blob: 61093dde115368d419dc088f5051b96c865ab236 [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/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4 *
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
11 */
12
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070013#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#include <linux/mm.h>
Ingo Molnar5b3cc152017-02-02 20:43:54 +010016#include <linux/sched/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Dave Hansen26aa2d12021-09-02 14:59:16 -070044#include <linux/migrate.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070045#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070046#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070047#include <linux/oom.h>
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +000048#include <linux/pagevec.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070049#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070050#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080051#include <linux/dax.h>
Johannes Weinereb414682018-10-26 15:06:27 -070052#include <linux/psi.h>
Yu Zhao5280d762022-01-27 19:23:01 -070053#include <linux/pagewalk.h>
54#include <linux/shmem_fs.h>
Yu Zhao76f7f072022-01-27 19:52:09 -070055#include <linux/ctype.h>
Yu Zhao3d18c9e2022-01-27 20:12:41 -070056#include <linux/debugfs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
58#include <asm/tlbflush.h>
59#include <asm/div64.h>
60
61#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070062#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Nick Piggin0f8053a2006-03-22 00:08:33 -080064#include "internal.h"
65
Mel Gorman33906bc2010-08-09 17:19:16 -070066#define CREATE_TRACE_POINTS
67#include <trace/events/vmscan.h>
68
Sudarshan Rajagopalan55b46b62021-06-01 04:05:01 -070069#undef CREATE_TRACE_POINTS
70#include <trace/hooks/vmscan.h>
71
Linus Torvalds1da177e2005-04-16 15:20:36 -070072struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080073 /* How many pages shrink_list() should reclaim */
74 unsigned long nr_to_reclaim;
75
Johannes Weineree814fe2014-08-06 16:06:19 -070076 /*
77 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * are scanned.
79 */
80 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070081
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070082 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080083 * The memory cgroup that hit its limit and as a result is the
84 * primary target of this reclaim invocation.
85 */
86 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080087
Johannes Weiner7cf111b2020-06-03 16:03:06 -070088 /*
89 * Scan pressure balancing between anon and file LRUs
90 */
91 unsigned long anon_cost;
92 unsigned long file_cost;
93
Johannes Weinerb91ac372019-11-30 17:56:02 -080094 /* Can active pages be deactivated as part of reclaim? */
95#define DEACTIVATE_ANON 1
96#define DEACTIVATE_FILE 2
97 unsigned int may_deactivate:2;
98 unsigned int force_deactivate:1;
99 unsigned int skipped_deactivate:1;
100
Johannes Weiner1276ad62017-02-24 14:56:11 -0800101 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -0700102 unsigned int may_writepage:1;
103
104 /* Can mapped pages be reclaimed? */
105 unsigned int may_unmap:1;
106
107 /* Can pages be swapped as part of reclaim? */
108 unsigned int may_swap:1;
109
Yisheng Xied6622f62017-05-03 14:53:57 -0700110 /*
Johannes Weinerf56ce412021-08-19 19:04:21 -0700111 * Cgroup memory below memory.low is protected as long as we
112 * don't threaten to OOM. If any cgroup is reclaimed at
113 * reduced force or passed over entirely due to its memory.low
114 * setting (memcg_low_skipped), and nothing is reclaimed as a
115 * result, then go back for one more cycle that reclaims the protected
116 * memory (memcg_low_reclaim) to avert OOM.
Yisheng Xied6622f62017-05-03 14:53:57 -0700117 */
118 unsigned int memcg_low_reclaim:1;
119 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -0800120
Johannes Weineree814fe2014-08-06 16:06:19 -0700121 unsigned int hibernation_mode:1;
122
123 /* One of the zones is ready for compaction */
124 unsigned int compaction_ready:1;
125
Johannes Weinerb91ac372019-11-30 17:56:02 -0800126 /* There is easily reclaimable cold cache in the current node */
127 unsigned int cache_trim_mode:1;
128
Johannes Weiner53138ce2019-11-30 17:55:56 -0800129 /* The file pages on the current node are dangerously low */
130 unsigned int file_is_tiny:1;
131
Dave Hansen26aa2d12021-09-02 14:59:16 -0700132 /* Always discard instead of demoting to lower tier memory */
133 unsigned int no_demotion:1;
134
Yu Zhao4d905e92022-01-27 19:59:54 -0700135#ifdef CONFIG_LRU_GEN
136 /* help make better choices when multiple memcgs are available */
137 unsigned int memcgs_need_aging:1;
138 unsigned int memcgs_need_swapping:1;
139 unsigned int memcgs_avoid_swapping:1;
140#endif
141
Greg Thelenbb451fd2018-08-17 15:45:19 -0700142 /* Allocation order */
143 s8 order;
144
145 /* Scan (total_size >> priority) pages at once */
146 s8 priority;
147
148 /* The highest zone to isolate pages for reclaim from */
149 s8 reclaim_idx;
150
151 /* This context's GFP mask */
152 gfp_t gfp_mask;
153
Johannes Weineree814fe2014-08-06 16:06:19 -0700154 /* Incremented by the number of inactive pages that were scanned */
155 unsigned long nr_scanned;
156
157 /* Number of pages freed so far during a call to shrink_zones() */
158 unsigned long nr_reclaimed;
Andrey Ryabinind108c772018-04-10 16:27:59 -0700159
160 struct {
161 unsigned int dirty;
162 unsigned int unqueued_dirty;
163 unsigned int congested;
164 unsigned int writeback;
165 unsigned int immediate;
166 unsigned int file_taken;
167 unsigned int taken;
168 } nr;
Yafang Shaoe5ca8072019-07-16 16:26:09 -0700169
170 /* for recording the reclaimed slab by now */
171 struct reclaim_state reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172};
173
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174#ifdef ARCH_HAS_PREFETCHW
175#define prefetchw_prev_lru_page(_page, _base, _field) \
176 do { \
177 if ((_page)->lru.prev != _base) { \
178 struct page *prev; \
179 \
180 prev = lru_to_page(&(_page->lru)); \
181 prefetchw(&prev->_field); \
182 } \
183 } while (0)
184#else
185#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
186#endif
187
188/*
Johannes Weinerc8439662020-06-03 16:02:37 -0700189 * From 0 .. 200. Higher means more swappy.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 */
191int vm_swappiness = 60;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
Yang Shi0a432dc2019-09-23 15:38:12 -0700193static void set_task_reclaim_state(struct task_struct *task,
194 struct reclaim_state *rs)
195{
196 /* Check for an overwrite */
197 WARN_ON_ONCE(rs && task->reclaim_state);
198
199 /* Check for the nulling of an already-nulled member */
200 WARN_ON_ONCE(!rs && !task->reclaim_state);
201
202 task->reclaim_state = rs;
203}
204
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205static LIST_HEAD(shrinker_list);
206static DECLARE_RWSEM(shrinker_rwsem);
207
Yang Shi0a432dc2019-09-23 15:38:12 -0700208#ifdef CONFIG_MEMCG
Yang Shia2fb1262021-05-04 18:36:17 -0700209static int shrinker_nr_max;
Yang Shi2bfd3632021-05-04 18:36:11 -0700210
Yang Shi3c6f17e2021-05-04 18:36:33 -0700211/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
Yang Shia2fb1262021-05-04 18:36:17 -0700212static inline int shrinker_map_size(int nr_items)
213{
214 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
215}
Yang Shi2bfd3632021-05-04 18:36:11 -0700216
Yang Shi3c6f17e2021-05-04 18:36:33 -0700217static inline int shrinker_defer_size(int nr_items)
218{
219 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
220}
221
Yang Shi468ab842021-05-04 18:36:26 -0700222static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
223 int nid)
224{
225 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
226 lockdep_is_held(&shrinker_rwsem));
227}
228
Yang Shie4262c42021-05-04 18:36:23 -0700229static int expand_one_shrinker_info(struct mem_cgroup *memcg,
Yang Shi3c6f17e2021-05-04 18:36:33 -0700230 int map_size, int defer_size,
231 int old_map_size, int old_defer_size)
Yang Shi2bfd3632021-05-04 18:36:11 -0700232{
Yang Shie4262c42021-05-04 18:36:23 -0700233 struct shrinker_info *new, *old;
Yang Shi2bfd3632021-05-04 18:36:11 -0700234 struct mem_cgroup_per_node *pn;
235 int nid;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700236 int size = map_size + defer_size;
Yang Shi2bfd3632021-05-04 18:36:11 -0700237
Yang Shi2bfd3632021-05-04 18:36:11 -0700238 for_each_node(nid) {
239 pn = memcg->nodeinfo[nid];
Yang Shi468ab842021-05-04 18:36:26 -0700240 old = shrinker_info_protected(memcg, nid);
Yang Shi2bfd3632021-05-04 18:36:11 -0700241 /* Not yet online memcg */
242 if (!old)
243 return 0;
244
245 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
246 if (!new)
247 return -ENOMEM;
248
Yang Shi3c6f17e2021-05-04 18:36:33 -0700249 new->nr_deferred = (atomic_long_t *)(new + 1);
250 new->map = (void *)new->nr_deferred + defer_size;
251
252 /* map: set all old bits, clear all new bits */
253 memset(new->map, (int)0xff, old_map_size);
254 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
255 /* nr_deferred: copy old values, clear all new values */
256 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
257 memset((void *)new->nr_deferred + old_defer_size, 0,
258 defer_size - old_defer_size);
Yang Shi2bfd3632021-05-04 18:36:11 -0700259
Yang Shie4262c42021-05-04 18:36:23 -0700260 rcu_assign_pointer(pn->shrinker_info, new);
Yang Shi72673e82021-05-04 18:36:20 -0700261 kvfree_rcu(old, rcu);
Yang Shi2bfd3632021-05-04 18:36:11 -0700262 }
263
264 return 0;
265}
266
Yang Shie4262c42021-05-04 18:36:23 -0700267void free_shrinker_info(struct mem_cgroup *memcg)
Yang Shi2bfd3632021-05-04 18:36:11 -0700268{
269 struct mem_cgroup_per_node *pn;
Yang Shie4262c42021-05-04 18:36:23 -0700270 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700271 int nid;
272
Yang Shi2bfd3632021-05-04 18:36:11 -0700273 for_each_node(nid) {
274 pn = memcg->nodeinfo[nid];
Yang Shie4262c42021-05-04 18:36:23 -0700275 info = rcu_dereference_protected(pn->shrinker_info, true);
276 kvfree(info);
277 rcu_assign_pointer(pn->shrinker_info, NULL);
Yang Shi2bfd3632021-05-04 18:36:11 -0700278 }
279}
280
Yang Shie4262c42021-05-04 18:36:23 -0700281int alloc_shrinker_info(struct mem_cgroup *memcg)
Yang Shi2bfd3632021-05-04 18:36:11 -0700282{
Yang Shie4262c42021-05-04 18:36:23 -0700283 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700284 int nid, size, ret = 0;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700285 int map_size, defer_size = 0;
Yang Shi2bfd3632021-05-04 18:36:11 -0700286
Yang Shid27cf2a2021-05-04 18:36:14 -0700287 down_write(&shrinker_rwsem);
Yang Shi3c6f17e2021-05-04 18:36:33 -0700288 map_size = shrinker_map_size(shrinker_nr_max);
289 defer_size = shrinker_defer_size(shrinker_nr_max);
290 size = map_size + defer_size;
Yang Shi2bfd3632021-05-04 18:36:11 -0700291 for_each_node(nid) {
Yang Shie4262c42021-05-04 18:36:23 -0700292 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
293 if (!info) {
294 free_shrinker_info(memcg);
Yang Shi2bfd3632021-05-04 18:36:11 -0700295 ret = -ENOMEM;
296 break;
297 }
Yang Shi3c6f17e2021-05-04 18:36:33 -0700298 info->nr_deferred = (atomic_long_t *)(info + 1);
299 info->map = (void *)info->nr_deferred + defer_size;
Yang Shie4262c42021-05-04 18:36:23 -0700300 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
Yang Shi2bfd3632021-05-04 18:36:11 -0700301 }
Yang Shid27cf2a2021-05-04 18:36:14 -0700302 up_write(&shrinker_rwsem);
Yang Shi2bfd3632021-05-04 18:36:11 -0700303
304 return ret;
305}
306
Yang Shi3c6f17e2021-05-04 18:36:33 -0700307static inline bool need_expand(int nr_max)
308{
309 return round_up(nr_max, BITS_PER_LONG) >
310 round_up(shrinker_nr_max, BITS_PER_LONG);
311}
312
Yang Shie4262c42021-05-04 18:36:23 -0700313static int expand_shrinker_info(int new_id)
Yang Shi2bfd3632021-05-04 18:36:11 -0700314{
Yang Shi3c6f17e2021-05-04 18:36:33 -0700315 int ret = 0;
Yang Shia2fb1262021-05-04 18:36:17 -0700316 int new_nr_max = new_id + 1;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700317 int map_size, defer_size = 0;
318 int old_map_size, old_defer_size = 0;
Yang Shi2bfd3632021-05-04 18:36:11 -0700319 struct mem_cgroup *memcg;
320
Yang Shi3c6f17e2021-05-04 18:36:33 -0700321 if (!need_expand(new_nr_max))
Yang Shia2fb1262021-05-04 18:36:17 -0700322 goto out;
Yang Shi2bfd3632021-05-04 18:36:11 -0700323
Yang Shi2bfd3632021-05-04 18:36:11 -0700324 if (!root_mem_cgroup)
Yang Shid27cf2a2021-05-04 18:36:14 -0700325 goto out;
326
327 lockdep_assert_held(&shrinker_rwsem);
Yang Shi2bfd3632021-05-04 18:36:11 -0700328
Yang Shi3c6f17e2021-05-04 18:36:33 -0700329 map_size = shrinker_map_size(new_nr_max);
330 defer_size = shrinker_defer_size(new_nr_max);
331 old_map_size = shrinker_map_size(shrinker_nr_max);
332 old_defer_size = shrinker_defer_size(shrinker_nr_max);
333
Yang Shi2bfd3632021-05-04 18:36:11 -0700334 memcg = mem_cgroup_iter(NULL, NULL, NULL);
335 do {
Yang Shi3c6f17e2021-05-04 18:36:33 -0700336 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
337 old_map_size, old_defer_size);
Yang Shi2bfd3632021-05-04 18:36:11 -0700338 if (ret) {
339 mem_cgroup_iter_break(NULL, memcg);
Yang Shid27cf2a2021-05-04 18:36:14 -0700340 goto out;
Yang Shi2bfd3632021-05-04 18:36:11 -0700341 }
342 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
Yang Shid27cf2a2021-05-04 18:36:14 -0700343out:
Yang Shi2bfd3632021-05-04 18:36:11 -0700344 if (!ret)
Yang Shia2fb1262021-05-04 18:36:17 -0700345 shrinker_nr_max = new_nr_max;
Yang Shid27cf2a2021-05-04 18:36:14 -0700346
Yang Shi2bfd3632021-05-04 18:36:11 -0700347 return ret;
348}
349
350void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
351{
352 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
Yang Shie4262c42021-05-04 18:36:23 -0700353 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700354
355 rcu_read_lock();
Yang Shie4262c42021-05-04 18:36:23 -0700356 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
Yang Shi2bfd3632021-05-04 18:36:11 -0700357 /* Pairs with smp mb in shrink_slab() */
358 smp_mb__before_atomic();
Yang Shie4262c42021-05-04 18:36:23 -0700359 set_bit(shrinker_id, info->map);
Yang Shi2bfd3632021-05-04 18:36:11 -0700360 rcu_read_unlock();
361 }
362}
363
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700364static DEFINE_IDR(shrinker_idr);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700365
366static int prealloc_memcg_shrinker(struct shrinker *shrinker)
367{
368 int id, ret = -ENOMEM;
369
Yang Shi476b30a2021-05-04 18:36:39 -0700370 if (mem_cgroup_disabled())
371 return -ENOSYS;
372
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700373 down_write(&shrinker_rwsem);
374 /* This may call shrinker, so it must use down_read_trylock() */
Yang Shi41ca6682021-05-04 18:36:29 -0700375 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700376 if (id < 0)
377 goto unlock;
378
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700379 if (id >= shrinker_nr_max) {
Yang Shie4262c42021-05-04 18:36:23 -0700380 if (expand_shrinker_info(id)) {
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700381 idr_remove(&shrinker_idr, id);
382 goto unlock;
383 }
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700384 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700385 shrinker->id = id;
386 ret = 0;
387unlock:
388 up_write(&shrinker_rwsem);
389 return ret;
390}
391
392static void unregister_memcg_shrinker(struct shrinker *shrinker)
393{
394 int id = shrinker->id;
395
396 BUG_ON(id < 0);
397
Yang Shi41ca6682021-05-04 18:36:29 -0700398 lockdep_assert_held(&shrinker_rwsem);
399
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700400 idr_remove(&shrinker_idr, id);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700401}
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700402
Yang Shi86750832021-05-04 18:36:36 -0700403static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
404 struct mem_cgroup *memcg)
405{
406 struct shrinker_info *info;
407
408 info = shrinker_info_protected(memcg, nid);
409 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
410}
411
412static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
413 struct mem_cgroup *memcg)
414{
415 struct shrinker_info *info;
416
417 info = shrinker_info_protected(memcg, nid);
418 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
419}
420
Yang Shia1780152021-05-04 18:36:42 -0700421void reparent_shrinker_deferred(struct mem_cgroup *memcg)
422{
423 int i, nid;
424 long nr;
425 struct mem_cgroup *parent;
426 struct shrinker_info *child_info, *parent_info;
427
428 parent = parent_mem_cgroup(memcg);
429 if (!parent)
430 parent = root_mem_cgroup;
431
432 /* Prevent from concurrent shrinker_info expand */
433 down_read(&shrinker_rwsem);
434 for_each_node(nid) {
435 child_info = shrinker_info_protected(memcg, nid);
436 parent_info = shrinker_info_protected(parent, nid);
437 for (i = 0; i < shrinker_nr_max; i++) {
438 nr = atomic_long_read(&child_info->nr_deferred[i]);
439 atomic_long_add(nr, &parent_info->nr_deferred[i]);
440 }
441 }
442 up_read(&shrinker_rwsem);
443}
444
Johannes Weinerb5ead352019-11-30 17:55:40 -0800445static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800446{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800447 return sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800448}
Tejun Heo97c93412015-05-22 18:23:36 -0400449
450/**
Johannes Weinerb5ead352019-11-30 17:55:40 -0800451 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
Tejun Heo97c93412015-05-22 18:23:36 -0400452 * @sc: scan_control in question
453 *
454 * The normal page dirty throttling mechanism in balance_dirty_pages() is
455 * completely broken with the legacy memcg and direct stalling in
456 * shrink_page_list() is used for throttling instead, which lacks all the
457 * niceties such as fairness, adaptive pausing, bandwidth proportional
458 * allocation and configurability.
459 *
460 * This function tests whether the vmscan currently in progress can assume
461 * that the normal dirty throttling mechanism is operational.
462 */
Johannes Weinerb5ead352019-11-30 17:55:40 -0800463static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c93412015-05-22 18:23:36 -0400464{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800465 if (!cgroup_reclaim(sc))
Tejun Heo97c93412015-05-22 18:23:36 -0400466 return true;
467#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800468 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400469 return true;
470#endif
471 return false;
472}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800473#else
Yang Shi0a432dc2019-09-23 15:38:12 -0700474static int prealloc_memcg_shrinker(struct shrinker *shrinker)
475{
Yang Shi476b30a2021-05-04 18:36:39 -0700476 return -ENOSYS;
Yang Shi0a432dc2019-09-23 15:38:12 -0700477}
478
479static void unregister_memcg_shrinker(struct shrinker *shrinker)
480{
481}
482
Yang Shi86750832021-05-04 18:36:36 -0700483static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
484 struct mem_cgroup *memcg)
485{
486 return 0;
487}
488
489static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
490 struct mem_cgroup *memcg)
491{
492 return 0;
493}
494
Johannes Weinerb5ead352019-11-30 17:55:40 -0800495static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800496{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800497 return false;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800498}
Tejun Heo97c93412015-05-22 18:23:36 -0400499
Johannes Weinerb5ead352019-11-30 17:55:40 -0800500static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c93412015-05-22 18:23:36 -0400501{
502 return true;
503}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800504#endif
505
Yang Shi86750832021-05-04 18:36:36 -0700506static long xchg_nr_deferred(struct shrinker *shrinker,
507 struct shrink_control *sc)
508{
509 int nid = sc->nid;
510
511 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
512 nid = 0;
513
514 if (sc->memcg &&
515 (shrinker->flags & SHRINKER_MEMCG_AWARE))
516 return xchg_nr_deferred_memcg(nid, shrinker,
517 sc->memcg);
518
519 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
520}
521
522
523static long add_nr_deferred(long nr, struct shrinker *shrinker,
524 struct shrink_control *sc)
525{
526 int nid = sc->nid;
527
528 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
529 nid = 0;
530
531 if (sc->memcg &&
532 (shrinker->flags & SHRINKER_MEMCG_AWARE))
533 return add_nr_deferred_memcg(nr, nid, shrinker,
534 sc->memcg);
535
536 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
537}
538
Dave Hansen26aa2d12021-09-02 14:59:16 -0700539static bool can_demote(int nid, struct scan_control *sc)
540{
Huang Ying20b51af12021-09-02 14:59:33 -0700541 if (!numa_demotion_enabled)
542 return false;
Dave Hansen3a235692021-09-02 14:59:30 -0700543 if (sc) {
544 if (sc->no_demotion)
545 return false;
546 /* It is pointless to do demotion in memcg reclaim */
547 if (cgroup_reclaim(sc))
548 return false;
549 }
Dave Hansen26aa2d12021-09-02 14:59:16 -0700550 if (next_demotion_node(nid) == NUMA_NO_NODE)
551 return false;
552
Huang Ying20b51af12021-09-02 14:59:33 -0700553 return true;
Dave Hansen26aa2d12021-09-02 14:59:16 -0700554}
555
Keith Buscha2a36482021-09-02 14:59:26 -0700556static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
557 int nid,
558 struct scan_control *sc)
559{
560 if (memcg == NULL) {
561 /*
562 * For non-memcg reclaim, is there
563 * space in any swap device?
564 */
565 if (get_nr_swap_pages() > 0)
566 return true;
567 } else {
568 /* Is the memcg below its swap limit? */
569 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
570 return true;
571 }
572
573 /*
574 * The page can not be swapped.
575 *
576 * Can it be reclaimed from this node via demotion?
577 */
578 return can_demote(nid, sc);
579}
580
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700581/*
582 * This misses isolated pages which are not accounted for to save counters.
583 * As the data only determines if reclaim or compaction continues, it is
584 * not expected that isolated pages will be a dominating factor.
585 */
586unsigned long zone_reclaimable_pages(struct zone *zone)
587{
588 unsigned long nr;
589
590 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
591 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
Keith Buscha2a36482021-09-02 14:59:26 -0700592 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700593 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
594 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
595
596 return nr;
597}
598
Michal Hockofd538802017-02-22 15:45:58 -0800599/**
600 * lruvec_lru_size - Returns the number of pages on the given LRU list.
601 * @lruvec: lru vector
602 * @lru: lru to use
603 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
604 */
Yu Zhao20913392021-02-24 12:08:44 -0800605static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
606 int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800607{
Johannes Weinerde3b0152019-11-30 17:55:31 -0800608 unsigned long size = 0;
Michal Hockofd538802017-02-22 15:45:58 -0800609 int zid;
610
Johannes Weinerde3b0152019-11-30 17:55:31 -0800611 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
Michal Hockofd538802017-02-22 15:45:58 -0800612 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800613
Michal Hockofd538802017-02-22 15:45:58 -0800614 if (!managed_zone(zone))
615 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800616
Michal Hockofd538802017-02-22 15:45:58 -0800617 if (!mem_cgroup_disabled())
Johannes Weinerde3b0152019-11-30 17:55:31 -0800618 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
Michal Hockofd538802017-02-22 15:45:58 -0800619 else
Johannes Weinerde3b0152019-11-30 17:55:31 -0800620 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
Michal Hockofd538802017-02-22 15:45:58 -0800621 }
Johannes Weinerde3b0152019-11-30 17:55:31 -0800622 return size;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800623}
624
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000626 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 */
Tetsuo Handa8e049442018-04-04 19:53:07 +0900628int prealloc_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629{
Yang Shi476b30a2021-05-04 18:36:39 -0700630 unsigned int size;
631 int err;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000632
Yang Shi476b30a2021-05-04 18:36:39 -0700633 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
634 err = prealloc_memcg_shrinker(shrinker);
635 if (err != -ENOSYS)
636 return err;
637
638 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
639 }
640
641 size = sizeof(*shrinker->nr_deferred);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000642 if (shrinker->flags & SHRINKER_NUMA_AWARE)
643 size *= nr_node_ids;
644
645 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
646 if (!shrinker->nr_deferred)
647 return -ENOMEM;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700648
Tetsuo Handa8e049442018-04-04 19:53:07 +0900649 return 0;
650}
Glauber Costa1d3d4432013-08-28 10:18:04 +1000651
Tetsuo Handa8e049442018-04-04 19:53:07 +0900652void free_prealloced_shrinker(struct shrinker *shrinker)
653{
Yang Shi41ca6682021-05-04 18:36:29 -0700654 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
655 down_write(&shrinker_rwsem);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700656 unregister_memcg_shrinker(shrinker);
Yang Shi41ca6682021-05-04 18:36:29 -0700657 up_write(&shrinker_rwsem);
Yang Shi476b30a2021-05-04 18:36:39 -0700658 return;
Yang Shi41ca6682021-05-04 18:36:29 -0700659 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700660
Tetsuo Handa8e049442018-04-04 19:53:07 +0900661 kfree(shrinker->nr_deferred);
662 shrinker->nr_deferred = NULL;
663}
664
665void register_shrinker_prepared(struct shrinker *shrinker)
666{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700667 down_write(&shrinker_rwsem);
668 list_add_tail(&shrinker->list, &shrinker_list);
Yang Shi41ca6682021-05-04 18:36:29 -0700669 shrinker->flags |= SHRINKER_REGISTERED;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700670 up_write(&shrinker_rwsem);
Tetsuo Handa8e049442018-04-04 19:53:07 +0900671}
672
673int register_shrinker(struct shrinker *shrinker)
674{
675 int err = prealloc_shrinker(shrinker);
676
677 if (err)
678 return err;
679 register_shrinker_prepared(shrinker);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000680 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700682EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
684/*
685 * Remove one
686 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700687void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
Yang Shi41ca6682021-05-04 18:36:29 -0700689 if (!(shrinker->flags & SHRINKER_REGISTERED))
Tetsuo Handabb422a72017-12-18 20:31:41 +0900690 return;
Yang Shi41ca6682021-05-04 18:36:29 -0700691
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 down_write(&shrinker_rwsem);
693 list_del(&shrinker->list);
Yang Shi41ca6682021-05-04 18:36:29 -0700694 shrinker->flags &= ~SHRINKER_REGISTERED;
695 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
696 unregister_memcg_shrinker(shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 up_write(&shrinker_rwsem);
Yang Shi41ca6682021-05-04 18:36:29 -0700698
Andrew Vaginae393322013-10-16 13:46:46 -0700699 kfree(shrinker->nr_deferred);
Tetsuo Handabb422a72017-12-18 20:31:41 +0900700 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700702EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000705
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800706static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
Josef Bacik9092c712018-01-31 16:16:26 -0800707 struct shrinker *shrinker, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000709 unsigned long freed = 0;
710 unsigned long long delta;
711 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700712 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000713 long nr;
714 long new_nr;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000715 long batch_size = shrinker->batch ? shrinker->batch
716 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800717 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000718
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700719 freeable = shrinker->count_objects(shrinker, shrinkctl);
Kirill Tkhai9b996462018-08-17 15:48:21 -0700720 if (freeable == 0 || freeable == SHRINK_EMPTY)
721 return freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000722
723 /*
724 * copy the current shrinker scan count into a local variable
725 * and zero it so that other concurrent shrinker invocations
726 * don't also do this scanning work.
727 */
Yang Shi86750832021-05-04 18:36:36 -0700728 nr = xchg_nr_deferred(shrinker, shrinkctl);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000729
Johannes Weiner4b85afb2018-10-26 15:06:42 -0700730 if (shrinker->seeks) {
731 delta = freeable >> priority;
732 delta *= 4;
733 do_div(delta, shrinker->seeks);
734 } else {
735 /*
736 * These objects don't require any IO to create. Trim
737 * them aggressively under memory pressure to keep
738 * them from causing refetches in the IO caches.
739 */
740 delta = freeable / 2;
741 }
Roman Gushchin172b06c32018-09-20 12:22:46 -0700742
Yang Shi18bb4732021-05-04 18:36:45 -0700743 total_scan = nr >> priority;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000744 total_scan += delta;
Yang Shi18bb4732021-05-04 18:36:45 -0700745 total_scan = min(total_scan, (2 * freeable));
Glauber Costa1d3d4432013-08-28 10:18:04 +1000746
747 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Josef Bacik9092c712018-01-31 16:16:26 -0800748 freeable, delta, total_scan, priority);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000749
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800750 /*
751 * Normally, we should not scan less than batch_size objects in one
752 * pass to avoid too frequent shrinker calls, but if the slab has less
753 * than batch_size objects in total and we are really tight on memory,
754 * we will try to reclaim all available objects, otherwise we can end
755 * up failing allocations although there are plenty of reclaimable
756 * objects spread over several slabs with usage less than the
757 * batch_size.
758 *
759 * We detect the "tight on memory" situations by looking at the total
760 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700761 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800762 * scanning at high prio and therefore should try to reclaim as much as
763 * possible.
764 */
765 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700766 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000767 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800768 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000769
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800770 shrinkctl->nr_to_scan = nr_to_scan;
Chris Wilsond460acb2017-09-06 16:19:26 -0700771 shrinkctl->nr_scanned = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000772 ret = shrinker->scan_objects(shrinker, shrinkctl);
773 if (ret == SHRINK_STOP)
774 break;
775 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000776
Chris Wilsond460acb2017-09-06 16:19:26 -0700777 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
778 total_scan -= shrinkctl->nr_scanned;
779 scanned += shrinkctl->nr_scanned;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000780
781 cond_resched();
782 }
783
Yang Shi18bb4732021-05-04 18:36:45 -0700784 /*
785 * The deferred work is increased by any new work (delta) that wasn't
786 * done, decreased by old deferred work that was done now.
787 *
788 * And it is capped to two times of the freeable items.
789 */
790 next_deferred = max_t(long, (nr + delta - scanned), 0);
791 next_deferred = min(next_deferred, (2 * freeable));
792
Glauber Costa1d3d4432013-08-28 10:18:04 +1000793 /*
794 * move the unused scan count back into the shrinker in a
Yang Shi86750832021-05-04 18:36:36 -0700795 * manner that handles concurrent updates.
Glauber Costa1d3d4432013-08-28 10:18:04 +1000796 */
Yang Shi86750832021-05-04 18:36:36 -0700797 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000798
Yang Shi8efb4b52021-05-04 18:36:08 -0700799 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000800 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801}
802
Yang Shi0a432dc2019-09-23 15:38:12 -0700803#ifdef CONFIG_MEMCG
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700804static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
805 struct mem_cgroup *memcg, int priority)
806{
Yang Shie4262c42021-05-04 18:36:23 -0700807 struct shrinker_info *info;
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700808 unsigned long ret, freed = 0;
809 int i;
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700810
Yang Shi0a432dc2019-09-23 15:38:12 -0700811 if (!mem_cgroup_online(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700812 return 0;
813
814 if (!down_read_trylock(&shrinker_rwsem))
815 return 0;
816
Yang Shi468ab842021-05-04 18:36:26 -0700817 info = shrinker_info_protected(memcg, nid);
Yang Shie4262c42021-05-04 18:36:23 -0700818 if (unlikely(!info))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700819 goto unlock;
820
Yang Shie4262c42021-05-04 18:36:23 -0700821 for_each_set_bit(i, info->map, shrinker_nr_max) {
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700822 struct shrink_control sc = {
823 .gfp_mask = gfp_mask,
824 .nid = nid,
825 .memcg = memcg,
826 };
827 struct shrinker *shrinker;
828
829 shrinker = idr_find(&shrinker_idr, i);
Yang Shi41ca6682021-05-04 18:36:29 -0700830 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700831 if (!shrinker)
Yang Shie4262c42021-05-04 18:36:23 -0700832 clear_bit(i, info->map);
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700833 continue;
834 }
835
Yang Shi0a432dc2019-09-23 15:38:12 -0700836 /* Call non-slab shrinkers even though kmem is disabled */
837 if (!memcg_kmem_enabled() &&
838 !(shrinker->flags & SHRINKER_NONSLAB))
839 continue;
840
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700841 ret = do_shrink_slab(&sc, shrinker, priority);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700842 if (ret == SHRINK_EMPTY) {
Yang Shie4262c42021-05-04 18:36:23 -0700843 clear_bit(i, info->map);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700844 /*
845 * After the shrinker reported that it had no objects to
846 * free, but before we cleared the corresponding bit in
847 * the memcg shrinker map, a new object might have been
848 * added. To make sure, we have the bit set in this
849 * case, we invoke the shrinker one more time and reset
850 * the bit if it reports that it is not empty anymore.
851 * The memory barrier here pairs with the barrier in
Yang Shi2bfd3632021-05-04 18:36:11 -0700852 * set_shrinker_bit():
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700853 *
854 * list_lru_add() shrink_slab_memcg()
855 * list_add_tail() clear_bit()
856 * <MB> <MB>
857 * set_bit() do_shrink_slab()
858 */
859 smp_mb__after_atomic();
860 ret = do_shrink_slab(&sc, shrinker, priority);
861 if (ret == SHRINK_EMPTY)
862 ret = 0;
863 else
Yang Shi2bfd3632021-05-04 18:36:11 -0700864 set_shrinker_bit(memcg, nid, i);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700865 }
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700866 freed += ret;
867
868 if (rwsem_is_contended(&shrinker_rwsem)) {
869 freed = freed ? : 1;
870 break;
871 }
872 }
873unlock:
874 up_read(&shrinker_rwsem);
875 return freed;
876}
Yang Shi0a432dc2019-09-23 15:38:12 -0700877#else /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700878static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
879 struct mem_cgroup *memcg, int priority)
880{
881 return 0;
882}
Yang Shi0a432dc2019-09-23 15:38:12 -0700883#endif /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700884
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800885/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800886 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800887 * @gfp_mask: allocation context
888 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800889 * @memcg: memory cgroup whose slab caches to target
Josef Bacik9092c712018-01-31 16:16:26 -0800890 * @priority: the reclaim priority
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800892 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800894 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
895 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 *
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700897 * @memcg specifies the memory cgroup to target. Unaware shrinkers
898 * are called only if it is the root cgroup.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800899 *
Josef Bacik9092c712018-01-31 16:16:26 -0800900 * @priority is sc->priority, we take the number of objects and >> by priority
901 * in order to get the scan target.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800903 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800905static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
906 struct mem_cgroup *memcg,
Josef Bacik9092c712018-01-31 16:16:26 -0800907 int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908{
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700909 unsigned long ret, freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 struct shrinker *shrinker;
911
Yang Shifa1e5122019-08-02 21:48:44 -0700912 /*
913 * The root memcg might be allocated even though memcg is disabled
914 * via "cgroup_disable=memory" boot parameter. This could make
915 * mem_cgroup_is_root() return false, then just run memcg slab
916 * shrink, but skip global shrink. This may result in premature
917 * oom.
918 */
919 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700920 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800921
Tetsuo Handae830c632018-04-05 16:23:35 -0700922 if (!down_read_trylock(&shrinker_rwsem))
Minchan Kimf06590b2011-05-24 17:11:11 -0700923 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
925 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800926 struct shrink_control sc = {
927 .gfp_mask = gfp_mask,
928 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800929 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800930 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800931
Kirill Tkhai9b996462018-08-17 15:48:21 -0700932 ret = do_shrink_slab(&sc, shrinker, priority);
933 if (ret == SHRINK_EMPTY)
934 ret = 0;
935 freed += ret;
Minchan Kime4966122018-01-31 16:16:55 -0800936 /*
937 * Bail out if someone want to register a new shrinker to
Ethon Paul55b65a52020-06-04 16:49:10 -0700938 * prevent the registration from being stalled for long periods
Minchan Kime4966122018-01-31 16:16:55 -0800939 * by parallel ongoing shrinking.
940 */
941 if (rwsem_is_contended(&shrinker_rwsem)) {
942 freed = freed ? : 1;
943 break;
944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800946
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700948out:
949 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000950 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951}
952
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800953void drop_slab_node(int nid)
954{
955 unsigned long freed;
Vlastimil Babka1399af72021-09-02 14:59:53 -0700956 int shift = 0;
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800957
958 do {
959 struct mem_cgroup *memcg = NULL;
960
Chunxin Zang069c4112020-10-13 16:56:46 -0700961 if (fatal_signal_pending(current))
962 return;
963
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800964 freed = 0;
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700965 memcg = mem_cgroup_iter(NULL, NULL, NULL);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800966 do {
Josef Bacik9092c712018-01-31 16:16:26 -0800967 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800968 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
Vlastimil Babka1399af72021-09-02 14:59:53 -0700969 } while ((freed >> shift++) > 1);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800970}
971
972void drop_slab(void)
973{
974 int nid;
975
976 for_each_online_node(nid)
977 drop_slab_node(nid);
978}
979
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980static inline int is_page_cache_freeable(struct page *page)
981{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700982 /*
983 * A freeable page cache page is referenced only by the caller
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400984 * that isolated the page, the page cache and optional buffer
985 * heads at page->private.
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700986 */
Matthew Wilcox (Oracle)3efe62e2020-10-15 20:05:56 -0700987 int page_cache_pins = thp_nr_pages(page);
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400988 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989}
990
Yang Shicb165562019-11-30 17:55:28 -0800991static int may_write_to_inode(struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992{
Christoph Lameter930d9152006-01-08 01:00:47 -0800993 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400995 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400997 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 return 1;
999 return 0;
1000}
1001
1002/*
1003 * We detected a synchronous write error writing a page out. Probably
1004 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1005 * fsync(), msync() or close().
1006 *
1007 * The tricky part is that after writepage we cannot touch the mapping: nothing
1008 * prevents it from being freed up. But we have a ref on the page and once
1009 * that page is locked, the mapping is pinned.
1010 *
1011 * We're allowed to run sleeping lock_page() here because we know the caller has
1012 * __GFP_FS.
1013 */
1014static void handle_write_error(struct address_space *mapping,
1015 struct page *page, int error)
1016{
Jens Axboe7eaceac2011-03-10 08:52:07 +01001017 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -07001018 if (page_mapping(page) == mapping)
1019 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 unlock_page(page);
1021}
1022
Christoph Lameter04e62a22006-06-23 02:03:38 -07001023/* possible outcome of pageout() */
1024typedef enum {
1025 /* failed to write page out, page is locked */
1026 PAGE_KEEP,
1027 /* move page to the active list, page is locked */
1028 PAGE_ACTIVATE,
1029 /* page has been sent to the disk successfully, page is unlocked */
1030 PAGE_SUCCESS,
1031 /* page is clean and locked */
1032 PAGE_CLEAN,
1033} pageout_t;
1034
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035/*
Andrew Morton1742f192006-03-22 00:08:21 -08001036 * pageout is called by shrink_page_list() for each dirty page.
1037 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 */
Yang Shicb165562019-11-30 17:55:28 -08001039static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040{
1041 /*
1042 * If the page is dirty, only perform writeback if that write
1043 * will be non-blocking. To prevent this allocation from being
1044 * stalled by pagecache activity. But note that there may be
1045 * stalls if we need to run get_block(). We could test
1046 * PagePrivate for that.
1047 *
Al Viro81742022014-04-03 03:17:43 -04001048 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 * this page's queue, we can perform writeback even if that
1050 * will block.
1051 *
1052 * If the page is swapcache, write it back even if that would
1053 * block, for some throttling. This happens by accident, because
1054 * swap_backing_dev_info is bust: it doesn't reflect the
1055 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 */
1057 if (!is_page_cache_freeable(page))
1058 return PAGE_KEEP;
1059 if (!mapping) {
1060 /*
1061 * Some data journaling orphaned pages can have
1062 * page->mapping == NULL while being dirty with clean buffers.
1063 */
David Howells266cf652009-04-03 16:42:36 +01001064 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 if (try_to_free_buffers(page)) {
1066 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -07001067 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 return PAGE_CLEAN;
1069 }
1070 }
1071 return PAGE_KEEP;
1072 }
1073 if (mapping->a_ops->writepage == NULL)
1074 return PAGE_ACTIVATE;
Yang Shicb165562019-11-30 17:55:28 -08001075 if (!may_write_to_inode(mapping->host))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 return PAGE_KEEP;
1077
1078 if (clear_page_dirty_for_io(page)) {
1079 int res;
1080 struct writeback_control wbc = {
1081 .sync_mode = WB_SYNC_NONE,
1082 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001083 .range_start = 0,
1084 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 .for_reclaim = 1,
1086 };
1087
1088 SetPageReclaim(page);
1089 res = mapping->a_ops->writepage(page, &wbc);
1090 if (res < 0)
1091 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -08001092 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 ClearPageReclaim(page);
1094 return PAGE_ACTIVATE;
1095 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001096
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 if (!PageWriteback(page)) {
1098 /* synchronous write or broken a_ops? */
1099 ClearPageReclaim(page);
1100 }
yalin wang3aa23852016-01-14 15:18:30 -08001101 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -07001102 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 return PAGE_SUCCESS;
1104 }
1105
1106 return PAGE_CLEAN;
1107}
1108
Andrew Mortona649fd92006-10-17 00:09:36 -07001109/*
Nick Piggine2867812008-07-25 19:45:30 -07001110 * Same as remove_mapping, but if the page is removed from the mapping, it
1111 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -07001112 */
Johannes Weinera5289102014-04-03 14:47:51 -07001113static int __remove_mapping(struct address_space *mapping, struct page *page,
Johannes Weinerb9107182019-11-30 17:55:59 -08001114 bool reclaimed, struct mem_cgroup *target_memcg)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001115{
Huang Yingbd4c82c22017-09-06 16:22:49 -07001116 int refcount;
Joonsoo Kimaae466b2020-08-11 18:30:50 -07001117 void *shadow = NULL;
Greg Thelenc4843a72015-05-22 17:13:16 -04001118
Nick Piggin28e4d962006-09-25 23:31:23 -07001119 BUG_ON(!PageLocked(page));
1120 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001121
Johannes Weiner30472502021-09-02 14:53:18 -07001122 xa_lock_irq(&mapping->i_pages);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001123 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -07001124 * The non racy check for a busy page.
1125 *
1126 * Must be careful with the order of the tests. When someone has
1127 * a ref to the page, it may be possible that they dirty it then
1128 * drop the reference. So if PageDirty is tested before page_count
1129 * here, then the following race may occur:
1130 *
1131 * get_user_pages(&page);
1132 * [user mapping goes away]
1133 * write_to(page);
1134 * !PageDirty(page) [good]
1135 * SetPageDirty(page);
1136 * put_page(page);
1137 * !page_count(page) [good, discard it]
1138 *
1139 * [oops, our write_to data is lost]
1140 *
1141 * Reversing the order of the tests ensures such a situation cannot
1142 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001143 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -07001144 *
1145 * Note that if SetPageDirty is always performed via set_page_dirty,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001146 * and thus under the i_pages lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001147 */
William Kucharski906d2782019-10-18 20:20:33 -07001148 refcount = 1 + compound_nr(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -07001149 if (!page_ref_freeze(page, refcount))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001150 goto cannot_free;
Jiang Biao1c4c3b92018-08-21 21:53:13 -07001151 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
Nick Piggine2867812008-07-25 19:45:30 -07001152 if (unlikely(PageDirty(page))) {
Huang Yingbd4c82c22017-09-06 16:22:49 -07001153 page_ref_unfreeze(page, refcount);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001154 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -07001155 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001156
1157 if (PageSwapCache(page)) {
1158 swp_entry_t swap = { .val = page_private(page) };
Yu Zhaoa1537a62022-01-27 20:32:37 -07001159
1160 /* get a shadow entry before mem_cgroup_swapout() clears page_memcg() */
Joonsoo Kimaae466b2020-08-11 18:30:50 -07001161 if (reclaimed && !mapping_exiting(mapping))
1162 shadow = workingset_eviction(page, target_memcg);
Yu Zhaoa1537a62022-01-27 20:32:37 -07001163 mem_cgroup_swapout(page, swap);
Joonsoo Kimaae466b2020-08-11 18:30:50 -07001164 __delete_from_swap_cache(page, swap, shadow);
Johannes Weiner30472502021-09-02 14:53:18 -07001165 xa_unlock_irq(&mapping->i_pages);
Minchan Kim75f6d6d2017-07-06 15:37:21 -07001166 put_swap_page(page, swap);
Nick Piggine2867812008-07-25 19:45:30 -07001167 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -05001168 void (*freepage)(struct page *);
1169
1170 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -07001171 /*
1172 * Remember a shadow entry for reclaimed file cache in
1173 * order to detect refaults, thus thrashing, later on.
1174 *
1175 * But don't store shadows in an address space that is
dylan-meiners238c3042020-08-06 23:26:29 -07001176 * already exiting. This is not just an optimization,
Johannes Weinera5289102014-04-03 14:47:51 -07001177 * inode reclaim needs to empty out the radix tree or
1178 * the nodes are lost. Don't plant shadows behind its
1179 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -08001180 *
1181 * We also don't store shadows for DAX mappings because the
1182 * only page cache pages found in these are zero pages
1183 * covering holes, and because we don't want to mix DAX
1184 * exceptional entries and shadow exceptional entries in the
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001185 * same address_space.
Johannes Weinera5289102014-04-03 14:47:51 -07001186 */
Huang Ying9de4f222020-04-06 20:04:41 -07001187 if (reclaimed && page_is_file_lru(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -08001188 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinerb9107182019-11-30 17:55:59 -08001189 shadow = workingset_eviction(page, target_memcg);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001190 __delete_from_page_cache(page, shadow);
Johannes Weiner30472502021-09-02 14:53:18 -07001191 xa_unlock_irq(&mapping->i_pages);
Linus Torvalds6072d132010-12-01 13:35:19 -05001192
1193 if (freepage != NULL)
1194 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001195 }
1196
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001197 return 1;
1198
1199cannot_free:
Johannes Weiner30472502021-09-02 14:53:18 -07001200 xa_unlock_irq(&mapping->i_pages);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001201 return 0;
1202}
1203
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204/*
Nick Piggine2867812008-07-25 19:45:30 -07001205 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1206 * someone else has a ref on the page, abort and return 0. If it was
1207 * successfully detached, return 1. Assumes the caller has a single ref on
1208 * this page.
1209 */
1210int remove_mapping(struct address_space *mapping, struct page *page)
1211{
Johannes Weinerb9107182019-11-30 17:55:59 -08001212 if (__remove_mapping(mapping, page, false, NULL)) {
Nick Piggine2867812008-07-25 19:45:30 -07001213 /*
1214 * Unfreezing the refcount with 1 rather than 2 effectively
1215 * drops the pagecache ref for us without requiring another
1216 * atomic operation.
1217 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001218 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -07001219 return 1;
1220 }
1221 return 0;
1222}
1223
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001224/**
1225 * putback_lru_page - put previously isolated page onto appropriate LRU list
1226 * @page: page to be put back to appropriate lru list
1227 *
1228 * Add previously isolated @page to appropriate LRU list.
1229 * Page may still be unevictable for other reasons.
1230 *
1231 * lru_lock must not be held, interrupts must be enabled.
1232 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001233void putback_lru_page(struct page *page)
1234{
Shakeel Butt9c4e6b12018-02-21 14:45:28 -08001235 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001236 put_page(page); /* drop ref from isolate */
1237}
1238
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001239enum page_references {
1240 PAGEREF_RECLAIM,
1241 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -08001242 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001243 PAGEREF_ACTIVATE,
1244};
1245
1246static enum page_references page_check_references(struct page *page,
1247 struct scan_control *sc)
1248{
Johannes Weiner645747462010-03-05 13:42:22 -08001249 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001250 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001251
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001252 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1253 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -08001254 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001255
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001256 /*
1257 * Mlock lost the isolation race with us. Let try_to_unmap()
1258 * move the page to the unevictable list.
1259 */
1260 if (vm_flags & VM_LOCKED)
1261 return PAGEREF_RECLAIM;
1262
Johannes Weiner645747462010-03-05 13:42:22 -08001263 if (referenced_ptes) {
Johannes Weiner645747462010-03-05 13:42:22 -08001264 /*
1265 * All mapped pages start out with page table
1266 * references from the instantiating fault, so we need
1267 * to look twice if a mapped file page is used more
1268 * than once.
1269 *
1270 * Mark it and spare it for another trip around the
1271 * inactive list. Another page table reference will
1272 * lead to its activation.
1273 *
1274 * Note: the mark is set for activated pages as well
1275 * so that recently deactivated but used pages are
1276 * quickly recovered.
1277 */
1278 SetPageReferenced(page);
1279
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -08001280 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -08001281 return PAGEREF_ACTIVATE;
1282
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001283 /*
1284 * Activate file-backed executable pages after first usage.
1285 */
Joonsoo Kimb5181542020-08-11 18:30:40 -07001286 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001287 return PAGEREF_ACTIVATE;
1288
Johannes Weiner645747462010-03-05 13:42:22 -08001289 return PAGEREF_KEEP;
1290 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001291
1292 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -07001293 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001294 return PAGEREF_RECLAIM_CLEAN;
1295
1296 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001297}
1298
Mel Gormane2be15f2013-07-03 15:01:57 -07001299/* Check if a page is dirty or under writeback */
1300static void page_check_dirty_writeback(struct page *page,
1301 bool *dirty, bool *writeback)
1302{
Mel Gormanb4597222013-07-03 15:02:05 -07001303 struct address_space *mapping;
1304
Mel Gormane2be15f2013-07-03 15:01:57 -07001305 /*
1306 * Anonymous pages are not handled by flushers and must be written
1307 * from reclaim context. Do not stall reclaim based on them
1308 */
Huang Ying9de4f222020-04-06 20:04:41 -07001309 if (!page_is_file_lru(page) ||
Shaohua Li802a3a92017-05-03 14:52:32 -07001310 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -07001311 *dirty = false;
1312 *writeback = false;
1313 return;
1314 }
1315
1316 /* By default assume that the page flags are accurate */
1317 *dirty = PageDirty(page);
1318 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -07001319
1320 /* Verify dirty/writeback state if the filesystem supports it */
1321 if (!page_has_private(page))
1322 return;
1323
1324 mapping = page_mapping(page);
1325 if (mapping && mapping->a_ops->is_dirty_writeback)
1326 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -07001327}
1328
Dave Hansen26aa2d12021-09-02 14:59:16 -07001329static struct page *alloc_demote_page(struct page *page, unsigned long node)
1330{
1331 struct migration_target_control mtc = {
1332 /*
1333 * Allocate from 'node', or fail quickly and quietly.
1334 * When this happens, 'page' will likely just be discarded
1335 * instead of migrated.
1336 */
1337 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
1338 __GFP_THISNODE | __GFP_NOWARN |
1339 __GFP_NOMEMALLOC | GFP_NOWAIT,
1340 .nid = node
1341 };
1342
1343 return alloc_migration_target(page, (unsigned long)&mtc);
1344}
1345
1346/*
1347 * Take pages on @demote_list and attempt to demote them to
1348 * another node. Pages which are not demoted are left on
1349 * @demote_pages.
1350 */
1351static unsigned int demote_page_list(struct list_head *demote_pages,
1352 struct pglist_data *pgdat)
1353{
1354 int target_nid = next_demotion_node(pgdat->node_id);
1355 unsigned int nr_succeeded;
1356 int err;
1357
1358 if (list_empty(demote_pages))
1359 return 0;
1360
1361 if (target_nid == NUMA_NO_NODE)
1362 return 0;
1363
1364 /* Demotion ignores all cpuset and mempolicy settings */
1365 err = migrate_pages(demote_pages, alloc_demote_page, NULL,
1366 target_nid, MIGRATE_ASYNC, MR_DEMOTION,
1367 &nr_succeeded);
1368
Yang Shi668e4142021-09-02 14:59:19 -07001369 if (current_is_kswapd())
1370 __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1371 else
1372 __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1373
Dave Hansen26aa2d12021-09-02 14:59:16 -07001374 return nr_succeeded;
1375}
1376
Nick Piggine2867812008-07-25 19:45:30 -07001377/*
Andrew Morton1742f192006-03-22 00:08:21 -08001378 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 */
Maninder Singh730ec8c2020-06-03 16:01:18 -07001380static unsigned int shrink_page_list(struct list_head *page_list,
1381 struct pglist_data *pgdat,
1382 struct scan_control *sc,
Maninder Singh730ec8c2020-06-03 16:01:18 -07001383 struct reclaim_stat *stat,
1384 bool ignore_references)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385{
1386 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001387 LIST_HEAD(free_pages);
Dave Hansen26aa2d12021-09-02 14:59:16 -07001388 LIST_HEAD(demote_pages);
Maninder Singh730ec8c2020-06-03 16:01:18 -07001389 unsigned int nr_reclaimed = 0;
1390 unsigned int pgactivate = 0;
Dave Hansen26aa2d12021-09-02 14:59:16 -07001391 bool do_demote_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
Kirill Tkhai060f0052019-03-05 15:48:15 -08001393 memset(stat, 0, sizeof(*stat));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 cond_resched();
Dave Hansen26aa2d12021-09-02 14:59:16 -07001395 do_demote_pass = can_demote(pgdat->node_id, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Dave Hansen26aa2d12021-09-02 14:59:16 -07001397retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 while (!list_empty(page_list)) {
1399 struct address_space *mapping;
1400 struct page *page;
Minchan Kim8940b342019-09-25 16:49:11 -07001401 enum page_references references = PAGEREF_RECLAIM;
Kirill Tkhai4b793062020-04-01 21:10:18 -07001402 bool dirty, writeback, may_enter_fs;
Yang Shi98879b32019-07-11 20:59:30 -07001403 unsigned int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404
1405 cond_resched();
1406
1407 page = lru_to_page(page_list);
1408 list_del(&page->lru);
1409
Nick Piggin529ae9a2008-08-02 12:01:03 +02001410 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 goto keep;
1412
Sasha Levin309381fea2014-01-23 15:52:54 -08001413 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07001415 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07001416
1417 /* Account the number of base pages even though THP */
1418 sc->nr_scanned += nr_pages;
Christoph Lameter80e43422006-02-11 17:55:53 -08001419
Hugh Dickins39b5f292012-10-08 16:33:18 -07001420 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -07001421 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001422
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001423 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -08001424 goto keep_locked;
1425
Yu Zhaoafd94c92022-01-27 20:43:22 -07001426 /* page_update_gen() tried to promote this page? */
1427 if (lru_gen_enabled() && !ignore_references &&
1428 page_mapped(page) && PageReferenced(page))
1429 goto keep_locked;
1430
Andy Whitcroftc661b072007-08-22 14:01:26 -07001431 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1432 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1433
Mel Gorman283aba92013-07-03 15:01:51 -07001434 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07001435 * The number of dirty pages determines if a node is marked
Mel Gormane2be15f2013-07-03 15:01:57 -07001436 * reclaim_congested which affects wait_iff_congested. kswapd
1437 * will stall and start writing pages if the tail of the LRU
1438 * is all dirty unqueued pages.
1439 */
1440 page_check_dirty_writeback(page, &dirty, &writeback);
1441 if (dirty || writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001442 stat->nr_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001443
1444 if (dirty && !writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001445 stat->nr_unqueued_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001446
Mel Gormand04e8ac2013-07-03 15:02:03 -07001447 /*
1448 * Treat this page as congested if the underlying BDI is or if
1449 * pages are cycling through the LRU so quickly that the
1450 * pages marked for immediate reclaim are making it to the
1451 * end of the LRU a second time.
1452 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001453 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001454 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001455 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001456 (writeback && PageReclaim(page)))
Kirill Tkhai060f0052019-03-05 15:48:15 -08001457 stat->nr_congested++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001458
1459 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001460 * If a page at the tail of the LRU is under writeback, there
1461 * are three cases to consider.
1462 *
1463 * 1) If reclaim is encountering an excessive number of pages
1464 * under writeback and this page is both under writeback and
1465 * PageReclaim then it indicates that pages are being queued
1466 * for IO but are being recycled through the LRU before the
1467 * IO can complete. Waiting on the page itself risks an
1468 * indefinite stall if it is impossible to writeback the
1469 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001470 * note that the LRU is being scanned too quickly and the
1471 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001472 *
Tejun Heo97c93412015-05-22 18:23:36 -04001473 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001474 * not marked for immediate reclaim, or the caller does not
1475 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1476 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001477 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001478 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001479 * Require may_enter_fs because we would wait on fs, which
1480 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001481 * enter reclaim, and deadlock if it waits on a page for
1482 * which it is needed to do the write (loop masks off
1483 * __GFP_IO|__GFP_FS for this reason); but more thought
1484 * would probably show more reasons.
1485 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001486 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001487 * PageReclaim. memcg does not have any dirty pages
1488 * throttling so we could easily OOM just because too many
1489 * pages are in writeback and there is nothing else to
1490 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001491 *
1492 * In cases 1) and 2) we activate the pages to get them out of
1493 * the way while we continue scanning for clean pages on the
1494 * inactive list and refilling from the active list. The
1495 * observation here is that waiting for disk writes is more
1496 * expensive than potentially causing reloads down the line.
1497 * Since they're marked for immediate reclaim, they won't put
1498 * memory pressure on the cache working set any longer than it
1499 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001500 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001501 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001502 /* Case 1 above */
1503 if (current_is_kswapd() &&
1504 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001505 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Kirill Tkhai060f0052019-03-05 15:48:15 -08001506 stat->nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001507 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001508
1509 /* Case 2 above */
Johannes Weinerb5ead352019-11-30 17:55:40 -08001510 } else if (writeback_throttling_sane(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001511 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001512 /*
1513 * This is slightly racy - end_page_writeback()
1514 * might have just cleared PageReclaim, then
1515 * setting PageReclaim here end up interpreted
1516 * as PageReadahead - but that does not matter
1517 * enough to care. What we do want is for this
1518 * page to have PageReclaim set next time memcg
1519 * reclaim reaches the tests above, so it will
1520 * then wait_on_page_writeback() to avoid OOM;
1521 * and it's also appropriate in global reclaim.
1522 */
1523 SetPageReclaim(page);
Kirill Tkhai060f0052019-03-05 15:48:15 -08001524 stat->nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001525 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001526
1527 /* Case 3 above */
1528 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001529 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001530 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001531 /* then go back and try same page again */
1532 list_add_tail(&page->lru, page_list);
1533 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001534 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Minchan Kim8940b342019-09-25 16:49:11 -07001537 if (!ignore_references)
Minchan Kim02c6de82012-10-08 16:31:55 -07001538 references = page_check_references(page, sc);
1539
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001540 switch (references) {
1541 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001543 case PAGEREF_KEEP:
Yang Shi98879b32019-07-11 20:59:30 -07001544 stat->nr_ref_keep += nr_pages;
Johannes Weiner645747462010-03-05 13:42:22 -08001545 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001546 case PAGEREF_RECLAIM:
1547 case PAGEREF_RECLAIM_CLEAN:
1548 ; /* try to reclaim the page below */
1549 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 /*
Dave Hansen26aa2d12021-09-02 14:59:16 -07001552 * Before reclaiming the page, try to relocate
1553 * its contents to another node.
1554 */
1555 if (do_demote_pass &&
1556 (thp_migration_supported() || !PageTransHuge(page))) {
1557 list_add(&page->lru, &demote_pages);
1558 unlock_page(page);
1559 continue;
1560 }
1561
1562 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 * Anonymous process memory has backing store?
1564 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001565 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001567 if (PageAnon(page) && PageSwapBacked(page)) {
1568 if (!PageSwapCache(page)) {
1569 if (!(sc->gfp_mask & __GFP_IO))
1570 goto keep_locked;
Linus Torvaldsfeb889f2021-01-16 15:34:57 -08001571 if (page_maybe_dma_pinned(page))
1572 goto keep_locked;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001573 if (PageTransHuge(page)) {
1574 /* cannot split THP, skip it */
1575 if (!can_split_huge_page(page, NULL))
1576 goto activate_locked;
1577 /*
1578 * Split pages without a PMD map right
1579 * away. Chances are some or all of the
1580 * tail pages can be freed without IO.
1581 */
1582 if (!compound_mapcount(page) &&
1583 split_huge_page_to_list(page,
1584 page_list))
1585 goto activate_locked;
1586 }
1587 if (!add_to_swap(page)) {
1588 if (!PageTransHuge(page))
Yang Shi98879b32019-07-11 20:59:30 -07001589 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001590 /* Fallback to swap normal pages */
1591 if (split_huge_page_to_list(page,
1592 page_list))
1593 goto activate_locked;
Huang Yingfe490cc2017-09-06 16:22:52 -07001594#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1595 count_vm_event(THP_SWPOUT_FALLBACK);
1596#endif
Huang Yingbd4c82c22017-09-06 16:22:49 -07001597 if (!add_to_swap(page))
Yang Shi98879b32019-07-11 20:59:30 -07001598 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001599 }
Minchan Kim0f074652017-07-06 15:37:24 -07001600
Kirill Tkhai4b793062020-04-01 21:10:18 -07001601 may_enter_fs = true;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001602
1603 /* Adding to swap updated mapping */
1604 mapping = page_mapping(page);
Minchan Kim0f074652017-07-06 15:37:24 -07001605 }
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001606 } else if (unlikely(PageTransHuge(page))) {
1607 /* Split file THP */
1608 if (split_huge_page_to_list(page, page_list))
1609 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001610 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611
1612 /*
Yang Shi98879b32019-07-11 20:59:30 -07001613 * THP may get split above, need minus tail pages and update
1614 * nr_pages to avoid accounting tail pages twice.
1615 *
1616 * The tail pages that are added into swap cache successfully
1617 * reach here.
1618 */
1619 if ((nr_pages > 1) && !PageTransHuge(page)) {
1620 sc->nr_scanned -= (nr_pages - 1);
1621 nr_pages = 1;
1622 }
1623
1624 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 * The page is mapped into the page tables of one or more
1626 * processes. Try to unmap it here.
1627 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001628 if (page_mapped(page)) {
Shakeel Butt013339d2020-12-14 19:06:39 -08001629 enum ttu_flags flags = TTU_BATCH_FLUSH;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001630 bool was_swapbacked = PageSwapBacked(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -07001631
1632 if (unlikely(PageTransHuge(page)))
1633 flags |= TTU_SPLIT_HUGE_PMD;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001634
Yang Shi1fb08ac2021-06-30 18:52:01 -07001635 try_to_unmap(page, flags);
1636 if (page_mapped(page)) {
Yang Shi98879b32019-07-11 20:59:30 -07001637 stat->nr_unmap_fail += nr_pages;
Jaewon Kim1f318a92020-06-03 16:01:15 -07001638 if (!was_swapbacked && PageSwapBacked(page))
1639 stat->nr_lazyfree_fail += nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 }
1642 }
1643
1644 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001645 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001646 * Only kswapd can writeback filesystem pages
1647 * to avoid risk of stack overflow. But avoid
1648 * injecting inefficient single-page IO into
1649 * flusher writeback as much as possible: only
1650 * write pages when we've encountered many
1651 * dirty pages, and when we've already scanned
1652 * the rest of the LRU for clean pages and see
1653 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001654 */
Huang Ying9de4f222020-04-06 20:04:41 -07001655 if (page_is_file_lru(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001656 (!current_is_kswapd() || !PageReclaim(page) ||
1657 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001658 /*
1659 * Immediately reclaim when written back.
1660 * Similar in principal to deactivate_page()
1661 * except we already have the page isolated
1662 * and know it's dirty
1663 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001664 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001665 SetPageReclaim(page);
1666
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001667 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001668 }
1669
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001670 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001672 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001674 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 goto keep_locked;
1676
Mel Gormand950c942015-09-04 15:47:35 -07001677 /*
1678 * Page is dirty. Flush the TLB if a writable entry
1679 * potentially exists to avoid CPU writes after IO
1680 * starts and then write it out here.
1681 */
1682 try_to_unmap_flush_dirty();
Yang Shicb165562019-11-30 17:55:28 -08001683 switch (pageout(page, mapping)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 case PAGE_KEEP:
1685 goto keep_locked;
1686 case PAGE_ACTIVATE:
1687 goto activate_locked;
1688 case PAGE_SUCCESS:
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07001689 stat->nr_pageout += thp_nr_pages(page);
Johannes Weiner96f8bf42020-06-03 16:03:09 -07001690
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001691 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001692 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001693 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001695
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 /*
1697 * A synchronous write - probably a ramdisk. Go
1698 * ahead and try to reclaim the page.
1699 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001700 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 goto keep;
1702 if (PageDirty(page) || PageWriteback(page))
1703 goto keep_locked;
1704 mapping = page_mapping(page);
Gustavo A. R. Silva01359eb2020-12-14 19:15:00 -08001705 fallthrough;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 case PAGE_CLEAN:
1707 ; /* try to free the page below */
1708 }
1709 }
1710
1711 /*
1712 * If the page has buffers, try to free the buffer mappings
1713 * associated with this page. If we succeed we try to free
1714 * the page as well.
1715 *
1716 * We do this even if the page is PageDirty().
1717 * try_to_release_page() does not perform I/O, but it is
1718 * possible for a page to have PageDirty set, but it is actually
1719 * clean (all its buffers are clean). This happens if the
1720 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001721 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 * try_to_release_page() will discover that cleanness and will
1723 * drop the buffers and mark the page clean - it can be freed.
1724 *
1725 * Rarely, pages can have buffers and no ->mapping. These are
1726 * the pages which were not successfully invalidated in
Yang Shid12b8952020-12-14 19:13:02 -08001727 * truncate_cleanup_page(). We try to drop those buffers here
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 * and if that worked, and the page is no longer mapped into
1729 * process address space (page_count == 1) it can be freed.
1730 * Otherwise, leave the page on the LRU so it is swappable.
1731 */
David Howells266cf652009-04-03 16:42:36 +01001732 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 if (!try_to_release_page(page, sc->gfp_mask))
1734 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001735 if (!mapping && page_count(page) == 1) {
1736 unlock_page(page);
1737 if (put_page_testzero(page))
1738 goto free_it;
1739 else {
1740 /*
1741 * rare race with speculative reference.
1742 * the speculative reference will free
1743 * this page shortly, so we may
1744 * increment nr_reclaimed here (and
1745 * leave it off the LRU).
1746 */
1747 nr_reclaimed++;
1748 continue;
1749 }
1750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 }
1752
Shaohua Li802a3a92017-05-03 14:52:32 -07001753 if (PageAnon(page) && !PageSwapBacked(page)) {
1754 /* follow __remove_mapping for reference */
1755 if (!page_ref_freeze(page, 1))
1756 goto keep_locked;
Miaohe Lind17be2d2021-09-02 14:59:39 -07001757 /*
1758 * The page has only one reference left, which is
1759 * from the isolation. After the caller puts the
1760 * page back on lru and drops the reference, the
1761 * page will be freed anyway. It doesn't matter
1762 * which lru it goes. So we don't bother checking
1763 * PageDirty here.
1764 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001765 count_vm_event(PGLAZYFREED);
Roman Gushchin22621852017-07-06 15:40:25 -07001766 count_memcg_page_event(page, PGLAZYFREED);
Johannes Weinerb9107182019-11-30 17:55:59 -08001767 } else if (!mapping || !__remove_mapping(mapping, page, true,
1768 sc->target_mem_cgroup))
Shaohua Li802a3a92017-05-03 14:52:32 -07001769 goto keep_locked;
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001770
1771 unlock_page(page);
Nick Piggine2867812008-07-25 19:45:30 -07001772free_it:
Yang Shi98879b32019-07-11 20:59:30 -07001773 /*
1774 * THP may get swapped out in a whole, need account
1775 * all base pages.
1776 */
1777 nr_reclaimed += nr_pages;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001778
1779 /*
1780 * Is there need to periodically free_page_list? It would
1781 * appear not as the counts should be low
1782 */
Yang Shi7ae88532019-09-23 15:38:09 -07001783 if (unlikely(PageTransHuge(page)))
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07001784 destroy_compound_page(page);
Yang Shi7ae88532019-09-23 15:38:09 -07001785 else
Huang Yingbd4c82c22017-09-06 16:22:49 -07001786 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 continue;
1788
Yang Shi98879b32019-07-11 20:59:30 -07001789activate_locked_split:
1790 /*
1791 * The tail pages that are failed to add into swap cache
1792 * reach here. Fixup nr_scanned and nr_pages.
1793 */
1794 if (nr_pages > 1) {
1795 sc->nr_scanned -= (nr_pages - 1);
1796 nr_pages = 1;
1797 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001799 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001800 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1801 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001802 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001803 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001804 if (!PageMlocked(page)) {
Huang Ying9de4f222020-04-06 20:04:41 -07001805 int type = page_is_file_lru(page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001806 SetPageActive(page);
Yang Shi98879b32019-07-11 20:59:30 -07001807 stat->nr_activate[type] += nr_pages;
Roman Gushchin22621852017-07-06 15:40:25 -07001808 count_memcg_page_event(page, PGACTIVATE);
Minchan Kimad6b6702017-05-03 14:54:13 -07001809 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810keep_locked:
1811 unlock_page(page);
1812keep:
1813 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001814 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 }
Dave Hansen26aa2d12021-09-02 14:59:16 -07001816 /* 'page_list' is always empty here */
1817
1818 /* Migrate pages selected for demotion */
1819 nr_reclaimed += demote_page_list(&demote_pages, pgdat);
1820 /* Pages that could not be demoted are still in @demote_pages */
1821 if (!list_empty(&demote_pages)) {
1822 /* Pages which failed to demoted go back on @page_list for retry: */
1823 list_splice_init(&demote_pages, page_list);
1824 do_demote_pass = false;
1825 goto retry;
1826 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001827
Yang Shi98879b32019-07-11 20:59:30 -07001828 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1829
Johannes Weiner747db952014-08-08 14:19:24 -07001830 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001831 try_to_unmap_flush();
Mel Gorman2d4894b2017-11-15 17:37:59 -08001832 free_unref_page_list(&free_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001833
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 list_splice(&ret_pages, page_list);
Kirill Tkhai886cf192019-05-13 17:16:51 -07001835 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001836
Andrew Morton05ff5132006-03-22 00:08:20 -08001837 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838}
1839
Maninder Singh730ec8c2020-06-03 16:01:18 -07001840unsigned int reclaim_clean_pages_from_list(struct zone *zone,
Minchan Kim02c6de82012-10-08 16:31:55 -07001841 struct list_head *page_list)
1842{
1843 struct scan_control sc = {
1844 .gfp_mask = GFP_KERNEL,
Minchan Kim02c6de82012-10-08 16:31:55 -07001845 .may_unmap = 1,
1846 };
Jaewon Kim1f318a92020-06-03 16:01:15 -07001847 struct reclaim_stat stat;
Maninder Singh730ec8c2020-06-03 16:01:18 -07001848 unsigned int nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001849 struct page *page, *next;
1850 LIST_HEAD(clean_pages);
Yu Zhao2d2b8d22021-06-30 18:49:48 -07001851 unsigned int noreclaim_flag;
Minchan Kim02c6de82012-10-08 16:31:55 -07001852
1853 list_for_each_entry_safe(page, next, page_list, lru) {
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07001854 if (!PageHuge(page) && page_is_file_lru(page) &&
1855 !PageDirty(page) && !__PageMovable(page) &&
1856 !PageUnevictable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001857 ClearPageActive(page);
1858 list_move(&page->lru, &clean_pages);
1859 }
1860 }
1861
Yu Zhao2d2b8d22021-06-30 18:49:48 -07001862 /*
1863 * We should be safe here since we are only dealing with file pages and
1864 * we are not kswapd and therefore cannot write dirty file pages. But
1865 * call memalloc_noreclaim_save() anyway, just in case these conditions
1866 * change in the future.
1867 */
1868 noreclaim_flag = memalloc_noreclaim_save();
Jaewon Kim1f318a92020-06-03 16:01:15 -07001869 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Shakeel Butt013339d2020-12-14 19:06:39 -08001870 &stat, true);
Yu Zhao2d2b8d22021-06-30 18:49:48 -07001871 memalloc_noreclaim_restore(noreclaim_flag);
1872
Minchan Kim02c6de82012-10-08 16:31:55 -07001873 list_splice(&clean_pages, page_list);
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001874 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1875 -(long)nr_reclaimed);
Jaewon Kim1f318a92020-06-03 16:01:15 -07001876 /*
1877 * Since lazyfree pages are isolated from file LRU from the beginning,
1878 * they will rotate back to anonymous LRU in the end if it failed to
1879 * discard so isolated count will be mismatched.
1880 * Compensate the isolated count for both LRU lists.
1881 */
1882 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1883 stat.nr_lazyfree_fail);
1884 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001885 -(long)stat.nr_lazyfree_fail);
Jaewon Kim1f318a92020-06-03 16:01:15 -07001886 return nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001887}
1888
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001889/*
1890 * Attempt to remove the specified page from its LRU. Only take this page
1891 * if it is of the appropriate PageActive status. Pages which are being
1892 * freed elsewhere are also ignored.
1893 *
1894 * page: page to consider
1895 * mode: one of the LRU isolation modes defined above
1896 *
Alex Shic2135f72021-02-24 12:08:01 -08001897 * returns true on success, false on failure.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001898 */
Alex Shic2135f72021-02-24 12:08:01 -08001899bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001900{
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001901 /* Only take pages on the LRU. */
1902 if (!PageLRU(page))
Alex Shic2135f72021-02-24 12:08:01 -08001903 return false;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001904
Minchan Kime46a2872012-10-08 16:33:48 -07001905 /* Compaction should not handle unevictable pages but CMA can do so */
1906 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Alex Shic2135f72021-02-24 12:08:01 -08001907 return false;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001908
Mel Gormanc8244932012-01-12 17:19:38 -08001909 /*
1910 * To minimise LRU disruption, the caller can indicate that it only
1911 * wants to isolate pages it will be able to operate on without
1912 * blocking - clean pages for the most part.
1913 *
Mel Gormanc8244932012-01-12 17:19:38 -08001914 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1915 * that it is possible to migrate without blocking
1916 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001917 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001918 /* All the caller can do on PageWriteback is block */
1919 if (PageWriteback(page))
Alex Shic2135f72021-02-24 12:08:01 -08001920 return false;
Mel Gormanc8244932012-01-12 17:19:38 -08001921
1922 if (PageDirty(page)) {
1923 struct address_space *mapping;
Mel Gorman69d763f2018-01-31 16:19:52 -08001924 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001925
Mel Gormanc8244932012-01-12 17:19:38 -08001926 /*
1927 * Only pages without mappings or that have a
1928 * ->migratepage callback are possible to migrate
Mel Gorman69d763f2018-01-31 16:19:52 -08001929 * without blocking. However, we can be racing with
1930 * truncation so it's necessary to lock the page
1931 * to stabilise the mapping as truncation holds
1932 * the page lock until after the page is removed
1933 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001934 */
Mel Gorman69d763f2018-01-31 16:19:52 -08001935 if (!trylock_page(page))
Alex Shic2135f72021-02-24 12:08:01 -08001936 return false;
Mel Gorman69d763f2018-01-31 16:19:52 -08001937
Mel Gormanc8244932012-01-12 17:19:38 -08001938 mapping = page_mapping(page);
Hugh Dickins145e1a72018-06-01 16:50:50 -07001939 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gorman69d763f2018-01-31 16:19:52 -08001940 unlock_page(page);
1941 if (!migrate_dirty)
Alex Shic2135f72021-02-24 12:08:01 -08001942 return false;
Mel Gormanc8244932012-01-12 17:19:38 -08001943 }
1944 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001945
Minchan Kimf80c0672011-10-31 17:06:55 -07001946 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
Alex Shic2135f72021-02-24 12:08:01 -08001947 return false;
Minchan Kimf80c0672011-10-31 17:06:55 -07001948
Alex Shic2135f72021-02-24 12:08:01 -08001949 return true;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001950}
1951
Mel Gorman7ee36a12016-07-28 15:47:17 -07001952/*
1953 * Update LRU sizes after isolating pages. The LRU size updates must
Ethon Paul55b65a52020-06-04 16:49:10 -07001954 * be complete before mem_cgroup_update_lru_size due to a sanity check.
Mel Gorman7ee36a12016-07-28 15:47:17 -07001955 */
1956static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001957 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001958{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001959 int zid;
1960
Mel Gorman7ee36a12016-07-28 15:47:17 -07001961 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1962 if (!nr_zone_taken[zid])
1963 continue;
1964
Wei Yanga892cb62020-06-03 16:01:12 -07001965 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Mel Gorman7ee36a12016-07-28 15:47:17 -07001966 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001967
Mel Gorman7ee36a12016-07-28 15:47:17 -07001968}
1969
Mel Gormanf611fab2021-06-30 18:53:19 -07001970/*
Hugh Dickins15b44732020-12-15 14:21:31 -08001971 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1972 *
1973 * lruvec->lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 * shrink the lists perform better by taking out a batch of pages
1975 * and working on them outside the LRU lock.
1976 *
1977 * For pagecache intensive workloads, this function is the hottest
1978 * spot in the kernel (apart from copy_*_user functions).
1979 *
Hugh Dickins15b44732020-12-15 14:21:31 -08001980 * Lru_lock must be held before calling this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 *
Minchan Kim791b48b2017-05-12 15:47:06 -07001982 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001983 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001985 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001986 * @sc: The scan_control struct for this reclaim session
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001987 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 *
1989 * returns how many pages were moved onto *@dst.
1990 */
Andrew Morton69e05942006-03-22 00:08:19 -08001991static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001992 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001993 unsigned long *nr_scanned, struct scan_control *sc,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001994 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001996 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001997 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001998 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001999 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07002000 unsigned long skipped = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07002001 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07002002 LIST_HEAD(pages_skipped);
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002003 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004
Yang Shi98879b32019-07-11 20:59:30 -07002005 total_scan = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07002006 scan = 0;
Yang Shi98879b32019-07-11 20:59:30 -07002007 while (scan < nr_to_scan && !list_empty(src)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002008 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002009
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 page = lru_to_page(src);
2011 prefetchw_prev_lru_page(page, src, flags);
2012
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07002013 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07002014 total_scan += nr_pages;
2015
Mel Gormanb2e18752016-07-28 15:45:37 -07002016 if (page_zonenum(page) > sc->reclaim_idx) {
2017 list_move(&page->lru, &pages_skipped);
Yang Shi98879b32019-07-11 20:59:30 -07002018 nr_skipped[page_zonenum(page)] += nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07002019 continue;
2020 }
2021
Minchan Kim791b48b2017-05-12 15:47:06 -07002022 /*
2023 * Do not count skipped pages because that makes the function
2024 * return with no isolated pages if the LRU mostly contains
2025 * ineligible pages. This causes the VM to not reclaim any
2026 * pages, triggering a premature OOM.
Yang Shi98879b32019-07-11 20:59:30 -07002027 *
2028 * Account all tail pages of THP. This would not cause
2029 * premature OOM since __isolate_lru_page() returns -EBUSY
2030 * only when the page is being freed somewhere else.
Minchan Kim791b48b2017-05-12 15:47:06 -07002031 */
Yang Shi98879b32019-07-11 20:59:30 -07002032 scan += nr_pages;
Alex Shic2135f72021-02-24 12:08:01 -08002033 if (!__isolate_lru_page_prepare(page, mode)) {
2034 /* It is being freed elsewhere */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002035 list_move(&page->lru, src);
Alex Shic2135f72021-02-24 12:08:01 -08002036 continue;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002037 }
Alex Shic2135f72021-02-24 12:08:01 -08002038 /*
2039 * Be careful not to clear PageLRU until after we're
2040 * sure the page is not being freed elsewhere -- the
2041 * page release code relies on it.
2042 */
2043 if (unlikely(!get_page_unless_zero(page))) {
2044 list_move(&page->lru, src);
2045 continue;
2046 }
2047
2048 if (!TestClearPageLRU(page)) {
2049 /* Another thread is already isolating this page */
2050 put_page(page);
2051 list_move(&page->lru, src);
2052 continue;
2053 }
2054
2055 nr_taken += nr_pages;
2056 nr_zone_taken[page_zonenum(page)] += nr_pages;
2057 list_move(&page->lru, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 }
2059
Mel Gormanb2e18752016-07-28 15:45:37 -07002060 /*
2061 * Splice any skipped pages to the start of the LRU list. Note that
2062 * this disrupts the LRU order when reclaiming for lower zones but
2063 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2064 * scanning would soon rescan the same pages to skip and put the
2065 * system at risk of premature OOM.
2066 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002067 if (!list_empty(&pages_skipped)) {
2068 int zid;
2069
Johannes Weiner3db65812017-05-03 14:52:13 -07002070 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002071 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2072 if (!nr_skipped[zid])
2073 continue;
2074
2075 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08002076 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002077 }
2078 }
Minchan Kim791b48b2017-05-12 15:47:06 -07002079 *nr_scanned = total_scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08002080 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim791b48b2017-05-12 15:47:06 -07002081 total_scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 16:58:04 -08002082 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 return nr_taken;
2084}
2085
Nick Piggin62695a82008-10-18 20:26:09 -07002086/**
2087 * isolate_lru_page - tries to isolate a page from its LRU list
2088 * @page: page to isolate from its LRU list
2089 *
2090 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
2091 * vmstat statistic corresponding to whatever LRU list the page was on.
2092 *
2093 * Returns 0 if the page was removed from an LRU list.
2094 * Returns -EBUSY if the page was not on an LRU list.
2095 *
2096 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002097 * the active list, it will have PageActive set. If it was found on
2098 * the unevictable list, it will have the PageUnevictable bit set. That flag
2099 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07002100 *
2101 * The vmstat statistic corresponding to the list on which the page was
2102 * found will be decremented.
2103 *
2104 * Restrictions:
Mike Rapoporta5d09be2018-02-06 15:42:19 -08002105 *
Nick Piggin62695a82008-10-18 20:26:09 -07002106 * (1) Must be called with an elevated refcount on the page. This is a
Hui Su01c47762020-10-13 16:56:49 -07002107 * fundamental difference from isolate_lru_pages (which is called
Nick Piggin62695a82008-10-18 20:26:09 -07002108 * without a stable reference).
2109 * (2) the lru_lock must not be held.
2110 * (3) interrupts must be enabled.
2111 */
2112int isolate_lru_page(struct page *page)
2113{
2114 int ret = -EBUSY;
2115
Sasha Levin309381fea2014-01-23 15:52:54 -08002116 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08002117 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07002118
Alex Shid25b5bd2020-12-15 12:34:16 -08002119 if (TestClearPageLRU(page)) {
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002120 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07002121
Alex Shid25b5bd2020-12-15 12:34:16 -08002122 get_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08002123 lruvec = lock_page_lruvec_irq(page);
Yu Zhao46ae6b22021-02-24 12:08:25 -08002124 del_page_from_lru_list(page, lruvec);
Alex Shi6168d0d2020-12-15 12:34:29 -08002125 unlock_page_lruvec_irq(lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08002126 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07002127 }
Alex Shid25b5bd2020-12-15 12:34:16 -08002128
Nick Piggin62695a82008-10-18 20:26:09 -07002129 return ret;
2130}
2131
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002132/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08002133 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
Xianting Tian178821b2019-11-30 17:56:05 -08002134 * then get rescheduled. When there are massive number of tasks doing page
Fengguang Wud37dd5d2012-12-18 14:23:28 -08002135 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2136 * the LRU list will go small and be scanned faster than necessary, leading to
2137 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07002138 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002139static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07002140 struct scan_control *sc)
2141{
2142 unsigned long inactive, isolated;
2143
2144 if (current_is_kswapd())
2145 return 0;
2146
Johannes Weinerb5ead352019-11-30 17:55:40 -08002147 if (!writeback_throttling_sane(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07002148 return 0;
2149
2150 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002151 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2152 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07002153 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07002154 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2155 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07002156 }
2157
Fengguang Wu3cf23842012-12-18 14:23:31 -08002158 /*
2159 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2160 * won't get blocked by normal direct-reclaimers, forming a circular
2161 * deadlock.
2162 */
Mel Gormand0164ad2015-11-06 16:28:21 -08002163 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08002164 inactive >>= 3;
2165
Rik van Riel35cd7812009-09-21 17:01:38 -07002166 return isolated > inactive;
2167}
2168
Kirill Tkhaia222f342019-05-13 17:17:00 -07002169/*
Hugh Dickins15b44732020-12-15 14:21:31 -08002170 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2171 * On return, @list is reused as a list of pages to be freed by the caller.
Kirill Tkhaia222f342019-05-13 17:17:00 -07002172 *
2173 * Returns the number of pages moved to the given lruvec.
2174 */
Muchun Song9ef56b72021-06-28 19:38:09 -07002175static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2176 struct list_head *list)
Mel Gorman66635622010-08-09 17:19:30 -07002177{
Kirill Tkhaia222f342019-05-13 17:17:00 -07002178 int nr_pages, nr_moved = 0;
Hugh Dickins3f797682012-01-12 17:20:07 -08002179 LIST_HEAD(pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07002180 struct page *page;
Mel Gorman66635622010-08-09 17:19:30 -07002181
Kirill Tkhaia222f342019-05-13 17:17:00 -07002182 while (!list_empty(list)) {
2183 page = lru_to_page(list);
Sasha Levin309381fea2014-01-23 15:52:54 -08002184 VM_BUG_ON_PAGE(PageLRU(page), page);
Alex Shi3d06afa2020-12-15 12:33:37 -08002185 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07002186 if (unlikely(!page_evictable(page))) {
Alex Shi6168d0d2020-12-15 12:34:29 -08002187 spin_unlock_irq(&lruvec->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07002188 putback_lru_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08002189 spin_lock_irq(&lruvec->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07002190 continue;
2191 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002192
Alex Shi3d06afa2020-12-15 12:33:37 -08002193 /*
2194 * The SetPageLRU needs to be kept here for list integrity.
2195 * Otherwise:
2196 * #0 move_pages_to_lru #1 release_pages
2197 * if !put_page_testzero
2198 * if (put_page_testzero())
2199 * !PageLRU //skip lru_lock
2200 * SetPageLRU()
2201 * list_add(&page->lru,)
2202 * list_add(&page->lru,)
2203 */
Linus Torvalds7a608572011-01-17 14:42:19 -08002204 SetPageLRU(page);
Kirill Tkhaia222f342019-05-13 17:17:00 -07002205
Alex Shi3d06afa2020-12-15 12:33:37 -08002206 if (unlikely(put_page_testzero(page))) {
Yu Zhao87560172021-02-24 12:08:28 -08002207 __clear_page_lru_flags(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002208
2209 if (unlikely(PageCompound(page))) {
Alex Shi6168d0d2020-12-15 12:34:29 -08002210 spin_unlock_irq(&lruvec->lru_lock);
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07002211 destroy_compound_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08002212 spin_lock_irq(&lruvec->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002213 } else
2214 list_add(&page->lru, &pages_to_free);
Alex Shi3d06afa2020-12-15 12:33:37 -08002215
2216 continue;
Mel Gorman66635622010-08-09 17:19:30 -07002217 }
Alex Shi3d06afa2020-12-15 12:33:37 -08002218
Alex Shiafca9152020-12-15 12:34:02 -08002219 /*
2220 * All pages were isolated from the same lruvec (and isolation
2221 * inhibits memcg migration).
2222 */
Muchun Song7467c392021-06-28 19:37:59 -07002223 VM_BUG_ON_PAGE(!page_matches_lruvec(page, lruvec), page);
Yu Zhao3a9c9782021-02-24 12:08:17 -08002224 add_page_to_lru_list(page, lruvec);
Alex Shi3d06afa2020-12-15 12:33:37 -08002225 nr_pages = thp_nr_pages(page);
Alex Shi3d06afa2020-12-15 12:33:37 -08002226 nr_moved += nr_pages;
2227 if (PageActive(page))
2228 workingset_age_nonresident(lruvec, nr_pages);
Mel Gorman66635622010-08-09 17:19:30 -07002229 }
Mel Gorman66635622010-08-09 17:19:30 -07002230
Hugh Dickins3f797682012-01-12 17:20:07 -08002231 /*
2232 * To save our caller's stack, now use input list for pages to free.
2233 */
Kirill Tkhaia222f342019-05-13 17:17:00 -07002234 list_splice(&pages_to_free, list);
2235
2236 return nr_moved;
Mel Gorman66635622010-08-09 17:19:30 -07002237}
2238
2239/*
NeilBrown399ba0b2014-06-04 16:07:42 -07002240 * If a kernel thread (such as nfsd for loop-back mounts) services
NeilBrowna37b0712020-06-01 21:48:18 -07002241 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
NeilBrown399ba0b2014-06-04 16:07:42 -07002242 * In that case we should only throttle if the backing device it is
2243 * writing to is congested. In other cases it is safe to throttle.
2244 */
2245static int current_may_throttle(void)
2246{
NeilBrowna37b0712020-06-01 21:48:18 -07002247 return !(current->flags & PF_LOCAL_THROTTLE) ||
NeilBrown399ba0b2014-06-04 16:07:42 -07002248 current->backing_dev_info == NULL ||
2249 bdi_write_congested(current->backing_dev_info);
2250}
2251
2252/*
Mel Gormanb2e18752016-07-28 15:45:37 -07002253 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08002254 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 */
Muchun Song9ef56b72021-06-28 19:38:09 -07002256static unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002257shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002258 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259{
2260 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002261 unsigned long nr_scanned;
Maninder Singh730ec8c2020-06-03 16:01:18 -07002262 unsigned int nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002263 unsigned long nr_taken;
Kirill Tkhai060f0052019-03-05 15:48:15 -08002264 struct reclaim_stat stat;
Johannes Weiner497a6c12020-06-03 16:02:34 -07002265 bool file = is_file_lru(lru);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002266 enum vm_event_item item;
Mel Gorman599d0c92016-07-28 15:45:31 -07002267 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Michal Hockodb73ee02017-09-06 16:21:11 -07002268 bool stalled = false;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07002269
Mel Gorman599d0c92016-07-28 15:45:31 -07002270 while (unlikely(too_many_isolated(pgdat, file, sc))) {
Michal Hockodb73ee02017-09-06 16:21:11 -07002271 if (stalled)
2272 return 0;
2273
2274 /* wait a bit for the reclaimer. */
2275 msleep(100);
2276 stalled = true;
Rik van Riel35cd7812009-09-21 17:01:38 -07002277
2278 /* We are about to die and free our memory. Return now. */
2279 if (fatal_signal_pending(current))
2280 return SWAP_CLUSTER_MAX;
2281 }
2282
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002284
Alex Shi6168d0d2020-12-15 12:34:29 -08002285 spin_lock_irq(&lruvec->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002287 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002288 &nr_scanned, sc, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07002289
Mel Gorman599d0c92016-07-28 15:45:31 -07002290 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002291 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08002292 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002293 __count_vm_events(item, nr_scanned);
2294 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002295 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2296
Alex Shi6168d0d2020-12-15 12:34:29 -08002297 spin_unlock_irq(&lruvec->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002298
Hillf Dantond563c052012-03-21 16:34:02 -07002299 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07002300 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002301
Shakeel Butt013339d2020-12-14 19:06:39 -08002302 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002303
Alex Shi6168d0d2020-12-15 12:34:29 -08002304 spin_lock_irq(&lruvec->lru_lock);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002305 move_pages_to_lru(lruvec, &page_list);
2306
2307 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002308 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08002309 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002310 __count_vm_events(item, nr_reclaimed);
2311 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002312 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
Alex Shi6168d0d2020-12-15 12:34:29 -08002313 spin_unlock_irq(&lruvec->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08002314
Alex Shi75cc3c92020-12-15 14:20:50 -08002315 lru_note_cost(lruvec, file, stat.nr_pageout);
Johannes Weiner747db952014-08-08 14:19:24 -07002316 mem_cgroup_uncharge_list(&page_list);
Mel Gorman2d4894b2017-11-15 17:37:59 -08002317 free_unref_page_list(&page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07002318
Mel Gorman92df3a72011-10-31 17:07:56 -07002319 /*
Andrey Ryabinin1c610d52018-03-22 16:17:42 -07002320 * If dirty pages are scanned that are not queued for IO, it
2321 * implies that flushers are not doing their job. This can
2322 * happen when memory pressure pushes dirty pages to the end of
2323 * the LRU before the dirty limits are breached and the dirty
2324 * data has expired. It can also happen when the proportion of
2325 * dirty pages grows not through writes but through memory
2326 * pressure reclaiming all the clean cache. And in some cases,
2327 * the flushers simply cannot keep up with the allocation
2328 * rate. Nudge the flusher threads in case they are asleep.
2329 */
2330 if (stat.nr_unqueued_dirty == nr_taken)
2331 wakeup_flusher_threads(WB_REASON_VMSCAN);
2332
Andrey Ryabinind108c772018-04-10 16:27:59 -07002333 sc->nr.dirty += stat.nr_dirty;
2334 sc->nr.congested += stat.nr_congested;
2335 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2336 sc->nr.writeback += stat.nr_writeback;
2337 sc->nr.immediate += stat.nr_immediate;
2338 sc->nr.taken += nr_taken;
2339 if (file)
2340 sc->nr.file_taken += nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07002341
Mel Gorman599d0c92016-07-28 15:45:31 -07002342 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
Steven Rostedtd51d1e62018-04-10 16:28:07 -07002343 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08002344 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345}
2346
Hugh Dickins15b44732020-12-15 14:21:31 -08002347/*
2348 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2349 *
2350 * We move them the other way if the page is referenced by one or more
2351 * processes.
2352 *
2353 * If the pages are mostly unmapped, the processing is fast and it is
2354 * appropriate to hold lru_lock across the whole operation. But if
2355 * the pages are mapped, the processing is slow (page_referenced()), so
2356 * we should drop lru_lock around each page. It's impossible to balance
2357 * this, so instead we remove the pages from the LRU while processing them.
2358 * It is safe to rely on PG_active against the non-LRU pages in here because
2359 * nobody will play with that bit on a non-LRU page.
2360 *
2361 * The downside is that we have to touch page->_refcount against each page.
2362 * But we had to alter page->flags anyway.
2363 */
Hugh Dickinsf6260122012-01-12 17:20:06 -08002364static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002365 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002366 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002367 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002369 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08002370 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07002371 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002373 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002374 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 struct page *page;
Michal Hocko9d998b42017-02-22 15:44:18 -08002376 unsigned nr_deactivate, nr_activate;
2377 unsigned nr_rotated = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002378 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002379 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Liujie Xie6f98ceb2022-02-23 10:32:01 +08002380 bool bypass = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
2382 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002383
Alex Shi6168d0d2020-12-15 12:34:29 -08002384 spin_lock_irq(&lruvec->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002385
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002386 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002387 &nr_scanned, sc, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002388
Mel Gorman599d0c92016-07-28 15:45:31 -07002389 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002390
Shakeel Butt912c0572020-08-06 23:26:32 -07002391 if (!cgroup_reclaim(sc))
2392 __count_vm_events(PGREFILL, nr_scanned);
Yafang Shao2fa26902019-05-13 17:23:02 -07002393 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002394
Alex Shi6168d0d2020-12-15 12:34:29 -08002395 spin_unlock_irq(&lruvec->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 while (!list_empty(&l_hold)) {
2398 cond_resched();
2399 page = lru_to_page(&l_hold);
2400 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002401
Hugh Dickins39b5f292012-10-08 16:33:18 -07002402 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002403 putback_lru_page(page);
2404 continue;
2405 }
2406
Mel Gormancc715d92012-03-21 16:34:00 -07002407 if (unlikely(buffer_heads_over_limit)) {
2408 if (page_has_private(page) && trylock_page(page)) {
2409 if (page_has_private(page))
2410 try_to_release_page(page, 0);
2411 unlock_page(page);
2412 }
2413 }
2414
Liujie Xie6f98ceb2022-02-23 10:32:01 +08002415 trace_android_vh_page_referenced_check_bypass(page, nr_to_scan, lru, &bypass);
2416 if (bypass)
2417 goto skip_page_referenced;
2418
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002419 if (page_referenced(page, 0, sc->target_mem_cgroup,
2420 &vm_flags)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002421 /*
2422 * Identify referenced, file-backed active pages and
2423 * give them one more trip around the active list. So
2424 * that executable code get better chances to stay in
2425 * memory under moderate memory pressure. Anon pages
2426 * are not likely to be evicted by use-once streaming
2427 * IO, plus JVM can create lots of anon VM_EXEC pages,
2428 * so we ignore them here.
2429 */
Huang Ying9de4f222020-04-06 20:04:41 -07002430 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07002431 nr_rotated += thp_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002432 list_add(&page->lru, &l_active);
2433 continue;
2434 }
2435 }
Liujie Xie6f98ceb2022-02-23 10:32:01 +08002436skip_page_referenced:
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002437 ClearPageActive(page); /* we are de-activating */
Johannes Weiner1899ad12018-10-26 15:06:04 -07002438 SetPageWorkingset(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 list_add(&page->lru, &l_inactive);
2440 }
2441
Andrew Mortonb5557492009-01-06 14:40:13 -08002442 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002443 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002444 */
Alex Shi6168d0d2020-12-15 12:34:29 -08002445 spin_lock_irq(&lruvec->lru_lock);
Rik van Riel556adec2008-10-18 20:26:34 -07002446
Kirill Tkhaia222f342019-05-13 17:17:00 -07002447 nr_activate = move_pages_to_lru(lruvec, &l_active);
2448 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
Kirill Tkhaif372d892019-05-13 17:16:57 -07002449 /* Keep all free pages in l_active list */
2450 list_splice(&l_inactive, &l_active);
Kirill Tkhai9851ac12019-05-13 17:16:54 -07002451
2452 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2453 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2454
Mel Gorman599d0c92016-07-28 15:45:31 -07002455 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Alex Shi6168d0d2020-12-15 12:34:29 -08002456 spin_unlock_irq(&lruvec->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002457
Kirill Tkhaif372d892019-05-13 17:16:57 -07002458 mem_cgroup_uncharge_list(&l_active);
2459 free_unref_page_list(&l_active);
Michal Hocko9d998b42017-02-22 15:44:18 -08002460 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2461 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462}
2463
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002464unsigned long reclaim_pages(struct list_head *page_list)
2465{
Yang Shif661d002020-04-01 21:10:05 -07002466 int nid = NUMA_NO_NODE;
Maninder Singh730ec8c2020-06-03 16:01:18 -07002467 unsigned int nr_reclaimed = 0;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002468 LIST_HEAD(node_page_list);
2469 struct reclaim_stat dummy_stat;
2470 struct page *page;
Yu Zhao2d2b8d22021-06-30 18:49:48 -07002471 unsigned int noreclaim_flag;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002472 struct scan_control sc = {
2473 .gfp_mask = GFP_KERNEL,
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002474 .may_writepage = 1,
2475 .may_unmap = 1,
2476 .may_swap = 1,
Dave Hansen26aa2d12021-09-02 14:59:16 -07002477 .no_demotion = 1,
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002478 };
2479
Yu Zhao2d2b8d22021-06-30 18:49:48 -07002480 noreclaim_flag = memalloc_noreclaim_save();
2481
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002482 while (!list_empty(page_list)) {
2483 page = lru_to_page(page_list);
Yang Shif661d002020-04-01 21:10:05 -07002484 if (nid == NUMA_NO_NODE) {
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002485 nid = page_to_nid(page);
2486 INIT_LIST_HEAD(&node_page_list);
2487 }
2488
2489 if (nid == page_to_nid(page)) {
2490 ClearPageActive(page);
2491 list_move(&page->lru, &node_page_list);
2492 continue;
2493 }
2494
2495 nr_reclaimed += shrink_page_list(&node_page_list,
2496 NODE_DATA(nid),
Shakeel Butt013339d2020-12-14 19:06:39 -08002497 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002498 while (!list_empty(&node_page_list)) {
2499 page = lru_to_page(&node_page_list);
2500 list_del(&page->lru);
2501 putback_lru_page(page);
2502 }
2503
Yang Shif661d002020-04-01 21:10:05 -07002504 nid = NUMA_NO_NODE;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002505 }
2506
2507 if (!list_empty(&node_page_list)) {
2508 nr_reclaimed += shrink_page_list(&node_page_list,
2509 NODE_DATA(nid),
Shakeel Butt013339d2020-12-14 19:06:39 -08002510 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002511 while (!list_empty(&node_page_list)) {
2512 page = lru_to_page(&node_page_list);
2513 list_del(&page->lru);
2514 putback_lru_page(page);
2515 }
2516 }
2517
Yu Zhao2d2b8d22021-06-30 18:49:48 -07002518 memalloc_noreclaim_restore(noreclaim_flag);
2519
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002520 return nr_reclaimed;
2521}
2522
Johannes Weinerb91ac372019-11-30 17:56:02 -08002523static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2524 struct lruvec *lruvec, struct scan_control *sc)
2525{
2526 if (is_active_lru(lru)) {
2527 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2528 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2529 else
2530 sc->skipped_deactivate = 1;
2531 return 0;
2532 }
2533
2534 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2535}
2536
Rik van Riel59dc76b2016-05-20 16:56:31 -07002537/*
2538 * The inactive anon list should be small enough that the VM never has
2539 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002540 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002541 * The inactive file list should be small enough to leave most memory
2542 * to the established workingset on the scan-resistant active list,
2543 * but large enough to avoid thrashing the aggregate readahead window.
2544 *
2545 * Both inactive lists should also be large enough that each inactive
2546 * page has a chance to be referenced again before it is reclaimed.
2547 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002548 * If that fails and refaulting is observed, the inactive list grows.
2549 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002550 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
Andrey Ryabinin3a50d142017-11-15 17:34:15 -08002551 * on this LRU, maintained by the pageout code. An inactive_ratio
Rik van Riel59dc76b2016-05-20 16:56:31 -07002552 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2553 *
2554 * total target max
2555 * memory ratio inactive
2556 * -------------------------------------
2557 * 10MB 1 5MB
2558 * 100MB 1 50MB
2559 * 1GB 3 250MB
2560 * 10GB 10 0.9GB
2561 * 100GB 31 3GB
2562 * 1TB 101 10GB
2563 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002564 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002565static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002566{
Johannes Weinerb91ac372019-11-30 17:56:02 -08002567 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002568 unsigned long inactive, active;
2569 unsigned long inactive_ratio;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002570 unsigned long gb;
2571
Johannes Weinerb91ac372019-11-30 17:56:02 -08002572 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2573 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002574
Johannes Weinerb91ac372019-11-30 17:56:02 -08002575 gb = (inactive + active) >> (30 - PAGE_SHIFT);
Joonsoo Kim40025702020-08-11 18:30:54 -07002576 if (gb)
Johannes Weinerb91ac372019-11-30 17:56:02 -08002577 inactive_ratio = int_sqrt(10 * gb);
2578 else
2579 inactive_ratio = 1;
Michal Hockofd538802017-02-22 15:45:58 -08002580
Rik van Riel59dc76b2016-05-20 16:56:31 -07002581 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002582}
2583
Johannes Weiner9a265112013-02-22 16:32:17 -08002584enum scan_balance {
2585 SCAN_EQUAL,
2586 SCAN_FRACT,
2587 SCAN_ANON,
2588 SCAN_FILE,
2589};
2590
Yu Zhao3eb07702020-09-27 20:49:08 -06002591static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2592{
2593 unsigned long file;
2594 struct lruvec *target_lruvec;
2595
Yu Zhaoa1537a62022-01-27 20:32:37 -07002596 if (lru_gen_enabled())
2597 return;
2598
Yu Zhao3eb07702020-09-27 20:49:08 -06002599 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2600
2601 /*
2602 * Flush the memory cgroup stats, so that we read accurate per-memcg
2603 * lruvec stats for heuristics.
2604 */
2605 mem_cgroup_flush_stats();
2606
2607 /*
2608 * Determine the scan balance between anon and file LRUs.
2609 */
2610 spin_lock_irq(&target_lruvec->lru_lock);
2611 sc->anon_cost = target_lruvec->anon_cost;
2612 sc->file_cost = target_lruvec->file_cost;
2613 spin_unlock_irq(&target_lruvec->lru_lock);
2614
2615 /*
2616 * Target desirable inactive:active list ratios for the anon
2617 * and file LRU lists.
2618 */
2619 if (!sc->force_deactivate) {
2620 unsigned long refaults;
2621
2622 refaults = lruvec_page_state(target_lruvec,
2623 WORKINGSET_ACTIVATE_ANON);
2624 if (refaults != target_lruvec->refaults[0] ||
2625 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2626 sc->may_deactivate |= DEACTIVATE_ANON;
2627 else
2628 sc->may_deactivate &= ~DEACTIVATE_ANON;
2629
2630 /*
2631 * When refaults are being observed, it means a new
2632 * workingset is being established. Deactivate to get
2633 * rid of any stale active pages quickly.
2634 */
2635 refaults = lruvec_page_state(target_lruvec,
2636 WORKINGSET_ACTIVATE_FILE);
2637 if (refaults != target_lruvec->refaults[1] ||
2638 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2639 sc->may_deactivate |= DEACTIVATE_FILE;
2640 else
2641 sc->may_deactivate &= ~DEACTIVATE_FILE;
2642 } else
2643 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2644
2645 /*
2646 * If we have plenty of inactive file pages that aren't
2647 * thrashing, try to reclaim those first before touching
2648 * anonymous pages.
2649 */
2650 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2651 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2652 sc->cache_trim_mode = 1;
2653 else
2654 sc->cache_trim_mode = 0;
2655
2656 /*
2657 * Prevent the reclaimer from falling into the cache trap: as
2658 * cache pages start out inactive, every cache fault will tip
2659 * the scan balance towards the file LRU. And as the file LRU
2660 * shrinks, so does the window for rotation from references.
2661 * This means we have a runaway feedback loop where a tiny
2662 * thrashing file LRU becomes infinitely more attractive than
2663 * anon pages. Try to detect this based on file LRU size.
2664 */
2665 if (!cgroup_reclaim(sc)) {
2666 unsigned long total_high_wmark = 0;
2667 unsigned long free, anon;
2668 int z;
2669
2670 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2671 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2672 node_page_state(pgdat, NR_INACTIVE_FILE);
2673
2674 for (z = 0; z < MAX_NR_ZONES; z++) {
2675 struct zone *zone = &pgdat->node_zones[z];
2676
2677 if (!managed_zone(zone))
2678 continue;
2679
2680 total_high_wmark += high_wmark_pages(zone);
2681 }
2682
2683 /*
2684 * Consider anon: if that's low too, this isn't a
2685 * runaway file reclaim problem, but rather just
2686 * extreme pressure. Reclaim as per usual then.
2687 */
2688 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2689
2690 sc->file_is_tiny =
2691 file + free <= total_high_wmark &&
2692 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2693 anon >> sc->priority;
2694 }
2695}
2696
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002698 * Determine how aggressively the anon and file LRU lists should be
2699 * scanned. The relative value of each set of LRU lists is determined
2700 * by looking at the fraction of the pages scanned we did rotate back
2701 * onto the active list instead of evict.
2702 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002703 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2704 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002705 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002706static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2707 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002708{
Keith Buscha2a36482021-09-02 14:59:26 -07002709 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002710 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002711 unsigned long anon_cost, file_cost, total_cost;
Vladimir Davydov33377672016-01-20 15:02:59 -08002712 int swappiness = mem_cgroup_swappiness(memcg);
Yu Zhaoed017372020-10-15 20:09:55 -07002713 u64 fraction[ANON_AND_FILE];
Johannes Weiner9a265112013-02-22 16:32:17 -08002714 u64 denominator = 0; /* gcc */
Johannes Weiner9a265112013-02-22 16:32:17 -08002715 enum scan_balance scan_balance;
Johannes Weiner9a265112013-02-22 16:32:17 -08002716 unsigned long ap, fp;
2717 enum lru_list lru;
Sudarshan Rajagopalan55b46b62021-06-01 04:05:01 -07002718 bool balance_anon_file_reclaim = false;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002719
2720 /* If we have no swap space, do not bother scanning anon pages. */
Keith Buscha2a36482021-09-02 14:59:26 -07002721 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002722 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002723 goto out;
2724 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002725
Johannes Weiner10316b32013-02-22 16:32:14 -08002726 /*
2727 * Global reclaim will swap to prevent OOM even with no
2728 * swappiness, but memcg users want to use this knob to
2729 * disable swapping for individual groups completely when
2730 * using the memory controller's swap limit feature would be
2731 * too expensive.
2732 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08002733 if (cgroup_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002734 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002735 goto out;
2736 }
2737
2738 /*
2739 * Do not apply any pressure balancing cleverness when the
2740 * system is close to OOM, scan both anon and file equally
2741 * (unless the swappiness setting disagrees with swapping).
2742 */
Johannes Weiner02695172014-08-06 16:06:17 -07002743 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002744 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002745 goto out;
2746 }
2747
Johannes Weiner11d16c22013-02-22 16:32:15 -08002748 /*
Johannes Weiner53138ce2019-11-30 17:55:56 -08002749 * If the system is almost out of file pages, force-scan anon.
Johannes Weiner62376252014-05-06 12:50:07 -07002750 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002751 if (sc->file_is_tiny) {
Johannes Weiner53138ce2019-11-30 17:55:56 -08002752 scan_balance = SCAN_ANON;
2753 goto out;
Johannes Weiner62376252014-05-06 12:50:07 -07002754 }
2755
Sudarshan Rajagopalan55b46b62021-06-01 04:05:01 -07002756 trace_android_rvh_set_balance_anon_file_reclaim(&balance_anon_file_reclaim);
2757
Johannes Weiner62376252014-05-06 12:50:07 -07002758 /*
Johannes Weinerb91ac372019-11-30 17:56:02 -08002759 * If there is enough inactive page cache, we do not reclaim
Sudarshan Rajagopalan55b46b62021-06-01 04:05:01 -07002760 * anything from the anonymous working right now. But when balancing
2761 * anon and page cache files for reclaim, allow swapping of anon pages
2762 * even if there are a number of inactive file cache pages.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002763 */
Sudarshan Rajagopalan55b46b62021-06-01 04:05:01 -07002764 if (!balance_anon_file_reclaim && sc->cache_trim_mode) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002765 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002766 goto out;
2767 }
2768
Johannes Weiner9a265112013-02-22 16:32:17 -08002769 scan_balance = SCAN_FRACT;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002770 /*
Johannes Weiner314b57f2020-06-03 16:03:03 -07002771 * Calculate the pressure balance between anon and file pages.
2772 *
2773 * The amount of pressure we put on each LRU is inversely
2774 * proportional to the cost of reclaiming each list, as
2775 * determined by the share of pages that are refaulting, times
2776 * the relative IO cost of bringing back a swapped out
2777 * anonymous page vs reloading a filesystem page (swappiness).
2778 *
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002779 * Although we limit that influence to ensure no list gets
2780 * left behind completely: at least a third of the pressure is
2781 * applied, before swappiness.
2782 *
Johannes Weiner314b57f2020-06-03 16:03:03 -07002783 * With swappiness at 100, anon and file have equal IO cost.
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002784 */
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002785 total_cost = sc->anon_cost + sc->file_cost;
2786 anon_cost = total_cost + sc->anon_cost;
2787 file_cost = total_cost + sc->file_cost;
2788 total_cost = anon_cost + file_cost;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002789
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002790 ap = swappiness * (total_cost + 1);
2791 ap /= anon_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002792
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002793 fp = (200 - swappiness) * (total_cost + 1);
2794 fp /= file_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002795
Shaohua Li76a33fc2010-05-24 14:32:36 -07002796 fraction[0] = ap;
2797 fraction[1] = fp;
Johannes Weinera4fe1632020-06-03 16:02:50 -07002798 denominator = ap + fp;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002799out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002800 for_each_evictable_lru(lru) {
2801 int file = is_file_lru(lru);
Chris Down9783aa92019-10-06 17:58:32 -07002802 unsigned long lruvec_size;
Johannes Weinerf56ce412021-08-19 19:04:21 -07002803 unsigned long low, min;
Johannes Weiner688035f2017-05-03 14:52:07 -07002804 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002805
Chris Down9783aa92019-10-06 17:58:32 -07002806 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Johannes Weinerf56ce412021-08-19 19:04:21 -07002807 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2808 &min, &low);
Chris Down9783aa92019-10-06 17:58:32 -07002809
Johannes Weinerf56ce412021-08-19 19:04:21 -07002810 if (min || low) {
Chris Down9783aa92019-10-06 17:58:32 -07002811 /*
2812 * Scale a cgroup's reclaim pressure by proportioning
2813 * its current usage to its memory.low or memory.min
2814 * setting.
2815 *
2816 * This is important, as otherwise scanning aggression
2817 * becomes extremely binary -- from nothing as we
2818 * approach the memory protection threshold, to totally
2819 * nominal as we exceed it. This results in requiring
2820 * setting extremely liberal protection thresholds. It
2821 * also means we simply get no protection at all if we
2822 * set it too low, which is not ideal.
Chris Down1bc63fb2019-10-06 17:58:38 -07002823 *
2824 * If there is any protection in place, we reduce scan
2825 * pressure by how much of the total memory used is
2826 * within protection thresholds.
Chris Down9783aa92019-10-06 17:58:32 -07002827 *
Chris Down9de7ca42019-10-06 17:58:35 -07002828 * There is one special case: in the first reclaim pass,
2829 * we skip over all groups that are within their low
2830 * protection. If that fails to reclaim enough pages to
2831 * satisfy the reclaim goal, we come back and override
2832 * the best-effort low protection. However, we still
2833 * ideally want to honor how well-behaved groups are in
2834 * that case instead of simply punishing them all
2835 * equally. As such, we reclaim them based on how much
Chris Down1bc63fb2019-10-06 17:58:38 -07002836 * memory they are using, reducing the scan pressure
2837 * again by how much of the total memory used is under
2838 * hard protection.
Chris Down9783aa92019-10-06 17:58:32 -07002839 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002840 unsigned long cgroup_size = mem_cgroup_size(memcg);
Johannes Weinerf56ce412021-08-19 19:04:21 -07002841 unsigned long protection;
2842
2843 /* memory.low scaling, make sure we retry before OOM */
2844 if (!sc->memcg_low_reclaim && low > min) {
2845 protection = low;
2846 sc->memcg_low_skipped = 1;
2847 } else {
2848 protection = min;
2849 }
Chris Down1bc63fb2019-10-06 17:58:38 -07002850
2851 /* Avoid TOCTOU with earlier protection check */
2852 cgroup_size = max(cgroup_size, protection);
2853
2854 scan = lruvec_size - lruvec_size * protection /
Rik van Riel32d4f4b2021-09-08 18:10:08 -07002855 (cgroup_size + 1);
Chris Down9783aa92019-10-06 17:58:32 -07002856
2857 /*
Chris Down1bc63fb2019-10-06 17:58:38 -07002858 * Minimally target SWAP_CLUSTER_MAX pages to keep
Ethon Paul55b65a52020-06-04 16:49:10 -07002859 * reclaim moving forwards, avoiding decrementing
Chris Down9de7ca42019-10-06 17:58:35 -07002860 * sc->priority further than desirable.
Chris Down9783aa92019-10-06 17:58:32 -07002861 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002862 scan = max(scan, SWAP_CLUSTER_MAX);
Chris Down9783aa92019-10-06 17:58:32 -07002863 } else {
2864 scan = lruvec_size;
2865 }
2866
2867 scan >>= sc->priority;
2868
Johannes Weiner688035f2017-05-03 14:52:07 -07002869 /*
2870 * If the cgroup's already been deleted, make sure to
2871 * scrape out the remaining cache.
2872 */
2873 if (!scan && !mem_cgroup_online(memcg))
Chris Down9783aa92019-10-06 17:58:32 -07002874 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002875
Johannes Weiner688035f2017-05-03 14:52:07 -07002876 switch (scan_balance) {
2877 case SCAN_EQUAL:
2878 /* Scan lists relative to size */
2879 break;
2880 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002881 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002882 * Scan types proportional to swappiness and
2883 * their relative recent reclaim efficiency.
Gavin Shan76073c62020-02-20 20:04:24 -08002884 * Make sure we don't miss the last page on
2885 * the offlined memory cgroups because of a
2886 * round-off error.
Johannes Weiner9a265112013-02-22 16:32:17 -08002887 */
Gavin Shan76073c62020-02-20 20:04:24 -08002888 scan = mem_cgroup_online(memcg) ?
2889 div64_u64(scan * fraction[file], denominator) :
2890 DIV64_U64_ROUND_UP(scan * fraction[file],
Roman Gushchin68600f62018-10-26 15:03:27 -07002891 denominator);
Johannes Weiner688035f2017-05-03 14:52:07 -07002892 break;
2893 case SCAN_FILE:
2894 case SCAN_ANON:
2895 /* Scan one type exclusively */
Mateusz Noseke072bff2020-04-01 21:10:15 -07002896 if ((scan_balance == SCAN_FILE) != file)
Johannes Weiner688035f2017-05-03 14:52:07 -07002897 scan = 0;
Johannes Weiner688035f2017-05-03 14:52:07 -07002898 break;
2899 default:
2900 /* Look ma, no brain */
2901 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002902 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002903
Johannes Weiner688035f2017-05-03 14:52:07 -07002904 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002905 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002906}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002907
Dave Hansen2f368a92021-09-02 14:59:23 -07002908/*
2909 * Anonymous LRU management is a waste if there is
2910 * ultimately no way to reclaim the memory.
2911 */
2912static bool can_age_anon_pages(struct pglist_data *pgdat,
2913 struct scan_control *sc)
2914{
2915 /* Aging the anon LRU is valuable if swap is present: */
2916 if (total_swap_pages > 0)
2917 return true;
2918
2919 /* Also valuable if anon pages can be demoted: */
2920 return can_demote(pgdat->node_id, sc);
2921}
2922
Yu Zhaof88ed5a2021-01-25 21:12:33 -07002923#ifdef CONFIG_LRU_GEN
2924
Yu Zhao76f7f072022-01-27 19:52:09 -07002925#ifdef CONFIG_LRU_GEN_ENABLED
2926DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2927#else
2928DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2929#endif
2930
Yu Zhaof88ed5a2021-01-25 21:12:33 -07002931/******************************************************************************
2932 * shorthand helpers
2933 ******************************************************************************/
2934
Yu Zhaoa1537a62022-01-27 20:32:37 -07002935#define DEFINE_MAX_SEQ(lruvec) \
2936 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2937
2938#define DEFINE_MIN_SEQ(lruvec) \
2939 unsigned long min_seq[ANON_AND_FILE] = { \
2940 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2941 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2942 }
2943
Yu Zhaof88ed5a2021-01-25 21:12:33 -07002944#define for_each_gen_type_zone(gen, type, zone) \
2945 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2946 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2947 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2948
Yu Zhaoa1537a62022-01-27 20:32:37 -07002949static int page_lru_gen(struct page *page)
2950{
2951 unsigned long flags = READ_ONCE(page->flags);
2952
2953 return ((flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
2954}
2955
2956static int page_lru_tier(struct page *page)
2957{
2958 int refs;
2959 unsigned long flags = READ_ONCE(page->flags);
2960
2961 refs = (flags & LRU_REFS_FLAGS) == LRU_REFS_FLAGS ?
2962 ((flags & LRU_REFS_MASK) >> LRU_REFS_PGOFF) + 1 : 0;
2963
2964 return lru_tier_from_refs(refs);
2965}
2966
Yu Zhao76f7f072022-01-27 19:52:09 -07002967static bool get_cap(int cap)
2968{
2969#ifdef CONFIG_LRU_GEN_ENABLED
2970 return static_branch_likely(&lru_gen_caps[cap]);
2971#else
2972 return static_branch_unlikely(&lru_gen_caps[cap]);
2973#endif
2974}
2975
Yu Zhaof88ed5a2021-01-25 21:12:33 -07002976static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2977{
2978 struct pglist_data *pgdat = NODE_DATA(nid);
2979
2980#ifdef CONFIG_MEMCG
2981 if (memcg) {
2982 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2983
2984 /* for hotadd_new_pgdat() */
2985 if (!lruvec->pgdat)
2986 lruvec->pgdat = pgdat;
2987
2988 return lruvec;
2989 }
2990#endif
2991 VM_BUG_ON(!mem_cgroup_disabled());
2992
2993 return pgdat ? &pgdat->__lruvec : NULL;
2994}
2995
Yu Zhaoa1537a62022-01-27 20:32:37 -07002996static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2997{
2998 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2999 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3000
3001 if (!can_demote(pgdat->node_id, sc) &&
3002 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3003 return 0;
3004
3005 return mem_cgroup_swappiness(memcg);
3006}
3007
3008static int get_nr_gens(struct lruvec *lruvec, int type)
3009{
3010 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3011}
3012
3013static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3014{
3015 /* see the comment on lru_gen_struct */
3016 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3017 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3018 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3019}
3020
3021/******************************************************************************
Yu Zhao5280d762022-01-27 19:23:01 -07003022 * mm_struct list
3023 ******************************************************************************/
3024
3025static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3026{
3027 static struct lru_gen_mm_list mm_list = {
3028 .fifo = LIST_HEAD_INIT(mm_list.fifo),
3029 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3030 };
3031
3032#ifdef CONFIG_MEMCG
3033 if (memcg)
3034 return &memcg->mm_list;
3035#endif
3036 VM_BUG_ON(!mem_cgroup_disabled());
3037
3038 return &mm_list;
3039}
3040
3041void lru_gen_add_mm(struct mm_struct *mm)
3042{
3043 int nid;
3044 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3045 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3046
3047 VM_BUG_ON_MM(!list_empty(&mm->lru_gen.list), mm);
3048#ifdef CONFIG_MEMCG
3049 VM_BUG_ON_MM(mm->lru_gen.memcg, mm);
3050 mm->lru_gen.memcg = memcg;
3051#endif
3052 spin_lock(&mm_list->lock);
3053
3054 for_each_node_state(nid, N_MEMORY) {
3055 struct lruvec *lruvec = get_lruvec(memcg, nid);
3056
3057 if (!lruvec)
3058 continue;
3059
3060 if (lruvec->mm_state.tail == &mm_list->fifo)
3061 lruvec->mm_state.tail = &mm->lru_gen.list;
3062 }
3063
3064 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3065
3066 spin_unlock(&mm_list->lock);
3067}
3068
3069void lru_gen_del_mm(struct mm_struct *mm)
3070{
3071 int nid;
3072 struct lru_gen_mm_list *mm_list;
3073 struct mem_cgroup *memcg = NULL;
3074
3075 if (list_empty(&mm->lru_gen.list))
3076 return;
3077
3078#ifdef CONFIG_MEMCG
3079 memcg = mm->lru_gen.memcg;
3080#endif
3081 mm_list = get_mm_list(memcg);
3082
3083 spin_lock(&mm_list->lock);
3084
3085 for_each_node(nid) {
3086 struct lruvec *lruvec = get_lruvec(memcg, nid);
3087
3088 if (!lruvec)
3089 continue;
3090
3091 if (lruvec->mm_state.tail == &mm->lru_gen.list)
3092 lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3093
3094 if (lruvec->mm_state.head != &mm->lru_gen.list)
3095 continue;
3096
3097 lruvec->mm_state.head = lruvec->mm_state.head->next;
3098 if (lruvec->mm_state.head == &mm_list->fifo)
3099 WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
3100 }
3101
3102 list_del_init(&mm->lru_gen.list);
3103
3104 spin_unlock(&mm_list->lock);
3105
3106#ifdef CONFIG_MEMCG
3107 mem_cgroup_put(mm->lru_gen.memcg);
3108 mm->lru_gen.memcg = NULL;
3109#endif
3110}
3111
3112#ifdef CONFIG_MEMCG
3113void lru_gen_migrate_mm(struct mm_struct *mm)
3114{
3115 struct mem_cgroup *memcg;
3116
3117 lockdep_assert_held(&mm->owner->alloc_lock);
3118
3119 /* for mm_update_next_owner() */
3120 if (mem_cgroup_disabled())
3121 return;
3122
3123 rcu_read_lock();
3124 memcg = mem_cgroup_from_task(mm->owner);
3125 rcu_read_unlock();
3126 if (memcg == mm->lru_gen.memcg)
3127 return;
3128
3129 VM_BUG_ON_MM(!mm->lru_gen.memcg, mm);
3130 VM_BUG_ON_MM(list_empty(&mm->lru_gen.list), mm);
3131
3132 lru_gen_del_mm(mm);
3133 lru_gen_add_mm(mm);
3134}
3135#endif
3136
3137/*
3138 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3139 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3140 * bits in a bitmap, k is the number of hash functions and n is the number of
3141 * inserted items.
3142 *
3143 * Page table walkers use one of the two filters to reduce their search space.
3144 * To get rid of non-leaf entries that no longer have enough leaf entries, the
3145 * aging uses the double-buffering technique to flip to the other filter each
3146 * time it produces a new generation. For non-leaf entries that have enough
3147 * leaf entries, the aging carries them over to the next generation in
3148 * walk_pmd_range(); the eviction also report them when walking the rmap
3149 * in lru_gen_look_around().
3150 *
3151 * For future optimizations:
3152 * 1. It's not necessary to keep both filters all the time. The spare one can be
3153 * freed after the RCU grace period and reallocated if needed again.
3154 * 2. And when reallocating, it's worth scaling its size according to the number
3155 * of inserted entries in the other filter, to reduce the memory overhead on
3156 * small systems and false positives on large systems.
3157 * 3. Jenkins' hash function is an alternative to Knuth's.
3158 */
3159#define BLOOM_FILTER_SHIFT 15
3160
3161static inline int filter_gen_from_seq(unsigned long seq)
3162{
3163 return seq % NR_BLOOM_FILTERS;
3164}
3165
3166static void get_item_key(void *item, int *key)
3167{
3168 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3169
3170 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3171
3172 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3173 key[1] = hash >> BLOOM_FILTER_SHIFT;
3174}
3175
3176static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3177{
3178 unsigned long *filter;
3179 int gen = filter_gen_from_seq(seq);
3180
3181 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3182
3183 filter = lruvec->mm_state.filters[gen];
3184 if (filter) {
3185 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3186 return;
3187 }
3188
3189 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT), GFP_ATOMIC);
3190 WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3191}
3192
3193static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3194{
3195 int key[2];
3196 unsigned long *filter;
3197 int gen = filter_gen_from_seq(seq);
3198
3199 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3200 if (!filter)
3201 return;
3202
3203 get_item_key(item, key);
3204
3205 if (!test_bit(key[0], filter))
3206 set_bit(key[0], filter);
3207 if (!test_bit(key[1], filter))
3208 set_bit(key[1], filter);
3209}
3210
3211static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3212{
3213 int key[2];
3214 unsigned long *filter;
3215 int gen = filter_gen_from_seq(seq);
3216
3217 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3218 if (!filter)
3219 return true;
3220
3221 get_item_key(item, key);
3222
3223 return test_bit(key[0], filter) && test_bit(key[1], filter);
3224}
3225
3226static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3227{
3228 int i;
3229 int hist;
3230
3231 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3232
3233 if (walk) {
3234 hist = lru_hist_from_seq(walk->max_seq);
3235
3236 for (i = 0; i < NR_MM_STATS; i++) {
3237 WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3238 lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3239 walk->mm_stats[i] = 0;
3240 }
3241 }
3242
3243 if (NR_HIST_GENS > 1 && last) {
3244 hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3245
3246 for (i = 0; i < NR_MM_STATS; i++)
3247 WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3248 }
3249}
3250
3251static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3252{
3253 int type;
3254 unsigned long size = 0;
3255 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3256
3257 if (!walk->full_scan && cpumask_empty(mm_cpumask(mm)) &&
3258 !node_isset(pgdat->node_id, mm->lru_gen.nodes))
3259 return true;
3260
3261 node_clear(pgdat->node_id, mm->lru_gen.nodes);
3262
3263 for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3264 size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3265 get_mm_counter(mm, MM_ANONPAGES) +
3266 get_mm_counter(mm, MM_SHMEMPAGES);
3267 }
3268
3269 if (size < MIN_LRU_BATCH)
3270 return true;
3271
3272 if (mm_is_oom_victim(mm))
3273 return true;
3274
3275 return !mmget_not_zero(mm);
3276}
3277
3278static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3279 struct mm_struct **iter)
3280{
3281 bool first = false;
3282 bool last = true;
3283 struct mm_struct *mm = NULL;
3284 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3285 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3286 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3287
3288 /*
3289 * There are four interesting cases for this page table walker:
3290 * 1. It tries to start a new iteration of mm_list with a stale max_seq;
3291 * there is nothing to be done.
3292 * 2. It's the first of the current generation, and it needs to reset
3293 * the Bloom filter for the next generation.
3294 * 3. It reaches the end of mm_list, and it needs to increment
3295 * mm_state->seq; the iteration is done.
3296 * 4. It's the last of the current generation, and it needs to reset the
3297 * mm stats counters for the next generation.
3298 */
3299 if (*iter)
3300 mmput_async(*iter);
3301 else if (walk->max_seq <= READ_ONCE(mm_state->seq))
3302 return false;
3303
3304 spin_lock(&mm_list->lock);
3305
3306 VM_BUG_ON(mm_state->seq + 1 < walk->max_seq);
3307 VM_BUG_ON(*iter && mm_state->seq > walk->max_seq);
3308 VM_BUG_ON(*iter && !mm_state->nr_walkers);
3309
3310 if (walk->max_seq <= mm_state->seq) {
3311 if (!*iter)
3312 last = false;
3313 goto done;
3314 }
3315
3316 if (!mm_state->nr_walkers) {
3317 VM_BUG_ON(mm_state->head && mm_state->head != &mm_list->fifo);
3318
3319 mm_state->head = mm_list->fifo.next;
3320 first = true;
3321 }
3322
3323 while (!mm && mm_state->head != &mm_list->fifo) {
3324 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3325
3326 mm_state->head = mm_state->head->next;
3327
3328 /* full scan for those added after the last iteration */
3329 if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
3330 mm_state->tail = mm_state->head;
3331 walk->full_scan = true;
3332 }
3333
3334 if (should_skip_mm(mm, walk))
3335 mm = NULL;
3336 }
3337
3338 if (mm_state->head == &mm_list->fifo)
3339 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3340done:
3341 if (*iter && !mm)
3342 mm_state->nr_walkers--;
3343 if (!*iter && mm)
3344 mm_state->nr_walkers++;
3345
3346 if (mm_state->nr_walkers)
3347 last = false;
3348
3349 if (mm && first)
3350 reset_bloom_filter(lruvec, walk->max_seq + 1);
3351
3352 if (*iter || last)
3353 reset_mm_stats(lruvec, walk, last);
3354
3355 spin_unlock(&mm_list->lock);
3356
3357 *iter = mm;
3358
3359 return last;
3360}
3361
3362static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3363{
3364 bool success = false;
3365 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3366 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3367 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3368
3369 if (max_seq <= READ_ONCE(mm_state->seq))
3370 return false;
3371
3372 spin_lock(&mm_list->lock);
3373
3374 VM_BUG_ON(mm_state->seq + 1 < max_seq);
3375
3376 if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
3377 VM_BUG_ON(mm_state->head && mm_state->head != &mm_list->fifo);
3378
3379 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3380 reset_mm_stats(lruvec, NULL, true);
3381 success = true;
3382 }
3383
3384 spin_unlock(&mm_list->lock);
3385
3386 return success;
3387}
3388
3389/******************************************************************************
Yu Zhaoa1537a62022-01-27 20:32:37 -07003390 * refault feedback loop
3391 ******************************************************************************/
3392
3393/*
3394 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3395 *
3396 * The P term is refaulted/(evicted+protected) from a tier in the generation
3397 * currently being evicted; the I term is the exponential moving average of the
3398 * P term over the generations previously evicted, using the smoothing factor
3399 * 1/2; the D term isn't supported.
3400 *
3401 * The setpoint (SP) is always the first tier of one type; the process variable
3402 * (PV) is either any tier of the other type or any other tier of the same
3403 * type.
3404 *
3405 * The error is the difference between the SP and the PV; the correction is
3406 * turn off protection when SP>PV or turn on protection when SP<PV.
3407 *
3408 * For future optimizations:
3409 * 1. The D term may discount the other two terms over time so that long-lived
3410 * generations can resist stale information.
3411 */
3412struct ctrl_pos {
3413 unsigned long refaulted;
3414 unsigned long total;
3415 int gain;
3416};
3417
3418static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3419 struct ctrl_pos *pos)
3420{
3421 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3422 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3423
3424 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3425 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3426 pos->total = lrugen->avg_total[type][tier] +
3427 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3428 if (tier)
3429 pos->total += lrugen->protected[hist][type][tier - 1];
3430 pos->gain = gain;
3431}
3432
3433static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3434{
3435 int hist, tier;
3436 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3437 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3438 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3439
3440 lockdep_assert_held(&lruvec->lru_lock);
3441
3442 if (!carryover && !clear)
3443 return;
3444
3445 hist = lru_hist_from_seq(seq);
3446
3447 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3448 if (carryover) {
3449 unsigned long sum;
3450
3451 sum = lrugen->avg_refaulted[type][tier] +
3452 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3453 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3454
3455 sum = lrugen->avg_total[type][tier] +
3456 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3457 if (tier)
3458 sum += lrugen->protected[hist][type][tier - 1];
3459 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3460 }
3461
3462 if (clear) {
3463 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3464 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3465 if (tier)
3466 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3467 }
3468 }
3469}
3470
3471static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3472{
3473 /*
3474 * Return true if the PV has a limited number of refaults or a lower
3475 * refaulted/total than the SP.
3476 */
3477 return pv->refaulted < MIN_LRU_BATCH ||
3478 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3479 (sp->refaulted + 1) * pv->total * pv->gain;
3480}
3481
3482/******************************************************************************
3483 * the aging
3484 ******************************************************************************/
3485
Yu Zhaoafd94c92022-01-27 20:43:22 -07003486static int page_update_gen(struct page *page, int gen)
3487{
3488 unsigned long old_flags, new_flags;
3489
3490 VM_BUG_ON(gen >= MAX_NR_GENS);
3491 VM_BUG_ON(!rcu_read_lock_held());
3492
3493 do {
3494 new_flags = old_flags = READ_ONCE(page->flags);
3495
3496 /* for shrink_page_list() */
3497 if (!(new_flags & LRU_GEN_MASK)) {
3498 new_flags |= BIT(PG_referenced);
3499 continue;
3500 }
3501
3502 new_flags &= ~LRU_GEN_MASK;
3503 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3504 new_flags &= ~(LRU_REFS_MASK | LRU_REFS_FLAGS);
3505 } while (new_flags != old_flags &&
3506 cmpxchg(&page->flags, old_flags, new_flags) != old_flags);
3507
3508 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3509}
3510
Yu Zhaoa1537a62022-01-27 20:32:37 -07003511static int page_inc_gen(struct lruvec *lruvec, struct page *page, bool reclaiming)
3512{
3513 unsigned long old_flags, new_flags;
3514 int type = page_is_file_lru(page);
3515 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3516 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3517
3518 do {
3519 new_flags = old_flags = READ_ONCE(page->flags);
3520 VM_BUG_ON_PAGE(!(new_flags & LRU_GEN_MASK), page);
3521
Yu Zhaoafd94c92022-01-27 20:43:22 -07003522 new_gen = ((new_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3523 /* page_update_gen() has promoted this page? */
3524 if (new_gen >= 0 && new_gen != old_gen)
3525 return new_gen;
3526
Yu Zhaoa1537a62022-01-27 20:32:37 -07003527 new_gen = (old_gen + 1) % MAX_NR_GENS;
3528
3529 new_flags &= ~LRU_GEN_MASK;
3530 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3531 new_flags &= ~(LRU_REFS_MASK | LRU_REFS_FLAGS);
3532 /* for end_page_writeback() */
3533 if (reclaiming)
3534 new_flags |= BIT(PG_reclaim);
3535 } while (cmpxchg(&page->flags, old_flags, new_flags) != old_flags);
3536
3537 lru_gen_update_size(lruvec, page, old_gen, new_gen);
3538
3539 return new_gen;
3540}
3541
Yu Zhao5280d762022-01-27 19:23:01 -07003542static void update_batch_size(struct lru_gen_mm_walk *walk, struct page *page,
3543 int old_gen, int new_gen)
3544{
3545 int type = page_is_file_lru(page);
3546 int zone = page_zonenum(page);
3547 int delta = thp_nr_pages(page);
3548
3549 VM_BUG_ON(old_gen >= MAX_NR_GENS);
3550 VM_BUG_ON(new_gen >= MAX_NR_GENS);
3551
3552 walk->batched++;
3553
3554 walk->nr_pages[old_gen][type][zone] -= delta;
3555 walk->nr_pages[new_gen][type][zone] += delta;
3556}
3557
3558static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3559{
3560 int gen, type, zone;
3561 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3562
3563 walk->batched = 0;
3564
3565 for_each_gen_type_zone(gen, type, zone) {
3566 enum lru_list lru = type * LRU_INACTIVE_FILE;
3567 int delta = walk->nr_pages[gen][type][zone];
3568
3569 if (!delta)
3570 continue;
3571
3572 walk->nr_pages[gen][type][zone] = 0;
3573 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3574 lrugen->nr_pages[gen][type][zone] + delta);
3575
3576 if (lru_gen_is_active(lruvec, gen))
3577 lru += LRU_ACTIVE;
3578 __update_lru_size(lruvec, lru, zone, delta);
3579 }
3580}
3581
3582static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *walk)
3583{
3584 struct address_space *mapping;
3585 struct vm_area_struct *vma = walk->vma;
3586 struct lru_gen_mm_walk *priv = walk->private;
3587
3588 if (!vma_is_accessible(vma) || is_vm_hugetlb_page(vma) ||
3589 (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ)) ||
3590 vma == get_gate_vma(vma->vm_mm))
3591 return true;
3592
3593 if (vma_is_anonymous(vma))
3594 return !priv->can_swap;
3595
3596 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3597 return true;
3598
3599 mapping = vma->vm_file->f_mapping;
3600 if (mapping_unevictable(mapping))
3601 return true;
3602
3603 /* check readpage to exclude special mappings like dax, etc. */
3604 return shmem_mapping(mapping) ? !priv->can_swap : !mapping->a_ops->readpage;
3605}
3606
3607/*
3608 * Some userspace memory allocators map many single-page VMAs. Instead of
3609 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3610 * table to reduce zigzags and improve cache performance.
3611 */
3612static bool get_next_vma(struct mm_walk *walk, unsigned long mask, unsigned long size,
3613 unsigned long *start, unsigned long *end)
3614{
3615 unsigned long next = round_up(*end, size);
3616
3617 VM_BUG_ON(mask & size);
3618 VM_BUG_ON(*start >= *end);
3619 VM_BUG_ON((next & mask) != (*start & mask));
3620
3621 while (walk->vma) {
3622 if (next >= walk->vma->vm_end) {
3623 walk->vma = walk->vma->vm_next;
3624 continue;
3625 }
3626
3627 if ((next & mask) != (walk->vma->vm_start & mask))
3628 return false;
3629
3630 if (should_skip_vma(walk->vma->vm_start, walk->vma->vm_end, walk)) {
3631 walk->vma = walk->vma->vm_next;
3632 continue;
3633 }
3634
3635 *start = max(next, walk->vma->vm_start);
3636 next = (next | ~mask) + 1;
3637 /* rounded-up boundaries can wrap to 0 */
3638 *end = next && next < walk->vma->vm_end ? next : walk->vma->vm_end;
3639
3640 return true;
3641 }
3642
3643 return false;
3644}
3645
3646static bool suitable_to_scan(int total, int young)
3647{
3648 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3649
3650 /* suitable if the average number of young PTEs per cacheline is >=1 */
3651 return young * n >= total;
3652}
3653
3654static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3655 struct mm_walk *walk)
3656{
3657 int i;
3658 pte_t *pte;
3659 spinlock_t *ptl;
3660 unsigned long addr;
3661 int total = 0;
3662 int young = 0;
3663 struct lru_gen_mm_walk *priv = walk->private;
3664 struct mem_cgroup *memcg = lruvec_memcg(priv->lruvec);
3665 struct pglist_data *pgdat = lruvec_pgdat(priv->lruvec);
3666 int old_gen, new_gen = lru_gen_from_seq(priv->max_seq);
3667
3668 VM_BUG_ON(pmd_leaf(*pmd));
3669
3670 ptl = pte_lockptr(walk->mm, pmd);
3671 if (!spin_trylock(ptl))
3672 return false;
3673
3674 arch_enter_lazy_mmu_mode();
3675
3676 pte = pte_offset_map(pmd, start & PMD_MASK);
3677restart:
3678 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3679 struct page *page;
3680 unsigned long pfn = pte_pfn(pte[i]);
3681
3682 VM_BUG_ON(addr < walk->vma->vm_start || addr >= walk->vma->vm_end);
3683
3684 total++;
3685 priv->mm_stats[MM_PTE_TOTAL]++;
3686
3687 if (!pte_present(pte[i]) || is_zero_pfn(pfn))
3688 continue;
3689
3690 if (WARN_ON_ONCE(pte_devmap(pte[i]) || pte_special(pte[i])))
3691 continue;
3692
3693 if (!pte_young(pte[i])) {
3694 priv->mm_stats[MM_PTE_OLD]++;
3695 continue;
3696 }
3697
3698 VM_BUG_ON(!pfn_valid(pfn));
3699 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3700 continue;
3701
3702 page = compound_head(pfn_to_page(pfn));
3703 if (page_to_nid(page) != pgdat->node_id)
3704 continue;
3705
3706 if (page_memcg_rcu(page) != memcg)
3707 continue;
3708
3709 if (!ptep_test_and_clear_young(walk->vma, addr, pte + i))
3710 continue;
3711
3712 young++;
3713 priv->mm_stats[MM_PTE_YOUNG]++;
3714
3715 if (pte_dirty(pte[i]) && !PageDirty(page) &&
3716 !(PageAnon(page) && PageSwapBacked(page) && !PageSwapCache(page)))
3717 set_page_dirty(page);
3718
3719 old_gen = page_update_gen(page, new_gen);
3720 if (old_gen >= 0 && old_gen != new_gen)
3721 update_batch_size(priv, page, old_gen, new_gen);
3722 }
3723
3724 if (i < PTRS_PER_PTE && get_next_vma(walk, PMD_MASK, PAGE_SIZE, &start, &end))
3725 goto restart;
3726
3727 pte_unmap(pte);
3728
3729 arch_leave_lazy_mmu_mode();
3730 spin_unlock(ptl);
3731
3732 return suitable_to_scan(total, young);
3733}
3734
3735#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3736static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3737 struct mm_walk *walk, unsigned long *start)
3738{
3739 int i;
3740 pmd_t *pmd;
3741 spinlock_t *ptl;
3742 struct lru_gen_mm_walk *priv = walk->private;
3743 struct mem_cgroup *memcg = lruvec_memcg(priv->lruvec);
3744 struct pglist_data *pgdat = lruvec_pgdat(priv->lruvec);
3745 int old_gen, new_gen = lru_gen_from_seq(priv->max_seq);
3746
3747 VM_BUG_ON(pud_leaf(*pud));
3748
3749 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3750 if (*start == -1) {
3751 *start = next;
3752 return;
3753 }
3754
3755 i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start);
3756 if (i && i <= MIN_LRU_BATCH) {
3757 __set_bit(i - 1, priv->bitmap);
3758 return;
3759 }
3760
3761 pmd = pmd_offset(pud, *start);
3762
3763 ptl = pmd_lockptr(walk->mm, pmd);
3764 if (!spin_trylock(ptl))
3765 goto done;
3766
3767 arch_enter_lazy_mmu_mode();
3768
3769 do {
3770 struct page *page;
3771 unsigned long pfn = pmd_pfn(pmd[i]);
3772 unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start;
3773
3774 VM_BUG_ON(addr < vma->vm_start || addr >= vma->vm_end);
3775
3776 if (!pmd_present(pmd[i]) || is_huge_zero_pmd(pmd[i]))
3777 goto next;
3778
3779 if (WARN_ON_ONCE(pmd_devmap(pmd[i])))
3780 goto next;
3781
3782 if (!pmd_trans_huge(pmd[i])) {
Yu Zhao76f7f072022-01-27 19:52:09 -07003783 if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) &&
3784 get_cap(LRU_GEN_NONLEAF_YOUNG))
Yu Zhao5280d762022-01-27 19:23:01 -07003785 pmdp_test_and_clear_young(vma, addr, pmd + i);
3786 goto next;
3787 }
3788
3789 VM_BUG_ON(!pfn_valid(pfn));
3790 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3791 goto next;
3792
3793 page = pfn_to_page(pfn);
3794 VM_BUG_ON_PAGE(PageTail(page), page);
3795 if (page_to_nid(page) != pgdat->node_id)
3796 goto next;
3797
3798 if (page_memcg_rcu(page) != memcg)
3799 goto next;
3800
3801 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3802 goto next;
3803
3804 priv->mm_stats[MM_PTE_YOUNG]++;
3805
3806 if (pmd_dirty(pmd[i]) && !PageDirty(page) &&
3807 !(PageAnon(page) && PageSwapBacked(page) && !PageSwapCache(page)))
3808 set_page_dirty(page);
3809
3810 old_gen = page_update_gen(page, new_gen);
3811 if (old_gen >= 0 && old_gen != new_gen)
3812 update_batch_size(priv, page, old_gen, new_gen);
3813next:
3814 i = i > MIN_LRU_BATCH ? 0 :
3815 find_next_bit(priv->bitmap, MIN_LRU_BATCH, i) + 1;
3816 } while (i <= MIN_LRU_BATCH);
3817
3818 arch_leave_lazy_mmu_mode();
3819 spin_unlock(ptl);
3820done:
3821 *start = -1;
3822 bitmap_zero(priv->bitmap, MIN_LRU_BATCH);
3823}
3824#else
3825static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3826 struct mm_walk *walk, unsigned long *start)
3827{
3828}
3829#endif
3830
3831static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3832 struct mm_walk *walk)
3833{
3834 int i;
3835 pmd_t *pmd;
3836 unsigned long next;
3837 unsigned long addr;
3838 struct vm_area_struct *vma;
3839 unsigned long pos = -1;
3840 struct lru_gen_mm_walk *priv = walk->private;
3841
3842 VM_BUG_ON(pud_leaf(*pud));
3843
3844 /*
3845 * Finish an entire PMD in two passes: the first only reaches to PTE
3846 * tables to avoid taking the PMD lock; the second, if necessary, takes
3847 * the PMD lock to clear the accessed bit in PMD entries.
3848 */
3849 pmd = pmd_offset(pud, start & PUD_MASK);
3850restart:
3851 /* walk_pte_range() may call get_next_vma() */
3852 vma = walk->vma;
3853 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3854 pmd_t val = pmd_read_atomic(pmd + i);
3855
3856 /* for pmd_read_atomic() */
3857 barrier();
3858
3859 next = pmd_addr_end(addr, end);
3860
3861 if (!pmd_present(val)) {
3862 priv->mm_stats[MM_PTE_TOTAL]++;
3863 continue;
3864 }
3865
3866#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3867 if (pmd_trans_huge(val)) {
3868 unsigned long pfn = pmd_pfn(val);
3869 struct pglist_data *pgdat = lruvec_pgdat(priv->lruvec);
3870
3871 priv->mm_stats[MM_PTE_TOTAL]++;
3872
3873 if (is_huge_zero_pmd(val))
3874 continue;
3875
3876 if (!pmd_young(val)) {
3877 priv->mm_stats[MM_PTE_OLD]++;
3878 continue;
3879 }
3880
3881 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3882 continue;
3883
3884 walk_pmd_range_locked(pud, addr, vma, walk, &pos);
3885 continue;
3886 }
3887#endif
3888 priv->mm_stats[MM_PMD_TOTAL]++;
3889
3890#ifdef CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG
Yu Zhao76f7f072022-01-27 19:52:09 -07003891 if (get_cap(LRU_GEN_NONLEAF_YOUNG)) {
3892 if (!pmd_young(val))
3893 continue;
Yu Zhao5280d762022-01-27 19:23:01 -07003894
Yu Zhao76f7f072022-01-27 19:52:09 -07003895 walk_pmd_range_locked(pud, addr, vma, walk, &pos);
3896 }
Yu Zhao5280d762022-01-27 19:23:01 -07003897#endif
3898 if (!priv->full_scan && !test_bloom_filter(priv->lruvec, priv->max_seq, pmd + i))
3899 continue;
3900
3901 priv->mm_stats[MM_PMD_FOUND]++;
3902
3903 if (!walk_pte_range(&val, addr, next, walk))
3904 continue;
3905
3906 priv->mm_stats[MM_PMD_ADDED]++;
3907
3908 /* carry over to the next generation */
3909 update_bloom_filter(priv->lruvec, priv->max_seq + 1, pmd + i);
3910 }
3911
3912 walk_pmd_range_locked(pud, -1, vma, walk, &pos);
3913
3914 if (i < PTRS_PER_PMD && get_next_vma(walk, PUD_MASK, PMD_SIZE, &start, &end))
3915 goto restart;
3916}
3917
3918static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3919 struct mm_walk *walk)
3920{
3921 int i;
3922 pud_t *pud;
3923 unsigned long addr;
3924 unsigned long next;
3925 struct lru_gen_mm_walk *priv = walk->private;
3926
3927 VM_BUG_ON(p4d_leaf(*p4d));
3928
3929 pud = pud_offset(p4d, start & P4D_MASK);
3930restart:
3931 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3932 pud_t val = READ_ONCE(pud[i]);
3933
3934 next = pud_addr_end(addr, end);
3935
3936 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3937 continue;
3938
3939 walk_pmd_range(&val, addr, next, walk);
3940
3941 if (priv->batched >= MAX_LRU_BATCH) {
3942 end = (addr | ~PUD_MASK) + 1;
3943 goto done;
3944 }
3945 }
3946
3947 if (i < PTRS_PER_PUD && get_next_vma(walk, P4D_MASK, PUD_SIZE, &start, &end))
3948 goto restart;
3949
3950 end = round_up(end, P4D_SIZE);
3951done:
3952 /* rounded-up boundaries can wrap to 0 */
3953 priv->next_addr = end && walk->vma ? max(end, walk->vma->vm_start) : 0;
3954
3955 return -EAGAIN;
3956}
3957
3958static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3959{
3960 static const struct mm_walk_ops mm_walk_ops = {
3961 .test_walk = should_skip_vma,
3962 .p4d_entry = walk_pud_range,
3963 };
3964
3965 int err;
3966 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3967
3968 walk->next_addr = FIRST_USER_ADDRESS;
3969
3970 do {
3971 err = -EBUSY;
3972
3973 /* page_update_gen() requires stable page_memcg() */
3974 if (!mem_cgroup_trylock_pages(memcg))
3975 break;
3976
3977 /* the caller might be holding the lock for write */
3978 if (mmap_read_trylock(mm)) {
3979 unsigned long start = walk->next_addr;
3980 unsigned long end = mm->highest_vm_end;
3981
3982 err = walk_page_range(mm, start, end, &mm_walk_ops, walk);
3983
3984 mmap_read_unlock(mm);
3985
3986 if (walk->batched) {
3987 spin_lock_irq(&lruvec->lru_lock);
3988 reset_batch_size(lruvec, walk);
3989 spin_unlock_irq(&lruvec->lru_lock);
3990 }
3991 }
3992
3993 mem_cgroup_unlock_pages();
3994
3995 cond_resched();
3996 } while (err == -EAGAIN && walk->next_addr && !mm_is_oom_victim(mm));
3997}
3998
3999static struct lru_gen_mm_walk *alloc_mm_walk(void)
4000{
4001 if (current->reclaim_state && current->reclaim_state->mm_walk)
4002 return current->reclaim_state->mm_walk;
4003
4004 return kzalloc(sizeof(struct lru_gen_mm_walk),
4005 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4006}
4007
4008static void free_mm_walk(struct lru_gen_mm_walk *walk)
4009{
4010 if (!current->reclaim_state || !current->reclaim_state->mm_walk)
4011 kfree(walk);
4012}
4013
Yu Zhaoa1537a62022-01-27 20:32:37 -07004014static void inc_min_seq(struct lruvec *lruvec)
4015{
4016 int type;
4017 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4018
4019 VM_BUG_ON(!seq_is_valid(lruvec));
4020
4021 for (type = 0; type < ANON_AND_FILE; type++) {
4022 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4023 continue;
4024
4025 reset_ctrl_pos(lruvec, type, true);
4026 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4027 }
4028}
4029
4030static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4031{
4032 int gen, type, zone;
4033 bool success = false;
4034 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4035 DEFINE_MIN_SEQ(lruvec);
4036
4037 VM_BUG_ON(!seq_is_valid(lruvec));
4038
4039 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4040 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4041 gen = lru_gen_from_seq(min_seq[type]);
4042
4043 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4044 if (!list_empty(&lrugen->lists[gen][type][zone]))
4045 goto next;
4046 }
4047
4048 min_seq[type]++;
4049 }
4050next:
4051 ;
4052 }
4053
4054 /* see the comment on lru_gen_struct */
4055 if (can_swap) {
4056 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4057 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4058 }
4059
4060 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4061 if (min_seq[type] == lrugen->min_seq[type])
4062 continue;
4063
4064 reset_ctrl_pos(lruvec, type, true);
4065 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4066 success = true;
4067 }
4068
4069 return success;
4070}
4071
Yu Zhao5280d762022-01-27 19:23:01 -07004072static void inc_max_seq(struct lruvec *lruvec)
Yu Zhaoa1537a62022-01-27 20:32:37 -07004073{
4074 int prev, next;
4075 int type, zone;
4076 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4077
4078 spin_lock_irq(&lruvec->lru_lock);
4079
4080 VM_BUG_ON(!seq_is_valid(lruvec));
4081
Yu Zhaoa1537a62022-01-27 20:32:37 -07004082 inc_min_seq(lruvec);
4083
4084 /*
4085 * Update the active/inactive LRU sizes for compatibility. Both sides of
4086 * the current max_seq need to be covered, since max_seq+1 can overlap
4087 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4088 * overlap, cold/hot inversion happens. This can be solved by moving
4089 * pages from min_seq to min_seq+1 but is omitted for simplicity.
4090 */
4091 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4092 next = lru_gen_from_seq(lrugen->max_seq + 1);
4093
4094 for (type = 0; type < ANON_AND_FILE; type++) {
4095 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4096 enum lru_list lru = type * LRU_INACTIVE_FILE;
4097 long delta = lrugen->nr_pages[prev][type][zone] -
4098 lrugen->nr_pages[next][type][zone];
4099
4100 if (!delta)
4101 continue;
4102
4103 __update_lru_size(lruvec, lru, zone, delta);
4104 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4105 }
4106 }
4107
4108 for (type = 0; type < ANON_AND_FILE; type++)
4109 reset_ctrl_pos(lruvec, type, false);
4110
Yu Zhaoe8507812022-01-27 20:08:50 -07004111 WRITE_ONCE(lrugen->timestamps[next], jiffies);
Yu Zhaoa1537a62022-01-27 20:32:37 -07004112 /* make sure preceding modifications appear */
4113 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
Yu Zhao5280d762022-01-27 19:23:01 -07004114
Yu Zhaoa1537a62022-01-27 20:32:37 -07004115 spin_unlock_irq(&lruvec->lru_lock);
4116}
4117
Yu Zhao5280d762022-01-27 19:23:01 -07004118static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4119 struct scan_control *sc, bool can_swap, bool full_scan)
4120{
4121 bool success;
4122 struct lru_gen_mm_walk *walk;
4123 struct mm_struct *mm = NULL;
4124 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4125
4126 VM_BUG_ON(max_seq > READ_ONCE(lrugen->max_seq));
4127
4128 /*
4129 * If the hardware doesn't automatically set the accessed bit, fallback
4130 * to lru_gen_look_around(), which only clears the accessed bit in a
4131 * handful of PTEs. Spreading the work out over a period of time usually
4132 * is less efficient, but it avoids bursty page faults.
4133 */
Yu Zhao76f7f072022-01-27 19:52:09 -07004134 if (!full_scan && (!arch_has_hw_pte_young() || !get_cap(LRU_GEN_MM_WALK))) {
Yu Zhao5280d762022-01-27 19:23:01 -07004135 success = iterate_mm_list_nowalk(lruvec, max_seq);
4136 goto done;
4137 }
4138
4139 walk = alloc_mm_walk();
4140 if (!walk) {
4141 success = iterate_mm_list_nowalk(lruvec, max_seq);
4142 goto done;
4143 }
4144
4145 walk->lruvec = lruvec;
4146 walk->max_seq = max_seq;
4147 walk->can_swap = can_swap;
4148 walk->full_scan = full_scan;
4149
4150 do {
4151 success = iterate_mm_list(lruvec, walk, &mm);
4152 if (mm)
4153 walk_mm(lruvec, mm, walk);
4154
4155 cond_resched();
4156 } while (mm);
4157
4158 free_mm_walk(walk);
4159done:
4160 if (!success) {
4161 if (!current_is_kswapd() && !sc->priority)
4162 wait_event_killable(lruvec->mm_state.wait,
4163 max_seq < READ_ONCE(lrugen->max_seq));
4164
4165 return max_seq < READ_ONCE(lrugen->max_seq);
4166 }
4167
4168 VM_BUG_ON(max_seq != READ_ONCE(lrugen->max_seq));
4169
4170 inc_max_seq(lruvec);
4171 /* either this sees any waiters or they will see updated max_seq */
4172 if (wq_has_sleeper(&lruvec->mm_state.wait))
4173 wake_up_all(&lruvec->mm_state.wait);
4174
4175 wakeup_flusher_threads(WB_REASON_VMSCAN);
4176
4177 return true;
4178}
4179
Yu Zhaoa1537a62022-01-27 20:32:37 -07004180static long get_nr_evictable(struct lruvec *lruvec, unsigned long max_seq,
4181 unsigned long *min_seq, bool can_swap, bool *need_aging)
4182{
4183 int gen, type, zone;
4184 long old = 0;
4185 long young = 0;
4186 long total = 0;
4187 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4188
4189 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4190 unsigned long seq;
4191
4192 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4193 long size = 0;
4194
4195 gen = lru_gen_from_seq(seq);
4196
4197 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4198 size += READ_ONCE(lrugen->nr_pages[gen][type][zone]);
4199
4200 total += size;
4201 if (seq == max_seq)
4202 young += size;
4203 if (seq + MIN_NR_GENS == max_seq)
4204 old += size;
4205 }
4206 }
4207
4208 /*
4209 * The aging and the eviction is a typical producer-consumer model. The
4210 * aging tries to be lazy to reduce the unnecessary overhead. On the
4211 * other hand, the eviction stalls when the number of generations
4212 * reaches MIN_NR_GENS. So ideally, there should be MIN_NR_GENS+1
4213 * generations, hence the first two if's.
4214 *
4215 * In addition, it's ideal to spread pages out evenly, meaning
4216 * 1/(MIN_NR_GENS+1) of the total number of pages for each generation. A
4217 * reasonable range for this average portion would [1/MIN_NR_GENS,
4218 * 1/(MIN_NR_GENS+2)]. From the consumer's POV, the eviction only cares
4219 * about the lower bound of cold pages, i.e., 1/(MIN_NR_GENS+2), whereas
4220 * from the producer's POV, the aging only cares about the upper bound
4221 * of hot pages, i.e., 1/MIN_NR_GENS.
4222 */
4223 if (min_seq[LRU_GEN_FILE] + MIN_NR_GENS > max_seq)
4224 *need_aging = true;
4225 else if (min_seq[LRU_GEN_FILE] + MIN_NR_GENS < max_seq)
4226 *need_aging = false;
4227 else if (young * MIN_NR_GENS > total)
4228 *need_aging = true;
4229 else if (old * (MIN_NR_GENS + 2) < total)
4230 *need_aging = true;
4231 else
4232 *need_aging = false;
4233
4234 return total > 0 ? total : 0;
4235}
4236
Yu Zhaoe8507812022-01-27 20:08:50 -07004237static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc,
4238 unsigned long min_ttl)
Yu Zhaoa1537a62022-01-27 20:32:37 -07004239{
4240 bool need_aging;
4241 long nr_to_scan;
4242 int swappiness = get_swappiness(lruvec, sc);
4243 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4244 DEFINE_MAX_SEQ(lruvec);
4245 DEFINE_MIN_SEQ(lruvec);
4246
Yu Zhaoe8507812022-01-27 20:08:50 -07004247 if (min_ttl) {
4248 int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4249 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4250
4251 if (time_is_after_jiffies(birth + min_ttl))
4252 return false;
4253 }
4254
Yu Zhaoa1537a62022-01-27 20:32:37 -07004255 mem_cgroup_calculate_protection(NULL, memcg);
4256
4257 if (mem_cgroup_below_min(memcg))
Yu Zhaoe8507812022-01-27 20:08:50 -07004258 return false;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004259
4260 nr_to_scan = get_nr_evictable(lruvec, max_seq, min_seq, swappiness, &need_aging);
4261 if (!nr_to_scan)
Yu Zhaoe8507812022-01-27 20:08:50 -07004262 return false;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004263
4264 nr_to_scan >>= sc->priority;
4265
4266 if (!mem_cgroup_online(memcg))
4267 nr_to_scan++;
4268
4269 if (nr_to_scan && need_aging && (!mem_cgroup_below_low(memcg) || sc->memcg_low_reclaim))
Yu Zhao5280d762022-01-27 19:23:01 -07004270 try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false);
Yu Zhaoe8507812022-01-27 20:08:50 -07004271
4272 return true;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004273}
4274
Yu Zhaoe8507812022-01-27 20:08:50 -07004275/* to protect the working set of the last N jiffies */
4276static unsigned long lru_gen_min_ttl __read_mostly;
4277
Yu Zhaoa1537a62022-01-27 20:32:37 -07004278static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4279{
4280 struct mem_cgroup *memcg;
Yu Zhaoe8507812022-01-27 20:08:50 -07004281 bool success = false;
4282 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
Yu Zhaoa1537a62022-01-27 20:32:37 -07004283
4284 VM_BUG_ON(!current_is_kswapd());
4285
Yu Zhao4d905e92022-01-27 19:59:54 -07004286 /*
4287 * To reduce the chance of going into the aging path or swapping, which
4288 * can be costly, optimistically skip them unless their corresponding
4289 * flags were cleared in the eviction path. This improves the overall
4290 * performance when multiple memcgs are available.
4291 */
4292 if (!sc->memcgs_need_aging) {
4293 sc->memcgs_need_aging = true;
4294 sc->memcgs_avoid_swapping = !sc->memcgs_need_swapping;
4295 sc->memcgs_need_swapping = true;
4296 return;
4297 }
4298
4299 sc->memcgs_need_swapping = true;
4300 sc->memcgs_avoid_swapping = true;
4301
Yu Zhao5280d762022-01-27 19:23:01 -07004302 current->reclaim_state->mm_walk = &pgdat->mm_walk;
4303
Yu Zhaoa1537a62022-01-27 20:32:37 -07004304 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4305 do {
4306 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4307
Yu Zhaoe8507812022-01-27 20:08:50 -07004308 if (age_lruvec(lruvec, sc, min_ttl))
4309 success = true;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004310
4311 cond_resched();
4312 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
Yu Zhao5280d762022-01-27 19:23:01 -07004313
4314 current->reclaim_state->mm_walk = NULL;
Yu Zhaoe8507812022-01-27 20:08:50 -07004315
4316 /*
4317 * The main goal is to OOM kill if every generation from all memcgs is
4318 * younger than min_ttl. However, another theoretical possibility is all
4319 * memcgs are either below min or empty.
4320 */
4321 if (!success && mutex_trylock(&oom_lock)) {
4322 struct oom_control oc = {
4323 .gfp_mask = sc->gfp_mask,
4324 .order = sc->order,
4325 };
4326
4327 out_of_memory(&oc);
4328
4329 mutex_unlock(&oom_lock);
4330 }
Yu Zhaoa1537a62022-01-27 20:32:37 -07004331}
4332
Yu Zhaoafd94c92022-01-27 20:43:22 -07004333/*
4334 * This function exploits spatial locality when shrink_page_list() walks the
4335 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages.
Yu Zhao5280d762022-01-27 19:23:01 -07004336 * If the scan was done cacheline efficiently, it adds the PMD entry pointing
4337 * to the PTE table to the Bloom filter. This process is a feedback loop from
4338 * the eviction to the aging.
Yu Zhaoafd94c92022-01-27 20:43:22 -07004339 */
4340void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4341{
4342 int i;
4343 pte_t *pte;
4344 unsigned long start;
4345 unsigned long end;
4346 unsigned long addr;
4347 struct page *page;
Yu Zhao5280d762022-01-27 19:23:01 -07004348 struct lru_gen_mm_walk *walk;
4349 int young = 0;
Yu Zhaoafd94c92022-01-27 20:43:22 -07004350 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4351 struct mem_cgroup *memcg = page_memcg(pvmw->page);
4352 struct pglist_data *pgdat = page_pgdat(pvmw->page);
4353 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4354 DEFINE_MAX_SEQ(lruvec);
4355 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4356
4357 lockdep_assert_held(pvmw->ptl);
4358 VM_BUG_ON_PAGE(PageLRU(pvmw->page), pvmw->page);
4359
4360 if (spin_is_contended(pvmw->ptl))
4361 return;
4362
4363 start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
4364 end = pmd_addr_end(pvmw->address, pvmw->vma->vm_end);
4365
4366 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4367 if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4368 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4369 else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
4370 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4371 else {
4372 start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
4373 end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
4374 }
4375 }
4376
4377 pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
4378
4379 rcu_read_lock();
4380 arch_enter_lazy_mmu_mode();
4381
4382 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4383 unsigned long pfn = pte_pfn(pte[i]);
4384
4385 VM_BUG_ON(addr < pvmw->vma->vm_start || addr >= pvmw->vma->vm_end);
4386
4387 if (!pte_present(pte[i]) || is_zero_pfn(pfn))
4388 continue;
4389
4390 if (WARN_ON_ONCE(pte_devmap(pte[i]) || pte_special(pte[i])))
4391 continue;
4392
4393 if (!pte_young(pte[i]))
4394 continue;
4395
4396 VM_BUG_ON(!pfn_valid(pfn));
4397 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4398 continue;
4399
4400 page = compound_head(pfn_to_page(pfn));
4401 if (page_to_nid(page) != pgdat->node_id)
4402 continue;
4403
4404 if (page_memcg_rcu(page) != memcg)
4405 continue;
4406
4407 if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
4408 continue;
4409
Yu Zhao5280d762022-01-27 19:23:01 -07004410 young++;
4411
Yu Zhaoafd94c92022-01-27 20:43:22 -07004412 if (pte_dirty(pte[i]) && !PageDirty(page) &&
4413 !(PageAnon(page) && PageSwapBacked(page) && !PageSwapCache(page)))
4414 set_page_dirty(page);
4415
4416 old_gen = page_lru_gen(page);
4417 if (old_gen < 0)
4418 SetPageReferenced(page);
4419 else if (old_gen != new_gen)
4420 __set_bit(i, bitmap);
4421 }
4422
4423 arch_leave_lazy_mmu_mode();
4424 rcu_read_unlock();
4425
Yu Zhao5280d762022-01-27 19:23:01 -07004426 /* feedback from rmap walkers to page table walkers */
4427 if (suitable_to_scan(i, young))
4428 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4429
4430 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4431
4432 if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
Yu Zhaoafd94c92022-01-27 20:43:22 -07004433 for_each_set_bit(i, bitmap, MIN_LRU_BATCH)
4434 activate_page(pte_page(pte[i]));
4435 return;
4436 }
4437
4438 /* page_update_gen() requires stable page_memcg() */
4439 if (!mem_cgroup_trylock_pages(memcg))
4440 return;
4441
Yu Zhao5280d762022-01-27 19:23:01 -07004442 if (!walk) {
4443 spin_lock_irq(&lruvec->lru_lock);
4444 new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
4445 }
Yu Zhaoafd94c92022-01-27 20:43:22 -07004446
4447 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4448 page = compound_head(pte_page(pte[i]));
4449 if (page_memcg_rcu(page) != memcg)
4450 continue;
4451
4452 old_gen = page_update_gen(page, new_gen);
4453 if (old_gen < 0 || old_gen == new_gen)
4454 continue;
4455
Yu Zhao5280d762022-01-27 19:23:01 -07004456 if (walk)
4457 update_batch_size(walk, page, old_gen, new_gen);
4458 else
4459 lru_gen_update_size(lruvec, page, old_gen, new_gen);
Yu Zhaoafd94c92022-01-27 20:43:22 -07004460 }
4461
Yu Zhao5280d762022-01-27 19:23:01 -07004462 if (!walk)
4463 spin_unlock_irq(&lruvec->lru_lock);
Yu Zhaoafd94c92022-01-27 20:43:22 -07004464
4465 mem_cgroup_unlock_pages();
4466}
4467
Yu Zhaoa1537a62022-01-27 20:32:37 -07004468/******************************************************************************
4469 * the eviction
4470 ******************************************************************************/
4471
4472static bool sort_page(struct lruvec *lruvec, struct page *page, int tier_idx)
4473{
4474 bool success;
4475 int gen = page_lru_gen(page);
4476 int type = page_is_file_lru(page);
4477 int zone = page_zonenum(page);
4478 int tier = page_lru_tier(page);
4479 int delta = thp_nr_pages(page);
4480 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4481
4482 VM_BUG_ON_PAGE(gen >= MAX_NR_GENS, page);
4483
4484 if (!page_evictable(page)) {
4485 success = lru_gen_del_page(lruvec, page, true);
4486 VM_BUG_ON_PAGE(!success, page);
4487 SetPageUnevictable(page);
4488 add_page_to_lru_list(page, lruvec);
4489 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4490 return true;
4491 }
4492
4493 if (type == LRU_GEN_FILE && PageAnon(page) && PageDirty(page)) {
4494 success = lru_gen_del_page(lruvec, page, true);
4495 VM_BUG_ON_PAGE(!success, page);
4496 SetPageSwapBacked(page);
4497 add_page_to_lru_list_tail(page, lruvec);
4498 return true;
4499 }
4500
Yu Zhaoafd94c92022-01-27 20:43:22 -07004501 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4502 list_move(&page->lru, &lrugen->lists[gen][type][zone]);
4503 return true;
4504 }
4505
Yu Zhaoa1537a62022-01-27 20:32:37 -07004506 if (tier > tier_idx) {
4507 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4508
4509 gen = page_inc_gen(lruvec, page, false);
4510 list_move_tail(&page->lru, &lrugen->lists[gen][type][zone]);
4511
4512 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4513 lrugen->protected[hist][type][tier - 1] + delta);
4514 __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
4515 return true;
4516 }
4517
4518 if (PageLocked(page) || PageWriteback(page) ||
4519 (type == LRU_GEN_FILE && PageDirty(page))) {
4520 gen = page_inc_gen(lruvec, page, true);
4521 list_move(&page->lru, &lrugen->lists[gen][type][zone]);
4522 return true;
4523 }
4524
4525 return false;
4526}
4527
4528static bool isolate_page(struct lruvec *lruvec, struct page *page, struct scan_control *sc)
4529{
4530 bool success;
4531
4532 if (!sc->may_unmap && page_mapped(page))
4533 return false;
4534
4535 if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) &&
4536 (PageDirty(page) || (PageAnon(page) && !PageSwapCache(page))))
4537 return false;
4538
4539 if (!get_page_unless_zero(page))
4540 return false;
4541
4542 if (!TestClearPageLRU(page)) {
4543 put_page(page);
4544 return false;
4545 }
4546
4547 success = lru_gen_del_page(lruvec, page, true);
4548 VM_BUG_ON_PAGE(!success, page);
4549
4550 return true;
4551}
4552
4553static int scan_pages(struct lruvec *lruvec, struct scan_control *sc,
4554 int type, int tier, struct list_head *list)
4555{
4556 int gen, zone;
4557 enum vm_event_item item;
4558 int sorted = 0;
4559 int scanned = 0;
4560 int isolated = 0;
4561 int remaining = MAX_LRU_BATCH;
4562 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4563 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4564
4565 VM_BUG_ON(!list_empty(list));
4566
4567 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4568 return 0;
4569
4570 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4571
4572 for (zone = sc->reclaim_idx; zone >= 0; zone--) {
4573 LIST_HEAD(moved);
4574 int skipped = 0;
4575 struct list_head *head = &lrugen->lists[gen][type][zone];
4576
4577 while (!list_empty(head)) {
4578 struct page *page = lru_to_page(head);
4579 int delta = thp_nr_pages(page);
4580
4581 VM_BUG_ON_PAGE(PageTail(page), page);
4582 VM_BUG_ON_PAGE(PageUnevictable(page), page);
4583 VM_BUG_ON_PAGE(PageActive(page), page);
4584 VM_BUG_ON_PAGE(page_is_file_lru(page) != type, page);
4585 VM_BUG_ON_PAGE(page_zonenum(page) != zone, page);
4586
4587 prefetchw_prev_lru_page(page, head, flags);
4588
4589 scanned += delta;
4590
4591 if (sort_page(lruvec, page, tier))
4592 sorted += delta;
4593 else if (isolate_page(lruvec, page, sc)) {
4594 list_add(&page->lru, list);
4595 isolated += delta;
4596 } else {
4597 list_move(&page->lru, &moved);
4598 skipped += delta;
4599 }
4600
4601 if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
4602 break;
4603 }
4604
4605 if (skipped) {
4606 list_splice(&moved, head);
4607 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
4608 }
4609
4610 if (!remaining || isolated >= MIN_LRU_BATCH)
4611 break;
4612 }
4613
4614 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
4615 if (!cgroup_reclaim(sc)) {
4616 __count_vm_events(item, isolated);
4617 __count_vm_events(PGREFILL, sorted);
4618 }
4619 __count_memcg_events(memcg, item, isolated);
4620 __count_memcg_events(memcg, PGREFILL, sorted);
4621 __count_vm_events(PGSCAN_ANON + type, isolated);
4622
4623 /*
4624 * There might not be eligible pages due to reclaim_idx, may_unmap and
4625 * may_writepage. Check the remaining to prevent livelock if there is no
4626 * progress.
4627 */
4628 return isolated || !remaining ? scanned : 0;
4629}
4630
4631static int get_tier_idx(struct lruvec *lruvec, int type)
4632{
4633 int tier;
4634 struct ctrl_pos sp, pv;
4635
4636 /*
4637 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4638 * This value is chosen because any other tier would have at least twice
4639 * as many refaults as the first tier.
4640 */
4641 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4642 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4643 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4644 if (!positive_ctrl_err(&sp, &pv))
4645 break;
4646 }
4647
4648 return tier - 1;
4649}
4650
4651static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4652{
4653 int type, tier;
4654 struct ctrl_pos sp, pv;
4655 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4656
4657 /*
4658 * Compare the first tier of anon with that of file to determine which
4659 * type to scan. Also need to compare other tiers of the selected type
4660 * with the first tier of the other type to determine the last tier (of
4661 * the selected type) to evict.
4662 */
4663 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4664 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4665 type = positive_ctrl_err(&sp, &pv);
4666
4667 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4668 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4669 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4670 if (!positive_ctrl_err(&sp, &pv))
4671 break;
4672 }
4673
4674 *tier_idx = tier - 1;
4675
4676 return type;
4677}
4678
4679static int isolate_pages(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4680 int *type_scanned, struct list_head *list)
4681{
4682 int i;
4683 int type;
4684 int scanned;
4685 int tier = -1;
4686 DEFINE_MIN_SEQ(lruvec);
4687
4688 VM_BUG_ON(!seq_is_valid(lruvec));
4689
4690 /*
4691 * Try to make the obvious choice first. When anon and file are both
4692 * available from the same generation, interpret swappiness 1 as file
4693 * first and 200 as anon first.
4694 */
4695 if (!swappiness)
4696 type = LRU_GEN_FILE;
4697 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4698 type = LRU_GEN_ANON;
4699 else if (swappiness == 1)
4700 type = LRU_GEN_FILE;
4701 else if (swappiness == 200)
4702 type = LRU_GEN_ANON;
4703 else
4704 type = get_type_to_scan(lruvec, swappiness, &tier);
4705
4706 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4707 if (tier < 0)
4708 tier = get_tier_idx(lruvec, type);
4709
4710 scanned = scan_pages(lruvec, sc, type, tier, list);
4711 if (scanned)
4712 break;
4713
4714 type = !type;
4715 tier = -1;
4716 }
4717
4718 *type_scanned = type;
4719
4720 return scanned;
4721}
4722
Yu Zhao4d905e92022-01-27 19:59:54 -07004723static int evict_pages(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4724 bool *swapped)
Yu Zhaoa1537a62022-01-27 20:32:37 -07004725{
4726 int type;
4727 int scanned;
4728 int reclaimed;
4729 LIST_HEAD(list);
4730 struct page *page;
4731 enum vm_event_item item;
4732 struct reclaim_stat stat;
Yu Zhao5280d762022-01-27 19:23:01 -07004733 struct lru_gen_mm_walk *walk;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004734 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4735 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4736
4737 spin_lock_irq(&lruvec->lru_lock);
4738
4739 scanned = isolate_pages(lruvec, sc, swappiness, &type, &list);
4740
4741 if (try_to_inc_min_seq(lruvec, swappiness))
4742 scanned++;
4743
4744 if (get_nr_gens(lruvec, LRU_GEN_FILE) == MIN_NR_GENS)
4745 scanned = 0;
4746
4747 spin_unlock_irq(&lruvec->lru_lock);
4748
4749 if (list_empty(&list))
4750 return scanned;
4751
4752 reclaimed = shrink_page_list(&list, pgdat, sc, &stat, false);
4753
4754 /*
4755 * To avoid livelock, don't add rejected pages back to the same lists
4756 * they were isolated from. See lru_gen_add_page().
4757 */
4758 list_for_each_entry(page, &list, lru) {
4759 ClearPageReferenced(page);
4760 ClearPageWorkingset(page);
4761
4762 if (PageReclaim(page) && (PageDirty(page) || PageWriteback(page)))
4763 ClearPageActive(page);
4764 else if (page_is_file_lru(page) || PageSwapCache(page))
4765 SetPageActive(page);
4766 }
4767
4768 spin_lock_irq(&lruvec->lru_lock);
4769
4770 move_pages_to_lru(lruvec, &list);
4771
Yu Zhao5280d762022-01-27 19:23:01 -07004772 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4773 if (walk && walk->batched)
4774 reset_batch_size(lruvec, walk);
4775
Yu Zhaoa1537a62022-01-27 20:32:37 -07004776 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
4777 if (!cgroup_reclaim(sc))
4778 __count_vm_events(item, reclaimed);
4779 __count_memcg_events(memcg, item, reclaimed);
4780 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4781
4782 spin_unlock_irq(&lruvec->lru_lock);
4783
4784 mem_cgroup_uncharge_list(&list);
4785 free_unref_page_list(&list);
4786
4787 sc->nr_reclaimed += reclaimed;
4788
Yu Zhao4d905e92022-01-27 19:59:54 -07004789 if (type == LRU_GEN_ANON && swapped)
4790 *swapped = true;
4791
Yu Zhaoa1537a62022-01-27 20:32:37 -07004792 return scanned;
4793}
4794
4795static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4796{
4797 bool need_aging;
4798 long nr_to_scan;
4799 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4800 DEFINE_MAX_SEQ(lruvec);
4801 DEFINE_MIN_SEQ(lruvec);
4802
4803 if (mem_cgroup_below_min(memcg) ||
4804 (mem_cgroup_below_low(memcg) && !sc->memcg_low_reclaim))
4805 return 0;
4806
4807 nr_to_scan = get_nr_evictable(lruvec, max_seq, min_seq, can_swap, &need_aging);
4808 if (!nr_to_scan)
4809 return 0;
4810
4811 /* reset the priority if the target has been met */
4812 nr_to_scan >>= sc->nr_reclaimed < sc->nr_to_reclaim ? sc->priority : DEF_PRIORITY;
4813
4814 if (!mem_cgroup_online(memcg))
4815 nr_to_scan++;
4816
4817 if (!nr_to_scan)
4818 return 0;
4819
Yu Zhao4d905e92022-01-27 19:59:54 -07004820 if (!need_aging) {
4821 sc->memcgs_need_aging = false;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004822 return nr_to_scan;
Yu Zhao4d905e92022-01-27 19:59:54 -07004823 }
Yu Zhaoa1537a62022-01-27 20:32:37 -07004824
4825 /* leave the work to lru_gen_age_node() */
4826 if (current_is_kswapd())
4827 return 0;
4828
4829 /* try other memcgs before going to the aging path */
4830 if (!cgroup_reclaim(sc) && !sc->force_deactivate) {
4831 sc->skipped_deactivate = true;
4832 return 0;
4833 }
4834
Yu Zhao5280d762022-01-27 19:23:01 -07004835 if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false))
4836 return nr_to_scan;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004837
Yu Zhao5280d762022-01-27 19:23:01 -07004838 return min_seq[LRU_GEN_FILE] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
Yu Zhaoa1537a62022-01-27 20:32:37 -07004839}
4840
4841static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4842{
4843 struct blk_plug plug;
4844 long scanned = 0;
Yu Zhao4d905e92022-01-27 19:59:54 -07004845 bool swapped = false;
4846 unsigned long reclaimed = sc->nr_reclaimed;
Yu Zhao5280d762022-01-27 19:23:01 -07004847 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Yu Zhaoa1537a62022-01-27 20:32:37 -07004848
4849 lru_add_drain();
4850
4851 blk_start_plug(&plug);
4852
Yu Zhao5280d762022-01-27 19:23:01 -07004853 if (current_is_kswapd())
4854 current->reclaim_state->mm_walk = &pgdat->mm_walk;
4855
Yu Zhaoa1537a62022-01-27 20:32:37 -07004856 while (true) {
4857 int delta;
4858 int swappiness;
4859 long nr_to_scan;
4860
4861 if (sc->may_swap)
4862 swappiness = get_swappiness(lruvec, sc);
4863 else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc))
4864 swappiness = 1;
4865 else
4866 swappiness = 0;
4867
4868 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4869 if (!nr_to_scan)
4870 break;
4871
Yu Zhao4d905e92022-01-27 19:59:54 -07004872 delta = evict_pages(lruvec, sc, swappiness, &swapped);
Yu Zhaoa1537a62022-01-27 20:32:37 -07004873 if (!delta)
4874 break;
4875
Yu Zhao4d905e92022-01-27 19:59:54 -07004876 if (sc->memcgs_avoid_swapping && swappiness < 200 && swapped)
Yu Zhaoa1537a62022-01-27 20:32:37 -07004877 break;
4878
Yu Zhao4d905e92022-01-27 19:59:54 -07004879 scanned += delta;
4880 if (scanned >= nr_to_scan) {
4881 if (!swapped && sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH)
4882 sc->memcgs_need_swapping = false;
4883 break;
4884 }
4885
Yu Zhaoa1537a62022-01-27 20:32:37 -07004886 cond_resched();
4887 }
4888
Yu Zhao5280d762022-01-27 19:23:01 -07004889 if (current_is_kswapd())
4890 current->reclaim_state->mm_walk = NULL;
4891
Yu Zhaoa1537a62022-01-27 20:32:37 -07004892 blk_finish_plug(&plug);
4893}
4894
Yu Zhaof88ed5a2021-01-25 21:12:33 -07004895/******************************************************************************
Yu Zhao76f7f072022-01-27 19:52:09 -07004896 * state change
4897 ******************************************************************************/
4898
4899static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4900{
4901 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4902
4903 if (lrugen->enabled) {
4904 enum lru_list lru;
4905
4906 for_each_evictable_lru(lru) {
4907 if (!list_empty(&lruvec->lists[lru]))
4908 return false;
4909 }
4910 } else {
4911 int gen, type, zone;
4912
4913 for_each_gen_type_zone(gen, type, zone) {
4914 if (!list_empty(&lrugen->lists[gen][type][zone]))
4915 return false;
4916
4917 /* unlikely but not a bug when reset_batch_size() is pending */
4918 VM_WARN_ON(lrugen->nr_pages[gen][type][zone]);
4919 }
4920 }
4921
4922 return true;
4923}
4924
4925static bool fill_evictable(struct lruvec *lruvec)
4926{
4927 enum lru_list lru;
4928 int remaining = MAX_LRU_BATCH;
4929
4930 for_each_evictable_lru(lru) {
4931 int type = is_file_lru(lru);
4932 bool active = is_active_lru(lru);
4933 struct list_head *head = &lruvec->lists[lru];
4934
4935 while (!list_empty(head)) {
4936 bool success;
4937 struct page *page = lru_to_page(head);
4938
4939 VM_BUG_ON_PAGE(PageTail(page), page);
4940 VM_BUG_ON_PAGE(PageUnevictable(page), page);
4941 VM_BUG_ON_PAGE(PageActive(page) != active, page);
4942 VM_BUG_ON_PAGE(page_is_file_lru(page) != type, page);
4943 VM_BUG_ON_PAGE(page_lru_gen(page) < MAX_NR_GENS, page);
4944
4945 prefetchw_prev_lru_page(page, head, flags);
4946
4947 del_page_from_lru_list(page, lruvec);
4948 success = lru_gen_add_page(lruvec, page, false);
4949 VM_BUG_ON(!success);
4950
4951 if (!--remaining)
4952 return false;
4953 }
4954 }
4955
4956 return true;
4957}
4958
4959static bool drain_evictable(struct lruvec *lruvec)
4960{
4961 int gen, type, zone;
4962 int remaining = MAX_LRU_BATCH;
4963
4964 for_each_gen_type_zone(gen, type, zone) {
4965 struct list_head *head = &lruvec->lrugen.lists[gen][type][zone];
4966
4967 while (!list_empty(head)) {
4968 bool success;
4969 struct page *page = lru_to_page(head);
4970
4971 VM_BUG_ON_PAGE(PageTail(page), page);
4972 VM_BUG_ON_PAGE(PageUnevictable(page), page);
4973 VM_BUG_ON_PAGE(PageActive(page), page);
4974 VM_BUG_ON_PAGE(page_is_file_lru(page) != type, page);
4975 VM_BUG_ON_PAGE(page_zonenum(page) != zone, page);
4976
4977 prefetchw_prev_lru_page(page, head, flags);
4978
4979 success = lru_gen_del_page(lruvec, page, false);
4980 VM_BUG_ON(!success);
4981 add_page_to_lru_list(page, lruvec);
4982
4983 if (!--remaining)
4984 return false;
4985 }
4986 }
4987
4988 return true;
4989}
4990
4991static void lru_gen_change_state(bool enable)
4992{
4993 static DEFINE_MUTEX(state_mutex);
4994
4995 struct mem_cgroup *memcg;
4996
4997 cgroup_lock();
4998 cpus_read_lock();
4999 get_online_mems();
5000 mutex_lock(&state_mutex);
5001
5002 if (enable == lru_gen_enabled())
5003 goto unlock;
5004
5005 if (enable)
5006 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5007 else
5008 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5009
5010 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5011 do {
5012 int nid;
5013
5014 for_each_node(nid) {
5015 struct lruvec *lruvec = get_lruvec(memcg, nid);
5016
5017 if (!lruvec)
5018 continue;
5019
5020 spin_lock_irq(&lruvec->lru_lock);
5021
5022 VM_BUG_ON(!seq_is_valid(lruvec));
5023 VM_BUG_ON(!state_is_valid(lruvec));
5024
5025 lruvec->lrugen.enabled = enable;
5026
5027 while (!(enable ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5028 spin_unlock_irq(&lruvec->lru_lock);
5029 cond_resched();
5030 spin_lock_irq(&lruvec->lru_lock);
5031 }
5032
5033 spin_unlock_irq(&lruvec->lru_lock);
5034 }
5035
5036 cond_resched();
5037 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5038unlock:
5039 mutex_unlock(&state_mutex);
5040 put_online_mems();
5041 cpus_read_unlock();
5042 cgroup_unlock();
5043}
5044
5045/******************************************************************************
5046 * sysfs interface
5047 ******************************************************************************/
5048
Yu Zhaoe8507812022-01-27 20:08:50 -07005049static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5050{
5051 return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5052}
5053
5054static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
5055 const char *buf, size_t len)
5056{
5057 unsigned int msecs;
5058
5059 if (kstrtouint(buf, 0, &msecs))
5060 return -EINVAL;
5061
5062 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5063
5064 return len;
5065}
5066
5067static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
5068 min_ttl_ms, 0644, show_min_ttl, store_min_ttl
5069);
5070
Yu Zhao76f7f072022-01-27 19:52:09 -07005071static ssize_t show_enable(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5072{
5073 unsigned int caps = 0;
5074
5075 if (get_cap(LRU_GEN_CORE))
5076 caps |= BIT(LRU_GEN_CORE);
5077
5078 if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
5079 caps |= BIT(LRU_GEN_MM_WALK);
5080
5081 if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) && get_cap(LRU_GEN_NONLEAF_YOUNG))
5082 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5083
5084 return snprintf(buf, PAGE_SIZE, "0x%04x\n", caps);
5085}
5086
5087static ssize_t store_enable(struct kobject *kobj, struct kobj_attribute *attr,
5088 const char *buf, size_t len)
5089{
5090 int i;
5091 unsigned int caps;
5092
5093 if (tolower(*buf) == 'n')
5094 caps = 0;
5095 else if (tolower(*buf) == 'y')
5096 caps = -1;
5097 else if (kstrtouint(buf, 0, &caps))
5098 return -EINVAL;
5099
5100 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5101 bool enable = caps & BIT(i);
5102
5103 if (i == LRU_GEN_CORE)
5104 lru_gen_change_state(enable);
5105 else if (enable)
5106 static_branch_enable(&lru_gen_caps[i]);
5107 else
5108 static_branch_disable(&lru_gen_caps[i]);
5109 }
5110
5111 return len;
5112}
5113
5114static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
5115 enabled, 0644, show_enable, store_enable
5116);
5117
5118static struct attribute *lru_gen_attrs[] = {
Yu Zhaoe8507812022-01-27 20:08:50 -07005119 &lru_gen_min_ttl_attr.attr,
Yu Zhao76f7f072022-01-27 19:52:09 -07005120 &lru_gen_enabled_attr.attr,
5121 NULL
5122};
5123
5124static struct attribute_group lru_gen_attr_group = {
5125 .name = "lru_gen",
5126 .attrs = lru_gen_attrs,
5127};
5128
5129/******************************************************************************
Yu Zhao3d18c9e2022-01-27 20:12:41 -07005130 * debugfs interface
5131 ******************************************************************************/
5132
5133static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5134{
5135 struct mem_cgroup *memcg;
5136 loff_t nr_to_skip = *pos;
5137
5138 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5139 if (!m->private)
5140 return ERR_PTR(-ENOMEM);
5141
5142 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5143 do {
5144 int nid;
5145
5146 for_each_node_state(nid, N_MEMORY) {
5147 if (!nr_to_skip--)
5148 return get_lruvec(memcg, nid);
5149 }
5150 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5151
5152 return NULL;
5153}
5154
5155static void lru_gen_seq_stop(struct seq_file *m, void *v)
5156{
5157 if (!IS_ERR_OR_NULL(v))
5158 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5159
5160 kvfree(m->private);
5161 m->private = NULL;
5162}
5163
5164static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5165{
5166 int nid = lruvec_pgdat(v)->node_id;
5167 struct mem_cgroup *memcg = lruvec_memcg(v);
5168
5169 ++*pos;
5170
5171 nid = next_memory_node(nid);
5172 if (nid == MAX_NUMNODES) {
5173 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5174 if (!memcg)
5175 return NULL;
5176
5177 nid = first_memory_node;
5178 }
5179
5180 return get_lruvec(memcg, nid);
5181}
5182
5183static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5184 unsigned long max_seq, unsigned long *min_seq,
5185 unsigned long seq)
5186{
5187 int i;
5188 int type, tier;
5189 int hist = lru_hist_from_seq(seq);
5190 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5191
5192 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5193 seq_printf(m, " %10d", tier);
5194 for (type = 0; type < ANON_AND_FILE; type++) {
5195 unsigned long n[3] = {};
5196
5197 if (seq == max_seq) {
5198 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5199 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5200
5201 seq_printf(m, " %10luR %10luT %10lu ", n[0], n[1], n[2]);
5202 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5203 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5204 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5205 if (tier)
5206 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5207
5208 seq_printf(m, " %10lur %10lue %10lup", n[0], n[1], n[2]);
5209 } else
5210 seq_puts(m, " 0 0 0 ");
5211 }
5212 seq_putc(m, '\n');
5213 }
5214
5215 seq_puts(m, " ");
5216 for (i = 0; i < NR_MM_STATS; i++) {
5217 if (seq == max_seq && NR_HIST_GENS == 1)
5218 seq_printf(m, " %10lu%c", READ_ONCE(lruvec->mm_state.stats[hist][i]),
5219 toupper(MM_STAT_CODES[i]));
5220 else if (seq != max_seq && NR_HIST_GENS > 1)
5221 seq_printf(m, " %10lu%c", READ_ONCE(lruvec->mm_state.stats[hist][i]),
5222 MM_STAT_CODES[i]);
5223 else
5224 seq_puts(m, " 0 ");
5225 }
5226 seq_putc(m, '\n');
5227}
5228
5229static int lru_gen_seq_show(struct seq_file *m, void *v)
5230{
5231 unsigned long seq;
5232 bool full = !debugfs_real_fops(m->file)->write;
5233 struct lruvec *lruvec = v;
5234 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5235 int nid = lruvec_pgdat(lruvec)->node_id;
5236 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5237 DEFINE_MAX_SEQ(lruvec);
5238 DEFINE_MIN_SEQ(lruvec);
5239
5240 if (nid == first_memory_node) {
5241 const char *path = memcg ? m->private : "";
5242
5243#ifdef CONFIG_MEMCG
5244 if (memcg)
5245 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5246#endif
5247 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5248 }
5249
5250 seq_printf(m, " node %5d\n", nid);
5251
5252 if (!full)
5253 seq = min_seq[LRU_GEN_ANON];
5254 else if (max_seq >= MAX_NR_GENS)
5255 seq = max_seq - MAX_NR_GENS + 1;
5256 else
5257 seq = 0;
5258
5259 for (; seq <= max_seq; seq++) {
5260 int type, zone;
5261 int gen = lru_gen_from_seq(seq);
5262 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5263
5264 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5265
5266 for (type = 0; type < ANON_AND_FILE; type++) {
5267 long size = 0;
5268 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5269
5270 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5271 size += READ_ONCE(lrugen->nr_pages[gen][type][zone]);
5272
5273 seq_printf(m, " %10lu%c", max(size, 0L), mark);
5274 }
5275
5276 seq_putc(m, '\n');
5277
5278 if (full)
5279 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5280 }
5281
5282 return 0;
5283}
5284
5285static const struct seq_operations lru_gen_seq_ops = {
5286 .start = lru_gen_seq_start,
5287 .stop = lru_gen_seq_stop,
5288 .next = lru_gen_seq_next,
5289 .show = lru_gen_seq_show,
5290};
5291
5292static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5293 bool can_swap, bool full_scan)
5294{
5295 DEFINE_MAX_SEQ(lruvec);
5296
5297 if (seq == max_seq)
5298 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, full_scan);
5299
5300 return seq > max_seq ? -EINVAL : 0;
5301}
5302
5303static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5304 int swappiness, unsigned long nr_to_reclaim)
5305{
5306 struct blk_plug plug;
5307 int err = -EINTR;
5308 DEFINE_MAX_SEQ(lruvec);
5309
5310 if (seq + MIN_NR_GENS > max_seq)
5311 return -EINVAL;
5312
5313 sc->nr_reclaimed = 0;
5314
5315 blk_start_plug(&plug);
5316
5317 while (!signal_pending(current)) {
5318 DEFINE_MIN_SEQ(lruvec);
5319
5320 if (seq < min_seq[!swappiness] || sc->nr_reclaimed >= nr_to_reclaim ||
5321 !evict_pages(lruvec, sc, swappiness, NULL)) {
5322 err = 0;
5323 break;
5324 }
5325
5326 cond_resched();
5327 }
5328
5329 blk_finish_plug(&plug);
5330
5331 return err;
5332}
5333
5334static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5335 struct scan_control *sc, int swappiness, unsigned long opt)
5336{
5337 struct lruvec *lruvec;
5338 int err = -EINVAL;
5339 struct mem_cgroup *memcg = NULL;
5340
5341 if (!mem_cgroup_disabled()) {
5342 rcu_read_lock();
5343 memcg = mem_cgroup_from_id(memcg_id);
5344#ifdef CONFIG_MEMCG
5345 if (memcg && !css_tryget(&memcg->css))
5346 memcg = NULL;
5347#endif
5348 rcu_read_unlock();
5349
5350 if (!memcg)
5351 goto done;
5352 }
5353 if (memcg_id != mem_cgroup_id(memcg))
5354 goto done;
5355
5356 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5357 goto done;
5358
5359 lruvec = get_lruvec(memcg, nid);
5360
5361 if (swappiness < 0)
5362 swappiness = get_swappiness(lruvec, sc);
5363 else if (swappiness > 200)
5364 goto done;
5365
5366 switch (cmd) {
5367 case '+':
5368 err = run_aging(lruvec, seq, sc, swappiness, opt);
5369 break;
5370 case '-':
5371 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5372 break;
5373 }
5374done:
5375 mem_cgroup_put(memcg);
5376
5377 return err;
5378}
5379
5380static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5381 size_t len, loff_t *pos)
5382{
5383 void *buf;
5384 char *cur, *next;
5385 unsigned int flags;
5386 int err = 0;
5387 struct scan_control sc = {
5388 .may_writepage = true,
5389 .may_unmap = true,
5390 .may_swap = true,
5391 .reclaim_idx = MAX_NR_ZONES - 1,
5392 .gfp_mask = GFP_KERNEL,
5393 };
5394
5395 buf = kvmalloc(len + 1, GFP_KERNEL);
5396 if (!buf)
5397 return -ENOMEM;
5398
5399 if (copy_from_user(buf, src, len)) {
5400 kvfree(buf);
5401 return -EFAULT;
5402 }
5403
5404 next = buf;
5405 next[len] = '\0';
5406
5407 sc.reclaim_state.mm_walk = alloc_mm_walk();
5408 if (!sc.reclaim_state.mm_walk) {
5409 kvfree(buf);
5410 return -ENOMEM;
5411 }
5412
5413 set_task_reclaim_state(current, &sc.reclaim_state);
5414 flags = memalloc_noreclaim_save();
5415
5416 while ((cur = strsep(&next, ",;\n"))) {
5417 int n;
5418 int end;
5419 char cmd;
5420 unsigned int memcg_id;
5421 unsigned int nid;
5422 unsigned long seq;
5423 unsigned int swappiness = -1;
5424 unsigned long opt = -1;
5425
5426 cur = skip_spaces(cur);
5427 if (!*cur)
5428 continue;
5429
5430 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5431 &seq, &end, &swappiness, &end, &opt, &end);
5432 if (n < 4 || cur[end]) {
5433 err = -EINVAL;
5434 break;
5435 }
5436
5437 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5438 if (err)
5439 break;
5440 }
5441
5442 memalloc_noreclaim_restore(flags);
5443 set_task_reclaim_state(current, NULL);
5444
5445 free_mm_walk(sc.reclaim_state.mm_walk);
5446 kvfree(buf);
5447
5448 return err ? : len;
5449}
5450
5451static int lru_gen_seq_open(struct inode *inode, struct file *file)
5452{
5453 return seq_open(file, &lru_gen_seq_ops);
5454}
5455
5456static const struct file_operations lru_gen_rw_fops = {
5457 .open = lru_gen_seq_open,
5458 .read = seq_read,
5459 .write = lru_gen_seq_write,
5460 .llseek = seq_lseek,
5461 .release = seq_release,
5462};
5463
5464static const struct file_operations lru_gen_ro_fops = {
5465 .open = lru_gen_seq_open,
5466 .read = seq_read,
5467 .llseek = seq_lseek,
5468 .release = seq_release,
5469};
5470
5471/******************************************************************************
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005472 * initialization
5473 ******************************************************************************/
5474
5475void lru_gen_init_lruvec(struct lruvec *lruvec)
5476{
Yu Zhaoe8507812022-01-27 20:08:50 -07005477 int i;
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005478 int gen, type, zone;
5479 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5480
5481 lrugen->max_seq = MIN_NR_GENS + 1;
Yu Zhao76f7f072022-01-27 19:52:09 -07005482 lrugen->enabled = lru_gen_enabled();
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005483
Yu Zhaoe8507812022-01-27 20:08:50 -07005484 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5485 lrugen->timestamps[i] = jiffies;
5486
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005487 for_each_gen_type_zone(gen, type, zone)
5488 INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
Yu Zhao5280d762022-01-27 19:23:01 -07005489
5490 lruvec->mm_state.seq = MIN_NR_GENS;
5491 init_waitqueue_head(&lruvec->mm_state.wait);
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005492}
5493
5494#ifdef CONFIG_MEMCG
5495void lru_gen_init_memcg(struct mem_cgroup *memcg)
5496{
Yu Zhao5280d762022-01-27 19:23:01 -07005497 INIT_LIST_HEAD(&memcg->mm_list.fifo);
5498 spin_lock_init(&memcg->mm_list.lock);
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005499}
5500
5501void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5502{
Yu Zhao5280d762022-01-27 19:23:01 -07005503 int i;
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005504 int nid;
5505
5506 for_each_node(nid) {
5507 struct lruvec *lruvec = get_lruvec(memcg, nid);
5508
5509 VM_BUG_ON(memchr_inv(lruvec->lrugen.nr_pages, 0,
5510 sizeof(lruvec->lrugen.nr_pages)));
Yu Zhao5280d762022-01-27 19:23:01 -07005511
5512 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5513 bitmap_free(lruvec->mm_state.filters[i]);
5514 lruvec->mm_state.filters[i] = NULL;
5515 }
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005516 }
5517}
5518#endif
5519
5520static int __init init_lru_gen(void)
5521{
5522 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5523 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
Yu Zhao5280d762022-01-27 19:23:01 -07005524 BUILD_BUG_ON(sizeof(MM_STAT_CODES) != NR_MM_STATS + 1);
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005525
Yu Zhao76f7f072022-01-27 19:52:09 -07005526 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5527 pr_err("lru_gen: failed to create sysfs group\n");
5528
Yu Zhao3d18c9e2022-01-27 20:12:41 -07005529 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5530 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5531
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005532 return 0;
5533};
5534late_initcall(init_lru_gen);
5535
Yu Zhaoa1537a62022-01-27 20:32:37 -07005536#else
5537
5538static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5539{
5540}
5541
5542static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5543{
5544}
5545
Yu Zhaof88ed5a2021-01-25 21:12:33 -07005546#endif /* CONFIG_LRU_GEN */
5547
Johannes Weinerafaf07a2019-11-30 17:55:46 -08005548static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005549{
5550 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07005551 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005552 unsigned long nr_to_scan;
5553 enum lru_list lru;
5554 unsigned long nr_reclaimed = 0;
5555 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5556 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07005557 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005558
Yu Zhaoa1537a62022-01-27 20:32:37 -07005559 if (lru_gen_enabled()) {
5560 lru_gen_shrink_lruvec(lruvec, sc);
5561 return;
5562 }
5563
Johannes Weinerafaf07a2019-11-30 17:55:46 -08005564 get_scan_count(lruvec, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005565
Mel Gormane82e0562013-07-03 15:01:44 -07005566 /* Record the original scan target for proportional adjustments later */
5567 memcpy(targets, nr, sizeof(nr));
5568
Mel Gorman1a501902014-06-04 16:10:49 -07005569 /*
5570 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5571 * event that can occur when there is little memory pressure e.g.
5572 * multiple streaming readers/writers. Hence, we do not abort scanning
5573 * when the requested number of pages are reclaimed when scanning at
5574 * DEF_PRIORITY on the assumption that the fact we are direct
5575 * reclaiming implies that kswapd is not keeping up and it is best to
5576 * do a batch of work at once. For memcg reclaim one check is made to
5577 * abort proportional reclaim if either the file or anon lru has already
5578 * dropped to zero at the first pass.
5579 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08005580 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
Mel Gorman1a501902014-06-04 16:10:49 -07005581 sc->priority == DEF_PRIORITY);
5582
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005583 blk_start_plug(&plug);
5584 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5585 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07005586 unsigned long nr_anon, nr_file, percentage;
5587 unsigned long nr_scanned;
5588
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005589 for_each_evictable_lru(lru) {
5590 if (nr[lru]) {
5591 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5592 nr[lru] -= nr_to_scan;
5593
5594 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner3b991202019-04-18 17:50:34 -07005595 lruvec, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005596 }
5597 }
Mel Gormane82e0562013-07-03 15:01:44 -07005598
Michal Hockobd041732016-12-02 17:26:48 -08005599 cond_resched();
5600
Mel Gormane82e0562013-07-03 15:01:44 -07005601 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
5602 continue;
5603
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005604 /*
Mel Gormane82e0562013-07-03 15:01:44 -07005605 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07005606 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07005607 * proportionally what was requested by get_scan_count(). We
5608 * stop reclaiming one LRU and reduce the amount scanning
5609 * proportional to the original scan target.
5610 */
5611 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5612 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5613
Mel Gorman1a501902014-06-04 16:10:49 -07005614 /*
5615 * It's just vindictive to attack the larger once the smaller
5616 * has gone to zero. And given the way we stop scanning the
5617 * smaller below, this makes sure that we only make one nudge
5618 * towards proportionality once we've got nr_to_reclaim.
5619 */
5620 if (!nr_file || !nr_anon)
5621 break;
5622
Mel Gormane82e0562013-07-03 15:01:44 -07005623 if (nr_file > nr_anon) {
5624 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5625 targets[LRU_ACTIVE_ANON] + 1;
5626 lru = LRU_BASE;
5627 percentage = nr_anon * 100 / scan_target;
5628 } else {
5629 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5630 targets[LRU_ACTIVE_FILE] + 1;
5631 lru = LRU_FILE;
5632 percentage = nr_file * 100 / scan_target;
5633 }
5634
5635 /* Stop scanning the smaller of the LRU */
5636 nr[lru] = 0;
5637 nr[lru + LRU_ACTIVE] = 0;
5638
5639 /*
5640 * Recalculate the other LRU scan count based on its original
5641 * scan target and the percentage scanning already complete
5642 */
5643 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5644 nr_scanned = targets[lru] - nr[lru];
5645 nr[lru] = targets[lru] * (100 - percentage) / 100;
5646 nr[lru] -= min(nr[lru], nr_scanned);
5647
5648 lru += LRU_ACTIVE;
5649 nr_scanned = targets[lru] - nr[lru];
5650 nr[lru] = targets[lru] * (100 - percentage) / 100;
5651 nr[lru] -= min(nr[lru], nr_scanned);
5652
5653 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005654 }
5655 blk_finish_plug(&plug);
5656 sc->nr_reclaimed += nr_reclaimed;
5657
5658 /*
5659 * Even if we did not try to evict anon pages at all, we want to
5660 * rebalance the anon lru active/inactive ratio.
5661 */
Dave Hansen2f368a92021-09-02 14:59:23 -07005662 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5663 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005664 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5665 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08005666}
5667
Mel Gorman23b9da52012-05-29 15:06:20 -07005668/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07005669static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07005670{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08005671 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07005672 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07005673 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07005674 return true;
5675
5676 return false;
5677}
5678
Rik van Riel4f98a2f2008-10-18 20:26:32 -07005679/*
Mel Gorman23b9da52012-05-29 15:06:20 -07005680 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5681 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5682 * true if more pages should be reclaimed such that when the page allocator
Qiwu Chendf3a45f2020-06-03 16:01:21 -07005683 * calls try_to_compact_pages() that it will have enough free pages to succeed.
Mel Gorman23b9da52012-05-29 15:06:20 -07005684 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08005685 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07005686static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08005687 unsigned long nr_reclaimed,
Mel Gorman3e7d3442011-01-13 15:45:56 -08005688 struct scan_control *sc)
5689{
5690 unsigned long pages_for_compaction;
5691 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07005692 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08005693
5694 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07005695 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08005696 return false;
5697
Vlastimil Babka5ee04712019-09-23 15:37:29 -07005698 /*
5699 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5700 * number of pages that were scanned. This will return to the caller
5701 * with the risk reclaim/compaction and the resulting allocation attempt
5702 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5703 * allocations through requiring that the full LRU list has been scanned
5704 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5705 * scan, but that approximation was wrong, and there were corner cases
5706 * where always a non-zero amount of pages were scanned.
5707 */
5708 if (!nr_reclaimed)
5709 return false;
Mel Gorman3e7d3442011-01-13 15:45:56 -08005710
Mel Gorman3e7d3442011-01-13 15:45:56 -08005711 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07005712 for (z = 0; z <= sc->reclaim_idx; z++) {
5713 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07005714 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07005715 continue;
5716
5717 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07005718 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07005719 case COMPACT_CONTINUE:
5720 return false;
5721 default:
5722 /* check next zone */
5723 ;
5724 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08005725 }
Hillf Danton1c6c1592019-09-23 15:37:26 -07005726
5727 /*
5728 * If we have not reclaimed enough pages for compaction and the
5729 * inactive lists are large enough, continue reclaiming
5730 */
5731 pages_for_compaction = compact_gap(sc->order);
5732 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Keith Buscha2a36482021-09-02 14:59:26 -07005733 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
Hillf Danton1c6c1592019-09-23 15:37:26 -07005734 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5735
Vlastimil Babka5ee04712019-09-23 15:37:29 -07005736 return inactive_lru_pages > pages_for_compaction;
Mel Gorman3e7d3442011-01-13 15:45:56 -08005737}
5738
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005739static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08005740{
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005741 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
Johannes Weinerd2af3392019-11-30 17:55:43 -08005742 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08005743
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005744 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
Johannes Weiner56600482012-01-12 17:17:59 -08005745 do {
Johannes Weinerafaf07a2019-11-30 17:55:46 -08005746 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Johannes Weinerd2af3392019-11-30 17:55:43 -08005747 unsigned long reclaimed;
5748 unsigned long scanned;
Johannes Weiner56600482012-01-12 17:17:59 -08005749
Xunlei Pange3336ca2020-09-04 16:35:27 -07005750 /*
5751 * This loop can become CPU-bound when target memcgs
5752 * aren't eligible for reclaim - either because they
5753 * don't have any reclaimable pages, or because their
5754 * memory is explicitly protected. Avoid soft lockups.
5755 */
5756 cond_resched();
5757
Chris Down45c7f7e2020-08-06 23:22:05 -07005758 mem_cgroup_calculate_protection(target_memcg, memcg);
5759
5760 if (mem_cgroup_below_min(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08005761 /*
5762 * Hard protection.
5763 * If there is no reclaimable memory, OOM.
5764 */
5765 continue;
Chris Down45c7f7e2020-08-06 23:22:05 -07005766 } else if (mem_cgroup_below_low(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08005767 /*
5768 * Soft protection.
5769 * Respect the protection only as long as
5770 * there is an unprotected supply
5771 * of reclaimable memory from other cgroups.
5772 */
5773 if (!sc->memcg_low_reclaim) {
5774 sc->memcg_low_skipped = 1;
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07005775 continue;
Johannes Weiner241994ed2015-02-11 15:26:06 -08005776 }
Johannes Weinerd2af3392019-11-30 17:55:43 -08005777 memcg_memory_event(memcg, MEMCG_LOW);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08005778 }
5779
Johannes Weinerd2af3392019-11-30 17:55:43 -08005780 reclaimed = sc->nr_reclaimed;
5781 scanned = sc->nr_scanned;
Johannes Weinerafaf07a2019-11-30 17:55:46 -08005782
5783 shrink_lruvec(lruvec, sc);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07005784
Johannes Weinerd2af3392019-11-30 17:55:43 -08005785 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5786 sc->priority);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07005787
Johannes Weinerd2af3392019-11-30 17:55:43 -08005788 /* Record the group's reclaim efficiency */
5789 vmpressure(sc->gfp_mask, memcg, false,
5790 sc->nr_scanned - scanned,
5791 sc->nr_reclaimed - reclaimed);
Andrey Ryabinind108c772018-04-10 16:27:59 -07005792
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005793 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5794}
5795
Liu Song6c9e09072020-01-30 22:14:08 -08005796static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005797{
5798 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005799 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner1b051172019-11-30 17:55:52 -08005800 struct lruvec *target_lruvec;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005801 bool reclaimable = false;
5802
Johannes Weiner1b051172019-11-30 17:55:52 -08005803 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5804
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005805again:
5806 memset(&sc->nr, 0, sizeof(sc->nr));
5807
5808 nr_reclaimed = sc->nr_reclaimed;
5809 nr_scanned = sc->nr_scanned;
5810
Yu Zhao3eb07702020-09-27 20:49:08 -06005811 prepare_scan_count(pgdat, sc);
Johannes Weiner53138ce2019-11-30 17:55:56 -08005812
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08005813 shrink_node_memcgs(pgdat, sc);
Andrey Ryabinind108c772018-04-10 16:27:59 -07005814
Johannes Weinerd2af3392019-11-30 17:55:43 -08005815 if (reclaim_state) {
5816 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
5817 reclaim_state->reclaimed_slab = 0;
5818 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07005819
Johannes Weinerd2af3392019-11-30 17:55:43 -08005820 /* Record the subtree's reclaim efficiency */
Johannes Weiner1b051172019-11-30 17:55:52 -08005821 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Johannes Weinerd2af3392019-11-30 17:55:43 -08005822 sc->nr_scanned - nr_scanned,
5823 sc->nr_reclaimed - nr_reclaimed);
5824
5825 if (sc->nr_reclaimed - nr_reclaimed)
5826 reclaimable = true;
5827
5828 if (current_is_kswapd()) {
5829 /*
5830 * If reclaim is isolating dirty pages under writeback,
5831 * it implies that the long-lived page allocation rate
5832 * is exceeding the page laundering rate. Either the
5833 * global limits are not being effective at throttling
5834 * processes due to the page distribution throughout
5835 * zones or there is heavy usage of a slow backing
5836 * device. The only option is to throttle from reclaim
5837 * context which is not ideal as there is no guarantee
5838 * the dirtying process is throttled in the same way
5839 * balance_dirty_pages() manages.
5840 *
5841 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5842 * count the number of pages under pages flagged for
5843 * immediate reclaim and stall if any are encountered
5844 * in the nr_immediate check below.
5845 */
5846 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5847 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Andrey Ryabinind108c772018-04-10 16:27:59 -07005848
Johannes Weinerd2af3392019-11-30 17:55:43 -08005849 /* Allow kswapd to start writing pages during reclaim.*/
5850 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5851 set_bit(PGDAT_DIRTY, &pgdat->flags);
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07005852
5853 /*
Randy Dunlap1eba09c2020-08-11 18:33:26 -07005854 * If kswapd scans pages marked for immediate
Johannes Weinerd2af3392019-11-30 17:55:43 -08005855 * reclaim and under writeback (nr_immediate), it
5856 * implies that pages are cycling through the LRU
5857 * faster than they are written so also forcibly stall.
Andrey Ryabinind108c772018-04-10 16:27:59 -07005858 */
Johannes Weinerd2af3392019-11-30 17:55:43 -08005859 if (sc->nr.immediate)
5860 congestion_wait(BLK_RW_ASYNC, HZ/10);
5861 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07005862
Johannes Weinerd2af3392019-11-30 17:55:43 -08005863 /*
Johannes Weiner1b051172019-11-30 17:55:52 -08005864 * Tag a node/memcg as congested if all the dirty pages
5865 * scanned were backed by a congested BDI and
5866 * wait_iff_congested will stall.
5867 *
Johannes Weinerd2af3392019-11-30 17:55:43 -08005868 * Legacy memcg will stall in page writeback so avoid forcibly
5869 * stalling in wait_iff_congested().
5870 */
Johannes Weiner1b051172019-11-30 17:55:52 -08005871 if ((current_is_kswapd() ||
5872 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
Johannes Weinerd2af3392019-11-30 17:55:43 -08005873 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
Johannes Weiner1b051172019-11-30 17:55:52 -08005874 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
Johannes Weinerd2af3392019-11-30 17:55:43 -08005875
5876 /*
5877 * Stall direct reclaim for IO completions if underlying BDIs
5878 * and node is congested. Allow kswapd to continue until it
5879 * starts encountering unqueued dirty pages or cycling through
5880 * the LRU too quickly.
5881 */
Johannes Weiner1b051172019-11-30 17:55:52 -08005882 if (!current_is_kswapd() && current_may_throttle() &&
5883 !sc->hibernation_mode &&
5884 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
Johannes Weinerd2af3392019-11-30 17:55:43 -08005885 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
5886
5887 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
5888 sc))
5889 goto again;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07005890
Johannes Weinerc73322d2017-05-03 14:51:51 -07005891 /*
5892 * Kswapd gives up on balancing particular nodes after too
5893 * many failures to reclaim anything from them and goes to
5894 * sleep. On reclaim progress, reset the failure counter. A
5895 * successful direct reclaim run will revive a dormant kswapd.
5896 */
5897 if (reclaimable)
5898 pgdat->kswapd_failures = 0;
Johannes Weinerf16015f2012-01-12 17:17:52 -08005899}
5900
Vlastimil Babka53853e22014-10-09 15:27:02 -07005901/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07005902 * Returns true if compaction should go ahead for a costly-order request, or
5903 * the allocation would already succeed without compaction. Return false if we
5904 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07005905 */
Mel Gorman4f588332016-07-28 15:46:38 -07005906static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08005907{
Mel Gorman31483b62016-07-28 15:45:46 -07005908 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07005909 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08005910
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07005911 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
5912 if (suitable == COMPACT_SUCCESS)
5913 /* Allocation should succeed already. Don't reclaim. */
5914 return true;
5915 if (suitable == COMPACT_SKIPPED)
5916 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08005917 return false;
5918
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07005919 /*
5920 * Compaction is already possible, but it takes time to run and there
5921 * are potentially other callers using the pages just freed. So proceed
5922 * with reclaim to make a buffer of free pages available to give
5923 * compaction a reasonable chance of completing and allocating the page.
5924 * Note that we won't actually reclaim the whole buffer in one attempt
5925 * as the target watermark in should_continue_reclaim() is lower. But if
5926 * we are already above the high+gap watermark, don't reclaim at all.
5927 */
5928 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
5929
5930 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08005931}
5932
Linus Torvalds1da177e2005-04-16 15:20:36 -07005933/*
5934 * This is the direct reclaim path, for page-allocating processes. We only
5935 * try to reclaim pages from zones which will satisfy the caller's allocation
5936 * request.
5937 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07005938 * If a zone is deemed to be full of pinned pages then just give it a light
5939 * scan then give up on it.
5940 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07005941static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005942{
Mel Gormandd1a2392008-04-28 02:12:17 -07005943 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07005944 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07005945 unsigned long nr_soft_reclaimed;
5946 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07005947 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07005948 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08005949
Mel Gormancc715d92012-03-21 16:34:00 -07005950 /*
5951 * If the number of buffer_heads in the machine exceeds the maximum
5952 * allowed level, force direct reclaim to scan the highmem zone as
5953 * highmem pages could be pinning lowmem pages storing buffer_heads
5954 */
Weijie Yang619d0d762014-04-07 15:36:59 -07005955 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07005956 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07005957 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07005958 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07005959 }
Mel Gormancc715d92012-03-21 16:34:00 -07005960
Mel Gormand4debc62010-08-09 17:19:29 -07005961 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07005962 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07005963 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08005964 * Take care memory controller reclaiming has small influence
5965 * to global LRU.
5966 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08005967 if (!cgroup_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04005968 if (!cpuset_zone_allowed(zone,
5969 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08005970 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07005971
Johannes Weiner0b064962014-08-06 16:06:12 -07005972 /*
5973 * If we already have plenty of memory free for
5974 * compaction in this zone, don't free any more.
5975 * Even though compaction is invoked for any
5976 * non-zero order, only frequent costly order
5977 * reclamation is disruptive enough to become a
5978 * noticeable problem, like transparent huge
5979 * page allocations.
5980 */
5981 if (IS_ENABLED(CONFIG_COMPACTION) &&
5982 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07005983 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07005984 sc->compaction_ready = true;
5985 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07005986 }
Johannes Weiner0b064962014-08-06 16:06:12 -07005987
Andrew Morton0608f432013-09-24 15:27:41 -07005988 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07005989 * Shrink each node in the zonelist once. If the
5990 * zonelist is ordered by zone (not the default) then a
5991 * node may be shrunk multiple times but in that case
5992 * the user prefers lower zones being preserved.
5993 */
5994 if (zone->zone_pgdat == last_pgdat)
5995 continue;
5996
5997 /*
Andrew Morton0608f432013-09-24 15:27:41 -07005998 * This steals pages from memory cgroups over softlimit
5999 * and returns the number of reclaimed pages and
6000 * scanned pages. This works for global memory pressure
6001 * and balancing, not for a memcg's limit.
6002 */
6003 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07006004 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07006005 sc->order, sc->gfp_mask,
6006 &nr_soft_scanned);
6007 sc->nr_reclaimed += nr_soft_reclaimed;
6008 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07006009 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08006010 }
Nick Piggin408d8542006-09-25 23:31:27 -07006011
Mel Gorman79dafcd2016-07-28 15:45:53 -07006012 /* See comment about same check for global reclaim above */
6013 if (zone->zone_pgdat == last_pgdat)
6014 continue;
6015 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07006016 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006017 }
Mel Gormane0c23272011-10-31 17:09:33 -07006018
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07006019 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07006020 * Restore to original mask to avoid the impact on the caller if we
6021 * promoted it to __GFP_HIGHMEM.
6022 */
6023 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006024}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07006025
Johannes Weinerb9107182019-11-30 17:55:59 -08006026static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006027{
Johannes Weinerb9107182019-11-30 17:55:59 -08006028 struct lruvec *target_lruvec;
6029 unsigned long refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006030
Yu Zhaoa1537a62022-01-27 20:32:37 -07006031 if (lru_gen_enabled())
6032 return;
6033
Johannes Weinerb9107182019-11-30 17:55:59 -08006034 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
Joonsoo Kim170b04b72020-08-11 18:30:43 -07006035 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6036 target_lruvec->refaults[0] = refaults;
6037 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6038 target_lruvec->refaults[1] = refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006039}
6040
Linus Torvalds1da177e2005-04-16 15:20:36 -07006041/*
6042 * This is the main entry point to direct page reclaim.
6043 *
6044 * If a full scan of the inactive list fails to free enough memory then we
6045 * are "out of memory" and something needs to be killed.
6046 *
6047 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6048 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02006049 * caller can't do much about. We kick the writeback threads and take explicit
6050 * naps in the hope that some of these pages can be written. But if the
6051 * allocating task holds filesystem locks which prevent writeout this might not
6052 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07006053 *
6054 * returns: 0, if no pages reclaimed
6055 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07006056 */
Mel Gormandac1d272008-04-28 02:12:12 -07006057static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07006058 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006059{
Johannes Weiner241994ed2015-02-11 15:26:06 -08006060 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006061 pg_data_t *last_pgdat;
6062 struct zoneref *z;
6063 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08006064retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07006065 delayacct_freepages_start();
6066
Johannes Weinerb5ead352019-11-30 17:55:40 -08006067 if (!cgroup_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07006068 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006069
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07006070 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07006071 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6072 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08006073 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07006074 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07006075
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07006076 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07006077 break;
6078
6079 if (sc->compaction_ready)
6080 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006081
6082 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08006083 * If we're getting trouble reclaiming, start doing
6084 * writepage even in laptop mode.
6085 */
6086 if (sc->priority < DEF_PRIORITY - 2)
6087 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07006088 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07006089
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006090 last_pgdat = NULL;
6091 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6092 sc->nodemask) {
6093 if (zone->zone_pgdat == last_pgdat)
6094 continue;
6095 last_pgdat = zone->zone_pgdat;
Johannes Weiner1b051172019-11-30 17:55:52 -08006096
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006097 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
Johannes Weiner1b051172019-11-30 17:55:52 -08006098
6099 if (cgroup_reclaim(sc)) {
6100 struct lruvec *lruvec;
6101
6102 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6103 zone->zone_pgdat);
6104 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6105 }
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006106 }
6107
Keika Kobayashi873b4772008-07-25 01:48:52 -07006108 delayacct_freepages_end();
6109
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07006110 if (sc->nr_reclaimed)
6111 return sc->nr_reclaimed;
6112
Mel Gorman0cee34f2012-01-12 17:19:49 -08006113 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07006114 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08006115 return 1;
6116
Johannes Weinerb91ac372019-11-30 17:56:02 -08006117 /*
6118 * We make inactive:active ratio decisions based on the node's
6119 * composition of memory, but a restrictive reclaim_idx or a
6120 * memory.low cgroup setting can exempt large amounts of
6121 * memory from reclaim. Neither of which are very common, so
6122 * instead of doing costly eligibility calculations of the
6123 * entire cgroup subtree up front, we assume the estimates are
6124 * good, and retry with forcible deactivation if that fails.
6125 */
6126 if (sc->skipped_deactivate) {
6127 sc->priority = initial_priority;
6128 sc->force_deactivate = 1;
6129 sc->skipped_deactivate = 0;
6130 goto retry;
6131 }
6132
Johannes Weiner241994ed2015-02-11 15:26:06 -08006133 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07006134 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08006135 sc->priority = initial_priority;
Johannes Weinerb91ac372019-11-30 17:56:02 -08006136 sc->force_deactivate = 0;
Yisheng Xied6622f62017-05-03 14:53:57 -07006137 sc->memcg_low_reclaim = 1;
6138 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08006139 goto retry;
6140 }
6141
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07006142 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006143}
6144
Johannes Weinerc73322d2017-05-03 14:51:51 -07006145static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07006146{
6147 struct zone *zone;
6148 unsigned long pfmemalloc_reserve = 0;
6149 unsigned long free_pages = 0;
6150 int i;
6151 bool wmark_ok;
6152
Johannes Weinerc73322d2017-05-03 14:51:51 -07006153 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6154 return true;
6155
Mel Gorman55150612012-07-31 16:44:35 -07006156 for (i = 0; i <= ZONE_NORMAL; i++) {
6157 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd82017-05-03 14:51:54 -07006158 if (!managed_zone(zone))
6159 continue;
6160
6161 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07006162 continue;
6163
Mel Gorman55150612012-07-31 16:44:35 -07006164 pfmemalloc_reserve += min_wmark_pages(zone);
6165 free_pages += zone_page_state(zone, NR_FREE_PAGES);
6166 }
6167
Mel Gorman675becc2014-06-04 16:07:35 -07006168 /* If there are no reserves (unexpected config) then do not throttle */
6169 if (!pfmemalloc_reserve)
6170 return true;
6171
Mel Gorman55150612012-07-31 16:44:35 -07006172 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6173
6174 /* kswapd must be awake if processes are being throttled */
6175 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006176 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6177 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
Qian Cai5644e1fb2020-04-01 21:10:12 -07006178
Mel Gorman55150612012-07-31 16:44:35 -07006179 wake_up_interruptible(&pgdat->kswapd_wait);
6180 }
6181
6182 return wmark_ok;
6183}
6184
6185/*
6186 * Throttle direct reclaimers if backing storage is backed by the network
6187 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6188 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08006189 * when the low watermark is reached.
6190 *
6191 * Returns true if a fatal signal was delivered during throttling. If this
6192 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07006193 */
Mel Gorman50694c22012-11-26 16:29:48 -08006194static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07006195 nodemask_t *nodemask)
6196{
Mel Gorman675becc2014-06-04 16:07:35 -07006197 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07006198 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07006199 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07006200
6201 /*
6202 * Kernel threads should not be throttled as they may be indirectly
6203 * responsible for cleaning pages necessary for reclaim to make forward
6204 * progress. kjournald for example may enter direct reclaim while
6205 * committing a transaction where throttling it could forcing other
6206 * processes to block on log_wait_commit().
6207 */
6208 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08006209 goto out;
6210
6211 /*
6212 * If a fatal signal is pending, this process should not throttle.
6213 * It should return quickly so it can exit and free its memory
6214 */
6215 if (fatal_signal_pending(current))
6216 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07006217
Mel Gorman675becc2014-06-04 16:07:35 -07006218 /*
6219 * Check if the pfmemalloc reserves are ok by finding the first node
6220 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6221 * GFP_KERNEL will be required for allocating network buffers when
6222 * swapping over the network so ZONE_HIGHMEM is unusable.
6223 *
6224 * Throttling is based on the first usable node and throttled processes
6225 * wait on a queue until kswapd makes progress and wakes them. There
6226 * is an affinity then between processes waking up and where reclaim
6227 * progress has been made assuming the process wakes on the same node.
6228 * More importantly, processes running on remote nodes will not compete
6229 * for remote pfmemalloc reserves and processes on different nodes
6230 * should make reasonable progress.
6231 */
6232 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08006233 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07006234 if (zone_idx(zone) > ZONE_NORMAL)
6235 continue;
6236
6237 /* Throttle based on the first usable node */
6238 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07006239 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07006240 goto out;
6241 break;
6242 }
6243
6244 /* If no zone was usable by the allocation flags then do not throttle */
6245 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08006246 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07006247
Mel Gorman68243e72012-07-31 16:44:39 -07006248 /* Account for the throttling */
6249 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6250
Mel Gorman55150612012-07-31 16:44:35 -07006251 /*
6252 * If the caller cannot enter the filesystem, it's possible that it
6253 * is due to the caller holding an FS lock or performing a journal
6254 * transaction in the case of a filesystem like ext[3|4]. In this case,
6255 * it is not safe to block on pfmemalloc_wait as kswapd could be
6256 * blocked waiting on the same lock. Instead, throttle for up to a
6257 * second before continuing.
6258 */
Miaohe Lin2e786d92021-09-02 14:59:50 -07006259 if (!(gfp_mask & __GFP_FS))
Mel Gorman55150612012-07-31 16:44:35 -07006260 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07006261 allow_direct_reclaim(pgdat), HZ);
Miaohe Lin2e786d92021-09-02 14:59:50 -07006262 else
6263 /* Throttle until kswapd wakes the process */
6264 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6265 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08006266
Mel Gorman50694c22012-11-26 16:29:48 -08006267 if (fatal_signal_pending(current))
6268 return true;
6269
6270out:
6271 return false;
Mel Gorman55150612012-07-31 16:44:35 -07006272}
6273
Mel Gormandac1d272008-04-28 02:12:12 -07006274unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07006275 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08006276{
Mel Gorman33906bc2010-08-09 17:19:16 -07006277 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08006278 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08006279 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07006280 .gfp_mask = current_gfp_context(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07006281 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07006282 .order = order,
6283 .nodemask = nodemask,
6284 .priority = DEF_PRIORITY,
6285 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07006286 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07006287 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08006288 };
6289
Mel Gorman55150612012-07-31 16:44:35 -07006290 /*
Greg Thelenbb451fd2018-08-17 15:45:19 -07006291 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6292 * Confirm they are large enough for max values.
6293 */
6294 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
6295 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6296 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6297
6298 /*
Mel Gorman50694c22012-11-26 16:29:48 -08006299 * Do not enter reclaim if fatal signal was delivered while throttled.
6300 * 1 is returned so that the page allocator does not OOM kill at this
6301 * point.
Mel Gorman55150612012-07-31 16:44:35 -07006302 */
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07006303 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07006304 return 1;
6305
Andrew Morton1732d2b012019-07-16 16:26:15 -07006306 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07006307 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
Mel Gorman33906bc2010-08-09 17:19:16 -07006308
Vladimir Davydov3115cd92014-04-03 14:47:22 -07006309 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07006310
6311 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07006312 set_task_reclaim_state(current, NULL);
Mel Gorman33906bc2010-08-09 17:19:16 -07006313
6314 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08006315}
6316
Andrew Mortonc255a452012-07-31 16:43:02 -07006317#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08006318
Michal Hockod2e5fb92019-08-30 16:04:50 -07006319/* Only used by soft limit reclaim. Do not reuse for anything else. */
Mel Gormana9dd0a82016-07-28 15:46:02 -07006320unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07006321 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07006322 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07006323 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07006324{
Johannes Weinerafaf07a2019-11-30 17:55:46 -08006325 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Balbir Singh4e416952009-09-23 15:56:39 -07006326 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07006327 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07006328 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07006329 .may_writepage = !laptop_mode,
6330 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07006331 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07006332 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07006333 };
Ying Han0ae5e892011-05-26 16:25:25 -07006334
Michal Hockod2e5fb92019-08-30 16:04:50 -07006335 WARN_ON_ONCE(!current->reclaim_state);
6336
Balbir Singh4e416952009-09-23 15:56:39 -07006337 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6338 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006339
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07006340 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
Yafang Shao3481c372019-05-13 17:19:14 -07006341 sc.gfp_mask);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006342
Balbir Singh4e416952009-09-23 15:56:39 -07006343 /*
6344 * NOTE: Although we can get the priority field, using it
6345 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07006346 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07006347 * will pick up pages from other mem cgroup's as well. We hack
6348 * the priority and make it zero.
6349 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08006350 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006351
6352 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6353
Ying Han0ae5e892011-05-26 16:25:25 -07006354 *nr_scanned = sc.nr_scanned;
Yafang Shao0308f7c2019-07-16 16:26:12 -07006355
Balbir Singh4e416952009-09-23 15:56:39 -07006356 return sc.nr_reclaimed;
6357}
6358
Johannes Weiner72835c82012-01-12 17:18:32 -08006359unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07006360 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08006361 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07006362 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08006363{
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006364 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07006365 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08006366 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07006367 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07006368 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07006369 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07006370 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07006371 .target_mem_cgroup = memcg,
6372 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08006373 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07006374 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07006375 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07006376 };
Shakeel Buttfa40d1e2019-11-30 17:50:16 -08006377 /*
6378 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6379 * equal pressure on all the nodes. This is based on the assumption that
6380 * the reclaim does not bail out early.
6381 */
6382 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
Balbir Singh66e17072008-02-07 00:13:56 -08006383
Andrew Morton1732d2b012019-07-16 16:26:15 -07006384 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07006385 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
Vlastimil Babka499118e2017-05-08 15:59:50 -07006386 noreclaim_flag = memalloc_noreclaim_save();
Johannes Weinereb414682018-10-26 15:06:27 -07006387
Vladimir Davydov3115cd92014-04-03 14:47:22 -07006388 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weinereb414682018-10-26 15:06:27 -07006389
Vlastimil Babka499118e2017-05-08 15:59:50 -07006390 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006391 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07006392 set_task_reclaim_state(current, NULL);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07006393
6394 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08006395}
6396#endif
6397
Mel Gorman1d82de62016-07-28 15:45:43 -07006398static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07006399 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08006400{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08006401 struct mem_cgroup *memcg;
Johannes Weinerb91ac372019-11-30 17:56:02 -08006402 struct lruvec *lruvec;
Johannes Weinerf16015f2012-01-12 17:17:52 -08006403
Yu Zhaoa1537a62022-01-27 20:32:37 -07006404 if (lru_gen_enabled()) {
6405 lru_gen_age_node(pgdat, sc);
6406 return;
6407 }
6408
Dave Hansen2f368a92021-09-02 14:59:23 -07006409 if (!can_age_anon_pages(pgdat, sc))
Johannes Weinerb95a2f22012-01-12 17:18:06 -08006410 return;
6411
Johannes Weinerb91ac372019-11-30 17:56:02 -08006412 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6413 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6414 return;
6415
Johannes Weinerb95a2f22012-01-12 17:18:06 -08006416 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6417 do {
Johannes Weinerb91ac372019-11-30 17:56:02 -08006418 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6419 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6420 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08006421 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6422 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08006423}
6424
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006425static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
Mel Gorman1c308442018-12-28 00:35:52 -08006426{
6427 int i;
6428 struct zone *zone;
6429
6430 /*
6431 * Check for watermark boosts top-down as the higher zones
6432 * are more likely to be boosted. Both watermarks and boosts
Randy Dunlap1eba09c2020-08-11 18:33:26 -07006433 * should not be checked at the same time as reclaim would
Mel Gorman1c308442018-12-28 00:35:52 -08006434 * start prematurely when there is no boosting and a lower
6435 * zone is balanced.
6436 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006437 for (i = highest_zoneidx; i >= 0; i--) {
Mel Gorman1c308442018-12-28 00:35:52 -08006438 zone = pgdat->node_zones + i;
6439 if (!managed_zone(zone))
6440 continue;
6441
6442 if (zone->watermark_boost)
6443 return true;
6444 }
6445
6446 return false;
6447}
6448
Mel Gormane716f2e2017-05-03 14:53:45 -07006449/*
6450 * Returns true if there is an eligible zone balanced for the request order
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006451 * and highest_zoneidx
Mel Gormane716f2e2017-05-03 14:53:45 -07006452 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006453static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08006454{
Mel Gormane716f2e2017-05-03 14:53:45 -07006455 int i;
6456 unsigned long mark = -1;
6457 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08006458
Mel Gorman1c308442018-12-28 00:35:52 -08006459 /*
6460 * Check watermarks bottom-up as lower zones are more likely to
6461 * meet watermarks.
6462 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006463 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gormane716f2e2017-05-03 14:53:45 -07006464 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07006465
Mel Gormane716f2e2017-05-03 14:53:45 -07006466 if (!managed_zone(zone))
6467 continue;
6468
6469 mark = high_wmark_pages(zone);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006470 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
Mel Gormane716f2e2017-05-03 14:53:45 -07006471 return true;
6472 }
6473
6474 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006475 * If a node has no populated zone within highest_zoneidx, it does not
Mel Gormane716f2e2017-05-03 14:53:45 -07006476 * need balancing by definition. This can happen if a zone-restricted
6477 * allocation tries to wake a remote kswapd.
6478 */
6479 if (mark == -1)
6480 return true;
6481
6482 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08006483}
6484
Mel Gorman631b6e02017-05-03 14:53:41 -07006485/* Clear pgdat state for congested, dirty or under writeback. */
6486static void clear_pgdat_congested(pg_data_t *pgdat)
6487{
Johannes Weiner1b051172019-11-30 17:55:52 -08006488 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6489
6490 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
Mel Gorman631b6e02017-05-03 14:53:41 -07006491 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6492 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6493}
6494
Mel Gorman1741c872011-01-13 15:46:21 -08006495/*
Mel Gorman55150612012-07-31 16:44:35 -07006496 * Prepare kswapd for sleeping. This verifies that there are no processes
6497 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6498 *
6499 * Returns true if kswapd is ready to sleep
6500 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006501static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6502 int highest_zoneidx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08006503{
Mel Gorman55150612012-07-31 16:44:35 -07006504 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08006505 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07006506 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08006507 * race between when kswapd checks the watermarks and a process gets
6508 * throttled. There is also a potential race if processes get
6509 * throttled, kswapd wakes, a large process exits thereby balancing the
6510 * zones, which causes kswapd to exit balance_pgdat() before reaching
6511 * the wake up checks. If kswapd is going to sleep, no process should
6512 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6513 * the wake up is premature, processes will wake kswapd and get
6514 * throttled again. The difference from wake ups in balance_pgdat() is
6515 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07006516 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08006517 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6518 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08006519
Johannes Weinerc73322d2017-05-03 14:51:51 -07006520 /* Hopeless node, leave it to direct reclaim */
6521 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6522 return true;
6523
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006524 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
Mel Gormane716f2e2017-05-03 14:53:45 -07006525 clear_pgdat_congested(pgdat);
6526 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07006527 }
6528
Shantanu Goel333b0a42017-05-03 14:53:38 -07006529 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08006530}
6531
Linus Torvalds1da177e2005-04-16 15:20:36 -07006532/*
Mel Gorman1d82de62016-07-28 15:45:43 -07006533 * kswapd shrinks a node of pages that are at or below the highest usable
6534 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07006535 *
6536 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07006537 * reclaim or if the lack of progress was due to pages under writeback.
6538 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07006539 */
Mel Gorman1d82de62016-07-28 15:45:43 -07006540static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07006541 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07006542{
Mel Gorman1d82de62016-07-28 15:45:43 -07006543 struct zone *zone;
6544 int z;
Mel Gorman75485362013-07-03 15:01:42 -07006545
Mel Gorman1d82de62016-07-28 15:45:43 -07006546 /* Reclaim a number of pages proportional to the number of zones */
6547 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07006548 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07006549 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07006550 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07006551 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07006552
Mel Gorman1d82de62016-07-28 15:45:43 -07006553 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07006554 }
6555
Mel Gorman1d82de62016-07-28 15:45:43 -07006556 /*
6557 * Historically care was taken to put equal pressure on all zones but
6558 * now pressure is applied based on node LRU order.
6559 */
Mel Gorman970a39a2016-07-28 15:46:35 -07006560 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07006561
6562 /*
6563 * Fragmentation may mean that the system cannot be rebalanced for
6564 * high-order allocations. If twice the allocation size has been
6565 * reclaimed then recheck watermarks only at order-0 to prevent
6566 * excessive reclaim. Assume that a process requested a high-order
6567 * can direct reclaim/compact.
6568 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07006569 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07006570 sc->order = 0;
6571
Mel Gormanb8e83b92013-07-03 15:01:45 -07006572 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07006573}
6574
Mel Gormanc49c2c42021-06-28 19:42:21 -07006575/* Page allocator PCP high watermark is lowered if reclaim is active. */
6576static inline void
6577update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6578{
6579 int i;
6580 struct zone *zone;
6581
6582 for (i = 0; i <= highest_zoneidx; i++) {
6583 zone = pgdat->node_zones + i;
6584
6585 if (!managed_zone(zone))
6586 continue;
6587
6588 if (active)
6589 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6590 else
6591 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6592 }
6593}
6594
6595static inline void
6596set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6597{
6598 update_reclaim_active(pgdat, highest_zoneidx, true);
6599}
6600
6601static inline void
6602clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6603{
6604 update_reclaim_active(pgdat, highest_zoneidx, false);
6605}
6606
Mel Gorman75485362013-07-03 15:01:42 -07006607/*
Mel Gorman1d82de62016-07-28 15:45:43 -07006608 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6609 * that are eligible for use by the caller until at least one zone is
6610 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006611 *
Mel Gorman1d82de62016-07-28 15:45:43 -07006612 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006613 *
6614 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07006615 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Wei Yang8bb4e7a2019-03-05 15:46:22 -08006616 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
Mel Gorman1d82de62016-07-28 15:45:43 -07006617 * or lower is eligible for reclaim until at least one usable zone is
6618 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006620static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006621{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006622 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07006623 unsigned long nr_soft_reclaimed;
6624 unsigned long nr_soft_scanned;
Johannes Weinereb414682018-10-26 15:06:27 -07006625 unsigned long pflags;
Mel Gorman1c308442018-12-28 00:35:52 -08006626 unsigned long nr_boost_reclaim;
6627 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6628 bool boosted;
Mel Gorman1d82de62016-07-28 15:45:43 -07006629 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08006630 struct scan_control sc = {
6631 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07006632 .order = order,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07006633 .may_unmap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08006634 };
Omar Sandoval93781322018-06-07 17:07:02 -07006635
Andrew Morton1732d2b012019-07-16 16:26:15 -07006636 set_task_reclaim_state(current, &sc.reclaim_state);
Johannes Weinereb414682018-10-26 15:06:27 -07006637 psi_memstall_enter(&pflags);
Matthew Wilcox (Oracle)4f3eaf42021-09-02 14:52:58 -07006638 __fs_reclaim_acquire(_THIS_IP_);
Omar Sandoval93781322018-06-07 17:07:02 -07006639
Christoph Lameterf8891e52006-06-30 01:55:45 -07006640 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006641
Mel Gorman1c308442018-12-28 00:35:52 -08006642 /*
6643 * Account for the reclaim boost. Note that the zone boost is left in
6644 * place so that parallel allocations that are near the watermark will
6645 * stall or direct reclaim until kswapd is finished.
6646 */
6647 nr_boost_reclaim = 0;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006648 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08006649 zone = pgdat->node_zones + i;
6650 if (!managed_zone(zone))
6651 continue;
6652
6653 nr_boost_reclaim += zone->watermark_boost;
6654 zone_boosts[i] = zone->watermark_boost;
6655 }
6656 boosted = nr_boost_reclaim;
6657
6658restart:
Mel Gormanc49c2c42021-06-28 19:42:21 -07006659 set_reclaim_active(pgdat, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08006660 sc.priority = DEF_PRIORITY;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07006661 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07006662 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07006663 bool raise_priority = true;
Mel Gorman1c308442018-12-28 00:35:52 -08006664 bool balanced;
Omar Sandoval93781322018-06-07 17:07:02 -07006665 bool ret;
Mel Gormanb8e83b92013-07-03 15:01:45 -07006666
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006667 sc.reclaim_idx = highest_zoneidx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006668
Mel Gorman86c79f62016-07-28 15:45:59 -07006669 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07006670 * If the number of buffer_heads exceeds the maximum allowed
6671 * then consider reclaiming from all zones. This has a dual
6672 * purpose -- on 64-bit systems it is expected that
6673 * buffer_heads are stripped during active rotation. On 32-bit
6674 * systems, highmem pages can pin lowmem memory and shrinking
6675 * buffers can relieve lowmem pressure. Reclaim may still not
6676 * go ahead if all eligible zones for the original allocation
6677 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07006678 */
6679 if (buffer_heads_over_limit) {
6680 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6681 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07006682 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07006683 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006684
Mel Gorman970a39a2016-07-28 15:46:35 -07006685 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08006686 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006687 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006688 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08006689
Mel Gorman86c79f62016-07-28 15:45:59 -07006690 /*
Mel Gorman1c308442018-12-28 00:35:52 -08006691 * If the pgdat is imbalanced then ignore boosting and preserve
6692 * the watermarks for a later time and restart. Note that the
6693 * zone watermarks will be still reset at the end of balancing
6694 * on the grounds that the normal reclaim should be enough to
6695 * re-evaluate if boosting is required when kswapd next wakes.
Mel Gorman86c79f62016-07-28 15:45:59 -07006696 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006697 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08006698 if (!balanced && nr_boost_reclaim) {
6699 nr_boost_reclaim = 0;
6700 goto restart;
6701 }
6702
6703 /*
6704 * If boosting is not active then only reclaim if there are no
6705 * eligible zones. Note that sc.reclaim_idx is not used as
6706 * buffer_heads_over_limit may have adjusted it.
6707 */
6708 if (!nr_boost_reclaim && balanced)
Mel Gormane716f2e2017-05-03 14:53:45 -07006709 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08006710
Mel Gorman1c308442018-12-28 00:35:52 -08006711 /* Limit the priority of boosting to avoid reclaim writeback */
6712 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6713 raise_priority = false;
6714
6715 /*
6716 * Do not writeback or swap pages for boosted reclaim. The
6717 * intent is to relieve pressure not issue sub-optimal IO
6718 * from reclaim context. If no pages are reclaimed, the
6719 * reclaim will be aborted.
6720 */
6721 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6722 sc.may_swap = !nr_boost_reclaim;
Mel Gorman1c308442018-12-28 00:35:52 -08006723
Linus Torvalds1da177e2005-04-16 15:20:36 -07006724 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07006725 * Do some background aging of the anon list, to give
6726 * pages a chance to be referenced before reclaiming. All
6727 * pages are rotated regardless of classzone as this is
6728 * about consistent aging.
6729 */
Mel Gormanef8f2322016-07-28 15:46:05 -07006730 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07006731
6732 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07006733 * If we're getting trouble reclaiming, start doing writepage
6734 * even in laptop mode.
6735 */
Johannes Weiner047d72c2017-05-03 14:51:57 -07006736 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07006737 sc.may_writepage = 1;
6738
Mel Gorman1d82de62016-07-28 15:45:43 -07006739 /* Call soft limit reclaim before calling shrink_node. */
6740 sc.nr_scanned = 0;
6741 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07006742 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07006743 sc.gfp_mask, &nr_soft_scanned);
6744 sc.nr_reclaimed += nr_soft_reclaimed;
6745
Mel Gormanb7ea3c42013-07-03 15:01:53 -07006746 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07006747 * There should be no need to raise the scanning priority if
6748 * enough pages are already being scanned that that high
6749 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006750 */
Mel Gorman970a39a2016-07-28 15:46:35 -07006751 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07006752 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07006753
6754 /*
6755 * If the low watermark is met there is no need for processes
6756 * to be throttled on pfmemalloc_wait as they should not be
6757 * able to safely make forward progress. Wake them
6758 */
6759 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc73322d2017-05-03 14:51:51 -07006760 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08006761 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07006762
Mel Gormanb8e83b92013-07-03 15:01:45 -07006763 /* Check if kswapd should be suspending */
Matthew Wilcox (Oracle)4f3eaf42021-09-02 14:52:58 -07006764 __fs_reclaim_release(_THIS_IP_);
Omar Sandoval93781322018-06-07 17:07:02 -07006765 ret = try_to_freeze();
Matthew Wilcox (Oracle)4f3eaf42021-09-02 14:52:58 -07006766 __fs_reclaim_acquire(_THIS_IP_);
Omar Sandoval93781322018-06-07 17:07:02 -07006767 if (ret || kthread_should_stop())
Mel Gormanb8e83b92013-07-03 15:01:45 -07006768 break;
6769
6770 /*
6771 * Raise priority if scanning rate is too low or there was no
6772 * progress in reclaiming pages
6773 */
Johannes Weinerc73322d2017-05-03 14:51:51 -07006774 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
Mel Gorman1c308442018-12-28 00:35:52 -08006775 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6776
6777 /*
6778 * If reclaim made no progress for a boost, stop reclaim as
6779 * IO cannot be queued and it could be an infinite loop in
6780 * extreme circumstances.
6781 */
6782 if (nr_boost_reclaim && !nr_reclaimed)
6783 break;
6784
Johannes Weinerc73322d2017-05-03 14:51:51 -07006785 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07006786 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07006787 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006788
Johannes Weinerc73322d2017-05-03 14:51:51 -07006789 if (!sc.nr_reclaimed)
6790 pgdat->kswapd_failures++;
6791
Mel Gormanb8e83b92013-07-03 15:01:45 -07006792out:
Mel Gormanc49c2c42021-06-28 19:42:21 -07006793 clear_reclaim_active(pgdat, highest_zoneidx);
6794
Mel Gorman1c308442018-12-28 00:35:52 -08006795 /* If reclaim was boosted, account for the reclaim done in this pass */
6796 if (boosted) {
6797 unsigned long flags;
6798
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006799 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08006800 if (!zone_boosts[i])
6801 continue;
6802
6803 /* Increments are under the zone lock */
6804 zone = pgdat->node_zones + i;
6805 spin_lock_irqsave(&zone->lock, flags);
6806 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6807 spin_unlock_irqrestore(&zone->lock, flags);
6808 }
6809
6810 /*
6811 * As there is now likely space, wakeup kcompact to defragment
6812 * pageblocks.
6813 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006814 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08006815 }
6816
Johannes Weiner2a2e4882017-05-03 14:55:03 -07006817 snapshot_refaults(NULL, pgdat);
Matthew Wilcox (Oracle)4f3eaf42021-09-02 14:52:58 -07006818 __fs_reclaim_release(_THIS_IP_);
Johannes Weinereb414682018-10-26 15:06:27 -07006819 psi_memstall_leave(&pflags);
Andrew Morton1732d2b012019-07-16 16:26:15 -07006820 set_task_reclaim_state(current, NULL);
Yafang Shaoe5ca8072019-07-16 16:26:09 -07006821
Mel Gorman0abdee22011-01-13 15:46:22 -08006822 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07006823 * Return the order kswapd stopped reclaiming at as
6824 * prepare_kswapd_sleep() takes it into account. If another caller
6825 * entered the allocator slow path while kswapd was awake, order will
6826 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08006827 */
Mel Gorman1d82de62016-07-28 15:45:43 -07006828 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006829}
6830
Mel Gormane716f2e2017-05-03 14:53:45 -07006831/*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006832 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6833 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6834 * not a valid index then either kswapd runs for first time or kswapd couldn't
6835 * sleep after previous reclaim attempt (node is still unbalanced). In that
6836 * case return the zone index of the previous kswapd reclaim cycle.
Mel Gormane716f2e2017-05-03 14:53:45 -07006837 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006838static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6839 enum zone_type prev_highest_zoneidx)
Mel Gormane716f2e2017-05-03 14:53:45 -07006840{
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006841 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07006842
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006843 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
Mel Gormane716f2e2017-05-03 14:53:45 -07006844}
6845
Mel Gorman38087d92016-07-28 15:45:49 -07006846static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006847 unsigned int highest_zoneidx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08006848{
6849 long remaining = 0;
6850 DEFINE_WAIT(wait);
6851
6852 if (freezing(current) || kthread_should_stop())
6853 return;
6854
6855 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
6856
Shantanu Goel333b0a42017-05-03 14:53:38 -07006857 /*
6858 * Try to sleep for a short interval. Note that kcompactd will only be
6859 * woken if it is possible to sleep for a short interval. This is
6860 * deliberate on the assumption that if reclaim cannot keep an
6861 * eligible zone balanced that it's also unlikely that compaction will
6862 * succeed.
6863 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006864 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07006865 /*
6866 * Compaction records what page blocks it recently failed to
6867 * isolate pages from and skips them in the future scanning.
6868 * When kswapd is going to sleep, it is reasonable to assume
6869 * that pages and compaction may succeed so reset the cache.
6870 */
6871 reset_isolation_suitable(pgdat);
6872
6873 /*
6874 * We have freed the memory, now we should compact it to make
6875 * allocation of the requested order possible.
6876 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006877 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07006878
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08006879 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07006880
6881 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006882 * If woken prematurely then reset kswapd_highest_zoneidx and
Mel Gorman38087d92016-07-28 15:45:49 -07006883 * order. The values will either be from a wakeup request or
6884 * the previous request that slept prematurely.
6885 */
6886 if (remaining) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006887 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
6888 kswapd_highest_zoneidx(pgdat,
6889 highest_zoneidx));
Qian Cai5644e1fb2020-04-01 21:10:12 -07006890
6891 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
6892 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
Mel Gorman38087d92016-07-28 15:45:49 -07006893 }
6894
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08006895 finish_wait(&pgdat->kswapd_wait, &wait);
6896 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
6897 }
6898
6899 /*
6900 * After a short sleep, check if it was a premature sleep. If not, then
6901 * go fully to sleep until explicitly woken up.
6902 */
Mel Gormand9f21d42016-07-28 15:46:41 -07006903 if (!remaining &&
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006904 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08006905 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
6906
6907 /*
6908 * vmstat counters are not perfectly accurate and the estimated
6909 * value for counters such as NR_FREE_PAGES can deviate from the
6910 * true value by nr_online_cpus * threshold. To avoid the zone
6911 * watermarks being breached while under pressure, we reduce the
6912 * per-cpu vmstat threshold while kswapd is awake and restore
6913 * them before going back to sleep.
6914 */
6915 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07006916
6917 if (!kthread_should_stop())
6918 schedule();
6919
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08006920 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
6921 } else {
6922 if (remaining)
6923 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
6924 else
6925 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
6926 }
6927 finish_wait(&pgdat->kswapd_wait, &wait);
6928}
6929
Linus Torvalds1da177e2005-04-16 15:20:36 -07006930/*
6931 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07006932 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006933 *
6934 * This basically trickles out pages so that we have _some_
6935 * free memory available even if there is no other activity
6936 * that frees anything up. This is needed for things like routing
6937 * etc, where we otherwise might have all activity going on in
6938 * asynchronous contexts that cannot page things out.
6939 *
6940 * If there are applications that are active memory-allocators
6941 * (most normal use), this basically shouldn't matter.
6942 */
Vijayanand Jitta12972dd2022-03-23 12:37:28 +05306943int kswapd(void *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006944{
Mel Gormane716f2e2017-05-03 14:53:45 -07006945 unsigned int alloc_order, reclaim_order;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006946 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
Zhiyuan Dai68d68ff2021-05-04 18:40:12 -07006947 pg_data_t *pgdat = (pg_data_t *)p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006948 struct task_struct *tsk = current;
Rusty Russella70f7302009-03-13 14:49:46 +10306949 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006950
Rusty Russell174596a2009-01-01 10:12:29 +10306951 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07006952 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006953
6954 /*
6955 * Tell the memory management that we're a "memory allocator",
6956 * and that if we need more memory we should get access to it
6957 * regardless (see "__alloc_pages()"). "kswapd" should
6958 * never get caught in the normal page freeing logic.
6959 *
6960 * (Kswapd normally doesn't need memory anyway, but sometimes
6961 * you need a small amount of memory in order to be able to
6962 * page out something else, and this flag essentially protects
6963 * us from recursively trying to free more memory as we're
6964 * trying to free the first piece of memory in the first place).
6965 */
Christoph Lameter930d9152006-01-08 01:00:47 -08006966 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07006967 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006968
Qian Cai5644e1fb2020-04-01 21:10:12 -07006969 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006970 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006971 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08006972 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07006973
Qian Cai5644e1fb2020-04-01 21:10:12 -07006974 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006975 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
6976 highest_zoneidx);
Mel Gormane716f2e2017-05-03 14:53:45 -07006977
Mel Gorman38087d92016-07-28 15:45:49 -07006978kswapd_try_sleep:
6979 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006980 highest_zoneidx);
Mel Gorman215ddd62011-07-08 15:39:40 -07006981
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006982 /* Read the new order and highest_zoneidx */
Lukas Bulwahn2b47a242020-12-14 19:12:18 -08006983 alloc_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006984 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
6985 highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07006986 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07006987 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006988
David Rientjes8fe23e02009-12-14 17:58:33 -08006989 ret = try_to_freeze();
6990 if (kthread_should_stop())
6991 break;
6992
6993 /*
6994 * We can speed up thawing tasks if we don't call balance_pgdat
6995 * after returning from the refrigerator
6996 */
Mel Gorman38087d92016-07-28 15:45:49 -07006997 if (ret)
6998 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07006999
Mel Gorman38087d92016-07-28 15:45:49 -07007000 /*
7001 * Reclaim begins at the requested order but if a high-order
7002 * reclaim fails then kswapd falls back to reclaiming for
7003 * order-0. If that happens, kswapd will consider sleeping
7004 * for the order it finished reclaiming at (reclaim_order)
7005 * but kcompactd is woken to compact for the original
7006 * request (alloc_order).
7007 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007008 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
Mel Gormane5146b12016-07-28 15:46:47 -07007009 alloc_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007010 reclaim_order = balance_pgdat(pgdat, alloc_order,
7011 highest_zoneidx);
Mel Gorman38087d92016-07-28 15:45:49 -07007012 if (reclaim_order < alloc_order)
7013 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007014 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08007015
Johannes Weiner71abdc12014-06-06 14:35:35 -07007016 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Johannes Weiner71abdc12014-06-06 14:35:35 -07007017
Linus Torvalds1da177e2005-04-16 15:20:36 -07007018 return 0;
7019}
Vijayanand Jitta12972dd2022-03-23 12:37:28 +05307020EXPORT_SYMBOL_GPL(kswapd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007021
7022/*
David Rientjes5ecd9d42018-04-05 16:25:16 -07007023 * A zone is low on free memory or too fragmented for high-order memory. If
7024 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7025 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7026 * has failed or is not needed, still wake up kcompactd if only compaction is
7027 * needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007028 */
David Rientjes5ecd9d42018-04-05 16:25:16 -07007029void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007030 enum zone_type highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007031{
7032 pg_data_t *pgdat;
Qian Cai5644e1fb2020-04-01 21:10:12 -07007033 enum zone_type curr_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007034
Mel Gorman6aa303d2016-09-01 16:14:55 -07007035 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007036 return;
7037
David Rientjes5ecd9d42018-04-05 16:25:16 -07007038 if (!cpuset_zone_allowed(zone, gfp_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007039 return;
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07007040
Qian Cai5644e1fb2020-04-01 21:10:12 -07007041 pgdat = zone->zone_pgdat;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007042 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07007043
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007044 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7045 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07007046
7047 if (READ_ONCE(pgdat->kswapd_order) < order)
7048 WRITE_ONCE(pgdat->kswapd_order, order);
7049
Con Kolivas8d0986e2005-09-13 01:25:07 -07007050 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007051 return;
Mel Gormane1a55632016-07-28 15:46:26 -07007052
David Rientjes5ecd9d42018-04-05 16:25:16 -07007053 /* Hopeless node, leave it to direct reclaim if possible */
7054 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007055 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7056 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
David Rientjes5ecd9d42018-04-05 16:25:16 -07007057 /*
7058 * There may be plenty of free memory available, but it's too
7059 * fragmented for high-order allocations. Wake up kcompactd
7060 * and rely on compaction_suitable() to determine if it's
7061 * needed. If it fails, it will defer subsequent attempts to
7062 * ratelimit its work.
7063 */
7064 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007065 wakeup_kcompactd(pgdat, order, highest_zoneidx);
Johannes Weinerc73322d2017-05-03 14:51:51 -07007066 return;
David Rientjes5ecd9d42018-04-05 16:25:16 -07007067 }
Johannes Weinerc73322d2017-05-03 14:51:51 -07007068
Joonsoo Kim97a225e2020-06-03 15:59:01 -07007069 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
David Rientjes5ecd9d42018-04-05 16:25:16 -07007070 gfp_flags);
Con Kolivas8d0986e2005-09-13 01:25:07 -07007071 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007072}
7073
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02007074#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07007075/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08007076 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07007077 * freed pages.
7078 *
7079 * Rather than trying to age LRUs the aim is to preserve the overall
7080 * LRU order by reclaiming preferentially
7081 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07007082 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08007083unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007084{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07007085 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08007086 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07007087 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07007088 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07007089 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07007090 .may_writepage = 1,
7091 .may_unmap = 1,
7092 .may_swap = 1,
7093 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007094 };
Ying Hana09ed5e2011-05-24 17:12:26 -07007095 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08007096 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07007097 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007098
Peter Zijlstrad92a8cf2017-03-03 10:13:38 +01007099 fs_reclaim_acquire(sc.gfp_mask);
Omar Sandoval93781322018-06-07 17:07:02 -07007100 noreclaim_flag = memalloc_noreclaim_save();
Andrew Morton1732d2b012019-07-16 16:26:15 -07007101 set_task_reclaim_state(current, &sc.reclaim_state);
Andrew Morton69e05942006-03-22 00:08:19 -08007102
Vladimir Davydov3115cd92014-04-03 14:47:22 -07007103 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07007104
Andrew Morton1732d2b012019-07-16 16:26:15 -07007105 set_task_reclaim_state(current, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07007106 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07007107 fs_reclaim_release(sc.gfp_mask);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07007108
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08007109 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007110}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02007111#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007112
Yasunori Goto3218ae12006-06-27 02:53:33 -07007113/*
7114 * This kswapd start function will be called by init and node-hot-add.
7115 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
7116 */
Miaohe Linb87c517a2021-09-02 14:59:46 -07007117void kswapd_run(int nid)
Yasunori Goto3218ae12006-06-27 02:53:33 -07007118{
7119 pg_data_t *pgdat = NODE_DATA(nid);
Charan Teja Reddyd831f072021-02-05 17:47:57 +05307120 bool skip = false;
Yasunori Goto3218ae12006-06-27 02:53:33 -07007121
7122 if (pgdat->kswapd)
Miaohe Linb87c517a2021-09-02 14:59:46 -07007123 return;
Yasunori Goto3218ae12006-06-27 02:53:33 -07007124
Charan Teja Reddyd831f072021-02-05 17:47:57 +05307125 trace_android_vh_kswapd_per_node(nid, &skip, true);
7126 if (skip)
7127 return;
Yasunori Goto3218ae12006-06-27 02:53:33 -07007128 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7129 if (IS_ERR(pgdat->kswapd)) {
7130 /* failure at boot is fatal */
Thomas Gleixnerc6202ad2017-05-16 20:42:46 +02007131 BUG_ON(system_state < SYSTEM_RUNNING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07007132 pr_err("Failed to start kswapd on node %d\n", nid);
Xishi Qiud72515b2013-04-17 15:58:34 -07007133 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07007134 }
Yasunori Goto3218ae12006-06-27 02:53:33 -07007135}
7136
David Rientjes8fe23e02009-12-14 17:58:33 -08007137/*
Jiang Liud8adde12012-07-11 14:01:52 -07007138 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07007139 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08007140 */
7141void kswapd_stop(int nid)
7142{
7143 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
Charan Teja Reddyd831f072021-02-05 17:47:57 +05307144 bool skip = false;
David Rientjes8fe23e02009-12-14 17:58:33 -08007145
Charan Teja Reddyd831f072021-02-05 17:47:57 +05307146 trace_android_vh_kswapd_per_node(nid, &skip, false);
7147 if (skip)
7148 return;
Jiang Liud8adde12012-07-11 14:01:52 -07007149 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08007150 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07007151 NODE_DATA(nid)->kswapd = NULL;
7152 }
David Rientjes8fe23e02009-12-14 17:58:33 -08007153}
7154
Linus Torvalds1da177e2005-04-16 15:20:36 -07007155static int __init kswapd_init(void)
7156{
Wei Yang6b700b52020-04-01 21:10:09 -07007157 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08007158
Linus Torvalds1da177e2005-04-16 15:20:36 -07007159 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08007160 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07007161 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007162 return 0;
7163}
7164
7165module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08007166
7167#ifdef CONFIG_NUMA
7168/*
Mel Gormana5f5f912016-07-28 15:46:32 -07007169 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08007170 *
Mel Gormana5f5f912016-07-28 15:46:32 -07007171 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08007172 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08007173 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007174int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08007175
Dave Hansen51998362021-02-24 12:09:15 -08007176/*
Mel Gormana5f5f912016-07-28 15:46:32 -07007177 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08007178 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7179 * a zone.
7180 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007181#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08007182
Christoph Lameter9eeff232006-01-18 17:42:31 -08007183/*
Mel Gormana5f5f912016-07-28 15:46:32 -07007184 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07007185 * occur.
7186 */
7187int sysctl_min_unmapped_ratio = 1;
7188
7189/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07007190 * If the number of slab pages in a zone grows beyond this percentage then
7191 * slab reclaim needs to occur.
7192 */
7193int sysctl_min_slab_ratio = 5;
7194
Mel Gorman11fb9982016-07-28 15:46:20 -07007195static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07007196{
Mel Gorman11fb9982016-07-28 15:46:20 -07007197 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7198 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7199 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07007200
7201 /*
7202 * It's possible for there to be more file mapped pages than
7203 * accounted for by the pages on the file LRU lists because
7204 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7205 */
7206 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7207}
7208
7209/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07007210static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07007211{
Alexandru Moised031a152015-11-05 18:48:08 -08007212 unsigned long nr_pagecache_reclaimable;
7213 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07007214
7215 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07007216 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07007217 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07007218 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07007219 * a better estimate
7220 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007221 if (node_reclaim_mode & RECLAIM_UNMAP)
7222 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07007223 else
Mel Gormana5f5f912016-07-28 15:46:32 -07007224 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07007225
7226 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07007227 if (!(node_reclaim_mode & RECLAIM_WRITE))
7228 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07007229
7230 /* Watch for any possible underflows due to delta */
7231 if (unlikely(delta > nr_pagecache_reclaimable))
7232 delta = nr_pagecache_reclaimable;
7233
7234 return nr_pagecache_reclaimable - delta;
7235}
7236
Christoph Lameter0ff38492006-09-25 23:31:52 -07007237/*
Mel Gormana5f5f912016-07-28 15:46:32 -07007238 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08007239 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007240static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08007241{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08007242 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08007243 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08007244 struct task_struct *p = current;
Vlastimil Babka499118e2017-05-08 15:59:50 -07007245 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08007246 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08007247 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07007248 .gfp_mask = current_gfp_context(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07007249 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07007250 .priority = NODE_RECLAIM_PRIORITY,
7251 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7252 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07007253 .may_swap = 1,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07007254 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08007255 };
Johannes Weiner57f29762021-08-19 19:04:27 -07007256 unsigned long pflags;
Christoph Lameter9eeff232006-01-18 17:42:31 -08007257
Yafang Shao132bb8c2019-05-13 17:17:53 -07007258 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7259 sc.gfp_mask);
7260
Christoph Lameter9eeff232006-01-18 17:42:31 -08007261 cond_resched();
Johannes Weiner57f29762021-08-19 19:04:27 -07007262 psi_memstall_enter(&pflags);
Omar Sandoval93781322018-06-07 17:07:02 -07007263 fs_reclaim_acquire(sc.gfp_mask);
Christoph Lameterd4f77962006-02-24 13:04:22 -08007264 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07007265 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08007266 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07007267 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08007268 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07007269 noreclaim_flag = memalloc_noreclaim_save();
7270 p->flags |= PF_SWAPWRITE;
Andrew Morton1732d2b012019-07-16 16:26:15 -07007271 set_task_reclaim_state(p, &sc.reclaim_state);
Christoph Lameterc84db232006-02-01 03:05:29 -08007272
Mel Gormana5f5f912016-07-28 15:46:32 -07007273 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07007274 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07007275 * Free memory by calling shrink node with increasing
Christoph Lameter0ff38492006-09-25 23:31:52 -07007276 * priorities until we have enough memory freed.
7277 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07007278 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07007279 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07007280 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07007281 }
Christoph Lameterc84db232006-02-01 03:05:29 -08007282
Andrew Morton1732d2b012019-07-16 16:26:15 -07007283 set_task_reclaim_state(p, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07007284 current->flags &= ~PF_SWAPWRITE;
7285 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07007286 fs_reclaim_release(sc.gfp_mask);
Johannes Weiner57f29762021-08-19 19:04:27 -07007287 psi_memstall_leave(&pflags);
Yafang Shao132bb8c2019-05-13 17:17:53 -07007288
7289 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7290
Rik van Riela79311c2009-01-06 14:40:01 -08007291 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08007292}
Andrew Morton179e9632006-03-22 00:08:18 -08007293
Mel Gormana5f5f912016-07-28 15:46:32 -07007294int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08007295{
David Rientjesd773ed62007-10-16 23:26:01 -07007296 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08007297
7298 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07007299 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07007300 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07007301 *
Christoph Lameter96146342006-07-03 00:24:13 -07007302 * A small portion of unmapped file backed pages is needed for
7303 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07007304 * thrown out if the node is overallocated. So we do not reclaim
7305 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07007306 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08007307 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007308 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
Roman Gushchind42f3242020-08-06 23:20:39 -07007309 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7310 pgdat->min_slab_pages)
Mel Gormana5f5f912016-07-28 15:46:32 -07007311 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08007312
7313 /*
David Rientjesd773ed62007-10-16 23:26:01 -07007314 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08007315 */
Mel Gormand0164ad2015-11-06 16:28:21 -08007316 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07007317 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08007318
7319 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07007320 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08007321 * have associated processors. This will favor the local processor
7322 * over remote processors and spread off node memory allocations
7323 * as wide as possible.
7324 */
Mel Gormana5f5f912016-07-28 15:46:32 -07007325 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7326 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07007327
Mel Gormana5f5f912016-07-28 15:46:32 -07007328 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7329 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07007330
Mel Gormana5f5f912016-07-28 15:46:32 -07007331 ret = __node_reclaim(pgdat, gfp_mask, order);
7332 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07007333
Mel Gorman24cf725182009-06-16 15:33:23 -07007334 if (!ret)
7335 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7336
David Rientjesd773ed62007-10-16 23:26:01 -07007337 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08007338}
Christoph Lameter9eeff232006-01-18 17:42:31 -08007339#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07007340
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007341/**
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00007342 * check_move_unevictable_pages - check pages for evictability and move to
7343 * appropriate zone lru list
7344 * @pvec: pagevec with lru pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007345 *
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00007346 * Checks pages for evictability, if an evictable page is in the unevictable
7347 * lru list, moves it to the appropriate evictable lru list. This function
7348 * should be only used for lru pages.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007349 */
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00007350void check_move_unevictable_pages(struct pagevec *pvec)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007351{
Alex Shi6168d0d2020-12-15 12:34:29 -08007352 struct lruvec *lruvec = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08007353 int pgscanned = 0;
7354 int pgrescued = 0;
7355 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007356
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00007357 for (i = 0; i < pvec->nr; i++) {
7358 struct page *page = pvec->pages[i];
Hugh Dickins8d8869c2020-09-18 21:20:12 -07007359 int nr_pages;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07007360
Hugh Dickins8d8869c2020-09-18 21:20:12 -07007361 if (PageTransTail(page))
7362 continue;
7363
7364 nr_pages = thp_nr_pages(page);
7365 pgscanned += nr_pages;
7366
Alex Shid25b5bd2020-12-15 12:34:16 -08007367 /* block memcg migration during page moving between lru */
7368 if (!TestClearPageLRU(page))
7369 continue;
7370
Alexander Duyck2a5e4e32020-12-15 12:34:33 -08007371 lruvec = relock_page_lruvec_irq(page, lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08007372 if (page_evictable(page) && PageUnevictable(page)) {
Yu Zhao46ae6b22021-02-24 12:08:25 -08007373 del_page_from_lru_list(page, lruvec);
Hugh Dickins24513262012-01-20 14:34:21 -08007374 ClearPageUnevictable(page);
Yu Zhao3a9c9782021-02-24 12:08:17 -08007375 add_page_to_lru_list(page, lruvec);
Hugh Dickins8d8869c2020-09-18 21:20:12 -07007376 pgrescued += nr_pages;
Hugh Dickins24513262012-01-20 14:34:21 -08007377 }
Alex Shid25b5bd2020-12-15 12:34:16 -08007378 SetPageLRU(page);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07007379 }
Hugh Dickins24513262012-01-20 14:34:21 -08007380
Alex Shi6168d0d2020-12-15 12:34:29 -08007381 if (lruvec) {
Hugh Dickins24513262012-01-20 14:34:21 -08007382 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7383 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Alex Shi6168d0d2020-12-15 12:34:29 -08007384 unlock_page_lruvec_irq(lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08007385 } else if (pgscanned) {
7386 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Hugh Dickins24513262012-01-20 14:34:21 -08007387 }
Hugh Dickins850465792012-01-20 14:34:19 -08007388}
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00007389EXPORT_SYMBOL_GPL(check_move_unevictable_pages);