blob: b3c44fc17861afdf6767d6fe3beaa1034953c1f1 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080049#include <linux/dax.h>
Johannes Weiner3df0e592018-10-26 15:06:27 -070050#include <linux/psi.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 Weiner14d6ecf2017-02-24 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
Johannes Weiner241994e2015-02-11 15:26:06 -0800100 /* Can cgroups be reclaimed below their normal consumption range? */
101 unsigned int may_thrash:1;
102
Johannes Weineree814fe2014-08-06 16:06:19 -0700103 unsigned int hibernation_mode:1;
104
105 /* One of the zones is ready for compaction */
106 unsigned int compaction_ready:1;
107
108 /* Incremented by the number of inactive pages that were scanned */
109 unsigned long nr_scanned;
110
111 /* Number of pages freed so far during a call to shrink_zones() */
112 unsigned long nr_reclaimed;
Minchan Kimcd256bf2013-05-09 16:21:27 +0900113
114 /*
115 * Reclaim pages from a vma. If the page is shared by other tasks
116 * it is zapped from a vma without reclaim so it ends up remaining
117 * on memory until last task zap it.
118 */
119 struct vm_area_struct *target_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120};
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#ifdef ARCH_HAS_PREFETCH
123#define prefetch_prev_lru_page(_page, _base, _field) \
124 do { \
125 if ((_page)->lru.prev != _base) { \
126 struct page *prev; \
127 \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetch(&prev->_field); \
130 } \
131 } while (0)
132#else
133#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134#endif
135
136#ifdef ARCH_HAS_PREFETCHW
137#define prefetchw_prev_lru_page(_page, _base, _field) \
138 do { \
139 if ((_page)->lru.prev != _base) { \
140 struct page *prev; \
141 \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetchw(&prev->_field); \
144 } \
145 } while (0)
146#else
147#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
148#endif
149
150/*
151 * From 0 .. 100. Higher means more swappy.
152 */
153int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700154/*
155 * The total number of pages which are beyond the high watermark within all
156 * zones.
157 */
158unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
160static LIST_HEAD(shrinker_list);
161static DECLARE_RWSEM(shrinker_rwsem);
162
Andrew Mortonc255a452012-07-31 16:43:02 -0700163#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800164static bool global_reclaim(struct scan_control *sc)
165{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800166 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800167}
Tejun Heo97c93412015-05-22 18:23:36 -0400168
169/**
170 * sane_reclaim - is the usual dirty throttling mechanism operational?
171 * @sc: scan_control in question
172 *
173 * The normal page dirty throttling mechanism in balance_dirty_pages() is
174 * completely broken with the legacy memcg and direct stalling in
175 * shrink_page_list() is used for throttling instead, which lacks all the
176 * niceties such as fairness, adaptive pausing, bandwidth proportional
177 * allocation and configurability.
178 *
179 * This function tests whether the vmscan currently in progress can assume
180 * that the normal dirty throttling mechanism is operational.
181 */
182static bool sane_reclaim(struct scan_control *sc)
183{
184 struct mem_cgroup *memcg = sc->target_mem_cgroup;
185
186 if (!memcg)
187 return true;
188#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800189 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400190 return true;
191#endif
192 return false;
193}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800194#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800195static bool global_reclaim(struct scan_control *sc)
196{
197 return true;
198}
Tejun Heo97c93412015-05-22 18:23:36 -0400199
200static bool sane_reclaim(struct scan_control *sc)
201{
202 return true;
203}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800204#endif
205
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700206/*
207 * This misses isolated pages which are not accounted for to save counters.
208 * As the data only determines if reclaim or compaction continues, it is
209 * not expected that isolated pages will be a dominating factor.
210 */
211unsigned long zone_reclaimable_pages(struct zone *zone)
212{
213 unsigned long nr;
214
215 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
216 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
Vinayak Menon3771b882018-02-09 18:24:58 +0530217 if (get_nr_swap_pages() > 0
218 || IS_ENABLED(CONFIG_ANDROID_LOW_MEMORY_KILLER))
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700219 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
220 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
221
222 return nr;
223}
224
Mel Gorman599d0c92016-07-28 15:45:31 -0700225unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
Lisa Du6e543d52013-09-11 14:22:36 -0700226{
Mel Gorman599d0c92016-07-28 15:45:31 -0700227 unsigned long nr;
228
229 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
230 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
231 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
232
233 if (get_nr_swap_pages() > 0)
234 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
235 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
236 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
237
238 return nr;
239}
240
Michal Hocko71053132017-02-22 15:45:58 -0800241/**
242 * lruvec_lru_size - Returns the number of pages on the given LRU list.
243 * @lruvec: lru vector
244 * @lru: lru to use
245 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
246 */
247unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800248{
Michal Hocko71053132017-02-22 15:45:58 -0800249 unsigned long lru_size;
250 int zid;
251
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700252 if (!mem_cgroup_disabled())
Michal Hocko71053132017-02-22 15:45:58 -0800253 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
254 else
255 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800256
Michal Hocko71053132017-02-22 15:45:58 -0800257 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
258 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
259 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800260
Michal Hocko71053132017-02-22 15:45:58 -0800261 if (!managed_zone(zone))
262 continue;
Michal Hocko07fc9572017-01-10 16:58:04 -0800263
Michal Hocko71053132017-02-22 15:45:58 -0800264 if (!mem_cgroup_disabled())
265 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
266 else
267 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
268 NR_ZONE_LRU_BASE + lru);
269 lru_size -= min(size, lru_size);
270 }
271
272 return lru_size;
273
Michal Hocko07fc9572017-01-10 16:58:04 -0800274}
275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000277 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000279int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000281 size_t size = sizeof(*shrinker->nr_deferred);
282
Glauber Costa1d3d4432013-08-28 10:18:04 +1000283 if (shrinker->flags & SHRINKER_NUMA_AWARE)
284 size *= nr_node_ids;
285
286 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
287 if (!shrinker->nr_deferred)
288 return -ENOMEM;
289
Rusty Russell8e1f9362007-07-17 04:03:17 -0700290 down_write(&shrinker_rwsem);
291 list_add_tail(&shrinker->list, &shrinker_list);
292 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000293 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700295EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297/*
298 * Remove one
299 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700300void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301{
Tetsuo Handa274ee932017-12-18 20:31:41 +0900302 if (!shrinker->nr_deferred)
303 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 down_write(&shrinker_rwsem);
305 list_del(&shrinker->list);
306 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700307 kfree(shrinker->nr_deferred);
Tetsuo Handa274ee932017-12-18 20:31:41 +0900308 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700310EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
312#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000313
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800314static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
315 struct shrinker *shrinker,
316 unsigned long nr_scanned,
317 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000319 unsigned long freed = 0;
320 unsigned long long delta;
321 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700322 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000323 long nr;
324 long new_nr;
325 int nid = shrinkctl->nid;
326 long batch_size = shrinker->batch ? shrinker->batch
327 : SHRINK_BATCH;
Shaohua Li04597be2016-12-12 16:41:50 -0800328 long scanned = 0, next_deferred;
Liam Marka708a022014-11-25 14:43:42 -0800329 long min_cache_size = batch_size;
330
331 if (current_is_kswapd())
332 min_cache_size = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000333
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700334 freeable = shrinker->count_objects(shrinker, shrinkctl);
335 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000336 return 0;
337
338 /*
339 * copy the current shrinker scan count into a local variable
340 * and zero it so that other concurrent shrinker invocations
341 * don't also do this scanning work.
342 */
343 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
344
345 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800346 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700347 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800348 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000349 total_scan += delta;
350 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800351 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000352 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700353 total_scan = freeable;
Shaohua Li04597be2016-12-12 16:41:50 -0800354 next_deferred = nr;
355 } else
356 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000357
358 /*
359 * We need to avoid excessive windup on filesystem shrinkers
360 * due to large numbers of GFP_NOFS allocations causing the
361 * shrinkers to return -1 all the time. This results in a large
362 * nr being built up so when a shrink that can do some work
363 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700364 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000365 * memory.
366 *
367 * Hence only allow the shrinker to scan the entire cache when
368 * a large delta change is calculated directly.
369 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700370 if (delta < freeable / 4)
371 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000372
373 /*
374 * Avoid risking looping forever due to too large nr value:
375 * never try to free more than twice the estimate number of
376 * freeable entries.
377 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700378 if (total_scan > freeable * 2)
379 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000380
381 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800382 nr_scanned, nr_eligible,
383 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000384
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800385 /*
386 * Normally, we should not scan less than batch_size objects in one
387 * pass to avoid too frequent shrinker calls, but if the slab has less
388 * than batch_size objects in total and we are really tight on memory,
389 * we will try to reclaim all available objects, otherwise we can end
390 * up failing allocations although there are plenty of reclaimable
391 * objects spread over several slabs with usage less than the
392 * batch_size.
393 *
394 * We detect the "tight on memory" situations by looking at the total
395 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700396 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800397 * scanning at high prio and therefore should try to reclaim as much as
398 * possible.
399 */
Liam Marka708a022014-11-25 14:43:42 -0800400 while (total_scan > min_cache_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700401 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000402 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800403 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000404
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800405 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000406 ret = shrinker->scan_objects(shrinker, shrinkctl);
407 if (ret == SHRINK_STOP)
408 break;
409 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000410
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800411 count_vm_events(SLABS_SCANNED, nr_to_scan);
412 total_scan -= nr_to_scan;
Shaohua Li04597be2016-12-12 16:41:50 -0800413 scanned += nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000414
415 cond_resched();
416 }
417
Shaohua Li04597be2016-12-12 16:41:50 -0800418 if (next_deferred >= scanned)
419 next_deferred -= scanned;
420 else
421 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000422 /*
423 * move the unused scan count back into the shrinker in a
424 * manner that handles concurrent updates. If we exhausted the
425 * scan, there is no need to do an update.
426 */
Shaohua Li04597be2016-12-12 16:41:50 -0800427 if (next_deferred > 0)
428 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000429 &shrinker->nr_deferred[nid]);
430 else
431 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
432
Dave Hansendf9024a2014-06-04 16:08:07 -0700433 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000434 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435}
436
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800437/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800438 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800439 * @gfp_mask: allocation context
440 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800441 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800442 * @nr_scanned: pressure numerator
443 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800445 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800447 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
448 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800450 * @memcg specifies the memory cgroup to target. If it is not NULL,
451 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700452 * objects from the memory cgroup specified. Otherwise, only unaware
453 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800454 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800455 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
456 * the available objects should be scanned. Page reclaim for example
457 * passes the number of pages scanned and the number of pages on the
458 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
459 * when it encountered mapped pages. The ratio is further biased by
460 * the ->seeks setting of the shrink function, which indicates the
461 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800463 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800465static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
466 struct mem_cgroup *memcg,
467 unsigned long nr_scanned,
468 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
470 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000471 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700473 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800474 return 0;
475
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800476 if (nr_scanned == 0)
477 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
Minchan Kimf06590b2011-05-24 17:11:11 -0700479 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000480 /*
481 * If we would return 0, our callers would understand that we
482 * have nothing else to shrink and give up trying. By returning
483 * 1 we keep it going and assume we'll be able to shrink next
484 * time.
485 */
486 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700487 goto out;
488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
490 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800491 struct shrink_control sc = {
492 .gfp_mask = gfp_mask,
493 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800494 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800495 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800496
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700497 /*
498 * If kernel memory accounting is disabled, we ignore
499 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
500 * passing NULL for memcg.
501 */
502 if (memcg_kmem_enabled() &&
503 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800504 continue;
505
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800506 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
507 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000508
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800509 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800511
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700513out:
514 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000515 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516}
517
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800518void drop_slab_node(int nid)
519{
520 unsigned long freed;
521
522 do {
523 struct mem_cgroup *memcg = NULL;
524
525 freed = 0;
526 do {
527 freed += shrink_slab(GFP_KERNEL, nid, memcg,
528 1000, 1000);
529 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
530 } while (freed > 10);
531}
532
533void drop_slab(void)
534{
535 int nid;
536
537 for_each_online_node(nid)
538 drop_slab_node(nid);
539}
540
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541static inline int is_page_cache_freeable(struct page *page)
542{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700543 /*
544 * A freeable page cache page is referenced only by the caller
545 * that isolated the page, the page cache radix tree and
546 * optional buffer heads at page->private.
547 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700548 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
Tejun Heo703c2702015-05-22 17:13:44 -0400551static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552{
Christoph Lameter930d9152006-01-08 01:00:47 -0800553 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400555 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400557 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 return 1;
559 return 0;
560}
561
562/*
563 * We detected a synchronous write error writing a page out. Probably
564 * -ENOSPC. We need to propagate that into the address_space for a subsequent
565 * fsync(), msync() or close().
566 *
567 * The tricky part is that after writepage we cannot touch the mapping: nothing
568 * prevents it from being freed up. But we have a ref on the page and once
569 * that page is locked, the mapping is pinned.
570 *
571 * We're allowed to run sleeping lock_page() here because we know the caller has
572 * __GFP_FS.
573 */
574static void handle_write_error(struct address_space *mapping,
575 struct page *page, int error)
576{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100577 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700578 if (page_mapping(page) == mapping)
579 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 unlock_page(page);
581}
582
Christoph Lameter04e62a22006-06-23 02:03:38 -0700583/* possible outcome of pageout() */
584typedef enum {
585 /* failed to write page out, page is locked */
586 PAGE_KEEP,
587 /* move page to the active list, page is locked */
588 PAGE_ACTIVATE,
589 /* page has been sent to the disk successfully, page is unlocked */
590 PAGE_SUCCESS,
591 /* page is clean and locked */
592 PAGE_CLEAN,
593} pageout_t;
594
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595/*
Andrew Morton1742f192006-03-22 00:08:21 -0800596 * pageout is called by shrink_page_list() for each dirty page.
597 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700599static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700600 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601{
602 /*
603 * If the page is dirty, only perform writeback if that write
604 * will be non-blocking. To prevent this allocation from being
605 * stalled by pagecache activity. But note that there may be
606 * stalls if we need to run get_block(). We could test
607 * PagePrivate for that.
608 *
Al Viro81742022014-04-03 03:17:43 -0400609 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 * this page's queue, we can perform writeback even if that
611 * will block.
612 *
613 * If the page is swapcache, write it back even if that would
614 * block, for some throttling. This happens by accident, because
615 * swap_backing_dev_info is bust: it doesn't reflect the
616 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 */
618 if (!is_page_cache_freeable(page))
619 return PAGE_KEEP;
620 if (!mapping) {
621 /*
622 * Some data journaling orphaned pages can have
623 * page->mapping == NULL while being dirty with clean buffers.
624 */
David Howells266cf652009-04-03 16:42:36 +0100625 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 if (try_to_free_buffers(page)) {
627 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700628 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 return PAGE_CLEAN;
630 }
631 }
632 return PAGE_KEEP;
633 }
634 if (mapping->a_ops->writepage == NULL)
635 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400636 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 return PAGE_KEEP;
638
639 if (clear_page_dirty_for_io(page)) {
640 int res;
641 struct writeback_control wbc = {
642 .sync_mode = WB_SYNC_NONE,
643 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700644 .range_start = 0,
645 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 .for_reclaim = 1,
647 };
648
649 SetPageReclaim(page);
650 res = mapping->a_ops->writepage(page, &wbc);
651 if (res < 0)
652 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800653 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 ClearPageReclaim(page);
655 return PAGE_ACTIVATE;
656 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 if (!PageWriteback(page)) {
659 /* synchronous write or broken a_ops? */
660 ClearPageReclaim(page);
661 }
yalin wang3aa23852016-01-14 15:18:30 -0800662 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700663 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 return PAGE_SUCCESS;
665 }
666
667 return PAGE_CLEAN;
668}
669
Andrew Mortona649fd92006-10-17 00:09:36 -0700670/*
Nick Piggine2867812008-07-25 19:45:30 -0700671 * Same as remove_mapping, but if the page is removed from the mapping, it
672 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700673 */
Johannes Weinera5289102014-04-03 14:47:51 -0700674static int __remove_mapping(struct address_space *mapping, struct page *page,
675 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800676{
Greg Thelenc4843a72015-05-22 17:13:16 -0400677 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400678
Nick Piggin28e4d962006-09-25 23:31:23 -0700679 BUG_ON(!PageLocked(page));
680 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800681
Greg Thelenc4843a72015-05-22 17:13:16 -0400682 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800683 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700684 * The non racy check for a busy page.
685 *
686 * Must be careful with the order of the tests. When someone has
687 * a ref to the page, it may be possible that they dirty it then
688 * drop the reference. So if PageDirty is tested before page_count
689 * here, then the following race may occur:
690 *
691 * get_user_pages(&page);
692 * [user mapping goes away]
693 * write_to(page);
694 * !PageDirty(page) [good]
695 * SetPageDirty(page);
696 * put_page(page);
697 * !page_count(page) [good, discard it]
698 *
699 * [oops, our write_to data is lost]
700 *
701 * Reversing the order of the tests ensures such a situation cannot
702 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700703 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700704 *
705 * Note that if SetPageDirty is always performed via set_page_dirty,
706 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800707 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700708 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800709 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700710 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
711 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700712 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800713 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700714 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800715
716 if (PageSwapCache(page)) {
717 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700718 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800719 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400720 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700721 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700722 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500723 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700724 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500725
726 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700727 /*
728 * Remember a shadow entry for reclaimed file cache in
729 * order to detect refaults, thus thrashing, later on.
730 *
731 * But don't store shadows in an address space that is
732 * already exiting. This is not just an optizimation,
733 * inode reclaim needs to empty out the radix tree or
734 * the nodes are lost. Don't plant shadows behind its
735 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800736 *
737 * We also don't store shadows for DAX mappings because the
738 * only page cache pages found in these are zero pages
739 * covering holes, and because we don't want to mix DAX
740 * exceptional entries and shadow exceptional entries in the
741 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700742 */
743 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800744 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700745 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700746 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400747 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500748
749 if (freepage != NULL)
750 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800751 }
752
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800753 return 1;
754
755cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400756 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800757 return 0;
758}
759
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760/*
Nick Piggine2867812008-07-25 19:45:30 -0700761 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
762 * someone else has a ref on the page, abort and return 0. If it was
763 * successfully detached, return 1. Assumes the caller has a single ref on
764 * this page.
765 */
766int remove_mapping(struct address_space *mapping, struct page *page)
767{
Johannes Weinera5289102014-04-03 14:47:51 -0700768 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700769 /*
770 * Unfreezing the refcount with 1 rather than 2 effectively
771 * drops the pagecache ref for us without requiring another
772 * atomic operation.
773 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700774 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700775 return 1;
776 }
777 return 0;
778}
779
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700780/**
781 * putback_lru_page - put previously isolated page onto appropriate LRU list
782 * @page: page to be put back to appropriate lru list
783 *
784 * Add previously isolated @page to appropriate LRU list.
785 * Page may still be unevictable for other reasons.
786 *
787 * lru_lock must not be held, interrupts must be enabled.
788 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700789void putback_lru_page(struct page *page)
790{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700791 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700792 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700793
Sasha Levin309381fea2014-01-23 15:52:54 -0800794 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700795
796redo:
797 ClearPageUnevictable(page);
798
Hugh Dickins39b5f292012-10-08 16:33:18 -0700799 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700800 /*
801 * For evictable pages, we can use the cache.
802 * In event of a race, worst case is we end up with an
803 * unevictable page on [in]active list.
804 * We know how to handle that.
805 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700806 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700807 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700808 } else {
809 /*
810 * Put unevictable pages directly on zone's unevictable
811 * list.
812 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700813 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700814 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700815 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700816 * When racing with an mlock or AS_UNEVICTABLE clearing
817 * (page is unlocked) make sure that if the other thread
818 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800819 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700820 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700821 * the page back to the evictable list.
822 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700823 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700824 */
825 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700826 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700827
828 /*
829 * page's status can change while we move it among lru. If an evictable
830 * page is on unevictable list, it never be freed. To avoid that,
831 * check after we added it to the list, again.
832 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700833 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700834 if (!isolate_lru_page(page)) {
835 put_page(page);
836 goto redo;
837 }
838 /* This means someone else dropped this page from LRU
839 * So, it will be freed or putback to LRU again. There is
840 * nothing to do here.
841 */
842 }
843
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700844 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700845 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700846 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700847 count_vm_event(UNEVICTABLE_PGCULLED);
848
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700849 put_page(page); /* drop ref from isolate */
850}
851
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800852enum page_references {
853 PAGEREF_RECLAIM,
854 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800855 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800856 PAGEREF_ACTIVATE,
857};
858
859static enum page_references page_check_references(struct page *page,
860 struct scan_control *sc)
861{
Johannes Weiner645747462010-03-05 13:42:22 -0800862 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800863 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800864
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700865 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
866 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800867 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800868
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800869 /*
870 * Mlock lost the isolation race with us. Let try_to_unmap()
871 * move the page to the unevictable list.
872 */
873 if (vm_flags & VM_LOCKED)
874 return PAGEREF_RECLAIM;
875
Johannes Weiner645747462010-03-05 13:42:22 -0800876 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700877 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800878 return PAGEREF_ACTIVATE;
879 /*
880 * All mapped pages start out with page table
881 * references from the instantiating fault, so we need
882 * to look twice if a mapped file page is used more
883 * than once.
884 *
885 * Mark it and spare it for another trip around the
886 * inactive list. Another page table reference will
887 * lead to its activation.
888 *
889 * Note: the mark is set for activated pages as well
890 * so that recently deactivated but used pages are
891 * quickly recovered.
892 */
893 SetPageReferenced(page);
894
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800895 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800896 return PAGEREF_ACTIVATE;
897
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800898 /*
899 * Activate file-backed executable pages after first usage.
900 */
901 if (vm_flags & VM_EXEC)
902 return PAGEREF_ACTIVATE;
903
Johannes Weiner645747462010-03-05 13:42:22 -0800904 return PAGEREF_KEEP;
905 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800906
907 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700908 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800909 return PAGEREF_RECLAIM_CLEAN;
910
911 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800912}
913
Mel Gormane2be15f2013-07-03 15:01:57 -0700914/* Check if a page is dirty or under writeback */
915static void page_check_dirty_writeback(struct page *page,
916 bool *dirty, bool *writeback)
917{
Mel Gormanb4597222013-07-03 15:02:05 -0700918 struct address_space *mapping;
919
Mel Gormane2be15f2013-07-03 15:01:57 -0700920 /*
921 * Anonymous pages are not handled by flushers and must be written
922 * from reclaim context. Do not stall reclaim based on them
923 */
924 if (!page_is_file_cache(page)) {
925 *dirty = false;
926 *writeback = false;
927 return;
928 }
929
930 /* By default assume that the page flags are accurate */
931 *dirty = PageDirty(page);
932 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700933
934 /* Verify dirty/writeback state if the filesystem supports it */
935 if (!page_has_private(page))
936 return;
937
938 mapping = page_mapping(page);
939 if (mapping && mapping->a_ops->is_dirty_writeback)
940 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700941}
942
Nick Piggine2867812008-07-25 19:45:30 -0700943/*
Andrew Morton1742f192006-03-22 00:08:21 -0800944 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 */
Andrew Morton1742f192006-03-22 00:08:21 -0800946static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -0700947 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700948 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700949 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700950 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700951 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700952 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700953 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700954 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700955 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956{
957 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700958 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700960 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700961 unsigned long nr_dirty = 0;
962 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800963 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700964 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700965 unsigned long nr_immediate = 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 Kim13e7b912015-01-09 18:36:55 +0530973 enum page_references references = PAGEREF_RECLAIM;
Mel Gormane2be15f2013-07-03 15:01:57 -0700974 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800975 bool lazyfree = false;
976 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
978 cond_resched();
979
980 page = lru_to_page(page_list);
981 list_del(&page->lru);
982
Nick Piggin529ae9a2008-08-02 12:01:03 +0200983 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 goto keep;
985
Sasha Levin309381fea2014-01-23 15:52:54 -0800986 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kim07e6c8d2013-05-09 16:21:25 +0900987 if (pgdat)
988 VM_BUG_ON_PAGE(page_pgdat(page) != pgdat, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989
990 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800991
Hugh Dickins39b5f292012-10-08 16:33:18 -0700992 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700993 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700994
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700995 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800996 goto keep_locked;
997
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 /* Double the slab pressure for mapped and swapcache pages */
999 if (page_mapped(page) || PageSwapCache(page))
1000 sc->nr_scanned++;
1001
Andy Whitcroftc661b072007-08-22 14:01:26 -07001002 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1003 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1004
Mel Gorman283aba92013-07-03 15:01:51 -07001005 /*
Mel Gormane2be15f2013-07-03 15:01:57 -07001006 * The number of dirty pages determines if a zone is marked
1007 * reclaim_congested which affects wait_iff_congested. kswapd
1008 * will stall and start writing pages if the tail of the LRU
1009 * is all dirty unqueued pages.
1010 */
1011 page_check_dirty_writeback(page, &dirty, &writeback);
1012 if (dirty || writeback)
1013 nr_dirty++;
1014
1015 if (dirty && !writeback)
1016 nr_unqueued_dirty++;
1017
Mel Gormand04e8ac2013-07-03 15:02:03 -07001018 /*
1019 * Treat this page as congested if the underlying BDI is or if
1020 * pages are cycling through the LRU so quickly that the
1021 * pages marked for immediate reclaim are making it to the
1022 * end of the LRU a second time.
1023 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001024 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001025 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001026 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001027 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001028 nr_congested++;
1029
1030 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001031 * If a page at the tail of the LRU is under writeback, there
1032 * are three cases to consider.
1033 *
1034 * 1) If reclaim is encountering an excessive number of pages
1035 * under writeback and this page is both under writeback and
1036 * PageReclaim then it indicates that pages are being queued
1037 * for IO but are being recycled through the LRU before the
1038 * IO can complete. Waiting on the page itself risks an
1039 * indefinite stall if it is impossible to writeback the
1040 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001041 * note that the LRU is being scanned too quickly and the
1042 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001043 *
Tejun Heo97c93412015-05-22 18:23:36 -04001044 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001045 * not marked for immediate reclaim, or the caller does not
1046 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1047 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001048 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001049 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001050 * Require may_enter_fs because we would wait on fs, which
1051 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001052 * enter reclaim, and deadlock if it waits on a page for
1053 * which it is needed to do the write (loop masks off
1054 * __GFP_IO|__GFP_FS for this reason); but more thought
1055 * would probably show more reasons.
1056 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001057 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001058 * PageReclaim. memcg does not have any dirty pages
1059 * throttling so we could easily OOM just because too many
1060 * pages are in writeback and there is nothing else to
1061 * reclaim. Wait for the writeback to complete.
Johannes Weiner6b25f412017-02-24 14:56:23 -08001062 *
1063 * In cases 1) and 2) we activate the pages to get them out of
1064 * the way while we continue scanning for clean pages on the
1065 * inactive list and refilling from the active list. The
1066 * observation here is that waiting for disk writes is more
1067 * expensive than potentially causing reloads down the line.
1068 * Since they're marked for immediate reclaim, they won't put
1069 * memory pressure on the cache working set any longer than it
1070 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001071 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001072 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001073 /* Case 1 above */
1074 if (current_is_kswapd() &&
1075 PageReclaim(page) &&
Minchan Kim07e6c8d2013-05-09 16:21:25 +09001076 (pgdat && test_bit(PGDAT_WRITEBACK, &pgdat->flags))) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001077 nr_immediate++;
Johannes Weiner6b25f412017-02-24 14:56:23 -08001078 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001079
1080 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001081 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001082 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001083 /*
1084 * This is slightly racy - end_page_writeback()
1085 * might have just cleared PageReclaim, then
1086 * setting PageReclaim here end up interpreted
1087 * as PageReadahead - but that does not matter
1088 * enough to care. What we do want is for this
1089 * page to have PageReclaim set next time memcg
1090 * reclaim reaches the tests above, so it will
1091 * then wait_on_page_writeback() to avoid OOM;
1092 * and it's also appropriate in global reclaim.
1093 */
1094 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001095 nr_writeback++;
Johannes Weiner6b25f412017-02-24 14:56:23 -08001096 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001097
1098 /* Case 3 above */
1099 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001100 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001101 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001102 /* then go back and try same page again */
1103 list_add_tail(&page->lru, page_list);
1104 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001105 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107
Minchan Kim02c6de82012-10-08 16:31:55 -07001108 if (!force_reclaim)
1109 references = page_check_references(page, sc);
1110
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001111 switch (references) {
1112 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001114 case PAGEREF_KEEP:
1115 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001116 case PAGEREF_RECLAIM:
1117 case PAGEREF_RECLAIM_CLEAN:
1118 ; /* try to reclaim the page below */
1119 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 /*
1122 * Anonymous process memory has backing store?
1123 * Try to allocate it some swap space here.
1124 */
Nick Pigginb291f002008-10-18 20:26:44 -07001125 if (PageAnon(page) && !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;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001130 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001131 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Mel Gormane2be15f2013-07-03 15:01:57 -07001133 /* Adding to swap updated mapping */
1134 mapping = page_mapping(page);
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001135 } else if (unlikely(PageTransHuge(page))) {
1136 /* Split file THP */
1137 if (split_huge_page_to_list(page, page_list))
1138 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001139 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001141 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1142
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 /*
1144 * The page is mapped into the page tables of one or more
1145 * processes. Try to unmap it here.
1146 */
1147 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001148 switch (ret = try_to_unmap(page, lazyfree ?
1149 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
Minchan Kimcd256bf2013-05-09 16:21:27 +09001150 (ttu_flags | TTU_BATCH_FLUSH),
1151 sc->target_vma)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 case SWAP_FAIL:
1153 goto activate_locked;
1154 case SWAP_AGAIN:
1155 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001156 case SWAP_MLOCK:
1157 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001158 case SWAP_LZFREE:
1159 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 case SWAP_SUCCESS:
1161 ; /* try to free the page below */
1162 }
1163 }
1164
1165 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001166 /*
Johannes Weiner319400b2017-02-24 14:56:20 -08001167 * Only kswapd can writeback filesystem pages
1168 * to avoid risk of stack overflow. But avoid
1169 * injecting inefficient single-page IO into
1170 * flusher writeback as much as possible: only
1171 * write pages when we've encountered many
1172 * dirty pages, and when we've already scanned
1173 * the rest of the LRU for clean pages and see
1174 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001175 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001176 if (page_is_file_cache(page) &&
Johannes Weiner319400b2017-02-24 14:56:20 -08001177 (!current_is_kswapd() || !PageReclaim(page) ||
1178 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001179 /*
1180 * Immediately reclaim when written back.
1181 * Similar in principal to deactivate_page()
1182 * except we already have the page isolated
1183 * and know it's dirty
1184 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001185 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001186 SetPageReclaim(page);
1187
Johannes Weiner6b25f412017-02-24 14:56:23 -08001188 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001189 }
1190
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001191 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001193 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001195 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 goto keep_locked;
1197
Mel Gormand950c942015-09-04 15:47:35 -07001198 /*
1199 * Page is dirty. Flush the TLB if a writable entry
1200 * potentially exists to avoid CPU writes after IO
1201 * starts and then write it out here.
1202 */
1203 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001204 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 case PAGE_KEEP:
1206 goto keep_locked;
1207 case PAGE_ACTIVATE:
1208 goto activate_locked;
1209 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001210 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001211 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001212 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001214
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 /*
1216 * A synchronous write - probably a ramdisk. Go
1217 * ahead and try to reclaim the page.
1218 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001219 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 goto keep;
1221 if (PageDirty(page) || PageWriteback(page))
1222 goto keep_locked;
1223 mapping = page_mapping(page);
1224 case PAGE_CLEAN:
1225 ; /* try to free the page below */
1226 }
1227 }
1228
1229 /*
1230 * If the page has buffers, try to free the buffer mappings
1231 * associated with this page. If we succeed we try to free
1232 * the page as well.
1233 *
1234 * We do this even if the page is PageDirty().
1235 * try_to_release_page() does not perform I/O, but it is
1236 * possible for a page to have PageDirty set, but it is actually
1237 * clean (all its buffers are clean). This happens if the
1238 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001239 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 * try_to_release_page() will discover that cleanness and will
1241 * drop the buffers and mark the page clean - it can be freed.
1242 *
1243 * Rarely, pages can have buffers and no ->mapping. These are
1244 * the pages which were not successfully invalidated in
1245 * truncate_complete_page(). We try to drop those buffers here
1246 * and if that worked, and the page is no longer mapped into
1247 * process address space (page_count == 1) it can be freed.
1248 * Otherwise, leave the page on the LRU so it is swappable.
1249 */
David Howells266cf652009-04-03 16:42:36 +01001250 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 if (!try_to_release_page(page, sc->gfp_mask))
1252 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001253 if (!mapping && page_count(page) == 1) {
1254 unlock_page(page);
1255 if (put_page_testzero(page))
1256 goto free_it;
1257 else {
1258 /*
1259 * rare race with speculative reference.
1260 * the speculative reference will free
1261 * this page shortly, so we may
1262 * increment nr_reclaimed here (and
1263 * leave it off the LRU).
1264 */
1265 nr_reclaimed++;
1266 continue;
1267 }
1268 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 }
1270
Minchan Kim854e9ed2016-01-15 16:54:53 -08001271lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001272 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001273 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274
Nick Piggina978d6f2008-10-18 20:26:58 -07001275 /*
1276 * At this point, we have no other references and there is
1277 * no way to pick any more up (removed from LRU, removed
1278 * from pagecache). Can use non-atomic bitops now (and
1279 * we obviously don't have to worry about waking up a process
1280 * waiting on the page lock, because there are no references.
1281 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001282 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001283free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001284 if (ret == SWAP_LZFREE)
1285 count_vm_event(PGLAZYFREED);
1286
Andrew Morton05ff5132006-03-22 00:08:20 -08001287 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001288
1289 /*
1290 * Is there need to periodically free_page_list? It would
1291 * appear not as the counts should be low
1292 */
1293 list_add(&page->lru, &free_pages);
Minchan Kim74ac8822013-05-09 16:21:26 +09001294 /*
1295 * If pagelist are from multiple zones, we should decrease
1296 * NR_ISOLATED_ANON + x on freed pages in here.
1297 */
1298 if (!pgdat)
1299 dec_node_page_state(page, NR_ISOLATED_ANON +
1300 page_is_file_cache(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 continue;
1302
Nick Pigginb291f002008-10-18 20:26:44 -07001303cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001304 if (PageSwapCache(page))
1305 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001306 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001307 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001308 continue;
1309
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001311 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001312 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001313 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001314 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 SetPageActive(page);
1316 pgactivate++;
1317keep_locked:
1318 unlock_page(page);
1319keep:
1320 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001321 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001323
Johannes Weiner747db952014-08-08 14:19:24 -07001324 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001325 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001326 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001327
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001329 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001330
Mel Gorman8e950282013-07-03 15:02:02 -07001331 *ret_nr_dirty += nr_dirty;
1332 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001333 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001334 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001335 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001336 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337}
1338
Minchan Kim02c6de82012-10-08 16:31:55 -07001339unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1340 struct list_head *page_list)
1341{
1342 struct scan_control sc = {
1343 .gfp_mask = GFP_KERNEL,
1344 .priority = DEF_PRIORITY,
1345 .may_unmap = 1,
Minchan Kim13e7b912015-01-09 18:36:55 +05301346 /* Doesn't allow to write out dirty page */
1347 .may_writepage = 0,
Minchan Kim02c6de82012-10-08 16:31:55 -07001348 };
Mel Gorman8e950282013-07-03 15:02:02 -07001349 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001350 struct page *page, *next;
1351 LIST_HEAD(clean_pages);
1352
1353 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001354 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kimb1123ea62016-07-26 15:23:09 -07001355 !__PageMovable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001356 ClearPageActive(page);
1357 list_move(&page->lru, &clean_pages);
1358 }
1359 }
1360
Mel Gorman599d0c92016-07-28 15:45:31 -07001361 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001362 TTU_UNMAP|TTU_IGNORE_ACCESS,
1363 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001364 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001365 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001366 return ret;
1367}
1368
Minchan Kimff926892013-05-09 16:21:24 +09001369#ifdef CONFIG_PROCESS_RECLAIM
Minchan Kimcd256bf2013-05-09 16:21:27 +09001370unsigned long reclaim_pages_from_list(struct list_head *page_list,
1371 struct vm_area_struct *vma)
Minchan Kimff926892013-05-09 16:21:24 +09001372{
1373 struct scan_control sc = {
1374 .gfp_mask = GFP_KERNEL,
1375 .priority = DEF_PRIORITY,
1376 .may_writepage = 1,
1377 .may_unmap = 1,
1378 .may_swap = 1,
Minchan Kimcd256bf2013-05-09 16:21:27 +09001379 .target_vma = vma,
Minchan Kimff926892013-05-09 16:21:24 +09001380 };
1381
Minchan Kim74ac8822013-05-09 16:21:26 +09001382 unsigned long nr_reclaimed;
Minchan Kimff926892013-05-09 16:21:24 +09001383 struct page *page;
1384 unsigned long dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim74ac8822013-05-09 16:21:26 +09001385
1386 list_for_each_entry(page, page_list, lru)
1387 ClearPageActive(page);
1388
1389 nr_reclaimed = shrink_page_list(page_list, NULL, &sc,
1390 TTU_UNMAP|TTU_IGNORE_ACCESS,
1391 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kimff926892013-05-09 16:21:24 +09001392
1393 while (!list_empty(page_list)) {
1394 page = lru_to_page(page_list);
1395 list_del(&page->lru);
Minchan Kimff926892013-05-09 16:21:24 +09001396 dec_node_page_state(page, NR_ISOLATED_ANON +
1397 page_is_file_cache(page));
1398 putback_lru_page(page);
1399 }
1400
1401 return nr_reclaimed;
1402}
1403#endif
1404
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001405/*
1406 * Attempt to remove the specified page from its LRU. Only take this page
1407 * if it is of the appropriate PageActive status. Pages which are being
1408 * freed elsewhere are also ignored.
1409 *
1410 * page: page to consider
1411 * mode: one of the LRU isolation modes defined above
1412 *
1413 * returns 0 on success, -ve errno on failure.
1414 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001415int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001416{
1417 int ret = -EINVAL;
1418
1419 /* Only take pages on the LRU. */
1420 if (!PageLRU(page))
1421 return ret;
1422
Minchan Kime46a2872012-10-08 16:33:48 -07001423 /* Compaction should not handle unevictable pages but CMA can do so */
1424 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001425 return ret;
1426
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001427 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001428
Mel Gormanc8244932012-01-12 17:19:38 -08001429 /*
1430 * To minimise LRU disruption, the caller can indicate that it only
1431 * wants to isolate pages it will be able to operate on without
1432 * blocking - clean pages for the most part.
1433 *
Mel Gormanc8244932012-01-12 17:19:38 -08001434 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1435 * that it is possible to migrate without blocking
1436 */
Johannes Weiner14d6ecf2017-02-24 14:56:11 -08001437 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001438 /* All the caller can do on PageWriteback is block */
1439 if (PageWriteback(page))
1440 return ret;
1441
1442 if (PageDirty(page)) {
1443 struct address_space *mapping;
Mel Gormanab88b8a2018-01-31 16:19:52 -08001444 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001445
Mel Gormanc8244932012-01-12 17:19:38 -08001446 /*
1447 * Only pages without mappings or that have a
1448 * ->migratepage callback are possible to migrate
Mel Gormanab88b8a2018-01-31 16:19:52 -08001449 * without blocking. However, we can be racing with
1450 * truncation so it's necessary to lock the page
1451 * to stabilise the mapping as truncation holds
1452 * the page lock until after the page is removed
1453 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001454 */
Mel Gormanab88b8a2018-01-31 16:19:52 -08001455 if (!trylock_page(page))
1456 return ret;
1457
Mel Gormanc8244932012-01-12 17:19:38 -08001458 mapping = page_mapping(page);
Hugh Dickins93960f92018-06-01 16:50:50 -07001459 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gormanab88b8a2018-01-31 16:19:52 -08001460 unlock_page(page);
1461 if (!migrate_dirty)
Mel Gormanc8244932012-01-12 17:19:38 -08001462 return ret;
1463 }
1464 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001465
Minchan Kimf80c0672011-10-31 17:06:55 -07001466 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1467 return ret;
1468
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001469 if (likely(get_page_unless_zero(page))) {
1470 /*
1471 * Be careful not to clear PageLRU until after we're
1472 * sure the page is not being freed elsewhere -- the
1473 * page release code relies on it.
1474 */
1475 ClearPageLRU(page);
1476 ret = 0;
1477 }
1478
1479 return ret;
1480}
1481
Mel Gorman7ee36a12016-07-28 15:47:17 -07001482
1483/*
1484 * Update LRU sizes after isolating pages. The LRU size updates must
1485 * be complete before mem_cgroup_update_lru_size due to a santity check.
1486 */
1487static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hocko07fc9572017-01-10 16:58:04 -08001488 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001489{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001490 int zid;
1491
Mel Gorman7ee36a12016-07-28 15:47:17 -07001492 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1493 if (!nr_zone_taken[zid])
1494 continue;
1495
1496 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hocko07fc9572017-01-10 16:58:04 -08001497#ifdef CONFIG_MEMCG
1498 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1499#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001500 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001501
Mel Gorman7ee36a12016-07-28 15:47:17 -07001502}
1503
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001504/*
Mel Gormana52633d2016-07-28 15:45:28 -07001505 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 * shrink the lists perform better by taking out a batch of pages
1507 * and working on them outside the LRU lock.
1508 *
1509 * For pagecache intensive workloads, this function is the hottest
1510 * spot in the kernel (apart from copy_*_user functions).
1511 *
1512 * Appropriate locks must be held before calling this function.
1513 *
Minchan Kim470e5b62017-05-12 15:47:06 -07001514 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001515 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001517 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001518 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001519 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001520 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 *
1522 * returns how many pages were moved onto *@dst.
1523 */
Andrew Morton69e05942006-03-22 00:08:19 -08001524static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001525 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001526 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001527 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001529 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001530 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001531 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001532 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner061d6da2017-05-03 14:52:13 -07001533 unsigned long skipped = 0;
Minchan Kim470e5b62017-05-12 15:47:06 -07001534 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001535 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Minchan Kim470e5b62017-05-12 15:47:06 -07001537 scan = 0;
1538 for (total_scan = 0;
1539 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1540 total_scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001541 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001542
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 page = lru_to_page(src);
1544 prefetchw_prev_lru_page(page, src, flags);
1545
Sasha Levin309381fea2014-01-23 15:52:54 -08001546 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001547
Mel Gormanb2e18752016-07-28 15:45:37 -07001548 if (page_zonenum(page) > sc->reclaim_idx) {
1549 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001550 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001551 continue;
1552 }
1553
Minchan Kim470e5b62017-05-12 15:47:06 -07001554 /*
1555 * Do not count skipped pages because that makes the function
1556 * return with no isolated pages if the LRU mostly contains
1557 * ineligible pages. This causes the VM to not reclaim any
1558 * pages, triggering a premature OOM.
1559 */
1560 scan++;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001561 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001562 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001563 nr_pages = hpage_nr_pages(page);
1564 nr_taken += nr_pages;
1565 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001566 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001567 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001568
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001569 case -EBUSY:
1570 /* else it is being freed elsewhere */
1571 list_move(&page->lru, src);
1572 continue;
1573
1574 default:
1575 BUG();
1576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 }
1578
Mel Gormanb2e18752016-07-28 15:45:37 -07001579 /*
1580 * Splice any skipped pages to the start of the LRU list. Note that
1581 * this disrupts the LRU order when reclaiming for lower zones but
1582 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1583 * scanning would soon rescan the same pages to skip and put the
1584 * system at risk of premature OOM.
1585 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001586 if (!list_empty(&pages_skipped)) {
1587 int zid;
1588
Johannes Weiner061d6da2017-05-03 14:52:13 -07001589 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001590 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1591 if (!nr_skipped[zid])
1592 continue;
1593
1594 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko4d7658d2017-02-22 15:44:21 -08001595 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001596 }
1597 }
Minchan Kim470e5b62017-05-12 15:47:06 -07001598 *nr_scanned = total_scan;
Michal Hocko4d7658d2017-02-22 15:44:21 -08001599 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim470e5b62017-05-12 15:47:06 -07001600 total_scan, skipped, nr_taken, mode,
1601 is_file_lru(lru));
Michal Hocko07fc9572017-01-10 16:58:04 -08001602 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 return nr_taken;
1604}
1605
Nick Piggin62695a82008-10-18 20:26:09 -07001606/**
1607 * isolate_lru_page - tries to isolate a page from its LRU list
1608 * @page: page to isolate from its LRU list
1609 *
1610 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1611 * vmstat statistic corresponding to whatever LRU list the page was on.
1612 *
1613 * Returns 0 if the page was removed from an LRU list.
1614 * Returns -EBUSY if the page was not on an LRU list.
1615 *
1616 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001617 * the active list, it will have PageActive set. If it was found on
1618 * the unevictable list, it will have the PageUnevictable bit set. That flag
1619 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001620 *
1621 * The vmstat statistic corresponding to the list on which the page was
1622 * found will be decremented.
1623 *
1624 * Restrictions:
1625 * (1) Must be called with an elevated refcount on the page. This is a
1626 * fundamentnal difference from isolate_lru_pages (which is called
1627 * without a stable reference).
1628 * (2) the lru_lock must not be held.
1629 * (3) interrupts must be enabled.
1630 */
1631int isolate_lru_page(struct page *page)
1632{
1633 int ret = -EBUSY;
1634
Sasha Levin309381fea2014-01-23 15:52:54 -08001635 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001636 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001637
Nick Piggin62695a82008-10-18 20:26:09 -07001638 if (PageLRU(page)) {
1639 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001640 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001641
Mel Gormana52633d2016-07-28 15:45:28 -07001642 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001643 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001644 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001645 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001646 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001647 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001648 del_page_from_lru_list(page, lruvec, lru);
1649 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001650 }
Mel Gormana52633d2016-07-28 15:45:28 -07001651 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001652 }
1653 return ret;
1654}
1655
Vinayak Menondd68ed22014-12-26 19:29:41 +05301656static int __too_many_isolated(struct pglist_data *pgdat, int file,
1657 struct scan_control *sc, int safe)
Rik van Riel35cd7812009-09-21 17:01:38 -07001658{
1659 unsigned long inactive, isolated;
1660
Rik van Riel35cd7812009-09-21 17:01:38 -07001661 if (file) {
Vinayak Menondd68ed22014-12-26 19:29:41 +05301662 if (safe) {
1663 inactive = node_page_state_snapshot(pgdat,
1664 NR_INACTIVE_FILE);
1665 isolated = node_page_state_snapshot(pgdat,
1666 NR_ISOLATED_FILE);
1667 } else {
1668 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1669 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1670 }
Rik van Riel35cd7812009-09-21 17:01:38 -07001671 } else {
Vinayak Menondd68ed22014-12-26 19:29:41 +05301672 if (safe) {
1673 inactive = node_page_state_snapshot(pgdat,
1674 NR_INACTIVE_ANON);
1675 isolated = node_page_state_snapshot(pgdat,
1676 NR_ISOLATED_ANON);
1677 } else {
1678 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1679 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1680 }
Rik van Riel35cd7812009-09-21 17:01:38 -07001681 }
1682
Fengguang Wu3cf23842012-12-18 14:23:31 -08001683 /*
1684 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1685 * won't get blocked by normal direct-reclaimers, forming a circular
1686 * deadlock.
1687 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001688 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001689 inactive >>= 3;
1690
Rik van Riel35cd7812009-09-21 17:01:38 -07001691 return isolated > inactive;
1692}
1693
Vinayak Menondd68ed22014-12-26 19:29:41 +05301694/*
1695 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1696 * then get resheduled. When there are massive number of tasks doing page
1697 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1698 * the LRU list will go small and be scanned faster than necessary, leading to
1699 * unnecessary swapping, thrashing and OOM.
1700 */
1701static int too_many_isolated(struct pglist_data *pgdat, int file,
1702 struct scan_control *sc, int safe)
1703{
1704 if (current_is_kswapd())
1705 return 0;
1706
1707 if (!sane_reclaim(sc))
1708 return 0;
1709
1710 if (unlikely(__too_many_isolated(pgdat, file, sc, 0))) {
1711 if (safe)
1712 return __too_many_isolated(pgdat, file, sc, safe);
1713 else
1714 return 1;
1715 }
1716
1717 return 0;
1718}
1719
Mel Gorman66635622010-08-09 17:19:30 -07001720static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001721putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001722{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001723 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001724 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001725 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001726
Mel Gorman66635622010-08-09 17:19:30 -07001727 /*
1728 * Put back any unfreeable pages.
1729 */
Mel Gorman66635622010-08-09 17:19:30 -07001730 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001731 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001732 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001733
Sasha Levin309381fea2014-01-23 15:52:54 -08001734 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001735 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001736 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001737 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001738 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001739 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001740 continue;
1741 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001742
Mel Gorman599d0c92016-07-28 15:45:31 -07001743 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001744
Linus Torvalds7a608572011-01-17 14:42:19 -08001745 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001746 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001747 add_page_to_lru_list(page, lruvec, lru);
1748
Mel Gorman66635622010-08-09 17:19:30 -07001749 if (is_active_lru(lru)) {
1750 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001751 int numpages = hpage_nr_pages(page);
1752 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001753 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001754 if (put_page_testzero(page)) {
1755 __ClearPageLRU(page);
1756 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001757 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001758
1759 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001760 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001761 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001762 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001763 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001764 } else
1765 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001766 }
1767 }
Mel Gorman66635622010-08-09 17:19:30 -07001768
Hugh Dickins3f797682012-01-12 17:20:07 -08001769 /*
1770 * To save our caller's stack, now use input list for pages to free.
1771 */
1772 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001773}
1774
1775/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001776 * If a kernel thread (such as nfsd for loop-back mounts) services
1777 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1778 * In that case we should only throttle if the backing device it is
1779 * writing to is congested. In other cases it is safe to throttle.
1780 */
1781static int current_may_throttle(void)
1782{
1783 return !(current->flags & PF_LESS_THROTTLE) ||
1784 current->backing_dev_info == NULL ||
1785 bdi_write_congested(current->backing_dev_info);
1786}
1787
Minchan Kim91dcade2016-07-28 15:47:40 -07001788static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1789 struct scan_control *sc, enum lru_list lru)
1790{
1791 int zid;
1792 struct zone *zone;
1793 int file = is_file_lru(lru);
1794 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1795
1796 if (!global_reclaim(sc))
1797 return true;
1798
1799 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1800 zone = &pgdat->node_zones[zid];
Mel Gorman6aa303d2016-09-01 16:14:55 -07001801 if (!managed_zone(zone))
Minchan Kim91dcade2016-07-28 15:47:40 -07001802 continue;
1803
1804 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1805 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1806 return true;
1807 }
1808
1809 return false;
1810}
1811
NeilBrown399ba0b2014-06-04 16:07:42 -07001812/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001813 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001814 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 */
Mel Gorman66635622010-08-09 17:19:30 -07001816static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001817shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001818 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819{
1820 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001821 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001822 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001823 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001824 unsigned long nr_dirty = 0;
1825 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001826 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001827 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001828 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001829 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001830 int file = is_file_lru(lru);
Vinayak Menondd68ed22014-12-26 19:29:41 +05301831 int safe = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001832 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001833 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001834
Minchan Kim91dcade2016-07-28 15:47:40 -07001835 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1836 return 0;
1837
Vinayak Menondd68ed22014-12-26 19:29:41 +05301838 while (unlikely(too_many_isolated(pgdat, file, sc, safe))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001839 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001840
1841 /* We are about to die and free our memory. Return now. */
1842 if (fatal_signal_pending(current))
1843 return SWAP_CLUSTER_MAX;
Vinayak Menondd68ed22014-12-26 19:29:41 +05301844
1845 safe = 1;
Rik van Riel35cd7812009-09-21 17:01:38 -07001846 }
1847
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001849
1850 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001851 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001852
Mel Gorman599d0c92016-07-28 15:45:31 -07001853 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001855 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1856 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001857
Mel Gorman599d0c92016-07-28 15:45:31 -07001858 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001859 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001860
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001861 if (global_reclaim(sc)) {
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001862 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001863 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001864 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001865 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001866 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001867 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001868
Hillf Dantond563c052012-03-21 16:34:02 -07001869 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001870 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001871
Mel Gorman599d0c92016-07-28 15:45:31 -07001872 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001873 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1874 &nr_writeback, &nr_immediate,
1875 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001876
Mel Gorman599d0c92016-07-28 15:45:31 -07001877 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001878
Ying Han904249a2012-04-25 16:01:48 -07001879 if (global_reclaim(sc)) {
1880 if (current_is_kswapd())
Mel Gorman599d0c92016-07-28 15:45:31 -07001881 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001882 else
Mel Gorman599d0c92016-07-28 15:45:31 -07001883 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001884 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001885
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001886 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001887
Mel Gorman599d0c92016-07-28 15:45:31 -07001888 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001889
Mel Gorman599d0c92016-07-28 15:45:31 -07001890 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001891
Johannes Weiner747db952014-08-08 14:19:24 -07001892 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001893 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001894
Mel Gorman92df3a72011-10-31 17:07:56 -07001895 /*
1896 * If reclaim is isolating dirty pages under writeback, it implies
1897 * that the long-lived page allocation rate is exceeding the page
1898 * laundering rate. Either the global limits are not being effective
1899 * at throttling processes due to the page distribution throughout
1900 * zones or there is heavy usage of a slow backing device. The
1901 * only option is to throttle from reclaim context which is not ideal
1902 * as there is no guarantee the dirtying process is throttled in the
1903 * same way balance_dirty_pages() manages.
1904 *
Mel Gorman8e950282013-07-03 15:02:02 -07001905 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1906 * of pages under pages flagged for immediate reclaim and stall if any
1907 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001908 */
Mel Gorman918fc712013-07-08 16:00:25 -07001909 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001910 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001911
Mel Gormand43006d2013-07-03 15:01:50 -07001912 /*
Andrey Ryabinin82ed8342018-03-22 16:17:42 -07001913 * If dirty pages are scanned that are not queued for IO, it
1914 * implies that flushers are not doing their job. This can
1915 * happen when memory pressure pushes dirty pages to the end of
1916 * the LRU before the dirty limits are breached and the dirty
1917 * data has expired. It can also happen when the proportion of
1918 * dirty pages grows not through writes but through memory
1919 * pressure reclaiming all the clean cache. And in some cases,
1920 * the flushers simply cannot keep up with the allocation
1921 * rate. Nudge the flusher threads in case they are asleep.
1922 */
1923 if (nr_unqueued_dirty == nr_taken)
1924 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1925
1926 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001927 * Legacy memcg will stall in page writeback so avoid forcibly
1928 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001929 */
Tejun Heo97c93412015-05-22 18:23:36 -04001930 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001931 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001932 * Tag a zone as congested if all the dirty pages scanned were
1933 * backed by a congested BDI and wait_iff_congested will stall.
1934 */
1935 if (nr_dirty && nr_dirty == nr_congested)
Mel Gorman599d0c92016-07-28 15:45:31 -07001936 set_bit(PGDAT_CONGESTED, &pgdat->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001937
Andrey Ryabinin82ed8342018-03-22 16:17:42 -07001938 /* Allow kswapd to start writing pages during reclaim. */
1939 if (nr_unqueued_dirty == nr_taken)
Mel Gorman599d0c92016-07-28 15:45:31 -07001940 set_bit(PGDAT_DIRTY, &pgdat->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001941
1942 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001943 * If kswapd scans pages marked marked for immediate
1944 * reclaim and under writeback (nr_immediate), it implies
1945 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001946 * they are written so also forcibly stall.
1947 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001948 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001949 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001950 }
Mel Gormand43006d2013-07-03 15:01:50 -07001951
Mel Gorman8e950282013-07-03 15:02:02 -07001952 /*
1953 * Stall direct reclaim for IO completions if underlying BDIs or zone
1954 * is congested. Allow kswapd to continue until it starts encountering
1955 * unqueued dirty pages or cycling through the LRU too quickly.
1956 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001957 if (!sc->hibernation_mode && !current_is_kswapd() &&
1958 current_may_throttle())
Mel Gorman599d0c92016-07-28 15:45:31 -07001959 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
Mel Gorman8e950282013-07-03 15:02:02 -07001960
Mel Gorman599d0c92016-07-28 15:45:31 -07001961 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1962 nr_scanned, nr_reclaimed,
yalin wangba5e9572016-01-14 15:18:48 -08001963 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001964 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965}
1966
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001967/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 * This moves pages from the active list to the inactive list.
1969 *
1970 * We move them the other way if the page is referenced by one or more
1971 * processes, from rmap.
1972 *
1973 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07001974 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07001976 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 * this, so instead we remove the pages from the LRU while processing them.
1978 * It is safe to rely on PG_active against the non-LRU pages in here because
1979 * nobody will play with that bit on a non-LRU page.
1980 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001981 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 * But we had to alter page->flags anyway.
1983 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001984
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001985static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001986 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001987 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001988 enum lru_list lru)
1989{
Mel Gorman599d0c92016-07-28 15:45:31 -07001990 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001991 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001992 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001993 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001994
Wu Fengguang3eb41402009-06-16 15:33:13 -07001995 while (!list_empty(list)) {
1996 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001997 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001998
Sasha Levin309381fea2014-01-23 15:52:54 -08001999 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002000 SetPageLRU(page);
2001
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002002 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07002003 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08002004 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002005 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07002006
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002007 if (put_page_testzero(page)) {
2008 __ClearPageLRU(page);
2009 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002010 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002011
2012 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002013 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07002014 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002015 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07002016 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002017 } else
2018 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002019 }
2020 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002021
Wu Fengguang3eb41402009-06-16 15:33:13 -07002022 if (!is_active_lru(lru))
2023 __count_vm_events(PGDEACTIVATE, pgmoved);
2024}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002025
Hugh Dickinsf6260122012-01-12 17:20:06 -08002026static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002027 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002028 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002029 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002031 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08002032 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07002033 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002035 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002036 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002038 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002039 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07002040 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002041 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002042 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
2044 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002045
2046 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07002047 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07002048
Mel Gorman599d0c92016-07-28 15:45:31 -07002049 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002050
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002051 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2052 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002053
Mel Gorman599d0c92016-07-28 15:45:31 -07002054 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002055 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002056
Mel Gorman599d0c92016-07-28 15:45:31 -07002057 __count_vm_events(PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002058
Mel Gorman599d0c92016-07-28 15:45:31 -07002059 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 while (!list_empty(&l_hold)) {
2062 cond_resched();
2063 page = lru_to_page(&l_hold);
2064 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002065
Hugh Dickins39b5f292012-10-08 16:33:18 -07002066 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002067 putback_lru_page(page);
2068 continue;
2069 }
2070
Mel Gormancc715d92012-03-21 16:34:00 -07002071 if (unlikely(buffer_heads_over_limit)) {
2072 if (page_has_private(page) && trylock_page(page)) {
2073 if (page_has_private(page))
2074 try_to_release_page(page, 0);
2075 unlock_page(page);
2076 }
2077 }
2078
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002079 if (page_referenced(page, 0, sc->target_mem_cgroup,
2080 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08002081 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002082 /*
2083 * Identify referenced, file-backed active pages and
2084 * give them one more trip around the active list. So
2085 * that executable code get better chances to stay in
2086 * memory under moderate memory pressure. Anon pages
2087 * are not likely to be evicted by use-once streaming
2088 * IO, plus JVM can create lots of anon VM_EXEC pages,
2089 * so we ignore them here.
2090 */
Wu Fengguang41e20982009-10-26 16:49:53 -07002091 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002092 list_add(&page->lru, &l_active);
2093 continue;
2094 }
2095 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002096
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002097 ClearPageActive(page); /* we are de-activating */
Johannes Weinera2383322018-10-26 15:06:04 -07002098 SetPageWorkingset(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 list_add(&page->lru, &l_inactive);
2100 }
2101
Andrew Mortonb5557492009-01-06 14:40:13 -08002102 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002103 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002104 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002105 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002106 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002107 * Count referenced pages from currently used mappings as rotated,
2108 * even though only some of them are actually re-activated. This
2109 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002110 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002111 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002112 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002113
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002114 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2115 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07002116 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2117 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002118
Johannes Weiner747db952014-08-08 14:19:24 -07002119 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07002120 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121}
2122
Rik van Riel59dc76b2016-05-20 16:56:31 -07002123/*
2124 * The inactive anon list should be small enough that the VM never has
2125 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002126 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002127 * The inactive file list should be small enough to leave most memory
2128 * to the established workingset on the scan-resistant active list,
2129 * but large enough to avoid thrashing the aggregate readahead window.
2130 *
2131 * Both inactive lists should also be large enough that each inactive
2132 * page has a chance to be referenced again before it is reclaimed.
2133 *
2134 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2135 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2136 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2137 *
2138 * total target max
2139 * memory ratio inactive
2140 * -------------------------------------
2141 * 10MB 1 5MB
2142 * 100MB 1 50MB
2143 * 1GB 3 250MB
2144 * 10GB 10 0.9GB
2145 * 100GB 31 3GB
2146 * 1TB 101 10GB
2147 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002148 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002149static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2150 struct scan_control *sc)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002151{
Rik van Riel59dc76b2016-05-20 16:56:31 -07002152 unsigned long inactive_ratio;
Michal Hocko71053132017-02-22 15:45:58 -08002153 unsigned long inactive, active;
2154 enum lru_list inactive_lru = file * LRU_FILE;
2155 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002156 unsigned long gb;
2157
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002158 /*
2159 * If we don't have swap space, anonymous page deactivation
2160 * is pointless.
2161 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002162 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002163 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002164
Michal Hocko71053132017-02-22 15:45:58 -08002165 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2166 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002167
Rik van Riel59dc76b2016-05-20 16:56:31 -07002168 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2169 if (gb)
2170 inactive_ratio = int_sqrt(10 * gb);
Rik van Rielb39415b2009-12-14 17:59:48 -08002171 else
Rik van Riel59dc76b2016-05-20 16:56:31 -07002172 inactive_ratio = 1;
2173
2174 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002175}
2176
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002177static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002178 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002179{
Rik van Rielb39415b2009-12-14 17:59:48 -08002180 if (is_active_lru(lru)) {
Mel Gormanf8d1a312016-07-28 15:47:34 -07002181 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002182 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002183 return 0;
2184 }
2185
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002186 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002187}
2188
Johannes Weiner9a265112013-02-22 16:32:17 -08002189enum scan_balance {
2190 SCAN_EQUAL,
2191 SCAN_FRACT,
2192 SCAN_ANON,
2193 SCAN_FILE,
2194};
2195
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002197 * Determine how aggressively the anon and file LRU lists should be
2198 * scanned. The relative value of each set of LRU lists is determined
2199 * by looking at the fraction of the pages scanned we did rotate back
2200 * onto the active list instead of evict.
2201 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002202 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2203 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002204 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002205static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002206 struct scan_control *sc, unsigned long *nr,
2207 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002208{
Vladimir Davydov33377672016-01-20 15:02:59 -08002209 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002210 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002211 u64 fraction[2];
2212 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002213 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002214 unsigned long anon_prio, file_prio;
2215 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002216 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002217 unsigned long ap, fp;
2218 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002219
2220 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002221 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002222 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002223 goto out;
2224 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002225
Johannes Weiner10316b32013-02-22 16:32:14 -08002226 /*
2227 * Global reclaim will swap to prevent OOM even with no
2228 * swappiness, but memcg users want to use this knob to
2229 * disable swapping for individual groups completely when
2230 * using the memory controller's swap limit feature would be
2231 * too expensive.
2232 */
Johannes Weiner02695172014-08-06 16:06:17 -07002233 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002234 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002235 goto out;
2236 }
2237
2238 /*
2239 * Do not apply any pressure balancing cleverness when the
2240 * system is close to OOM, scan both anon and file equally
2241 * (unless the swappiness setting disagrees with swapping).
2242 */
Johannes Weiner02695172014-08-06 16:06:17 -07002243 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002244 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002245 goto out;
2246 }
2247
Johannes Weiner11d16c22013-02-22 16:32:15 -08002248 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002249 * Prevent the reclaimer from falling into the cache trap: as
2250 * cache pages start out inactive, every cache fault will tip
2251 * the scan balance towards the file LRU. And as the file LRU
2252 * shrinks, so does the window for rotation from references.
2253 * This means we have a runaway feedback loop where a tiny
2254 * thrashing file LRU becomes infinitely more attractive than
2255 * anon pages. Try to detect this based on file LRU size.
2256 */
2257 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002258 unsigned long pgdatfile;
2259 unsigned long pgdatfree;
2260 int z;
2261 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002262
Mel Gorman599d0c92016-07-28 15:45:31 -07002263 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2264 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2265 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002266
Mel Gorman599d0c92016-07-28 15:45:31 -07002267 for (z = 0; z < MAX_NR_ZONES; z++) {
2268 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002269 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002270 continue;
2271
2272 total_high_wmark += high_wmark_pages(zone);
2273 }
2274
2275 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
Johannes Weiner62376252014-05-06 12:50:07 -07002276 scan_balance = SCAN_ANON;
2277 goto out;
2278 }
2279 }
2280
2281 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002282 * If there is enough inactive page cache, i.e. if the size of the
2283 * inactive list is greater than that of the active list *and* the
2284 * inactive list actually has some pages to scan on this priority, we
2285 * do not reclaim anything from the anonymous working set right now.
2286 * Without the second condition we could end up never scanning an
2287 * lruvec even if it has plenty of old anonymous pages unless the
2288 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002289 */
Patrick Daly72312122017-03-14 18:30:41 -07002290 if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) &&
2291 !inactive_list_is_low(lruvec, true, sc) &&
Michal Hocko521e92b2017-02-22 15:46:01 -08002292 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002293 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002294 goto out;
2295 }
2296
Johannes Weiner9a265112013-02-22 16:32:17 -08002297 scan_balance = SCAN_FRACT;
2298
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002299 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002300 * With swappiness at 100, anonymous and file have the same priority.
2301 * This scanning priority is essentially the inverse of IO cost.
2302 */
Johannes Weiner02695172014-08-06 16:06:17 -07002303 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002304 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002305
2306 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002307 * OK, so we have swap space and a fair amount of page cache
2308 * pages. We use the recently rotated / recently scanned
2309 * ratios to determine how valuable each cache is.
2310 *
2311 * Because workloads change over time (and to avoid overflow)
2312 * we keep these statistics as a floating average, which ends
2313 * up weighing recent references more than old ones.
2314 *
2315 * anon in [0], file in [1]
2316 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002317
Michal Hocko71053132017-02-22 15:45:58 -08002318 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2319 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2320 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2321 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002322
Mel Gorman599d0c92016-07-28 15:45:31 -07002323 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002324 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002325 reclaim_stat->recent_scanned[0] /= 2;
2326 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002327 }
2328
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002329 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002330 reclaim_stat->recent_scanned[1] /= 2;
2331 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002332 }
2333
2334 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002335 * The amount of pressure on anon vs file pages is inversely
2336 * proportional to the fraction of recently scanned pages on
2337 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002338 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002339 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002340 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002341
Satoru Moriyafe350042012-05-29 15:06:47 -07002342 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002343 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002344 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002345
Shaohua Li76a33fc2010-05-24 14:32:36 -07002346 fraction[0] = ap;
2347 fraction[1] = fp;
2348 denominator = ap + fp + 1;
2349out:
Johannes Weinerc383d082017-05-03 14:52:07 -07002350 *lru_pages = 0;
2351 for_each_evictable_lru(lru) {
2352 int file = is_file_lru(lru);
2353 unsigned long size;
2354 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002355
Johannes Weinerc383d082017-05-03 14:52:07 -07002356 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2357 scan = size >> sc->priority;
2358 /*
2359 * If the cgroup's already been deleted, make sure to
2360 * scrape out the remaining cache.
2361 */
2362 if (!scan && !mem_cgroup_online(memcg))
2363 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002364
Johannes Weinerc383d082017-05-03 14:52:07 -07002365 switch (scan_balance) {
2366 case SCAN_EQUAL:
2367 /* Scan lists relative to size */
2368 break;
2369 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002370 /*
Johannes Weinerc383d082017-05-03 14:52:07 -07002371 * Scan types proportional to swappiness and
2372 * their relative recent reclaim efficiency.
Johannes Weiner9a265112013-02-22 16:32:17 -08002373 */
Johannes Weinerc383d082017-05-03 14:52:07 -07002374 scan = div64_u64(scan * fraction[file],
2375 denominator);
2376 break;
2377 case SCAN_FILE:
2378 case SCAN_ANON:
2379 /* Scan one type exclusively */
2380 if ((scan_balance == SCAN_FILE) != file) {
2381 size = 0;
2382 scan = 0;
2383 }
2384 break;
2385 default:
2386 /* Look ma, no brain */
2387 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002388 }
Johannes Weinerc383d082017-05-03 14:52:07 -07002389
2390 *lru_pages += size;
2391 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002392 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002393}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002394
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002395/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002396 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002397 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002398static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002399 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002400{
Mel Gormanef8f2322016-07-28 15:46:05 -07002401 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002402 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002403 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002404 unsigned long nr_to_scan;
2405 enum lru_list lru;
2406 unsigned long nr_reclaimed = 0;
2407 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2408 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002409 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002410
Vladimir Davydov33377672016-01-20 15:02:59 -08002411 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002412
Mel Gormane82e0562013-07-03 15:01:44 -07002413 /* Record the original scan target for proportional adjustments later */
2414 memcpy(targets, nr, sizeof(nr));
2415
Mel Gorman1a501902014-06-04 16:10:49 -07002416 /*
2417 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2418 * event that can occur when there is little memory pressure e.g.
2419 * multiple streaming readers/writers. Hence, we do not abort scanning
2420 * when the requested number of pages are reclaimed when scanning at
2421 * DEF_PRIORITY on the assumption that the fact we are direct
2422 * reclaiming implies that kswapd is not keeping up and it is best to
2423 * do a batch of work at once. For memcg reclaim one check is made to
2424 * abort proportional reclaim if either the file or anon lru has already
2425 * dropped to zero at the first pass.
2426 */
2427 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2428 sc->priority == DEF_PRIORITY);
2429
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002430 blk_start_plug(&plug);
2431 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2432 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002433 unsigned long nr_anon, nr_file, percentage;
2434 unsigned long nr_scanned;
2435
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002436 for_each_evictable_lru(lru) {
2437 if (nr[lru]) {
2438 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2439 nr[lru] -= nr_to_scan;
2440
2441 nr_reclaimed += shrink_list(lru, nr_to_scan,
2442 lruvec, sc);
2443 }
2444 }
Mel Gormane82e0562013-07-03 15:01:44 -07002445
Michal Hockobd041732016-12-02 17:26:48 -08002446 cond_resched();
2447
Mel Gormane82e0562013-07-03 15:01:44 -07002448 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2449 continue;
2450
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002451 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002452 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002453 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002454 * proportionally what was requested by get_scan_count(). We
2455 * stop reclaiming one LRU and reduce the amount scanning
2456 * proportional to the original scan target.
2457 */
2458 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2459 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2460
Mel Gorman1a501902014-06-04 16:10:49 -07002461 /*
2462 * It's just vindictive to attack the larger once the smaller
2463 * has gone to zero. And given the way we stop scanning the
2464 * smaller below, this makes sure that we only make one nudge
2465 * towards proportionality once we've got nr_to_reclaim.
2466 */
2467 if (!nr_file || !nr_anon)
2468 break;
2469
Mel Gormane82e0562013-07-03 15:01:44 -07002470 if (nr_file > nr_anon) {
2471 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2472 targets[LRU_ACTIVE_ANON] + 1;
2473 lru = LRU_BASE;
2474 percentage = nr_anon * 100 / scan_target;
2475 } else {
2476 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2477 targets[LRU_ACTIVE_FILE] + 1;
2478 lru = LRU_FILE;
2479 percentage = nr_file * 100 / scan_target;
2480 }
2481
2482 /* Stop scanning the smaller of the LRU */
2483 nr[lru] = 0;
2484 nr[lru + LRU_ACTIVE] = 0;
2485
2486 /*
2487 * Recalculate the other LRU scan count based on its original
2488 * scan target and the percentage scanning already complete
2489 */
2490 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2491 nr_scanned = targets[lru] - nr[lru];
2492 nr[lru] = targets[lru] * (100 - percentage) / 100;
2493 nr[lru] -= min(nr[lru], nr_scanned);
2494
2495 lru += LRU_ACTIVE;
2496 nr_scanned = targets[lru] - nr[lru];
2497 nr[lru] = targets[lru] * (100 - percentage) / 100;
2498 nr[lru] -= min(nr[lru], nr_scanned);
2499
2500 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002501 }
2502 blk_finish_plug(&plug);
2503 sc->nr_reclaimed += nr_reclaimed;
2504
2505 /*
2506 * Even if we did not try to evict anon pages at all, we want to
2507 * rebalance the anon lru active/inactive ratio.
2508 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002509 if (inactive_list_is_low(lruvec, false, sc))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002510 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2511 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002512}
2513
Mel Gorman23b9da52012-05-29 15:06:20 -07002514/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002515static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002516{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002517 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002518 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002519 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002520 return true;
2521
2522 return false;
2523}
2524
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002525/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002526 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2527 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2528 * true if more pages should be reclaimed such that when the page allocator
2529 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2530 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002531 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002532static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002533 unsigned long nr_reclaimed,
2534 unsigned long nr_scanned,
2535 struct scan_control *sc)
2536{
2537 unsigned long pages_for_compaction;
2538 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002539 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002540
2541 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002542 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002543 return false;
2544
Mel Gorman28765922011-02-25 14:44:20 -08002545 /* Consider stopping depending on scan and reclaim activity */
2546 if (sc->gfp_mask & __GFP_REPEAT) {
2547 /*
2548 * For __GFP_REPEAT allocations, stop reclaiming if the
2549 * full LRU list has been scanned and we are still failing
2550 * to reclaim pages. This full LRU scan is potentially
2551 * expensive but a __GFP_REPEAT caller really wants to succeed
2552 */
2553 if (!nr_reclaimed && !nr_scanned)
2554 return false;
2555 } else {
2556 /*
2557 * For non-__GFP_REPEAT allocations which can presumably
2558 * fail without consequence, stop if we failed to reclaim
2559 * any pages from the last SWAP_CLUSTER_MAX number of
2560 * pages that were scanned. This will return to the
2561 * caller faster at the risk reclaim/compaction and
2562 * the resulting allocation attempt fails
2563 */
2564 if (!nr_reclaimed)
2565 return false;
2566 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002567
2568 /*
2569 * If we have not reclaimed enough pages for compaction and the
2570 * inactive lists are large enough, continue reclaiming
2571 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002572 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002573 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002574 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002575 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002576 if (sc->nr_reclaimed < pages_for_compaction &&
2577 inactive_lru_pages > pages_for_compaction)
2578 return true;
2579
2580 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002581 for (z = 0; z <= sc->reclaim_idx; z++) {
2582 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002583 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002584 continue;
2585
2586 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002587 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002588 case COMPACT_CONTINUE:
2589 return false;
2590 default:
2591 /* check next zone */
2592 ;
2593 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002594 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002595 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002596}
2597
Mel Gorman970a39a2016-07-28 15:46:35 -07002598static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002599{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002600 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002601 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002602 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002603
Johannes Weiner56600482012-01-12 17:17:59 -08002604 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002605 struct mem_cgroup *root = sc->target_mem_cgroup;
2606 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002607 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002608 .priority = sc->priority,
2609 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002610 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002611 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002612
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002613 nr_reclaimed = sc->nr_reclaimed;
2614 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002615
Andrew Morton694fbc02013-09-24 15:27:37 -07002616 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2617 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002618 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002619 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002620 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002621
Johannes Weiner241994e2015-02-11 15:26:06 -08002622 if (mem_cgroup_low(root, memcg)) {
2623 if (!sc->may_thrash)
2624 continue;
2625 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2626 }
2627
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002628 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002629 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002630
Mel Gormana9dd0a82016-07-28 15:46:02 -07002631 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2632 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002633
Vladimir Davydovb5afba22016-08-02 14:03:04 -07002634 if (memcg)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002635 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002636 memcg, sc->nr_scanned - scanned,
2637 lru_pages);
2638
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002639 /* Record the group's reclaim efficiency */
2640 vmpressure(sc->gfp_mask, memcg, false,
2641 sc->nr_scanned - scanned,
2642 sc->nr_reclaimed - reclaimed);
2643
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002644 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002645 * Direct reclaim and kswapd have to scan all memory
2646 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002647 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002648 *
2649 * Limit reclaim, on the other hand, only cares about
2650 * nr_to_reclaim pages to be reclaimed and it will
2651 * retry with decreasing priority if one round over the
2652 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002653 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002654 if (!global_reclaim(sc) &&
2655 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002656 mem_cgroup_iter_break(root, memcg);
2657 break;
2658 }
Johannes Weiner241994e2015-02-11 15:26:06 -08002659 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002660
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002661 /*
2662 * Shrink the slab caches in the same proportion that
2663 * the eligible LRU pages were scanned.
2664 */
Mel Gormanb2e18752016-07-28 15:45:37 -07002665 if (global_reclaim(sc))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002666 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002667 sc->nr_scanned - nr_scanned,
Mel Gormana9dd0a82016-07-28 15:46:02 -07002668 node_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002669
Vinayak Menon57e6f442017-02-19 20:00:31 +11002670 /*
2671 * Record the subtree's reclaim efficiency. The reclaimed
2672 * pages from slab is excluded here because the corresponding
2673 * scanned pages is not accounted. Moreover, freeing a page
2674 * by slab shrinking depends on each slab's object population,
2675 * making the cost model (i.e. scan:free) different from that
2676 * of LRU.
2677 */
2678 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2679 sc->nr_scanned - nr_scanned,
2680 sc->nr_reclaimed - nr_reclaimed);
2681
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002682 if (reclaim_state) {
2683 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2684 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002685 }
2686
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002687 if (sc->nr_reclaimed - nr_reclaimed)
2688 reclaimable = true;
2689
Mel Gormana9dd0a82016-07-28 15:46:02 -07002690 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002691 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002692
Johannes Weinerc40db802017-05-03 14:51:51 -07002693 /*
2694 * Kswapd gives up on balancing particular nodes after too
2695 * many failures to reclaim anything from them and goes to
2696 * sleep. On reclaim progress, reset the failure counter. A
2697 * successful direct reclaim run will revive a dormant kswapd.
2698 */
2699 if (reclaimable)
2700 pgdat->kswapd_failures = 0;
2701
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002702 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002703}
2704
Vlastimil Babka53853e22014-10-09 15:27:02 -07002705/*
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002706 * Returns true if compaction should go ahead for a costly-order request, or
2707 * the allocation would already succeed without compaction. Return false if we
2708 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002709 */
Mel Gorman4f588332016-07-28 15:46:38 -07002710static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002711{
Mel Gorman31483b62016-07-28 15:45:46 -07002712 unsigned long watermark;
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002713 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002714
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002715 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2716 if (suitable == COMPACT_SUCCESS)
2717 /* Allocation should succeed already. Don't reclaim. */
2718 return true;
2719 if (suitable == COMPACT_SKIPPED)
2720 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002721 return false;
2722
Vlastimil Babkafdd4c612016-10-07 16:58:03 -07002723 /*
2724 * Compaction is already possible, but it takes time to run and there
2725 * are potentially other callers using the pages just freed. So proceed
2726 * with reclaim to make a buffer of free pages available to give
2727 * compaction a reasonable chance of completing and allocating the page.
2728 * Note that we won't actually reclaim the whole buffer in one attempt
2729 * as the target watermark in should_continue_reclaim() is lower. But if
2730 * we are already above the high+gap watermark, don't reclaim at all.
2731 */
2732 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2733
2734 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002735}
2736
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737/*
2738 * This is the direct reclaim path, for page-allocating processes. We only
2739 * try to reclaim pages from zones which will satisfy the caller's allocation
2740 * request.
2741 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 * If a zone is deemed to be full of pinned pages then just give it a light
2743 * scan then give up on it.
2744 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002745static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746{
Mel Gormandd1a2392008-04-28 02:12:17 -07002747 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002748 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002749 unsigned long nr_soft_reclaimed;
2750 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002751 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002752 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002753
Mel Gormancc715d92012-03-21 16:34:00 -07002754 /*
2755 * If the number of buffer_heads in the machine exceeds the maximum
2756 * allowed level, force direct reclaim to scan the highmem zone as
2757 * highmem pages could be pinning lowmem pages storing buffer_heads
2758 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002759 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002760 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002761 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002762 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002763 }
Mel Gormancc715d92012-03-21 16:34:00 -07002764
Mel Gormand4debc62010-08-09 17:19:29 -07002765 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002766 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002767 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002768 * Take care memory controller reclaiming has small influence
2769 * to global LRU.
2770 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002771 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002772 if (!cpuset_zone_allowed(zone,
2773 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002774 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002775
Johannes Weiner0b064962014-08-06 16:06:12 -07002776 /*
2777 * If we already have plenty of memory free for
2778 * compaction in this zone, don't free any more.
2779 * Even though compaction is invoked for any
2780 * non-zero order, only frequent costly order
2781 * reclamation is disruptive enough to become a
2782 * noticeable problem, like transparent huge
2783 * page allocations.
2784 */
2785 if (IS_ENABLED(CONFIG_COMPACTION) &&
2786 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002787 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002788 sc->compaction_ready = true;
2789 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002790 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002791
Andrew Morton0608f432013-09-24 15:27:41 -07002792 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002793 * Shrink each node in the zonelist once. If the
2794 * zonelist is ordered by zone (not the default) then a
2795 * node may be shrunk multiple times but in that case
2796 * the user prefers lower zones being preserved.
2797 */
2798 if (zone->zone_pgdat == last_pgdat)
2799 continue;
2800
2801 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002802 * This steals pages from memory cgroups over softlimit
2803 * and returns the number of reclaimed pages and
2804 * scanned pages. This works for global memory pressure
2805 * and balancing, not for a memcg's limit.
2806 */
2807 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002808 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002809 sc->order, sc->gfp_mask,
2810 &nr_soft_scanned);
2811 sc->nr_reclaimed += nr_soft_reclaimed;
2812 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002813 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002814 }
Nick Piggin408d8542006-09-25 23:31:27 -07002815
Mel Gorman79dafcd2016-07-28 15:45:53 -07002816 /* See comment about same check for global reclaim above */
2817 if (zone->zone_pgdat == last_pgdat)
2818 continue;
2819 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002820 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 }
Mel Gormane0c23272011-10-31 17:09:33 -07002822
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002823 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002824 * Restore to original mask to avoid the impact on the caller if we
2825 * promoted it to __GFP_HIGHMEM.
2826 */
2827 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002829
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830/*
2831 * This is the main entry point to direct page reclaim.
2832 *
2833 * If a full scan of the inactive list fails to free enough memory then we
2834 * are "out of memory" and something needs to be killed.
2835 *
2836 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2837 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002838 * caller can't do much about. We kick the writeback threads and take explicit
2839 * naps in the hope that some of these pages can be written. But if the
2840 * allocating task holds filesystem locks which prevent writeout this might not
2841 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002842 *
2843 * returns: 0, if no pages reclaimed
2844 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 */
Mel Gormandac1d272008-04-28 02:12:12 -07002846static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002847 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848{
Johannes Weiner241994e2015-02-11 15:26:06 -08002849 int initial_priority = sc->priority;
Johannes Weiner241994e2015-02-11 15:26:06 -08002850retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002851 delayacct_freepages_start();
2852
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002853 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002854 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002856 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002857 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2858 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002859 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002860 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002861
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002862 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002863 break;
2864
2865 if (sc->compaction_ready)
2866 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
2868 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002869 * If we're getting trouble reclaiming, start doing
2870 * writepage even in laptop mode.
2871 */
2872 if (sc->priority < DEF_PRIORITY - 2)
2873 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07002874 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002875
Keika Kobayashi873b4772008-07-25 01:48:52 -07002876 delayacct_freepages_end();
2877
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002878 if (sc->nr_reclaimed)
2879 return sc->nr_reclaimed;
2880
Mel Gorman0cee34f2012-01-12 17:19:49 -08002881 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002882 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002883 return 1;
2884
Johannes Weiner241994e2015-02-11 15:26:06 -08002885 /* Untapped cgroup reserves? Don't OOM, retry. */
2886 if (!sc->may_thrash) {
2887 sc->priority = initial_priority;
2888 sc->may_thrash = 1;
2889 goto retry;
2890 }
2891
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002892 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893}
2894
Johannes Weinerc40db802017-05-03 14:51:51 -07002895static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07002896{
2897 struct zone *zone;
2898 unsigned long pfmemalloc_reserve = 0;
2899 unsigned long free_pages = 0;
2900 int i;
2901 bool wmark_ok;
2902
Johannes Weinerc40db802017-05-03 14:51:51 -07002903 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2904 return true;
2905
Mel Gorman55150612012-07-31 16:44:35 -07002906 for (i = 0; i <= ZONE_NORMAL; i++) {
2907 zone = &pgdat->node_zones[i];
Johannes Weinercbe59562017-05-03 14:51:54 -07002908 if (!managed_zone(zone))
2909 continue;
2910
2911 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07002912 continue;
2913
Mel Gorman55150612012-07-31 16:44:35 -07002914 pfmemalloc_reserve += min_wmark_pages(zone);
2915 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2916 }
2917
Mel Gorman675becc2014-06-04 16:07:35 -07002918 /* If there are no reserves (unexpected config) then do not throttle */
2919 if (!pfmemalloc_reserve)
2920 return true;
2921
Mel Gorman55150612012-07-31 16:44:35 -07002922 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2923
2924 /* kswapd must be awake if processes are being throttled */
2925 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07002926 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07002927 (enum zone_type)ZONE_NORMAL);
2928 wake_up_interruptible(&pgdat->kswapd_wait);
2929 }
2930
2931 return wmark_ok;
2932}
2933
2934/*
2935 * Throttle direct reclaimers if backing storage is backed by the network
2936 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2937 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002938 * when the low watermark is reached.
2939 *
2940 * Returns true if a fatal signal was delivered during throttling. If this
2941 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002942 */
Mel Gorman50694c22012-11-26 16:29:48 -08002943static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002944 nodemask_t *nodemask)
2945{
Mel Gorman675becc2014-06-04 16:07:35 -07002946 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002947 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002948 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002949
2950 /*
2951 * Kernel threads should not be throttled as they may be indirectly
2952 * responsible for cleaning pages necessary for reclaim to make forward
2953 * progress. kjournald for example may enter direct reclaim while
2954 * committing a transaction where throttling it could forcing other
2955 * processes to block on log_wait_commit().
2956 */
2957 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002958 goto out;
2959
2960 /*
2961 * If a fatal signal is pending, this process should not throttle.
2962 * It should return quickly so it can exit and free its memory
2963 */
2964 if (fatal_signal_pending(current))
2965 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002966
Mel Gorman675becc2014-06-04 16:07:35 -07002967 /*
2968 * Check if the pfmemalloc reserves are ok by finding the first node
2969 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2970 * GFP_KERNEL will be required for allocating network buffers when
2971 * swapping over the network so ZONE_HIGHMEM is unusable.
2972 *
2973 * Throttling is based on the first usable node and throttled processes
2974 * wait on a queue until kswapd makes progress and wakes them. There
2975 * is an affinity then between processes waking up and where reclaim
2976 * progress has been made assuming the process wakes on the same node.
2977 * More importantly, processes running on remote nodes will not compete
2978 * for remote pfmemalloc reserves and processes on different nodes
2979 * should make reasonable progress.
2980 */
2981 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002982 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002983 if (zone_idx(zone) > ZONE_NORMAL)
2984 continue;
2985
2986 /* Throttle based on the first usable node */
2987 pgdat = zone->zone_pgdat;
Johannes Weinerc40db802017-05-03 14:51:51 -07002988 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07002989 goto out;
2990 break;
2991 }
2992
2993 /* If no zone was usable by the allocation flags then do not throttle */
2994 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002995 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002996
Mel Gorman68243e72012-07-31 16:44:39 -07002997 /* Account for the throttling */
2998 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2999
Mel Gorman55150612012-07-31 16:44:35 -07003000 /*
3001 * If the caller cannot enter the filesystem, it's possible that it
3002 * is due to the caller holding an FS lock or performing a journal
3003 * transaction in the case of a filesystem like ext[3|4]. In this case,
3004 * it is not safe to block on pfmemalloc_wait as kswapd could be
3005 * blocked waiting on the same lock. Instead, throttle for up to a
3006 * second before continuing.
3007 */
3008 if (!(gfp_mask & __GFP_FS)) {
3009 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc40db802017-05-03 14:51:51 -07003010 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003011
3012 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003013 }
3014
3015 /* Throttle until kswapd wakes the process */
3016 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc40db802017-05-03 14:51:51 -07003017 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003018
3019check_pending:
3020 if (fatal_signal_pending(current))
3021 return true;
3022
3023out:
3024 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003025}
3026
Mel Gormandac1d272008-04-28 02:12:12 -07003027unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003028 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003029{
Mel Gorman33906bc2010-08-09 17:19:16 -07003030 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003031 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003032 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003033 .gfp_mask = memalloc_noio_flags(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07003034 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003035 .order = order,
3036 .nodemask = nodemask,
3037 .priority = DEF_PRIORITY,
3038 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003039 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003040 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003041 };
3042
Mel Gorman55150612012-07-31 16:44:35 -07003043 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003044 * Do not enter reclaim if fatal signal was delivered while throttled.
3045 * 1 is returned so that the page allocator does not OOM kill at this
3046 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003047 */
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003048 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003049 return 1;
3050
Mel Gorman33906bc2010-08-09 17:19:16 -07003051 trace_mm_vmscan_direct_reclaim_begin(order,
3052 sc.may_writepage,
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003053 sc.gfp_mask,
Mel Gormane5146b12016-07-28 15:46:47 -07003054 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003055
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003056 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003057
3058 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3059
3060 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003061}
3062
Andrew Mortonc255a452012-07-31 16:43:02 -07003063#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003064
Mel Gormana9dd0a82016-07-28 15:46:02 -07003065unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003066 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003067 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003068 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003069{
3070 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003071 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003072 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003073 .may_writepage = !laptop_mode,
3074 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003075 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003076 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003077 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003078 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003079
Balbir Singh4e416952009-09-23 15:56:39 -07003080 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3081 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003082
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003083 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003084 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003085 sc.gfp_mask,
3086 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003087
Balbir Singh4e416952009-09-23 15:56:39 -07003088 /*
3089 * NOTE: Although we can get the priority field, using it
3090 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003091 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003092 * will pick up pages from other mem cgroup's as well. We hack
3093 * the priority and make it zero.
3094 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003095 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003096
3097 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3098
Ying Han0ae5e892011-05-26 16:25:25 -07003099 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003100 return sc.nr_reclaimed;
3101}
3102
Johannes Weiner72835c82012-01-12 17:18:32 -08003103unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003104 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003105 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003106 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003107{
Balbir Singh4e416952009-09-23 15:56:39 -07003108 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003109 unsigned long nr_reclaimed;
Johannes Weiner3df0e592018-10-26 15:06:27 -07003110 unsigned long pflags;
Ying Han889976d2011-05-26 16:25:33 -07003111 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08003112 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003113 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07003114 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3115 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003116 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003117 .target_mem_cgroup = memcg,
3118 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003119 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003120 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003121 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003122 };
Balbir Singh66e17072008-02-07 00:13:56 -08003123
Ying Han889976d2011-05-26 16:25:33 -07003124 /*
3125 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3126 * take care of from where we get pages. So the node where we start the
3127 * scan does not need to be the current node.
3128 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003129 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003130
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003131 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003132
3133 trace_mm_vmscan_memcg_reclaim_begin(0,
3134 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003135 sc.gfp_mask,
3136 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003137
Johannes Weiner3df0e592018-10-26 15:06:27 -07003138 psi_memstall_enter(&pflags);
Johannes Weiner89a28482016-10-27 17:46:56 -07003139 current->flags |= PF_MEMALLOC;
Johannes Weiner3df0e592018-10-26 15:06:27 -07003140
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003141 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weiner3df0e592018-10-26 15:06:27 -07003142
Johannes Weiner89a28482016-10-27 17:46:56 -07003143 current->flags &= ~PF_MEMALLOC;
Johannes Weiner3df0e592018-10-26 15:06:27 -07003144 psi_memstall_leave(&pflags);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003145
3146 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3147
3148 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003149}
3150#endif
3151
Mel Gorman1d82de62016-07-28 15:45:43 -07003152static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003153 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003154{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003155 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003156
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003157 if (!total_swap_pages)
3158 return;
3159
3160 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3161 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003162 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003163
Mel Gormanf8d1a312016-07-28 15:47:34 -07003164 if (inactive_list_is_low(lruvec, false, sc))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003165 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003166 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003167
3168 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3169 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003170}
3171
Mel Gorman7fcc2022017-05-03 14:53:45 -07003172/*
3173 * Returns true if there is an eligible zone balanced for the request order
3174 * and classzone_idx
3175 */
3176static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003177{
Mel Gorman7fcc2022017-05-03 14:53:45 -07003178 int i;
3179 unsigned long mark = -1;
3180 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003181
Mel Gorman7fcc2022017-05-03 14:53:45 -07003182 for (i = 0; i <= classzone_idx; i++) {
3183 zone = pgdat->node_zones + i;
Mel Gorman6256c6b42016-07-28 15:45:56 -07003184
Mel Gorman7fcc2022017-05-03 14:53:45 -07003185 if (!managed_zone(zone))
3186 continue;
3187
3188 mark = high_wmark_pages(zone);
3189 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3190 return true;
3191 }
3192
3193 /*
3194 * If a node has no populated zone within classzone_idx, it does not
3195 * need balancing by definition. This can happen if a zone-restricted
3196 * allocation tries to wake a remote kswapd.
3197 */
3198 if (mark == -1)
3199 return true;
3200
3201 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003202}
3203
Mel Gorman62cd5882017-05-03 14:53:41 -07003204/* Clear pgdat state for congested, dirty or under writeback. */
3205static void clear_pgdat_congested(pg_data_t *pgdat)
3206{
3207 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3208 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3209 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3210}
3211
Mel Gorman1741c872011-01-13 15:46:21 -08003212/*
Mel Gorman55150612012-07-31 16:44:35 -07003213 * Prepare kswapd for sleeping. This verifies that there are no processes
3214 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3215 *
3216 * Returns true if kswapd is ready to sleep
3217 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003218static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003219{
Mel Gorman55150612012-07-31 16:44:35 -07003220 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003221 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc40db802017-05-03 14:51:51 -07003222 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003223 * race between when kswapd checks the watermarks and a process gets
3224 * throttled. There is also a potential race if processes get
3225 * throttled, kswapd wakes, a large process exits thereby balancing the
3226 * zones, which causes kswapd to exit balance_pgdat() before reaching
3227 * the wake up checks. If kswapd is going to sleep, no process should
3228 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3229 * the wake up is premature, processes will wake kswapd and get
3230 * throttled again. The difference from wake ups in balance_pgdat() is
3231 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003232 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003233 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3234 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003235
Johannes Weinerc40db802017-05-03 14:51:51 -07003236 /* Hopeless node, leave it to direct reclaim */
3237 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3238 return true;
3239
Mel Gorman7fcc2022017-05-03 14:53:45 -07003240 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3241 clear_pgdat_congested(pgdat);
3242 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003243 }
3244
Shantanu Goeleeb5a6a2017-05-03 14:53:38 -07003245 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003246}
3247
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003249 * kswapd shrinks a node of pages that are at or below the highest usable
3250 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003251 *
3252 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003253 * reclaim or if the lack of progress was due to pages under writeback.
3254 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003255 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003256static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003257 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003258{
Mel Gorman1d82de62016-07-28 15:45:43 -07003259 struct zone *zone;
3260 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003261
Mel Gorman1d82de62016-07-28 15:45:43 -07003262 /* Reclaim a number of pages proportional to the number of zones */
3263 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003264 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003265 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003266 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003267 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003268
Mel Gorman1d82de62016-07-28 15:45:43 -07003269 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003270 }
3271
Mel Gorman1d82de62016-07-28 15:45:43 -07003272 /*
3273 * Historically care was taken to put equal pressure on all zones but
3274 * now pressure is applied based on node LRU order.
3275 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003276 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003277
3278 /*
3279 * Fragmentation may mean that the system cannot be rebalanced for
3280 * high-order allocations. If twice the allocation size has been
3281 * reclaimed then recheck watermarks only at order-0 to prevent
3282 * excessive reclaim. Assume that a process requested a high-order
3283 * can direct reclaim/compact.
3284 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003285 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003286 sc->order = 0;
3287
Mel Gormanb8e83b92013-07-03 15:01:45 -07003288 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003289}
3290
3291/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003292 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3293 * that are eligible for use by the caller until at least one zone is
3294 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003296 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 *
3298 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003299 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003300 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3301 * or lower is eligible for reclaim until at least one usable zone is
3302 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003304static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003307 unsigned long nr_soft_reclaimed;
3308 unsigned long nr_soft_scanned;
Johannes Weiner3df0e592018-10-26 15:06:27 -07003309 unsigned long pflags;
Mel Gorman1d82de62016-07-28 15:45:43 -07003310 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003311 struct scan_control sc = {
3312 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003313 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003314 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003315 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003316 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003317 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003318 };
Johannes Weiner3df0e592018-10-26 15:06:27 -07003319
3320 psi_memstall_enter(&pflags);
Christoph Lameterf8891e52006-06-30 01:55:45 -07003321 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003323 do {
Johannes Weinerc40db802017-05-03 14:51:51 -07003324 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003325 bool raise_priority = true;
3326
Mel Gorman84c7a772016-07-28 15:46:44 -07003327 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328
Mel Gorman86c79f62016-07-28 15:45:59 -07003329 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003330 * If the number of buffer_heads exceeds the maximum allowed
3331 * then consider reclaiming from all zones. This has a dual
3332 * purpose -- on 64-bit systems it is expected that
3333 * buffer_heads are stripped during active rotation. On 32-bit
3334 * systems, highmem pages can pin lowmem memory and shrinking
3335 * buffers can relieve lowmem pressure. Reclaim may still not
3336 * go ahead if all eligible zones for the original allocation
3337 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003338 */
3339 if (buffer_heads_over_limit) {
3340 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3341 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003342 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003343 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344
Mel Gorman970a39a2016-07-28 15:46:35 -07003345 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003346 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003349
Mel Gorman86c79f62016-07-28 15:45:59 -07003350 /*
Mel Gorman7fcc2022017-05-03 14:53:45 -07003351 * Only reclaim if there are no eligible zones. Note that
3352 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3353 * have adjusted it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003354 */
Mel Gorman7fcc2022017-05-03 14:53:45 -07003355 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3356 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003357
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003359 * Do some background aging of the anon list, to give
3360 * pages a chance to be referenced before reclaiming. All
3361 * pages are rotated regardless of classzone as this is
3362 * about consistent aging.
3363 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003364 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003365
3366 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003367 * If we're getting trouble reclaiming, start doing writepage
3368 * even in laptop mode.
3369 */
Johannes Weiner9da1b5e2017-05-03 14:51:57 -07003370 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003371 sc.may_writepage = 1;
3372
Mel Gorman1d82de62016-07-28 15:45:43 -07003373 /* Call soft limit reclaim before calling shrink_node. */
3374 sc.nr_scanned = 0;
3375 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003376 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003377 sc.gfp_mask, &nr_soft_scanned);
3378 sc.nr_reclaimed += nr_soft_reclaimed;
3379
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003380 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003381 * There should be no need to raise the scanning priority if
3382 * enough pages are already being scanned that that high
3383 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003385 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003386 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003387
3388 /*
3389 * If the low watermark is met there is no need for processes
3390 * to be throttled on pfmemalloc_wait as they should not be
3391 * able to safely make forward progress. Wake them
3392 */
3393 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc40db802017-05-03 14:51:51 -07003394 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003395 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003396
Mel Gormanb8e83b92013-07-03 15:01:45 -07003397 /* Check if kswapd should be suspending */
3398 if (try_to_freeze() || kthread_should_stop())
3399 break;
3400
3401 /*
3402 * Raise priority if scanning rate is too low or there was no
3403 * progress in reclaiming pages
3404 */
Johannes Weinerc40db802017-05-03 14:51:51 -07003405 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3406 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003407 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003408 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
Johannes Weinerc40db802017-05-03 14:51:51 -07003410 if (!sc.nr_reclaimed)
3411 pgdat->kswapd_failures++;
3412
Mel Gormanb8e83b92013-07-03 15:01:45 -07003413out:
Johannes Weiner3df0e592018-10-26 15:06:27 -07003414 psi_memstall_leave(&pflags);
Mel Gorman0abdee22011-01-13 15:46:22 -08003415 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003416 * Return the order kswapd stopped reclaiming at as
3417 * prepare_kswapd_sleep() takes it into account. If another caller
3418 * entered the allocator slow path while kswapd was awake, order will
3419 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003420 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003421 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422}
3423
Mel Gorman7fcc2022017-05-03 14:53:45 -07003424/*
3425 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3426 * allocation request woke kswapd for. When kswapd has not woken recently,
3427 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3428 * given classzone and returns it or the highest classzone index kswapd
3429 * was recently woke for.
3430 */
3431static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3432 enum zone_type classzone_idx)
3433{
3434 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3435 return classzone_idx;
3436
3437 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3438}
3439
Mel Gorman38087d92016-07-28 15:45:49 -07003440static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3441 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003442{
3443 long remaining = 0;
3444 DEFINE_WAIT(wait);
3445
3446 if (freezing(current) || kthread_should_stop())
3447 return;
3448
3449 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3450
Shantanu Goeleeb5a6a2017-05-03 14:53:38 -07003451 /*
3452 * Try to sleep for a short interval. Note that kcompactd will only be
3453 * woken if it is possible to sleep for a short interval. This is
3454 * deliberate on the assumption that if reclaim cannot keep an
3455 * eligible zone balanced that it's also unlikely that compaction will
3456 * succeed.
3457 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003458 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003459 /*
3460 * Compaction records what page blocks it recently failed to
3461 * isolate pages from and skips them in the future scanning.
3462 * When kswapd is going to sleep, it is reasonable to assume
3463 * that pages and compaction may succeed so reset the cache.
3464 */
3465 reset_isolation_suitable(pgdat);
3466
3467 /*
3468 * We have freed the memory, now we should compact it to make
3469 * allocation of the requested order possible.
3470 */
Mel Gorman38087d92016-07-28 15:45:49 -07003471 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003472
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003473 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003474
3475 /*
3476 * If woken prematurely then reset kswapd_classzone_idx and
3477 * order. The values will either be from a wakeup request or
3478 * the previous request that slept prematurely.
3479 */
3480 if (remaining) {
Mel Gorman7fcc2022017-05-03 14:53:45 -07003481 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003482 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3483 }
3484
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003485 finish_wait(&pgdat->kswapd_wait, &wait);
3486 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3487 }
3488
3489 /*
3490 * After a short sleep, check if it was a premature sleep. If not, then
3491 * go fully to sleep until explicitly woken up.
3492 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003493 if (!remaining &&
3494 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003495 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3496
3497 /*
3498 * vmstat counters are not perfectly accurate and the estimated
3499 * value for counters such as NR_FREE_PAGES can deviate from the
3500 * true value by nr_online_cpus * threshold. To avoid the zone
3501 * watermarks being breached while under pressure, we reduce the
3502 * per-cpu vmstat threshold while kswapd is awake and restore
3503 * them before going back to sleep.
3504 */
3505 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003506
3507 if (!kthread_should_stop())
3508 schedule();
3509
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003510 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3511 } else {
3512 if (remaining)
3513 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3514 else
3515 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3516 }
3517 finish_wait(&pgdat->kswapd_wait, &wait);
3518}
3519
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520/*
3521 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003522 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 *
3524 * This basically trickles out pages so that we have _some_
3525 * free memory available even if there is no other activity
3526 * that frees anything up. This is needed for things like routing
3527 * etc, where we otherwise might have all activity going on in
3528 * asynchronous contexts that cannot page things out.
3529 *
3530 * If there are applications that are active memory-allocators
3531 * (most normal use), this basically shouldn't matter.
3532 */
3533static int kswapd(void *p)
3534{
Mel Gorman7fcc2022017-05-03 14:53:45 -07003535 unsigned int alloc_order, reclaim_order;
3536 unsigned int classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 pg_data_t *pgdat = (pg_data_t*)p;
3538 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003539
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 struct reclaim_state reclaim_state = {
3541 .reclaimed_slab = 0,
3542 };
Rusty Russella70f7302009-03-13 14:49:46 +10303543 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544
Nick Piggincf40bd12009-01-21 08:12:39 +01003545 lockdep_set_current_reclaim_state(GFP_KERNEL);
3546
Rusty Russell174596a2009-01-01 10:12:29 +10303547 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003548 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 current->reclaim_state = &reclaim_state;
3550
3551 /*
3552 * Tell the memory management that we're a "memory allocator",
3553 * and that if we need more memory we should get access to it
3554 * regardless (see "__alloc_pages()"). "kswapd" should
3555 * never get caught in the normal page freeing logic.
3556 *
3557 * (Kswapd normally doesn't need memory anyway, but sometimes
3558 * you need a small amount of memory in order to be able to
3559 * page out something else, and this flag essentially protects
3560 * us from recursively trying to free more memory as we're
3561 * trying to free the first piece of memory in the first place).
3562 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003563 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003564 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565
Mel Gorman7fcc2022017-05-03 14:53:45 -07003566 pgdat->kswapd_order = 0;
3567 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003569 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003570
Mel Gorman7fcc2022017-05-03 14:53:45 -07003571 alloc_order = reclaim_order = pgdat->kswapd_order;
3572 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3573
Mel Gorman38087d92016-07-28 15:45:49 -07003574kswapd_try_sleep:
3575 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3576 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003577
Mel Gorman38087d92016-07-28 15:45:49 -07003578 /* Read the new order and classzone_idx */
3579 alloc_order = reclaim_order = pgdat->kswapd_order;
Mel Gorman7fcc2022017-05-03 14:53:45 -07003580 classzone_idx = kswapd_classzone_idx(pgdat, 0);
Mel Gorman38087d92016-07-28 15:45:49 -07003581 pgdat->kswapd_order = 0;
Mel Gorman7fcc2022017-05-03 14:53:45 -07003582 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
David Rientjes8fe23e02009-12-14 17:58:33 -08003584 ret = try_to_freeze();
3585 if (kthread_should_stop())
3586 break;
3587
3588 /*
3589 * We can speed up thawing tasks if we don't call balance_pgdat
3590 * after returning from the refrigerator
3591 */
Mel Gorman38087d92016-07-28 15:45:49 -07003592 if (ret)
3593 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003594
Mel Gorman38087d92016-07-28 15:45:49 -07003595 /*
3596 * Reclaim begins at the requested order but if a high-order
3597 * reclaim fails then kswapd falls back to reclaiming for
3598 * order-0. If that happens, kswapd will consider sleeping
3599 * for the order it finished reclaiming at (reclaim_order)
3600 * but kcompactd is woken to compact for the original
3601 * request (alloc_order).
3602 */
Mel Gormane5146b12016-07-28 15:46:47 -07003603 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3604 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003605 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3606 if (reclaim_order < alloc_order)
3607 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003609
Johannes Weiner71abdc12014-06-06 14:35:35 -07003610 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003611 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003612 lockdep_clear_current_reclaim_state();
3613
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 return 0;
3615}
3616
3617/*
3618 * A zone is low on free memory, so wake its kswapd task to service it.
3619 */
Mel Gorman99504742011-01-13 15:46:20 -08003620void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621{
3622 pg_data_t *pgdat;
3623
Mel Gorman6aa303d2016-09-01 16:14:55 -07003624 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 return;
3626
Vladimir Davydov344736f2014-10-20 15:50:30 +04003627 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003629 pgdat = zone->zone_pgdat;
Mel Gorman7fcc2022017-05-03 14:53:45 -07003630 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3631 classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003632 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003633 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003635
Johannes Weinerc40db802017-05-03 14:51:51 -07003636 /* Hopeless node, leave it to direct reclaim */
3637 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3638 return;
3639
Mel Gorman7fcc2022017-05-03 14:53:45 -07003640 if (pgdat_balanced(pgdat, order, classzone_idx))
3641 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003642
Mel Gorman7fcc2022017-05-03 14:53:45 -07003643 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003644 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645}
3646
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003647#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003649 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003650 * freed pages.
3651 *
3652 * Rather than trying to age LRUs the aim is to preserve the overall
3653 * LRU order by reclaiming preferentially
3654 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003656unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003658 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003659 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003660 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003661 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003662 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003663 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003664 .may_writepage = 1,
3665 .may_unmap = 1,
3666 .may_swap = 1,
3667 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003669 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003670 struct task_struct *p = current;
3671 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003673 p->flags |= PF_MEMALLOC;
3674 lockdep_set_current_reclaim_state(sc.gfp_mask);
3675 reclaim_state.reclaimed_slab = 0;
3676 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003677
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003678 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003679
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003680 p->reclaim_state = NULL;
3681 lockdep_clear_current_reclaim_state();
3682 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003683
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003684 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003686#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688/* It's optimal to keep kswapds on the same CPUs as their memory, but
3689 not required for correctness. So if the last cpu in a node goes
3690 away, we get changed to run anywhere: as the first one comes back,
3691 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003692static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3693 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003695 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003697 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003698 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003699 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303700 const struct cpumask *mask;
3701
3702 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003703
Rusty Russell3e597942009-01-01 10:12:24 +10303704 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003706 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 }
3708 }
3709 return NOTIFY_OK;
3710}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
Yasunori Goto3218ae12006-06-27 02:53:33 -07003712/*
3713 * This kswapd start function will be called by init and node-hot-add.
3714 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3715 */
3716int kswapd_run(int nid)
3717{
3718 pg_data_t *pgdat = NODE_DATA(nid);
3719 int ret = 0;
3720
3721 if (pgdat->kswapd)
3722 return 0;
3723
3724 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3725 if (IS_ERR(pgdat->kswapd)) {
3726 /* failure at boot is fatal */
3727 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003728 pr_err("Failed to start kswapd on node %d\n", nid);
3729 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003730 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003731 }
3732 return ret;
3733}
3734
David Rientjes8fe23e02009-12-14 17:58:33 -08003735/*
Jiang Liud8adde12012-07-11 14:01:52 -07003736 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003737 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003738 */
3739void kswapd_stop(int nid)
3740{
3741 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3742
Jiang Liud8adde12012-07-11 14:01:52 -07003743 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003744 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003745 NODE_DATA(nid)->kswapd = NULL;
3746 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003747}
3748
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749static int __init kswapd_init(void)
3750{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003751 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003752
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003754 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003755 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 hotcpu_notifier(cpu_callback, 0);
3757 return 0;
3758}
3759
3760module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003761
3762#ifdef CONFIG_NUMA
3763/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003764 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003765 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003766 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003767 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003768 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003769int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003770
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003771#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003772#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003773#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003774#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003775
Christoph Lameter9eeff232006-01-18 17:42:31 -08003776/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003777 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003778 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3779 * a zone.
3780 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003781#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003782
Christoph Lameter9eeff232006-01-18 17:42:31 -08003783/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003784 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003785 * occur.
3786 */
3787int sysctl_min_unmapped_ratio = 1;
3788
3789/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003790 * If the number of slab pages in a zone grows beyond this percentage then
3791 * slab reclaim needs to occur.
3792 */
3793int sysctl_min_slab_ratio = 5;
3794
Mel Gorman11fb9982016-07-28 15:46:20 -07003795static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003796{
Mel Gorman11fb9982016-07-28 15:46:20 -07003797 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3798 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3799 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003800
3801 /*
3802 * It's possible for there to be more file mapped pages than
3803 * accounted for by the pages on the file LRU lists because
3804 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3805 */
3806 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3807}
3808
3809/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07003810static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003811{
Alexandru Moised031a152015-11-05 18:48:08 -08003812 unsigned long nr_pagecache_reclaimable;
3813 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003814
3815 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003816 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003817 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07003818 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07003819 * a better estimate
3820 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003821 if (node_reclaim_mode & RECLAIM_UNMAP)
3822 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003823 else
Mel Gormana5f5f912016-07-28 15:46:32 -07003824 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003825
3826 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07003827 if (!(node_reclaim_mode & RECLAIM_WRITE))
3828 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003829
3830 /* Watch for any possible underflows due to delta */
3831 if (unlikely(delta > nr_pagecache_reclaimable))
3832 delta = nr_pagecache_reclaimable;
3833
3834 return nr_pagecache_reclaimable - delta;
3835}
3836
Christoph Lameter0ff38492006-09-25 23:31:52 -07003837/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003838 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003839 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003840static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003841{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003842 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003843 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003844 struct task_struct *p = current;
3845 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003846 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003847 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003848 .gfp_mask = memalloc_noio_flags(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003849 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07003850 .priority = NODE_RECLAIM_PRIORITY,
3851 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3852 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003853 .may_swap = 1,
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003854 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08003855 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003856
Christoph Lameter9eeff232006-01-18 17:42:31 -08003857 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003858 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003859 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003860 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003861 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003862 */
3863 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Nick Desaulniersc2a3e4f2017-07-06 15:36:50 -07003864 lockdep_set_current_reclaim_state(sc.gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003865 reclaim_state.reclaimed_slab = 0;
3866 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003867
Mel Gormana5f5f912016-07-28 15:46:32 -07003868 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003869 /*
3870 * Free memory by calling shrink zone with increasing
3871 * priorities until we have enough memory freed.
3872 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003873 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07003874 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003875 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003876 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003877
Christoph Lameter9eeff232006-01-18 17:42:31 -08003878 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003879 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003880 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003881 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003882}
Andrew Morton179e9632006-03-22 00:08:18 -08003883
Mel Gormana5f5f912016-07-28 15:46:32 -07003884int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08003885{
David Rientjesd773ed62007-10-16 23:26:01 -07003886 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003887
3888 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003889 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07003890 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003891 *
Christoph Lameter96146342006-07-03 00:24:13 -07003892 * A small portion of unmapped file backed pages is needed for
3893 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07003894 * thrown out if the node is overallocated. So we do not reclaim
3895 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07003896 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003897 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003898 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3899 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3900 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003901
3902 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003903 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003904 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003905 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07003906 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003907
3908 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07003909 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08003910 * have associated processors. This will favor the local processor
3911 * over remote processors and spread off node memory allocations
3912 * as wide as possible.
3913 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003914 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3915 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003916
Mel Gormana5f5f912016-07-28 15:46:32 -07003917 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3918 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07003919
Mel Gormana5f5f912016-07-28 15:46:32 -07003920 ret = __node_reclaim(pgdat, gfp_mask, order);
3921 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003922
Mel Gorman24cf725182009-06-16 15:33:23 -07003923 if (!ret)
3924 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3925
David Rientjesd773ed62007-10-16 23:26:01 -07003926 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003927}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003928#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003929
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003930/*
3931 * page_evictable - test whether a page is evictable
3932 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003933 *
3934 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003935 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003936 *
3937 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003938 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003939 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003940 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003941 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003942int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003943{
Huang Yingd7f4e942018-04-05 16:23:20 -07003944 int ret;
3945
3946 /* Prevent address_space of inode and swap cache from being freed */
3947 rcu_read_lock();
3948 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3949 rcu_read_unlock();
3950 return ret;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003951}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003952
Hugh Dickins850465792012-01-20 14:34:19 -08003953#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003954/**
Hugh Dickins24513262012-01-20 14:34:21 -08003955 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3956 * @pages: array of pages to check
3957 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003958 *
Hugh Dickins24513262012-01-20 14:34:21 -08003959 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003960 *
3961 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003962 */
Hugh Dickins24513262012-01-20 14:34:21 -08003963void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003964{
Johannes Weiner925b7672012-01-12 17:18:15 -08003965 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07003966 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08003967 int pgscanned = 0;
3968 int pgrescued = 0;
3969 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003970
Hugh Dickins24513262012-01-20 14:34:21 -08003971 for (i = 0; i < nr_pages; i++) {
3972 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07003973 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003974
Hugh Dickins24513262012-01-20 14:34:21 -08003975 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07003976 if (pagepgdat != pgdat) {
3977 if (pgdat)
3978 spin_unlock_irq(&pgdat->lru_lock);
3979 pgdat = pagepgdat;
3980 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003981 }
Mel Gorman785b99f2016-07-28 15:47:23 -07003982 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003983
Hugh Dickins24513262012-01-20 14:34:21 -08003984 if (!PageLRU(page) || !PageUnevictable(page))
3985 continue;
3986
Hugh Dickins39b5f292012-10-08 16:33:18 -07003987 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003988 enum lru_list lru = page_lru_base_type(page);
3989
Sasha Levin309381fea2014-01-23 15:52:54 -08003990 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003991 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003992 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3993 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003994 pgrescued++;
3995 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003996 }
Hugh Dickins24513262012-01-20 14:34:21 -08003997
Mel Gorman785b99f2016-07-28 15:47:23 -07003998 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08003999 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4000 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07004001 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08004002 }
Hugh Dickins850465792012-01-20 14:34:19 -08004003}
4004#endif /* CONFIG_SHMEM */