851
David Rientjesd773ed62007-10-16 23:26:01 -07003852 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003853}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003854#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003855
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003856/*
3857 * page_evictable - test whether a page is evictable
3858 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003859 *
3860 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003861 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003862 *
3863 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003864 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003865 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003866 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003867 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003868int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003869{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003870 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003871}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003872
Hugh Dickins850465792012-01-20 14:34:19 -08003873#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003874/**
Hugh Dickins24513262012-01-20 14:34:21 -08003875 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3876 * @pages: array of pages to check
3877 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003878 *
Hugh Dickins24513262012-01-20 14:34:21 -08003879 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003880 *
3881 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003882 */
Hugh Dickins24513262012-01-20 14:34:21 -08003883void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003884{
Johannes Weiner925b7672012-01-12 17:18:15 -08003885 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003886 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003887 int pgscanned = 0;
3888 int pgrescued = 0;
3889 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003890
Hugh Dickins24513262012-01-20 14:34:21 -08003891 for (i = 0; i < nr_pages; i++) {
3892 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003893 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003894
Hugh Dickins24513262012-01-20 14:34:21 -08003895 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003896 if (pagepgdat != pgdat) {
3897 if (pgdat)
3898 spin_unlock_irq(&pgdat->lru_lock);
3899 pgdat = pagepgdat;
3900 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003901 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003902 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003903
Hugh Dickins24513262012-01-20 14:34:21 -08003904 if (!PageLRU(page) || !PageUnevictable(page))
3905 continue;
3906
Hugh Dickins39b5f292012-10-08 16:33:18 -07003907 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003908 enum lru_list lru = page_lru_base_type(page);
3909
Sasha Levin309381fea2014-01-23 15:52:54 -08003910 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003911 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003912 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3913 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003914 pgrescued++;
3915 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003916 }
Hugh Dickins24513262012-01-20 14:34:21 -08003917
Mel Gorman785b99f2016-07-28 15:47:23 -07003918 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003919 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3920 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07003921 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08003922 }
Hugh Dickins850465792012-01-20 14:34:19 -08003923}
3924#endif /* CONFIG_SHMEM */