blob: 5199b9696babf7a44c1dc500d63e2d2aefe23986 [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>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +000047#include <linux/pagevec.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070048#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070049#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080050#include <linux/dax.h>
Johannes Weinereb414682018-10-26 15:06:27 -070051#include <linux/psi.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
53#include <asm/tlbflush.h>
54#include <asm/div64.h>
55
56#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070057#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
Nick Piggin0f8053a2006-03-22 00:08:33 -080059#include "internal.h"
60
Mel Gorman33906bc2010-08-09 17:19:16 -070061#define CREATE_TRACE_POINTS
62#include <trace/events/vmscan.h>
63
Linus Torvalds1da177e2005-04-16 15:20:36 -070064struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080065 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
Johannes Weineree814fe2014-08-06 16:06:19 -070068 /*
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
70 * are scanned.
71 */
72 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070073
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070074 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080075 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
77 */
78 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080079
Johannes Weiner7cf111b2020-06-03 16:03:06 -070080 /*
81 * Scan pressure balancing between anon and file LRUs
82 */
83 unsigned long anon_cost;
84 unsigned long file_cost;
85
Johannes Weinerb91ac372019-11-30 17:56:02 -080086 /* Can active pages be deactivated as part of reclaim? */
87#define DEACTIVATE_ANON 1
88#define DEACTIVATE_FILE 2
89 unsigned int may_deactivate:2;
90 unsigned int force_deactivate:1;
91 unsigned int skipped_deactivate:1;
92
Johannes Weiner1276ad62017-02-24 14:56:11 -080093 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -070094 unsigned int may_writepage:1;
95
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap:1;
98
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap:1;
101
Yisheng Xied6622f62017-05-03 14:53:57 -0700102 /*
103 * Cgroups are not reclaimed below their configured memory.low,
104 * unless we threaten to OOM. If any cgroups are skipped due to
105 * memory.low and nothing was reclaimed, go back for memory.low.
106 */
107 unsigned int memcg_low_reclaim:1;
108 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -0800109
Johannes Weineree814fe2014-08-06 16:06:19 -0700110 unsigned int hibernation_mode:1;
111
112 /* One of the zones is ready for compaction */
113 unsigned int compaction_ready:1;
114
Johannes Weinerb91ac372019-11-30 17:56:02 -0800115 /* There is easily reclaimable cold cache in the current node */
116 unsigned int cache_trim_mode:1;
117
Johannes Weiner53138ce2019-11-30 17:55:56 -0800118 /* The file pages on the current node are dangerously low */
119 unsigned int file_is_tiny:1;
120
Greg Thelenbb451fd2018-08-17 15:45:19 -0700121 /* Allocation order */
122 s8 order;
123
124 /* Scan (total_size >> priority) pages at once */
125 s8 priority;
126
127 /* The highest zone to isolate pages for reclaim from */
128 s8 reclaim_idx;
129
130 /* This context's GFP mask */
131 gfp_t gfp_mask;
132
Johannes Weineree814fe2014-08-06 16:06:19 -0700133 /* Incremented by the number of inactive pages that were scanned */
134 unsigned long nr_scanned;
135
136 /* Number of pages freed so far during a call to shrink_zones() */
137 unsigned long nr_reclaimed;
Andrey Ryabinind108c772018-04-10 16:27:59 -0700138
139 struct {
140 unsigned int dirty;
141 unsigned int unqueued_dirty;
142 unsigned int congested;
143 unsigned int writeback;
144 unsigned int immediate;
145 unsigned int file_taken;
146 unsigned int taken;
147 } nr;
Yafang Shaoe5ca8072019-07-16 16:26:09 -0700148
149 /* for recording the reclaimed slab by now */
150 struct reclaim_state reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151};
152
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153#ifdef ARCH_HAS_PREFETCHW
154#define prefetchw_prev_lru_page(_page, _base, _field) \
155 do { \
156 if ((_page)->lru.prev != _base) { \
157 struct page *prev; \
158 \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
161 } \
162 } while (0)
163#else
164#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165#endif
166
167/*
Johannes Weinerc8439662020-06-03 16:02:37 -0700168 * From 0 .. 200. Higher means more swappy.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 */
170int vm_swappiness = 60;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
Yang Shi0a432dcb2019-09-23 15:38:12 -0700172static void set_task_reclaim_state(struct task_struct *task,
173 struct reclaim_state *rs)
174{
175 /* Check for an overwrite */
176 WARN_ON_ONCE(rs && task->reclaim_state);
177
178 /* Check for the nulling of an already-nulled member */
179 WARN_ON_ONCE(!rs && !task->reclaim_state);
180
181 task->reclaim_state = rs;
182}
183
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184static LIST_HEAD(shrinker_list);
185static DECLARE_RWSEM(shrinker_rwsem);
186
Yang Shi0a432dcb2019-09-23 15:38:12 -0700187#ifdef CONFIG_MEMCG
Yang Shia2fb1262021-05-04 18:36:17 -0700188static int shrinker_nr_max;
Yang Shi2bfd3632021-05-04 18:36:11 -0700189
Yang Shi3c6f17e2021-05-04 18:36:33 -0700190/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
Yang Shia2fb1262021-05-04 18:36:17 -0700191static inline int shrinker_map_size(int nr_items)
192{
193 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
194}
Yang Shi2bfd3632021-05-04 18:36:11 -0700195
Yang Shi3c6f17e2021-05-04 18:36:33 -0700196static inline int shrinker_defer_size(int nr_items)
197{
198 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
199}
200
Yang Shi468ab842021-05-04 18:36:26 -0700201static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
202 int nid)
203{
204 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
205 lockdep_is_held(&shrinker_rwsem));
206}
207
Yang Shie4262c42021-05-04 18:36:23 -0700208static int expand_one_shrinker_info(struct mem_cgroup *memcg,
Yang Shi3c6f17e2021-05-04 18:36:33 -0700209 int map_size, int defer_size,
210 int old_map_size, int old_defer_size)
Yang Shi2bfd3632021-05-04 18:36:11 -0700211{
Yang Shie4262c42021-05-04 18:36:23 -0700212 struct shrinker_info *new, *old;
Yang Shi2bfd3632021-05-04 18:36:11 -0700213 struct mem_cgroup_per_node *pn;
214 int nid;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700215 int size = map_size + defer_size;
Yang Shi2bfd3632021-05-04 18:36:11 -0700216
Yang Shi2bfd3632021-05-04 18:36:11 -0700217 for_each_node(nid) {
218 pn = memcg->nodeinfo[nid];
Yang Shi468ab842021-05-04 18:36:26 -0700219 old = shrinker_info_protected(memcg, nid);
Yang Shi2bfd3632021-05-04 18:36:11 -0700220 /* Not yet online memcg */
221 if (!old)
222 return 0;
223
224 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
225 if (!new)
226 return -ENOMEM;
227
Yang Shi3c6f17e2021-05-04 18:36:33 -0700228 new->nr_deferred = (atomic_long_t *)(new + 1);
229 new->map = (void *)new->nr_deferred + defer_size;
230
231 /* map: set all old bits, clear all new bits */
232 memset(new->map, (int)0xff, old_map_size);
233 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
234 /* nr_deferred: copy old values, clear all new values */
235 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
236 memset((void *)new->nr_deferred + old_defer_size, 0,
237 defer_size - old_defer_size);
Yang Shi2bfd3632021-05-04 18:36:11 -0700238
Yang Shie4262c42021-05-04 18:36:23 -0700239 rcu_assign_pointer(pn->shrinker_info, new);
Yang Shi72673e82021-05-04 18:36:20 -0700240 kvfree_rcu(old, rcu);
Yang Shi2bfd3632021-05-04 18:36:11 -0700241 }
242
243 return 0;
244}
245
Yang Shie4262c42021-05-04 18:36:23 -0700246void free_shrinker_info(struct mem_cgroup *memcg)
Yang Shi2bfd3632021-05-04 18:36:11 -0700247{
248 struct mem_cgroup_per_node *pn;
Yang Shie4262c42021-05-04 18:36:23 -0700249 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700250 int nid;
251
Yang Shi2bfd3632021-05-04 18:36:11 -0700252 for_each_node(nid) {
253 pn = memcg->nodeinfo[nid];
Yang Shie4262c42021-05-04 18:36:23 -0700254 info = rcu_dereference_protected(pn->shrinker_info, true);
255 kvfree(info);
256 rcu_assign_pointer(pn->shrinker_info, NULL);
Yang Shi2bfd3632021-05-04 18:36:11 -0700257 }
258}
259
Yang Shie4262c42021-05-04 18:36:23 -0700260int alloc_shrinker_info(struct mem_cgroup *memcg)
Yang Shi2bfd3632021-05-04 18:36:11 -0700261{
Yang Shie4262c42021-05-04 18:36:23 -0700262 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700263 int nid, size, ret = 0;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700264 int map_size, defer_size = 0;
Yang Shi2bfd3632021-05-04 18:36:11 -0700265
Yang Shid27cf2a2021-05-04 18:36:14 -0700266 down_write(&shrinker_rwsem);
Yang Shi3c6f17e2021-05-04 18:36:33 -0700267 map_size = shrinker_map_size(shrinker_nr_max);
268 defer_size = shrinker_defer_size(shrinker_nr_max);
269 size = map_size + defer_size;
Yang Shi2bfd3632021-05-04 18:36:11 -0700270 for_each_node(nid) {
Yang Shie4262c42021-05-04 18:36:23 -0700271 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
272 if (!info) {
273 free_shrinker_info(memcg);
Yang Shi2bfd3632021-05-04 18:36:11 -0700274 ret = -ENOMEM;
275 break;
276 }
Yang Shi3c6f17e2021-05-04 18:36:33 -0700277 info->nr_deferred = (atomic_long_t *)(info + 1);
278 info->map = (void *)info->nr_deferred + defer_size;
Yang Shie4262c42021-05-04 18:36:23 -0700279 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
Yang Shi2bfd3632021-05-04 18:36:11 -0700280 }
Yang Shid27cf2a2021-05-04 18:36:14 -0700281 up_write(&shrinker_rwsem);
Yang Shi2bfd3632021-05-04 18:36:11 -0700282
283 return ret;
284}
285
Yang Shi3c6f17e2021-05-04 18:36:33 -0700286static inline bool need_expand(int nr_max)
287{
288 return round_up(nr_max, BITS_PER_LONG) >
289 round_up(shrinker_nr_max, BITS_PER_LONG);
290}
291
Yang Shie4262c42021-05-04 18:36:23 -0700292static int expand_shrinker_info(int new_id)
Yang Shi2bfd3632021-05-04 18:36:11 -0700293{
Yang Shi3c6f17e2021-05-04 18:36:33 -0700294 int ret = 0;
Yang Shia2fb1262021-05-04 18:36:17 -0700295 int new_nr_max = new_id + 1;
Yang Shi3c6f17e2021-05-04 18:36:33 -0700296 int map_size, defer_size = 0;
297 int old_map_size, old_defer_size = 0;
Yang Shi2bfd3632021-05-04 18:36:11 -0700298 struct mem_cgroup *memcg;
299
Yang Shi3c6f17e2021-05-04 18:36:33 -0700300 if (!need_expand(new_nr_max))
Yang Shia2fb1262021-05-04 18:36:17 -0700301 goto out;
Yang Shi2bfd3632021-05-04 18:36:11 -0700302
Yang Shi2bfd3632021-05-04 18:36:11 -0700303 if (!root_mem_cgroup)
Yang Shid27cf2a2021-05-04 18:36:14 -0700304 goto out;
305
306 lockdep_assert_held(&shrinker_rwsem);
Yang Shi2bfd3632021-05-04 18:36:11 -0700307
Yang Shi3c6f17e2021-05-04 18:36:33 -0700308 map_size = shrinker_map_size(new_nr_max);
309 defer_size = shrinker_defer_size(new_nr_max);
310 old_map_size = shrinker_map_size(shrinker_nr_max);
311 old_defer_size = shrinker_defer_size(shrinker_nr_max);
312
Yang Shi2bfd3632021-05-04 18:36:11 -0700313 memcg = mem_cgroup_iter(NULL, NULL, NULL);
314 do {
Yang Shi3c6f17e2021-05-04 18:36:33 -0700315 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
316 old_map_size, old_defer_size);
Yang Shi2bfd3632021-05-04 18:36:11 -0700317 if (ret) {
318 mem_cgroup_iter_break(NULL, memcg);
Yang Shid27cf2a2021-05-04 18:36:14 -0700319 goto out;
Yang Shi2bfd3632021-05-04 18:36:11 -0700320 }
321 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
Yang Shid27cf2a2021-05-04 18:36:14 -0700322out:
Yang Shi2bfd3632021-05-04 18:36:11 -0700323 if (!ret)
Yang Shia2fb1262021-05-04 18:36:17 -0700324 shrinker_nr_max = new_nr_max;
Yang Shid27cf2a2021-05-04 18:36:14 -0700325
Yang Shi2bfd3632021-05-04 18:36:11 -0700326 return ret;
327}
328
329void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
330{
331 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
Yang Shie4262c42021-05-04 18:36:23 -0700332 struct shrinker_info *info;
Yang Shi2bfd3632021-05-04 18:36:11 -0700333
334 rcu_read_lock();
Yang Shie4262c42021-05-04 18:36:23 -0700335 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
Yang Shi2bfd3632021-05-04 18:36:11 -0700336 /* Pairs with smp mb in shrink_slab() */
337 smp_mb__before_atomic();
Yang Shie4262c42021-05-04 18:36:23 -0700338 set_bit(shrinker_id, info->map);
Yang Shi2bfd3632021-05-04 18:36:11 -0700339 rcu_read_unlock();
340 }
341}
342
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700343static DEFINE_IDR(shrinker_idr);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700344
345static int prealloc_memcg_shrinker(struct shrinker *shrinker)
346{
347 int id, ret = -ENOMEM;
348
Yang Shi476b30a2021-05-04 18:36:39 -0700349 if (mem_cgroup_disabled())
350 return -ENOSYS;
351
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700352 down_write(&shrinker_rwsem);
353 /* This may call shrinker, so it must use down_read_trylock() */
Yang Shi41ca6682021-05-04 18:36:29 -0700354 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700355 if (id < 0)
356 goto unlock;
357
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700358 if (id >= shrinker_nr_max) {
Yang Shie4262c42021-05-04 18:36:23 -0700359 if (expand_shrinker_info(id)) {
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700360 idr_remove(&shrinker_idr, id);
361 goto unlock;
362 }
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700363 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700364 shrinker->id = id;
365 ret = 0;
366unlock:
367 up_write(&shrinker_rwsem);
368 return ret;
369}
370
371static void unregister_memcg_shrinker(struct shrinker *shrinker)
372{
373 int id = shrinker->id;
374
375 BUG_ON(id < 0);
376
Yang Shi41ca6682021-05-04 18:36:29 -0700377 lockdep_assert_held(&shrinker_rwsem);
378
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700379 idr_remove(&shrinker_idr, id);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700380}
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700381
Yang Shi86750832021-05-04 18:36:36 -0700382static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
383 struct mem_cgroup *memcg)
384{
385 struct shrinker_info *info;
386
387 info = shrinker_info_protected(memcg, nid);
388 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
389}
390
391static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
392 struct mem_cgroup *memcg)
393{
394 struct shrinker_info *info;
395
396 info = shrinker_info_protected(memcg, nid);
397 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
398}
399
Yang Shia1780152021-05-04 18:36:42 -0700400void reparent_shrinker_deferred(struct mem_cgroup *memcg)
401{
402 int i, nid;
403 long nr;
404 struct mem_cgroup *parent;
405 struct shrinker_info *child_info, *parent_info;
406
407 parent = parent_mem_cgroup(memcg);
408 if (!parent)
409 parent = root_mem_cgroup;
410
411 /* Prevent from concurrent shrinker_info expand */
412 down_read(&shrinker_rwsem);
413 for_each_node(nid) {
414 child_info = shrinker_info_protected(memcg, nid);
415 parent_info = shrinker_info_protected(parent, nid);
416 for (i = 0; i < shrinker_nr_max; i++) {
417 nr = atomic_long_read(&child_info->nr_deferred[i]);
418 atomic_long_add(nr, &parent_info->nr_deferred[i]);
419 }
420 }
421 up_read(&shrinker_rwsem);
422}
423
Johannes Weinerb5ead352019-11-30 17:55:40 -0800424static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800425{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800426 return sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800427}
Tejun Heo97c9341f2015-05-22 18:23:36 -0400428
429/**
Johannes Weinerb5ead352019-11-30 17:55:40 -0800430 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
Tejun Heo97c9341f2015-05-22 18:23:36 -0400431 * @sc: scan_control in question
432 *
433 * The normal page dirty throttling mechanism in balance_dirty_pages() is
434 * completely broken with the legacy memcg and direct stalling in
435 * shrink_page_list() is used for throttling instead, which lacks all the
436 * niceties such as fairness, adaptive pausing, bandwidth proportional
437 * allocation and configurability.
438 *
439 * This function tests whether the vmscan currently in progress can assume
440 * that the normal dirty throttling mechanism is operational.
441 */
Johannes Weinerb5ead352019-11-30 17:55:40 -0800442static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c9341f2015-05-22 18:23:36 -0400443{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800444 if (!cgroup_reclaim(sc))
Tejun Heo97c9341f2015-05-22 18:23:36 -0400445 return true;
446#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800447 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c9341f2015-05-22 18:23:36 -0400448 return true;
449#endif
450 return false;
451}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800452#else
Yang Shi0a432dcb2019-09-23 15:38:12 -0700453static int prealloc_memcg_shrinker(struct shrinker *shrinker)
454{
Yang Shi476b30a2021-05-04 18:36:39 -0700455 return -ENOSYS;
Yang Shi0a432dcb2019-09-23 15:38:12 -0700456}
457
458static void unregister_memcg_shrinker(struct shrinker *shrinker)
459{
460}
461
Yang Shi86750832021-05-04 18:36:36 -0700462static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
463 struct mem_cgroup *memcg)
464{
465 return 0;
466}
467
468static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
469 struct mem_cgroup *memcg)
470{
471 return 0;
472}
473
Johannes Weinerb5ead352019-11-30 17:55:40 -0800474static bool cgroup_reclaim(struct scan_control *sc)
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800475{
Johannes Weinerb5ead352019-11-30 17:55:40 -0800476 return false;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800477}
Tejun Heo97c9341f2015-05-22 18:23:36 -0400478
Johannes Weinerb5ead352019-11-30 17:55:40 -0800479static bool writeback_throttling_sane(struct scan_control *sc)
Tejun Heo97c9341f2015-05-22 18:23:36 -0400480{
481 return true;
482}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800483#endif
484
Yang Shi86750832021-05-04 18:36:36 -0700485static long xchg_nr_deferred(struct shrinker *shrinker,
486 struct shrink_control *sc)
487{
488 int nid = sc->nid;
489
490 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
491 nid = 0;
492
493 if (sc->memcg &&
494 (shrinker->flags & SHRINKER_MEMCG_AWARE))
495 return xchg_nr_deferred_memcg(nid, shrinker,
496 sc->memcg);
497
498 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
499}
500
501
502static long add_nr_deferred(long nr, struct shrinker *shrinker,
503 struct shrink_control *sc)
504{
505 int nid = sc->nid;
506
507 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
508 nid = 0;
509
510 if (sc->memcg &&
511 (shrinker->flags & SHRINKER_MEMCG_AWARE))
512 return add_nr_deferred_memcg(nr, nid, shrinker,
513 sc->memcg);
514
515 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
516}
517
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700518/*
519 * This misses isolated pages which are not accounted for to save counters.
520 * As the data only determines if reclaim or compaction continues, it is
521 * not expected that isolated pages will be a dominating factor.
522 */
523unsigned long zone_reclaimable_pages(struct zone *zone)
524{
525 unsigned long nr;
526
527 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
528 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
529 if (get_nr_swap_pages() > 0)
530 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
531 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
532
533 return nr;
534}
535
Michal Hockofd538802017-02-22 15:45:58 -0800536/**
537 * lruvec_lru_size - Returns the number of pages on the given LRU list.
538 * @lruvec: lru vector
539 * @lru: lru to use
540 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
541 */
Yu Zhao20913392021-02-24 12:08:44 -0800542static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
543 int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800544{
Johannes Weinerde3b0152019-11-30 17:55:31 -0800545 unsigned long size = 0;
Michal Hockofd538802017-02-22 15:45:58 -0800546 int zid;
547
Johannes Weinerde3b0152019-11-30 17:55:31 -0800548 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
Michal Hockofd538802017-02-22 15:45:58 -0800549 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800550
Michal Hockofd538802017-02-22 15:45:58 -0800551 if (!managed_zone(zone))
552 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800553
Michal Hockofd538802017-02-22 15:45:58 -0800554 if (!mem_cgroup_disabled())
Johannes Weinerde3b0152019-11-30 17:55:31 -0800555 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
Michal Hockofd538802017-02-22 15:45:58 -0800556 else
Johannes Weinerde3b0152019-11-30 17:55:31 -0800557 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
Michal Hockofd538802017-02-22 15:45:58 -0800558 }
Johannes Weinerde3b0152019-11-30 17:55:31 -0800559 return size;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800560}
561
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000563 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 */
Tetsuo Handa8e049442018-04-04 19:53:07 +0900565int prealloc_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566{
Yang Shi476b30a2021-05-04 18:36:39 -0700567 unsigned int size;
568 int err;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000569
Yang Shi476b30a2021-05-04 18:36:39 -0700570 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
571 err = prealloc_memcg_shrinker(shrinker);
572 if (err != -ENOSYS)
573 return err;
574
575 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
576 }
577
578 size = sizeof(*shrinker->nr_deferred);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000579 if (shrinker->flags & SHRINKER_NUMA_AWARE)
580 size *= nr_node_ids;
581
582 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
583 if (!shrinker->nr_deferred)
584 return -ENOMEM;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700585
Tetsuo Handa8e049442018-04-04 19:53:07 +0900586 return 0;
587}
Glauber Costa1d3d4432013-08-28 10:18:04 +1000588
Tetsuo Handa8e049442018-04-04 19:53:07 +0900589void free_prealloced_shrinker(struct shrinker *shrinker)
590{
Yang Shi41ca6682021-05-04 18:36:29 -0700591 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
592 down_write(&shrinker_rwsem);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700593 unregister_memcg_shrinker(shrinker);
Yang Shi41ca6682021-05-04 18:36:29 -0700594 up_write(&shrinker_rwsem);
Yang Shi476b30a2021-05-04 18:36:39 -0700595 return;
Yang Shi41ca6682021-05-04 18:36:29 -0700596 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700597
Tetsuo Handa8e049442018-04-04 19:53:07 +0900598 kfree(shrinker->nr_deferred);
599 shrinker->nr_deferred = NULL;
600}
601
602void register_shrinker_prepared(struct shrinker *shrinker)
603{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700604 down_write(&shrinker_rwsem);
605 list_add_tail(&shrinker->list, &shrinker_list);
Yang Shi41ca6682021-05-04 18:36:29 -0700606 shrinker->flags |= SHRINKER_REGISTERED;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700607 up_write(&shrinker_rwsem);
Tetsuo Handa8e049442018-04-04 19:53:07 +0900608}
609
610int register_shrinker(struct shrinker *shrinker)
611{
612 int err = prealloc_shrinker(shrinker);
613
614 if (err)
615 return err;
616 register_shrinker_prepared(shrinker);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000617 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700619EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621/*
622 * Remove one
623 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700624void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625{
Yang Shi41ca6682021-05-04 18:36:29 -0700626 if (!(shrinker->flags & SHRINKER_REGISTERED))
Tetsuo Handabb422a72017-12-18 20:31:41 +0900627 return;
Yang Shi41ca6682021-05-04 18:36:29 -0700628
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 down_write(&shrinker_rwsem);
630 list_del(&shrinker->list);
Yang Shi41ca6682021-05-04 18:36:29 -0700631 shrinker->flags &= ~SHRINKER_REGISTERED;
632 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
633 unregister_memcg_shrinker(shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 up_write(&shrinker_rwsem);
Yang Shi41ca6682021-05-04 18:36:29 -0700635
Andrew Vaginae393322013-10-16 13:46:46 -0700636 kfree(shrinker->nr_deferred);
Tetsuo Handabb422a72017-12-18 20:31:41 +0900637 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700639EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
641#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000642
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800643static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
Josef Bacik9092c712018-01-31 16:16:26 -0800644 struct shrinker *shrinker, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000646 unsigned long freed = 0;
647 unsigned long long delta;
648 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700649 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000650 long nr;
651 long new_nr;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000652 long batch_size = shrinker->batch ? shrinker->batch
653 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800654 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000655
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700656 freeable = shrinker->count_objects(shrinker, shrinkctl);
Kirill Tkhai9b996462018-08-17 15:48:21 -0700657 if (freeable == 0 || freeable == SHRINK_EMPTY)
658 return freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000659
660 /*
661 * copy the current shrinker scan count into a local variable
662 * and zero it so that other concurrent shrinker invocations
663 * don't also do this scanning work.
664 */
Yang Shi86750832021-05-04 18:36:36 -0700665 nr = xchg_nr_deferred(shrinker, shrinkctl);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000666
Johannes Weiner4b85afb2018-10-26 15:06:42 -0700667 if (shrinker->seeks) {
668 delta = freeable >> priority;
669 delta *= 4;
670 do_div(delta, shrinker->seeks);
671 } else {
672 /*
673 * These objects don't require any IO to create. Trim
674 * them aggressively under memory pressure to keep
675 * them from causing refetches in the IO caches.
676 */
677 delta = freeable / 2;
678 }
Roman Gushchin172b06c32018-09-20 12:22:46 -0700679
Yang Shi18bb4732021-05-04 18:36:45 -0700680 total_scan = nr >> priority;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000681 total_scan += delta;
Yang Shi18bb4732021-05-04 18:36:45 -0700682 total_scan = min(total_scan, (2 * freeable));
Glauber Costa1d3d4432013-08-28 10:18:04 +1000683
684 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Josef Bacik9092c712018-01-31 16:16:26 -0800685 freeable, delta, total_scan, priority);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000686
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800687 /*
688 * Normally, we should not scan less than batch_size objects in one
689 * pass to avoid too frequent shrinker calls, but if the slab has less
690 * than batch_size objects in total and we are really tight on memory,
691 * we will try to reclaim all available objects, otherwise we can end
692 * up failing allocations although there are plenty of reclaimable
693 * objects spread over several slabs with usage less than the
694 * batch_size.
695 *
696 * We detect the "tight on memory" situations by looking at the total
697 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700698 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800699 * scanning at high prio and therefore should try to reclaim as much as
700 * possible.
701 */
702 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700703 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000704 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800705 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000706
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800707 shrinkctl->nr_to_scan = nr_to_scan;
Chris Wilsond460acb2017-09-06 16:19:26 -0700708 shrinkctl->nr_scanned = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000709 ret = shrinker->scan_objects(shrinker, shrinkctl);
710 if (ret == SHRINK_STOP)
711 break;
712 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000713
Chris Wilsond460acb2017-09-06 16:19:26 -0700714 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
715 total_scan -= shrinkctl->nr_scanned;
716 scanned += shrinkctl->nr_scanned;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000717
718 cond_resched();
719 }
720
Yang Shi18bb4732021-05-04 18:36:45 -0700721 /*
722 * The deferred work is increased by any new work (delta) that wasn't
723 * done, decreased by old deferred work that was done now.
724 *
725 * And it is capped to two times of the freeable items.
726 */
727 next_deferred = max_t(long, (nr + delta - scanned), 0);
728 next_deferred = min(next_deferred, (2 * freeable));
729
Glauber Costa1d3d4432013-08-28 10:18:04 +1000730 /*
731 * move the unused scan count back into the shrinker in a
Yang Shi86750832021-05-04 18:36:36 -0700732 * manner that handles concurrent updates.
Glauber Costa1d3d4432013-08-28 10:18:04 +1000733 */
Yang Shi86750832021-05-04 18:36:36 -0700734 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000735
Yang Shi8efb4b52021-05-04 18:36:08 -0700736 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000737 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738}
739
Yang Shi0a432dcb2019-09-23 15:38:12 -0700740#ifdef CONFIG_MEMCG
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700741static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
742 struct mem_cgroup *memcg, int priority)
743{
Yang Shie4262c42021-05-04 18:36:23 -0700744 struct shrinker_info *info;
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700745 unsigned long ret, freed = 0;
746 int i;
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700747
Yang Shi0a432dcb2019-09-23 15:38:12 -0700748 if (!mem_cgroup_online(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700749 return 0;
750
751 if (!down_read_trylock(&shrinker_rwsem))
752 return 0;
753
Yang Shi468ab842021-05-04 18:36:26 -0700754 info = shrinker_info_protected(memcg, nid);
Yang Shie4262c42021-05-04 18:36:23 -0700755 if (unlikely(!info))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700756 goto unlock;
757
Yang Shie4262c42021-05-04 18:36:23 -0700758 for_each_set_bit(i, info->map, shrinker_nr_max) {
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700759 struct shrink_control sc = {
760 .gfp_mask = gfp_mask,
761 .nid = nid,
762 .memcg = memcg,
763 };
764 struct shrinker *shrinker;
765
766 shrinker = idr_find(&shrinker_idr, i);
Yang Shi41ca6682021-05-04 18:36:29 -0700767 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700768 if (!shrinker)
Yang Shie4262c42021-05-04 18:36:23 -0700769 clear_bit(i, info->map);
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700770 continue;
771 }
772
Yang Shi0a432dcb2019-09-23 15:38:12 -0700773 /* Call non-slab shrinkers even though kmem is disabled */
774 if (!memcg_kmem_enabled() &&
775 !(shrinker->flags & SHRINKER_NONSLAB))
776 continue;
777
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700778 ret = do_shrink_slab(&sc, shrinker, priority);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700779 if (ret == SHRINK_EMPTY) {
Yang Shie4262c42021-05-04 18:36:23 -0700780 clear_bit(i, info->map);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700781 /*
782 * After the shrinker reported that it had no objects to
783 * free, but before we cleared the corresponding bit in
784 * the memcg shrinker map, a new object might have been
785 * added. To make sure, we have the bit set in this
786 * case, we invoke the shrinker one more time and reset
787 * the bit if it reports that it is not empty anymore.
788 * The memory barrier here pairs with the barrier in
Yang Shi2bfd3632021-05-04 18:36:11 -0700789 * set_shrinker_bit():
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700790 *
791 * list_lru_add() shrink_slab_memcg()
792 * list_add_tail() clear_bit()
793 * <MB> <MB>
794 * set_bit() do_shrink_slab()
795 */
796 smp_mb__after_atomic();
797 ret = do_shrink_slab(&sc, shrinker, priority);
798 if (ret == SHRINK_EMPTY)
799 ret = 0;
800 else
Yang Shi2bfd3632021-05-04 18:36:11 -0700801 set_shrinker_bit(memcg, nid, i);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700802 }
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700803 freed += ret;
804
805 if (rwsem_is_contended(&shrinker_rwsem)) {
806 freed = freed ? : 1;
807 break;
808 }
809 }
810unlock:
811 up_read(&shrinker_rwsem);
812 return freed;
813}
Yang Shi0a432dcb2019-09-23 15:38:12 -0700814#else /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700815static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
816 struct mem_cgroup *memcg, int priority)
817{
818 return 0;
819}
Yang Shi0a432dcb2019-09-23 15:38:12 -0700820#endif /* CONFIG_MEMCG */
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700821
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800822/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800823 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800824 * @gfp_mask: allocation context
825 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800826 * @memcg: memory cgroup whose slab caches to target
Josef Bacik9092c712018-01-31 16:16:26 -0800827 * @priority: the reclaim priority
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800829 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800831 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
832 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 *
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700834 * @memcg specifies the memory cgroup to target. Unaware shrinkers
835 * are called only if it is the root cgroup.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800836 *
Josef Bacik9092c712018-01-31 16:16:26 -0800837 * @priority is sc->priority, we take the number of objects and >> by priority
838 * in order to get the scan target.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800840 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800842static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
843 struct mem_cgroup *memcg,
Josef Bacik9092c712018-01-31 16:16:26 -0800844 int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845{
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700846 unsigned long ret, freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 struct shrinker *shrinker;
848
Yang Shifa1e5122019-08-02 21:48:44 -0700849 /*
850 * The root memcg might be allocated even though memcg is disabled
851 * via "cgroup_disable=memory" boot parameter. This could make
852 * mem_cgroup_is_root() return false, then just run memcg slab
853 * shrink, but skip global shrink. This may result in premature
854 * oom.
855 */
856 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700857 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800858
Tetsuo Handae830c632018-04-05 16:23:35 -0700859 if (!down_read_trylock(&shrinker_rwsem))
Minchan Kimf06590b2011-05-24 17:11:11 -0700860 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800863 struct shrink_control sc = {
864 .gfp_mask = gfp_mask,
865 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800866 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800867 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800868
Kirill Tkhai9b996462018-08-17 15:48:21 -0700869 ret = do_shrink_slab(&sc, shrinker, priority);
870 if (ret == SHRINK_EMPTY)
871 ret = 0;
872 freed += ret;
Minchan Kime4966122018-01-31 16:16:55 -0800873 /*
874 * Bail out if someone want to register a new shrinker to
Ethon Paul55b65a52020-06-04 16:49:10 -0700875 * prevent the registration from being stalled for long periods
Minchan Kime4966122018-01-31 16:16:55 -0800876 * by parallel ongoing shrinking.
877 */
878 if (rwsem_is_contended(&shrinker_rwsem)) {
879 freed = freed ? : 1;
880 break;
881 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800883
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700885out:
886 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000887 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888}
889
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800890void drop_slab_node(int nid)
891{
892 unsigned long freed;
893
894 do {
895 struct mem_cgroup *memcg = NULL;
896
Chunxin Zang069c4112020-10-13 16:56:46 -0700897 if (fatal_signal_pending(current))
898 return;
899
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800900 freed = 0;
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700901 memcg = mem_cgroup_iter(NULL, NULL, NULL);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800902 do {
Josef Bacik9092c712018-01-31 16:16:26 -0800903 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800904 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
905 } while (freed > 10);
906}
907
908void drop_slab(void)
909{
910 int nid;
911
912 for_each_online_node(nid)
913 drop_slab_node(nid);
914}
915
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916static inline int is_page_cache_freeable(struct page *page)
917{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700918 /*
919 * A freeable page cache page is referenced only by the caller
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400920 * that isolated the page, the page cache and optional buffer
921 * heads at page->private.
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700922 */
Matthew Wilcox (Oracle)3efe62e2020-10-15 20:05:56 -0700923 int page_cache_pins = thp_nr_pages(page);
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400924 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925}
926
Yang Shicb165562019-11-30 17:55:28 -0800927static int may_write_to_inode(struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928{
Christoph Lameter930d9152006-01-08 01:00:47 -0800929 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400931 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400933 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 return 1;
935 return 0;
936}
937
938/*
939 * We detected a synchronous write error writing a page out. Probably
940 * -ENOSPC. We need to propagate that into the address_space for a subsequent
941 * fsync(), msync() or close().
942 *
943 * The tricky part is that after writepage we cannot touch the mapping: nothing
944 * prevents it from being freed up. But we have a ref on the page and once
945 * that page is locked, the mapping is pinned.
946 *
947 * We're allowed to run sleeping lock_page() here because we know the caller has
948 * __GFP_FS.
949 */
950static void handle_write_error(struct address_space *mapping,
951 struct page *page, int error)
952{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100953 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700954 if (page_mapping(page) == mapping)
955 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 unlock_page(page);
957}
958
Christoph Lameter04e62a22006-06-23 02:03:38 -0700959/* possible outcome of pageout() */
960typedef enum {
961 /* failed to write page out, page is locked */
962 PAGE_KEEP,
963 /* move page to the active list, page is locked */
964 PAGE_ACTIVATE,
965 /* page has been sent to the disk successfully, page is unlocked */
966 PAGE_SUCCESS,
967 /* page is clean and locked */
968 PAGE_CLEAN,
969} pageout_t;
970
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971/*
Andrew Morton1742f192006-03-22 00:08:21 -0800972 * pageout is called by shrink_page_list() for each dirty page.
973 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 */
Yang Shicb165562019-11-30 17:55:28 -0800975static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976{
977 /*
978 * If the page is dirty, only perform writeback if that write
979 * will be non-blocking. To prevent this allocation from being
980 * stalled by pagecache activity. But note that there may be
981 * stalls if we need to run get_block(). We could test
982 * PagePrivate for that.
983 *
Al Viro81742022014-04-03 03:17:43 -0400984 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 * this page's queue, we can perform writeback even if that
986 * will block.
987 *
988 * If the page is swapcache, write it back even if that would
989 * block, for some throttling. This happens by accident, because
990 * swap_backing_dev_info is bust: it doesn't reflect the
991 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 */
993 if (!is_page_cache_freeable(page))
994 return PAGE_KEEP;
995 if (!mapping) {
996 /*
997 * Some data journaling orphaned pages can have
998 * page->mapping == NULL while being dirty with clean buffers.
999 */
David Howells266cf652009-04-03 16:42:36 +01001000 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 if (try_to_free_buffers(page)) {
1002 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -07001003 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 return PAGE_CLEAN;
1005 }
1006 }
1007 return PAGE_KEEP;
1008 }
1009 if (mapping->a_ops->writepage == NULL)
1010 return PAGE_ACTIVATE;
Yang Shicb165562019-11-30 17:55:28 -08001011 if (!may_write_to_inode(mapping->host))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 return PAGE_KEEP;
1013
1014 if (clear_page_dirty_for_io(page)) {
1015 int res;
1016 struct writeback_control wbc = {
1017 .sync_mode = WB_SYNC_NONE,
1018 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001019 .range_start = 0,
1020 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 .for_reclaim = 1,
1022 };
1023
1024 SetPageReclaim(page);
1025 res = mapping->a_ops->writepage(page, &wbc);
1026 if (res < 0)
1027 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -08001028 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 ClearPageReclaim(page);
1030 return PAGE_ACTIVATE;
1031 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001032
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 if (!PageWriteback(page)) {
1034 /* synchronous write or broken a_ops? */
1035 ClearPageReclaim(page);
1036 }
yalin wang3aa23852016-01-14 15:18:30 -08001037 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -07001038 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 return PAGE_SUCCESS;
1040 }
1041
1042 return PAGE_CLEAN;
1043}
1044
Andrew Mortona649fd92006-10-17 00:09:36 -07001045/*
Nick Piggine2867812008-07-25 19:45:30 -07001046 * Same as remove_mapping, but if the page is removed from the mapping, it
1047 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -07001048 */
Johannes Weinera5289102014-04-03 14:47:51 -07001049static int __remove_mapping(struct address_space *mapping, struct page *page,
Johannes Weinerb9107182019-11-30 17:55:59 -08001050 bool reclaimed, struct mem_cgroup *target_memcg)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001051{
Greg Thelenc4843a72015-05-22 17:13:16 -04001052 unsigned long flags;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001053 int refcount;
Joonsoo Kimaae466b2020-08-11 18:30:50 -07001054 void *shadow = NULL;
Greg Thelenc4843a72015-05-22 17:13:16 -04001055
Nick Piggin28e4d962006-09-25 23:31:23 -07001056 BUG_ON(!PageLocked(page));
1057 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001058
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001059 xa_lock_irqsave(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001060 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -07001061 * The non racy check for a busy page.
1062 *
1063 * Must be careful with the order of the tests. When someone has
1064 * a ref to the page, it may be possible that they dirty it then
1065 * drop the reference. So if PageDirty is tested before page_count
1066 * here, then the following race may occur:
1067 *
1068 * get_user_pages(&page);
1069 * [user mapping goes away]
1070 * write_to(page);
1071 * !PageDirty(page) [good]
1072 * SetPageDirty(page);
1073 * put_page(page);
1074 * !page_count(page) [good, discard it]
1075 *
1076 * [oops, our write_to data is lost]
1077 *
1078 * Reversing the order of the tests ensures such a situation cannot
1079 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001080 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -07001081 *
1082 * Note that if SetPageDirty is always performed via set_page_dirty,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001083 * and thus under the i_pages lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001084 */
William Kucharski906d2782019-10-18 20:20:33 -07001085 refcount = 1 + compound_nr(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -07001086 if (!page_ref_freeze(page, refcount))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001087 goto cannot_free;
Jiang Biao1c4c3b92018-08-21 21:53:13 -07001088 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
Nick Piggine2867812008-07-25 19:45:30 -07001089 if (unlikely(PageDirty(page))) {
Huang Yingbd4c82c22017-09-06 16:22:49 -07001090 page_ref_unfreeze(page, refcount);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001091 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -07001092 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001093
1094 if (PageSwapCache(page)) {
1095 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001096 mem_cgroup_swapout(page, swap);
Joonsoo Kimaae466b2020-08-11 18:30:50 -07001097 if (reclaimed && !mapping_exiting(mapping))
1098 shadow = workingset_eviction(page, target_memcg);
1099 __delete_from_swap_cache(page, swap, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001100 xa_unlock_irqrestore(&mapping->i_pages, flags);
Minchan Kim75f6d6d2017-07-06 15:37:21 -07001101 put_swap_page(page, swap);
Nick Piggine2867812008-07-25 19:45:30 -07001102 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -05001103 void (*freepage)(struct page *);
1104
1105 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -07001106 /*
1107 * Remember a shadow entry for reclaimed file cache in
1108 * order to detect refaults, thus thrashing, later on.
1109 *
1110 * But don't store shadows in an address space that is
dylan-meiners238c3042020-08-06 23:26:29 -07001111 * already exiting. This is not just an optimization,
Johannes Weinera5289102014-04-03 14:47:51 -07001112 * inode reclaim needs to empty out the radix tree or
1113 * the nodes are lost. Don't plant shadows behind its
1114 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -08001115 *
1116 * We also don't store shadows for DAX mappings because the
1117 * only page cache pages found in these are zero pages
1118 * covering holes, and because we don't want to mix DAX
1119 * exceptional entries and shadow exceptional entries in the
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001120 * same address_space.
Johannes Weinera5289102014-04-03 14:47:51 -07001121 */
Huang Ying9de4f222020-04-06 20:04:41 -07001122 if (reclaimed && page_is_file_lru(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -08001123 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinerb9107182019-11-30 17:55:59 -08001124 shadow = workingset_eviction(page, target_memcg);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001125 __delete_from_page_cache(page, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001126 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -05001127
1128 if (freepage != NULL)
1129 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001130 }
1131
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001132 return 1;
1133
1134cannot_free:
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07001135 xa_unlock_irqrestore(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001136 return 0;
1137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139/*
Nick Piggine2867812008-07-25 19:45:30 -07001140 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1141 * someone else has a ref on the page, abort and return 0. If it was
1142 * successfully detached, return 1. Assumes the caller has a single ref on
1143 * this page.
1144 */
1145int remove_mapping(struct address_space *mapping, struct page *page)
1146{
Johannes Weinerb9107182019-11-30 17:55:59 -08001147 if (__remove_mapping(mapping, page, false, NULL)) {
Nick Piggine2867812008-07-25 19:45:30 -07001148 /*
1149 * Unfreezing the refcount with 1 rather than 2 effectively
1150 * drops the pagecache ref for us without requiring another
1151 * atomic operation.
1152 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001153 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -07001154 return 1;
1155 }
1156 return 0;
1157}
1158
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001159/**
1160 * putback_lru_page - put previously isolated page onto appropriate LRU list
1161 * @page: page to be put back to appropriate lru list
1162 *
1163 * Add previously isolated @page to appropriate LRU list.
1164 * Page may still be unevictable for other reasons.
1165 *
1166 * lru_lock must not be held, interrupts must be enabled.
1167 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001168void putback_lru_page(struct page *page)
1169{
Shakeel Butt9c4e6b12018-02-21 14:45:28 -08001170 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001171 put_page(page); /* drop ref from isolate */
1172}
1173
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001174enum page_references {
1175 PAGEREF_RECLAIM,
1176 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -08001177 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001178 PAGEREF_ACTIVATE,
1179};
1180
1181static enum page_references page_check_references(struct page *page,
1182 struct scan_control *sc)
1183{
Johannes Weiner645747462010-03-05 13:42:22 -08001184 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001185 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001186
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001187 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1188 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -08001189 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001190
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001191 /*
1192 * Mlock lost the isolation race with us. Let try_to_unmap()
1193 * move the page to the unevictable list.
1194 */
1195 if (vm_flags & VM_LOCKED)
1196 return PAGEREF_RECLAIM;
1197
Johannes Weiner645747462010-03-05 13:42:22 -08001198 if (referenced_ptes) {
Johannes Weiner645747462010-03-05 13:42:22 -08001199 /*
1200 * All mapped pages start out with page table
1201 * references from the instantiating fault, so we need
1202 * to look twice if a mapped file page is used more
1203 * than once.
1204 *
1205 * Mark it and spare it for another trip around the
1206 * inactive list. Another page table reference will
1207 * lead to its activation.
1208 *
1209 * Note: the mark is set for activated pages as well
1210 * so that recently deactivated but used pages are
1211 * quickly recovered.
1212 */
1213 SetPageReferenced(page);
1214
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -08001215 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -08001216 return PAGEREF_ACTIVATE;
1217
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001218 /*
1219 * Activate file-backed executable pages after first usage.
1220 */
Joonsoo Kimb5181542020-08-11 18:30:40 -07001221 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001222 return PAGEREF_ACTIVATE;
1223
Johannes Weiner645747462010-03-05 13:42:22 -08001224 return PAGEREF_KEEP;
1225 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001226
1227 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -07001228 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001229 return PAGEREF_RECLAIM_CLEAN;
1230
1231 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001232}
1233
Mel Gormane2be15f2013-07-03 15:01:57 -07001234/* Check if a page is dirty or under writeback */
1235static void page_check_dirty_writeback(struct page *page,
1236 bool *dirty, bool *writeback)
1237{
Mel Gormanb4597222013-07-03 15:02:05 -07001238 struct address_space *mapping;
1239
Mel Gormane2be15f2013-07-03 15:01:57 -07001240 /*
1241 * Anonymous pages are not handled by flushers and must be written
1242 * from reclaim context. Do not stall reclaim based on them
1243 */
Huang Ying9de4f222020-04-06 20:04:41 -07001244 if (!page_is_file_lru(page) ||
Shaohua Li802a3a92017-05-03 14:52:32 -07001245 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -07001246 *dirty = false;
1247 *writeback = false;
1248 return;
1249 }
1250
1251 /* By default assume that the page flags are accurate */
1252 *dirty = PageDirty(page);
1253 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -07001254
1255 /* Verify dirty/writeback state if the filesystem supports it */
1256 if (!page_has_private(page))
1257 return;
1258
1259 mapping = page_mapping(page);
1260 if (mapping && mapping->a_ops->is_dirty_writeback)
1261 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -07001262}
1263
Nick Piggine2867812008-07-25 19:45:30 -07001264/*
Andrew Morton1742f192006-03-22 00:08:21 -08001265 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 */
Maninder Singh730ec8c2020-06-03 16:01:18 -07001267static unsigned int shrink_page_list(struct list_head *page_list,
1268 struct pglist_data *pgdat,
1269 struct scan_control *sc,
Maninder Singh730ec8c2020-06-03 16:01:18 -07001270 struct reclaim_stat *stat,
1271 bool ignore_references)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272{
1273 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001274 LIST_HEAD(free_pages);
Maninder Singh730ec8c2020-06-03 16:01:18 -07001275 unsigned int nr_reclaimed = 0;
1276 unsigned int pgactivate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Kirill Tkhai060f0052019-03-05 15:48:15 -08001278 memset(stat, 0, sizeof(*stat));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 cond_resched();
1280
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 while (!list_empty(page_list)) {
1282 struct address_space *mapping;
1283 struct page *page;
Minchan Kim8940b342019-09-25 16:49:11 -07001284 enum page_references references = PAGEREF_RECLAIM;
Kirill Tkhai4b793062020-04-01 21:10:18 -07001285 bool dirty, writeback, may_enter_fs;
Yang Shi98879b32019-07-11 20:59:30 -07001286 unsigned int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
1288 cond_resched();
1289
1290 page = lru_to_page(page_list);
1291 list_del(&page->lru);
1292
Nick Piggin529ae9a2008-08-02 12:01:03 +02001293 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 goto keep;
1295
Sasha Levin309381fea2014-01-23 15:52:54 -08001296 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07001298 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07001299
1300 /* Account the number of base pages even though THP */
1301 sc->nr_scanned += nr_pages;
Christoph Lameter80e43422006-02-11 17:55:53 -08001302
Hugh Dickins39b5f292012-10-08 16:33:18 -07001303 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -07001304 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001305
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001306 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -08001307 goto keep_locked;
1308
Andy Whitcroftc661b072007-08-22 14:01:26 -07001309 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1310 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1311
Mel Gorman283aba92013-07-03 15:01:51 -07001312 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07001313 * The number of dirty pages determines if a node is marked
Mel Gormane2be15f2013-07-03 15:01:57 -07001314 * reclaim_congested which affects wait_iff_congested. kswapd
1315 * will stall and start writing pages if the tail of the LRU
1316 * is all dirty unqueued pages.
1317 */
1318 page_check_dirty_writeback(page, &dirty, &writeback);
1319 if (dirty || writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001320 stat->nr_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001321
1322 if (dirty && !writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001323 stat->nr_unqueued_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001324
Mel Gormand04e8ac2013-07-03 15:02:03 -07001325 /*
1326 * Treat this page as congested if the underlying BDI is or if
1327 * pages are cycling through the LRU so quickly that the
1328 * pages marked for immediate reclaim are making it to the
1329 * end of the LRU a second time.
1330 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001331 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001332 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001333 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001334 (writeback && PageReclaim(page)))
Kirill Tkhai060f0052019-03-05 15:48:15 -08001335 stat->nr_congested++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001336
1337 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001338 * If a page at the tail of the LRU is under writeback, there
1339 * are three cases to consider.
1340 *
1341 * 1) If reclaim is encountering an excessive number of pages
1342 * under writeback and this page is both under writeback and
1343 * PageReclaim then it indicates that pages are being queued
1344 * for IO but are being recycled through the LRU before the
1345 * IO can complete. Waiting on the page itself risks an
1346 * indefinite stall if it is impossible to writeback the
1347 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001348 * note that the LRU is being scanned too quickly and the
1349 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001350 *
Tejun Heo97c9341f2015-05-22 18:23:36 -04001351 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001352 * not marked for immediate reclaim, or the caller does not
1353 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1354 * not to fs). In this case mark the page for immediate
Tejun Heo97c9341f2015-05-22 18:23:36 -04001355 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001356 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001357 * Require may_enter_fs because we would wait on fs, which
1358 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001359 * enter reclaim, and deadlock if it waits on a page for
1360 * which it is needed to do the write (loop masks off
1361 * __GFP_IO|__GFP_FS for this reason); but more thought
1362 * would probably show more reasons.
1363 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001364 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001365 * PageReclaim. memcg does not have any dirty pages
1366 * throttling so we could easily OOM just because too many
1367 * pages are in writeback and there is nothing else to
1368 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001369 *
1370 * In cases 1) and 2) we activate the pages to get them out of
1371 * the way while we continue scanning for clean pages on the
1372 * inactive list and refilling from the active list. The
1373 * observation here is that waiting for disk writes is more
1374 * expensive than potentially causing reloads down the line.
1375 * Since they're marked for immediate reclaim, they won't put
1376 * memory pressure on the cache working set any longer than it
1377 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001378 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001379 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001380 /* Case 1 above */
1381 if (current_is_kswapd() &&
1382 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001383 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Kirill Tkhai060f0052019-03-05 15:48:15 -08001384 stat->nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001385 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001386
1387 /* Case 2 above */
Johannes Weinerb5ead352019-11-30 17:55:40 -08001388 } else if (writeback_throttling_sane(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001389 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001390 /*
1391 * This is slightly racy - end_page_writeback()
1392 * might have just cleared PageReclaim, then
1393 * setting PageReclaim here end up interpreted
1394 * as PageReadahead - but that does not matter
1395 * enough to care. What we do want is for this
1396 * page to have PageReclaim set next time memcg
1397 * reclaim reaches the tests above, so it will
1398 * then wait_on_page_writeback() to avoid OOM;
1399 * and it's also appropriate in global reclaim.
1400 */
1401 SetPageReclaim(page);
Kirill Tkhai060f0052019-03-05 15:48:15 -08001402 stat->nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001403 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001404
1405 /* Case 3 above */
1406 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001407 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001408 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001409 /* then go back and try same page again */
1410 list_add_tail(&page->lru, page_list);
1411 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001412 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001413 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414
Minchan Kim8940b342019-09-25 16:49:11 -07001415 if (!ignore_references)
Minchan Kim02c6de82012-10-08 16:31:55 -07001416 references = page_check_references(page, sc);
1417
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001418 switch (references) {
1419 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001421 case PAGEREF_KEEP:
Yang Shi98879b32019-07-11 20:59:30 -07001422 stat->nr_ref_keep += nr_pages;
Johannes Weiner645747462010-03-05 13:42:22 -08001423 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001424 case PAGEREF_RECLAIM:
1425 case PAGEREF_RECLAIM_CLEAN:
1426 ; /* try to reclaim the page below */
1427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429 /*
1430 * Anonymous process memory has backing store?
1431 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001432 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001434 if (PageAnon(page) && PageSwapBacked(page)) {
1435 if (!PageSwapCache(page)) {
1436 if (!(sc->gfp_mask & __GFP_IO))
1437 goto keep_locked;
Linus Torvaldsfeb889f2021-01-16 15:34:57 -08001438 if (page_maybe_dma_pinned(page))
1439 goto keep_locked;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001440 if (PageTransHuge(page)) {
1441 /* cannot split THP, skip it */
1442 if (!can_split_huge_page(page, NULL))
1443 goto activate_locked;
1444 /*
1445 * Split pages without a PMD map right
1446 * away. Chances are some or all of the
1447 * tail pages can be freed without IO.
1448 */
1449 if (!compound_mapcount(page) &&
1450 split_huge_page_to_list(page,
1451 page_list))
1452 goto activate_locked;
1453 }
1454 if (!add_to_swap(page)) {
1455 if (!PageTransHuge(page))
Yang Shi98879b32019-07-11 20:59:30 -07001456 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001457 /* Fallback to swap normal pages */
1458 if (split_huge_page_to_list(page,
1459 page_list))
1460 goto activate_locked;
Huang Yingfe490cc2017-09-06 16:22:52 -07001461#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1462 count_vm_event(THP_SWPOUT_FALLBACK);
1463#endif
Huang Yingbd4c82c22017-09-06 16:22:49 -07001464 if (!add_to_swap(page))
Yang Shi98879b32019-07-11 20:59:30 -07001465 goto activate_locked_split;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001466 }
Minchan Kim0f074652017-07-06 15:37:24 -07001467
Kirill Tkhai4b793062020-04-01 21:10:18 -07001468 may_enter_fs = true;
Huang Yingbd4c82c22017-09-06 16:22:49 -07001469
1470 /* Adding to swap updated mapping */
1471 mapping = page_mapping(page);
Minchan Kim0f074652017-07-06 15:37:24 -07001472 }
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001473 } else if (unlikely(PageTransHuge(page))) {
1474 /* Split file THP */
1475 if (split_huge_page_to_list(page, page_list))
1476 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478
1479 /*
Yang Shi98879b32019-07-11 20:59:30 -07001480 * THP may get split above, need minus tail pages and update
1481 * nr_pages to avoid accounting tail pages twice.
1482 *
1483 * The tail pages that are added into swap cache successfully
1484 * reach here.
1485 */
1486 if ((nr_pages > 1) && !PageTransHuge(page)) {
1487 sc->nr_scanned -= (nr_pages - 1);
1488 nr_pages = 1;
1489 }
1490
1491 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 * The page is mapped into the page tables of one or more
1493 * processes. Try to unmap it here.
1494 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001495 if (page_mapped(page)) {
Shakeel Butt013339d2020-12-14 19:06:39 -08001496 enum ttu_flags flags = TTU_BATCH_FLUSH;
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001497 bool was_swapbacked = PageSwapBacked(page);
Huang Yingbd4c82c22017-09-06 16:22:49 -07001498
1499 if (unlikely(PageTransHuge(page)))
1500 flags |= TTU_SPLIT_HUGE_PMD;
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001501
Huang Yingbd4c82c22017-09-06 16:22:49 -07001502 if (!try_to_unmap(page, flags)) {
Yang Shi98879b32019-07-11 20:59:30 -07001503 stat->nr_unmap_fail += nr_pages;
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001504 if (!was_swapbacked && PageSwapBacked(page))
1505 stat->nr_lazyfree_fail += nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 }
1508 }
1509
1510 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001511 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001512 * Only kswapd can writeback filesystem pages
1513 * to avoid risk of stack overflow. But avoid
1514 * injecting inefficient single-page IO into
1515 * flusher writeback as much as possible: only
1516 * write pages when we've encountered many
1517 * dirty pages, and when we've already scanned
1518 * the rest of the LRU for clean pages and see
1519 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001520 */
Huang Ying9de4f222020-04-06 20:04:41 -07001521 if (page_is_file_lru(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001522 (!current_is_kswapd() || !PageReclaim(page) ||
1523 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001524 /*
1525 * Immediately reclaim when written back.
1526 * Similar in principal to deactivate_page()
1527 * except we already have the page isolated
1528 * and know it's dirty
1529 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001530 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001531 SetPageReclaim(page);
1532
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001533 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001534 }
1535
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001536 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001538 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001540 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 goto keep_locked;
1542
Mel Gormand950c942015-09-04 15:47:35 -07001543 /*
1544 * Page is dirty. Flush the TLB if a writable entry
1545 * potentially exists to avoid CPU writes after IO
1546 * starts and then write it out here.
1547 */
1548 try_to_unmap_flush_dirty();
Yang Shicb165562019-11-30 17:55:28 -08001549 switch (pageout(page, mapping)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 case PAGE_KEEP:
1551 goto keep_locked;
1552 case PAGE_ACTIVATE:
1553 goto activate_locked;
1554 case PAGE_SUCCESS:
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07001555 stat->nr_pageout += thp_nr_pages(page);
Johannes Weiner96f8bf42020-06-03 16:03:09 -07001556
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001557 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001558 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001559 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001561
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 /*
1563 * A synchronous write - probably a ramdisk. Go
1564 * ahead and try to reclaim the page.
1565 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001566 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 goto keep;
1568 if (PageDirty(page) || PageWriteback(page))
1569 goto keep_locked;
1570 mapping = page_mapping(page);
Gustavo A. R. Silva01359eb2020-12-14 19:15:00 -08001571 fallthrough;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 case PAGE_CLEAN:
1573 ; /* try to free the page below */
1574 }
1575 }
1576
1577 /*
1578 * If the page has buffers, try to free the buffer mappings
1579 * associated with this page. If we succeed we try to free
1580 * the page as well.
1581 *
1582 * We do this even if the page is PageDirty().
1583 * try_to_release_page() does not perform I/O, but it is
1584 * possible for a page to have PageDirty set, but it is actually
1585 * clean (all its buffers are clean). This happens if the
1586 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001587 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 * try_to_release_page() will discover that cleanness and will
1589 * drop the buffers and mark the page clean - it can be freed.
1590 *
1591 * Rarely, pages can have buffers and no ->mapping. These are
1592 * the pages which were not successfully invalidated in
Yang Shid12b8952020-12-14 19:13:02 -08001593 * truncate_cleanup_page(). We try to drop those buffers here
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 * and if that worked, and the page is no longer mapped into
1595 * process address space (page_count == 1) it can be freed.
1596 * Otherwise, leave the page on the LRU so it is swappable.
1597 */
David Howells266cf652009-04-03 16:42:36 +01001598 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 if (!try_to_release_page(page, sc->gfp_mask))
1600 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001601 if (!mapping && page_count(page) == 1) {
1602 unlock_page(page);
1603 if (put_page_testzero(page))
1604 goto free_it;
1605 else {
1606 /*
1607 * rare race with speculative reference.
1608 * the speculative reference will free
1609 * this page shortly, so we may
1610 * increment nr_reclaimed here (and
1611 * leave it off the LRU).
1612 */
1613 nr_reclaimed++;
1614 continue;
1615 }
1616 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 }
1618
Shaohua Li802a3a92017-05-03 14:52:32 -07001619 if (PageAnon(page) && !PageSwapBacked(page)) {
1620 /* follow __remove_mapping for reference */
1621 if (!page_ref_freeze(page, 1))
1622 goto keep_locked;
1623 if (PageDirty(page)) {
1624 page_ref_unfreeze(page, 1);
1625 goto keep_locked;
1626 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Shaohua Li802a3a92017-05-03 14:52:32 -07001628 count_vm_event(PGLAZYFREED);
Roman Gushchin22621852017-07-06 15:40:25 -07001629 count_memcg_page_event(page, PGLAZYFREED);
Johannes Weinerb9107182019-11-30 17:55:59 -08001630 } else if (!mapping || !__remove_mapping(mapping, page, true,
1631 sc->target_mem_cgroup))
Shaohua Li802a3a92017-05-03 14:52:32 -07001632 goto keep_locked;
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001633
1634 unlock_page(page);
Nick Piggine2867812008-07-25 19:45:30 -07001635free_it:
Yang Shi98879b32019-07-11 20:59:30 -07001636 /*
1637 * THP may get swapped out in a whole, need account
1638 * all base pages.
1639 */
1640 nr_reclaimed += nr_pages;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001641
1642 /*
1643 * Is there need to periodically free_page_list? It would
1644 * appear not as the counts should be low
1645 */
Yang Shi7ae88532019-09-23 15:38:09 -07001646 if (unlikely(PageTransHuge(page)))
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07001647 destroy_compound_page(page);
Yang Shi7ae88532019-09-23 15:38:09 -07001648 else
Huang Yingbd4c82c22017-09-06 16:22:49 -07001649 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 continue;
1651
Yang Shi98879b32019-07-11 20:59:30 -07001652activate_locked_split:
1653 /*
1654 * The tail pages that are failed to add into swap cache
1655 * reach here. Fixup nr_scanned and nr_pages.
1656 */
1657 if (nr_pages > 1) {
1658 sc->nr_scanned -= (nr_pages - 1);
1659 nr_pages = 1;
1660 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001662 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001663 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1664 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001665 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001666 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001667 if (!PageMlocked(page)) {
Huang Ying9de4f222020-04-06 20:04:41 -07001668 int type = page_is_file_lru(page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001669 SetPageActive(page);
Yang Shi98879b32019-07-11 20:59:30 -07001670 stat->nr_activate[type] += nr_pages;
Roman Gushchin22621852017-07-06 15:40:25 -07001671 count_memcg_page_event(page, PGACTIVATE);
Minchan Kimad6b6702017-05-03 14:54:13 -07001672 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673keep_locked:
1674 unlock_page(page);
1675keep:
1676 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001677 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001679
Yang Shi98879b32019-07-11 20:59:30 -07001680 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1681
Johannes Weiner747db952014-08-08 14:19:24 -07001682 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001683 try_to_unmap_flush();
Mel Gorman2d4894b2017-11-15 17:37:59 -08001684 free_unref_page_list(&free_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001685
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 list_splice(&ret_pages, page_list);
Kirill Tkhai886cf192019-05-13 17:16:51 -07001687 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001688
Andrew Morton05ff5132006-03-22 00:08:20 -08001689 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690}
1691
Maninder Singh730ec8c2020-06-03 16:01:18 -07001692unsigned int reclaim_clean_pages_from_list(struct zone *zone,
Minchan Kim02c6de82012-10-08 16:31:55 -07001693 struct list_head *page_list)
1694{
1695 struct scan_control sc = {
1696 .gfp_mask = GFP_KERNEL,
1697 .priority = DEF_PRIORITY,
1698 .may_unmap = 1,
1699 };
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001700 struct reclaim_stat stat;
Maninder Singh730ec8c2020-06-03 16:01:18 -07001701 unsigned int nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001702 struct page *page, *next;
1703 LIST_HEAD(clean_pages);
1704
1705 list_for_each_entry_safe(page, next, page_list, lru) {
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07001706 if (!PageHuge(page) && page_is_file_lru(page) &&
1707 !PageDirty(page) && !__PageMovable(page) &&
1708 !PageUnevictable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001709 ClearPageActive(page);
1710 list_move(&page->lru, &clean_pages);
1711 }
1712 }
1713
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001714 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Shakeel Butt013339d2020-12-14 19:06:39 -08001715 &stat, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001716 list_splice(&clean_pages, page_list);
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001717 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1718 -(long)nr_reclaimed);
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001719 /*
1720 * Since lazyfree pages are isolated from file LRU from the beginning,
1721 * they will rotate back to anonymous LRU in the end if it failed to
1722 * discard so isolated count will be mismatched.
1723 * Compensate the isolated count for both LRU lists.
1724 */
1725 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1726 stat.nr_lazyfree_fail);
1727 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
Nicholas Piggin2da9f632020-11-13 22:51:46 -08001728 -(long)stat.nr_lazyfree_fail);
Jaewon Kim1f318a9b2020-06-03 16:01:15 -07001729 return nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07001730}
1731
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001732/*
1733 * Attempt to remove the specified page from its LRU. Only take this page
1734 * if it is of the appropriate PageActive status. Pages which are being
1735 * freed elsewhere are also ignored.
1736 *
1737 * page: page to consider
1738 * mode: one of the LRU isolation modes defined above
1739 *
Alex Shic2135f72021-02-24 12:08:01 -08001740 * returns true on success, false on failure.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001741 */
Alex Shic2135f72021-02-24 12:08:01 -08001742bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001743{
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001744 /* Only take pages on the LRU. */
1745 if (!PageLRU(page))
Alex Shic2135f72021-02-24 12:08:01 -08001746 return false;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001747
Minchan Kime46a2872012-10-08 16:33:48 -07001748 /* Compaction should not handle unevictable pages but CMA can do so */
1749 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Alex Shic2135f72021-02-24 12:08:01 -08001750 return false;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001751
Mel Gormanc8244932012-01-12 17:19:38 -08001752 /*
1753 * To minimise LRU disruption, the caller can indicate that it only
1754 * wants to isolate pages it will be able to operate on without
1755 * blocking - clean pages for the most part.
1756 *
Mel Gormanc8244932012-01-12 17:19:38 -08001757 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1758 * that it is possible to migrate without blocking
1759 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001760 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001761 /* All the caller can do on PageWriteback is block */
1762 if (PageWriteback(page))
Alex Shic2135f72021-02-24 12:08:01 -08001763 return false;
Mel Gormanc8244932012-01-12 17:19:38 -08001764
1765 if (PageDirty(page)) {
1766 struct address_space *mapping;
Mel Gorman69d763f2018-01-31 16:19:52 -08001767 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001768
Mel Gormanc8244932012-01-12 17:19:38 -08001769 /*
1770 * Only pages without mappings or that have a
1771 * ->migratepage callback are possible to migrate
Mel Gorman69d763f2018-01-31 16:19:52 -08001772 * without blocking. However, we can be racing with
1773 * truncation so it's necessary to lock the page
1774 * to stabilise the mapping as truncation holds
1775 * the page lock until after the page is removed
1776 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001777 */
Mel Gorman69d763f2018-01-31 16:19:52 -08001778 if (!trylock_page(page))
Alex Shic2135f72021-02-24 12:08:01 -08001779 return false;
Mel Gorman69d763f2018-01-31 16:19:52 -08001780
Mel Gormanc8244932012-01-12 17:19:38 -08001781 mapping = page_mapping(page);
Hugh Dickins145e1a72018-06-01 16:50:50 -07001782 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gorman69d763f2018-01-31 16:19:52 -08001783 unlock_page(page);
1784 if (!migrate_dirty)
Alex Shic2135f72021-02-24 12:08:01 -08001785 return false;
Mel Gormanc8244932012-01-12 17:19:38 -08001786 }
1787 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001788
Minchan Kimf80c0672011-10-31 17:06:55 -07001789 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
Alex Shic2135f72021-02-24 12:08:01 -08001790 return false;
Minchan Kimf80c0672011-10-31 17:06:55 -07001791
Alex Shic2135f72021-02-24 12:08:01 -08001792 return true;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001793}
1794
Mel Gorman7ee36a12016-07-28 15:47:17 -07001795/*
1796 * Update LRU sizes after isolating pages. The LRU size updates must
Ethon Paul55b65a52020-06-04 16:49:10 -07001797 * be complete before mem_cgroup_update_lru_size due to a sanity check.
Mel Gorman7ee36a12016-07-28 15:47:17 -07001798 */
1799static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001800 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001801{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001802 int zid;
1803
Mel Gorman7ee36a12016-07-28 15:47:17 -07001804 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1805 if (!nr_zone_taken[zid])
1806 continue;
1807
Wei Yanga892cb62020-06-03 16:01:12 -07001808 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Mel Gorman7ee36a12016-07-28 15:47:17 -07001809 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001810
Mel Gorman7ee36a12016-07-28 15:47:17 -07001811}
1812
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001813/**
Hugh Dickins15b44732020-12-15 14:21:31 -08001814 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1815 *
1816 * lruvec->lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 * shrink the lists perform better by taking out a batch of pages
1818 * and working on them outside the LRU lock.
1819 *
1820 * For pagecache intensive workloads, this function is the hottest
1821 * spot in the kernel (apart from copy_*_user functions).
1822 *
Hugh Dickins15b44732020-12-15 14:21:31 -08001823 * Lru_lock must be held before calling this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 *
Minchan Kim791b48b2017-05-12 15:47:06 -07001825 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001826 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001828 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001829 * @sc: The scan_control struct for this reclaim session
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001830 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 *
1832 * returns how many pages were moved onto *@dst.
1833 */
Andrew Morton69e05942006-03-22 00:08:19 -08001834static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001835 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001836 unsigned long *nr_scanned, struct scan_control *sc,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001837 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001839 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001840 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001841 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001842 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07001843 unsigned long skipped = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001844 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001845 LIST_HEAD(pages_skipped);
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001846 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
Yang Shi98879b32019-07-11 20:59:30 -07001848 total_scan = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001849 scan = 0;
Yang Shi98879b32019-07-11 20:59:30 -07001850 while (scan < nr_to_scan && !list_empty(src)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001851 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001852
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 page = lru_to_page(src);
1854 prefetchw_prev_lru_page(page, src, flags);
1855
Matthew Wilcox (Oracle)d8c65462019-09-23 15:34:30 -07001856 nr_pages = compound_nr(page);
Yang Shi98879b32019-07-11 20:59:30 -07001857 total_scan += nr_pages;
1858
Mel Gormanb2e18752016-07-28 15:45:37 -07001859 if (page_zonenum(page) > sc->reclaim_idx) {
1860 list_move(&page->lru, &pages_skipped);
Yang Shi98879b32019-07-11 20:59:30 -07001861 nr_skipped[page_zonenum(page)] += nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001862 continue;
1863 }
1864
Minchan Kim791b48b2017-05-12 15:47:06 -07001865 /*
1866 * Do not count skipped pages because that makes the function
1867 * return with no isolated pages if the LRU mostly contains
1868 * ineligible pages. This causes the VM to not reclaim any
1869 * pages, triggering a premature OOM.
Yang Shi98879b32019-07-11 20:59:30 -07001870 *
1871 * Account all tail pages of THP. This would not cause
1872 * premature OOM since __isolate_lru_page() returns -EBUSY
1873 * only when the page is being freed somewhere else.
Minchan Kim791b48b2017-05-12 15:47:06 -07001874 */
Yang Shi98879b32019-07-11 20:59:30 -07001875 scan += nr_pages;
Alex Shic2135f72021-02-24 12:08:01 -08001876 if (!__isolate_lru_page_prepare(page, mode)) {
1877 /* It is being freed elsewhere */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001878 list_move(&page->lru, src);
Alex Shic2135f72021-02-24 12:08:01 -08001879 continue;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001880 }
Alex Shic2135f72021-02-24 12:08:01 -08001881 /*
1882 * Be careful not to clear PageLRU until after we're
1883 * sure the page is not being freed elsewhere -- the
1884 * page release code relies on it.
1885 */
1886 if (unlikely(!get_page_unless_zero(page))) {
1887 list_move(&page->lru, src);
1888 continue;
1889 }
1890
1891 if (!TestClearPageLRU(page)) {
1892 /* Another thread is already isolating this page */
1893 put_page(page);
1894 list_move(&page->lru, src);
1895 continue;
1896 }
1897
1898 nr_taken += nr_pages;
1899 nr_zone_taken[page_zonenum(page)] += nr_pages;
1900 list_move(&page->lru, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 }
1902
Mel Gormanb2e18752016-07-28 15:45:37 -07001903 /*
1904 * Splice any skipped pages to the start of the LRU list. Note that
1905 * this disrupts the LRU order when reclaiming for lower zones but
1906 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1907 * scanning would soon rescan the same pages to skip and put the
1908 * system at risk of premature OOM.
1909 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001910 if (!list_empty(&pages_skipped)) {
1911 int zid;
1912
Johannes Weiner3db65812017-05-03 14:52:13 -07001913 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001914 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1915 if (!nr_skipped[zid])
1916 continue;
1917
1918 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08001919 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001920 }
1921 }
Minchan Kim791b48b2017-05-12 15:47:06 -07001922 *nr_scanned = total_scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08001923 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim791b48b2017-05-12 15:47:06 -07001924 total_scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001925 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 return nr_taken;
1927}
1928
Nick Piggin62695a82008-10-18 20:26:09 -07001929/**
1930 * isolate_lru_page - tries to isolate a page from its LRU list
1931 * @page: page to isolate from its LRU list
1932 *
1933 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1934 * vmstat statistic corresponding to whatever LRU list the page was on.
1935 *
1936 * Returns 0 if the page was removed from an LRU list.
1937 * Returns -EBUSY if the page was not on an LRU list.
1938 *
1939 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001940 * the active list, it will have PageActive set. If it was found on
1941 * the unevictable list, it will have the PageUnevictable bit set. That flag
1942 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001943 *
1944 * The vmstat statistic corresponding to the list on which the page was
1945 * found will be decremented.
1946 *
1947 * Restrictions:
Mike Rapoporta5d09be2018-02-06 15:42:19 -08001948 *
Nick Piggin62695a82008-10-18 20:26:09 -07001949 * (1) Must be called with an elevated refcount on the page. This is a
Hui Su01c47762020-10-13 16:56:49 -07001950 * fundamental difference from isolate_lru_pages (which is called
Nick Piggin62695a82008-10-18 20:26:09 -07001951 * without a stable reference).
1952 * (2) the lru_lock must not be held.
1953 * (3) interrupts must be enabled.
1954 */
1955int isolate_lru_page(struct page *page)
1956{
1957 int ret = -EBUSY;
1958
Sasha Levin309381fea2014-01-23 15:52:54 -08001959 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001960 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001961
Alex Shid25b5bd2020-12-15 12:34:16 -08001962 if (TestClearPageLRU(page)) {
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001963 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001964
Alex Shid25b5bd2020-12-15 12:34:16 -08001965 get_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08001966 lruvec = lock_page_lruvec_irq(page);
Yu Zhao46ae6b22021-02-24 12:08:25 -08001967 del_page_from_lru_list(page, lruvec);
Alex Shi6168d0d2020-12-15 12:34:29 -08001968 unlock_page_lruvec_irq(lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08001969 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001970 }
Alex Shid25b5bd2020-12-15 12:34:16 -08001971
Nick Piggin62695a82008-10-18 20:26:09 -07001972 return ret;
1973}
1974
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001975/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001976 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
Xianting Tian178821b2019-11-30 17:56:05 -08001977 * then get rescheduled. When there are massive number of tasks doing page
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001978 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1979 * the LRU list will go small and be scanned faster than necessary, leading to
1980 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001981 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001982static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001983 struct scan_control *sc)
1984{
1985 unsigned long inactive, isolated;
1986
1987 if (current_is_kswapd())
1988 return 0;
1989
Johannes Weinerb5ead352019-11-30 17:55:40 -08001990 if (!writeback_throttling_sane(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001991 return 0;
1992
1993 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001994 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1995 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001996 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001997 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1998 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001999 }
2000
Fengguang Wu3cf23842012-12-18 14:23:31 -08002001 /*
2002 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2003 * won't get blocked by normal direct-reclaimers, forming a circular
2004 * deadlock.
2005 */
Mel Gormand0164ad2015-11-06 16:28:21 -08002006 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08002007 inactive >>= 3;
2008
Rik van Riel35cd7812009-09-21 17:01:38 -07002009 return isolated > inactive;
2010}
2011
Kirill Tkhaia222f342019-05-13 17:17:00 -07002012/*
Hugh Dickins15b44732020-12-15 14:21:31 -08002013 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2014 * On return, @list is reused as a list of pages to be freed by the caller.
Kirill Tkhaia222f342019-05-13 17:17:00 -07002015 *
2016 * Returns the number of pages moved to the given lruvec.
2017 */
Kirill Tkhaia222f342019-05-13 17:17:00 -07002018static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
2019 struct list_head *list)
Mel Gorman66635622010-08-09 17:19:30 -07002020{
Kirill Tkhaia222f342019-05-13 17:17:00 -07002021 int nr_pages, nr_moved = 0;
Hugh Dickins3f797682012-01-12 17:20:07 -08002022 LIST_HEAD(pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07002023 struct page *page;
Mel Gorman66635622010-08-09 17:19:30 -07002024
Kirill Tkhaia222f342019-05-13 17:17:00 -07002025 while (!list_empty(list)) {
2026 page = lru_to_page(list);
Sasha Levin309381fea2014-01-23 15:52:54 -08002027 VM_BUG_ON_PAGE(PageLRU(page), page);
Alex Shi3d06afa2020-12-15 12:33:37 -08002028 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07002029 if (unlikely(!page_evictable(page))) {
Alex Shi6168d0d2020-12-15 12:34:29 -08002030 spin_unlock_irq(&lruvec->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07002031 putback_lru_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08002032 spin_lock_irq(&lruvec->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07002033 continue;
2034 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002035
Alex Shi3d06afa2020-12-15 12:33:37 -08002036 /*
2037 * The SetPageLRU needs to be kept here for list integrity.
2038 * Otherwise:
2039 * #0 move_pages_to_lru #1 release_pages
2040 * if !put_page_testzero
2041 * if (put_page_testzero())
2042 * !PageLRU //skip lru_lock
2043 * SetPageLRU()
2044 * list_add(&page->lru,)
2045 * list_add(&page->lru,)
2046 */
Linus Torvalds7a608572011-01-17 14:42:19 -08002047 SetPageLRU(page);
Kirill Tkhaia222f342019-05-13 17:17:00 -07002048
Alex Shi3d06afa2020-12-15 12:33:37 -08002049 if (unlikely(put_page_testzero(page))) {
Yu Zhao87560172021-02-24 12:08:28 -08002050 __clear_page_lru_flags(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002051
2052 if (unlikely(PageCompound(page))) {
Alex Shi6168d0d2020-12-15 12:34:29 -08002053 spin_unlock_irq(&lruvec->lru_lock);
Matthew Wilcox (Oracle)ff45fc32020-06-03 16:01:09 -07002054 destroy_compound_page(page);
Alex Shi6168d0d2020-12-15 12:34:29 -08002055 spin_lock_irq(&lruvec->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002056 } else
2057 list_add(&page->lru, &pages_to_free);
Alex Shi3d06afa2020-12-15 12:33:37 -08002058
2059 continue;
Mel Gorman66635622010-08-09 17:19:30 -07002060 }
Alex Shi3d06afa2020-12-15 12:33:37 -08002061
Alex Shiafca9152020-12-15 12:34:02 -08002062 /*
2063 * All pages were isolated from the same lruvec (and isolation
2064 * inhibits memcg migration).
2065 */
Alexander Duyck2a5e4e32020-12-15 12:34:33 -08002066 VM_BUG_ON_PAGE(!lruvec_holds_page_lru_lock(page, lruvec), page);
Yu Zhao3a9c9782021-02-24 12:08:17 -08002067 add_page_to_lru_list(page, lruvec);
Alex Shi3d06afa2020-12-15 12:33:37 -08002068 nr_pages = thp_nr_pages(page);
Alex Shi3d06afa2020-12-15 12:33:37 -08002069 nr_moved += nr_pages;
2070 if (PageActive(page))
2071 workingset_age_nonresident(lruvec, nr_pages);
Mel Gorman66635622010-08-09 17:19:30 -07002072 }
Mel Gorman66635622010-08-09 17:19:30 -07002073
Hugh Dickins3f797682012-01-12 17:20:07 -08002074 /*
2075 * To save our caller's stack, now use input list for pages to free.
2076 */
Kirill Tkhaia222f342019-05-13 17:17:00 -07002077 list_splice(&pages_to_free, list);
2078
2079 return nr_moved;
Mel Gorman66635622010-08-09 17:19:30 -07002080}
2081
2082/*
NeilBrown399ba0b2014-06-04 16:07:42 -07002083 * If a kernel thread (such as nfsd for loop-back mounts) services
NeilBrowna37b0712020-06-01 21:48:18 -07002084 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
NeilBrown399ba0b2014-06-04 16:07:42 -07002085 * In that case we should only throttle if the backing device it is
2086 * writing to is congested. In other cases it is safe to throttle.
2087 */
2088static int current_may_throttle(void)
2089{
NeilBrowna37b0712020-06-01 21:48:18 -07002090 return !(current->flags & PF_LOCAL_THROTTLE) ||
NeilBrown399ba0b2014-06-04 16:07:42 -07002091 current->backing_dev_info == NULL ||
2092 bdi_write_congested(current->backing_dev_info);
2093}
2094
2095/*
Mel Gormanb2e18752016-07-28 15:45:37 -07002096 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08002097 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 */
Mel Gorman66635622010-08-09 17:19:30 -07002099static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002100shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002101 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102{
2103 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002104 unsigned long nr_scanned;
Maninder Singh730ec8c2020-06-03 16:01:18 -07002105 unsigned int nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002106 unsigned long nr_taken;
Kirill Tkhai060f0052019-03-05 15:48:15 -08002107 struct reclaim_stat stat;
Johannes Weiner497a6c12020-06-03 16:02:34 -07002108 bool file = is_file_lru(lru);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002109 enum vm_event_item item;
Mel Gorman599d0c92016-07-28 15:45:31 -07002110 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Michal Hockodb73ee02017-09-06 16:21:11 -07002111 bool stalled = false;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07002112
Mel Gorman599d0c92016-07-28 15:45:31 -07002113 while (unlikely(too_many_isolated(pgdat, file, sc))) {
Michal Hockodb73ee02017-09-06 16:21:11 -07002114 if (stalled)
2115 return 0;
2116
2117 /* wait a bit for the reclaimer. */
2118 msleep(100);
2119 stalled = true;
Rik van Riel35cd7812009-09-21 17:01:38 -07002120
2121 /* We are about to die and free our memory. Return now. */
2122 if (fatal_signal_pending(current))
2123 return SWAP_CLUSTER_MAX;
2124 }
2125
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002127
Alex Shi6168d0d2020-12-15 12:34:29 -08002128 spin_lock_irq(&lruvec->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002130 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002131 &nr_scanned, sc, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07002132
Mel Gorman599d0c92016-07-28 15:45:31 -07002133 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002134 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08002135 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002136 __count_vm_events(item, nr_scanned);
2137 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002138 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2139
Alex Shi6168d0d2020-12-15 12:34:29 -08002140 spin_unlock_irq(&lruvec->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002141
Hillf Dantond563c052012-03-21 16:34:02 -07002142 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07002143 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002144
Shakeel Butt013339d2020-12-14 19:06:39 -08002145 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07002146
Alex Shi6168d0d2020-12-15 12:34:29 -08002147 spin_lock_irq(&lruvec->lru_lock);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002148 move_pages_to_lru(lruvec, &page_list);
2149
2150 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002151 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
Johannes Weinerb5ead352019-11-30 17:55:40 -08002152 if (!cgroup_reclaim(sc))
Kirill Tkhaif46b7912019-05-13 17:22:33 -07002153 __count_vm_events(item, nr_reclaimed);
2154 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
Johannes Weiner497a6c12020-06-03 16:02:34 -07002155 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
Alex Shi6168d0d2020-12-15 12:34:29 -08002156 spin_unlock_irq(&lruvec->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08002157
Alex Shi75cc3c92020-12-15 14:20:50 -08002158 lru_note_cost(lruvec, file, stat.nr_pageout);
Johannes Weiner747db952014-08-08 14:19:24 -07002159 mem_cgroup_uncharge_list(&page_list);
Mel Gorman2d4894b2017-11-15 17:37:59 -08002160 free_unref_page_list(&page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07002161
Mel Gorman92df3a72011-10-31 17:07:56 -07002162 /*
Andrey Ryabinin1c610d52018-03-22 16:17:42 -07002163 * If dirty pages are scanned that are not queued for IO, it
2164 * implies that flushers are not doing their job. This can
2165 * happen when memory pressure pushes dirty pages to the end of
2166 * the LRU before the dirty limits are breached and the dirty
2167 * data has expired. It can also happen when the proportion of
2168 * dirty pages grows not through writes but through memory
2169 * pressure reclaiming all the clean cache. And in some cases,
2170 * the flushers simply cannot keep up with the allocation
2171 * rate. Nudge the flusher threads in case they are asleep.
2172 */
2173 if (stat.nr_unqueued_dirty == nr_taken)
2174 wakeup_flusher_threads(WB_REASON_VMSCAN);
2175
Andrey Ryabinind108c772018-04-10 16:27:59 -07002176 sc->nr.dirty += stat.nr_dirty;
2177 sc->nr.congested += stat.nr_congested;
2178 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2179 sc->nr.writeback += stat.nr_writeback;
2180 sc->nr.immediate += stat.nr_immediate;
2181 sc->nr.taken += nr_taken;
2182 if (file)
2183 sc->nr.file_taken += nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07002184
Mel Gorman599d0c92016-07-28 15:45:31 -07002185 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
Steven Rostedtd51d1e62018-04-10 16:28:07 -07002186 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08002187 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188}
2189
Hugh Dickins15b44732020-12-15 14:21:31 -08002190/*
2191 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2192 *
2193 * We move them the other way if the page is referenced by one or more
2194 * processes.
2195 *
2196 * If the pages are mostly unmapped, the processing is fast and it is
2197 * appropriate to hold lru_lock across the whole operation. But if
2198 * the pages are mapped, the processing is slow (page_referenced()), so
2199 * we should drop lru_lock around each page. It's impossible to balance
2200 * this, so instead we remove the pages from the LRU while processing them.
2201 * It is safe to rely on PG_active against the non-LRU pages in here because
2202 * nobody will play with that bit on a non-LRU page.
2203 *
2204 * The downside is that we have to touch page->_refcount against each page.
2205 * But we had to alter page->flags anyway.
2206 */
Hugh Dickinsf6260122012-01-12 17:20:06 -08002207static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002208 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002209 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002210 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002212 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08002213 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07002214 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002216 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002217 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 struct page *page;
Michal Hocko9d998b42017-02-22 15:44:18 -08002219 unsigned nr_deactivate, nr_activate;
2220 unsigned nr_rotated = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002221 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002222 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
2224 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002225
Alex Shi6168d0d2020-12-15 12:34:29 -08002226 spin_lock_irq(&lruvec->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002227
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002228 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002229 &nr_scanned, sc, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002230
Mel Gorman599d0c92016-07-28 15:45:31 -07002231 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002232
Shakeel Butt912c0572020-08-06 23:26:32 -07002233 if (!cgroup_reclaim(sc))
2234 __count_vm_events(PGREFILL, nr_scanned);
Yafang Shao2fa26902019-05-13 17:23:02 -07002235 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002236
Alex Shi6168d0d2020-12-15 12:34:29 -08002237 spin_unlock_irq(&lruvec->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 while (!list_empty(&l_hold)) {
2240 cond_resched();
2241 page = lru_to_page(&l_hold);
2242 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002243
Hugh Dickins39b5f292012-10-08 16:33:18 -07002244 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002245 putback_lru_page(page);
2246 continue;
2247 }
2248
Mel Gormancc715d92012-03-21 16:34:00 -07002249 if (unlikely(buffer_heads_over_limit)) {
2250 if (page_has_private(page) && trylock_page(page)) {
2251 if (page_has_private(page))
2252 try_to_release_page(page, 0);
2253 unlock_page(page);
2254 }
2255 }
2256
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002257 if (page_referenced(page, 0, sc->target_mem_cgroup,
2258 &vm_flags)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002259 /*
2260 * Identify referenced, file-backed active pages and
2261 * give them one more trip around the active list. So
2262 * that executable code get better chances to stay in
2263 * memory under moderate memory pressure. Anon pages
2264 * are not likely to be evicted by use-once streaming
2265 * IO, plus JVM can create lots of anon VM_EXEC pages,
2266 * so we ignore them here.
2267 */
Huang Ying9de4f222020-04-06 20:04:41 -07002268 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
Matthew Wilcox (Oracle)6c357842020-08-14 17:30:37 -07002269 nr_rotated += thp_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002270 list_add(&page->lru, &l_active);
2271 continue;
2272 }
2273 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002274
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002275 ClearPageActive(page); /* we are de-activating */
Johannes Weiner1899ad12018-10-26 15:06:04 -07002276 SetPageWorkingset(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 list_add(&page->lru, &l_inactive);
2278 }
2279
Andrew Mortonb5557492009-01-06 14:40:13 -08002280 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002281 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002282 */
Alex Shi6168d0d2020-12-15 12:34:29 -08002283 spin_lock_irq(&lruvec->lru_lock);
Rik van Riel556adec2008-10-18 20:26:34 -07002284
Kirill Tkhaia222f342019-05-13 17:17:00 -07002285 nr_activate = move_pages_to_lru(lruvec, &l_active);
2286 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
Kirill Tkhaif372d892019-05-13 17:16:57 -07002287 /* Keep all free pages in l_active list */
2288 list_splice(&l_inactive, &l_active);
Kirill Tkhai9851ac12019-05-13 17:16:54 -07002289
2290 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2291 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2292
Mel Gorman599d0c92016-07-28 15:45:31 -07002293 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Alex Shi6168d0d2020-12-15 12:34:29 -08002294 spin_unlock_irq(&lruvec->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002295
Kirill Tkhaif372d892019-05-13 17:16:57 -07002296 mem_cgroup_uncharge_list(&l_active);
2297 free_unref_page_list(&l_active);
Michal Hocko9d998b42017-02-22 15:44:18 -08002298 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2299 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300}
2301
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002302unsigned long reclaim_pages(struct list_head *page_list)
2303{
Yang Shif661d002020-04-01 21:10:05 -07002304 int nid = NUMA_NO_NODE;
Maninder Singh730ec8c2020-06-03 16:01:18 -07002305 unsigned int nr_reclaimed = 0;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002306 LIST_HEAD(node_page_list);
2307 struct reclaim_stat dummy_stat;
2308 struct page *page;
2309 struct scan_control sc = {
2310 .gfp_mask = GFP_KERNEL,
2311 .priority = DEF_PRIORITY,
2312 .may_writepage = 1,
2313 .may_unmap = 1,
2314 .may_swap = 1,
2315 };
2316
2317 while (!list_empty(page_list)) {
2318 page = lru_to_page(page_list);
Yang Shif661d002020-04-01 21:10:05 -07002319 if (nid == NUMA_NO_NODE) {
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002320 nid = page_to_nid(page);
2321 INIT_LIST_HEAD(&node_page_list);
2322 }
2323
2324 if (nid == page_to_nid(page)) {
2325 ClearPageActive(page);
2326 list_move(&page->lru, &node_page_list);
2327 continue;
2328 }
2329
2330 nr_reclaimed += shrink_page_list(&node_page_list,
2331 NODE_DATA(nid),
Shakeel Butt013339d2020-12-14 19:06:39 -08002332 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002333 while (!list_empty(&node_page_list)) {
2334 page = lru_to_page(&node_page_list);
2335 list_del(&page->lru);
2336 putback_lru_page(page);
2337 }
2338
Yang Shif661d002020-04-01 21:10:05 -07002339 nid = NUMA_NO_NODE;
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002340 }
2341
2342 if (!list_empty(&node_page_list)) {
2343 nr_reclaimed += shrink_page_list(&node_page_list,
2344 NODE_DATA(nid),
Shakeel Butt013339d2020-12-14 19:06:39 -08002345 &sc, &dummy_stat, false);
Minchan Kim1a4e58c2019-09-25 16:49:15 -07002346 while (!list_empty(&node_page_list)) {
2347 page = lru_to_page(&node_page_list);
2348 list_del(&page->lru);
2349 putback_lru_page(page);
2350 }
2351 }
2352
2353 return nr_reclaimed;
2354}
2355
Johannes Weinerb91ac372019-11-30 17:56:02 -08002356static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2357 struct lruvec *lruvec, struct scan_control *sc)
2358{
2359 if (is_active_lru(lru)) {
2360 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2361 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2362 else
2363 sc->skipped_deactivate = 1;
2364 return 0;
2365 }
2366
2367 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2368}
2369
Rik van Riel59dc76b2016-05-20 16:56:31 -07002370/*
2371 * The inactive anon list should be small enough that the VM never has
2372 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002373 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002374 * The inactive file list should be small enough to leave most memory
2375 * to the established workingset on the scan-resistant active list,
2376 * but large enough to avoid thrashing the aggregate readahead window.
2377 *
2378 * Both inactive lists should also be large enough that each inactive
2379 * page has a chance to be referenced again before it is reclaimed.
2380 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002381 * If that fails and refaulting is observed, the inactive list grows.
2382 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002383 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
Andrey Ryabinin3a50d142017-11-15 17:34:15 -08002384 * on this LRU, maintained by the pageout code. An inactive_ratio
Rik van Riel59dc76b2016-05-20 16:56:31 -07002385 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2386 *
2387 * total target max
2388 * memory ratio inactive
2389 * -------------------------------------
2390 * 10MB 1 5MB
2391 * 100MB 1 50MB
2392 * 1GB 3 250MB
2393 * 10GB 10 0.9GB
2394 * 100GB 31 3GB
2395 * 1TB 101 10GB
2396 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002397 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002398static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002399{
Johannes Weinerb91ac372019-11-30 17:56:02 -08002400 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002401 unsigned long inactive, active;
2402 unsigned long inactive_ratio;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002403 unsigned long gb;
2404
Johannes Weinerb91ac372019-11-30 17:56:02 -08002405 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2406 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002407
Johannes Weinerb91ac372019-11-30 17:56:02 -08002408 gb = (inactive + active) >> (30 - PAGE_SHIFT);
Joonsoo Kim40025702020-08-11 18:30:54 -07002409 if (gb)
Johannes Weinerb91ac372019-11-30 17:56:02 -08002410 inactive_ratio = int_sqrt(10 * gb);
2411 else
2412 inactive_ratio = 1;
Michal Hockofd538802017-02-22 15:45:58 -08002413
Rik van Riel59dc76b2016-05-20 16:56:31 -07002414 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002415}
2416
Johannes Weiner9a265112013-02-22 16:32:17 -08002417enum scan_balance {
2418 SCAN_EQUAL,
2419 SCAN_FRACT,
2420 SCAN_ANON,
2421 SCAN_FILE,
2422};
2423
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002425 * Determine how aggressively the anon and file LRU lists should be
2426 * scanned. The relative value of each set of LRU lists is determined
2427 * by looking at the fraction of the pages scanned we did rotate back
2428 * onto the active list instead of evict.
2429 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002430 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2431 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002432 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002433static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2434 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002435{
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002436 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002437 unsigned long anon_cost, file_cost, total_cost;
Vladimir Davydov33377672016-01-20 15:02:59 -08002438 int swappiness = mem_cgroup_swappiness(memcg);
Yu Zhaoed017372020-10-15 20:09:55 -07002439 u64 fraction[ANON_AND_FILE];
Johannes Weiner9a265112013-02-22 16:32:17 -08002440 u64 denominator = 0; /* gcc */
Johannes Weiner9a265112013-02-22 16:32:17 -08002441 enum scan_balance scan_balance;
Johannes Weiner9a265112013-02-22 16:32:17 -08002442 unsigned long ap, fp;
2443 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002444
2445 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002446 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002447 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002448 goto out;
2449 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002450
Johannes Weiner10316b32013-02-22 16:32:14 -08002451 /*
2452 * Global reclaim will swap to prevent OOM even with no
2453 * swappiness, but memcg users want to use this knob to
2454 * disable swapping for individual groups completely when
2455 * using the memory controller's swap limit feature would be
2456 * too expensive.
2457 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08002458 if (cgroup_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002459 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002460 goto out;
2461 }
2462
2463 /*
2464 * Do not apply any pressure balancing cleverness when the
2465 * system is close to OOM, scan both anon and file equally
2466 * (unless the swappiness setting disagrees with swapping).
2467 */
Johannes Weiner02695172014-08-06 16:06:17 -07002468 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002469 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002470 goto out;
2471 }
2472
Johannes Weiner11d16c22013-02-22 16:32:15 -08002473 /*
Johannes Weiner53138ce2019-11-30 17:55:56 -08002474 * If the system is almost out of file pages, force-scan anon.
Johannes Weiner62376252014-05-06 12:50:07 -07002475 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002476 if (sc->file_is_tiny) {
Johannes Weiner53138ce2019-11-30 17:55:56 -08002477 scan_balance = SCAN_ANON;
2478 goto out;
Johannes Weiner62376252014-05-06 12:50:07 -07002479 }
2480
2481 /*
Johannes Weinerb91ac372019-11-30 17:56:02 -08002482 * If there is enough inactive page cache, we do not reclaim
2483 * anything from the anonymous working right now.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002484 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002485 if (sc->cache_trim_mode) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002486 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002487 goto out;
2488 }
2489
Johannes Weiner9a265112013-02-22 16:32:17 -08002490 scan_balance = SCAN_FRACT;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002491 /*
Johannes Weiner314b57f2020-06-03 16:03:03 -07002492 * Calculate the pressure balance between anon and file pages.
2493 *
2494 * The amount of pressure we put on each LRU is inversely
2495 * proportional to the cost of reclaiming each list, as
2496 * determined by the share of pages that are refaulting, times
2497 * the relative IO cost of bringing back a swapped out
2498 * anonymous page vs reloading a filesystem page (swappiness).
2499 *
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002500 * Although we limit that influence to ensure no list gets
2501 * left behind completely: at least a third of the pressure is
2502 * applied, before swappiness.
2503 *
Johannes Weiner314b57f2020-06-03 16:03:03 -07002504 * With swappiness at 100, anon and file have equal IO cost.
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002505 */
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002506 total_cost = sc->anon_cost + sc->file_cost;
2507 anon_cost = total_cost + sc->anon_cost;
2508 file_cost = total_cost + sc->file_cost;
2509 total_cost = anon_cost + file_cost;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002510
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002511 ap = swappiness * (total_cost + 1);
2512 ap /= anon_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002513
Johannes Weinerd483a5d2020-06-03 16:03:13 -07002514 fp = (200 - swappiness) * (total_cost + 1);
2515 fp /= file_cost + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002516
Shaohua Li76a33fc2010-05-24 14:32:36 -07002517 fraction[0] = ap;
2518 fraction[1] = fp;
Johannes Weinera4fe1632020-06-03 16:02:50 -07002519 denominator = ap + fp;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002520out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002521 for_each_evictable_lru(lru) {
2522 int file = is_file_lru(lru);
Chris Down9783aa92019-10-06 17:58:32 -07002523 unsigned long lruvec_size;
Johannes Weiner688035f2017-05-03 14:52:07 -07002524 unsigned long scan;
Chris Down1bc63fb2019-10-06 17:58:38 -07002525 unsigned long protection;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002526
Chris Down9783aa92019-10-06 17:58:32 -07002527 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
Yafang Shao22f74962020-08-06 23:22:01 -07002528 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2529 memcg,
Chris Down1bc63fb2019-10-06 17:58:38 -07002530 sc->memcg_low_reclaim);
Chris Down9783aa92019-10-06 17:58:32 -07002531
Chris Down1bc63fb2019-10-06 17:58:38 -07002532 if (protection) {
Chris Down9783aa92019-10-06 17:58:32 -07002533 /*
2534 * Scale a cgroup's reclaim pressure by proportioning
2535 * its current usage to its memory.low or memory.min
2536 * setting.
2537 *
2538 * This is important, as otherwise scanning aggression
2539 * becomes extremely binary -- from nothing as we
2540 * approach the memory protection threshold, to totally
2541 * nominal as we exceed it. This results in requiring
2542 * setting extremely liberal protection thresholds. It
2543 * also means we simply get no protection at all if we
2544 * set it too low, which is not ideal.
Chris Down1bc63fb2019-10-06 17:58:38 -07002545 *
2546 * If there is any protection in place, we reduce scan
2547 * pressure by how much of the total memory used is
2548 * within protection thresholds.
Chris Down9783aa92019-10-06 17:58:32 -07002549 *
Chris Down9de7ca42019-10-06 17:58:35 -07002550 * There is one special case: in the first reclaim pass,
2551 * we skip over all groups that are within their low
2552 * protection. If that fails to reclaim enough pages to
2553 * satisfy the reclaim goal, we come back and override
2554 * the best-effort low protection. However, we still
2555 * ideally want to honor how well-behaved groups are in
2556 * that case instead of simply punishing them all
2557 * equally. As such, we reclaim them based on how much
Chris Down1bc63fb2019-10-06 17:58:38 -07002558 * memory they are using, reducing the scan pressure
2559 * again by how much of the total memory used is under
2560 * hard protection.
Chris Down9783aa92019-10-06 17:58:32 -07002561 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002562 unsigned long cgroup_size = mem_cgroup_size(memcg);
2563
2564 /* Avoid TOCTOU with earlier protection check */
2565 cgroup_size = max(cgroup_size, protection);
2566
2567 scan = lruvec_size - lruvec_size * protection /
2568 cgroup_size;
Chris Down9783aa92019-10-06 17:58:32 -07002569
2570 /*
Chris Down1bc63fb2019-10-06 17:58:38 -07002571 * Minimally target SWAP_CLUSTER_MAX pages to keep
Ethon Paul55b65a52020-06-04 16:49:10 -07002572 * reclaim moving forwards, avoiding decrementing
Chris Down9de7ca42019-10-06 17:58:35 -07002573 * sc->priority further than desirable.
Chris Down9783aa92019-10-06 17:58:32 -07002574 */
Chris Down1bc63fb2019-10-06 17:58:38 -07002575 scan = max(scan, SWAP_CLUSTER_MAX);
Chris Down9783aa92019-10-06 17:58:32 -07002576 } else {
2577 scan = lruvec_size;
2578 }
2579
2580 scan >>= sc->priority;
2581
Johannes Weiner688035f2017-05-03 14:52:07 -07002582 /*
2583 * If the cgroup's already been deleted, make sure to
2584 * scrape out the remaining cache.
2585 */
2586 if (!scan && !mem_cgroup_online(memcg))
Chris Down9783aa92019-10-06 17:58:32 -07002587 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002588
Johannes Weiner688035f2017-05-03 14:52:07 -07002589 switch (scan_balance) {
2590 case SCAN_EQUAL:
2591 /* Scan lists relative to size */
2592 break;
2593 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002594 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002595 * Scan types proportional to swappiness and
2596 * their relative recent reclaim efficiency.
Gavin Shan76073c62020-02-20 20:04:24 -08002597 * Make sure we don't miss the last page on
2598 * the offlined memory cgroups because of a
2599 * round-off error.
Johannes Weiner9a265112013-02-22 16:32:17 -08002600 */
Gavin Shan76073c62020-02-20 20:04:24 -08002601 scan = mem_cgroup_online(memcg) ?
2602 div64_u64(scan * fraction[file], denominator) :
2603 DIV64_U64_ROUND_UP(scan * fraction[file],
Roman Gushchin68600f62018-10-26 15:03:27 -07002604 denominator);
Johannes Weiner688035f2017-05-03 14:52:07 -07002605 break;
2606 case SCAN_FILE:
2607 case SCAN_ANON:
2608 /* Scan one type exclusively */
Mateusz Noseke072bff2020-04-01 21:10:15 -07002609 if ((scan_balance == SCAN_FILE) != file)
Johannes Weiner688035f2017-05-03 14:52:07 -07002610 scan = 0;
Johannes Weiner688035f2017-05-03 14:52:07 -07002611 break;
2612 default:
2613 /* Look ma, no brain */
2614 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002615 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002616
Johannes Weiner688035f2017-05-03 14:52:07 -07002617 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002618 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002619}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002620
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002621static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002622{
2623 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002624 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002625 unsigned long nr_to_scan;
2626 enum lru_list lru;
2627 unsigned long nr_reclaimed = 0;
2628 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2629 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002630 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002631
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002632 get_scan_count(lruvec, sc, nr);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002633
Mel Gormane82e0562013-07-03 15:01:44 -07002634 /* Record the original scan target for proportional adjustments later */
2635 memcpy(targets, nr, sizeof(nr));
2636
Mel Gorman1a501902014-06-04 16:10:49 -07002637 /*
2638 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2639 * event that can occur when there is little memory pressure e.g.
2640 * multiple streaming readers/writers. Hence, we do not abort scanning
2641 * when the requested number of pages are reclaimed when scanning at
2642 * DEF_PRIORITY on the assumption that the fact we are direct
2643 * reclaiming implies that kswapd is not keeping up and it is best to
2644 * do a batch of work at once. For memcg reclaim one check is made to
2645 * abort proportional reclaim if either the file or anon lru has already
2646 * dropped to zero at the first pass.
2647 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08002648 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
Mel Gorman1a501902014-06-04 16:10:49 -07002649 sc->priority == DEF_PRIORITY);
2650
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002651 blk_start_plug(&plug);
2652 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2653 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002654 unsigned long nr_anon, nr_file, percentage;
2655 unsigned long nr_scanned;
2656
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002657 for_each_evictable_lru(lru) {
2658 if (nr[lru]) {
2659 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2660 nr[lru] -= nr_to_scan;
2661
2662 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner3b991202019-04-18 17:50:34 -07002663 lruvec, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002664 }
2665 }
Mel Gormane82e0562013-07-03 15:01:44 -07002666
Michal Hockobd041732016-12-02 17:26:48 -08002667 cond_resched();
2668
Mel Gormane82e0562013-07-03 15:01:44 -07002669 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2670 continue;
2671
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002672 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002673 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002674 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002675 * proportionally what was requested by get_scan_count(). We
2676 * stop reclaiming one LRU and reduce the amount scanning
2677 * proportional to the original scan target.
2678 */
2679 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2680 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2681
Mel Gorman1a501902014-06-04 16:10:49 -07002682 /*
2683 * It's just vindictive to attack the larger once the smaller
2684 * has gone to zero. And given the way we stop scanning the
2685 * smaller below, this makes sure that we only make one nudge
2686 * towards proportionality once we've got nr_to_reclaim.
2687 */
2688 if (!nr_file || !nr_anon)
2689 break;
2690
Mel Gormane82e0562013-07-03 15:01:44 -07002691 if (nr_file > nr_anon) {
2692 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2693 targets[LRU_ACTIVE_ANON] + 1;
2694 lru = LRU_BASE;
2695 percentage = nr_anon * 100 / scan_target;
2696 } else {
2697 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2698 targets[LRU_ACTIVE_FILE] + 1;
2699 lru = LRU_FILE;
2700 percentage = nr_file * 100 / scan_target;
2701 }
2702
2703 /* Stop scanning the smaller of the LRU */
2704 nr[lru] = 0;
2705 nr[lru + LRU_ACTIVE] = 0;
2706
2707 /*
2708 * Recalculate the other LRU scan count based on its original
2709 * scan target and the percentage scanning already complete
2710 */
2711 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2712 nr_scanned = targets[lru] - nr[lru];
2713 nr[lru] = targets[lru] * (100 - percentage) / 100;
2714 nr[lru] -= min(nr[lru], nr_scanned);
2715
2716 lru += LRU_ACTIVE;
2717 nr_scanned = targets[lru] - nr[lru];
2718 nr[lru] = targets[lru] * (100 - percentage) / 100;
2719 nr[lru] -= min(nr[lru], nr_scanned);
2720
2721 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002722 }
2723 blk_finish_plug(&plug);
2724 sc->nr_reclaimed += nr_reclaimed;
2725
2726 /*
2727 * Even if we did not try to evict anon pages at all, we want to
2728 * rebalance the anon lru active/inactive ratio.
2729 */
Johannes Weinerb91ac372019-11-30 17:56:02 -08002730 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002731 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2732 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002733}
2734
Mel Gorman23b9da52012-05-29 15:06:20 -07002735/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002736static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002737{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002738 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002739 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002740 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002741 return true;
2742
2743 return false;
2744}
2745
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002746/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002747 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2748 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2749 * true if more pages should be reclaimed such that when the page allocator
Qiwu Chendf3a45f2020-06-03 16:01:21 -07002750 * calls try_to_compact_pages() that it will have enough free pages to succeed.
Mel Gorman23b9da52012-05-29 15:06:20 -07002751 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002752 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002753static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002754 unsigned long nr_reclaimed,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002755 struct scan_control *sc)
2756{
2757 unsigned long pages_for_compaction;
2758 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002759 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002760
2761 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002762 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002763 return false;
2764
Vlastimil Babka5ee04712019-09-23 15:37:29 -07002765 /*
2766 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2767 * number of pages that were scanned. This will return to the caller
2768 * with the risk reclaim/compaction and the resulting allocation attempt
2769 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2770 * allocations through requiring that the full LRU list has been scanned
2771 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2772 * scan, but that approximation was wrong, and there were corner cases
2773 * where always a non-zero amount of pages were scanned.
2774 */
2775 if (!nr_reclaimed)
2776 return false;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002777
Mel Gorman3e7d3442011-01-13 15:45:56 -08002778 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002779 for (z = 0; z <= sc->reclaim_idx; z++) {
2780 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002781 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002782 continue;
2783
2784 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002785 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002786 case COMPACT_CONTINUE:
2787 return false;
2788 default:
2789 /* check next zone */
2790 ;
2791 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002792 }
Hillf Danton1c6c1592019-09-23 15:37:26 -07002793
2794 /*
2795 * If we have not reclaimed enough pages for compaction and the
2796 * inactive lists are large enough, continue reclaiming
2797 */
2798 pages_for_compaction = compact_gap(sc->order);
2799 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2800 if (get_nr_swap_pages() > 0)
2801 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2802
Vlastimil Babka5ee04712019-09-23 15:37:29 -07002803 return inactive_lru_pages > pages_for_compaction;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002804}
2805
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002806static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002807{
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002808 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
Johannes Weinerd2af3392019-11-30 17:55:43 -08002809 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002810
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002811 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
Johannes Weiner56600482012-01-12 17:17:59 -08002812 do {
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002813 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Johannes Weinerd2af3392019-11-30 17:55:43 -08002814 unsigned long reclaimed;
2815 unsigned long scanned;
Johannes Weiner56600482012-01-12 17:17:59 -08002816
Xunlei Pange3336ca2020-09-04 16:35:27 -07002817 /*
2818 * This loop can become CPU-bound when target memcgs
2819 * aren't eligible for reclaim - either because they
2820 * don't have any reclaimable pages, or because their
2821 * memory is explicitly protected. Avoid soft lockups.
2822 */
2823 cond_resched();
2824
Chris Down45c7f7e2020-08-06 23:22:05 -07002825 mem_cgroup_calculate_protection(target_memcg, memcg);
2826
2827 if (mem_cgroup_below_min(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08002828 /*
2829 * Hard protection.
2830 * If there is no reclaimable memory, OOM.
2831 */
2832 continue;
Chris Down45c7f7e2020-08-06 23:22:05 -07002833 } else if (mem_cgroup_below_low(memcg)) {
Johannes Weinerd2af3392019-11-30 17:55:43 -08002834 /*
2835 * Soft protection.
2836 * Respect the protection only as long as
2837 * there is an unprotected supply
2838 * of reclaimable memory from other cgroups.
2839 */
2840 if (!sc->memcg_low_reclaim) {
2841 sc->memcg_low_skipped = 1;
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002842 continue;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002843 }
Johannes Weinerd2af3392019-11-30 17:55:43 -08002844 memcg_memory_event(memcg, MEMCG_LOW);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002845 }
2846
Johannes Weinerd2af3392019-11-30 17:55:43 -08002847 reclaimed = sc->nr_reclaimed;
2848 scanned = sc->nr_scanned;
Johannes Weinerafaf07a2019-11-30 17:55:46 -08002849
2850 shrink_lruvec(lruvec, sc);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002851
Johannes Weinerd2af3392019-11-30 17:55:43 -08002852 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2853 sc->priority);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002854
Johannes Weinerd2af3392019-11-30 17:55:43 -08002855 /* Record the group's reclaim efficiency */
2856 vmpressure(sc->gfp_mask, memcg, false,
2857 sc->nr_scanned - scanned,
2858 sc->nr_reclaimed - reclaimed);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002859
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002860 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2861}
2862
Liu Song6c9e09072020-01-30 22:14:08 -08002863static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002864{
2865 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002866 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner1b051172019-11-30 17:55:52 -08002867 struct lruvec *target_lruvec;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002868 bool reclaimable = false;
Johannes Weinerb91ac372019-11-30 17:56:02 -08002869 unsigned long file;
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002870
Johannes Weiner1b051172019-11-30 17:55:52 -08002871 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2872
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002873again:
2874 memset(&sc->nr, 0, sizeof(sc->nr));
2875
2876 nr_reclaimed = sc->nr_reclaimed;
2877 nr_scanned = sc->nr_scanned;
2878
Johannes Weiner53138ce2019-11-30 17:55:56 -08002879 /*
Johannes Weiner7cf111b2020-06-03 16:03:06 -07002880 * Determine the scan balance between anon and file LRUs.
2881 */
Alex Shi6168d0d2020-12-15 12:34:29 -08002882 spin_lock_irq(&target_lruvec->lru_lock);
Johannes Weiner7cf111b2020-06-03 16:03:06 -07002883 sc->anon_cost = target_lruvec->anon_cost;
2884 sc->file_cost = target_lruvec->file_cost;
Alex Shi6168d0d2020-12-15 12:34:29 -08002885 spin_unlock_irq(&target_lruvec->lru_lock);
Johannes Weiner7cf111b2020-06-03 16:03:06 -07002886
2887 /*
Johannes Weinerb91ac372019-11-30 17:56:02 -08002888 * Target desirable inactive:active list ratios for the anon
2889 * and file LRU lists.
2890 */
2891 if (!sc->force_deactivate) {
2892 unsigned long refaults;
2893
Joonsoo Kim170b04b72020-08-11 18:30:43 -07002894 refaults = lruvec_page_state(target_lruvec,
2895 WORKINGSET_ACTIVATE_ANON);
2896 if (refaults != target_lruvec->refaults[0] ||
2897 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
Johannes Weinerb91ac372019-11-30 17:56:02 -08002898 sc->may_deactivate |= DEACTIVATE_ANON;
2899 else
2900 sc->may_deactivate &= ~DEACTIVATE_ANON;
2901
2902 /*
2903 * When refaults are being observed, it means a new
2904 * workingset is being established. Deactivate to get
2905 * rid of any stale active pages quickly.
2906 */
2907 refaults = lruvec_page_state(target_lruvec,
Joonsoo Kim170b04b72020-08-11 18:30:43 -07002908 WORKINGSET_ACTIVATE_FILE);
2909 if (refaults != target_lruvec->refaults[1] ||
Johannes Weinerb91ac372019-11-30 17:56:02 -08002910 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2911 sc->may_deactivate |= DEACTIVATE_FILE;
2912 else
2913 sc->may_deactivate &= ~DEACTIVATE_FILE;
2914 } else
2915 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2916
2917 /*
2918 * If we have plenty of inactive file pages that aren't
2919 * thrashing, try to reclaim those first before touching
2920 * anonymous pages.
2921 */
2922 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2923 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2924 sc->cache_trim_mode = 1;
2925 else
2926 sc->cache_trim_mode = 0;
2927
2928 /*
Johannes Weiner53138ce2019-11-30 17:55:56 -08002929 * Prevent the reclaimer from falling into the cache trap: as
2930 * cache pages start out inactive, every cache fault will tip
2931 * the scan balance towards the file LRU. And as the file LRU
2932 * shrinks, so does the window for rotation from references.
2933 * This means we have a runaway feedback loop where a tiny
2934 * thrashing file LRU becomes infinitely more attractive than
2935 * anon pages. Try to detect this based on file LRU size.
2936 */
2937 if (!cgroup_reclaim(sc)) {
Johannes Weiner53138ce2019-11-30 17:55:56 -08002938 unsigned long total_high_wmark = 0;
Johannes Weinerb91ac372019-11-30 17:56:02 -08002939 unsigned long free, anon;
2940 int z;
Johannes Weiner53138ce2019-11-30 17:55:56 -08002941
2942 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2943 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2944 node_page_state(pgdat, NR_INACTIVE_FILE);
2945
2946 for (z = 0; z < MAX_NR_ZONES; z++) {
2947 struct zone *zone = &pgdat->node_zones[z];
2948 if (!managed_zone(zone))
2949 continue;
2950
2951 total_high_wmark += high_wmark_pages(zone);
2952 }
2953
Johannes Weinerb91ac372019-11-30 17:56:02 -08002954 /*
2955 * Consider anon: if that's low too, this isn't a
2956 * runaway file reclaim problem, but rather just
2957 * extreme pressure. Reclaim as per usual then.
2958 */
2959 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2960
2961 sc->file_is_tiny =
2962 file + free <= total_high_wmark &&
2963 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2964 anon >> sc->priority;
Johannes Weiner53138ce2019-11-30 17:55:56 -08002965 }
2966
Johannes Weiner0f6a5cf2019-11-30 17:55:49 -08002967 shrink_node_memcgs(pgdat, sc);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002968
Johannes Weinerd2af3392019-11-30 17:55:43 -08002969 if (reclaim_state) {
2970 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2971 reclaim_state->reclaimed_slab = 0;
2972 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07002973
Johannes Weinerd2af3392019-11-30 17:55:43 -08002974 /* Record the subtree's reclaim efficiency */
Johannes Weiner1b051172019-11-30 17:55:52 -08002975 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Johannes Weinerd2af3392019-11-30 17:55:43 -08002976 sc->nr_scanned - nr_scanned,
2977 sc->nr_reclaimed - nr_reclaimed);
2978
2979 if (sc->nr_reclaimed - nr_reclaimed)
2980 reclaimable = true;
2981
2982 if (current_is_kswapd()) {
2983 /*
2984 * If reclaim is isolating dirty pages under writeback,
2985 * it implies that the long-lived page allocation rate
2986 * is exceeding the page laundering rate. Either the
2987 * global limits are not being effective at throttling
2988 * processes due to the page distribution throughout
2989 * zones or there is heavy usage of a slow backing
2990 * device. The only option is to throttle from reclaim
2991 * context which is not ideal as there is no guarantee
2992 * the dirtying process is throttled in the same way
2993 * balance_dirty_pages() manages.
2994 *
2995 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2996 * count the number of pages under pages flagged for
2997 * immediate reclaim and stall if any are encountered
2998 * in the nr_immediate check below.
2999 */
3000 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3001 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Andrey Ryabinind108c772018-04-10 16:27:59 -07003002
Johannes Weinerd2af3392019-11-30 17:55:43 -08003003 /* Allow kswapd to start writing pages during reclaim.*/
3004 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3005 set_bit(PGDAT_DIRTY, &pgdat->flags);
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07003006
3007 /*
Randy Dunlap1eba09c2020-08-11 18:33:26 -07003008 * If kswapd scans pages marked for immediate
Johannes Weinerd2af3392019-11-30 17:55:43 -08003009 * reclaim and under writeback (nr_immediate), it
3010 * implies that pages are cycling through the LRU
3011 * faster than they are written so also forcibly stall.
Andrey Ryabinind108c772018-04-10 16:27:59 -07003012 */
Johannes Weinerd2af3392019-11-30 17:55:43 -08003013 if (sc->nr.immediate)
3014 congestion_wait(BLK_RW_ASYNC, HZ/10);
3015 }
Andrey Ryabinind108c772018-04-10 16:27:59 -07003016
Johannes Weinerd2af3392019-11-30 17:55:43 -08003017 /*
Johannes Weiner1b051172019-11-30 17:55:52 -08003018 * Tag a node/memcg as congested if all the dirty pages
3019 * scanned were backed by a congested BDI and
3020 * wait_iff_congested will stall.
3021 *
Johannes Weinerd2af3392019-11-30 17:55:43 -08003022 * Legacy memcg will stall in page writeback so avoid forcibly
3023 * stalling in wait_iff_congested().
3024 */
Johannes Weiner1b051172019-11-30 17:55:52 -08003025 if ((current_is_kswapd() ||
3026 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
Johannes Weinerd2af3392019-11-30 17:55:43 -08003027 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
Johannes Weiner1b051172019-11-30 17:55:52 -08003028 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
Johannes Weinerd2af3392019-11-30 17:55:43 -08003029
3030 /*
3031 * Stall direct reclaim for IO completions if underlying BDIs
3032 * and node is congested. Allow kswapd to continue until it
3033 * starts encountering unqueued dirty pages or cycling through
3034 * the LRU too quickly.
3035 */
Johannes Weiner1b051172019-11-30 17:55:52 -08003036 if (!current_is_kswapd() && current_may_throttle() &&
3037 !sc->hibernation_mode &&
3038 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
Johannes Weinerd2af3392019-11-30 17:55:43 -08003039 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
3040
3041 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3042 sc))
3043 goto again;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07003044
Johannes Weinerc73322d2017-05-03 14:51:51 -07003045 /*
3046 * Kswapd gives up on balancing particular nodes after too
3047 * many failures to reclaim anything from them and goes to
3048 * sleep. On reclaim progress, reset the failure counter. A
3049 * successful direct reclaim run will revive a dormant kswapd.
3050 */
3051 if (reclaimable)
3052 pgdat->kswapd_failures = 0;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003053}
3054
Vlastimil Babka53853e22014-10-09 15:27:02 -07003055/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07003056 * Returns true if compaction should go ahead for a costly-order request, or
3057 * the allocation would already succeed without compaction. Return false if we
3058 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07003059 */
Mel Gorman4f588332016-07-28 15:46:38 -07003060static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08003061{
Mel Gorman31483b62016-07-28 15:45:46 -07003062 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07003063 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08003064
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07003065 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3066 if (suitable == COMPACT_SUCCESS)
3067 /* Allocation should succeed already. Don't reclaim. */
3068 return true;
3069 if (suitable == COMPACT_SKIPPED)
3070 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08003071 return false;
3072
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07003073 /*
3074 * Compaction is already possible, but it takes time to run and there
3075 * are potentially other callers using the pages just freed. So proceed
3076 * with reclaim to make a buffer of free pages available to give
3077 * compaction a reasonable chance of completing and allocating the page.
3078 * Note that we won't actually reclaim the whole buffer in one attempt
3079 * as the target watermark in should_continue_reclaim() is lower. But if
3080 * we are already above the high+gap watermark, don't reclaim at all.
3081 */
3082 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3083
3084 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08003085}
3086
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087/*
3088 * This is the direct reclaim path, for page-allocating processes. We only
3089 * try to reclaim pages from zones which will satisfy the caller's allocation
3090 * request.
3091 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 * If a zone is deemed to be full of pinned pages then just give it a light
3093 * scan then give up on it.
3094 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07003095static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096{
Mel Gormandd1a2392008-04-28 02:12:17 -07003097 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003098 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07003099 unsigned long nr_soft_reclaimed;
3100 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07003101 gfp_t orig_mask;
Mel Gorman79dafcdc2016-07-28 15:45:53 -07003102 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08003103
Mel Gormancc715d92012-03-21 16:34:00 -07003104 /*
3105 * If the number of buffer_heads in the machine exceeds the maximum
3106 * allowed level, force direct reclaim to scan the highmem zone as
3107 * highmem pages could be pinning lowmem pages storing buffer_heads
3108 */
Weijie Yang619d0d762014-04-07 15:36:59 -07003109 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07003110 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07003111 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07003112 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07003113 }
Mel Gormancc715d92012-03-21 16:34:00 -07003114
Mel Gormand4debc62010-08-09 17:19:29 -07003115 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07003116 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07003117 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08003118 * Take care memory controller reclaiming has small influence
3119 * to global LRU.
3120 */
Johannes Weinerb5ead352019-11-30 17:55:40 -08003121 if (!cgroup_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04003122 if (!cpuset_zone_allowed(zone,
3123 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08003124 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07003125
Johannes Weiner0b064962014-08-06 16:06:12 -07003126 /*
3127 * If we already have plenty of memory free for
3128 * compaction in this zone, don't free any more.
3129 * Even though compaction is invoked for any
3130 * non-zero order, only frequent costly order
3131 * reclamation is disruptive enough to become a
3132 * noticeable problem, like transparent huge
3133 * page allocations.
3134 */
3135 if (IS_ENABLED(CONFIG_COMPACTION) &&
3136 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07003137 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07003138 sc->compaction_ready = true;
3139 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07003140 }
Johannes Weiner0b064962014-08-06 16:06:12 -07003141
Andrew Morton0608f432013-09-24 15:27:41 -07003142 /*
Mel Gorman79dafcdc2016-07-28 15:45:53 -07003143 * Shrink each node in the zonelist once. If the
3144 * zonelist is ordered by zone (not the default) then a
3145 * node may be shrunk multiple times but in that case
3146 * the user prefers lower zones being preserved.
3147 */
3148 if (zone->zone_pgdat == last_pgdat)
3149 continue;
3150
3151 /*
Andrew Morton0608f432013-09-24 15:27:41 -07003152 * This steals pages from memory cgroups over softlimit
3153 * and returns the number of reclaimed pages and
3154 * scanned pages. This works for global memory pressure
3155 * and balancing, not for a memcg's limit.
3156 */
3157 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003158 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07003159 sc->order, sc->gfp_mask,
3160 &nr_soft_scanned);
3161 sc->nr_reclaimed += nr_soft_reclaimed;
3162 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07003163 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08003164 }
Nick Piggin408d8542006-09-25 23:31:27 -07003165
Mel Gorman79dafcdc2016-07-28 15:45:53 -07003166 /* See comment about same check for global reclaim above */
3167 if (zone->zone_pgdat == last_pgdat)
3168 continue;
3169 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07003170 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171 }
Mel Gormane0c23272011-10-31 17:09:33 -07003172
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07003173 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07003174 * Restore to original mask to avoid the impact on the caller if we
3175 * promoted it to __GFP_HIGHMEM.
3176 */
3177 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003179
Johannes Weinerb9107182019-11-30 17:55:59 -08003180static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003181{
Johannes Weinerb9107182019-11-30 17:55:59 -08003182 struct lruvec *target_lruvec;
3183 unsigned long refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003184
Johannes Weinerb9107182019-11-30 17:55:59 -08003185 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
Joonsoo Kim170b04b72020-08-11 18:30:43 -07003186 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3187 target_lruvec->refaults[0] = refaults;
3188 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3189 target_lruvec->refaults[1] = refaults;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003190}
3191
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192/*
3193 * This is the main entry point to direct page reclaim.
3194 *
3195 * If a full scan of the inactive list fails to free enough memory then we
3196 * are "out of memory" and something needs to be killed.
3197 *
3198 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3199 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02003200 * caller can't do much about. We kick the writeback threads and take explicit
3201 * naps in the hope that some of these pages can be written. But if the
3202 * allocating task holds filesystem locks which prevent writeout this might not
3203 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07003204 *
3205 * returns: 0, if no pages reclaimed
3206 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 */
Mel Gormandac1d272008-04-28 02:12:12 -07003208static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003209 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210{
Johannes Weiner241994ed2015-02-11 15:26:06 -08003211 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003212 pg_data_t *last_pgdat;
3213 struct zoneref *z;
3214 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003215retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07003216 delayacct_freepages_start();
3217
Johannes Weinerb5ead352019-11-30 17:55:40 -08003218 if (!cgroup_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07003219 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003221 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07003222 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3223 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08003224 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07003225 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07003226
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003227 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07003228 break;
3229
3230 if (sc->compaction_ready)
3231 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232
3233 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08003234 * If we're getting trouble reclaiming, start doing
3235 * writepage even in laptop mode.
3236 */
3237 if (sc->priority < DEF_PRIORITY - 2)
3238 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07003239 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003240
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003241 last_pgdat = NULL;
3242 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3243 sc->nodemask) {
3244 if (zone->zone_pgdat == last_pgdat)
3245 continue;
3246 last_pgdat = zone->zone_pgdat;
Johannes Weiner1b051172019-11-30 17:55:52 -08003247
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003248 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
Johannes Weiner1b051172019-11-30 17:55:52 -08003249
3250 if (cgroup_reclaim(sc)) {
3251 struct lruvec *lruvec;
3252
3253 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3254 zone->zone_pgdat);
3255 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3256 }
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003257 }
3258
Keika Kobayashi873b4772008-07-25 01:48:52 -07003259 delayacct_freepages_end();
3260
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003261 if (sc->nr_reclaimed)
3262 return sc->nr_reclaimed;
3263
Mel Gorman0cee34f2012-01-12 17:19:49 -08003264 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07003265 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08003266 return 1;
3267
Johannes Weinerb91ac372019-11-30 17:56:02 -08003268 /*
3269 * We make inactive:active ratio decisions based on the node's
3270 * composition of memory, but a restrictive reclaim_idx or a
3271 * memory.low cgroup setting can exempt large amounts of
3272 * memory from reclaim. Neither of which are very common, so
3273 * instead of doing costly eligibility calculations of the
3274 * entire cgroup subtree up front, we assume the estimates are
3275 * good, and retry with forcible deactivation if that fails.
3276 */
3277 if (sc->skipped_deactivate) {
3278 sc->priority = initial_priority;
3279 sc->force_deactivate = 1;
3280 sc->skipped_deactivate = 0;
3281 goto retry;
3282 }
3283
Johannes Weiner241994ed2015-02-11 15:26:06 -08003284 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07003285 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08003286 sc->priority = initial_priority;
Johannes Weinerb91ac372019-11-30 17:56:02 -08003287 sc->force_deactivate = 0;
Yisheng Xied6622f62017-05-03 14:53:57 -07003288 sc->memcg_low_reclaim = 1;
3289 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003290 goto retry;
3291 }
3292
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003293 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294}
3295
Johannes Weinerc73322d2017-05-03 14:51:51 -07003296static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07003297{
3298 struct zone *zone;
3299 unsigned long pfmemalloc_reserve = 0;
3300 unsigned long free_pages = 0;
3301 int i;
3302 bool wmark_ok;
3303
Johannes Weinerc73322d2017-05-03 14:51:51 -07003304 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3305 return true;
3306
Mel Gorman55150612012-07-31 16:44:35 -07003307 for (i = 0; i <= ZONE_NORMAL; i++) {
3308 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd82017-05-03 14:51:54 -07003309 if (!managed_zone(zone))
3310 continue;
3311
3312 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07003313 continue;
3314
Mel Gorman55150612012-07-31 16:44:35 -07003315 pfmemalloc_reserve += min_wmark_pages(zone);
3316 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3317 }
3318
Mel Gorman675becc2014-06-04 16:07:35 -07003319 /* If there are no reserves (unexpected config) then do not throttle */
3320 if (!pfmemalloc_reserve)
3321 return true;
3322
Mel Gorman55150612012-07-31 16:44:35 -07003323 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3324
3325 /* kswapd must be awake if processes are being throttled */
3326 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003327 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3328 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003329
Mel Gorman55150612012-07-31 16:44:35 -07003330 wake_up_interruptible(&pgdat->kswapd_wait);
3331 }
3332
3333 return wmark_ok;
3334}
3335
3336/*
3337 * Throttle direct reclaimers if backing storage is backed by the network
3338 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3339 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08003340 * when the low watermark is reached.
3341 *
3342 * Returns true if a fatal signal was delivered during throttling. If this
3343 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07003344 */
Mel Gorman50694c22012-11-26 16:29:48 -08003345static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07003346 nodemask_t *nodemask)
3347{
Mel Gorman675becc2014-06-04 16:07:35 -07003348 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07003349 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07003350 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07003351
3352 /*
3353 * Kernel threads should not be throttled as they may be indirectly
3354 * responsible for cleaning pages necessary for reclaim to make forward
3355 * progress. kjournald for example may enter direct reclaim while
3356 * committing a transaction where throttling it could forcing other
3357 * processes to block on log_wait_commit().
3358 */
3359 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08003360 goto out;
3361
3362 /*
3363 * If a fatal signal is pending, this process should not throttle.
3364 * It should return quickly so it can exit and free its memory
3365 */
3366 if (fatal_signal_pending(current))
3367 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003368
Mel Gorman675becc2014-06-04 16:07:35 -07003369 /*
3370 * Check if the pfmemalloc reserves are ok by finding the first node
3371 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3372 * GFP_KERNEL will be required for allocating network buffers when
3373 * swapping over the network so ZONE_HIGHMEM is unusable.
3374 *
3375 * Throttling is based on the first usable node and throttled processes
3376 * wait on a queue until kswapd makes progress and wakes them. There
3377 * is an affinity then between processes waking up and where reclaim
3378 * progress has been made assuming the process wakes on the same node.
3379 * More importantly, processes running on remote nodes will not compete
3380 * for remote pfmemalloc reserves and processes on different nodes
3381 * should make reasonable progress.
3382 */
3383 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08003384 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07003385 if (zone_idx(zone) > ZONE_NORMAL)
3386 continue;
3387
3388 /* Throttle based on the first usable node */
3389 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07003390 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07003391 goto out;
3392 break;
3393 }
3394
3395 /* If no zone was usable by the allocation flags then do not throttle */
3396 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08003397 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003398
Mel Gorman68243e72012-07-31 16:44:39 -07003399 /* Account for the throttling */
3400 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3401
Mel Gorman55150612012-07-31 16:44:35 -07003402 /*
3403 * If the caller cannot enter the filesystem, it's possible that it
3404 * is due to the caller holding an FS lock or performing a journal
3405 * transaction in the case of a filesystem like ext[3|4]. In this case,
3406 * it is not safe to block on pfmemalloc_wait as kswapd could be
3407 * blocked waiting on the same lock. Instead, throttle for up to a
3408 * second before continuing.
3409 */
3410 if (!(gfp_mask & __GFP_FS)) {
3411 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003412 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003413
3414 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003415 }
3416
3417 /* Throttle until kswapd wakes the process */
3418 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003419 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003420
3421check_pending:
3422 if (fatal_signal_pending(current))
3423 return true;
3424
3425out:
3426 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003427}
3428
Mel Gormandac1d272008-04-28 02:12:12 -07003429unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003430 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003431{
Mel Gorman33906bc2010-08-09 17:19:16 -07003432 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003433 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003434 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003435 .gfp_mask = current_gfp_context(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07003436 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003437 .order = order,
3438 .nodemask = nodemask,
3439 .priority = DEF_PRIORITY,
3440 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003441 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003442 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003443 };
3444
Mel Gorman55150612012-07-31 16:44:35 -07003445 /*
Greg Thelenbb451fd2018-08-17 15:45:19 -07003446 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3447 * Confirm they are large enough for max values.
3448 */
3449 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3450 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3451 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3452
3453 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003454 * Do not enter reclaim if fatal signal was delivered while throttled.
3455 * 1 is returned so that the page allocator does not OOM kill at this
3456 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003457 */
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003458 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003459 return 1;
3460
Andrew Morton1732d2b012019-07-16 16:26:15 -07003461 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07003462 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
Mel Gorman33906bc2010-08-09 17:19:16 -07003463
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003464 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003465
3466 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003467 set_task_reclaim_state(current, NULL);
Mel Gorman33906bc2010-08-09 17:19:16 -07003468
3469 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003470}
3471
Andrew Mortonc255a452012-07-31 16:43:02 -07003472#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003473
Michal Hockod2e5fb92019-08-30 16:04:50 -07003474/* Only used by soft limit reclaim. Do not reuse for anything else. */
Mel Gormana9dd0a82016-07-28 15:46:02 -07003475unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003476 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003477 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003478 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003479{
Johannes Weinerafaf07a2019-11-30 17:55:46 -08003480 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
Balbir Singh4e416952009-09-23 15:56:39 -07003481 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003482 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003483 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003484 .may_writepage = !laptop_mode,
3485 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003486 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003487 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003488 };
Ying Han0ae5e892011-05-26 16:25:25 -07003489
Michal Hockod2e5fb92019-08-30 16:04:50 -07003490 WARN_ON_ONCE(!current->reclaim_state);
3491
Balbir Singh4e416952009-09-23 15:56:39 -07003492 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3493 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003494
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003495 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
Yafang Shao3481c372019-05-13 17:19:14 -07003496 sc.gfp_mask);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003497
Balbir Singh4e416952009-09-23 15:56:39 -07003498 /*
3499 * NOTE: Although we can get the priority field, using it
3500 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003501 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003502 * will pick up pages from other mem cgroup's as well. We hack
3503 * the priority and make it zero.
3504 */
Johannes Weinerafaf07a2019-11-30 17:55:46 -08003505 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003506
3507 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3508
Ying Han0ae5e892011-05-26 16:25:25 -07003509 *nr_scanned = sc.nr_scanned;
Yafang Shao0308f7c2019-07-16 16:26:12 -07003510
Balbir Singh4e416952009-09-23 15:56:39 -07003511 return sc.nr_reclaimed;
3512}
3513
Johannes Weiner72835c82012-01-12 17:18:32 -08003514unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003515 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003516 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003517 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003518{
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003519 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003520 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08003521 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003522 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003523 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07003524 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003525 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003526 .target_mem_cgroup = memcg,
3527 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003528 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003529 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003530 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003531 };
Shakeel Buttfa40d1e2019-11-30 17:50:16 -08003532 /*
3533 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3534 * equal pressure on all the nodes. This is based on the assumption that
3535 * the reclaim does not bail out early.
3536 */
3537 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
Balbir Singh66e17072008-02-07 00:13:56 -08003538
Andrew Morton1732d2b012019-07-16 16:26:15 -07003539 set_task_reclaim_state(current, &sc.reclaim_state);
Yafang Shao3481c372019-05-13 17:19:14 -07003540 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003541 noreclaim_flag = memalloc_noreclaim_save();
Johannes Weinereb414682018-10-26 15:06:27 -07003542
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003543 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weinereb414682018-10-26 15:06:27 -07003544
Vlastimil Babka499118e2017-05-08 15:59:50 -07003545 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003546 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003547 set_task_reclaim_state(current, NULL);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003548
3549 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003550}
3551#endif
3552
Mel Gorman1d82de62016-07-28 15:45:43 -07003553static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003554 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003555{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003556 struct mem_cgroup *memcg;
Johannes Weinerb91ac372019-11-30 17:56:02 -08003557 struct lruvec *lruvec;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003558
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003559 if (!total_swap_pages)
3560 return;
3561
Johannes Weinerb91ac372019-11-30 17:56:02 -08003562 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3563 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3564 return;
3565
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003566 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3567 do {
Johannes Weinerb91ac372019-11-30 17:56:02 -08003568 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3569 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3570 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003571 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3572 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003573}
3574
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003575static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
Mel Gorman1c308442018-12-28 00:35:52 -08003576{
3577 int i;
3578 struct zone *zone;
3579
3580 /*
3581 * Check for watermark boosts top-down as the higher zones
3582 * are more likely to be boosted. Both watermarks and boosts
Randy Dunlap1eba09c2020-08-11 18:33:26 -07003583 * should not be checked at the same time as reclaim would
Mel Gorman1c308442018-12-28 00:35:52 -08003584 * start prematurely when there is no boosting and a lower
3585 * zone is balanced.
3586 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003587 for (i = highest_zoneidx; i >= 0; i--) {
Mel Gorman1c308442018-12-28 00:35:52 -08003588 zone = pgdat->node_zones + i;
3589 if (!managed_zone(zone))
3590 continue;
3591
3592 if (zone->watermark_boost)
3593 return true;
3594 }
3595
3596 return false;
3597}
3598
Mel Gormane716f2e2017-05-03 14:53:45 -07003599/*
3600 * Returns true if there is an eligible zone balanced for the request order
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003601 * and highest_zoneidx
Mel Gormane716f2e2017-05-03 14:53:45 -07003602 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003603static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003604{
Mel Gormane716f2e2017-05-03 14:53:45 -07003605 int i;
3606 unsigned long mark = -1;
3607 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003608
Mel Gorman1c308442018-12-28 00:35:52 -08003609 /*
3610 * Check watermarks bottom-up as lower zones are more likely to
3611 * meet watermarks.
3612 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003613 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003614 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07003615
Mel Gormane716f2e2017-05-03 14:53:45 -07003616 if (!managed_zone(zone))
3617 continue;
3618
3619 mark = high_wmark_pages(zone);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003620 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
Mel Gormane716f2e2017-05-03 14:53:45 -07003621 return true;
3622 }
3623
3624 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003625 * If a node has no populated zone within highest_zoneidx, it does not
Mel Gormane716f2e2017-05-03 14:53:45 -07003626 * need balancing by definition. This can happen if a zone-restricted
3627 * allocation tries to wake a remote kswapd.
3628 */
3629 if (mark == -1)
3630 return true;
3631
3632 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003633}
3634
Mel Gorman631b6e02017-05-03 14:53:41 -07003635/* Clear pgdat state for congested, dirty or under writeback. */
3636static void clear_pgdat_congested(pg_data_t *pgdat)
3637{
Johannes Weiner1b051172019-11-30 17:55:52 -08003638 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3639
3640 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
Mel Gorman631b6e02017-05-03 14:53:41 -07003641 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3642 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3643}
3644
Mel Gorman1741c872011-01-13 15:46:21 -08003645/*
Mel Gorman55150612012-07-31 16:44:35 -07003646 * Prepare kswapd for sleeping. This verifies that there are no processes
3647 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3648 *
3649 * Returns true if kswapd is ready to sleep
3650 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003651static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3652 int highest_zoneidx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003653{
Mel Gorman55150612012-07-31 16:44:35 -07003654 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003655 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07003656 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003657 * race between when kswapd checks the watermarks and a process gets
3658 * throttled. There is also a potential race if processes get
3659 * throttled, kswapd wakes, a large process exits thereby balancing the
3660 * zones, which causes kswapd to exit balance_pgdat() before reaching
3661 * the wake up checks. If kswapd is going to sleep, no process should
3662 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3663 * the wake up is premature, processes will wake kswapd and get
3664 * throttled again. The difference from wake ups in balance_pgdat() is
3665 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003666 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003667 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3668 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003669
Johannes Weinerc73322d2017-05-03 14:51:51 -07003670 /* Hopeless node, leave it to direct reclaim */
3671 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3672 return true;
3673
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003674 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003675 clear_pgdat_congested(pgdat);
3676 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003677 }
3678
Shantanu Goel333b0a42017-05-03 14:53:38 -07003679 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003680}
3681
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003683 * kswapd shrinks a node of pages that are at or below the highest usable
3684 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003685 *
3686 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003687 * reclaim or if the lack of progress was due to pages under writeback.
3688 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003689 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003690static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003691 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003692{
Mel Gorman1d82de62016-07-28 15:45:43 -07003693 struct zone *zone;
3694 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003695
Mel Gorman1d82de62016-07-28 15:45:43 -07003696 /* Reclaim a number of pages proportional to the number of zones */
3697 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003698 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003699 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003700 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003701 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003702
Mel Gorman1d82de62016-07-28 15:45:43 -07003703 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003704 }
3705
Mel Gorman1d82de62016-07-28 15:45:43 -07003706 /*
3707 * Historically care was taken to put equal pressure on all zones but
3708 * now pressure is applied based on node LRU order.
3709 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003710 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003711
3712 /*
3713 * Fragmentation may mean that the system cannot be rebalanced for
3714 * high-order allocations. If twice the allocation size has been
3715 * reclaimed then recheck watermarks only at order-0 to prevent
3716 * excessive reclaim. Assume that a process requested a high-order
3717 * can direct reclaim/compact.
3718 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003719 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003720 sc->order = 0;
3721
Mel Gormanb8e83b92013-07-03 15:01:45 -07003722 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003723}
3724
3725/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003726 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3727 * that are eligible for use by the caller until at least one zone is
3728 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003730 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 *
3732 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003733 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Wei Yang8bb4e7a2019-03-05 15:46:22 -08003734 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
Mel Gorman1d82de62016-07-28 15:45:43 -07003735 * or lower is eligible for reclaim until at least one usable zone is
3736 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003738static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003741 unsigned long nr_soft_reclaimed;
3742 unsigned long nr_soft_scanned;
Johannes Weinereb414682018-10-26 15:06:27 -07003743 unsigned long pflags;
Mel Gorman1c308442018-12-28 00:35:52 -08003744 unsigned long nr_boost_reclaim;
3745 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3746 bool boosted;
Mel Gorman1d82de62016-07-28 15:45:43 -07003747 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003748 struct scan_control sc = {
3749 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003750 .order = order,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003751 .may_unmap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003752 };
Omar Sandoval93781322018-06-07 17:07:02 -07003753
Andrew Morton1732d2b012019-07-16 16:26:15 -07003754 set_task_reclaim_state(current, &sc.reclaim_state);
Johannes Weinereb414682018-10-26 15:06:27 -07003755 psi_memstall_enter(&pflags);
Omar Sandoval93781322018-06-07 17:07:02 -07003756 __fs_reclaim_acquire();
3757
Christoph Lameterf8891e52006-06-30 01:55:45 -07003758 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759
Mel Gorman1c308442018-12-28 00:35:52 -08003760 /*
3761 * Account for the reclaim boost. Note that the zone boost is left in
3762 * place so that parallel allocations that are near the watermark will
3763 * stall or direct reclaim until kswapd is finished.
3764 */
3765 nr_boost_reclaim = 0;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003766 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08003767 zone = pgdat->node_zones + i;
3768 if (!managed_zone(zone))
3769 continue;
3770
3771 nr_boost_reclaim += zone->watermark_boost;
3772 zone_boosts[i] = zone->watermark_boost;
3773 }
3774 boosted = nr_boost_reclaim;
3775
3776restart:
3777 sc.priority = DEF_PRIORITY;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003778 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07003779 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003780 bool raise_priority = true;
Mel Gorman1c308442018-12-28 00:35:52 -08003781 bool balanced;
Omar Sandoval93781322018-06-07 17:07:02 -07003782 bool ret;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003783
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003784 sc.reclaim_idx = highest_zoneidx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785
Mel Gorman86c79f62016-07-28 15:45:59 -07003786 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003787 * If the number of buffer_heads exceeds the maximum allowed
3788 * then consider reclaiming from all zones. This has a dual
3789 * purpose -- on 64-bit systems it is expected that
3790 * buffer_heads are stripped during active rotation. On 32-bit
3791 * systems, highmem pages can pin lowmem memory and shrinking
3792 * buffers can relieve lowmem pressure. Reclaim may still not
3793 * go ahead if all eligible zones for the original allocation
3794 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003795 */
3796 if (buffer_heads_over_limit) {
3797 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3798 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003799 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003800 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801
Mel Gorman970a39a2016-07-28 15:46:35 -07003802 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003803 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003806
Mel Gorman86c79f62016-07-28 15:45:59 -07003807 /*
Mel Gorman1c308442018-12-28 00:35:52 -08003808 * If the pgdat is imbalanced then ignore boosting and preserve
3809 * the watermarks for a later time and restart. Note that the
3810 * zone watermarks will be still reset at the end of balancing
3811 * on the grounds that the normal reclaim should be enough to
3812 * re-evaluate if boosting is required when kswapd next wakes.
Mel Gorman86c79f62016-07-28 15:45:59 -07003813 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003814 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08003815 if (!balanced && nr_boost_reclaim) {
3816 nr_boost_reclaim = 0;
3817 goto restart;
3818 }
3819
3820 /*
3821 * If boosting is not active then only reclaim if there are no
3822 * eligible zones. Note that sc.reclaim_idx is not used as
3823 * buffer_heads_over_limit may have adjusted it.
3824 */
3825 if (!nr_boost_reclaim && balanced)
Mel Gormane716f2e2017-05-03 14:53:45 -07003826 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003827
Mel Gorman1c308442018-12-28 00:35:52 -08003828 /* Limit the priority of boosting to avoid reclaim writeback */
3829 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3830 raise_priority = false;
3831
3832 /*
3833 * Do not writeback or swap pages for boosted reclaim. The
3834 * intent is to relieve pressure not issue sub-optimal IO
3835 * from reclaim context. If no pages are reclaimed, the
3836 * reclaim will be aborted.
3837 */
3838 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3839 sc.may_swap = !nr_boost_reclaim;
Mel Gorman1c308442018-12-28 00:35:52 -08003840
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003842 * Do some background aging of the anon list, to give
3843 * pages a chance to be referenced before reclaiming. All
3844 * pages are rotated regardless of classzone as this is
3845 * about consistent aging.
3846 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003847 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003848
3849 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003850 * If we're getting trouble reclaiming, start doing writepage
3851 * even in laptop mode.
3852 */
Johannes Weiner047d72c2017-05-03 14:51:57 -07003853 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003854 sc.may_writepage = 1;
3855
Mel Gorman1d82de62016-07-28 15:45:43 -07003856 /* Call soft limit reclaim before calling shrink_node. */
3857 sc.nr_scanned = 0;
3858 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003859 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003860 sc.gfp_mask, &nr_soft_scanned);
3861 sc.nr_reclaimed += nr_soft_reclaimed;
3862
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003863 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003864 * There should be no need to raise the scanning priority if
3865 * enough pages are already being scanned that that high
3866 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003868 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003869 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003870
3871 /*
3872 * If the low watermark is met there is no need for processes
3873 * to be throttled on pfmemalloc_wait as they should not be
3874 * able to safely make forward progress. Wake them
3875 */
3876 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc73322d2017-05-03 14:51:51 -07003877 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003878 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003879
Mel Gormanb8e83b92013-07-03 15:01:45 -07003880 /* Check if kswapd should be suspending */
Omar Sandoval93781322018-06-07 17:07:02 -07003881 __fs_reclaim_release();
3882 ret = try_to_freeze();
3883 __fs_reclaim_acquire();
3884 if (ret || kthread_should_stop())
Mel Gormanb8e83b92013-07-03 15:01:45 -07003885 break;
3886
3887 /*
3888 * Raise priority if scanning rate is too low or there was no
3889 * progress in reclaiming pages
3890 */
Johannes Weinerc73322d2017-05-03 14:51:51 -07003891 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
Mel Gorman1c308442018-12-28 00:35:52 -08003892 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3893
3894 /*
3895 * If reclaim made no progress for a boost, stop reclaim as
3896 * IO cannot be queued and it could be an infinite loop in
3897 * extreme circumstances.
3898 */
3899 if (nr_boost_reclaim && !nr_reclaimed)
3900 break;
3901
Johannes Weinerc73322d2017-05-03 14:51:51 -07003902 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003903 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003904 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905
Johannes Weinerc73322d2017-05-03 14:51:51 -07003906 if (!sc.nr_reclaimed)
3907 pgdat->kswapd_failures++;
3908
Mel Gormanb8e83b92013-07-03 15:01:45 -07003909out:
Mel Gorman1c308442018-12-28 00:35:52 -08003910 /* If reclaim was boosted, account for the reclaim done in this pass */
3911 if (boosted) {
3912 unsigned long flags;
3913
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003914 for (i = 0; i <= highest_zoneidx; i++) {
Mel Gorman1c308442018-12-28 00:35:52 -08003915 if (!zone_boosts[i])
3916 continue;
3917
3918 /* Increments are under the zone lock */
3919 zone = pgdat->node_zones + i;
3920 spin_lock_irqsave(&zone->lock, flags);
3921 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3922 spin_unlock_irqrestore(&zone->lock, flags);
3923 }
3924
3925 /*
3926 * As there is now likely space, wakeup kcompact to defragment
3927 * pageblocks.
3928 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003929 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
Mel Gorman1c308442018-12-28 00:35:52 -08003930 }
3931
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003932 snapshot_refaults(NULL, pgdat);
Omar Sandoval93781322018-06-07 17:07:02 -07003933 __fs_reclaim_release();
Johannes Weinereb414682018-10-26 15:06:27 -07003934 psi_memstall_leave(&pflags);
Andrew Morton1732d2b012019-07-16 16:26:15 -07003935 set_task_reclaim_state(current, NULL);
Yafang Shaoe5ca8072019-07-16 16:26:09 -07003936
Mel Gorman0abdee22011-01-13 15:46:22 -08003937 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003938 * Return the order kswapd stopped reclaiming at as
3939 * prepare_kswapd_sleep() takes it into account. If another caller
3940 * entered the allocator slow path while kswapd was awake, order will
3941 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003942 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003943 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944}
3945
Mel Gormane716f2e2017-05-03 14:53:45 -07003946/*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003947 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3948 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3949 * not a valid index then either kswapd runs for first time or kswapd couldn't
3950 * sleep after previous reclaim attempt (node is still unbalanced). In that
3951 * case return the zone index of the previous kswapd reclaim cycle.
Mel Gormane716f2e2017-05-03 14:53:45 -07003952 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003953static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3954 enum zone_type prev_highest_zoneidx)
Mel Gormane716f2e2017-05-03 14:53:45 -07003955{
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003956 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07003957
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003958 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
Mel Gormane716f2e2017-05-03 14:53:45 -07003959}
3960
Mel Gorman38087d92016-07-28 15:45:49 -07003961static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003962 unsigned int highest_zoneidx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003963{
3964 long remaining = 0;
3965 DEFINE_WAIT(wait);
3966
3967 if (freezing(current) || kthread_should_stop())
3968 return;
3969
3970 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3971
Shantanu Goel333b0a42017-05-03 14:53:38 -07003972 /*
3973 * Try to sleep for a short interval. Note that kcompactd will only be
3974 * woken if it is possible to sleep for a short interval. This is
3975 * deliberate on the assumption that if reclaim cannot keep an
3976 * eligible zone balanced that it's also unlikely that compaction will
3977 * succeed.
3978 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003979 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003980 /*
3981 * Compaction records what page blocks it recently failed to
3982 * isolate pages from and skips them in the future scanning.
3983 * When kswapd is going to sleep, it is reasonable to assume
3984 * that pages and compaction may succeed so reset the cache.
3985 */
3986 reset_isolation_suitable(pgdat);
3987
3988 /*
3989 * We have freed the memory, now we should compact it to make
3990 * allocation of the requested order possible.
3991 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003992 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003993
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003994 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003995
3996 /*
Joonsoo Kim97a225e2020-06-03 15:59:01 -07003997 * If woken prematurely then reset kswapd_highest_zoneidx and
Mel Gorman38087d92016-07-28 15:45:49 -07003998 * order. The values will either be from a wakeup request or
3999 * the previous request that slept prematurely.
4000 */
4001 if (remaining) {
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004002 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4003 kswapd_highest_zoneidx(pgdat,
4004 highest_zoneidx));
Qian Cai5644e1fb2020-04-01 21:10:12 -07004005
4006 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4007 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
Mel Gorman38087d92016-07-28 15:45:49 -07004008 }
4009
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08004010 finish_wait(&pgdat->kswapd_wait, &wait);
4011 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4012 }
4013
4014 /*
4015 * After a short sleep, check if it was a premature sleep. If not, then
4016 * go fully to sleep until explicitly woken up.
4017 */
Mel Gormand9f21d42016-07-28 15:46:41 -07004018 if (!remaining &&
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004019 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08004020 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4021
4022 /*
4023 * vmstat counters are not perfectly accurate and the estimated
4024 * value for counters such as NR_FREE_PAGES can deviate from the
4025 * true value by nr_online_cpus * threshold. To avoid the zone
4026 * watermarks being breached while under pressure, we reduce the
4027 * per-cpu vmstat threshold while kswapd is awake and restore
4028 * them before going back to sleep.
4029 */
4030 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07004031
4032 if (!kthread_should_stop())
4033 schedule();
4034
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08004035 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4036 } else {
4037 if (remaining)
4038 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4039 else
4040 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4041 }
4042 finish_wait(&pgdat->kswapd_wait, &wait);
4043}
4044
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045/*
4046 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07004047 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 *
4049 * This basically trickles out pages so that we have _some_
4050 * free memory available even if there is no other activity
4051 * that frees anything up. This is needed for things like routing
4052 * etc, where we otherwise might have all activity going on in
4053 * asynchronous contexts that cannot page things out.
4054 *
4055 * If there are applications that are active memory-allocators
4056 * (most normal use), this basically shouldn't matter.
4057 */
4058static int kswapd(void *p)
4059{
Mel Gormane716f2e2017-05-03 14:53:45 -07004060 unsigned int alloc_order, reclaim_order;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004061 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
Zhiyuan Dai68d68ff2021-05-04 18:40:12 -07004062 pg_data_t *pgdat = (pg_data_t *)p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 struct task_struct *tsk = current;
Rusty Russella70f7302009-03-13 14:49:46 +10304064 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065
Rusty Russell174596a2009-01-01 10:12:29 +10304066 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07004067 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068
4069 /*
4070 * Tell the memory management that we're a "memory allocator",
4071 * and that if we need more memory we should get access to it
4072 * regardless (see "__alloc_pages()"). "kswapd" should
4073 * never get caught in the normal page freeing logic.
4074 *
4075 * (Kswapd normally doesn't need memory anyway, but sometimes
4076 * you need a small amount of memory in order to be able to
4077 * page out something else, and this flag essentially protects
4078 * us from recursively trying to free more memory as we're
4079 * trying to free the first piece of memory in the first place).
4080 */
Christoph Lameter930d9152006-01-08 01:00:47 -08004081 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07004082 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083
Qian Cai5644e1fb2020-04-01 21:10:12 -07004084 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004085 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08004087 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07004088
Qian Cai5644e1fb2020-04-01 21:10:12 -07004089 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004090 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4091 highest_zoneidx);
Mel Gormane716f2e2017-05-03 14:53:45 -07004092
Mel Gorman38087d92016-07-28 15:45:49 -07004093kswapd_try_sleep:
4094 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004095 highest_zoneidx);
Mel Gorman215ddd62011-07-08 15:39:40 -07004096
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004097 /* Read the new order and highest_zoneidx */
Lukas Bulwahn2b47a242020-12-14 19:12:18 -08004098 alloc_order = READ_ONCE(pgdat->kswapd_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004099 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4100 highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07004101 WRITE_ONCE(pgdat->kswapd_order, 0);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004102 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103
David Rientjes8fe23e02009-12-14 17:58:33 -08004104 ret = try_to_freeze();
4105 if (kthread_should_stop())
4106 break;
4107
4108 /*
4109 * We can speed up thawing tasks if we don't call balance_pgdat
4110 * after returning from the refrigerator
4111 */
Mel Gorman38087d92016-07-28 15:45:49 -07004112 if (ret)
4113 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07004114
Mel Gorman38087d92016-07-28 15:45:49 -07004115 /*
4116 * Reclaim begins at the requested order but if a high-order
4117 * reclaim fails then kswapd falls back to reclaiming for
4118 * order-0. If that happens, kswapd will consider sleeping
4119 * for the order it finished reclaiming at (reclaim_order)
4120 * but kcompactd is woken to compact for the original
4121 * request (alloc_order).
4122 */
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004123 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
Mel Gormane5146b12016-07-28 15:46:47 -07004124 alloc_order);
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004125 reclaim_order = balance_pgdat(pgdat, alloc_order,
4126 highest_zoneidx);
Mel Gorman38087d92016-07-28 15:45:49 -07004127 if (reclaim_order < alloc_order)
4128 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08004130
Johannes Weiner71abdc12014-06-06 14:35:35 -07004131 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Johannes Weiner71abdc12014-06-06 14:35:35 -07004132
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133 return 0;
4134}
4135
4136/*
David Rientjes5ecd9d42018-04-05 16:25:16 -07004137 * A zone is low on free memory or too fragmented for high-order memory. If
4138 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4139 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4140 * has failed or is not needed, still wake up kcompactd if only compaction is
4141 * needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 */
David Rientjes5ecd9d42018-04-05 16:25:16 -07004143void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004144 enum zone_type highest_zoneidx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145{
4146 pg_data_t *pgdat;
Qian Cai5644e1fb2020-04-01 21:10:12 -07004147 enum zone_type curr_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148
Mel Gorman6aa303d2016-09-01 16:14:55 -07004149 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 return;
4151
David Rientjes5ecd9d42018-04-05 16:25:16 -07004152 if (!cpuset_zone_allowed(zone, gfp_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 return;
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07004154
Qian Cai5644e1fb2020-04-01 21:10:12 -07004155 pgdat = zone->zone_pgdat;
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004156 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07004157
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004158 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4159 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
Qian Cai5644e1fb2020-04-01 21:10:12 -07004160
4161 if (READ_ONCE(pgdat->kswapd_order) < order)
4162 WRITE_ONCE(pgdat->kswapd_order, order);
4163
Con Kolivas8d0986e2005-09-13 01:25:07 -07004164 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 return;
Mel Gormane1a55632016-07-28 15:46:26 -07004166
David Rientjes5ecd9d42018-04-05 16:25:16 -07004167 /* Hopeless node, leave it to direct reclaim if possible */
4168 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004169 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4170 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
David Rientjes5ecd9d42018-04-05 16:25:16 -07004171 /*
4172 * There may be plenty of free memory available, but it's too
4173 * fragmented for high-order allocations. Wake up kcompactd
4174 * and rely on compaction_suitable() to determine if it's
4175 * needed. If it fails, it will defer subsequent attempts to
4176 * ratelimit its work.
4177 */
4178 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004179 wakeup_kcompactd(pgdat, order, highest_zoneidx);
Johannes Weinerc73322d2017-05-03 14:51:51 -07004180 return;
David Rientjes5ecd9d42018-04-05 16:25:16 -07004181 }
Johannes Weinerc73322d2017-05-03 14:51:51 -07004182
Joonsoo Kim97a225e2020-06-03 15:59:01 -07004183 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
David Rientjes5ecd9d42018-04-05 16:25:16 -07004184 gfp_flags);
Con Kolivas8d0986e2005-09-13 01:25:07 -07004185 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186}
4187
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02004188#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004190 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004191 * freed pages.
4192 *
4193 * Rather than trying to age LRUs the aim is to preserve the overall
4194 * LRU order by reclaiming preferentially
4195 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004197unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004199 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004200 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07004201 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07004202 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004203 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07004204 .may_writepage = 1,
4205 .may_unmap = 1,
4206 .may_swap = 1,
4207 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 };
Ying Hana09ed5e2011-05-24 17:12:26 -07004209 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004210 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004211 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212
Peter Zijlstrad92a8cf2017-03-03 10:13:38 +01004213 fs_reclaim_acquire(sc.gfp_mask);
Omar Sandoval93781322018-06-07 17:07:02 -07004214 noreclaim_flag = memalloc_noreclaim_save();
Andrew Morton1732d2b012019-07-16 16:26:15 -07004215 set_task_reclaim_state(current, &sc.reclaim_state);
Andrew Morton69e05942006-03-22 00:08:19 -08004216
Vladimir Davydov3115cd92014-04-03 14:47:22 -07004217 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004218
Andrew Morton1732d2b012019-07-16 16:26:15 -07004219 set_task_reclaim_state(current, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07004220 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004221 fs_reclaim_release(sc.gfp_mask);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07004222
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08004223 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02004225#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226
Yasunori Goto3218ae12006-06-27 02:53:33 -07004227/*
4228 * This kswapd start function will be called by init and node-hot-add.
4229 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4230 */
4231int kswapd_run(int nid)
4232{
4233 pg_data_t *pgdat = NODE_DATA(nid);
4234 int ret = 0;
4235
4236 if (pgdat->kswapd)
4237 return 0;
4238
4239 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4240 if (IS_ERR(pgdat->kswapd)) {
4241 /* failure at boot is fatal */
Thomas Gleixnerc6202ad2017-05-16 20:42:46 +02004242 BUG_ON(system_state < SYSTEM_RUNNING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07004243 pr_err("Failed to start kswapd on node %d\n", nid);
4244 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07004245 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07004246 }
4247 return ret;
4248}
4249
David Rientjes8fe23e02009-12-14 17:58:33 -08004250/*
Jiang Liud8adde12012-07-11 14:01:52 -07004251 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004252 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08004253 */
4254void kswapd_stop(int nid)
4255{
4256 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4257
Jiang Liud8adde12012-07-11 14:01:52 -07004258 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08004259 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07004260 NODE_DATA(nid)->kswapd = NULL;
4261 }
David Rientjes8fe23e02009-12-14 17:58:33 -08004262}
4263
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264static int __init kswapd_init(void)
4265{
Wei Yang6b700b52020-04-01 21:10:09 -07004266 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08004267
Linus Torvalds1da177e2005-04-16 15:20:36 -07004268 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08004269 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07004270 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004271 return 0;
4272}
4273
4274module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004275
4276#ifdef CONFIG_NUMA
4277/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004278 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08004279 *
Mel Gormana5f5f912016-07-28 15:46:32 -07004280 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08004281 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004282 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004283int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004284
Dave Hansen51998362021-02-24 12:09:15 -08004285/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004286 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08004287 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4288 * a zone.
4289 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004290#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08004291
Christoph Lameter9eeff232006-01-18 17:42:31 -08004292/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004293 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07004294 * occur.
4295 */
4296int sysctl_min_unmapped_ratio = 1;
4297
4298/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07004299 * If the number of slab pages in a zone grows beyond this percentage then
4300 * slab reclaim needs to occur.
4301 */
4302int sysctl_min_slab_ratio = 5;
4303
Mel Gorman11fb9982016-07-28 15:46:20 -07004304static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5de2009-06-16 15:33:20 -07004305{
Mel Gorman11fb9982016-07-28 15:46:20 -07004306 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4307 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4308 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5de2009-06-16 15:33:20 -07004309
4310 /*
4311 * It's possible for there to be more file mapped pages than
4312 * accounted for by the pages on the file LRU lists because
4313 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4314 */
4315 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4316}
4317
4318/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07004319static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5de2009-06-16 15:33:20 -07004320{
Alexandru Moised031a152015-11-05 18:48:08 -08004321 unsigned long nr_pagecache_reclaimable;
4322 unsigned long delta = 0;
Mel Gorman90afa5de2009-06-16 15:33:20 -07004323
4324 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004325 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5de2009-06-16 15:33:20 -07004326 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07004327 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5de2009-06-16 15:33:20 -07004328 * a better estimate
4329 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004330 if (node_reclaim_mode & RECLAIM_UNMAP)
4331 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5de2009-06-16 15:33:20 -07004332 else
Mel Gormana5f5f912016-07-28 15:46:32 -07004333 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5de2009-06-16 15:33:20 -07004334
4335 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07004336 if (!(node_reclaim_mode & RECLAIM_WRITE))
4337 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5de2009-06-16 15:33:20 -07004338
4339 /* Watch for any possible underflows due to delta */
4340 if (unlikely(delta > nr_pagecache_reclaimable))
4341 delta = nr_pagecache_reclaimable;
4342
4343 return nr_pagecache_reclaimable - delta;
4344}
4345
Christoph Lameter0ff38492006-09-25 23:31:52 -07004346/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004347 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004348 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004349static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004350{
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08004351 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08004352 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004353 struct task_struct *p = current;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004354 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08004355 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08004356 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004357 .gfp_mask = current_gfp_context(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07004358 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07004359 .priority = NODE_RECLAIM_PRIORITY,
4360 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4361 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07004362 .may_swap = 1,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004363 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08004364 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08004365
Yafang Shao132bb8c2019-05-13 17:17:53 -07004366 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4367 sc.gfp_mask);
4368
Christoph Lameter9eeff232006-01-18 17:42:31 -08004369 cond_resched();
Omar Sandoval93781322018-06-07 17:07:02 -07004370 fs_reclaim_acquire(sc.gfp_mask);
Christoph Lameterd4f77962006-02-24 13:04:22 -08004371 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004372 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08004373 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004374 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08004375 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07004376 noreclaim_flag = memalloc_noreclaim_save();
4377 p->flags |= PF_SWAPWRITE;
Andrew Morton1732d2b012019-07-16 16:26:15 -07004378 set_task_reclaim_state(p, &sc.reclaim_state);
Christoph Lameterc84db232006-02-01 03:05:29 -08004379
Mel Gormana5f5f912016-07-28 15:46:32 -07004380 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07004381 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07004382 * Free memory by calling shrink node with increasing
Christoph Lameter0ff38492006-09-25 23:31:52 -07004383 * priorities until we have enough memory freed.
4384 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07004385 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07004386 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004387 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07004388 }
Christoph Lameterc84db232006-02-01 03:05:29 -08004389
Andrew Morton1732d2b012019-07-16 16:26:15 -07004390 set_task_reclaim_state(p, NULL);
Vlastimil Babka499118e2017-05-08 15:59:50 -07004391 current->flags &= ~PF_SWAPWRITE;
4392 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004393 fs_reclaim_release(sc.gfp_mask);
Yafang Shao132bb8c2019-05-13 17:17:53 -07004394
4395 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4396
Rik van Riela79311c2009-01-06 14:40:01 -08004397 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004398}
Andrew Morton179e9632006-03-22 00:08:18 -08004399
Mel Gormana5f5f912016-07-28 15:46:32 -07004400int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08004401{
David Rientjesd773ed62007-10-16 23:26:01 -07004402 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004403
4404 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004405 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07004406 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07004407 *
Christoph Lameter96146342006-07-03 00:24:13 -07004408 * A small portion of unmapped file backed pages is needed for
4409 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07004410 * thrown out if the node is overallocated. So we do not reclaim
4411 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07004412 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08004413 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004414 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
Roman Gushchind42f3242020-08-06 23:20:39 -07004415 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4416 pgdat->min_slab_pages)
Mel Gormana5f5f912016-07-28 15:46:32 -07004417 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08004418
4419 /*
David Rientjesd773ed62007-10-16 23:26:01 -07004420 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08004421 */
Mel Gormand0164ad2015-11-06 16:28:21 -08004422 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07004423 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08004424
4425 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004426 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08004427 * have associated processors. This will favor the local processor
4428 * over remote processors and spread off node memory allocations
4429 * as wide as possible.
4430 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004431 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4432 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07004433
Mel Gormana5f5f912016-07-28 15:46:32 -07004434 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4435 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07004436
Mel Gormana5f5f912016-07-28 15:46:32 -07004437 ret = __node_reclaim(pgdat, gfp_mask, order);
4438 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07004439
Mel Gorman24cf725182009-06-16 15:33:23 -07004440 if (!ret)
4441 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4442
David Rientjesd773ed62007-10-16 23:26:01 -07004443 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004444}
Christoph Lameter9eeff232006-01-18 17:42:31 -08004445#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004446
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004447/**
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004448 * check_move_unevictable_pages - check pages for evictability and move to
4449 * appropriate zone lru list
4450 * @pvec: pagevec with lru pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004451 *
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004452 * Checks pages for evictability, if an evictable page is in the unevictable
4453 * lru list, moves it to the appropriate evictable lru list. This function
4454 * should be only used for lru pages.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004455 */
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004456void check_move_unevictable_pages(struct pagevec *pvec)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004457{
Alex Shi6168d0d2020-12-15 12:34:29 -08004458 struct lruvec *lruvec = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08004459 int pgscanned = 0;
4460 int pgrescued = 0;
4461 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004462
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004463 for (i = 0; i < pvec->nr; i++) {
4464 struct page *page = pvec->pages[i];
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004465 int nr_pages;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07004466
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004467 if (PageTransTail(page))
4468 continue;
4469
4470 nr_pages = thp_nr_pages(page);
4471 pgscanned += nr_pages;
4472
Alex Shid25b5bd2020-12-15 12:34:16 -08004473 /* block memcg migration during page moving between lru */
4474 if (!TestClearPageLRU(page))
4475 continue;
4476
Alexander Duyck2a5e4e32020-12-15 12:34:33 -08004477 lruvec = relock_page_lruvec_irq(page, lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08004478 if (page_evictable(page) && PageUnevictable(page)) {
Yu Zhao46ae6b22021-02-24 12:08:25 -08004479 del_page_from_lru_list(page, lruvec);
Hugh Dickins24513262012-01-20 14:34:21 -08004480 ClearPageUnevictable(page);
Yu Zhao3a9c9782021-02-24 12:08:17 -08004481 add_page_to_lru_list(page, lruvec);
Hugh Dickins8d8869c2020-09-18 21:20:12 -07004482 pgrescued += nr_pages;
Hugh Dickins24513262012-01-20 14:34:21 -08004483 }
Alex Shid25b5bd2020-12-15 12:34:16 -08004484 SetPageLRU(page);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004485 }
Hugh Dickins24513262012-01-20 14:34:21 -08004486
Alex Shi6168d0d2020-12-15 12:34:29 -08004487 if (lruvec) {
Hugh Dickins24513262012-01-20 14:34:21 -08004488 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4489 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Alex Shi6168d0d2020-12-15 12:34:29 -08004490 unlock_page_lruvec_irq(lruvec);
Alex Shid25b5bd2020-12-15 12:34:16 -08004491 } else if (pgscanned) {
4492 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Hugh Dickins24513262012-01-20 14:34:21 -08004493 }
Hugh Dickins850465792012-01-20 14:34:19 -08004494}
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004495EXPORT_SYMBOL_GPL(check_move_unevictable_pages);