blob: 51ccd80e70b6a29b6369b25a527f7e1ec180912d [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070015#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/mm.h>
Ingo Molnar5b3cc152017-02-02 20:43:54 +010018#include <linux/sched/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070026#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070027#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080039#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/notifier.h>
41#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080042#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070043#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080044#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080045#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070046#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070047#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070048#include <linux/oom.h>
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +000049#include <linux/pagevec.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070050#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070051#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080052#include <linux/dax.h>
Johannes Weinereb414682018-10-26 15:06:27 -070053#include <linux/psi.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070059#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Nick Piggin0f8053a2006-03-22 00:08:33 -080061#include "internal.h"
62
Mel Gorman33906bc2010-08-09 17:19:16 -070063#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080067 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
Johannes Weineree814fe2014-08-06 16:06:19 -070070 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070075
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070076 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080077 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080081
Johannes Weiner7cf111b2020-06-03 16:03:06 -070082 /*
83 * Scan pressure balancing between anon and file LRUs
84 */
85 unsigned long anon_cost;
86 unsigned long file_cost;
87
Johannes Weinerb91ac372019-11-30 17:56:02 -080088 /* Can active pages be deactivated as part of reclaim? */
89#define DEACTIVATE_ANON 1
90#define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
94
Johannes Weiner1276ad62017-02-24 14:56:11 -080095 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -070096 unsigned int may_writepage:1;
97
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
100
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
103
Yisheng Xied6622f62017-05-03 14:53:57 -0700104 /*
Johannes Weiner8132fc22021-08-19 19:04:21 -0700105 * Cgroup memory below memory.low is protected as long as we
106 * don't threaten to OOM. If any cgroup is reclaimed at
107 * reduced force or passed over entirely due to its memory.low
108 * setting (memcg_low_skipped), and nothing is reclaimed as a
109 * result, then go back for one more cycle that reclaims the protected
110 * memory (memcg_low_reclaim) to avert OOM.
Yisheng Xied6622f62017-05-03 14:53:57 -0700111 */
112 unsigned int memcg_low_reclaim:1;
113 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -0800114
Johannes Weineree814fe2014-08-06 16:06:19 -0700115 unsigned int hibernation_mode:1;
116
117 /* One of the zones is ready for compaction */
118 unsigned int compaction_ready:1;
119
Johannes Weinerb91ac372019-11-30 17:56:02 -0800120 /* There is easily reclaimable cold cache in the current node */
121 unsigned int cache_trim_mode:1;
122
Johannes Weiner53138ce2019-11-30 17:55:56 -0800123 /* The file pages on the current node are dangerously low */
124 unsigned int file_is_tiny:1;
125
Greg Thelenbb451fd2018-08-17 15:45:19 -0700126 /* Allocation order */
127 s8 order;
128
129 /* Scan (total_size >> priority) pages at once */
130 s8 priority;
131
132 /* The highest zone to isolate pages for reclaim from */
133 s8 reclaim_idx;
134
135 /* This context's GFP mask */
136 gfp_t gfp_mask;
137
Johannes Weineree814fe2014-08-06 16:06:19 -0700138 /* Incremented by the number of inactive pages that were scanned */
139 unsigned long nr_scanned;
140
141 /* Number of pages freed so far during a call to shrink_zones() */
142 unsigned long nr_reclaimed;
Andrey Ryabinind108c772018-04-10 16:27:59 -0700143
144 struct {
145 unsigned int dirty;
146 unsigned int unqueued_dirty;
147 unsigned int congested;
148 unsigned int writeback;
149 unsigned int immediate;
150 unsigned int file_taken;
151 unsigned int taken;
152 } nr;
Yafang Shaoe5ca8072019-07-16 16:26:09 -0700153
154 /* for recording the reclaimed slab by now */
155 struct reclaim_state reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156};
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158#ifdef ARCH_HAS_PREFETCHW
159#define prefetchw_prev_lru_page(_page, _base, _field) \
160 do { \
161 if ((_page)->lru.prev != _base) { \
162 struct page *prev; \
163 \
164 prev = lru_to_page(&(_page->lru)); \
165 prefetchw(&prev->_field); \
166 } \
167 } while (0)
168#else
169#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
170#endif
171
172/*
Johannes Weinerc8439662020-06-03 16:02:37 -0700173 * From 0 .. 200. Higher means more swappy.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 */
175int vm_swappiness = 60;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176
Yang Shi0a432dc2019-09-23 15:38:12 -0700177static void set_task_reclaim_state(struct task_struct *task,
178 struct reclaim_state *rs)
179{
180 /* Check for an overwrite */
181 WARN_ON_ONCE(rs && task->reclaim_state);
182
183 /* Check for the nulling of an already-nulled member */
184 WARN_ON_ONCE(!rs && !task->reclaim_state);
185
186 task->reclaim_state = rs;
187}
188
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189static LIST_HEAD(shrinker_list);
190static DECLARE_RWSEM(shrinker_rwsem);
191
Yang Shi0a432dc2019-09-23 15:38:12 -0700192#ifdef CONFIG_MEMCG
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700193/*
194 * We allow subsystems to populate their shrinker-related
195 * LRU lists before register_shrinker_prepared() is called
196 * for the shrinker, since we don't want to impose
197 * restrictions on their internal registration order.
198 * In this case shrink_slab_memcg() may find corresponding
199 * bit is set in the shrinkers map.
200 *
201 * This value is used by the function to detect registering
202 * shrinkers and to skip do_shrink_slab() calls for them.
203 */
204#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
205
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700206static DEFINE_IDR(shrinker_idr);
207static int shrinker_nr_max;
208
209static int prealloc_memcg_shrinker(struct shrinker *shrinker)
210{
211 int id, ret = -ENOMEM;
212
213 down_write(&shrinker_rwsem);
214 /* This may call shrinker, so it must use down_read_trylock() */
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700215 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700216 if (id < 0)
217 goto unlock;
218
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700219 if (id >= shrinker_nr_max) {
220 if (memcg_expand_shrinker_maps(id)) {
221 idr_remove(&shrinker_idr, id);
222 goto unlock;
223 }
224
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700225 shrinker_nr_max = id + 1;
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700226 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700227 shrinker->id = id;
228 ret = 0;
229unlock:
230 up_write(&shrinker_rwsem);
231 return ret;
232}
233
234static void unregister_memcg_shrinker(struct shrinker *shrinker)
235{
236 int id = shrinker->id;
237
238 BUG_ON(id < 0);
239
240 down_write(&shrinker_rwsem);
241 idr_remove(&shrinker_idr, id);
242 up_write(&shrinker_rwsem);
243}
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700244
Johannes Weinerb5ead352019-11-30 17:55:40 -0800245static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800246{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800247 return sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800248}
Tejun Heo97c93412015-05-22 18:23:36 -0400249
250/**
Johannes Weinerb5ead352019-11-30 17:55:40 -0800251 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
Tejun Heo97c93412015-05-22 18:23:36 -0400252 * @sc: scan_control in question
253 *
254 * The normal page dirty throttling mechanism in balance_dirty_pages() is
255 * completely broken with the legacy memcg and direct stalling in
256 * shrink_page_list() is used for throttling instead, which lacks all the
257 * niceties such as fairness, adaptive pausing, bandwidth proportional
258 * allocation and configurability.
259 *
260 * This function tests whether the vmscan currently in progress can assume
261 * that the normal dirty throttling mechanism is operational.
262 */
Johannes Weinerb5ead352019-11-30 17:55:40 -0800263static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c93412015-05-22 18:23:36 -0400264{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800265 if (!cgroup_reclaim(sc))
Tejun Heo97c93412015-05-22 18:23:36 -0400266 return true;
267#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800268 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400269 return true;
270#endif
271 return false;
272}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800273#else
Yang Shi0a432dc2019-09-23 15:38:12 -0700274static int prealloc_memcg_shrinker(struct shrinker *shrinker)
275{
276 return 0;
277}
278
279static void unregister_memcg_shrinker(struct shrinker *shrinker)
280{
281}
282
Johannes Weinerb5ead352019-11-30 17:55:40 -0800283static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800284{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800285 return false;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800286}
Tejun Heo97c93412015-05-22 18:23:36 -0400287
Johannes Weinerb5ead352019-11-30 17:55:40 -0800288static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c93412015-05-22 18:23:36 -0400289{
290 return true;
291}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800292#endif
293
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700294/*
295 * This misses isolated pages which are not accounted for to save counters.
296 * As the data only determines if reclaim or compaction continues, it is
297 * not expected that isolated pages will be a dominating factor.
298 */
299unsigned long zone_reclaimable_pages(struct zone *zone)
300{
301 unsigned long nr;
302
303 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
305 if (get_nr_swap_pages() > 0)
306 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
307 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
308
309 return nr;
310}
311
Michal Hockofd538802017-02-22 15:45:58 -0800312/**
313 * lruvec_lru_size - Returns the number of pages on the given LRU list.
314 * @lruvec: lru vector
315 * @lru: lru to use
316 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
317 */
318unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800319{
Johannes Weinerde3b0152019-11-30 17:55:31 -0800320 unsigned long size = 0;
Michal Hockofd538802017-02-22 15:45:58 -0800321 int zid;
322
Johannes Weinerde3b0152019-11-30 17:55:31 -0800323 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
Michal Hockofd538802017-02-22 15:45:58 -0800324 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800325
Michal Hockofd538802017-02-22 15:45:58 -0800326 if (!managed_zone(zone))
327 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800328
Michal Hockofd538802017-02-22 15:45:58 -0800329 if (!mem_cgroup_disabled())
Johannes Weinerde3b0152019-11-30 17:55:31 -0800330 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
Michal Hockofd538802017-02-22 15:45:58 -0800331 else
Johannes Weinerde3b0152019-11-30 17:55:31 -0800332 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
Michal Hockofd538802017-02-22 15:45:58 -0800333 }
Johannes Weinerde3b0152019-11-30 17:55:31 -0800334 return size;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800335}
336
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000338 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 */
Tetsuo Handa8e049442018-04-04 19:53:07 +0900340int prealloc_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
Alexey Dobriyanb9726c22019-03-05 15:48:26 -0800342 unsigned int size = sizeof(*shrinker->nr_deferred);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000343
Glauber Costa1d3d4432013-08-28 10:18:04 +1000344 if (shrinker->flags & SHRINKER_NUMA_AWARE)
345 size *= nr_node_ids;
346
347 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
348 if (!shrinker->nr_deferred)
349 return -ENOMEM;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700350
351 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
352 if (prealloc_memcg_shrinker(shrinker))
353 goto free_deferred;
354 }
355
Tetsuo Handa8e049442018-04-04 19:53:07 +0900356 return 0;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700357
358free_deferred:
359 kfree(shrinker->nr_deferred);
360 shrinker->nr_deferred = NULL;
361 return -ENOMEM;
Tetsuo Handa8e049442018-04-04 19:53:07 +0900362}
Glauber Costa1d3d4432013-08-28 10:18:04 +1000363
Tetsuo Handa8e049442018-04-04 19:53:07 +0900364void free_prealloced_shrinker(struct shrinker *shrinker)
365{
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700366 if (!shrinker->nr_deferred)
367 return;
368
369 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
370 unregister_memcg_shrinker(shrinker);
371
Tetsuo Handa8e049442018-04-04 19:53:07 +0900372 kfree(shrinker->nr_deferred);
373 shrinker->nr_deferred = NULL;
374}
375
376void register_shrinker_prepared(struct shrinker *shrinker)
377{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700378 down_write(&shrinker_rwsem);
379 list_add_tail(&shrinker->list, &shrinker_list);
Yang Shi42a9a532019-12-17 20:51:52 -0800380#ifdef CONFIG_MEMCG
Kirill Tkhai8df4a442018-08-21 21:51:49 -0700381 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
382 idr_replace(&shrinker_idr, shrinker, shrinker->id);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700383#endif
Rusty Russell8e1f9362007-07-17 04:03:17 -0700384 up_write(&shrinker_rwsem);
Tetsuo Handa8e049442018-04-04 19:53:07 +0900385}
386
387int register_shrinker(struct shrinker *shrinker)
388{
389 int err = prealloc_shrinker(shrinker);
390
391 if (err)
392 return err;
393 register_shrinker_prepared(shrinker);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000394 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700396EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398/*
399 * Remove one
400 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700401void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402{
Tetsuo Handabb422a72017-12-18 20:31:41 +0900403 if (!shrinker->nr_deferred)
404 return;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700405 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
406 unregister_memcg_shrinker(shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 down_write(&shrinker_rwsem);
408 list_del(&shrinker->list);
409 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700410 kfree(shrinker->nr_deferred);
Tetsuo Handabb422a72017-12-18 20:31:41 +0900411 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700413EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
415#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000416
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800417static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
Josef Bacik9092c712018-01-31 16:16:26 -0800418 struct shrinker *shrinker, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000420 unsigned long freed = 0;
421 unsigned long long delta;
422 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700423 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000424 long nr;
425 long new_nr;
426 int nid = shrinkctl->nid;
427 long batch_size = shrinker->batch ? shrinker->batch
428 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800429 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000430
Kirill Tkhaiac7fb3a2018-08-17 15:48:30 -0700431 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
432 nid = 0;
433
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700434 freeable = shrinker->count_objects(shrinker, shrinkctl);
Kirill Tkhai9b996462018-08-17 15:48:21 -0700435 if (freeable == 0 || freeable == SHRINK_EMPTY)
436 return freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000437
438 /*
439 * copy the current shrinker scan count into a local variable
440 * and zero it so that other concurrent shrinker invocations
441 * don't also do this scanning work.
442 */
443 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
444
445 total_scan = nr;
Johannes Weiner4b85afb2018-10-26 15:06:42 -0700446 if (shrinker->seeks) {
447 delta = freeable >> priority;
448 delta *= 4;
449 do_div(delta, shrinker->seeks);
450 } else {
451 /*
452 * These objects don't require any IO to create. Trim
453 * them aggressively under memory pressure to keep
454 * them from causing refetches in the IO caches.
455 */
456 delta = freeable / 2;
457 }
Roman Gushchin172b06c32018-09-20 12:22:46 -0700458
Glauber Costa1d3d4432013-08-28 10:18:04 +1000459 total_scan += delta;
460 if (total_scan < 0) {
Sakari Ailusd75f7732019-03-25 21:32:28 +0200461 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000462 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700463 total_scan = freeable;
Shaohua Li5f33a082016-12-12 16:41:50 -0800464 next_deferred = nr;
465 } else
466 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000467
468 /*
469 * We need to avoid excessive windup on filesystem shrinkers
470 * due to large numbers of GFP_NOFS allocations causing the
471 * shrinkers to return -1 all the time. This results in a large
472 * nr being built up so when a shrink that can do some work
473 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700474 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000475 * memory.
476 *
477 * Hence only allow the shrinker to scan the entire cache when
478 * a large delta change is calculated directly.
479 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700480 if (delta < freeable / 4)
481 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000482
483 /*
484 * Avoid risking looping forever due to too large nr value:
485 * never try to free more than twice the estimate number of
486 * freeable entries.
487 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700488 if (total_scan > freeable * 2)
489 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000490
491 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Josef Bacik9092c712018-01-31 16:16:26 -0800492 freeable, delta, total_scan, priority);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000493
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800494 /*
495 * Normally, we should not scan less than batch_size objects in one
496 * pass to avoid too frequent shrinker calls, but if the slab has less
497 * than batch_size objects in total and we are really tight on memory,
498 * we will try to reclaim all available objects, otherwise we can end
499 * up failing allocations although there are plenty of reclaimable
500 * objects spread over several slabs with usage less than the
501 * batch_size.
502 *
503 * We detect the "tight on memory" situations by looking at the total
504 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700505 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800506 * scanning at high prio and therefore should try to reclaim as much as
507 * possible.
508 */
509 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700510 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000511 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800512 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000513
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800514 shrinkctl->nr_to_scan = nr_to_scan;
Chris Wilsond460acb2017-09-06 16:19:26 -0700515 shrinkctl->nr_scanned = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000516 ret = shrinker->scan_objects(shrinker, shrinkctl);
517 if (ret == SHRINK_STOP)
518 break;
519 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000520
Chris Wilsond460acb2017-09-06 16:19:26 -0700521 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
522 total_scan -= shrinkctl->nr_scanned;
523 scanned += shrinkctl->nr_scanned;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000524
525 cond_resched();
526 }
527
Shaohua Li5f33a082016-12-12 16:41:50 -0800528 if (next_deferred >= scanned)
529 next_deferred -= scanned;
530 else
531 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000532 /*
533 * move the unused scan count back into the shrinker in a
534 * manner that handles concurrent updates. If we exhausted the
535 * scan, there is no need to do an update.
536 */
Shaohua Li5f33a082016-12-12 16:41:50 -0800537 if (next_deferred > 0)
538 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000539 &shrinker->nr_deferred[nid]);
540 else
541 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
542
Dave Hansendf9024a2014-06-04 16:08:07 -0700543 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000544 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545}
546
Yang Shi0a432dc2019-09-23 15:38:12 -0700547#ifdef CONFIG_MEMCG
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700548static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
549 struct mem_cgroup *memcg, int priority)
550{
551 struct memcg_shrinker_map *map;
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700552 unsigned long ret, freed = 0;
553 int i;
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700554
Yang Shi0a432dc2019-09-23 15:38:12 -0700555 if (!mem_cgroup_online(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700556 return 0;
557
558 if (!down_read_trylock(&shrinker_rwsem))
559 return 0;
560
561 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
562 true);
563 if (unlikely(!map))
564 goto unlock;
565
566 for_each_set_bit(i, map->map, shrinker_nr_max) {
567 struct shrink_control sc = {
568 .gfp_mask = gfp_mask,
569 .nid = nid,
570 .memcg = memcg,
571 };
572 struct shrinker *shrinker;
573
574 shrinker = idr_find(&shrinker_idr, i);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700575 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
576 if (!shrinker)
577 clear_bit(i, map->map);
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700578 continue;
579 }
580
Yang Shi0a432dc2019-09-23 15:38:12 -0700581 /* Call non-slab shrinkers even though kmem is disabled */
582 if (!memcg_kmem_enabled() &&
583 !(shrinker->flags & SHRINKER_NONSLAB))
584 continue;
585
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700586 ret = do_shrink_slab(&sc, shrinker, priority);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700587 if (ret == SHRINK_EMPTY) {
588 clear_bit(i, map->map);
589 /*
590 * After the shrinker reported that it had no objects to
591 * free, but before we cleared the corresponding bit in
592 * the memcg shrinker map, a new object might have been
593 * added. To make sure, we have the bit set in this
594 * case, we invoke the shrinker one more time and reset
595 * the bit if it reports that it is not empty anymore.
596 * The memory barrier here pairs with the barrier in
597 * memcg_set_shrinker_bit():
598 *
599 * list_lru_add() shrink_slab_memcg()
600 * list_add_tail() clear_bit()
601 * <MB> <MB>
602 * set_bit() do_shrink_slab()
603 */
604 smp_mb__after_atomic();
605 ret = do_shrink_slab(&sc, shrinker, priority);
606 if (ret == SHRINK_EMPTY)
607 ret = 0;
608 else
609 memcg_set_shrinker_bit(memcg, nid, i);
610 }
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700611 freed += ret;
612
613 if (rwsem_is_contended(&shrinker_rwsem)) {
614 freed = freed ? : 1;
615 break;
616 }
617 }
618unlock:
619 up_read(&shrinker_rwsem);
620 return freed;
621}
Yang Shi0a432dc2019-09-23 15:38:12 -0700622#else /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700623static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
624 struct mem_cgroup *memcg, int priority)
625{
626 return 0;
627}
Yang Shi0a432dc2019-09-23 15:38:12 -0700628#endif /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700629
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800630/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800631 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800632 * @gfp_mask: allocation context
633 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800634 * @memcg: memory cgroup whose slab caches to target
Josef Bacik9092c712018-01-31 16:16:26 -0800635 * @priority: the reclaim priority
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800637 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800639 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
640 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 *
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700642 * @memcg specifies the memory cgroup to target. Unaware shrinkers
643 * are called only if it is the root cgroup.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800644 *
Josef Bacik9092c712018-01-31 16:16:26 -0800645 * @priority is sc->priority, we take the number of objects and >> by priority
646 * in order to get the scan target.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800648 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800650static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
651 struct mem_cgroup *memcg,
Josef Bacik9092c712018-01-31 16:16:26 -0800652 int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700654 unsigned long ret, freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 struct shrinker *shrinker;
656
Yang Shifa1e5122019-08-02 21:48:44 -0700657 /*
658 * The root memcg might be allocated even though memcg is disabled
659 * via "cgroup_disable=memory" boot parameter. This could make
660 * mem_cgroup_is_root() return false, then just run memcg slab
661 * shrink, but skip global shrink. This may result in premature
662 * oom.
663 */
664 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700665 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800666
Tetsuo Handae830c632018-04-05 16:23:35 -0700667 if (!down_read_trylock(&shrinker_rwsem))
Minchan Kimf06590b2011-05-24 17:11:11 -0700668 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800671 struct shrink_control sc = {
672 .gfp_mask = gfp_mask,
673 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800674 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800675 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800676
Kirill Tkhai9b996462018-08-17 15:48:21 -0700677 ret = do_shrink_slab(&sc, shrinker, priority);
678 if (ret == SHRINK_EMPTY)
679 ret = 0;
680 freed += ret;
Minchan Kime4966122018-01-31 16:16:55 -0800681 /*
682 * Bail out if someone want to register a new shrinker to
Ethon Paul55b65a52020-06-04 16:49:10 -0700683 * prevent the registration from being stalled for long periods
Minchan Kime4966122018-01-31 16:16:55 -0800684 * by parallel ongoing shrinking.
685 */
686 if (rwsem_is_contended(&shrinker_rwsem)) {
687 freed = freed ? : 1;
688 break;
689 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800691
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700693out:
694 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000695 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800698void drop_slab_node(int nid)
699{
700 unsigned long freed;
701
702 do {
703 struct mem_cgroup *memcg = NULL;
704
Chunxin Zang069c4112020-10-13 16:56:46 -0700705 if (fatal_signal_pending(current))
706 return;
707
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800708 freed = 0;
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700709 memcg = mem_cgroup_iter(NULL, NULL, NULL);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800710 do {
Josef Bacik9092c712018-01-31 16:16:26 -0800711 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800712 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
713 } while (freed > 10);
714}
715
716void drop_slab(void)
717{
718 int nid;
719
720 for_each_online_node(nid)
721 drop_slab_node(nid);
722}
723
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724static inline int is_page_cache_freeable(struct page *page)
725{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700726 /*
727 * A freeable page cache page is referenced only by the caller
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400728 * that isolated the page, the page cache and optional buffer
729 * heads at page->private.
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700730 */
Matthew Wilcox (Oracle)3efe62e2020-10-15 20:05:56 -0700731 int page_cache_pins = thp_nr_pages(page);
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400732 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733}
734
Yang Shicb165562019-11-30 17:55:28 -0800735static int may_write_to_inode(struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
Christoph Lameter930d9152006-01-08 01:00:47 -0800737 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400739 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400741 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 return 1;
743 return 0;
744}
745
746/*
747 * We detected a synchronous write error writing a page out. Probably
748 * -ENOSPC. We need to propagate that into the address_space for a subsequent
749 * fsync(), msync() or close().
750 *
751 * The tricky part is that after writepage we cannot touch the mapping: nothing
752 * prevents it from being freed up. But we have a ref on the page and once
753 * that page is locked, the mapping is pinned.
754 *
755 * We're allowed to run sleeping lock_page() here because we know the caller has
756 * __GFP_FS.
757 */
758static void handle_write_error(struct address_space *mapping,
759 struct page *page, int error)
760{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100761 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700762 if (page_mapping(page) == mapping)
763 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 unlock_page(page);
765}
766
Christoph Lameter04e62a22006-06-23 02:03:38 -0700767/* possible outcome of pageout() */
768typedef enum {
769 /* failed to write page out, page is locked */
770 PAGE_KEEP,
771 /* move page to the active list, page is locked */
772 PAGE_ACTIVATE,
773 /* page has been sent to the disk successfully, page is unlocked */
774 PAGE_SUCCESS,
775 /* page is clean and locked */
776 PAGE_CLEAN,
777} pageout_t;
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779/*
Andrew Morton1742f192006-03-22 00:08:21 -0800780 * pageout is called by shrink_page_list() for each dirty page.
781 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 */
Yang Shicb165562019-11-30 17:55:28 -0800783static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784{
785 /*
786 * If the page is dirty, only perform writeback if that write
787 * will be non-blocking. To prevent this allocation from being
788 * stalled by pagecache activity. But note that there may be
789 * stalls if we need to run get_block(). We could test
790 * PagePrivate for that.
791 *
Al Viro81742022014-04-03 03:17:43 -0400792 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * this page's queue, we can perform writeback even if that
794 * will block.
795 *
796 * If the page is swapcache, write it back even if that would
797 * block, for some throttling. This happens by accident, because
798 * swap_backing_dev_info is bust: it doesn't reflect the
799 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 */
801 if (!is_page_cache_freeable(page))
802 return PAGE_KEEP;
803 if (!mapping) {
804 /*
805 * Some data journaling orphaned pages can have
806 * page->mapping == NULL while being dirty with clean buffers.
807 */
David Howells266cf652009-04-03 16:42:36 +0100808 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 if (try_to_free_buffers(page)) {
810 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700811 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 return PAGE_CLEAN;
813 }
814 }
815 return PAGE_KEEP;
816 }
817 if (mapping->a_ops->writepage == NULL)
818 return PAGE_ACTIVATE;
Yang Shicb165562019-11-30 17:55:28 -0800819 if (!may_write_to_inode(mapping->host))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 return PAGE_KEEP;
821
822 if (clear_page_dirty_for_io(page)) {
823 int res;
824 struct writeback_control wbc = {
825 .sync_mode = WB_SYNC_NONE,
826 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700827 .range_start = 0,
828 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 .for_reclaim = 1,
830 };
831
832 SetPageReclaim(page);
833 res = mapping->a_ops->writepage(page, &wbc);
834 if (res < 0)
835 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800836 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 ClearPageReclaim(page);
838 return PAGE_ACTIVATE;
839 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700840
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 if (!PageWriteback(page)) {
842 /* synchronous write or broken a_ops? */
843 ClearPageReclaim(page);
844 }
yalin wang3aa23852016-01-14 15:18:30 -0800845 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700846 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 return PAGE_SUCCESS;
848 }
849
850 return PAGE_CLEAN;
851}
852
Andrew Mortona649fd92006-10-17 00:09:36 -0700853/*
Nick Piggine2867812008-07-25 19:45:30 -0700854 * Same as remove_mapping, but if the page is removed from the mapping, it
855 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700856 */
Johannes Weinera5289102014-04-03 14:47:51 -0700857static int __remove_mapping(struct address_space *mapping, struct page *page,
Johannes Weinerb9107182019-11-30 17:55:59 -0800858 bool reclaimed, struct mem_cgroup *target_memcg)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800859{
Greg Thelenc4843a72015-05-22 17:13:16 -0400860 unsigned long flags;
Huang Yingbd4c82c22017-09-06 16:22:49 -0700861 int refcount;
Joonsoo Kimaae466b2020-08-11 18:30:50 -0700862 void *shadow = NULL;
Greg Thelenc4843a72015-05-22 17:13:16 -0400863
Nick Piggin28e4d962006-09-25 23:31:23 -0700864 BUG_ON(!PageLocked(page));
865 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800866
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700867 xa_lock_irqsave(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800868 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700869 * The non racy check for a busy page.
870 *
871 * Must be careful with the order of the tests. When someone has
872 * a ref to the page, it may be possible that they dirty it then
873 * drop the reference. So if PageDirty is tested before page_count
874 * here, then the following race may occur:
875 *
876 * get_user_pages(&page);
877 * [user mapping goes away]
878 * write_to(page);
879 * !PageDirty(page) [good]
880 * SetPageDirty(page);
881 * put_page(page);
882 * !page_count(page) [good, discard it]
883 *
884 * [oops, our write_to data is lost]
885 *
886 * Reversing the order of the tests ensures such a situation cannot
887 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700888 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700889 *
890 * Note that if SetPageDirty is always performed via set_page_dirty,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700891 * and thus under the i_pages lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800892 */
William Kucharski906d2782019-10-18 20:20:33 -0700893 refcount = 1 + compound_nr(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -0700894 if (!page_ref_freeze(page, refcount))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800895 goto cannot_free;
Jiang Biao1c4c3b92018-08-21 21:53:13 -0700896 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
Nick Piggine2867812008-07-25 19:45:30 -0700897 if (unlikely(PageDirty(page))) {
Huang Yingbd4c82c22017-09-06 16:22:49 -0700898 page_ref_unfreeze(page, refcount);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800899 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700900 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800901
902 if (PageSwapCache(page)) {
903 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700904 mem_cgroup_swapout(page, swap);
Joonsoo Kimaae466b2020-08-11 18:30:50 -0700905 if (reclaimed && !mapping_exiting(mapping))
906 shadow = workingset_eviction(page, target_memcg);
907 __delete_from_swap_cache(page, swap, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700908 xa_unlock_irqrestore(&mapping->i_pages, flags);
Minchan Kim75f6d6d2017-07-06 15:37:21 -0700909 put_swap_page(page, swap);
Nick Piggine2867812008-07-25 19:45:30 -0700910 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500911 void (*freepage)(struct page *);
912
913 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700914 /*
915 * Remember a shadow entry for reclaimed file cache in
916 * order to detect refaults, thus thrashing, later on.
917 *
918 * But don't store shadows in an address space that is
dylan-meiners238c3042020-08-06 23:26:29 -0700919 * already exiting. This is not just an optimization,
Johannes Weinera5289102014-04-03 14:47:51 -0700920 * inode reclaim needs to empty out the radix tree or
921 * the nodes are lost. Don't plant shadows behind its
922 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800923 *
924 * We also don't store shadows for DAX mappings because the
925 * only page cache pages found in these are zero pages
926 * covering holes, and because we don't want to mix DAX
927 * exceptional entries and shadow exceptional entries in the
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700928 * same address_space.
Johannes Weinera5289102014-04-03 14:47:51 -0700929 */
Huang Ying9de4f222020-04-06 20:04:41 -0700930 if (reclaimed && page_is_file_lru(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800931 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinerb9107182019-11-30 17:55:59 -0800932 shadow = workingset_eviction(page, target_memcg);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700933 __delete_from_page_cache(page, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700934 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500935
936 if (freepage != NULL)
937 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800938 }
939
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800940 return 1;
941
942cannot_free:
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700943 xa_unlock_irqrestore(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800944 return 0;
945}
946
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947/*
Nick Piggine2867812008-07-25 19:45:30 -0700948 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
949 * someone else has a ref on the page, abort and return 0. If it was
950 * successfully detached, return 1. Assumes the caller has a single ref on
951 * this page.
952 */
953int remove_mapping(struct address_space *mapping, struct page *page)
954{
Johannes Weinerb9107182019-11-30 17:55:59 -0800955 if (__remove_mapping(mapping, page, false, NULL)) {
Nick Piggine2867812008-07-25 19:45:30 -0700956 /*
957 * Unfreezing the refcount with 1 rather than 2 effectively
958 * drops the pagecache ref for us without requiring another
959 * atomic operation.
960 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700961 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700962 return 1;
963 }
964 return 0;
965}
966
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700967/**
968 * putback_lru_page - put previously isolated page onto appropriate LRU list
969 * @page: page to be put back to appropriate lru list
970 *
971 * Add previously isolated @page to appropriate LRU list.
972 * Page may still be unevictable for other reasons.
973 *
974 * lru_lock must not be held, interrupts must be enabled.
975 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700976void putback_lru_page(struct page *page)
977{
Shakeel Butt9c4e6b12018-02-21 14:45:28 -0800978 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700979 put_page(page); /* drop ref from isolate */
980}
981
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800982enum page_references {
983 PAGEREF_RECLAIM,
984 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800985 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800986 PAGEREF_ACTIVATE,
987};
988
989static enum page_references page_check_references(struct page *page,
990 struct scan_control *sc)
991{
Johannes Weiner645747462010-03-05 13:42:22 -0800992 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800993 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800994
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700995 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
996 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800997 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800998
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800999 /*
1000 * Mlock lost the isolation race with us. Let try_to_unmap()
1001 * move the page to the unevictable list.
1002 */
1003 if (vm_flags & VM_LOCKED)
1004 return PAGEREF_RECLAIM;
1005
Johannes Weiner645747462010-03-05 13:42:22 -08001006 if (referenced_ptes) {
Johannes Weiner645747462010-03-05 13:42:22 -08001007 /*
1008 * All mapped pages start out with page table
1009 * references from the instantiating fault, so we need
1010 * to look twice if a mapped file page is used more
1011 * than once.
1012 *
1013 * Mark it and spare it for another trip around the
1014 * inactive list. Another page table reference will
1015 * lead to its activation.
1016 *
1017 * Note: the mark is set for activated pages as well
1018 * so that recently deactivated but used pages are
1019 * quickly recovered.
1020 */
1021 SetPageReferenced(page);
1022
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -08001023 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -08001024 return PAGEREF_ACTIVATE;
1025
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001026 /*
1027 * Activate file-backed executable pages after first usage.
1028 */
Joonsoo Kimb5181542020-08-11 18:30:40 -07001029 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001030 return PAGEREF_ACTIVATE;
1031
Johannes Weiner645747462010-03-05 13:42:22 -08001032 return PAGEREF_KEEP;
1033 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001034
1035 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -07001036 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001037 return PAGEREF_RECLAIM_CLEAN;
1038
1039 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001040}
1041
Mel Gormane2be15f2013-07-03 15:01:57 -07001042/* Check if a page is dirty or under writeback */
1043static void page_check_dirty_writeback(struct page *page,
1044 bool *dirty, bool *writeback)
1045{
Mel Gormanb4597222013-07-03 15:02:05 -07001046 struct address_space *mapping;
1047
Mel Gormane2be15f2013-07-03 15:01:57 -07001048 /*
1049 * Anonymous pages are not handled by flushers and must be written
1050 * from reclaim context. Do not stall reclaim based on them
1051 */
Huang Ying9de4f222020-04-06 20:04:41 -07001052 if (!page_is_file_lru(page) ||
Shaohua Li802a3a92017-05-03 14:52:32 -07001053 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -07001054 *dirty = false;
1055 *writeback = false;
1056 return;
1057 }
1058
1059 /* By default assume that the page flags are accurate */
1060 *dirty = PageDirty(page);
1061 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -07001062
1063 /* Verify dirty/writeback state if the filesystem supports it */
1064 if (!page_has_private(page))
1065 return;
1066
1067 mapping = page_mapping(page);
1068 if (mapping && mapping->a_ops->is_dirty_writeback)
1069 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -07001070}
1071
Nick Piggine2867812008-07-25 19:45:30 -07001072/*
Andrew Morton1742f192006-03-22 00:08:21 -08001073 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 */
Maninder Singh730ec8c2020-06-03 16:01:18 -07001075static unsigned int shrink_page_list(struct list_head *page_list,
1076 struct pglist_data *pgdat,
1077 struct scan_control *sc,
Maninder Singh730ec8c2020-06-03 16:01:18 -07001078 struct reclaim_stat *stat,
1079 bool ignore_references)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080{
1081 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001082 LIST_HEAD(free_pages);
Maninder Singh730ec8c2020-06-03 16:01:18 -07001083 unsigned int nr_reclaimed = 0;
1084 unsigned int pgactivate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
Kirill Tkhai060f0052019-03-05 15:48:15 -08001086 memset(stat, 0, sizeof(*stat));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 cond_resched();
1088
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 while (!list_empty(page_list)) {
1090 struct address_space *mapping;
1091 struct page *page;
Minchan Kim8940b342019-09-25 16:49:11 -07001092 enum page_references references = PAGEREF_RECLAIM;
Kirill Tkhai4b793062020-04-01 21:10:18 -07001093 bool dirty, writeback, may_enter_fs;
Yang Shi98879b32019-07-11 20:59:30 -07001094 unsigned int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095
1096 cond_resched();
1097
1098 page = lru_to_page(page_list);
1099 list_del(&page->lru);
1100
Nick Piggin529ae9a2008-08-02 12:01:03 +02001101 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 goto keep;
1103
Sasha Levin309381fea2014-01-23 15:52:54 -08001104 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07001106 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07001107
1108 /* Account the number of base pages even though THP */
1109 sc->nr_scanned += nr_pages;
Christoph Lameter80e43422006-02-11 17:55:53 -08001110
Hugh Dickins39b5f292012-10-08 16:33:18 -07001111 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -07001112 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001113
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001114 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -08001115 goto keep_locked;
1116
Andy Whitcroftc661b072007-08-22 14:01:26 -07001117 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1118 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1119
Mel Gorman283aba92013-07-03 15:01:51 -07001120 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07001121 * The number of dirty pages determines if a node is marked
Mel Gormane2be15f2013-07-03 15:01:57 -07001122 * reclaim_congested which affects wait_iff_congested. kswapd
1123 * will stall and start writing pages if the tail of the LRU
1124 * is all dirty unqueued pages.
1125 */
1126 page_check_dirty_writeback(page, &dirty, &writeback);
1127 if (dirty || writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001128 stat->nr_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001129
1130 if (dirty && !writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001131 stat->nr_unqueued_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001132
Mel Gormand04e8ac2013-07-03 15:02:03 -07001133 /*
1134 * Treat this page as congested if the underlying BDI is or if
1135 * pages are cycling through the LRU so quickly that the
1136 * pages marked for immediate reclaim are making it to the
1137 * end of the LRU a second time.
1138 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001139 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001140 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001141 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001142 (writeback && PageReclaim(page)))
Kirill Tkhai060f0052019-03-05 15:48:15 -08001143 stat->nr_congested++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001144
1145 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001146 * If a page at the tail of the LRU is under writeback, there
1147 * are three cases to consider.
1148 *
1149 * 1) If reclaim is encountering an excessive number of pages
1150 * under writeback and this page is both under writeback and
1151 * PageReclaim then it indicates that pages are being queued
1152 * for IO but are being recycled through the LRU before the
1153 * IO can complete. Waiting on the page itself risks an
1154 * indefinite stall if it is impossible to writeback the
1155 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001156 * note that the LRU is being scanned too quickly and the
1157 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001158 *
Tejun Heo97c93412015-05-22 18:23:36 -04001159 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001160 * not marked for immediate reclaim, or the caller does not
1161 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1162 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001163 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001164 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001165 * Require may_enter_fs because we would wait on fs, which
1166 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001167 * enter reclaim, and deadlock if it waits on a page for
1168 * which it is needed to do the write (loop masks off
1169 * __GFP_IO|__GFP_FS for this reason); but more thought
1170 * would probably show more reasons.
1171 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001172 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001173 * PageReclaim. memcg does not have any dirty pages
1174 * throttling so we could easily OOM just because too many
1175 * pages are in writeback and there is nothing else to
1176 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001177 *
1178 * In cases 1) and 2) we activate the pages to get them out of
1179 * the way while we continue scanning for clean pages on the
1180 * inactive list and refilling from the active list. The
1181 * observation here is that waiting for disk writes is more
1182 * expensive than potentially causing reloads down the line.
1183 * Since they're marked for immediate reclaim, they won't put
1184 * memory pressure on the cache working set any longer than it
1185 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001186 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001187 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001188 /* Case 1 above */
1189 if (current_is_kswapd() &&
1190 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001191 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Kirill Tkhai060f0052019-03-05 15:48:15 -08001192 stat->nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001193 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001194
1195 /* Case 2 above */
Johannes Weinerb5ead352019-11-30 17:55:40 -08001196 } else if (writeback_throttling_sane(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001197 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001198 /*
1199 * This is slightly racy - end_page_writeback()
1200 * might have just cleared PageReclaim, then
1201 * setting PageReclaim here end up interpreted
1202 * as PageReadahead - but that does not matter
1203 * enough to care. What we do want is for this
1204 * page to have PageReclaim set next time memcg
1205 * reclaim reaches the tests above, so it will
1206 * then wait_on_page_writeback() to avoid OOM;
1207 * and it's also appropriate in global reclaim.
1208 */
1209 SetPageReclaim(page);
Kirill Tkhai060f0052019-03-05 15:48:15 -08001210 stat->nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001211 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001212
1213 /* Case 3 above */
1214 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001215 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001216 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001217 /* then go back and try same page again */
1218 list_add_tail(&page->lru, page_list);
1219 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001220 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001221 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
Minchan Kim8940b342019-09-25 16:49:11 -07001223 if (!ignore_references)
Minchan Kim02c6de82012-10-08 16:31:55 -07001224 references = page_check_references(page, sc);
1225
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001226 switch (references) {
1227 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001229 case PAGEREF_KEEP:
Yang Shi98879b32019-07-11 20:59:30 -07001230 stat->nr_ref_keep += nr_pages;
Johannes Weiner645747462010-03-05 13:42:22 -08001231 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001232 case PAGEREF_RECLAIM:
1233 case PAGEREF_RECLAIM_CLEAN:
1234 ; /* try to reclaim the page below */
1235 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 /*
1238 * Anonymous process memory has backing store?
1239 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001240 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001242 if (PageAnon(page) && PageSwapBacked(page)) {
1243 if (!PageSwapCache(page)) {
1244 if (!(sc->gfp_mask & __GFP_IO))
1245 goto keep_locked;
Linus Torvalds72c5ce82021-01-16 15:34:57 -08001246 if (page_maybe_dma_pinned(page))
1247 goto keep_locked;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001248 if (PageTransHuge(page)) {
1249 /* cannot split THP, skip it */
1250 if (!can_split_huge_page(page, NULL))
1251 goto activate_locked;
1252 /*
1253 * Split pages without a PMD map right
1254 * away. Chances are some or all of the
1255 * tail pages can be freed without IO.
1256 */
1257 if (!compound_mapcount(page) &&
1258 split_huge_page_to_list(page,
1259 page_list))
1260 goto activate_locked;
1261 }
1262 if (!add_to_swap(page)) {
1263 if (!PageTransHuge(page))
Yang Shi98879b32019-07-11 20:59:30 -07001264 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001265 /* Fallback to swap normal pages */
1266 if (split_huge_page_to_list(page,
1267 page_list))
1268 goto activate_locked;
Huang Yingfe490cc2017-09-06 16:22:52 -07001269#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1270 count_vm_event(THP_SWPOUT_FALLBACK);
1271#endif
Huang Yingbd4c82c22017-09-06 16:22:49 -07001272 if (!add_to_swap(page))
Yang Shi98879b32019-07-11 20:59:30 -07001273 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001274 }
Minchan Kim0f074652017-07-06 15:37:24 -07001275
Kirill Tkhai4b793062020-04-01 21:10:18 -07001276 may_enter_fs = true;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001277
1278 /* Adding to swap updated mapping */
1279 mapping = page_mapping(page);
Minchan Kim0f074652017-07-06 15:37:24 -07001280 }
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001281 } else if (unlikely(PageTransHuge(page))) {
1282 /* Split file THP */
1283 if (split_huge_page_to_list(page, page_list))
1284 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286
1287 /*
Yang Shi98879b32019-07-11 20:59:30 -07001288 * THP may get split above, need minus tail pages and update
1289 * nr_pages to avoid accounting tail pages twice.
1290 *
1291 * The tail pages that are added into swap cache successfully
1292 * reach here.
1293 */
1294 if ((nr_pages > 1) && !PageTransHuge(page)) {
1295 sc->nr_scanned -= (nr_pages - 1);
1296 nr_pages = 1;
1297 }
1298
1299 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 * The page is mapped into the page tables of one or more
1301 * processes. Try to unmap it here.
1302 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001303 if (page_mapped(page)) {
Shakeel Buttdd156e32020-12-14 19:06:39 -08001304 enum ttu_flags flags = TTU_BATCH_FLUSH;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001305 bool was_swapbacked = PageSwapBacked(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -07001306
1307 if (unlikely(PageTransHuge(page)))
1308 flags |= TTU_SPLIT_HUGE_PMD;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001309
Huang Yingbd4c82c22017-09-06 16:22:49 -07001310 if (!try_to_unmap(page, flags)) {
Yang Shi98879b32019-07-11 20:59:30 -07001311 stat->nr_unmap_fail += nr_pages;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001312 if (!was_swapbacked && PageSwapBacked(page))
1313 stat->nr_lazyfree_fail += nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 }
1316 }
1317
1318 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001319 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001320 * Only kswapd can writeback filesystem pages
1321 * to avoid risk of stack overflow. But avoid
1322 * injecting inefficient single-page IO into
1323 * flusher writeback as much as possible: only
1324 * write pages when we've encountered many
1325 * dirty pages, and when we've already scanned
1326 * the rest of the LRU for clean pages and see
1327 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001328 */
Huang Ying9de4f222020-04-06 20:04:41 -07001329 if (page_is_file_lru(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001330 (!current_is_kswapd() || !PageReclaim(page) ||
1331 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001332 /*
1333 * Immediately reclaim when written back.
1334 * Similar in principal to deactivate_page()
1335 * except we already have the page isolated
1336 * and know it's dirty
1337 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001338 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001339 SetPageReclaim(page);
1340
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001341 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001342 }
1343
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001344 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001346 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001348 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 goto keep_locked;
1350
Mel Gormand950c942015-09-04 15:47:35 -07001351 /*
1352 * Page is dirty. Flush the TLB if a writable entry
1353 * potentially exists to avoid CPU writes after IO
1354 * starts and then write it out here.
1355 */
1356 try_to_unmap_flush_dirty();
Yang Shicb165562019-11-30 17:55:28 -08001357 switch (pageout(page, mapping)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 case PAGE_KEEP:
1359 goto keep_locked;
1360 case PAGE_ACTIVATE:
1361 goto activate_locked;
1362 case PAGE_SUCCESS:
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07001363 stat->nr_pageout += thp_nr_pages(page);
Johannes Weiner96f8bf42020-06-03 16:03:09 -07001364
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001365 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001366 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001367 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 /*
1371 * A synchronous write - probably a ramdisk. Go
1372 * ahead and try to reclaim the page.
1373 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001374 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 goto keep;
1376 if (PageDirty(page) || PageWriteback(page))
1377 goto keep_locked;
1378 mapping = page_mapping(page);
1379 case PAGE_CLEAN:
1380 ; /* try to free the page below */
1381 }
1382 }
1383
1384 /*
1385 * If the page has buffers, try to free the buffer mappings
1386 * associated with this page. If we succeed we try to free
1387 * the page as well.
1388 *
1389 * We do this even if the page is PageDirty().
1390 * try_to_release_page() does not perform I/O, but it is
1391 * possible for a page to have PageDirty set, but it is actually
1392 * clean (all its buffers are clean). This happens if the
1393 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001394 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 * try_to_release_page() will discover that cleanness and will
1396 * drop the buffers and mark the page clean - it can be freed.
1397 *
1398 * Rarely, pages can have buffers and no ->mapping. These are
1399 * the pages which were not successfully invalidated in
1400 * truncate_complete_page(). We try to drop those buffers here
1401 * and if that worked, and the page is no longer mapped into
1402 * process address space (page_count == 1) it can be freed.
1403 * Otherwise, leave the page on the LRU so it is swappable.
1404 */
David Howells266cf652009-04-03 16:42:36 +01001405 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 if (!try_to_release_page(page, sc->gfp_mask))
1407 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001408 if (!mapping && page_count(page) == 1) {
1409 unlock_page(page);
1410 if (put_page_testzero(page))
1411 goto free_it;
1412 else {
1413 /*
1414 * rare race with speculative reference.
1415 * the speculative reference will free
1416 * this page shortly, so we may
1417 * increment nr_reclaimed here (and
1418 * leave it off the LRU).
1419 */
1420 nr_reclaimed++;
1421 continue;
1422 }
1423 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 }
1425
Shaohua Li802a3a92017-05-03 14:52:32 -07001426 if (PageAnon(page) && !PageSwapBacked(page)) {
1427 /* follow __remove_mapping for reference */
1428 if (!page_ref_freeze(page, 1))
1429 goto keep_locked;
1430 if (PageDirty(page)) {
1431 page_ref_unfreeze(page, 1);
1432 goto keep_locked;
1433 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Shaohua Li802a3a92017-05-03 14:52:32 -07001435 count_vm_event(PGLAZYFREED);
Roman Gushchin22621852017-07-06 15:40:25 -07001436 count_memcg_page_event(page, PGLAZYFREED);
Johannes Weinerb9107182019-11-30 17:55:59 -08001437 } else if (!mapping || !__remove_mapping(mapping, page, true,
1438 sc->target_mem_cgroup))
Shaohua Li802a3a92017-05-03 14:52:32 -07001439 goto keep_locked;
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001440
1441 unlock_page(page);
Nick Piggine2867812008-07-25 19:45:30 -07001442free_it:
Yang Shi98879b32019-07-11 20:59:30 -07001443 /*
1444 * THP may get swapped out in a whole, need account
1445 * all base pages.
1446 */
1447 nr_reclaimed += nr_pages;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001448
1449 /*
1450 * Is there need to periodically free_page_list? It would
1451 * appear not as the counts should be low
1452 */
Yang Shi7ae88532019-09-23 15:38:09 -07001453 if (unlikely(PageTransHuge(page)))
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07001454 destroy_compound_page(page);
Yang Shi7ae88532019-09-23 15:38:09 -07001455 else
Huang Yingbd4c82c22017-09-06 16:22:49 -07001456 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 continue;
1458
Yang Shi98879b32019-07-11 20:59:30 -07001459activate_locked_split:
1460 /*
1461 * The tail pages that are failed to add into swap cache
1462 * reach here. Fixup nr_scanned and nr_pages.
1463 */
1464 if (nr_pages > 1) {
1465 sc->nr_scanned -= (nr_pages - 1);
1466 nr_pages = 1;
1467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001469 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001470 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1471 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001472 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001473 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001474 if (!PageMlocked(page)) {
Huang Ying9de4f222020-04-06 20:04:41 -07001475 int type = page_is_file_lru(page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001476 SetPageActive(page);
Yang Shi98879b32019-07-11 20:59:30 -07001477 stat->nr_activate[type] += nr_pages;
Roman Gushchin22621852017-07-06 15:40:25 -07001478 count_memcg_page_event(page, PGACTIVATE);
Minchan Kimad6b6702017-05-03 14:54:13 -07001479 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480keep_locked:
1481 unlock_page(page);
1482keep:
1483 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001484 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001486
Yang Shi98879b32019-07-11 20:59:30 -07001487 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1488
Johannes Weiner747db952014-08-08 14:19:24 -07001489 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001490 try_to_unmap_flush();
Mel Gorman2d4894b2017-11-15 17:37:59 -08001491 free_unref_page_list(&free_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001492
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 list_splice(&ret_pages, page_list);
Kirill Tkhai886cf192019-05-13 17:16:51 -07001494 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001495
Andrew Morton05ff5132006-03-22 00:08:20 -08001496 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497}
1498
Maninder Singh730ec8c2020-06-03 16:01:18 -07001499unsigned int reclaim_clean_pages_from_list(struct zone *zone,
Minchan Kim02c6de82012-10-08 16:31:55 -07001500 struct list_head *page_list)
1501{
1502 struct scan_control sc = {
1503 .gfp_mask = GFP_KERNEL,
1504 .priority = DEF_PRIORITY,
1505 .may_unmap = 1,
1506 };
Jaewon Kim1f318a92020-06-03 16:01:15 -07001507 struct reclaim_stat stat;
Maninder Singh730ec8c2020-06-03 16:01:18 -07001508 unsigned int nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001509 struct page *page, *next;
1510 LIST_HEAD(clean_pages);
1511
1512 list_for_each_entry_safe(page, next, page_list, lru) {
Huang Ying9de4f222020-04-06 20:04:41 -07001513 if (page_is_file_lru(page) && !PageDirty(page) &&
Minchan Kima58f2ce2019-06-13 15:56:15 -07001514 !__PageMovable(page) && !PageUnevictable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001515 ClearPageActive(page);
1516 list_move(&page->lru, &clean_pages);
1517 }
1518 }
1519
Jaewon Kim1f318a92020-06-03 16:01:15 -07001520 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Shakeel Buttdd156e32020-12-14 19:06:39 -08001521 &stat, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001522 list_splice(&clean_pages, page_list);
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001523 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1524 -(long)nr_reclaimed);
Jaewon Kim1f318a92020-06-03 16:01:15 -07001525 /*
1526 * Since lazyfree pages are isolated from file LRU from the beginning,
1527 * they will rotate back to anonymous LRU in the end if it failed to
1528 * discard so isolated count will be mismatched.
1529 * Compensate the isolated count for both LRU lists.
1530 */
1531 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1532 stat.nr_lazyfree_fail);
1533 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001534 -(long)stat.nr_lazyfree_fail);
Jaewon Kim1f318a92020-06-03 16:01:15 -07001535 return nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001536}
1537
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001538/*
1539 * Attempt to remove the specified page from its LRU. Only take this page
1540 * if it is of the appropriate PageActive status. Pages which are being
1541 * freed elsewhere are also ignored.
1542 *
1543 * page: page to consider
1544 * mode: one of the LRU isolation modes defined above
1545 *
1546 * returns 0 on success, -ve errno on failure.
1547 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001548int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001549{
1550 int ret = -EINVAL;
1551
1552 /* Only take pages on the LRU. */
1553 if (!PageLRU(page))
1554 return ret;
1555
Minchan Kime46a2872012-10-08 16:33:48 -07001556 /* Compaction should not handle unevictable pages but CMA can do so */
1557 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001558 return ret;
1559
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001560 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001561
Mel Gormanc8244932012-01-12 17:19:38 -08001562 /*
1563 * To minimise LRU disruption, the caller can indicate that it only
1564 * wants to isolate pages it will be able to operate on without
1565 * blocking - clean pages for the most part.
1566 *
Mel Gormanc8244932012-01-12 17:19:38 -08001567 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1568 * that it is possible to migrate without blocking
1569 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001570 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001571 /* All the caller can do on PageWriteback is block */
1572 if (PageWriteback(page))
1573 return ret;
1574
1575 if (PageDirty(page)) {
1576 struct address_space *mapping;
Mel Gorman69d763f2018-01-31 16:19:52 -08001577 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001578
Mel Gormanc8244932012-01-12 17:19:38 -08001579 /*
1580 * Only pages without mappings or that have a
1581 * ->migratepage callback are possible to migrate
Mel Gorman69d763f2018-01-31 16:19:52 -08001582 * without blocking. However, we can be racing with
1583 * truncation so it's necessary to lock the page
1584 * to stabilise the mapping as truncation holds
1585 * the page lock until after the page is removed
1586 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001587 */
Mel Gorman69d763f2018-01-31 16:19:52 -08001588 if (!trylock_page(page))
1589 return ret;
1590
Mel Gormanc8244932012-01-12 17:19:38 -08001591 mapping = page_mapping(page);
Hugh Dickins145e1a72018-06-01 16:50:50 -07001592 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gorman69d763f2018-01-31 16:19:52 -08001593 unlock_page(page);
1594 if (!migrate_dirty)
Mel Gormanc8244932012-01-12 17:19:38 -08001595 return ret;
1596 }
1597 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001598
Minchan Kimf80c0672011-10-31 17:06:55 -07001599 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1600 return ret;
1601
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001602 if (likely(get_page_unless_zero(page))) {
1603 /*
1604 * Be careful not to clear PageLRU until after we're
1605 * sure the page is not being freed elsewhere -- the
1606 * page release code relies on it.
1607 */
1608 ClearPageLRU(page);
1609 ret = 0;
1610 }
1611
1612 return ret;
1613}
1614
Mel Gorman7ee36a12016-07-28 15:47:17 -07001615
1616/*
1617 * Update LRU sizes after isolating pages. The LRU size updates must
Ethon Paul55b65a52020-06-04 16:49:10 -07001618 * be complete before mem_cgroup_update_lru_size due to a sanity check.
Mel Gorman7ee36a12016-07-28 15:47:17 -07001619 */
1620static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001621 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001622{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001623 int zid;
1624
Mel Gorman7ee36a12016-07-28 15:47:17 -07001625 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1626 if (!nr_zone_taken[zid])
1627 continue;
1628
Wei Yanga892cb62020-06-03 16:01:12 -07001629 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Mel Gorman7ee36a12016-07-28 15:47:17 -07001630 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001631
Mel Gorman7ee36a12016-07-28 15:47:17 -07001632}
1633
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001634/**
1635 * pgdat->lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 * shrink the lists perform better by taking out a batch of pages
1637 * and working on them outside the LRU lock.
1638 *
1639 * For pagecache intensive workloads, this function is the hottest
1640 * spot in the kernel (apart from copy_*_user functions).
1641 *
1642 * Appropriate locks must be held before calling this function.
1643 *
Minchan Kim791b48b2017-05-12 15:47:06 -07001644 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001645 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001647 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001648 * @sc: The scan_control struct for this reclaim session
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001649 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 *
1651 * returns how many pages were moved onto *@dst.
1652 */
Andrew Morton69e05942006-03-22 00:08:19 -08001653static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001654 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001655 unsigned long *nr_scanned, struct scan_control *sc,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001656 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001658 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001659 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001660 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001661 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07001662 unsigned long skipped = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001663 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001664 LIST_HEAD(pages_skipped);
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001665 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666
Yang Shi98879b32019-07-11 20:59:30 -07001667 total_scan = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001668 scan = 0;
Yang Shi98879b32019-07-11 20:59:30 -07001669 while (scan < nr_to_scan && !list_empty(src)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001670 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001671
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 page = lru_to_page(src);
1673 prefetchw_prev_lru_page(page, src, flags);
1674
Sasha Levin309381fea2014-01-23 15:52:54 -08001675 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001676
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07001677 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07001678 total_scan += nr_pages;
1679
Mel Gormanb2e18752016-07-28 15:45:37 -07001680 if (page_zonenum(page) > sc->reclaim_idx) {
1681 list_move(&page->lru, &pages_skipped);
Yang Shi98879b32019-07-11 20:59:30 -07001682 nr_skipped[page_zonenum(page)] += nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001683 continue;
1684 }
1685
Minchan Kim791b48b2017-05-12 15:47:06 -07001686 /*
1687 * Do not count skipped pages because that makes the function
1688 * return with no isolated pages if the LRU mostly contains
1689 * ineligible pages. This causes the VM to not reclaim any
1690 * pages, triggering a premature OOM.
Yang Shi98879b32019-07-11 20:59:30 -07001691 *
1692 * Account all tail pages of THP. This would not cause
1693 * premature OOM since __isolate_lru_page() returns -EBUSY
1694 * only when the page is being freed somewhere else.
Minchan Kim791b48b2017-05-12 15:47:06 -07001695 */
Yang Shi98879b32019-07-11 20:59:30 -07001696 scan += nr_pages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001697 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001698 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001699 nr_taken += nr_pages;
1700 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001701 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001702 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001703
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001704 case -EBUSY:
1705 /* else it is being freed elsewhere */
1706 list_move(&page->lru, src);
1707 continue;
1708
1709 default:
1710 BUG();
1711 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 }
1713
Mel Gormanb2e18752016-07-28 15:45:37 -07001714 /*
1715 * Splice any skipped pages to the start of the LRU list. Note that
1716 * this disrupts the LRU order when reclaiming for lower zones but
1717 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1718 * scanning would soon rescan the same pages to skip and put the
1719 * system at risk of premature OOM.
1720 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001721 if (!list_empty(&pages_skipped)) {
1722 int zid;
1723
Johannes Weiner3db65812017-05-03 14:52:13 -07001724 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001725 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1726 if (!nr_skipped[zid])
1727 continue;
1728
1729 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08001730 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001731 }
1732 }
Minchan Kim791b48b2017-05-12 15:47:06 -07001733 *nr_scanned = total_scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08001734 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim791b48b2017-05-12 15:47:06 -07001735 total_scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001736 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 return nr_taken;
1738}
1739
Nick Piggin62695a82008-10-18 20:26:09 -07001740/**
1741 * isolate_lru_page - tries to isolate a page from its LRU list
1742 * @page: page to isolate from its LRU list
1743 *
1744 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1745 * vmstat statistic corresponding to whatever LRU list the page was on.
1746 *
1747 * Returns 0 if the page was removed from an LRU list.
1748 * Returns -EBUSY if the page was not on an LRU list.
1749 *
1750 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001751 * the active list, it will have PageActive set. If it was found on
1752 * the unevictable list, it will have the PageUnevictable bit set. That flag
1753 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001754 *
1755 * The vmstat statistic corresponding to the list on which the page was
1756 * found will be decremented.
1757 *
1758 * Restrictions:
Mike Rapoporta5d09be2018-02-06 15:42:19 -08001759 *
Nick Piggin62695a82008-10-18 20:26:09 -07001760 * (1) Must be called with an elevated refcount on the page. This is a
Hui Su01c47762020-10-13 16:56:49 -07001761 * fundamental difference from isolate_lru_pages (which is called
Nick Piggin62695a82008-10-18 20:26:09 -07001762 * without a stable reference).
1763 * (2) the lru_lock must not be held.
1764 * (3) interrupts must be enabled.
1765 */
1766int isolate_lru_page(struct page *page)
1767{
1768 int ret = -EBUSY;
1769
Sasha Levin309381fea2014-01-23 15:52:54 -08001770 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001771 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001772
Nick Piggin62695a82008-10-18 20:26:09 -07001773 if (PageLRU(page)) {
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001774 pg_data_t *pgdat = page_pgdat(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001775 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001776
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001777 spin_lock_irq(&pgdat->lru_lock);
1778 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001779 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001780 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001781 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001782 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001783 del_page_from_lru_list(page, lruvec, lru);
1784 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001785 }
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001786 spin_unlock_irq(&pgdat->lru_lock);
Nick Piggin62695a82008-10-18 20:26:09 -07001787 }
1788 return ret;
1789}
1790
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001791/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001792 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
Xianting Tian178821b2019-11-30 17:56:05 -08001793 * then get rescheduled. When there are massive number of tasks doing page
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001794 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1795 * the LRU list will go small and be scanned faster than necessary, leading to
1796 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001797 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001798static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001799 struct scan_control *sc)
1800{
1801 unsigned long inactive, isolated;
1802
1803 if (current_is_kswapd())
1804 return 0;
1805
Johannes Weinerb5ead352019-11-30 17:55:40 -08001806 if (!writeback_throttling_sane(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001807 return 0;
1808
1809 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001810 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1811 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001812 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001813 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1814 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001815 }
1816
Fengguang Wu3cf23842012-12-18 14:23:31 -08001817 /*
1818 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1819 * won't get blocked by normal direct-reclaimers, forming a circular
1820 * deadlock.
1821 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001822 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001823 inactive >>= 3;
1824
Rik van Riel35cd7812009-09-21 17:01:38 -07001825 return isolated > inactive;
1826}
1827
Kirill Tkhaia222f342019-05-13 17:17:00 -07001828/*
1829 * This moves pages from @list to corresponding LRU list.
1830 *
1831 * We move them the other way if the page is referenced by one or more
1832 * processes, from rmap.
1833 *
1834 * If the pages are mostly unmapped, the processing is fast and it is
1835 * appropriate to hold zone_lru_lock across the whole operation. But if
1836 * the pages are mapped, the processing is slow (page_referenced()) so we
1837 * should drop zone_lru_lock around each page. It's impossible to balance
1838 * this, so instead we remove the pages from the LRU while processing them.
1839 * It is safe to rely on PG_active against the non-LRU pages in here because
1840 * nobody will play with that bit on a non-LRU page.
1841 *
1842 * The downside is that we have to touch page->_refcount against each page.
1843 * But we had to alter page->flags anyway.
1844 *
1845 * Returns the number of pages moved to the given lruvec.
1846 */
1847
1848static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1849 struct list_head *list)
Mel Gorman66635622010-08-09 17:19:30 -07001850{
Mel Gorman599d0c92016-07-28 15:45:31 -07001851 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001852 int nr_pages, nr_moved = 0;
Hugh Dickins3f797682012-01-12 17:20:07 -08001853 LIST_HEAD(pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001854 struct page *page;
1855 enum lru_list lru;
Mel Gorman66635622010-08-09 17:19:30 -07001856
Kirill Tkhaia222f342019-05-13 17:17:00 -07001857 while (!list_empty(list)) {
1858 page = lru_to_page(list);
Sasha Levin309381fea2014-01-23 15:52:54 -08001859 VM_BUG_ON_PAGE(PageLRU(page), page);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001860 if (unlikely(!page_evictable(page))) {
Kirill Tkhaia222f342019-05-13 17:17:00 -07001861 list_del(&page->lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001862 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001863 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001864 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001865 continue;
1866 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001867 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001868
Linus Torvalds7a608572011-01-17 14:42:19 -08001869 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001870 lru = page_lru(page);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001871
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07001872 nr_pages = thp_nr_pages(page);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001873 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1874 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001875
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001876 if (put_page_testzero(page)) {
1877 __ClearPageLRU(page);
1878 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001879 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001880
1881 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001882 spin_unlock_irq(&pgdat->lru_lock);
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07001883 destroy_compound_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001884 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001885 } else
1886 list_add(&page->lru, &pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001887 } else {
1888 nr_moved += nr_pages;
Johannes Weiner31d8fca2020-06-25 20:30:31 -07001889 if (PageActive(page))
1890 workingset_age_nonresident(lruvec, nr_pages);
Mel Gorman66635622010-08-09 17:19:30 -07001891 }
1892 }
Mel Gorman66635622010-08-09 17:19:30 -07001893
Hugh Dickins3f797682012-01-12 17:20:07 -08001894 /*
1895 * To save our caller's stack, now use input list for pages to free.
1896 */
Kirill Tkhaia222f342019-05-13 17:17:00 -07001897 list_splice(&pages_to_free, list);
1898
1899 return nr_moved;
Mel Gorman66635622010-08-09 17:19:30 -07001900}
1901
1902/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001903 * If a kernel thread (such as nfsd for loop-back mounts) services
NeilBrowna37b0712020-06-01 21:48:18 -07001904 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
NeilBrown399ba0b2014-06-04 16:07:42 -07001905 * In that case we should only throttle if the backing device it is
1906 * writing to is congested. In other cases it is safe to throttle.
1907 */
1908static int current_may_throttle(void)
1909{
NeilBrowna37b0712020-06-01 21:48:18 -07001910 return !(current->flags & PF_LOCAL_THROTTLE) ||
NeilBrown399ba0b2014-06-04 16:07:42 -07001911 current->backing_dev_info == NULL ||
1912 bdi_write_congested(current->backing_dev_info);
1913}
1914
1915/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001916 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001917 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 */
Mel Gorman66635622010-08-09 17:19:30 -07001919static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001920shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001921 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
1923 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001924 unsigned long nr_scanned;
Maninder Singh730ec8c2020-06-03 16:01:18 -07001925 unsigned int nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001926 unsigned long nr_taken;
Kirill Tkhai060f0052019-03-05 15:48:15 -08001927 struct reclaim_stat stat;
Johannes Weiner497a6c12020-06-03 16:02:34 -07001928 bool file = is_file_lru(lru);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001929 enum vm_event_item item;
Mel Gorman599d0c92016-07-28 15:45:31 -07001930 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Michal Hockodb73ee02017-09-06 16:21:11 -07001931 bool stalled = false;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001932
Mel Gorman599d0c92016-07-28 15:45:31 -07001933 while (unlikely(too_many_isolated(pgdat, file, sc))) {
Michal Hockodb73ee02017-09-06 16:21:11 -07001934 if (stalled)
1935 return 0;
1936
1937 /* wait a bit for the reclaimer. */
1938 msleep(100);
1939 stalled = true;
Rik van Riel35cd7812009-09-21 17:01:38 -07001940
1941 /* We are about to die and free our memory. Return now. */
1942 if (fatal_signal_pending(current))
1943 return SWAP_CLUSTER_MAX;
1944 }
1945
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001947
Mel Gorman599d0c92016-07-28 15:45:31 -07001948 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001950 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001951 &nr_scanned, sc, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001952
Mel Gorman599d0c92016-07-28 15:45:31 -07001953 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001954 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08001955 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001956 __count_vm_events(item, nr_scanned);
1957 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
Johannes Weiner497a6c12020-06-03 16:02:34 -07001958 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1959
Mel Gorman599d0c92016-07-28 15:45:31 -07001960 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001961
Hillf Dantond563c052012-03-21 16:34:02 -07001962 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001963 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001964
Shakeel Buttdd156e32020-12-14 19:06:39 -08001965 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001966
Mel Gorman599d0c92016-07-28 15:45:31 -07001967 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001968
Johannes Weiner497a6c12020-06-03 16:02:34 -07001969 move_pages_to_lru(lruvec, &page_list);
1970
1971 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Johannes Weiner96f8bf42020-06-03 16:03:09 -07001972 lru_note_cost(lruvec, file, stat.nr_pageout);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001973 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08001974 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001975 __count_vm_events(item, nr_reclaimed);
1976 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
Johannes Weiner497a6c12020-06-03 16:02:34 -07001977 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
Hugh Dickins3f797682012-01-12 17:20:07 -08001978
Mel Gorman599d0c92016-07-28 15:45:31 -07001979 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001980
Johannes Weiner747db952014-08-08 14:19:24 -07001981 mem_cgroup_uncharge_list(&page_list);
Mel Gorman2d4894b2017-11-15 17:37:59 -08001982 free_unref_page_list(&page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001983
Mel Gorman92df3a72011-10-31 17:07:56 -07001984 /*
Andrey Ryabinin1c610d52018-03-22 16:17:42 -07001985 * If dirty pages are scanned that are not queued for IO, it
1986 * implies that flushers are not doing their job. This can
1987 * happen when memory pressure pushes dirty pages to the end of
1988 * the LRU before the dirty limits are breached and the dirty
1989 * data has expired. It can also happen when the proportion of
1990 * dirty pages grows not through writes but through memory
1991 * pressure reclaiming all the clean cache. And in some cases,
1992 * the flushers simply cannot keep up with the allocation
1993 * rate. Nudge the flusher threads in case they are asleep.
1994 */
1995 if (stat.nr_unqueued_dirty == nr_taken)
1996 wakeup_flusher_threads(WB_REASON_VMSCAN);
1997
Andrey Ryabinind108c772018-04-10 16:27:59 -07001998 sc->nr.dirty += stat.nr_dirty;
1999 sc->nr.congested += stat.nr_congested;
2000 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2001 sc->nr.writeback += stat.nr_writeback;
2002 sc->nr.immediate += stat.nr_immediate;
2003 sc->nr.taken += nr_taken;
2004 if (file)
2005 sc->nr.file_taken += nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07002006
Mel Gorman599d0c92016-07-28 15:45:31 -07002007 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
Steven Rostedtd51d1e62018-04-10 16:28:07 -07002008 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08002009 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010}
2011
Hugh Dickinsf6260122012-01-12 17:20:06 -08002012static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002013 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002014 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002015 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002017 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08002018 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07002019 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002021 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002022 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 struct page *page;
Michal Hocko9d998b42017-02-22 15:44:18 -08002024 unsigned nr_deactivate, nr_activate;
2025 unsigned nr_rotated = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002026 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002027 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
2029 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002030
Mel Gorman599d0c92016-07-28 15:45:31 -07002031 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002032
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002033 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002034 &nr_scanned, sc, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002035
Mel Gorman599d0c92016-07-28 15:45:31 -07002036 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002037
Shakeel Butt912c0572020-08-06 23:26:32 -07002038 if (!cgroup_reclaim(sc))
2039 __count_vm_events(PGREFILL, nr_scanned);
Yafang Shao2fa26902019-05-13 17:23:02 -07002040 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002041
Mel Gorman599d0c92016-07-28 15:45:31 -07002042 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 while (!list_empty(&l_hold)) {
2045 cond_resched();
2046 page = lru_to_page(&l_hold);
2047 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002048
Hugh Dickins39b5f292012-10-08 16:33:18 -07002049 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002050 putback_lru_page(page);
2051 continue;
2052 }
2053
Mel Gormancc715d92012-03-21 16:34:00 -07002054 if (unlikely(buffer_heads_over_limit)) {
2055 if (page_has_private(page) && trylock_page(page)) {
2056 if (page_has_private(page))
2057 try_to_release_page(page, 0);
2058 unlock_page(page);
2059 }
2060 }
2061
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002062 if (page_referenced(page, 0, sc->target_mem_cgroup,
2063 &vm_flags)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002064 /*
2065 * Identify referenced, file-backed active pages and
2066 * give them one more trip around the active list. So
2067 * that executable code get better chances to stay in
2068 * memory under moderate memory pressure. Anon pages
2069 * are not likely to be evicted by use-once streaming
2070 * IO, plus JVM can create lots of anon VM_EXEC pages,
2071 * so we ignore them here.
2072 */
Huang Ying9de4f222020-04-06 20:04:41 -07002073 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07002074 nr_rotated += thp_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002075 list_add(&page->lru, &l_active);
2076 continue;
2077 }
2078 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002079
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002080 ClearPageActive(page); /* we are de-activating */
Johannes Weiner1899ad12018-10-26 15:06:04 -07002081 SetPageWorkingset(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 list_add(&page->lru, &l_inactive);
2083 }
2084
Andrew Mortonb5557492009-01-06 14:40:13 -08002085 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002086 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002087 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002088 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel556adec2008-10-18 20:26:34 -07002089
Kirill Tkhaia222f342019-05-13 17:17:00 -07002090 nr_activate = move_pages_to_lru(lruvec, &l_active);
2091 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
Kirill Tkhaif372d892019-05-13 17:16:57 -07002092 /* Keep all free pages in l_active list */
2093 list_splice(&l_inactive, &l_active);
Kirill Tkhai9851ac12019-05-13 17:16:54 -07002094
2095 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2096 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2097
Mel Gorman599d0c92016-07-28 15:45:31 -07002098 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2099 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002100
Kirill Tkhaif372d892019-05-13 17:16:57 -07002101 mem_cgroup_uncharge_list(&l_active);
2102 free_unref_page_list(&l_active);
Michal Hocko9d998b42017-02-22 15:44:18 -08002103 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2104 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105}
2106
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002107unsigned long reclaim_pages(struct list_head *page_list)
2108{
Yang Shif661d002020-04-01 21:10:05 -07002109 int nid = NUMA_NO_NODE;
Maninder Singh730ec8c2020-06-03 16:01:18 -07002110 unsigned int nr_reclaimed = 0;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002111 LIST_HEAD(node_page_list);
2112 struct reclaim_stat dummy_stat;
2113 struct page *page;
2114 struct scan_control sc = {
2115 .gfp_mask = GFP_KERNEL,
2116 .priority = DEF_PRIORITY,
2117 .may_writepage = 1,
2118 .may_unmap = 1,
2119 .may_swap = 1,
2120 };
2121
2122 while (!list_empty(page_list)) {
2123 page = lru_to_page(page_list);
Yang Shif661d002020-04-01 21:10:05 -07002124 if (nid == NUMA_NO_NODE) {
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002125 nid = page_to_nid(page);
2126 INIT_LIST_HEAD(&node_page_list);
2127 }
2128
2129 if (nid == page_to_nid(page)) {
2130 ClearPageActive(page);
2131 list_move(&page->lru, &node_page_list);
2132 continue;
2133 }
2134
2135 nr_reclaimed += shrink_page_list(&node_page_list,
2136 NODE_DATA(nid),
Shakeel Buttdd156e32020-12-14 19:06:39 -08002137 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002138 while (!list_empty(&node_page_list)) {
2139 page = lru_to_page(&node_page_list);
2140 list_del(&page->lru);
2141 putback_lru_page(page);
2142 }
2143
Yang Shif661d002020-04-01 21:10:05 -07002144 nid = NUMA_NO_NODE;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002145 }
2146
2147 if (!list_empty(&node_page_list)) {
2148 nr_reclaimed += shrink_page_list(&node_page_list,
2149 NODE_DATA(nid),
Shakeel Buttdd156e32020-12-14 19:06:39 -08002150 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002151 while (!list_empty(&node_page_list)) {
2152 page = lru_to_page(&node_page_list);
2153 list_del(&page->lru);
2154 putback_lru_page(page);
2155 }
2156 }
2157
2158 return nr_reclaimed;
2159}
2160
Johannes Weinerb91ac372019-11-30 17:56:02 -08002161static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2162 struct lruvec *lruvec, struct scan_control *sc)
2163{
2164 if (is_active_lru(lru)) {
2165 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2166 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2167 else
2168 sc->skipped_deactivate = 1;
2169 return 0;
2170 }
2171
2172 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2173}
2174
Rik van Riel59dc76b2016-05-20 16:56:31 -07002175/*
2176 * The inactive anon list should be small enough that the VM never has
2177 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002178 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002179 * The inactive file list should be small enough to leave most memory
2180 * to the established workingset on the scan-resistant active list,
2181 * but large enough to avoid thrashing the aggregate readahead window.
2182 *
2183 * Both inactive lists should also be large enough that each inactive
2184 * page has a chance to be referenced again before it is reclaimed.
2185 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002186 * If that fails and refaulting is observed, the inactive list grows.
2187 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002188 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
Andrey Ryabinin3a50d142017-11-15 17:34:15 -08002189 * on this LRU, maintained by the pageout code. An inactive_ratio
Rik van Riel59dc76b2016-05-20 16:56:31 -07002190 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2191 *
2192 * total target max
2193 * memory ratio inactive
2194 * -------------------------------------
2195 * 10MB 1 5MB
2196 * 100MB 1 50MB
2197 * 1GB 3 250MB
2198 * 10GB 10 0.9GB
2199 * 100GB 31 3GB
2200 * 1TB 101 10GB
2201 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002202 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002203static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002204{
Johannes Weinerb91ac372019-11-30 17:56:02 -08002205 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002206 unsigned long inactive, active;
2207 unsigned long inactive_ratio;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002208 unsigned long gb;
2209
Johannes Weinerb91ac372019-11-30 17:56:02 -08002210 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2211 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002212
Johannes Weinerb91ac372019-11-30 17:56:02 -08002213 gb = (inactive + active) >> (30 - PAGE_SHIFT);
Joonsoo Kim40025702020-08-11 18:30:54 -07002214 if (gb)
Johannes Weinerb91ac372019-11-30 17:56:02 -08002215 inactive_ratio = int_sqrt(10 * gb);
2216 else
2217 inactive_ratio = 1;
Michal Hockofd538802017-02-22 15:45:58 -08002218
Rik van Riel59dc76b2016-05-20 16:56:31 -07002219 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002220}
2221
Johannes Weiner9a265112013-02-22 16:32:17 -08002222enum scan_balance {
2223 SCAN_EQUAL,
2224 SCAN_FRACT,
2225 SCAN_ANON,
2226 SCAN_FILE,
2227};
2228
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002230 * Determine how aggressively the anon and file LRU lists should be
2231 * scanned. The relative value of each set of LRU lists is determined
2232 * by looking at the fraction of the pages scanned we did rotate back
2233 * onto the active list instead of evict.
2234 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002235 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2236 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002237 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002238static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2239 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002240{
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002241 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002242 unsigned long anon_cost, file_cost, total_cost;
Vladimir Davydov33377672016-01-20 15:02:59 -08002243 int swappiness = mem_cgroup_swappiness(memcg);
Yu Zhaoed017372020-10-15 20:09:55 -07002244 u64 fraction[ANON_AND_FILE];
Johannes Weiner9a265112013-02-22 16:32:17 -08002245 u64 denominator = 0; /* gcc */
Johannes Weiner9a265112013-02-22 16:32:17 -08002246 enum scan_balance scan_balance;
Johannes Weiner9a265112013-02-22 16:32:17 -08002247 unsigned long ap, fp;
2248 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002249
2250 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002251 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002252 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002253 goto out;
2254 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002255
Johannes Weiner10316b32013-02-22 16:32:14 -08002256 /*
2257 * Global reclaim will swap to prevent OOM even with no
2258 * swappiness, but memcg users want to use this knob to
2259 * disable swapping for individual groups completely when
2260 * using the memory controller's swap limit feature would be
2261 * too expensive.
2262 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08002263 if (cgroup_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002264 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002265 goto out;
2266 }
2267
2268 /*
2269 * Do not apply any pressure balancing cleverness when the
2270 * system is close to OOM, scan both anon and file equally
2271 * (unless the swappiness setting disagrees with swapping).
2272 */
Johannes Weiner02695172014-08-06 16:06:17 -07002273 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002274 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002275 goto out;
2276 }
2277
Johannes Weiner11d16c22013-02-22 16:32:15 -08002278 /*
Johannes Weiner53138ce2019-11-30 17:55:56 -08002279 * If the system is almost out of file pages, force-scan anon.
Johannes Weiner62376252014-05-06 12:50:07 -07002280 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002281 if (sc->file_is_tiny) {
Johannes Weiner53138ce2019-11-30 17:55:56 -08002282 scan_balance = SCAN_ANON;
2283 goto out;
Johannes Weiner62376252014-05-06 12:50:07 -07002284 }
2285
2286 /*
Johannes Weinerb91ac372019-11-30 17:56:02 -08002287 * If there is enough inactive page cache, we do not reclaim
2288 * anything from the anonymous working right now.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002289 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002290 if (sc->cache_trim_mode) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002291 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002292 goto out;
2293 }
2294
Johannes Weiner9a265112013-02-22 16:32:17 -08002295 scan_balance = SCAN_FRACT;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002296 /*
Johannes Weiner314b57f2020-06-03 16:03:03 -07002297 * Calculate the pressure balance between anon and file pages.
2298 *
2299 * The amount of pressure we put on each LRU is inversely
2300 * proportional to the cost of reclaiming each list, as
2301 * determined by the share of pages that are refaulting, times
2302 * the relative IO cost of bringing back a swapped out
2303 * anonymous page vs reloading a filesystem page (swappiness).
2304 *
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002305 * Although we limit that influence to ensure no list gets
2306 * left behind completely: at least a third of the pressure is
2307 * applied, before swappiness.
2308 *
Johannes Weiner314b57f2020-06-03 16:03:03 -07002309 * With swappiness at 100, anon and file have equal IO cost.
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002310 */
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002311 total_cost = sc->anon_cost + sc->file_cost;
2312 anon_cost = total_cost + sc->anon_cost;
2313 file_cost = total_cost + sc->file_cost;
2314 total_cost = anon_cost + file_cost;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002315
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002316 ap = swappiness * (total_cost + 1);
2317 ap /= anon_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002318
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002319 fp = (200 - swappiness) * (total_cost + 1);
2320 fp /= file_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002321
Shaohua Li76a33fc2010-05-24 14:32:36 -07002322 fraction[0] = ap;
2323 fraction[1] = fp;
Johannes Weinera4fe1632020-06-03 16:02:50 -07002324 denominator = ap + fp;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002325out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002326 for_each_evictable_lru(lru) {
2327 int file = is_file_lru(lru);
Chris Down9783aa92019-10-06 17:58:32 -07002328 unsigned long lruvec_size;
Johannes Weiner8132fc22021-08-19 19:04:21 -07002329 unsigned long low, min;
Johannes Weiner688035f2017-05-03 14:52:07 -07002330 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002331
Chris Down9783aa92019-10-06 17:58:32 -07002332 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Johannes Weiner8132fc22021-08-19 19:04:21 -07002333 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2334 &min, &low);
Chris Down9783aa92019-10-06 17:58:32 -07002335
Johannes Weiner8132fc22021-08-19 19:04:21 -07002336 if (min || low) {
Chris Down9783aa92019-10-06 17:58:32 -07002337 /*
2338 * Scale a cgroup's reclaim pressure by proportioning
2339 * its current usage to its memory.low or memory.min
2340 * setting.
2341 *
2342 * This is important, as otherwise scanning aggression
2343 * becomes extremely binary -- from nothing as we
2344 * approach the memory protection threshold, to totally
2345 * nominal as we exceed it. This results in requiring
2346 * setting extremely liberal protection thresholds. It
2347 * also means we simply get no protection at all if we
2348 * set it too low, which is not ideal.
Chris Down1bc63fb2019-10-06 17:58:38 -07002349 *
2350 * If there is any protection in place, we reduce scan
2351 * pressure by how much of the total memory used is
2352 * within protection thresholds.
Chris Down9783aa92019-10-06 17:58:32 -07002353 *
Chris Down9de7ca42019-10-06 17:58:35 -07002354 * There is one special case: in the first reclaim pass,
2355 * we skip over all groups that are within their low
2356 * protection. If that fails to reclaim enough pages to
2357 * satisfy the reclaim goal, we come back and override
2358 * the best-effort low protection. However, we still
2359 * ideally want to honor how well-behaved groups are in
2360 * that case instead of simply punishing them all
2361 * equally. As such, we reclaim them based on how much
Chris Down1bc63fb2019-10-06 17:58:38 -07002362 * memory they are using, reducing the scan pressure
2363 * again by how much of the total memory used is under
2364 * hard protection.
Chris Down9783aa92019-10-06 17:58:32 -07002365 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002366 unsigned long cgroup_size = mem_cgroup_size(memcg);
Johannes Weiner8132fc22021-08-19 19:04:21 -07002367 unsigned long protection;
2368
2369 /* memory.low scaling, make sure we retry before OOM */
2370 if (!sc->memcg_low_reclaim && low > min) {
2371 protection = low;
2372 sc->memcg_low_skipped = 1;
2373 } else {
2374 protection = min;
2375 }
Chris Down1bc63fb2019-10-06 17:58:38 -07002376
2377 /* Avoid TOCTOU with earlier protection check */
2378 cgroup_size = max(cgroup_size, protection);
2379
2380 scan = lruvec_size - lruvec_size * protection /
Rik van Riel388f12d2021-09-08 18:10:08 -07002381 (cgroup_size + 1);
Chris Down9783aa92019-10-06 17:58:32 -07002382
2383 /*
Chris Down1bc63fb2019-10-06 17:58:38 -07002384 * Minimally target SWAP_CLUSTER_MAX pages to keep
Ethon Paul55b65a52020-06-04 16:49:10 -07002385 * reclaim moving forwards, avoiding decrementing
Chris Down9de7ca42019-10-06 17:58:35 -07002386 * sc->priority further than desirable.
Chris Down9783aa92019-10-06 17:58:32 -07002387 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002388 scan = max(scan, SWAP_CLUSTER_MAX);
Chris Down9783aa92019-10-06 17:58:32 -07002389 } else {
2390 scan = lruvec_size;
2391 }
2392
2393 scan >>= sc->priority;
2394
Johannes Weiner688035f2017-05-03 14:52:07 -07002395 /*
2396 * If the cgroup's already been deleted, make sure to
2397 * scrape out the remaining cache.
2398 */
2399 if (!scan && !mem_cgroup_online(memcg))
Chris Down9783aa92019-10-06 17:58:32 -07002400 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002401
Johannes Weiner688035f2017-05-03 14:52:07 -07002402 switch (scan_balance) {
2403 case SCAN_EQUAL:
2404 /* Scan lists relative to size */
2405 break;
2406 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002407 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002408 * Scan types proportional to swappiness and
2409 * their relative recent reclaim efficiency.
Gavin Shan76073c62020-02-20 20:04:24 -08002410 * Make sure we don't miss the last page on
2411 * the offlined memory cgroups because of a
2412 * round-off error.
Johannes Weiner9a265112013-02-22 16:32:17 -08002413 */
Gavin Shan76073c62020-02-20 20:04:24 -08002414 scan = mem_cgroup_online(memcg) ?
2415 div64_u64(scan * fraction[file], denominator) :
2416 DIV64_U64_ROUND_UP(scan * fraction[file],
Roman Gushchin68600f62018-10-26 15:03:27 -07002417 denominator);
Johannes Weiner688035f2017-05-03 14:52:07 -07002418 break;
2419 case SCAN_FILE:
2420 case SCAN_ANON:
2421 /* Scan one type exclusively */
Mateusz Noseke072bff2020-04-01 21:10:15 -07002422 if ((scan_balance == SCAN_FILE) != file)
Johannes Weiner688035f2017-05-03 14:52:07 -07002423 scan = 0;
Johannes Weiner688035f2017-05-03 14:52:07 -07002424 break;
2425 default:
2426 /* Look ma, no brain */
2427 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002428 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002429
Johannes Weiner688035f2017-05-03 14:52:07 -07002430 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002431 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002432}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002433
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002434static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002435{
2436 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002437 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002438 unsigned long nr_to_scan;
2439 enum lru_list lru;
2440 unsigned long nr_reclaimed = 0;
2441 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Johannes Weinerd925dd32022-08-02 12:28:11 -04002442 bool proportional_reclaim;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002443 struct blk_plug plug;
2444
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002445 get_scan_count(lruvec, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002446
Mel Gormane82e0562013-07-03 15:01:44 -07002447 /* Record the original scan target for proportional adjustments later */
2448 memcpy(targets, nr, sizeof(nr));
2449
Mel Gorman1a501902014-06-04 16:10:49 -07002450 /*
2451 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2452 * event that can occur when there is little memory pressure e.g.
2453 * multiple streaming readers/writers. Hence, we do not abort scanning
2454 * when the requested number of pages are reclaimed when scanning at
2455 * DEF_PRIORITY on the assumption that the fact we are direct
2456 * reclaiming implies that kswapd is not keeping up and it is best to
2457 * do a batch of work at once. For memcg reclaim one check is made to
2458 * abort proportional reclaim if either the file or anon lru has already
2459 * dropped to zero at the first pass.
2460 */
Johannes Weinerd925dd32022-08-02 12:28:11 -04002461 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2462 sc->priority == DEF_PRIORITY);
Mel Gorman1a501902014-06-04 16:10:49 -07002463
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002464 blk_start_plug(&plug);
2465 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2466 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002467 unsigned long nr_anon, nr_file, percentage;
2468 unsigned long nr_scanned;
2469
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002470 for_each_evictable_lru(lru) {
2471 if (nr[lru]) {
2472 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2473 nr[lru] -= nr_to_scan;
2474
2475 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner3b991202019-04-18 17:50:34 -07002476 lruvec, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002477 }
2478 }
Mel Gormane82e0562013-07-03 15:01:44 -07002479
Michal Hockobd041732016-12-02 17:26:48 -08002480 cond_resched();
2481
Johannes Weinerd925dd32022-08-02 12:28:11 -04002482 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
Mel Gormane82e0562013-07-03 15:01:44 -07002483 continue;
2484
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002485 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002486 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002487 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002488 * proportionally what was requested by get_scan_count(). We
2489 * stop reclaiming one LRU and reduce the amount scanning
2490 * proportional to the original scan target.
2491 */
2492 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2493 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2494
Mel Gorman1a501902014-06-04 16:10:49 -07002495 /*
2496 * It's just vindictive to attack the larger once the smaller
2497 * has gone to zero. And given the way we stop scanning the
2498 * smaller below, this makes sure that we only make one nudge
2499 * towards proportionality once we've got nr_to_reclaim.
2500 */
2501 if (!nr_file || !nr_anon)
2502 break;
2503
Mel Gormane82e0562013-07-03 15:01:44 -07002504 if (nr_file > nr_anon) {
2505 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2506 targets[LRU_ACTIVE_ANON] + 1;
2507 lru = LRU_BASE;
2508 percentage = nr_anon * 100 / scan_target;
2509 } else {
2510 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2511 targets[LRU_ACTIVE_FILE] + 1;
2512 lru = LRU_FILE;
2513 percentage = nr_file * 100 / scan_target;
2514 }
2515
2516 /* Stop scanning the smaller of the LRU */
2517 nr[lru] = 0;
2518 nr[lru + LRU_ACTIVE] = 0;
2519
2520 /*
2521 * Recalculate the other LRU scan count based on its original
2522 * scan target and the percentage scanning already complete
2523 */
2524 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2525 nr_scanned = targets[lru] - nr[lru];
2526 nr[lru] = targets[lru] * (100 - percentage) / 100;
2527 nr[lru] -= min(nr[lru], nr_scanned);
2528
2529 lru += LRU_ACTIVE;
2530 nr_scanned = targets[lru] - nr[lru];
2531 nr[lru] = targets[lru] * (100 - percentage) / 100;
2532 nr[lru] -= min(nr[lru], nr_scanned);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002533 }
2534 blk_finish_plug(&plug);
2535 sc->nr_reclaimed += nr_reclaimed;
2536
2537 /*
2538 * Even if we did not try to evict anon pages at all, we want to
2539 * rebalance the anon lru active/inactive ratio.
2540 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002541 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002542 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2543 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002544}
2545
Mel Gorman23b9da52012-05-29 15:06:20 -07002546/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002547static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002548{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002549 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002550 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002551 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002552 return true;
2553
2554 return false;
2555}
2556
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002557/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002558 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2559 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2560 * true if more pages should be reclaimed such that when the page allocator
Qiwu Chendf3a45f2020-06-03 16:01:21 -07002561 * calls try_to_compact_pages() that it will have enough free pages to succeed.
Mel Gorman23b9da52012-05-29 15:06:20 -07002562 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002563 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002564static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002565 unsigned long nr_reclaimed,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002566 struct scan_control *sc)
2567{
2568 unsigned long pages_for_compaction;
2569 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002570 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002571
2572 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002573 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002574 return false;
2575
Vlastimil Babka5ee04712019-09-23 15:37:29 -07002576 /*
2577 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2578 * number of pages that were scanned. This will return to the caller
2579 * with the risk reclaim/compaction and the resulting allocation attempt
2580 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2581 * allocations through requiring that the full LRU list has been scanned
2582 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2583 * scan, but that approximation was wrong, and there were corner cases
2584 * where always a non-zero amount of pages were scanned.
2585 */
2586 if (!nr_reclaimed)
2587 return false;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002588
Mel Gorman3e7d3442011-01-13 15:45:56 -08002589 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002590 for (z = 0; z <= sc->reclaim_idx; z++) {
2591 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002592 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002593 continue;
2594
2595 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002596 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002597 case COMPACT_CONTINUE:
2598 return false;
2599 default:
2600 /* check next zone */
2601 ;
2602 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002603 }
Hillf Danton1c6c1592019-09-23 15:37:26 -07002604
2605 /*
2606 * If we have not reclaimed enough pages for compaction and the
2607 * inactive lists are large enough, continue reclaiming
2608 */
2609 pages_for_compaction = compact_gap(sc->order);
2610 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2611 if (get_nr_swap_pages() > 0)
2612 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2613
Vlastimil Babka5ee04712019-09-23 15:37:29 -07002614 return inactive_lru_pages > pages_for_compaction;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002615}
2616
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002617static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002618{
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002619 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
Johannes Weinerd2af3392019-11-30 17:55:43 -08002620 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002621
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002622 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
Johannes Weiner56600482012-01-12 17:17:59 -08002623 do {
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002624 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Johannes Weinerd2af3392019-11-30 17:55:43 -08002625 unsigned long reclaimed;
2626 unsigned long scanned;
Johannes Weiner56600482012-01-12 17:17:59 -08002627
Xunlei Pange3336ca2020-09-04 16:35:27 -07002628 /*
2629 * This loop can become CPU-bound when target memcgs
2630 * aren't eligible for reclaim - either because they
2631 * don't have any reclaimable pages, or because their
2632 * memory is explicitly protected. Avoid soft lockups.
2633 */
2634 cond_resched();
2635
Chris Down45c7f7e2020-08-06 23:22:05 -07002636 mem_cgroup_calculate_protection(target_memcg, memcg);
2637
2638 if (mem_cgroup_below_min(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08002639 /*
2640 * Hard protection.
2641 * If there is no reclaimable memory, OOM.
2642 */
2643 continue;
Chris Down45c7f7e2020-08-06 23:22:05 -07002644 } else if (mem_cgroup_below_low(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08002645 /*
2646 * Soft protection.
2647 * Respect the protection only as long as
2648 * there is an unprotected supply
2649 * of reclaimable memory from other cgroups.
2650 */
2651 if (!sc->memcg_low_reclaim) {
2652 sc->memcg_low_skipped = 1;
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002653 continue;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002654 }
Johannes Weinerd2af3392019-11-30 17:55:43 -08002655 memcg_memory_event(memcg, MEMCG_LOW);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002656 }
2657
Johannes Weinerd2af3392019-11-30 17:55:43 -08002658 reclaimed = sc->nr_reclaimed;
2659 scanned = sc->nr_scanned;
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002660
2661 shrink_lruvec(lruvec, sc);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002662
Johannes Weinerd2af3392019-11-30 17:55:43 -08002663 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2664 sc->priority);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002665
Johannes Weinerd2af3392019-11-30 17:55:43 -08002666 /* Record the group's reclaim efficiency */
2667 vmpressure(sc->gfp_mask, memcg, false,
2668 sc->nr_scanned - scanned,
2669 sc->nr_reclaimed - reclaimed);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002670
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002671 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2672}
2673
Liu Song6c9e09072020-01-30 22:14:08 -08002674static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002675{
2676 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002677 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner1b051172019-11-30 17:55:52 -08002678 struct lruvec *target_lruvec;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002679 bool reclaimable = false;
Johannes Weinerb91ac372019-11-30 17:56:02 -08002680 unsigned long file;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002681
Johannes Weiner1b051172019-11-30 17:55:52 -08002682 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2683
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002684again:
2685 memset(&sc->nr, 0, sizeof(sc->nr));
2686
2687 nr_reclaimed = sc->nr_reclaimed;
2688 nr_scanned = sc->nr_scanned;
2689
Johannes Weiner53138ce2019-11-30 17:55:56 -08002690 /*
Johannes Weiner7cf111b2020-06-03 16:03:06 -07002691 * Determine the scan balance between anon and file LRUs.
2692 */
2693 spin_lock_irq(&pgdat->lru_lock);
2694 sc->anon_cost = target_lruvec->anon_cost;
2695 sc->file_cost = target_lruvec->file_cost;
2696 spin_unlock_irq(&pgdat->lru_lock);
2697
2698 /*
Johannes Weinerb91ac372019-11-30 17:56:02 -08002699 * Target desirable inactive:active list ratios for the anon
2700 * and file LRU lists.
2701 */
2702 if (!sc->force_deactivate) {
2703 unsigned long refaults;
2704
Joonsoo Kim170b04b72020-08-11 18:30:43 -07002705 refaults = lruvec_page_state(target_lruvec,
2706 WORKINGSET_ACTIVATE_ANON);
2707 if (refaults != target_lruvec->refaults[0] ||
2708 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
Johannes Weinerb91ac372019-11-30 17:56:02 -08002709 sc->may_deactivate |= DEACTIVATE_ANON;
2710 else
2711 sc->may_deactivate &= ~DEACTIVATE_ANON;
2712
2713 /*
2714 * When refaults are being observed, it means a new
2715 * workingset is being established. Deactivate to get
2716 * rid of any stale active pages quickly.
2717 */
2718 refaults = lruvec_page_state(target_lruvec,
Joonsoo Kim170b04b72020-08-11 18:30:43 -07002719 WORKINGSET_ACTIVATE_FILE);
2720 if (refaults != target_lruvec->refaults[1] ||
Johannes Weinerb91ac372019-11-30 17:56:02 -08002721 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2722 sc->may_deactivate |= DEACTIVATE_FILE;
2723 else
2724 sc->may_deactivate &= ~DEACTIVATE_FILE;
2725 } else
2726 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2727
2728 /*
2729 * If we have plenty of inactive file pages that aren't
2730 * thrashing, try to reclaim those first before touching
2731 * anonymous pages.
2732 */
2733 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2734 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2735 sc->cache_trim_mode = 1;
2736 else
2737 sc->cache_trim_mode = 0;
2738
2739 /*
Johannes Weiner53138ce2019-11-30 17:55:56 -08002740 * Prevent the reclaimer from falling into the cache trap: as
2741 * cache pages start out inactive, every cache fault will tip
2742 * the scan balance towards the file LRU. And as the file LRU
2743 * shrinks, so does the window for rotation from references.
2744 * This means we have a runaway feedback loop where a tiny
2745 * thrashing file LRU becomes infinitely more attractive than
2746 * anon pages. Try to detect this based on file LRU size.
2747 */
2748 if (!cgroup_reclaim(sc)) {
Johannes Weiner53138ce2019-11-30 17:55:56 -08002749 unsigned long total_high_wmark = 0;
Johannes Weinerb91ac372019-11-30 17:56:02 -08002750 unsigned long free, anon;
2751 int z;
Johannes Weiner53138ce2019-11-30 17:55:56 -08002752
2753 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2754 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2755 node_page_state(pgdat, NR_INACTIVE_FILE);
2756
2757 for (z = 0; z < MAX_NR_ZONES; z++) {
2758 struct zone *zone = &pgdat->node_zones[z];
2759 if (!managed_zone(zone))
2760 continue;
2761
2762 total_high_wmark += high_wmark_pages(zone);
2763 }
2764
Johannes Weinerb91ac372019-11-30 17:56:02 -08002765 /*
2766 * Consider anon: if that's low too, this isn't a
2767 * runaway file reclaim problem, but rather just
2768 * extreme pressure. Reclaim as per usual then.
2769 */
2770 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2771
2772 sc->file_is_tiny =
2773 file + free <= total_high_wmark &&
2774 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2775 anon >> sc->priority;
Johannes Weiner53138ce2019-11-30 17:55:56 -08002776 }
2777
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002778 shrink_node_memcgs(pgdat, sc);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002779
Johannes Weinerd2af3392019-11-30 17:55:43 -08002780 if (reclaim_state) {
2781 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2782 reclaim_state->reclaimed_slab = 0;
2783 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07002784
Johannes Weinerd2af3392019-11-30 17:55:43 -08002785 /* Record the subtree's reclaim efficiency */
Johannes Weiner1b051172019-11-30 17:55:52 -08002786 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Johannes Weinerd2af3392019-11-30 17:55:43 -08002787 sc->nr_scanned - nr_scanned,
2788 sc->nr_reclaimed - nr_reclaimed);
2789
2790 if (sc->nr_reclaimed - nr_reclaimed)
2791 reclaimable = true;
2792
2793 if (current_is_kswapd()) {
2794 /*
2795 * If reclaim is isolating dirty pages under writeback,
2796 * it implies that the long-lived page allocation rate
2797 * is exceeding the page laundering rate. Either the
2798 * global limits are not being effective at throttling
2799 * processes due to the page distribution throughout
2800 * zones or there is heavy usage of a slow backing
2801 * device. The only option is to throttle from reclaim
2802 * context which is not ideal as there is no guarantee
2803 * the dirtying process is throttled in the same way
2804 * balance_dirty_pages() manages.
2805 *
2806 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2807 * count the number of pages under pages flagged for
2808 * immediate reclaim and stall if any are encountered
2809 * in the nr_immediate check below.
2810 */
2811 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2812 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002813
Johannes Weinerd2af3392019-11-30 17:55:43 -08002814 /* Allow kswapd to start writing pages during reclaim.*/
2815 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2816 set_bit(PGDAT_DIRTY, &pgdat->flags);
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002817
2818 /*
Randy Dunlap1eba09c2020-08-11 18:33:26 -07002819 * If kswapd scans pages marked for immediate
Johannes Weinerd2af3392019-11-30 17:55:43 -08002820 * reclaim and under writeback (nr_immediate), it
2821 * implies that pages are cycling through the LRU
2822 * faster than they are written so also forcibly stall.
Andrey Ryabinind108c772018-04-10 16:27:59 -07002823 */
Johannes Weinerd2af3392019-11-30 17:55:43 -08002824 if (sc->nr.immediate)
2825 congestion_wait(BLK_RW_ASYNC, HZ/10);
2826 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07002827
Johannes Weinerd2af3392019-11-30 17:55:43 -08002828 /*
Johannes Weiner1b051172019-11-30 17:55:52 -08002829 * Tag a node/memcg as congested if all the dirty pages
2830 * scanned were backed by a congested BDI and
2831 * wait_iff_congested will stall.
2832 *
Johannes Weinerd2af3392019-11-30 17:55:43 -08002833 * Legacy memcg will stall in page writeback so avoid forcibly
2834 * stalling in wait_iff_congested().
2835 */
Johannes Weiner1b051172019-11-30 17:55:52 -08002836 if ((current_is_kswapd() ||
2837 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
Johannes Weinerd2af3392019-11-30 17:55:43 -08002838 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
Johannes Weiner1b051172019-11-30 17:55:52 -08002839 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
Johannes Weinerd2af3392019-11-30 17:55:43 -08002840
2841 /*
2842 * Stall direct reclaim for IO completions if underlying BDIs
2843 * and node is congested. Allow kswapd to continue until it
2844 * starts encountering unqueued dirty pages or cycling through
2845 * the LRU too quickly.
2846 */
Johannes Weiner1b051172019-11-30 17:55:52 -08002847 if (!current_is_kswapd() && current_may_throttle() &&
2848 !sc->hibernation_mode &&
2849 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
Johannes Weinerd2af3392019-11-30 17:55:43 -08002850 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2851
2852 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2853 sc))
2854 goto again;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002855
Johannes Weinerc73322d2017-05-03 14:51:51 -07002856 /*
2857 * Kswapd gives up on balancing particular nodes after too
2858 * many failures to reclaim anything from them and goes to
2859 * sleep. On reclaim progress, reset the failure counter. A
2860 * successful direct reclaim run will revive a dormant kswapd.
2861 */
2862 if (reclaimable)
2863 pgdat->kswapd_failures = 0;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002864}
2865
Vlastimil Babka53853e22014-10-09 15:27:02 -07002866/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002867 * Returns true if compaction should go ahead for a costly-order request, or
2868 * the allocation would already succeed without compaction. Return false if we
2869 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002870 */
Mel Gorman4f588332016-07-28 15:46:38 -07002871static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002872{
Mel Gorman31483b62016-07-28 15:45:46 -07002873 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002874 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002875
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002876 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2877 if (suitable == COMPACT_SUCCESS)
2878 /* Allocation should succeed already. Don't reclaim. */
2879 return true;
2880 if (suitable == COMPACT_SKIPPED)
2881 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002882 return false;
2883
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002884 /*
2885 * Compaction is already possible, but it takes time to run and there
2886 * are potentially other callers using the pages just freed. So proceed
2887 * with reclaim to make a buffer of free pages available to give
2888 * compaction a reasonable chance of completing and allocating the page.
2889 * Note that we won't actually reclaim the whole buffer in one attempt
2890 * as the target watermark in should_continue_reclaim() is lower. But if
2891 * we are already above the high+gap watermark, don't reclaim at all.
2892 */
2893 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2894
2895 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002896}
2897
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898/*
2899 * This is the direct reclaim path, for page-allocating processes. We only
2900 * try to reclaim pages from zones which will satisfy the caller's allocation
2901 * request.
2902 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 * If a zone is deemed to be full of pinned pages then just give it a light
2904 * scan then give up on it.
2905 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002906static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907{
Mel Gormandd1a2392008-04-28 02:12:17 -07002908 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002909 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002910 unsigned long nr_soft_reclaimed;
2911 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002912 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002913 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002914
Mel Gormancc715d92012-03-21 16:34:00 -07002915 /*
2916 * If the number of buffer_heads in the machine exceeds the maximum
2917 * allowed level, force direct reclaim to scan the highmem zone as
2918 * highmem pages could be pinning lowmem pages storing buffer_heads
2919 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002920 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002921 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002922 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002923 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002924 }
Mel Gormancc715d92012-03-21 16:34:00 -07002925
Mel Gormand4debc62010-08-09 17:19:29 -07002926 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002927 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002928 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002929 * Take care memory controller reclaiming has small influence
2930 * to global LRU.
2931 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08002932 if (!cgroup_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002933 if (!cpuset_zone_allowed(zone,
2934 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002935 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002936
Johannes Weiner0b064962014-08-06 16:06:12 -07002937 /*
2938 * If we already have plenty of memory free for
2939 * compaction in this zone, don't free any more.
2940 * Even though compaction is invoked for any
2941 * non-zero order, only frequent costly order
2942 * reclamation is disruptive enough to become a
2943 * noticeable problem, like transparent huge
2944 * page allocations.
2945 */
2946 if (IS_ENABLED(CONFIG_COMPACTION) &&
2947 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002948 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002949 sc->compaction_ready = true;
2950 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002951 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002952
Andrew Morton0608f432013-09-24 15:27:41 -07002953 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002954 * Shrink each node in the zonelist once. If the
2955 * zonelist is ordered by zone (not the default) then a
2956 * node may be shrunk multiple times but in that case
2957 * the user prefers lower zones being preserved.
2958 */
2959 if (zone->zone_pgdat == last_pgdat)
2960 continue;
2961
2962 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002963 * This steals pages from memory cgroups over softlimit
2964 * and returns the number of reclaimed pages and
2965 * scanned pages. This works for global memory pressure
2966 * and balancing, not for a memcg's limit.
2967 */
2968 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002969 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002970 sc->order, sc->gfp_mask,
2971 &nr_soft_scanned);
2972 sc->nr_reclaimed += nr_soft_reclaimed;
2973 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002974 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002975 }
Nick Piggin408d8542006-09-25 23:31:27 -07002976
Mel Gorman79dafcd2016-07-28 15:45:53 -07002977 /* See comment about same check for global reclaim above */
2978 if (zone->zone_pgdat == last_pgdat)
2979 continue;
2980 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002981 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 }
Mel Gormane0c23272011-10-31 17:09:33 -07002983
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002984 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002985 * Restore to original mask to avoid the impact on the caller if we
2986 * promoted it to __GFP_HIGHMEM.
2987 */
2988 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002990
Johannes Weinerb9107182019-11-30 17:55:59 -08002991static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002992{
Johannes Weinerb9107182019-11-30 17:55:59 -08002993 struct lruvec *target_lruvec;
2994 unsigned long refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002995
Johannes Weinerb9107182019-11-30 17:55:59 -08002996 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
Joonsoo Kim170b04b72020-08-11 18:30:43 -07002997 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2998 target_lruvec->refaults[0] = refaults;
2999 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3000 target_lruvec->refaults[1] = refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003001}
3002
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003/*
3004 * This is the main entry point to direct page reclaim.
3005 *
3006 * If a full scan of the inactive list fails to free enough memory then we
3007 * are "out of memory" and something needs to be killed.
3008 *
3009 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3010 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02003011 * caller can't do much about. We kick the writeback threads and take explicit
3012 * naps in the hope that some of these pages can be written. But if the
3013 * allocating task holds filesystem locks which prevent writeout this might not
3014 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07003015 *
3016 * returns: 0, if no pages reclaimed
3017 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 */
Mel Gormandac1d272008-04-28 02:12:12 -07003019static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003020 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021{
Johannes Weiner241994ed2015-02-11 15:26:06 -08003022 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003023 pg_data_t *last_pgdat;
3024 struct zoneref *z;
3025 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003026retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07003027 delayacct_freepages_start();
3028
Johannes Weinerb5ead352019-11-30 17:55:40 -08003029 if (!cgroup_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07003030 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003032 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07003033 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3034 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08003035 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07003036 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07003037
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003038 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07003039 break;
3040
3041 if (sc->compaction_ready)
3042 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
3044 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08003045 * If we're getting trouble reclaiming, start doing
3046 * writepage even in laptop mode.
3047 */
3048 if (sc->priority < DEF_PRIORITY - 2)
3049 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07003050 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003051
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003052 last_pgdat = NULL;
3053 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3054 sc->nodemask) {
3055 if (zone->zone_pgdat == last_pgdat)
3056 continue;
3057 last_pgdat = zone->zone_pgdat;
Johannes Weiner1b051172019-11-30 17:55:52 -08003058
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003059 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
Johannes Weiner1b051172019-11-30 17:55:52 -08003060
3061 if (cgroup_reclaim(sc)) {
3062 struct lruvec *lruvec;
3063
3064 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3065 zone->zone_pgdat);
3066 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3067 }
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003068 }
3069
Keika Kobayashi873b4772008-07-25 01:48:52 -07003070 delayacct_freepages_end();
3071
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003072 if (sc->nr_reclaimed)
3073 return sc->nr_reclaimed;
3074
Mel Gorman0cee34f2012-01-12 17:19:49 -08003075 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07003076 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08003077 return 1;
3078
Johannes Weinerb91ac372019-11-30 17:56:02 -08003079 /*
3080 * We make inactive:active ratio decisions based on the node's
3081 * composition of memory, but a restrictive reclaim_idx or a
3082 * memory.low cgroup setting can exempt large amounts of
3083 * memory from reclaim. Neither of which are very common, so
3084 * instead of doing costly eligibility calculations of the
3085 * entire cgroup subtree up front, we assume the estimates are
3086 * good, and retry with forcible deactivation if that fails.
3087 */
3088 if (sc->skipped_deactivate) {
3089 sc->priority = initial_priority;
3090 sc->force_deactivate = 1;
3091 sc->skipped_deactivate = 0;
3092 goto retry;
3093 }
3094
Johannes Weiner241994ed2015-02-11 15:26:06 -08003095 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07003096 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08003097 sc->priority = initial_priority;
Johannes Weinerb91ac372019-11-30 17:56:02 -08003098 sc->force_deactivate = 0;
Yisheng Xied6622f62017-05-03 14:53:57 -07003099 sc->memcg_low_reclaim = 1;
3100 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003101 goto retry;
3102 }
3103
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003104 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105}
3106
Johannes Weinerc73322d2017-05-03 14:51:51 -07003107static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07003108{
3109 struct zone *zone;
3110 unsigned long pfmemalloc_reserve = 0;
3111 unsigned long free_pages = 0;
3112 int i;
3113 bool wmark_ok;
3114
Johannes Weinerc73322d2017-05-03 14:51:51 -07003115 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3116 return true;
3117
Mel Gorman55150612012-07-31 16:44:35 -07003118 for (i = 0; i <= ZONE_NORMAL; i++) {
3119 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd2017-05-03 14:51:54 -07003120 if (!managed_zone(zone))
3121 continue;
3122
3123 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07003124 continue;
3125
Mel Gorman55150612012-07-31 16:44:35 -07003126 pfmemalloc_reserve += min_wmark_pages(zone);
3127 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3128 }
3129
Mel Gorman675becc2014-06-04 16:07:35 -07003130 /* If there are no reserves (unexpected config) then do not throttle */
3131 if (!pfmemalloc_reserve)
3132 return true;
3133
Mel Gorman55150612012-07-31 16:44:35 -07003134 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3135
3136 /* kswapd must be awake if processes are being throttled */
3137 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003138 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3139 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003140
Mel Gorman55150612012-07-31 16:44:35 -07003141 wake_up_interruptible(&pgdat->kswapd_wait);
3142 }
3143
3144 return wmark_ok;
3145}
3146
3147/*
3148 * Throttle direct reclaimers if backing storage is backed by the network
3149 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3150 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08003151 * when the low watermark is reached.
3152 *
3153 * Returns true if a fatal signal was delivered during throttling. If this
3154 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07003155 */
Mel Gorman50694c22012-11-26 16:29:48 -08003156static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07003157 nodemask_t *nodemask)
3158{
Mel Gorman675becc2014-06-04 16:07:35 -07003159 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07003160 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07003161 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07003162
3163 /*
3164 * Kernel threads should not be throttled as they may be indirectly
3165 * responsible for cleaning pages necessary for reclaim to make forward
3166 * progress. kjournald for example may enter direct reclaim while
3167 * committing a transaction where throttling it could forcing other
3168 * processes to block on log_wait_commit().
3169 */
3170 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08003171 goto out;
3172
3173 /*
3174 * If a fatal signal is pending, this process should not throttle.
3175 * It should return quickly so it can exit and free its memory
3176 */
3177 if (fatal_signal_pending(current))
3178 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003179
Mel Gorman675becc2014-06-04 16:07:35 -07003180 /*
3181 * Check if the pfmemalloc reserves are ok by finding the first node
3182 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3183 * GFP_KERNEL will be required for allocating network buffers when
3184 * swapping over the network so ZONE_HIGHMEM is unusable.
3185 *
3186 * Throttling is based on the first usable node and throttled processes
3187 * wait on a queue until kswapd makes progress and wakes them. There
3188 * is an affinity then between processes waking up and where reclaim
3189 * progress has been made assuming the process wakes on the same node.
3190 * More importantly, processes running on remote nodes will not compete
3191 * for remote pfmemalloc reserves and processes on different nodes
3192 * should make reasonable progress.
3193 */
3194 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08003195 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07003196 if (zone_idx(zone) > ZONE_NORMAL)
3197 continue;
3198
3199 /* Throttle based on the first usable node */
3200 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07003201 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07003202 goto out;
3203 break;
3204 }
3205
3206 /* If no zone was usable by the allocation flags then do not throttle */
3207 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08003208 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003209
Mel Gorman68243e72012-07-31 16:44:39 -07003210 /* Account for the throttling */
3211 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3212
Mel Gorman55150612012-07-31 16:44:35 -07003213 /*
3214 * If the caller cannot enter the filesystem, it's possible that it
3215 * is due to the caller holding an FS lock or performing a journal
3216 * transaction in the case of a filesystem like ext[3|4]. In this case,
3217 * it is not safe to block on pfmemalloc_wait as kswapd could be
3218 * blocked waiting on the same lock. Instead, throttle for up to a
3219 * second before continuing.
3220 */
3221 if (!(gfp_mask & __GFP_FS)) {
3222 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003223 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003224
3225 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003226 }
3227
3228 /* Throttle until kswapd wakes the process */
3229 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003230 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003231
3232check_pending:
3233 if (fatal_signal_pending(current))
3234 return true;
3235
3236out:
3237 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003238}
3239
Mel Gormandac1d272008-04-28 02:12:12 -07003240unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003241 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003242{
Mel Gorman33906bc2010-08-09 17:19:16 -07003243 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003244 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003245 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003246 .gfp_mask = current_gfp_context(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07003247 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003248 .order = order,
3249 .nodemask = nodemask,
3250 .priority = DEF_PRIORITY,
3251 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003252 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003253 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003254 };
3255
Mel Gorman55150612012-07-31 16:44:35 -07003256 /*
Greg Thelenbb451fd2018-08-17 15:45:19 -07003257 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3258 * Confirm they are large enough for max values.
3259 */
3260 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3261 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3262 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3263
3264 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003265 * Do not enter reclaim if fatal signal was delivered while throttled.
3266 * 1 is returned so that the page allocator does not OOM kill at this
3267 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003268 */
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003269 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003270 return 1;
3271
Andrew Morton1732d2b012019-07-16 16:26:15 -07003272 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07003273 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
Mel Gorman33906bc2010-08-09 17:19:16 -07003274
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003275 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003276
3277 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003278 set_task_reclaim_state(current, NULL);
Mel Gorman33906bc2010-08-09 17:19:16 -07003279
3280 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003281}
3282
Andrew Mortonc255a452012-07-31 16:43:02 -07003283#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003284
Michal Hockod2e5fb92019-08-30 16:04:50 -07003285/* Only used by soft limit reclaim. Do not reuse for anything else. */
Mel Gormana9dd0a82016-07-28 15:46:02 -07003286unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003287 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003288 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003289 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003290{
Johannes Weinerafaf07a2019-11-30 17:55:46 -08003291 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Balbir Singh4e416952009-09-23 15:56:39 -07003292 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003293 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003294 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003295 .may_writepage = !laptop_mode,
3296 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003297 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003298 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003299 };
Ying Han0ae5e892011-05-26 16:25:25 -07003300
Michal Hockod2e5fb92019-08-30 16:04:50 -07003301 WARN_ON_ONCE(!current->reclaim_state);
3302
Balbir Singh4e416952009-09-23 15:56:39 -07003303 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3304 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003305
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003306 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
Yafang Shao3481c372019-05-13 17:19:14 -07003307 sc.gfp_mask);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003308
Balbir Singh4e416952009-09-23 15:56:39 -07003309 /*
3310 * NOTE: Although we can get the priority field, using it
3311 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003312 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003313 * will pick up pages from other mem cgroup's as well. We hack
3314 * the priority and make it zero.
3315 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08003316 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003317
3318 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3319
Ying Han0ae5e892011-05-26 16:25:25 -07003320 *nr_scanned = sc.nr_scanned;
Yafang Shao0308f7c2019-07-16 16:26:12 -07003321
Balbir Singh4e416952009-09-23 15:56:39 -07003322 return sc.nr_reclaimed;
3323}
3324
Johannes Weiner72835c82012-01-12 17:18:32 -08003325unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003326 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003327 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003328 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003329{
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003330 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003331 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08003332 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003333 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003334 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07003335 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003336 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003337 .target_mem_cgroup = memcg,
3338 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003339 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003340 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003341 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003342 };
Shakeel Buttfa40d1e2019-11-30 17:50:16 -08003343 /*
3344 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3345 * equal pressure on all the nodes. This is based on the assumption that
3346 * the reclaim does not bail out early.
3347 */
3348 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
Balbir Singh66e17072008-02-07 00:13:56 -08003349
Andrew Morton1732d2b012019-07-16 16:26:15 -07003350 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07003351 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003352 noreclaim_flag = memalloc_noreclaim_save();
Johannes Weinereb414682018-10-26 15:06:27 -07003353
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003354 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weinereb414682018-10-26 15:06:27 -07003355
Vlastimil Babka499118e2017-05-08 15:59:50 -07003356 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003357 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003358 set_task_reclaim_state(current, NULL);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003359
3360 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003361}
3362#endif
3363
Mel Gorman1d82de62016-07-28 15:45:43 -07003364static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003365 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003366{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003367 struct mem_cgroup *memcg;
Johannes Weinerb91ac372019-11-30 17:56:02 -08003368 struct lruvec *lruvec;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003369
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003370 if (!total_swap_pages)
3371 return;
3372
Johannes Weinerb91ac372019-11-30 17:56:02 -08003373 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3374 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3375 return;
3376
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003377 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3378 do {
Johannes Weinerb91ac372019-11-30 17:56:02 -08003379 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3380 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3381 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003382 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3383 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003384}
3385
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003386static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
Mel Gorman1c308442018-12-28 00:35:52 -08003387{
3388 int i;
3389 struct zone *zone;
3390
3391 /*
3392 * Check for watermark boosts top-down as the higher zones
3393 * are more likely to be boosted. Both watermarks and boosts
Randy Dunlap1eba09c2020-08-11 18:33:26 -07003394 * should not be checked at the same time as reclaim would
Mel Gorman1c308442018-12-28 00:35:52 -08003395 * start prematurely when there is no boosting and a lower
3396 * zone is balanced.
3397 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003398 for (i = highest_zoneidx; i >= 0; i--) {
Mel Gorman1c308442018-12-28 00:35:52 -08003399 zone = pgdat->node_zones + i;
3400 if (!managed_zone(zone))
3401 continue;
3402
3403 if (zone->watermark_boost)
3404 return true;
3405 }
3406
3407 return false;
3408}
3409
Mel Gormane716f2e2017-05-03 14:53:45 -07003410/*
3411 * Returns true if there is an eligible zone balanced for the request order
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003412 * and highest_zoneidx
Mel Gormane716f2e2017-05-03 14:53:45 -07003413 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003414static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003415{
Mel Gormane716f2e2017-05-03 14:53:45 -07003416 int i;
3417 unsigned long mark = -1;
3418 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003419
Mel Gorman1c308442018-12-28 00:35:52 -08003420 /*
3421 * Check watermarks bottom-up as lower zones are more likely to
3422 * meet watermarks.
3423 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003424 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003425 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07003426
Mel Gormane716f2e2017-05-03 14:53:45 -07003427 if (!managed_zone(zone))
3428 continue;
3429
3430 mark = high_wmark_pages(zone);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003431 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
Mel Gormane716f2e2017-05-03 14:53:45 -07003432 return true;
3433 }
3434
3435 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003436 * If a node has no populated zone within highest_zoneidx, it does not
Mel Gormane716f2e2017-05-03 14:53:45 -07003437 * need balancing by definition. This can happen if a zone-restricted
3438 * allocation tries to wake a remote kswapd.
3439 */
3440 if (mark == -1)
3441 return true;
3442
3443 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003444}
3445
Mel Gorman631b6e02017-05-03 14:53:41 -07003446/* Clear pgdat state for congested, dirty or under writeback. */
3447static void clear_pgdat_congested(pg_data_t *pgdat)
3448{
Johannes Weiner1b051172019-11-30 17:55:52 -08003449 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3450
3451 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
Mel Gorman631b6e02017-05-03 14:53:41 -07003452 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3453 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3454}
3455
Mel Gorman1741c872011-01-13 15:46:21 -08003456/*
Mel Gorman55150612012-07-31 16:44:35 -07003457 * Prepare kswapd for sleeping. This verifies that there are no processes
3458 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3459 *
3460 * Returns true if kswapd is ready to sleep
3461 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003462static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3463 int highest_zoneidx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003464{
Mel Gorman55150612012-07-31 16:44:35 -07003465 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003466 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07003467 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003468 * race between when kswapd checks the watermarks and a process gets
3469 * throttled. There is also a potential race if processes get
3470 * throttled, kswapd wakes, a large process exits thereby balancing the
3471 * zones, which causes kswapd to exit balance_pgdat() before reaching
3472 * the wake up checks. If kswapd is going to sleep, no process should
3473 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3474 * the wake up is premature, processes will wake kswapd and get
3475 * throttled again. The difference from wake ups in balance_pgdat() is
3476 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003477 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003478 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3479 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003480
Johannes Weinerc73322d2017-05-03 14:51:51 -07003481 /* Hopeless node, leave it to direct reclaim */
3482 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3483 return true;
3484
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003485 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003486 clear_pgdat_congested(pgdat);
3487 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003488 }
3489
Shantanu Goel333b0a42017-05-03 14:53:38 -07003490 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003491}
3492
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003494 * kswapd shrinks a node of pages that are at or below the highest usable
3495 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003496 *
3497 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003498 * reclaim or if the lack of progress was due to pages under writeback.
3499 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003500 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003501static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003502 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003503{
Mel Gorman1d82de62016-07-28 15:45:43 -07003504 struct zone *zone;
3505 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003506
Mel Gorman1d82de62016-07-28 15:45:43 -07003507 /* Reclaim a number of pages proportional to the number of zones */
3508 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003509 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003510 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003511 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003512 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003513
Mel Gorman1d82de62016-07-28 15:45:43 -07003514 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003515 }
3516
Mel Gorman1d82de62016-07-28 15:45:43 -07003517 /*
3518 * Historically care was taken to put equal pressure on all zones but
3519 * now pressure is applied based on node LRU order.
3520 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003521 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003522
3523 /*
3524 * Fragmentation may mean that the system cannot be rebalanced for
3525 * high-order allocations. If twice the allocation size has been
3526 * reclaimed then recheck watermarks only at order-0 to prevent
3527 * excessive reclaim. Assume that a process requested a high-order
3528 * can direct reclaim/compact.
3529 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003530 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003531 sc->order = 0;
3532
Mel Gormanb8e83b92013-07-03 15:01:45 -07003533 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003534}
3535
3536/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003537 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3538 * that are eligible for use by the caller until at least one zone is
3539 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003541 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 *
3543 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003544 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Wei Yang8bb4e7a2019-03-05 15:46:22 -08003545 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
Mel Gorman1d82de62016-07-28 15:45:43 -07003546 * or lower is eligible for reclaim until at least one usable zone is
3547 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003549static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003552 unsigned long nr_soft_reclaimed;
3553 unsigned long nr_soft_scanned;
Johannes Weinereb414682018-10-26 15:06:27 -07003554 unsigned long pflags;
Mel Gorman1c308442018-12-28 00:35:52 -08003555 unsigned long nr_boost_reclaim;
3556 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3557 bool boosted;
Mel Gorman1d82de62016-07-28 15:45:43 -07003558 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003559 struct scan_control sc = {
3560 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003561 .order = order,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003562 .may_unmap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003563 };
Omar Sandoval93781322018-06-07 17:07:02 -07003564
Andrew Morton1732d2b012019-07-16 16:26:15 -07003565 set_task_reclaim_state(current, &sc.reclaim_state);
Johannes Weinereb414682018-10-26 15:06:27 -07003566 psi_memstall_enter(&pflags);
Omar Sandoval93781322018-06-07 17:07:02 -07003567 __fs_reclaim_acquire();
3568
Christoph Lameterf8891e52006-06-30 01:55:45 -07003569 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570
Mel Gorman1c308442018-12-28 00:35:52 -08003571 /*
3572 * Account for the reclaim boost. Note that the zone boost is left in
3573 * place so that parallel allocations that are near the watermark will
3574 * stall or direct reclaim until kswapd is finished.
3575 */
3576 nr_boost_reclaim = 0;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003577 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08003578 zone = pgdat->node_zones + i;
3579 if (!managed_zone(zone))
3580 continue;
3581
3582 nr_boost_reclaim += zone->watermark_boost;
3583 zone_boosts[i] = zone->watermark_boost;
3584 }
3585 boosted = nr_boost_reclaim;
3586
3587restart:
3588 sc.priority = DEF_PRIORITY;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003589 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07003590 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003591 bool raise_priority = true;
Mel Gorman1c308442018-12-28 00:35:52 -08003592 bool balanced;
Omar Sandoval93781322018-06-07 17:07:02 -07003593 bool ret;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003594
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003595 sc.reclaim_idx = highest_zoneidx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596
Mel Gorman86c79f62016-07-28 15:45:59 -07003597 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003598 * If the number of buffer_heads exceeds the maximum allowed
3599 * then consider reclaiming from all zones. This has a dual
3600 * purpose -- on 64-bit systems it is expected that
3601 * buffer_heads are stripped during active rotation. On 32-bit
3602 * systems, highmem pages can pin lowmem memory and shrinking
3603 * buffers can relieve lowmem pressure. Reclaim may still not
3604 * go ahead if all eligible zones for the original allocation
3605 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003606 */
3607 if (buffer_heads_over_limit) {
3608 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3609 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003610 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003611 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612
Mel Gorman970a39a2016-07-28 15:46:35 -07003613 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003614 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003617
Mel Gorman86c79f62016-07-28 15:45:59 -07003618 /*
Mel Gorman1c308442018-12-28 00:35:52 -08003619 * If the pgdat is imbalanced then ignore boosting and preserve
3620 * the watermarks for a later time and restart. Note that the
3621 * zone watermarks will be still reset at the end of balancing
3622 * on the grounds that the normal reclaim should be enough to
3623 * re-evaluate if boosting is required when kswapd next wakes.
Mel Gorman86c79f62016-07-28 15:45:59 -07003624 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003625 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08003626 if (!balanced && nr_boost_reclaim) {
3627 nr_boost_reclaim = 0;
3628 goto restart;
3629 }
3630
3631 /*
3632 * If boosting is not active then only reclaim if there are no
3633 * eligible zones. Note that sc.reclaim_idx is not used as
3634 * buffer_heads_over_limit may have adjusted it.
3635 */
3636 if (!nr_boost_reclaim && balanced)
Mel Gormane716f2e2017-05-03 14:53:45 -07003637 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003638
Mel Gorman1c308442018-12-28 00:35:52 -08003639 /* Limit the priority of boosting to avoid reclaim writeback */
3640 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3641 raise_priority = false;
3642
3643 /*
3644 * Do not writeback or swap pages for boosted reclaim. The
3645 * intent is to relieve pressure not issue sub-optimal IO
3646 * from reclaim context. If no pages are reclaimed, the
3647 * reclaim will be aborted.
3648 */
3649 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3650 sc.may_swap = !nr_boost_reclaim;
Mel Gorman1c308442018-12-28 00:35:52 -08003651
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003653 * Do some background aging of the anon list, to give
3654 * pages a chance to be referenced before reclaiming. All
3655 * pages are rotated regardless of classzone as this is
3656 * about consistent aging.
3657 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003658 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003659
3660 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003661 * If we're getting trouble reclaiming, start doing writepage
3662 * even in laptop mode.
3663 */
Johannes Weiner047d72c2017-05-03 14:51:57 -07003664 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003665 sc.may_writepage = 1;
3666
Mel Gorman1d82de62016-07-28 15:45:43 -07003667 /* Call soft limit reclaim before calling shrink_node. */
3668 sc.nr_scanned = 0;
3669 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003670 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003671 sc.gfp_mask, &nr_soft_scanned);
3672 sc.nr_reclaimed += nr_soft_reclaimed;
3673
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003674 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003675 * There should be no need to raise the scanning priority if
3676 * enough pages are already being scanned that that high
3677 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003679 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003680 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003681
3682 /*
3683 * If the low watermark is met there is no need for processes
3684 * to be throttled on pfmemalloc_wait as they should not be
3685 * able to safely make forward progress. Wake them
3686 */
3687 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc73322d2017-05-03 14:51:51 -07003688 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003689 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003690
Mel Gormanb8e83b92013-07-03 15:01:45 -07003691 /* Check if kswapd should be suspending */
Omar Sandoval93781322018-06-07 17:07:02 -07003692 __fs_reclaim_release();
3693 ret = try_to_freeze();
3694 __fs_reclaim_acquire();
3695 if (ret || kthread_should_stop())
Mel Gormanb8e83b92013-07-03 15:01:45 -07003696 break;
3697
3698 /*
3699 * Raise priority if scanning rate is too low or there was no
3700 * progress in reclaiming pages
3701 */
Johannes Weinerc73322d2017-05-03 14:51:51 -07003702 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
Mel Gorman1c308442018-12-28 00:35:52 -08003703 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3704
3705 /*
3706 * If reclaim made no progress for a boost, stop reclaim as
3707 * IO cannot be queued and it could be an infinite loop in
3708 * extreme circumstances.
3709 */
3710 if (nr_boost_reclaim && !nr_reclaimed)
3711 break;
3712
Johannes Weinerc73322d2017-05-03 14:51:51 -07003713 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003714 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003715 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716
Johannes Weinerc73322d2017-05-03 14:51:51 -07003717 if (!sc.nr_reclaimed)
3718 pgdat->kswapd_failures++;
3719
Mel Gormanb8e83b92013-07-03 15:01:45 -07003720out:
Mel Gorman1c308442018-12-28 00:35:52 -08003721 /* If reclaim was boosted, account for the reclaim done in this pass */
3722 if (boosted) {
3723 unsigned long flags;
3724
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003725 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08003726 if (!zone_boosts[i])
3727 continue;
3728
3729 /* Increments are under the zone lock */
3730 zone = pgdat->node_zones + i;
3731 spin_lock_irqsave(&zone->lock, flags);
3732 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3733 spin_unlock_irqrestore(&zone->lock, flags);
3734 }
3735
3736 /*
3737 * As there is now likely space, wakeup kcompact to defragment
3738 * pageblocks.
3739 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003740 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08003741 }
3742
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003743 snapshot_refaults(NULL, pgdat);
Omar Sandoval93781322018-06-07 17:07:02 -07003744 __fs_reclaim_release();
Johannes Weinereb414682018-10-26 15:06:27 -07003745 psi_memstall_leave(&pflags);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003746 set_task_reclaim_state(current, NULL);
Yafang Shaoe5ca8072019-07-16 16:26:09 -07003747
Mel Gorman0abdee22011-01-13 15:46:22 -08003748 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003749 * Return the order kswapd stopped reclaiming at as
3750 * prepare_kswapd_sleep() takes it into account. If another caller
3751 * entered the allocator slow path while kswapd was awake, order will
3752 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003753 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003754 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755}
3756
Mel Gormane716f2e2017-05-03 14:53:45 -07003757/*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003758 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3759 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3760 * not a valid index then either kswapd runs for first time or kswapd couldn't
3761 * sleep after previous reclaim attempt (node is still unbalanced). In that
3762 * case return the zone index of the previous kswapd reclaim cycle.
Mel Gormane716f2e2017-05-03 14:53:45 -07003763 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003764static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3765 enum zone_type prev_highest_zoneidx)
Mel Gormane716f2e2017-05-03 14:53:45 -07003766{
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003767 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003768
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003769 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
Mel Gormane716f2e2017-05-03 14:53:45 -07003770}
3771
Mel Gorman38087d92016-07-28 15:45:49 -07003772static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003773 unsigned int highest_zoneidx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003774{
3775 long remaining = 0;
3776 DEFINE_WAIT(wait);
3777
3778 if (freezing(current) || kthread_should_stop())
3779 return;
3780
3781 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3782
Shantanu Goel333b0a42017-05-03 14:53:38 -07003783 /*
3784 * Try to sleep for a short interval. Note that kcompactd will only be
3785 * woken if it is possible to sleep for a short interval. This is
3786 * deliberate on the assumption that if reclaim cannot keep an
3787 * eligible zone balanced that it's also unlikely that compaction will
3788 * succeed.
3789 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003790 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003791 /*
3792 * Compaction records what page blocks it recently failed to
3793 * isolate pages from and skips them in the future scanning.
3794 * When kswapd is going to sleep, it is reasonable to assume
3795 * that pages and compaction may succeed so reset the cache.
3796 */
3797 reset_isolation_suitable(pgdat);
3798
3799 /*
3800 * We have freed the memory, now we should compact it to make
3801 * allocation of the requested order possible.
3802 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003803 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003804
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003805 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003806
3807 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003808 * If woken prematurely then reset kswapd_highest_zoneidx and
Mel Gorman38087d92016-07-28 15:45:49 -07003809 * order. The values will either be from a wakeup request or
3810 * the previous request that slept prematurely.
3811 */
3812 if (remaining) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003813 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3814 kswapd_highest_zoneidx(pgdat,
3815 highest_zoneidx));
Qian Cai5644e1fb2020-04-01 21:10:12 -07003816
3817 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3818 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003819 }
3820
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003821 finish_wait(&pgdat->kswapd_wait, &wait);
3822 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3823 }
3824
3825 /*
3826 * After a short sleep, check if it was a premature sleep. If not, then
3827 * go fully to sleep until explicitly woken up.
3828 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003829 if (!remaining &&
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003830 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003831 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3832
3833 /*
3834 * vmstat counters are not perfectly accurate and the estimated
3835 * value for counters such as NR_FREE_PAGES can deviate from the
3836 * true value by nr_online_cpus * threshold. To avoid the zone
3837 * watermarks being breached while under pressure, we reduce the
3838 * per-cpu vmstat threshold while kswapd is awake and restore
3839 * them before going back to sleep.
3840 */
3841 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003842
3843 if (!kthread_should_stop())
3844 schedule();
3845
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003846 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3847 } else {
3848 if (remaining)
3849 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3850 else
3851 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3852 }
3853 finish_wait(&pgdat->kswapd_wait, &wait);
3854}
3855
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856/*
3857 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003858 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003859 *
3860 * This basically trickles out pages so that we have _some_
3861 * free memory available even if there is no other activity
3862 * that frees anything up. This is needed for things like routing
3863 * etc, where we otherwise might have all activity going on in
3864 * asynchronous contexts that cannot page things out.
3865 *
3866 * If there are applications that are active memory-allocators
3867 * (most normal use), this basically shouldn't matter.
3868 */
3869static int kswapd(void *p)
3870{
Mel Gormane716f2e2017-05-03 14:53:45 -07003871 unsigned int alloc_order, reclaim_order;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003872 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 pg_data_t *pgdat = (pg_data_t*)p;
3874 struct task_struct *tsk = current;
Rusty Russella70f7302009-03-13 14:49:46 +10303875 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876
Rusty Russell174596a2009-01-01 10:12:29 +10303877 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003878 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879
3880 /*
3881 * Tell the memory management that we're a "memory allocator",
3882 * and that if we need more memory we should get access to it
3883 * regardless (see "__alloc_pages()"). "kswapd" should
3884 * never get caught in the normal page freeing logic.
3885 *
3886 * (Kswapd normally doesn't need memory anyway, but sometimes
3887 * you need a small amount of memory in order to be able to
3888 * page out something else, and this flag essentially protects
3889 * us from recursively trying to free more memory as we're
3890 * trying to free the first piece of memory in the first place).
3891 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003892 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003893 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894
Qian Cai5644e1fb2020-04-01 21:10:12 -07003895 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003896 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003898 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003899
Qian Cai5644e1fb2020-04-01 21:10:12 -07003900 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003901 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3902 highest_zoneidx);
Mel Gormane716f2e2017-05-03 14:53:45 -07003903
Mel Gorman38087d92016-07-28 15:45:49 -07003904kswapd_try_sleep:
3905 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003906 highest_zoneidx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003907
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003908 /* Read the new order and highest_zoneidx */
Qian Cai5644e1fb2020-04-01 21:10:12 -07003909 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003910 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3911 highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003912 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003913 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914
David Rientjes8fe23e02009-12-14 17:58:33 -08003915 ret = try_to_freeze();
3916 if (kthread_should_stop())
3917 break;
3918
3919 /*
3920 * We can speed up thawing tasks if we don't call balance_pgdat
3921 * after returning from the refrigerator
3922 */
Mel Gorman38087d92016-07-28 15:45:49 -07003923 if (ret)
3924 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003925
Mel Gorman38087d92016-07-28 15:45:49 -07003926 /*
3927 * Reclaim begins at the requested order but if a high-order
3928 * reclaim fails then kswapd falls back to reclaiming for
3929 * order-0. If that happens, kswapd will consider sleeping
3930 * for the order it finished reclaiming at (reclaim_order)
3931 * but kcompactd is woken to compact for the original
3932 * request (alloc_order).
3933 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003934 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
Mel Gormane5146b12016-07-28 15:46:47 -07003935 alloc_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003936 reclaim_order = balance_pgdat(pgdat, alloc_order,
3937 highest_zoneidx);
Mel Gorman38087d92016-07-28 15:45:49 -07003938 if (reclaim_order < alloc_order)
3939 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003941
Johannes Weiner71abdc12014-06-06 14:35:35 -07003942 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Johannes Weiner71abdc12014-06-06 14:35:35 -07003943
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 return 0;
3945}
3946
3947/*
David Rientjes5ecd9d42018-04-05 16:25:16 -07003948 * A zone is low on free memory or too fragmented for high-order memory. If
3949 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3950 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3951 * has failed or is not needed, still wake up kcompactd if only compaction is
3952 * needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 */
David Rientjes5ecd9d42018-04-05 16:25:16 -07003954void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003955 enum zone_type highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956{
3957 pg_data_t *pgdat;
Qian Cai5644e1fb2020-04-01 21:10:12 -07003958 enum zone_type curr_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959
Mel Gorman6aa303d2016-09-01 16:14:55 -07003960 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 return;
3962
David Rientjes5ecd9d42018-04-05 16:25:16 -07003963 if (!cpuset_zone_allowed(zone, gfp_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964 return;
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003965
Qian Cai5644e1fb2020-04-01 21:10:12 -07003966 pgdat = zone->zone_pgdat;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003967 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003968
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003969 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3970 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003971
3972 if (READ_ONCE(pgdat->kswapd_order) < order)
3973 WRITE_ONCE(pgdat->kswapd_order, order);
3974
Con Kolivas8d0986e2005-09-13 01:25:07 -07003975 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003977
David Rientjes5ecd9d42018-04-05 16:25:16 -07003978 /* Hopeless node, leave it to direct reclaim if possible */
3979 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003980 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3981 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
David Rientjes5ecd9d42018-04-05 16:25:16 -07003982 /*
3983 * There may be plenty of free memory available, but it's too
3984 * fragmented for high-order allocations. Wake up kcompactd
3985 * and rely on compaction_suitable() to determine if it's
3986 * needed. If it fails, it will defer subsequent attempts to
3987 * ratelimit its work.
3988 */
3989 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003990 wakeup_kcompactd(pgdat, order, highest_zoneidx);
Johannes Weinerc73322d2017-05-03 14:51:51 -07003991 return;
David Rientjes5ecd9d42018-04-05 16:25:16 -07003992 }
Johannes Weinerc73322d2017-05-03 14:51:51 -07003993
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003994 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
David Rientjes5ecd9d42018-04-05 16:25:16 -07003995 gfp_flags);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003996 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997}
3998
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003999#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004001 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004002 * freed pages.
4003 *
4004 * Rather than trying to age LRUs the aim is to preserve the overall
4005 * LRU order by reclaiming preferentially
4006 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004008unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004010 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004011 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07004012 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07004013 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004014 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07004015 .may_writepage = 1,
4016 .may_unmap = 1,
4017 .may_swap = 1,
4018 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 };
Ying Hana09ed5e2011-05-24 17:12:26 -07004020 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004021 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004022 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023
Peter Zijlstrad92a8cf2017-03-03 10:13:38 +01004024 fs_reclaim_acquire(sc.gfp_mask);
Omar Sandoval93781322018-06-07 17:07:02 -07004025 noreclaim_flag = memalloc_noreclaim_save();
Andrew Morton1732d2b012019-07-16 16:26:15 -07004026 set_task_reclaim_state(current, &sc.reclaim_state);
Andrew Morton69e05942006-03-22 00:08:19 -08004027
Vladimir Davydov3115cd92014-04-03 14:47:22 -07004028 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004029
Andrew Morton1732d2b012019-07-16 16:26:15 -07004030 set_task_reclaim_state(current, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07004031 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004032 fs_reclaim_release(sc.gfp_mask);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004033
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004034 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02004036#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037
Yasunori Goto3218ae12006-06-27 02:53:33 -07004038/*
4039 * This kswapd start function will be called by init and node-hot-add.
4040 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4041 */
4042int kswapd_run(int nid)
4043{
4044 pg_data_t *pgdat = NODE_DATA(nid);
4045 int ret = 0;
4046
4047 if (pgdat->kswapd)
4048 return 0;
4049
4050 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4051 if (IS_ERR(pgdat->kswapd)) {
4052 /* failure at boot is fatal */
Thomas Gleixnerc6202ad2017-05-16 20:42:46 +02004053 BUG_ON(system_state < SYSTEM_RUNNING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07004054 pr_err("Failed to start kswapd on node %d\n", nid);
4055 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07004056 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07004057 }
4058 return ret;
4059}
4060
David Rientjes8fe23e02009-12-14 17:58:33 -08004061/*
Jiang Liud8adde12012-07-11 14:01:52 -07004062 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004063 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08004064 */
4065void kswapd_stop(int nid)
4066{
4067 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4068
Jiang Liud8adde12012-07-11 14:01:52 -07004069 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08004070 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07004071 NODE_DATA(nid)->kswapd = NULL;
4072 }
David Rientjes8fe23e02009-12-14 17:58:33 -08004073}
4074
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075static int __init kswapd_init(void)
4076{
Wei Yang6b700b52020-04-01 21:10:09 -07004077 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08004078
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08004080 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07004081 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 return 0;
4083}
4084
4085module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004086
4087#ifdef CONFIG_NUMA
4088/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004089 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08004090 *
Mel Gormana5f5f912016-07-28 15:46:32 -07004091 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08004092 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004093 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004094int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004095
Dave Hansen54683f82021-02-24 12:09:15 -08004096/*
4097 * These bit locations are exposed in the vm.zone_reclaim_mode sysctl
4098 * ABI. New bits are OK, but existing bits can never change.
4099 */
4100#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4101#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4102#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08004103
Christoph Lameter9eeff232006-01-18 17:42:31 -08004104/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004105 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08004106 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4107 * a zone.
4108 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004109#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08004110
Christoph Lameter9eeff232006-01-18 17:42:31 -08004111/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004112 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07004113 * occur.
4114 */
4115int sysctl_min_unmapped_ratio = 1;
4116
4117/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07004118 * If the number of slab pages in a zone grows beyond this percentage then
4119 * slab reclaim needs to occur.
4120 */
4121int sysctl_min_slab_ratio = 5;
4122
Mel Gorman11fb9982016-07-28 15:46:20 -07004123static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07004124{
Mel Gorman11fb9982016-07-28 15:46:20 -07004125 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4126 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4127 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004128
4129 /*
4130 * It's possible for there to be more file mapped pages than
4131 * accounted for by the pages on the file LRU lists because
4132 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4133 */
4134 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4135}
4136
4137/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07004138static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07004139{
Alexandru Moised031a152015-11-05 18:48:08 -08004140 unsigned long nr_pagecache_reclaimable;
4141 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07004142
4143 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004144 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07004145 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07004146 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07004147 * a better estimate
4148 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004149 if (node_reclaim_mode & RECLAIM_UNMAP)
4150 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004151 else
Mel Gormana5f5f912016-07-28 15:46:32 -07004152 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004153
4154 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07004155 if (!(node_reclaim_mode & RECLAIM_WRITE))
4156 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004157
4158 /* Watch for any possible underflows due to delta */
4159 if (unlikely(delta > nr_pagecache_reclaimable))
4160 delta = nr_pagecache_reclaimable;
4161
4162 return nr_pagecache_reclaimable - delta;
4163}
4164
Christoph Lameter0ff38492006-09-25 23:31:52 -07004165/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004166 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004167 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004168static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004169{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08004170 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08004171 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004172 struct task_struct *p = current;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004173 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08004174 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08004175 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004176 .gfp_mask = current_gfp_context(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07004177 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07004178 .priority = NODE_RECLAIM_PRIORITY,
4179 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4180 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07004181 .may_swap = 1,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004182 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08004183 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08004184
Yafang Shao132bb8c2019-05-13 17:17:53 -07004185 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4186 sc.gfp_mask);
4187
Christoph Lameter9eeff232006-01-18 17:42:31 -08004188 cond_resched();
Omar Sandoval93781322018-06-07 17:07:02 -07004189 fs_reclaim_acquire(sc.gfp_mask);
Christoph Lameterd4f77962006-02-24 13:04:22 -08004190 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004191 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08004192 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004193 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08004194 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07004195 noreclaim_flag = memalloc_noreclaim_save();
4196 p->flags |= PF_SWAPWRITE;
Andrew Morton1732d2b012019-07-16 16:26:15 -07004197 set_task_reclaim_state(p, &sc.reclaim_state);
Christoph Lameterc84db23c2006-02-01 03:05:29 -08004198
Mel Gormana5f5f912016-07-28 15:46:32 -07004199 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07004200 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07004201 * Free memory by calling shrink node with increasing
Christoph Lameter0ff38492006-09-25 23:31:52 -07004202 * priorities until we have enough memory freed.
4203 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07004204 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07004205 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004206 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07004207 }
Christoph Lameterc84db23c2006-02-01 03:05:29 -08004208
Andrew Morton1732d2b012019-07-16 16:26:15 -07004209 set_task_reclaim_state(p, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07004210 current->flags &= ~PF_SWAPWRITE;
4211 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004212 fs_reclaim_release(sc.gfp_mask);
Yafang Shao132bb8c2019-05-13 17:17:53 -07004213
4214 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4215
Rik van Riela79311c2009-01-06 14:40:01 -08004216 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004217}
Andrew Morton179e9632006-03-22 00:08:18 -08004218
Mel Gormana5f5f912016-07-28 15:46:32 -07004219int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08004220{
David Rientjesd773ed62007-10-16 23:26:01 -07004221 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004222
4223 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004224 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07004225 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07004226 *
Christoph Lameter96146342006-07-03 00:24:13 -07004227 * A small portion of unmapped file backed pages is needed for
4228 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07004229 * thrown out if the node is overallocated. So we do not reclaim
4230 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07004231 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08004232 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004233 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
Roman Gushchind42f3242020-08-06 23:20:39 -07004234 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4235 pgdat->min_slab_pages)
Mel Gormana5f5f912016-07-28 15:46:32 -07004236 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08004237
4238 /*
David Rientjesd773ed62007-10-16 23:26:01 -07004239 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08004240 */
Mel Gormand0164ad2015-11-06 16:28:21 -08004241 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07004242 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08004243
4244 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004245 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08004246 * have associated processors. This will favor the local processor
4247 * over remote processors and spread off node memory allocations
4248 * as wide as possible.
4249 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004250 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4251 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07004252
Mel Gormana5f5f912016-07-28 15:46:32 -07004253 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4254 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07004255
Mel Gormana5f5f912016-07-28 15:46:32 -07004256 ret = __node_reclaim(pgdat, gfp_mask, order);
4257 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07004258
Mel Gorman24cf725182009-06-16 15:33:23 -07004259 if (!ret)
4260 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4261
David Rientjesd773ed62007-10-16 23:26:01 -07004262 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004263}
Christoph Lameter9eeff232006-01-18 17:42:31 -08004264#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004265
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004266/**
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004267 * check_move_unevictable_pages - check pages for evictability and move to
4268 * appropriate zone lru list
4269 * @pvec: pagevec with lru pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004270 *
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004271 * Checks pages for evictability, if an evictable page is in the unevictable
4272 * lru list, moves it to the appropriate evictable lru list. This function
4273 * should be only used for lru pages.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004274 */
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004275void check_move_unevictable_pages(struct pagevec *pvec)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004276{
Johannes Weiner925b7672012-01-12 17:18:15 -08004277 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07004278 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08004279 int pgscanned = 0;
4280 int pgrescued = 0;
4281 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004282
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004283 for (i = 0; i < pvec->nr; i++) {
4284 struct page *page = pvec->pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07004285 struct pglist_data *pagepgdat = page_pgdat(page);
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004286 int nr_pages;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07004287
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004288 if (PageTransTail(page))
4289 continue;
4290
4291 nr_pages = thp_nr_pages(page);
4292 pgscanned += nr_pages;
4293
Mel Gorman785b99f2016-07-28 15:47:23 -07004294 if (pagepgdat != pgdat) {
4295 if (pgdat)
4296 spin_unlock_irq(&pgdat->lru_lock);
4297 pgdat = pagepgdat;
4298 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004299 }
Mel Gorman785b99f2016-07-28 15:47:23 -07004300 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004301
Hugh Dickins24513262012-01-20 14:34:21 -08004302 if (!PageLRU(page) || !PageUnevictable(page))
4303 continue;
4304
Hugh Dickins39b5f292012-10-08 16:33:18 -07004305 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08004306 enum lru_list lru = page_lru_base_type(page);
4307
Sasha Levin309381fea2014-01-23 15:52:54 -08004308 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08004309 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07004310 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4311 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004312 pgrescued += nr_pages;
Hugh Dickins24513262012-01-20 14:34:21 -08004313 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004314 }
Hugh Dickins24513262012-01-20 14:34:21 -08004315
Mel Gorman785b99f2016-07-28 15:47:23 -07004316 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08004317 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4318 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07004319 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08004320 }
Hugh Dickins850465792012-01-20 14:34:19 -08004321}
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004322EXPORT_SYMBOL_GPL(check_move_unevictable_pages);