blob: b37610c0eac626158bae78a2b2c202d3dfbe793d [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070015#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/mm.h>
Ingo Molnar5b3cc152017-02-02 20:43:54 +010018#include <linux/sched/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070026#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070027#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080039#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/notifier.h>
41#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080042#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070043#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080044#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080045#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070046#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070047#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070048#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070049#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070050#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080051#include <linux/dax.h>
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 Weiner1276ad62017-02-24 14:56:11 -080080 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -070081 unsigned int may_writepage:1;
82
83 /* Can mapped pages be reclaimed? */
84 unsigned int may_unmap:1;
85
86 /* Can pages be swapped as part of reclaim? */
87 unsigned int may_swap:1;
88
Yisheng Xied6622f62017-05-03 14:53:57 -070089 /*
90 * Cgroups are not reclaimed below their configured memory.low,
91 * unless we threaten to OOM. If any cgroups are skipped due to
92 * memory.low and nothing was reclaimed, go back for memory.low.
93 */
94 unsigned int memcg_low_reclaim:1;
95 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -080096
Johannes Weineree814fe2014-08-06 16:06:19 -070097 unsigned int hibernation_mode:1;
98
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
101
Greg Thelenbb451fd2018-08-17 15:45:19 -0700102 /* Allocation order */
103 s8 order;
104
105 /* Scan (total_size >> priority) pages at once */
106 s8 priority;
107
108 /* The highest zone to isolate pages for reclaim from */
109 s8 reclaim_idx;
110
111 /* This context's GFP mask */
112 gfp_t gfp_mask;
113
Johannes Weineree814fe2014-08-06 16:06:19 -0700114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned;
116
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed;
Andrey Ryabinind108c772018-04-10 16:27:59 -0700119
120 struct {
121 unsigned int dirty;
122 unsigned int unqueued_dirty;
123 unsigned int congested;
124 unsigned int writeback;
125 unsigned int immediate;
126 unsigned int file_taken;
127 unsigned int taken;
128 } nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129};
130
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131#ifdef ARCH_HAS_PREFETCH
132#define prefetch_prev_lru_page(_page, _base, _field) \
133 do { \
134 if ((_page)->lru.prev != _base) { \
135 struct page *prev; \
136 \
137 prev = lru_to_page(&(_page->lru)); \
138 prefetch(&prev->_field); \
139 } \
140 } while (0)
141#else
142#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
143#endif
144
145#ifdef ARCH_HAS_PREFETCHW
146#define prefetchw_prev_lru_page(_page, _base, _field) \
147 do { \
148 if ((_page)->lru.prev != _base) { \
149 struct page *prev; \
150 \
151 prev = lru_to_page(&(_page->lru)); \
152 prefetchw(&prev->_field); \
153 } \
154 } while (0)
155#else
156#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
157#endif
158
159/*
160 * From 0 .. 100. Higher means more swappy.
161 */
162int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700163/*
164 * The total number of pages which are beyond the high watermark within all
165 * zones.
166 */
167unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169static LIST_HEAD(shrinker_list);
170static DECLARE_RWSEM(shrinker_rwsem);
171
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700172#ifdef CONFIG_MEMCG_KMEM
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700173
174/*
175 * We allow subsystems to populate their shrinker-related
176 * LRU lists before register_shrinker_prepared() is called
177 * for the shrinker, since we don't want to impose
178 * restrictions on their internal registration order.
179 * In this case shrink_slab_memcg() may find corresponding
180 * bit is set in the shrinkers map.
181 *
182 * This value is used by the function to detect registering
183 * shrinkers and to skip do_shrink_slab() calls for them.
184 */
185#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
186
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700187static DEFINE_IDR(shrinker_idr);
188static int shrinker_nr_max;
189
190static int prealloc_memcg_shrinker(struct shrinker *shrinker)
191{
192 int id, ret = -ENOMEM;
193
194 down_write(&shrinker_rwsem);
195 /* This may call shrinker, so it must use down_read_trylock() */
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700196 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700197 if (id < 0)
198 goto unlock;
199
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700200 if (id >= shrinker_nr_max) {
201 if (memcg_expand_shrinker_maps(id)) {
202 idr_remove(&shrinker_idr, id);
203 goto unlock;
204 }
205
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700206 shrinker_nr_max = id + 1;
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700207 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700208 shrinker->id = id;
209 ret = 0;
210unlock:
211 up_write(&shrinker_rwsem);
212 return ret;
213}
214
215static void unregister_memcg_shrinker(struct shrinker *shrinker)
216{
217 int id = shrinker->id;
218
219 BUG_ON(id < 0);
220
221 down_write(&shrinker_rwsem);
222 idr_remove(&shrinker_idr, id);
223 up_write(&shrinker_rwsem);
224}
225#else /* CONFIG_MEMCG_KMEM */
226static int prealloc_memcg_shrinker(struct shrinker *shrinker)
227{
228 return 0;
229}
230
231static void unregister_memcg_shrinker(struct shrinker *shrinker)
232{
233}
234#endif /* CONFIG_MEMCG_KMEM */
235
Andrew Mortonc255a452012-07-31 16:43:02 -0700236#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800237static bool global_reclaim(struct scan_control *sc)
238{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800239 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800240}
Tejun Heo97c93412015-05-22 18:23:36 -0400241
242/**
243 * sane_reclaim - is the usual dirty throttling mechanism operational?
244 * @sc: scan_control in question
245 *
246 * The normal page dirty throttling mechanism in balance_dirty_pages() is
247 * completely broken with the legacy memcg and direct stalling in
248 * shrink_page_list() is used for throttling instead, which lacks all the
249 * niceties such as fairness, adaptive pausing, bandwidth proportional
250 * allocation and configurability.
251 *
252 * This function tests whether the vmscan currently in progress can assume
253 * that the normal dirty throttling mechanism is operational.
254 */
255static bool sane_reclaim(struct scan_control *sc)
256{
257 struct mem_cgroup *memcg = sc->target_mem_cgroup;
258
259 if (!memcg)
260 return true;
261#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800262 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400263 return true;
264#endif
265 return false;
266}
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -0700267
268static void set_memcg_congestion(pg_data_t *pgdat,
269 struct mem_cgroup *memcg,
270 bool congested)
271{
272 struct mem_cgroup_per_node *mn;
273
274 if (!memcg)
275 return;
276
277 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
278 WRITE_ONCE(mn->congested, congested);
279}
280
281static bool memcg_congested(pg_data_t *pgdat,
282 struct mem_cgroup *memcg)
283{
284 struct mem_cgroup_per_node *mn;
285
286 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
287 return READ_ONCE(mn->congested);
288
289}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800290#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800291static bool global_reclaim(struct scan_control *sc)
292{
293 return true;
294}
Tejun Heo97c93412015-05-22 18:23:36 -0400295
296static bool sane_reclaim(struct scan_control *sc)
297{
298 return true;
299}
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -0700300
301static inline void set_memcg_congestion(struct pglist_data *pgdat,
302 struct mem_cgroup *memcg, bool congested)
303{
304}
305
306static inline bool memcg_congested(struct pglist_data *pgdat,
307 struct mem_cgroup *memcg)
308{
309 return false;
310
311}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800312#endif
313
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700314/*
315 * This misses isolated pages which are not accounted for to save counters.
316 * As the data only determines if reclaim or compaction continues, it is
317 * not expected that isolated pages will be a dominating factor.
318 */
319unsigned long zone_reclaimable_pages(struct zone *zone)
320{
321 unsigned long nr;
322
323 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
324 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
325 if (get_nr_swap_pages() > 0)
326 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
327 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
328
329 return nr;
330}
331
Michal Hockofd538802017-02-22 15:45:58 -0800332/**
333 * lruvec_lru_size - Returns the number of pages on the given LRU list.
334 * @lruvec: lru vector
335 * @lru: lru to use
336 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
337 */
338unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800339{
Michal Hockofd538802017-02-22 15:45:58 -0800340 unsigned long lru_size;
341 int zid;
342
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700343 if (!mem_cgroup_disabled())
Michal Hockofd538802017-02-22 15:45:58 -0800344 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
345 else
346 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800347
Michal Hockofd538802017-02-22 15:45:58 -0800348 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
349 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
350 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800351
Michal Hockofd538802017-02-22 15:45:58 -0800352 if (!managed_zone(zone))
353 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800354
Michal Hockofd538802017-02-22 15:45:58 -0800355 if (!mem_cgroup_disabled())
356 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
357 else
358 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
359 NR_ZONE_LRU_BASE + lru);
360 lru_size -= min(size, lru_size);
361 }
362
363 return lru_size;
364
Michal Hockob4536f0c82017-01-10 16:58:04 -0800365}
366
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000368 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 */
Tetsuo Handa8e049442018-04-04 19:53:07 +0900370int prealloc_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000372 size_t size = sizeof(*shrinker->nr_deferred);
373
Glauber Costa1d3d4432013-08-28 10:18:04 +1000374 if (shrinker->flags & SHRINKER_NUMA_AWARE)
375 size *= nr_node_ids;
376
377 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
378 if (!shrinker->nr_deferred)
379 return -ENOMEM;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700380
381 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
382 if (prealloc_memcg_shrinker(shrinker))
383 goto free_deferred;
384 }
385
Tetsuo Handa8e049442018-04-04 19:53:07 +0900386 return 0;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700387
388free_deferred:
389 kfree(shrinker->nr_deferred);
390 shrinker->nr_deferred = NULL;
391 return -ENOMEM;
Tetsuo Handa8e049442018-04-04 19:53:07 +0900392}
Glauber Costa1d3d4432013-08-28 10:18:04 +1000393
Tetsuo Handa8e049442018-04-04 19:53:07 +0900394void free_prealloced_shrinker(struct shrinker *shrinker)
395{
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700396 if (!shrinker->nr_deferred)
397 return;
398
399 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
400 unregister_memcg_shrinker(shrinker);
401
Tetsuo Handa8e049442018-04-04 19:53:07 +0900402 kfree(shrinker->nr_deferred);
403 shrinker->nr_deferred = NULL;
404}
405
406void register_shrinker_prepared(struct shrinker *shrinker)
407{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700408 down_write(&shrinker_rwsem);
409 list_add_tail(&shrinker->list, &shrinker_list);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700410#ifdef CONFIG_MEMCG_KMEM
Kirill Tkhai8df4a442018-08-21 21:51:49 -0700411 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
412 idr_replace(&shrinker_idr, shrinker, shrinker->id);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700413#endif
Rusty Russell8e1f9362007-07-17 04:03:17 -0700414 up_write(&shrinker_rwsem);
Tetsuo Handa8e049442018-04-04 19:53:07 +0900415}
416
417int register_shrinker(struct shrinker *shrinker)
418{
419 int err = prealloc_shrinker(shrinker);
420
421 if (err)
422 return err;
423 register_shrinker_prepared(shrinker);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000424 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700426EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427
428/*
429 * Remove one
430 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700431void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Tetsuo Handabb422a72017-12-18 20:31:41 +0900433 if (!shrinker->nr_deferred)
434 return;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700435 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
436 unregister_memcg_shrinker(shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 down_write(&shrinker_rwsem);
438 list_del(&shrinker->list);
439 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700440 kfree(shrinker->nr_deferred);
Tetsuo Handabb422a72017-12-18 20:31:41 +0900441 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700443EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
445#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000446
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800447static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
Josef Bacik9092c712018-01-31 16:16:26 -0800448 struct shrinker *shrinker, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000450 unsigned long freed = 0;
451 unsigned long long delta;
452 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700453 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000454 long nr;
455 long new_nr;
456 int nid = shrinkctl->nid;
457 long batch_size = shrinker->batch ? shrinker->batch
458 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800459 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000460
Kirill Tkhaiac7fb3a2018-08-17 15:48:30 -0700461 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
462 nid = 0;
463
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700464 freeable = shrinker->count_objects(shrinker, shrinkctl);
Kirill Tkhai9b996462018-08-17 15:48:21 -0700465 if (freeable == 0 || freeable == SHRINK_EMPTY)
466 return freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000467
468 /*
469 * copy the current shrinker scan count into a local variable
470 * and zero it so that other concurrent shrinker invocations
471 * don't also do this scanning work.
472 */
473 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
474
475 total_scan = nr;
Josef Bacik9092c712018-01-31 16:16:26 -0800476 delta = freeable >> priority;
477 delta *= 4;
478 do_div(delta, shrinker->seeks);
Roman Gushchin172b06c32018-09-20 12:22:46 -0700479
Glauber Costa1d3d4432013-08-28 10:18:04 +1000480 total_scan += delta;
481 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800482 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000483 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700484 total_scan = freeable;
Shaohua Li5f33a082016-12-12 16:41:50 -0800485 next_deferred = nr;
486 } else
487 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000488
489 /*
490 * We need to avoid excessive windup on filesystem shrinkers
491 * due to large numbers of GFP_NOFS allocations causing the
492 * shrinkers to return -1 all the time. This results in a large
493 * nr being built up so when a shrink that can do some work
494 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700495 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000496 * memory.
497 *
498 * Hence only allow the shrinker to scan the entire cache when
499 * a large delta change is calculated directly.
500 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700501 if (delta < freeable / 4)
502 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000503
504 /*
505 * Avoid risking looping forever due to too large nr value:
506 * never try to free more than twice the estimate number of
507 * freeable entries.
508 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700509 if (total_scan > freeable * 2)
510 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000511
512 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Josef Bacik9092c712018-01-31 16:16:26 -0800513 freeable, delta, total_scan, priority);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000514
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800515 /*
516 * Normally, we should not scan less than batch_size objects in one
517 * pass to avoid too frequent shrinker calls, but if the slab has less
518 * than batch_size objects in total and we are really tight on memory,
519 * we will try to reclaim all available objects, otherwise we can end
520 * up failing allocations although there are plenty of reclaimable
521 * objects spread over several slabs with usage less than the
522 * batch_size.
523 *
524 * We detect the "tight on memory" situations by looking at the total
525 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700526 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800527 * scanning at high prio and therefore should try to reclaim as much as
528 * possible.
529 */
530 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700531 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000532 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800533 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000534
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800535 shrinkctl->nr_to_scan = nr_to_scan;
Chris Wilsond460acb2017-09-06 16:19:26 -0700536 shrinkctl->nr_scanned = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000537 ret = shrinker->scan_objects(shrinker, shrinkctl);
538 if (ret == SHRINK_STOP)
539 break;
540 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000541
Chris Wilsond460acb2017-09-06 16:19:26 -0700542 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
543 total_scan -= shrinkctl->nr_scanned;
544 scanned += shrinkctl->nr_scanned;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000545
546 cond_resched();
547 }
548
Shaohua Li5f33a082016-12-12 16:41:50 -0800549 if (next_deferred >= scanned)
550 next_deferred -= scanned;
551 else
552 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000553 /*
554 * move the unused scan count back into the shrinker in a
555 * manner that handles concurrent updates. If we exhausted the
556 * scan, there is no need to do an update.
557 */
Shaohua Li5f33a082016-12-12 16:41:50 -0800558 if (next_deferred > 0)
559 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000560 &shrinker->nr_deferred[nid]);
561 else
562 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
563
Dave Hansendf9024a2014-06-04 16:08:07 -0700564 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000565 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566}
567
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700568#ifdef CONFIG_MEMCG_KMEM
569static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
570 struct mem_cgroup *memcg, int priority)
571{
572 struct memcg_shrinker_map *map;
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700573 unsigned long ret, freed = 0;
574 int i;
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700575
576 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
577 return 0;
578
579 if (!down_read_trylock(&shrinker_rwsem))
580 return 0;
581
582 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
583 true);
584 if (unlikely(!map))
585 goto unlock;
586
587 for_each_set_bit(i, map->map, shrinker_nr_max) {
588 struct shrink_control sc = {
589 .gfp_mask = gfp_mask,
590 .nid = nid,
591 .memcg = memcg,
592 };
593 struct shrinker *shrinker;
594
595 shrinker = idr_find(&shrinker_idr, i);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700596 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
597 if (!shrinker)
598 clear_bit(i, map->map);
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700599 continue;
600 }
601
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700602 ret = do_shrink_slab(&sc, shrinker, priority);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700603 if (ret == SHRINK_EMPTY) {
604 clear_bit(i, map->map);
605 /*
606 * After the shrinker reported that it had no objects to
607 * free, but before we cleared the corresponding bit in
608 * the memcg shrinker map, a new object might have been
609 * added. To make sure, we have the bit set in this
610 * case, we invoke the shrinker one more time and reset
611 * the bit if it reports that it is not empty anymore.
612 * The memory barrier here pairs with the barrier in
613 * memcg_set_shrinker_bit():
614 *
615 * list_lru_add() shrink_slab_memcg()
616 * list_add_tail() clear_bit()
617 * <MB> <MB>
618 * set_bit() do_shrink_slab()
619 */
620 smp_mb__after_atomic();
621 ret = do_shrink_slab(&sc, shrinker, priority);
622 if (ret == SHRINK_EMPTY)
623 ret = 0;
624 else
625 memcg_set_shrinker_bit(memcg, nid, i);
626 }
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700627 freed += ret;
628
629 if (rwsem_is_contended(&shrinker_rwsem)) {
630 freed = freed ? : 1;
631 break;
632 }
633 }
634unlock:
635 up_read(&shrinker_rwsem);
636 return freed;
637}
638#else /* CONFIG_MEMCG_KMEM */
639static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
640 struct mem_cgroup *memcg, int priority)
641{
642 return 0;
643}
644#endif /* CONFIG_MEMCG_KMEM */
645
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800646/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800647 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800648 * @gfp_mask: allocation context
649 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800650 * @memcg: memory cgroup whose slab caches to target
Josef Bacik9092c712018-01-31 16:16:26 -0800651 * @priority: the reclaim priority
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800653 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800655 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
656 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 *
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700658 * @memcg specifies the memory cgroup to target. Unaware shrinkers
659 * are called only if it is the root cgroup.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800660 *
Josef Bacik9092c712018-01-31 16:16:26 -0800661 * @priority is sc->priority, we take the number of objects and >> by priority
662 * in order to get the scan target.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800664 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800666static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
667 struct mem_cgroup *memcg,
Josef Bacik9092c712018-01-31 16:16:26 -0800668 int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700670 unsigned long ret, freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 struct shrinker *shrinker;
672
Yang Shibeb0cc782019-08-02 21:48:44 -0700673 /*
674 * The root memcg might be allocated even though memcg is disabled
675 * via "cgroup_disable=memory" boot parameter. This could make
676 * mem_cgroup_is_root() return false, then just run memcg slab
677 * shrink, but skip global shrink. This may result in premature
678 * oom.
679 */
680 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700681 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800682
Tetsuo Handae830c632018-04-05 16:23:35 -0700683 if (!down_read_trylock(&shrinker_rwsem))
Minchan Kimf06590b2011-05-24 17:11:11 -0700684 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
686 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800687 struct shrink_control sc = {
688 .gfp_mask = gfp_mask,
689 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800690 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800691 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800692
Kirill Tkhai9b996462018-08-17 15:48:21 -0700693 ret = do_shrink_slab(&sc, shrinker, priority);
694 if (ret == SHRINK_EMPTY)
695 ret = 0;
696 freed += ret;
Minchan Kime4966122018-01-31 16:16:55 -0800697 /*
698 * Bail out if someone want to register a new shrinker to
699 * prevent the regsitration from being stalled for long periods
700 * by parallel ongoing shrinking.
701 */
702 if (rwsem_is_contended(&shrinker_rwsem)) {
703 freed = freed ? : 1;
704 break;
705 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800707
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700709out:
710 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000711 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712}
713
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800714void drop_slab_node(int nid)
715{
716 unsigned long freed;
717
718 do {
719 struct mem_cgroup *memcg = NULL;
720
721 freed = 0;
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700722 memcg = mem_cgroup_iter(NULL, NULL, NULL);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800723 do {
Josef Bacik9092c712018-01-31 16:16:26 -0800724 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800725 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
726 } while (freed > 10);
727}
728
729void drop_slab(void)
730{
731 int nid;
732
733 for_each_online_node(nid)
734 drop_slab_node(nid);
735}
736
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737static inline int is_page_cache_freeable(struct page *page)
738{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700739 /*
740 * A freeable page cache page is referenced only by the caller
741 * that isolated the page, the page cache radix tree and
742 * optional buffer heads at page->private.
743 */
Huang Yingbd4c82c22017-09-06 16:22:49 -0700744 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
745 HPAGE_PMD_NR : 1;
746 return page_count(page) - page_has_private(page) == 1 + radix_pins;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747}
748
Tejun Heo703c2702015-05-22 17:13:44 -0400749static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750{
Christoph Lameter930d9152006-01-08 01:00:47 -0800751 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400753 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400755 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 return 1;
757 return 0;
758}
759
760/*
761 * We detected a synchronous write error writing a page out. Probably
762 * -ENOSPC. We need to propagate that into the address_space for a subsequent
763 * fsync(), msync() or close().
764 *
765 * The tricky part is that after writepage we cannot touch the mapping: nothing
766 * prevents it from being freed up. But we have a ref on the page and once
767 * that page is locked, the mapping is pinned.
768 *
769 * We're allowed to run sleeping lock_page() here because we know the caller has
770 * __GFP_FS.
771 */
772static void handle_write_error(struct address_space *mapping,
773 struct page *page, int error)
774{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100775 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700776 if (page_mapping(page) == mapping)
777 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 unlock_page(page);
779}
780
Christoph Lameter04e62a22006-06-23 02:03:38 -0700781/* possible outcome of pageout() */
782typedef enum {
783 /* failed to write page out, page is locked */
784 PAGE_KEEP,
785 /* move page to the active list, page is locked */
786 PAGE_ACTIVATE,
787 /* page has been sent to the disk successfully, page is unlocked */
788 PAGE_SUCCESS,
789 /* page is clean and locked */
790 PAGE_CLEAN,
791} pageout_t;
792
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793/*
Andrew Morton1742f192006-03-22 00:08:21 -0800794 * pageout is called by shrink_page_list() for each dirty page.
795 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700797static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700798 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
800 /*
801 * If the page is dirty, only perform writeback if that write
802 * will be non-blocking. To prevent this allocation from being
803 * stalled by pagecache activity. But note that there may be
804 * stalls if we need to run get_block(). We could test
805 * PagePrivate for that.
806 *
Al Viro81742022014-04-03 03:17:43 -0400807 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 * this page's queue, we can perform writeback even if that
809 * will block.
810 *
811 * If the page is swapcache, write it back even if that would
812 * block, for some throttling. This happens by accident, because
813 * swap_backing_dev_info is bust: it doesn't reflect the
814 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 */
816 if (!is_page_cache_freeable(page))
817 return PAGE_KEEP;
818 if (!mapping) {
819 /*
820 * Some data journaling orphaned pages can have
821 * page->mapping == NULL while being dirty with clean buffers.
822 */
David Howells266cf652009-04-03 16:42:36 +0100823 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 if (try_to_free_buffers(page)) {
825 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700826 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 return PAGE_CLEAN;
828 }
829 }
830 return PAGE_KEEP;
831 }
832 if (mapping->a_ops->writepage == NULL)
833 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400834 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 return PAGE_KEEP;
836
837 if (clear_page_dirty_for_io(page)) {
838 int res;
839 struct writeback_control wbc = {
840 .sync_mode = WB_SYNC_NONE,
841 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700842 .range_start = 0,
843 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 .for_reclaim = 1,
845 };
846
847 SetPageReclaim(page);
848 res = mapping->a_ops->writepage(page, &wbc);
849 if (res < 0)
850 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800851 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 ClearPageReclaim(page);
853 return PAGE_ACTIVATE;
854 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700855
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 if (!PageWriteback(page)) {
857 /* synchronous write or broken a_ops? */
858 ClearPageReclaim(page);
859 }
yalin wang3aa23852016-01-14 15:18:30 -0800860 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700861 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 return PAGE_SUCCESS;
863 }
864
865 return PAGE_CLEAN;
866}
867
Andrew Mortona649fd92006-10-17 00:09:36 -0700868/*
Nick Piggine2867812008-07-25 19:45:30 -0700869 * Same as remove_mapping, but if the page is removed from the mapping, it
870 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700871 */
Johannes Weinera5289102014-04-03 14:47:51 -0700872static int __remove_mapping(struct address_space *mapping, struct page *page,
873 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800874{
Greg Thelenc4843a72015-05-22 17:13:16 -0400875 unsigned long flags;
Huang Yingbd4c82c22017-09-06 16:22:49 -0700876 int refcount;
Greg Thelenc4843a72015-05-22 17:13:16 -0400877
Nick Piggin28e4d962006-09-25 23:31:23 -0700878 BUG_ON(!PageLocked(page));
879 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800880
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700881 xa_lock_irqsave(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800882 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700883 * The non racy check for a busy page.
884 *
885 * Must be careful with the order of the tests. When someone has
886 * a ref to the page, it may be possible that they dirty it then
887 * drop the reference. So if PageDirty is tested before page_count
888 * here, then the following race may occur:
889 *
890 * get_user_pages(&page);
891 * [user mapping goes away]
892 * write_to(page);
893 * !PageDirty(page) [good]
894 * SetPageDirty(page);
895 * put_page(page);
896 * !page_count(page) [good, discard it]
897 *
898 * [oops, our write_to data is lost]
899 *
900 * Reversing the order of the tests ensures such a situation cannot
901 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700902 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700903 *
904 * Note that if SetPageDirty is always performed via set_page_dirty,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700905 * and thus under the i_pages lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800906 */
Huang Yingbd4c82c22017-09-06 16:22:49 -0700907 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
908 refcount = 1 + HPAGE_PMD_NR;
909 else
910 refcount = 2;
911 if (!page_ref_freeze(page, refcount))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800912 goto cannot_free;
Jiang Biao1c4c3b92018-08-21 21:53:13 -0700913 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
Nick Piggine2867812008-07-25 19:45:30 -0700914 if (unlikely(PageDirty(page))) {
Huang Yingbd4c82c22017-09-06 16:22:49 -0700915 page_ref_unfreeze(page, refcount);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800916 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700917 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800918
919 if (PageSwapCache(page)) {
920 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700921 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800922 __delete_from_swap_cache(page);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700923 xa_unlock_irqrestore(&mapping->i_pages, flags);
Minchan Kim75f6d6d2017-07-06 15:37:21 -0700924 put_swap_page(page, swap);
Nick Piggine2867812008-07-25 19:45:30 -0700925 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500926 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700927 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500928
929 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700930 /*
931 * Remember a shadow entry for reclaimed file cache in
932 * order to detect refaults, thus thrashing, later on.
933 *
934 * But don't store shadows in an address space that is
935 * already exiting. This is not just an optizimation,
936 * inode reclaim needs to empty out the radix tree or
937 * the nodes are lost. Don't plant shadows behind its
938 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800939 *
940 * We also don't store shadows for DAX mappings because the
941 * only page cache pages found in these are zero pages
942 * covering holes, and because we don't want to mix DAX
943 * exceptional entries and shadow exceptional entries in the
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700944 * same address_space.
Johannes Weinera5289102014-04-03 14:47:51 -0700945 */
946 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800947 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700948 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700949 __delete_from_page_cache(page, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700950 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500951
952 if (freepage != NULL)
953 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800954 }
955
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800956 return 1;
957
958cannot_free:
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700959 xa_unlock_irqrestore(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800960 return 0;
961}
962
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963/*
Nick Piggine2867812008-07-25 19:45:30 -0700964 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
965 * someone else has a ref on the page, abort and return 0. If it was
966 * successfully detached, return 1. Assumes the caller has a single ref on
967 * this page.
968 */
969int remove_mapping(struct address_space *mapping, struct page *page)
970{
Johannes Weinera5289102014-04-03 14:47:51 -0700971 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700972 /*
973 * Unfreezing the refcount with 1 rather than 2 effectively
974 * drops the pagecache ref for us without requiring another
975 * atomic operation.
976 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700977 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700978 return 1;
979 }
980 return 0;
981}
982
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700983/**
984 * putback_lru_page - put previously isolated page onto appropriate LRU list
985 * @page: page to be put back to appropriate lru list
986 *
987 * Add previously isolated @page to appropriate LRU list.
988 * Page may still be unevictable for other reasons.
989 *
990 * lru_lock must not be held, interrupts must be enabled.
991 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700992void putback_lru_page(struct page *page)
993{
Shakeel Butt9c4e6b12018-02-21 14:45:28 -0800994 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700995 put_page(page); /* drop ref from isolate */
996}
997
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800998enum page_references {
999 PAGEREF_RECLAIM,
1000 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -08001001 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001002 PAGEREF_ACTIVATE,
1003};
1004
1005static enum page_references page_check_references(struct page *page,
1006 struct scan_control *sc)
1007{
Johannes Weiner645747462010-03-05 13:42:22 -08001008 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001009 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001010
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001011 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1012 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -08001013 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001014
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001015 /*
1016 * Mlock lost the isolation race with us. Let try_to_unmap()
1017 * move the page to the unevictable list.
1018 */
1019 if (vm_flags & VM_LOCKED)
1020 return PAGEREF_RECLAIM;
1021
Johannes Weiner645747462010-03-05 13:42:22 -08001022 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -07001023 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001024 return PAGEREF_ACTIVATE;
1025 /*
1026 * All mapped pages start out with page table
1027 * references from the instantiating fault, so we need
1028 * to look twice if a mapped file page is used more
1029 * than once.
1030 *
1031 * Mark it and spare it for another trip around the
1032 * inactive list. Another page table reference will
1033 * lead to its activation.
1034 *
1035 * Note: the mark is set for activated pages as well
1036 * so that recently deactivated but used pages are
1037 * quickly recovered.
1038 */
1039 SetPageReferenced(page);
1040
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -08001041 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -08001042 return PAGEREF_ACTIVATE;
1043
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001044 /*
1045 * Activate file-backed executable pages after first usage.
1046 */
1047 if (vm_flags & VM_EXEC)
1048 return PAGEREF_ACTIVATE;
1049
Johannes Weiner645747462010-03-05 13:42:22 -08001050 return PAGEREF_KEEP;
1051 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001052
1053 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -07001054 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001055 return PAGEREF_RECLAIM_CLEAN;
1056
1057 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001058}
1059
Mel Gormane2be15f2013-07-03 15:01:57 -07001060/* Check if a page is dirty or under writeback */
1061static void page_check_dirty_writeback(struct page *page,
1062 bool *dirty, bool *writeback)
1063{
Mel Gormanb4597222013-07-03 15:02:05 -07001064 struct address_space *mapping;
1065
Mel Gormane2be15f2013-07-03 15:01:57 -07001066 /*
1067 * Anonymous pages are not handled by flushers and must be written
1068 * from reclaim context. Do not stall reclaim based on them
1069 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001070 if (!page_is_file_cache(page) ||
1071 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -07001072 *dirty = false;
1073 *writeback = false;
1074 return;
1075 }
1076
1077 /* By default assume that the page flags are accurate */
1078 *dirty = PageDirty(page);
1079 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -07001080
1081 /* Verify dirty/writeback state if the filesystem supports it */
1082 if (!page_has_private(page))
1083 return;
1084
1085 mapping = page_mapping(page);
1086 if (mapping && mapping->a_ops->is_dirty_writeback)
1087 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -07001088}
1089
Nick Piggine2867812008-07-25 19:45:30 -07001090/*
Andrew Morton1742f192006-03-22 00:08:21 -08001091 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 */
Andrew Morton1742f192006-03-22 00:08:21 -08001093static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -07001094 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -07001095 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -07001096 enum ttu_flags ttu_flags,
Michal Hocko3c710c12017-02-22 15:44:27 -08001097 struct reclaim_stat *stat,
Minchan Kim02c6de82012-10-08 16:31:55 -07001098 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099{
1100 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001101 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 int pgactivate = 0;
Michal Hocko3c710c12017-02-22 15:44:27 -08001103 unsigned nr_unqueued_dirty = 0;
1104 unsigned nr_dirty = 0;
1105 unsigned nr_congested = 0;
1106 unsigned nr_reclaimed = 0;
1107 unsigned nr_writeback = 0;
1108 unsigned nr_immediate = 0;
Michal Hocko5bccd162017-02-22 15:44:30 -08001109 unsigned nr_ref_keep = 0;
1110 unsigned nr_unmap_fail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
1112 cond_resched();
1113
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 while (!list_empty(page_list)) {
1115 struct address_space *mapping;
1116 struct page *page;
1117 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -07001118 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -07001119 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
1121 cond_resched();
1122
1123 page = lru_to_page(page_list);
1124 list_del(&page->lru);
1125
Nick Piggin529ae9a2008-08-02 12:01:03 +02001126 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 goto keep;
1128
Sasha Levin309381fea2014-01-23 15:52:54 -08001129 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
1131 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -08001132
Hugh Dickins39b5f292012-10-08 16:33:18 -07001133 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -07001134 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001135
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001136 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -08001137 goto keep_locked;
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 /* Double the slab pressure for mapped and swapcache pages */
Shaohua Li802a3a92017-05-03 14:52:32 -07001140 if ((page_mapped(page) || PageSwapCache(page)) &&
1141 !(PageAnon(page) && !PageSwapBacked(page)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 sc->nr_scanned++;
1143
Andy Whitcroftc661b072007-08-22 14:01:26 -07001144 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1145 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1146
Mel Gorman283aba92013-07-03 15:01:51 -07001147 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07001148 * The number of dirty pages determines if a node is marked
Mel Gormane2be15f2013-07-03 15:01:57 -07001149 * reclaim_congested which affects wait_iff_congested. kswapd
1150 * will stall and start writing pages if the tail of the LRU
1151 * is all dirty unqueued pages.
1152 */
1153 page_check_dirty_writeback(page, &dirty, &writeback);
1154 if (dirty || writeback)
1155 nr_dirty++;
1156
1157 if (dirty && !writeback)
1158 nr_unqueued_dirty++;
1159
Mel Gormand04e8ac2013-07-03 15:02:03 -07001160 /*
1161 * Treat this page as congested if the underlying BDI is or if
1162 * pages are cycling through the LRU so quickly that the
1163 * pages marked for immediate reclaim are making it to the
1164 * end of the LRU a second time.
1165 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001166 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001167 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001168 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001169 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -07001170 nr_congested++;
1171
1172 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001173 * If a page at the tail of the LRU is under writeback, there
1174 * are three cases to consider.
1175 *
1176 * 1) If reclaim is encountering an excessive number of pages
1177 * under writeback and this page is both under writeback and
1178 * PageReclaim then it indicates that pages are being queued
1179 * for IO but are being recycled through the LRU before the
1180 * IO can complete. Waiting on the page itself risks an
1181 * indefinite stall if it is impossible to writeback the
1182 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001183 * note that the LRU is being scanned too quickly and the
1184 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001185 *
Tejun Heo97c93412015-05-22 18:23:36 -04001186 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001187 * not marked for immediate reclaim, or the caller does not
1188 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1189 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001190 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001191 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001192 * Require may_enter_fs because we would wait on fs, which
1193 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001194 * enter reclaim, and deadlock if it waits on a page for
1195 * which it is needed to do the write (loop masks off
1196 * __GFP_IO|__GFP_FS for this reason); but more thought
1197 * would probably show more reasons.
1198 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001199 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001200 * PageReclaim. memcg does not have any dirty pages
1201 * throttling so we could easily OOM just because too many
1202 * pages are in writeback and there is nothing else to
1203 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001204 *
1205 * In cases 1) and 2) we activate the pages to get them out of
1206 * the way while we continue scanning for clean pages on the
1207 * inactive list and refilling from the active list. The
1208 * observation here is that waiting for disk writes is more
1209 * expensive than potentially causing reloads down the line.
1210 * Since they're marked for immediate reclaim, they won't put
1211 * memory pressure on the cache working set any longer than it
1212 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001213 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001214 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001215 /* Case 1 above */
1216 if (current_is_kswapd() &&
1217 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001218 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001219 nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001220 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001221
1222 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001223 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001224 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001225 /*
1226 * This is slightly racy - end_page_writeback()
1227 * might have just cleared PageReclaim, then
1228 * setting PageReclaim here end up interpreted
1229 * as PageReadahead - but that does not matter
1230 * enough to care. What we do want is for this
1231 * page to have PageReclaim set next time memcg
1232 * reclaim reaches the tests above, so it will
1233 * then wait_on_page_writeback() to avoid OOM;
1234 * and it's also appropriate in global reclaim.
1235 */
1236 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001237 nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001238 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001239
1240 /* Case 3 above */
1241 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001242 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001243 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001244 /* then go back and try same page again */
1245 list_add_tail(&page->lru, page_list);
1246 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001247 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001248 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249
Minchan Kim02c6de82012-10-08 16:31:55 -07001250 if (!force_reclaim)
1251 references = page_check_references(page, sc);
1252
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001253 switch (references) {
1254 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001256 case PAGEREF_KEEP:
Michal Hocko5bccd162017-02-22 15:44:30 -08001257 nr_ref_keep++;
Johannes Weiner645747462010-03-05 13:42:22 -08001258 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001259 case PAGEREF_RECLAIM:
1260 case PAGEREF_RECLAIM_CLEAN:
1261 ; /* try to reclaim the page below */
1262 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 /*
1265 * Anonymous process memory has backing store?
1266 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001267 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001269 if (PageAnon(page) && PageSwapBacked(page)) {
1270 if (!PageSwapCache(page)) {
1271 if (!(sc->gfp_mask & __GFP_IO))
1272 goto keep_locked;
1273 if (PageTransHuge(page)) {
1274 /* cannot split THP, skip it */
1275 if (!can_split_huge_page(page, NULL))
1276 goto activate_locked;
1277 /*
1278 * Split pages without a PMD map right
1279 * away. Chances are some or all of the
1280 * tail pages can be freed without IO.
1281 */
1282 if (!compound_mapcount(page) &&
1283 split_huge_page_to_list(page,
1284 page_list))
1285 goto activate_locked;
1286 }
1287 if (!add_to_swap(page)) {
1288 if (!PageTransHuge(page))
1289 goto activate_locked;
1290 /* Fallback to swap normal pages */
1291 if (split_huge_page_to_list(page,
1292 page_list))
1293 goto activate_locked;
Huang Yingfe490cc2017-09-06 16:22:52 -07001294#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1295 count_vm_event(THP_SWPOUT_FALLBACK);
1296#endif
Huang Yingbd4c82c22017-09-06 16:22:49 -07001297 if (!add_to_swap(page))
1298 goto activate_locked;
1299 }
Minchan Kim0f074652017-07-06 15:37:24 -07001300
Huang Yingbd4c82c22017-09-06 16:22:49 -07001301 may_enter_fs = 1;
1302
1303 /* Adding to swap updated mapping */
1304 mapping = page_mapping(page);
Minchan Kim0f074652017-07-06 15:37:24 -07001305 }
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001306 } else if (unlikely(PageTransHuge(page))) {
1307 /* Split file THP */
1308 if (split_huge_page_to_list(page, page_list))
1309 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001310 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
1312 /*
1313 * The page is mapped into the page tables of one or more
1314 * processes. Try to unmap it here.
1315 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001316 if (page_mapped(page)) {
Huang Yingbd4c82c22017-09-06 16:22:49 -07001317 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1318
1319 if (unlikely(PageTransHuge(page)))
1320 flags |= TTU_SPLIT_HUGE_PMD;
1321 if (!try_to_unmap(page, flags)) {
Michal Hocko5bccd162017-02-22 15:44:30 -08001322 nr_unmap_fail++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 }
1325 }
1326
1327 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001328 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001329 * Only kswapd can writeback filesystem pages
1330 * to avoid risk of stack overflow. But avoid
1331 * injecting inefficient single-page IO into
1332 * flusher writeback as much as possible: only
1333 * write pages when we've encountered many
1334 * dirty pages, and when we've already scanned
1335 * the rest of the LRU for clean pages and see
1336 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001337 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001338 if (page_is_file_cache(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001339 (!current_is_kswapd() || !PageReclaim(page) ||
1340 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001341 /*
1342 * Immediately reclaim when written back.
1343 * Similar in principal to deactivate_page()
1344 * except we already have the page isolated
1345 * and know it's dirty
1346 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001347 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001348 SetPageReclaim(page);
1349
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001350 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001351 }
1352
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001353 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001355 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001357 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 goto keep_locked;
1359
Mel Gormand950c942015-09-04 15:47:35 -07001360 /*
1361 * Page is dirty. Flush the TLB if a writable entry
1362 * potentially exists to avoid CPU writes after IO
1363 * starts and then write it out here.
1364 */
1365 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001366 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 case PAGE_KEEP:
1368 goto keep_locked;
1369 case PAGE_ACTIVATE:
1370 goto activate_locked;
1371 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001372 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001373 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001374 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001376
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 /*
1378 * A synchronous write - probably a ramdisk. Go
1379 * ahead and try to reclaim the page.
1380 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001381 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 goto keep;
1383 if (PageDirty(page) || PageWriteback(page))
1384 goto keep_locked;
1385 mapping = page_mapping(page);
1386 case PAGE_CLEAN:
1387 ; /* try to free the page below */
1388 }
1389 }
1390
1391 /*
1392 * If the page has buffers, try to free the buffer mappings
1393 * associated with this page. If we succeed we try to free
1394 * the page as well.
1395 *
1396 * We do this even if the page is PageDirty().
1397 * try_to_release_page() does not perform I/O, but it is
1398 * possible for a page to have PageDirty set, but it is actually
1399 * clean (all its buffers are clean). This happens if the
1400 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001401 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 * try_to_release_page() will discover that cleanness and will
1403 * drop the buffers and mark the page clean - it can be freed.
1404 *
1405 * Rarely, pages can have buffers and no ->mapping. These are
1406 * the pages which were not successfully invalidated in
1407 * truncate_complete_page(). We try to drop those buffers here
1408 * and if that worked, and the page is no longer mapped into
1409 * process address space (page_count == 1) it can be freed.
1410 * Otherwise, leave the page on the LRU so it is swappable.
1411 */
David Howells266cf652009-04-03 16:42:36 +01001412 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 if (!try_to_release_page(page, sc->gfp_mask))
1414 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001415 if (!mapping && page_count(page) == 1) {
1416 unlock_page(page);
1417 if (put_page_testzero(page))
1418 goto free_it;
1419 else {
1420 /*
1421 * rare race with speculative reference.
1422 * the speculative reference will free
1423 * this page shortly, so we may
1424 * increment nr_reclaimed here (and
1425 * leave it off the LRU).
1426 */
1427 nr_reclaimed++;
1428 continue;
1429 }
1430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 }
1432
Shaohua Li802a3a92017-05-03 14:52:32 -07001433 if (PageAnon(page) && !PageSwapBacked(page)) {
1434 /* follow __remove_mapping for reference */
1435 if (!page_ref_freeze(page, 1))
1436 goto keep_locked;
1437 if (PageDirty(page)) {
1438 page_ref_unfreeze(page, 1);
1439 goto keep_locked;
1440 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441
Shaohua Li802a3a92017-05-03 14:52:32 -07001442 count_vm_event(PGLAZYFREED);
Roman Gushchin22621852017-07-06 15:40:25 -07001443 count_memcg_page_event(page, PGLAZYFREED);
Shaohua Li802a3a92017-05-03 14:52:32 -07001444 } else if (!mapping || !__remove_mapping(mapping, page, true))
1445 goto keep_locked;
Nick Piggina978d6f2008-10-18 20:26:58 -07001446 /*
1447 * At this point, we have no other references and there is
1448 * no way to pick any more up (removed from LRU, removed
1449 * from pagecache). Can use non-atomic bitops now (and
1450 * we obviously don't have to worry about waking up a process
1451 * waiting on the page lock, because there are no references.
1452 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001453 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001454free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001455 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001456
1457 /*
1458 * Is there need to periodically free_page_list? It would
1459 * appear not as the counts should be low
1460 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001461 if (unlikely(PageTransHuge(page))) {
1462 mem_cgroup_uncharge(page);
1463 (*get_compound_page_dtor(page))(page);
1464 } else
1465 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 continue;
1467
1468activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001469 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001470 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1471 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001472 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001473 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001474 if (!PageMlocked(page)) {
1475 SetPageActive(page);
1476 pgactivate++;
Roman Gushchin22621852017-07-06 15:40:25 -07001477 count_memcg_page_event(page, PGACTIVATE);
Minchan Kimad6b6702017-05-03 14:54:13 -07001478 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479keep_locked:
1480 unlock_page(page);
1481keep:
1482 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001483 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001485
Johannes Weiner747db952014-08-08 14:19:24 -07001486 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001487 try_to_unmap_flush();
Mel Gorman2d4894b2017-11-15 17:37:59 -08001488 free_unref_page_list(&free_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001489
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001491 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001492
Michal Hocko3c710c12017-02-22 15:44:27 -08001493 if (stat) {
1494 stat->nr_dirty = nr_dirty;
1495 stat->nr_congested = nr_congested;
1496 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1497 stat->nr_writeback = nr_writeback;
1498 stat->nr_immediate = nr_immediate;
Michal Hocko5bccd162017-02-22 15:44:30 -08001499 stat->nr_activate = pgactivate;
1500 stat->nr_ref_keep = nr_ref_keep;
1501 stat->nr_unmap_fail = nr_unmap_fail;
Michal Hocko3c710c12017-02-22 15:44:27 -08001502 }
Andrew Morton05ff5132006-03-22 00:08:20 -08001503 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504}
1505
Minchan Kim02c6de82012-10-08 16:31:55 -07001506unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1507 struct list_head *page_list)
1508{
1509 struct scan_control sc = {
1510 .gfp_mask = GFP_KERNEL,
1511 .priority = DEF_PRIORITY,
1512 .may_unmap = 1,
1513 };
Michal Hocko3c710c12017-02-22 15:44:27 -08001514 unsigned long ret;
Minchan Kim02c6de82012-10-08 16:31:55 -07001515 struct page *page, *next;
1516 LIST_HEAD(clean_pages);
1517
1518 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001519 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kim54a20282019-06-13 15:56:15 -07001520 !__PageMovable(page) && !PageUnevictable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001521 ClearPageActive(page);
1522 list_move(&page->lru, &clean_pages);
1523 }
1524 }
1525
Mel Gorman599d0c92016-07-28 15:45:31 -07001526 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Shaohua Lia128ca72017-05-03 14:52:22 -07001527 TTU_IGNORE_ACCESS, NULL, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001528 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001529 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001530 return ret;
1531}
1532
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001533/*
1534 * Attempt to remove the specified page from its LRU. Only take this page
1535 * if it is of the appropriate PageActive status. Pages which are being
1536 * freed elsewhere are also ignored.
1537 *
1538 * page: page to consider
1539 * mode: one of the LRU isolation modes defined above
1540 *
1541 * returns 0 on success, -ve errno on failure.
1542 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001543int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001544{
1545 int ret = -EINVAL;
1546
1547 /* Only take pages on the LRU. */
1548 if (!PageLRU(page))
1549 return ret;
1550
Minchan Kime46a2872012-10-08 16:33:48 -07001551 /* Compaction should not handle unevictable pages but CMA can do so */
1552 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001553 return ret;
1554
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001555 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001556
Mel Gormanc8244932012-01-12 17:19:38 -08001557 /*
1558 * To minimise LRU disruption, the caller can indicate that it only
1559 * wants to isolate pages it will be able to operate on without
1560 * blocking - clean pages for the most part.
1561 *
Mel Gormanc8244932012-01-12 17:19:38 -08001562 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1563 * that it is possible to migrate without blocking
1564 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001565 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001566 /* All the caller can do on PageWriteback is block */
1567 if (PageWriteback(page))
1568 return ret;
1569
1570 if (PageDirty(page)) {
1571 struct address_space *mapping;
Mel Gorman69d763f2018-01-31 16:19:52 -08001572 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001573
Mel Gormanc8244932012-01-12 17:19:38 -08001574 /*
1575 * Only pages without mappings or that have a
1576 * ->migratepage callback are possible to migrate
Mel Gorman69d763f2018-01-31 16:19:52 -08001577 * without blocking. However, we can be racing with
1578 * truncation so it's necessary to lock the page
1579 * to stabilise the mapping as truncation holds
1580 * the page lock until after the page is removed
1581 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001582 */
Mel Gorman69d763f2018-01-31 16:19:52 -08001583 if (!trylock_page(page))
1584 return ret;
1585
Mel Gormanc8244932012-01-12 17:19:38 -08001586 mapping = page_mapping(page);
Hugh Dickins145e1a72018-06-01 16:50:50 -07001587 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gorman69d763f2018-01-31 16:19:52 -08001588 unlock_page(page);
1589 if (!migrate_dirty)
Mel Gormanc8244932012-01-12 17:19:38 -08001590 return ret;
1591 }
1592 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001593
Minchan Kimf80c0672011-10-31 17:06:55 -07001594 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1595 return ret;
1596
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001597 if (likely(get_page_unless_zero(page))) {
1598 /*
1599 * Be careful not to clear PageLRU until after we're
1600 * sure the page is not being freed elsewhere -- the
1601 * page release code relies on it.
1602 */
1603 ClearPageLRU(page);
1604 ret = 0;
1605 }
1606
1607 return ret;
1608}
1609
Mel Gorman7ee36a12016-07-28 15:47:17 -07001610
1611/*
1612 * Update LRU sizes after isolating pages. The LRU size updates must
1613 * be complete before mem_cgroup_update_lru_size due to a santity check.
1614 */
1615static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001616 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001617{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001618 int zid;
1619
Mel Gorman7ee36a12016-07-28 15:47:17 -07001620 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1621 if (!nr_zone_taken[zid])
1622 continue;
1623
1624 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001625#ifdef CONFIG_MEMCG
1626 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1627#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001628 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001629
Mel Gorman7ee36a12016-07-28 15:47:17 -07001630}
1631
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001632/*
Mel Gormana52633d2016-07-28 15:45:28 -07001633 * zone_lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 * shrink the lists perform better by taking out a batch of pages
1635 * and working on them outside the LRU lock.
1636 *
1637 * For pagecache intensive workloads, this function is the hottest
1638 * spot in the kernel (apart from copy_*_user functions).
1639 *
1640 * Appropriate locks must be held before calling this function.
1641 *
Minchan Kim791b48b2017-05-12 15:47:06 -07001642 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001643 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001645 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001646 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001647 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001648 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 *
1650 * returns how many pages were moved onto *@dst.
1651 */
Andrew Morton69e05942006-03-22 00:08:19 -08001652static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001653 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001654 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001655 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001657 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001658 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001659 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001660 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07001661 unsigned long skipped = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001662 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001663 LIST_HEAD(pages_skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664
Minchan Kim791b48b2017-05-12 15:47:06 -07001665 scan = 0;
1666 for (total_scan = 0;
1667 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1668 total_scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001669 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001670
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 page = lru_to_page(src);
1672 prefetchw_prev_lru_page(page, src, flags);
1673
Sasha Levin309381fea2014-01-23 15:52:54 -08001674 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001675
Mel Gormanb2e18752016-07-28 15:45:37 -07001676 if (page_zonenum(page) > sc->reclaim_idx) {
1677 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001678 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001679 continue;
1680 }
1681
Minchan Kim791b48b2017-05-12 15:47:06 -07001682 /*
1683 * Do not count skipped pages because that makes the function
1684 * return with no isolated pages if the LRU mostly contains
1685 * ineligible pages. This causes the VM to not reclaim any
1686 * pages, triggering a premature OOM.
1687 */
1688 scan++;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001689 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001690 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001691 nr_pages = hpage_nr_pages(page);
1692 nr_taken += nr_pages;
1693 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001694 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001695 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001696
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001697 case -EBUSY:
1698 /* else it is being freed elsewhere */
1699 list_move(&page->lru, src);
1700 continue;
1701
1702 default:
1703 BUG();
1704 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 }
1706
Mel Gormanb2e18752016-07-28 15:45:37 -07001707 /*
1708 * Splice any skipped pages to the start of the LRU list. Note that
1709 * this disrupts the LRU order when reclaiming for lower zones but
1710 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1711 * scanning would soon rescan the same pages to skip and put the
1712 * system at risk of premature OOM.
1713 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001714 if (!list_empty(&pages_skipped)) {
1715 int zid;
1716
Johannes Weiner3db65812017-05-03 14:52:13 -07001717 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001718 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1719 if (!nr_skipped[zid])
1720 continue;
1721
1722 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08001723 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001724 }
1725 }
Minchan Kim791b48b2017-05-12 15:47:06 -07001726 *nr_scanned = total_scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08001727 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim791b48b2017-05-12 15:47:06 -07001728 total_scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001729 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 return nr_taken;
1731}
1732
Nick Piggin62695a82008-10-18 20:26:09 -07001733/**
1734 * isolate_lru_page - tries to isolate a page from its LRU list
1735 * @page: page to isolate from its LRU list
1736 *
1737 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1738 * vmstat statistic corresponding to whatever LRU list the page was on.
1739 *
1740 * Returns 0 if the page was removed from an LRU list.
1741 * Returns -EBUSY if the page was not on an LRU list.
1742 *
1743 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001744 * the active list, it will have PageActive set. If it was found on
1745 * the unevictable list, it will have the PageUnevictable bit set. That flag
1746 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001747 *
1748 * The vmstat statistic corresponding to the list on which the page was
1749 * found will be decremented.
1750 *
1751 * Restrictions:
Mike Rapoporta5d09be2018-02-06 15:42:19 -08001752 *
Nick Piggin62695a82008-10-18 20:26:09 -07001753 * (1) Must be called with an elevated refcount on the page. This is a
1754 * fundamentnal difference from isolate_lru_pages (which is called
1755 * without a stable reference).
1756 * (2) the lru_lock must not be held.
1757 * (3) interrupts must be enabled.
1758 */
1759int isolate_lru_page(struct page *page)
1760{
1761 int ret = -EBUSY;
1762
Sasha Levin309381fea2014-01-23 15:52:54 -08001763 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001764 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001765
Nick Piggin62695a82008-10-18 20:26:09 -07001766 if (PageLRU(page)) {
1767 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001768 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001769
Mel Gormana52633d2016-07-28 15:45:28 -07001770 spin_lock_irq(zone_lru_lock(zone));
Mel Gorman599d0c92016-07-28 15:45:31 -07001771 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001772 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001773 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001774 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001775 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001776 del_page_from_lru_list(page, lruvec, lru);
1777 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001778 }
Mel Gormana52633d2016-07-28 15:45:28 -07001779 spin_unlock_irq(zone_lru_lock(zone));
Nick Piggin62695a82008-10-18 20:26:09 -07001780 }
1781 return ret;
1782}
1783
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001784/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001785 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1786 * then get resheduled. When there are massive number of tasks doing page
1787 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1788 * the LRU list will go small and be scanned faster than necessary, leading to
1789 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001790 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001791static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001792 struct scan_control *sc)
1793{
1794 unsigned long inactive, isolated;
1795
1796 if (current_is_kswapd())
1797 return 0;
1798
Tejun Heo97c93412015-05-22 18:23:36 -04001799 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001800 return 0;
1801
1802 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001803 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1804 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001805 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001806 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1807 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001808 }
1809
Fengguang Wu3cf23842012-12-18 14:23:31 -08001810 /*
1811 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1812 * won't get blocked by normal direct-reclaimers, forming a circular
1813 * deadlock.
1814 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001815 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001816 inactive >>= 3;
1817
Rik van Riel35cd7812009-09-21 17:01:38 -07001818 return isolated > inactive;
1819}
1820
Mel Gorman66635622010-08-09 17:19:30 -07001821static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001822putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001823{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001824 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Mel Gorman599d0c92016-07-28 15:45:31 -07001825 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001826 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001827
Mel Gorman66635622010-08-09 17:19:30 -07001828 /*
1829 * Put back any unfreeable pages.
1830 */
Mel Gorman66635622010-08-09 17:19:30 -07001831 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001832 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001833 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001834
Sasha Levin309381fea2014-01-23 15:52:54 -08001835 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001836 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001837 if (unlikely(!page_evictable(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001838 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001839 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001840 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001841 continue;
1842 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001843
Mel Gorman599d0c92016-07-28 15:45:31 -07001844 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001845
Linus Torvalds7a608572011-01-17 14:42:19 -08001846 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001847 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001848 add_page_to_lru_list(page, lruvec, lru);
1849
Mel Gorman66635622010-08-09 17:19:30 -07001850 if (is_active_lru(lru)) {
1851 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001852 int numpages = hpage_nr_pages(page);
1853 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001854 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001855 if (put_page_testzero(page)) {
1856 __ClearPageLRU(page);
1857 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001858 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001859
1860 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001861 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001862 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001863 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001864 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001865 } else
1866 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001867 }
1868 }
Mel Gorman66635622010-08-09 17:19:30 -07001869
Hugh Dickins3f797682012-01-12 17:20:07 -08001870 /*
1871 * To save our caller's stack, now use input list for pages to free.
1872 */
1873 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001874}
1875
1876/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001877 * If a kernel thread (such as nfsd for loop-back mounts) services
1878 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1879 * In that case we should only throttle if the backing device it is
1880 * writing to is congested. In other cases it is safe to throttle.
1881 */
1882static int current_may_throttle(void)
1883{
1884 return !(current->flags & PF_LESS_THROTTLE) ||
1885 current->backing_dev_info == NULL ||
1886 bdi_write_congested(current->backing_dev_info);
1887}
1888
1889/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001890 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001891 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 */
Mel Gorman66635622010-08-09 17:19:30 -07001893static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001894shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001895 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896{
1897 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001898 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001899 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001900 unsigned long nr_taken;
Michal Hocko3c710c12017-02-22 15:44:27 -08001901 struct reclaim_stat stat = {};
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001902 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001903 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001904 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001905 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Michal Hockodb73ee02017-09-06 16:21:11 -07001906 bool stalled = false;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001907
Mel Gorman599d0c92016-07-28 15:45:31 -07001908 while (unlikely(too_many_isolated(pgdat, file, sc))) {
Michal Hockodb73ee02017-09-06 16:21:11 -07001909 if (stalled)
1910 return 0;
1911
1912 /* wait a bit for the reclaimer. */
1913 msleep(100);
1914 stalled = true;
Rik van Riel35cd7812009-09-21 17:01:38 -07001915
1916 /* We are about to die and free our memory. Return now. */
1917 if (fatal_signal_pending(current))
1918 return SWAP_CLUSTER_MAX;
1919 }
1920
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001922
1923 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001924 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001925
Mel Gorman599d0c92016-07-28 15:45:31 -07001926 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001928 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1929 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001930
Mel Gorman599d0c92016-07-28 15:45:31 -07001931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001932 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001933
Roman Gushchin22621852017-07-06 15:40:25 -07001934 if (current_is_kswapd()) {
1935 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001936 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
Roman Gushchin22621852017-07-06 15:40:25 -07001937 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1938 nr_scanned);
1939 } else {
1940 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001941 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
Roman Gushchin22621852017-07-06 15:40:25 -07001942 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1943 nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001944 }
Mel Gorman599d0c92016-07-28 15:45:31 -07001945 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001946
Hillf Dantond563c052012-03-21 16:34:02 -07001947 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001948 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001949
Shaohua Lia128ca72017-05-03 14:52:22 -07001950 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
Michal Hocko3c710c12017-02-22 15:44:27 -08001951 &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001952
Mel Gorman599d0c92016-07-28 15:45:31 -07001953 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001954
Roman Gushchin22621852017-07-06 15:40:25 -07001955 if (current_is_kswapd()) {
1956 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001957 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
Roman Gushchin22621852017-07-06 15:40:25 -07001958 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1959 nr_reclaimed);
1960 } else {
1961 if (global_reclaim(sc))
Mel Gorman599d0c92016-07-28 15:45:31 -07001962 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
Roman Gushchin22621852017-07-06 15:40:25 -07001963 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1964 nr_reclaimed);
Ying Han904249a2012-04-25 16:01:48 -07001965 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001966
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001967 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001968
Mel Gorman599d0c92016-07-28 15:45:31 -07001969 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001970
Mel Gorman599d0c92016-07-28 15:45:31 -07001971 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001972
Johannes Weiner747db952014-08-08 14:19:24 -07001973 mem_cgroup_uncharge_list(&page_list);
Mel Gorman2d4894b2017-11-15 17:37:59 -08001974 free_unref_page_list(&page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001975
Mel Gorman92df3a72011-10-31 17:07:56 -07001976 /*
Andrey Ryabinin1c610d52018-03-22 16:17:42 -07001977 * If dirty pages are scanned that are not queued for IO, it
1978 * implies that flushers are not doing their job. This can
1979 * happen when memory pressure pushes dirty pages to the end of
1980 * the LRU before the dirty limits are breached and the dirty
1981 * data has expired. It can also happen when the proportion of
1982 * dirty pages grows not through writes but through memory
1983 * pressure reclaiming all the clean cache. And in some cases,
1984 * the flushers simply cannot keep up with the allocation
1985 * rate. Nudge the flusher threads in case they are asleep.
1986 */
1987 if (stat.nr_unqueued_dirty == nr_taken)
1988 wakeup_flusher_threads(WB_REASON_VMSCAN);
1989
Andrey Ryabinind108c772018-04-10 16:27:59 -07001990 sc->nr.dirty += stat.nr_dirty;
1991 sc->nr.congested += stat.nr_congested;
1992 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1993 sc->nr.writeback += stat.nr_writeback;
1994 sc->nr.immediate += stat.nr_immediate;
1995 sc->nr.taken += nr_taken;
1996 if (file)
1997 sc->nr.file_taken += nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001998
Mel Gorman599d0c92016-07-28 15:45:31 -07001999 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
Steven Rostedtd51d1e62018-04-10 16:28:07 -07002000 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08002001 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002}
2003
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002004/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 * This moves pages from the active list to the inactive list.
2006 *
2007 * We move them the other way if the page is referenced by one or more
2008 * processes, from rmap.
2009 *
2010 * If the pages are mostly unmapped, the processing is fast and it is
Mel Gormana52633d2016-07-28 15:45:28 -07002011 * appropriate to hold zone_lru_lock across the whole operation. But if
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 * the pages are mapped, the processing is slow (page_referenced()) so we
Mel Gormana52633d2016-07-28 15:45:28 -07002013 * should drop zone_lru_lock around each page. It's impossible to balance
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 * this, so instead we remove the pages from the LRU while processing them.
2015 * It is safe to rely on PG_active against the non-LRU pages in here because
2016 * nobody will play with that bit on a non-LRU page.
2017 *
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002018 * The downside is that we have to touch page->_refcount against each page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 * But we had to alter page->flags anyway.
Michal Hocko9d998b42017-02-22 15:44:18 -08002020 *
2021 * Returns the number of pages moved to the given lru.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002023
Michal Hocko9d998b42017-02-22 15:44:18 -08002024static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07002025 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002026 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07002027 enum lru_list lru)
2028{
Mel Gorman599d0c92016-07-28 15:45:31 -07002029 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002030 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002031 int nr_pages;
Michal Hocko9d998b42017-02-22 15:44:18 -08002032 int nr_moved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07002033
Wu Fengguang3eb41402009-06-16 15:33:13 -07002034 while (!list_empty(list)) {
2035 page = lru_to_page(list);
Mel Gorman599d0c92016-07-28 15:45:31 -07002036 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002037
Sasha Levin309381fea2014-01-23 15:52:54 -08002038 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002039 SetPageLRU(page);
2040
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002041 nr_pages = hpage_nr_pages(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07002042 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08002043 list_move(&page->lru, &lruvec->lists[lru]);
Wu Fengguang3eb41402009-06-16 15:33:13 -07002044
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002045 if (put_page_testzero(page)) {
2046 __ClearPageLRU(page);
2047 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07002048 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002049
2050 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002051 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07002052 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002053 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07002054 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002055 } else
2056 list_add(&page->lru, pages_to_free);
Michal Hocko9d998b42017-02-22 15:44:18 -08002057 } else {
2058 nr_moved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07002059 }
2060 }
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002061
Roman Gushchin22621852017-07-06 15:40:25 -07002062 if (!is_active_lru(lru)) {
Michal Hockof0958902017-02-22 15:45:55 -08002063 __count_vm_events(PGDEACTIVATE, nr_moved);
Roman Gushchin22621852017-07-06 15:40:25 -07002064 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
2065 nr_moved);
2066 }
Michal Hocko9d998b42017-02-22 15:44:18 -08002067
2068 return nr_moved;
Wu Fengguang3eb41402009-06-16 15:33:13 -07002069}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002070
Hugh Dickinsf6260122012-01-12 17:20:06 -08002071static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002072 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002073 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002074 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07002076 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08002077 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07002078 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002080 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002081 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002083 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Michal Hocko9d998b42017-02-22 15:44:18 -08002084 unsigned nr_deactivate, nr_activate;
2085 unsigned nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07002086 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002087 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002088 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089
2090 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002091
2092 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07002093 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07002094
Mel Gorman599d0c92016-07-28 15:45:31 -07002095 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002096
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002097 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2098 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002099
Mel Gorman599d0c92016-07-28 15:45:31 -07002100 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002101 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002102
Mel Gorman599d0c92016-07-28 15:45:31 -07002103 __count_vm_events(PGREFILL, nr_scanned);
Roman Gushchin22621852017-07-06 15:40:25 -07002104 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002105
Mel Gorman599d0c92016-07-28 15:45:31 -07002106 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 while (!list_empty(&l_hold)) {
2109 cond_resched();
2110 page = lru_to_page(&l_hold);
2111 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002112
Hugh Dickins39b5f292012-10-08 16:33:18 -07002113 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002114 putback_lru_page(page);
2115 continue;
2116 }
2117
Mel Gormancc715d92012-03-21 16:34:00 -07002118 if (unlikely(buffer_heads_over_limit)) {
2119 if (page_has_private(page) && trylock_page(page)) {
2120 if (page_has_private(page))
2121 try_to_release_page(page, 0);
2122 unlock_page(page);
2123 }
2124 }
2125
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002126 if (page_referenced(page, 0, sc->target_mem_cgroup,
2127 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08002128 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002129 /*
2130 * Identify referenced, file-backed active pages and
2131 * give them one more trip around the active list. So
2132 * that executable code get better chances to stay in
2133 * memory under moderate memory pressure. Anon pages
2134 * are not likely to be evicted by use-once streaming
2135 * IO, plus JVM can create lots of anon VM_EXEC pages,
2136 * so we ignore them here.
2137 */
Wu Fengguang41e20982009-10-26 16:49:53 -07002138 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002139 list_add(&page->lru, &l_active);
2140 continue;
2141 }
2142 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002143
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002144 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 list_add(&page->lru, &l_inactive);
2146 }
2147
Andrew Mortonb5557492009-01-06 14:40:13 -08002148 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002149 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002150 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002151 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002152 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002153 * Count referenced pages from currently used mappings as rotated,
2154 * even though only some of them are actually re-activated. This
2155 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002156 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002157 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002158 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002159
Michal Hocko9d998b42017-02-22 15:44:18 -08002160 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2161 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
Mel Gorman599d0c92016-07-28 15:45:31 -07002162 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2163 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002164
Johannes Weiner747db952014-08-08 14:19:24 -07002165 mem_cgroup_uncharge_list(&l_hold);
Mel Gorman2d4894b2017-11-15 17:37:59 -08002166 free_unref_page_list(&l_hold);
Michal Hocko9d998b42017-02-22 15:44:18 -08002167 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2168 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169}
2170
Rik van Riel59dc76b2016-05-20 16:56:31 -07002171/*
2172 * The inactive anon list should be small enough that the VM never has
2173 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002174 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002175 * The inactive file list should be small enough to leave most memory
2176 * to the established workingset on the scan-resistant active list,
2177 * but large enough to avoid thrashing the aggregate readahead window.
2178 *
2179 * Both inactive lists should also be large enough that each inactive
2180 * page has a chance to be referenced again before it is reclaimed.
2181 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002182 * If that fails and refaulting is observed, the inactive list grows.
2183 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002184 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
Andrey Ryabinin3a50d142017-11-15 17:34:15 -08002185 * on this LRU, maintained by the pageout code. An inactive_ratio
Rik van Riel59dc76b2016-05-20 16:56:31 -07002186 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2187 *
2188 * total target max
2189 * memory ratio inactive
2190 * -------------------------------------
2191 * 10MB 1 5MB
2192 * 100MB 1 50MB
2193 * 1GB 3 250MB
2194 * 10GB 10 0.9GB
2195 * 100GB 31 3GB
2196 * 1TB 101 10GB
2197 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002198 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002199static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
Kuo-Hsin Yangc1d98b72019-07-11 20:52:04 -07002200 struct scan_control *sc, bool trace)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002201{
Michal Hockofd538802017-02-22 15:45:58 -08002202 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002203 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2204 enum lru_list inactive_lru = file * LRU_FILE;
2205 unsigned long inactive, active;
2206 unsigned long inactive_ratio;
2207 unsigned long refaults;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002208 unsigned long gb;
2209
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002210 /*
2211 * If we don't have swap space, anonymous page deactivation
2212 * is pointless.
2213 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002214 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002215 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002216
Michal Hockofd538802017-02-22 15:45:58 -08002217 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2218 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002219
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002220 /*
2221 * When refaults are being observed, it means a new workingset
2222 * is being established. Disable active list protection to get
2223 * rid of the stale workingset quickly.
2224 */
Johannes Weiner6536de82019-04-18 17:50:34 -07002225 refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE);
Kuo-Hsin Yangc1d98b72019-07-11 20:52:04 -07002226 if (file && lruvec->refaults != refaults) {
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002227 inactive_ratio = 0;
2228 } else {
2229 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2230 if (gb)
2231 inactive_ratio = int_sqrt(10 * gb);
2232 else
2233 inactive_ratio = 1;
2234 }
2235
Kuo-Hsin Yangc1d98b72019-07-11 20:52:04 -07002236 if (trace)
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002237 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2238 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2239 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2240 inactive_ratio, file);
Michal Hockofd538802017-02-22 15:45:58 -08002241
Rik van Riel59dc76b2016-05-20 16:56:31 -07002242 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002243}
2244
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002245static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Johannes Weiner6536de82019-04-18 17:50:34 -07002246 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002247{
Rik van Rielb39415b2009-12-14 17:59:48 -08002248 if (is_active_lru(lru)) {
Johannes Weiner6536de82019-04-18 17:50:34 -07002249 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002250 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002251 return 0;
2252 }
2253
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002254 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002255}
2256
Johannes Weiner9a265112013-02-22 16:32:17 -08002257enum scan_balance {
2258 SCAN_EQUAL,
2259 SCAN_FRACT,
2260 SCAN_ANON,
2261 SCAN_FILE,
2262};
2263
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002265 * Determine how aggressively the anon and file LRU lists should be
2266 * scanned. The relative value of each set of LRU lists is determined
2267 * by looking at the fraction of the pages scanned we did rotate back
2268 * onto the active list instead of evict.
2269 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002270 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2271 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002272 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002273static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002274 struct scan_control *sc, unsigned long *nr,
2275 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002276{
Vladimir Davydov33377672016-01-20 15:02:59 -08002277 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002278 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002279 u64 fraction[2];
2280 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002281 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002282 unsigned long anon_prio, file_prio;
2283 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002284 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002285 unsigned long ap, fp;
2286 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002287
2288 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002289 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002290 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002291 goto out;
2292 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002293
Johannes Weiner10316b32013-02-22 16:32:14 -08002294 /*
2295 * Global reclaim will swap to prevent OOM even with no
2296 * swappiness, but memcg users want to use this knob to
2297 * disable swapping for individual groups completely when
2298 * using the memory controller's swap limit feature would be
2299 * too expensive.
2300 */
Johannes Weiner02695172014-08-06 16:06:17 -07002301 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002302 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002303 goto out;
2304 }
2305
2306 /*
2307 * Do not apply any pressure balancing cleverness when the
2308 * system is close to OOM, scan both anon and file equally
2309 * (unless the swappiness setting disagrees with swapping).
2310 */
Johannes Weiner02695172014-08-06 16:06:17 -07002311 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002312 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002313 goto out;
2314 }
2315
Johannes Weiner11d16c22013-02-22 16:32:15 -08002316 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002317 * Prevent the reclaimer from falling into the cache trap: as
2318 * cache pages start out inactive, every cache fault will tip
2319 * the scan balance towards the file LRU. And as the file LRU
2320 * shrinks, so does the window for rotation from references.
2321 * This means we have a runaway feedback loop where a tiny
2322 * thrashing file LRU becomes infinitely more attractive than
2323 * anon pages. Try to detect this based on file LRU size.
2324 */
2325 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002326 unsigned long pgdatfile;
2327 unsigned long pgdatfree;
2328 int z;
2329 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002330
Mel Gorman599d0c92016-07-28 15:45:31 -07002331 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2332 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2333 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002334
Mel Gorman599d0c92016-07-28 15:45:31 -07002335 for (z = 0; z < MAX_NR_ZONES; z++) {
2336 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002337 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002338 continue;
2339
2340 total_high_wmark += high_wmark_pages(zone);
2341 }
2342
2343 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
David Rientjes06226222017-07-10 15:47:20 -07002344 /*
2345 * Force SCAN_ANON if there are enough inactive
2346 * anonymous pages on the LRU in eligible zones.
2347 * Otherwise, the small LRU gets thrashed.
2348 */
Johannes Weiner6536de82019-04-18 17:50:34 -07002349 if (!inactive_list_is_low(lruvec, false, sc, false) &&
David Rientjes06226222017-07-10 15:47:20 -07002350 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2351 >> sc->priority) {
2352 scan_balance = SCAN_ANON;
2353 goto out;
2354 }
Johannes Weiner62376252014-05-06 12:50:07 -07002355 }
2356 }
2357
2358 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002359 * If there is enough inactive page cache, i.e. if the size of the
2360 * inactive list is greater than that of the active list *and* the
2361 * inactive list actually has some pages to scan on this priority, we
2362 * do not reclaim anything from the anonymous working set right now.
2363 * Without the second condition we could end up never scanning an
2364 * lruvec even if it has plenty of old anonymous pages unless the
2365 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002366 */
Johannes Weiner6536de82019-04-18 17:50:34 -07002367 if (!inactive_list_is_low(lruvec, true, sc, false) &&
Michal Hocko71ab6cf2017-02-22 15:46:01 -08002368 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002369 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002370 goto out;
2371 }
2372
Johannes Weiner9a265112013-02-22 16:32:17 -08002373 scan_balance = SCAN_FRACT;
2374
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002375 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002376 * With swappiness at 100, anonymous and file have the same priority.
2377 * This scanning priority is essentially the inverse of IO cost.
2378 */
Johannes Weiner02695172014-08-06 16:06:17 -07002379 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002380 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002381
2382 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002383 * OK, so we have swap space and a fair amount of page cache
2384 * pages. We use the recently rotated / recently scanned
2385 * ratios to determine how valuable each cache is.
2386 *
2387 * Because workloads change over time (and to avoid overflow)
2388 * we keep these statistics as a floating average, which ends
2389 * up weighing recent references more than old ones.
2390 *
2391 * anon in [0], file in [1]
2392 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002393
Michal Hockofd538802017-02-22 15:45:58 -08002394 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2395 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2396 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2397 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002398
Mel Gorman599d0c92016-07-28 15:45:31 -07002399 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002400 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002401 reclaim_stat->recent_scanned[0] /= 2;
2402 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002403 }
2404
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002405 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002406 reclaim_stat->recent_scanned[1] /= 2;
2407 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002408 }
2409
2410 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002411 * The amount of pressure on anon vs file pages is inversely
2412 * proportional to the fraction of recently scanned pages on
2413 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002414 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002415 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002416 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002417
Satoru Moriyafe350042012-05-29 15:06:47 -07002418 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002419 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002420 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002421
Shaohua Li76a33fc2010-05-24 14:32:36 -07002422 fraction[0] = ap;
2423 fraction[1] = fp;
2424 denominator = ap + fp + 1;
2425out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002426 *lru_pages = 0;
2427 for_each_evictable_lru(lru) {
2428 int file = is_file_lru(lru);
2429 unsigned long size;
2430 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002431
Johannes Weiner688035f2017-05-03 14:52:07 -07002432 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2433 scan = size >> sc->priority;
2434 /*
2435 * If the cgroup's already been deleted, make sure to
2436 * scrape out the remaining cache.
2437 */
2438 if (!scan && !mem_cgroup_online(memcg))
2439 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002440
Johannes Weiner688035f2017-05-03 14:52:07 -07002441 switch (scan_balance) {
2442 case SCAN_EQUAL:
2443 /* Scan lists relative to size */
2444 break;
2445 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002446 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002447 * Scan types proportional to swappiness and
2448 * their relative recent reclaim efficiency.
Roman Gushchina5e88092018-10-26 15:03:27 -07002449 * Make sure we don't miss the last page
2450 * because of a round-off error.
Johannes Weiner9a265112013-02-22 16:32:17 -08002451 */
Roman Gushchina5e88092018-10-26 15:03:27 -07002452 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2453 denominator);
Johannes Weiner688035f2017-05-03 14:52:07 -07002454 break;
2455 case SCAN_FILE:
2456 case SCAN_ANON:
2457 /* Scan one type exclusively */
2458 if ((scan_balance == SCAN_FILE) != file) {
2459 size = 0;
2460 scan = 0;
2461 }
2462 break;
2463 default:
2464 /* Look ma, no brain */
2465 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002466 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002467
2468 *lru_pages += size;
2469 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002470 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002471}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002472
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002473/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002474 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002475 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002476static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002477 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002478{
Mel Gormanef8f2322016-07-28 15:46:05 -07002479 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002480 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002481 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002482 unsigned long nr_to_scan;
2483 enum lru_list lru;
2484 unsigned long nr_reclaimed = 0;
2485 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2486 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002487 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002488
Vladimir Davydov33377672016-01-20 15:02:59 -08002489 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002490
Mel Gormane82e0562013-07-03 15:01:44 -07002491 /* Record the original scan target for proportional adjustments later */
2492 memcpy(targets, nr, sizeof(nr));
2493
Mel Gorman1a501902014-06-04 16:10:49 -07002494 /*
2495 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2496 * event that can occur when there is little memory pressure e.g.
2497 * multiple streaming readers/writers. Hence, we do not abort scanning
2498 * when the requested number of pages are reclaimed when scanning at
2499 * DEF_PRIORITY on the assumption that the fact we are direct
2500 * reclaiming implies that kswapd is not keeping up and it is best to
2501 * do a batch of work at once. For memcg reclaim one check is made to
2502 * abort proportional reclaim if either the file or anon lru has already
2503 * dropped to zero at the first pass.
2504 */
2505 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2506 sc->priority == DEF_PRIORITY);
2507
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002508 blk_start_plug(&plug);
2509 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2510 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002511 unsigned long nr_anon, nr_file, percentage;
2512 unsigned long nr_scanned;
2513
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002514 for_each_evictable_lru(lru) {
2515 if (nr[lru]) {
2516 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2517 nr[lru] -= nr_to_scan;
2518
2519 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner6536de82019-04-18 17:50:34 -07002520 lruvec, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002521 }
2522 }
Mel Gormane82e0562013-07-03 15:01:44 -07002523
Michal Hockobd041732016-12-02 17:26:48 -08002524 cond_resched();
2525
Mel Gormane82e0562013-07-03 15:01:44 -07002526 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2527 continue;
2528
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002529 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002530 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002531 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002532 * proportionally what was requested by get_scan_count(). We
2533 * stop reclaiming one LRU and reduce the amount scanning
2534 * proportional to the original scan target.
2535 */
2536 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2537 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2538
Mel Gorman1a501902014-06-04 16:10:49 -07002539 /*
2540 * It's just vindictive to attack the larger once the smaller
2541 * has gone to zero. And given the way we stop scanning the
2542 * smaller below, this makes sure that we only make one nudge
2543 * towards proportionality once we've got nr_to_reclaim.
2544 */
2545 if (!nr_file || !nr_anon)
2546 break;
2547
Mel Gormane82e0562013-07-03 15:01:44 -07002548 if (nr_file > nr_anon) {
2549 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2550 targets[LRU_ACTIVE_ANON] + 1;
2551 lru = LRU_BASE;
2552 percentage = nr_anon * 100 / scan_target;
2553 } else {
2554 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2555 targets[LRU_ACTIVE_FILE] + 1;
2556 lru = LRU_FILE;
2557 percentage = nr_file * 100 / scan_target;
2558 }
2559
2560 /* Stop scanning the smaller of the LRU */
2561 nr[lru] = 0;
2562 nr[lru + LRU_ACTIVE] = 0;
2563
2564 /*
2565 * Recalculate the other LRU scan count based on its original
2566 * scan target and the percentage scanning already complete
2567 */
2568 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2569 nr_scanned = targets[lru] - nr[lru];
2570 nr[lru] = targets[lru] * (100 - percentage) / 100;
2571 nr[lru] -= min(nr[lru], nr_scanned);
2572
2573 lru += LRU_ACTIVE;
2574 nr_scanned = targets[lru] - nr[lru];
2575 nr[lru] = targets[lru] * (100 - percentage) / 100;
2576 nr[lru] -= min(nr[lru], nr_scanned);
2577
2578 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002579 }
2580 blk_finish_plug(&plug);
2581 sc->nr_reclaimed += nr_reclaimed;
2582
2583 /*
2584 * Even if we did not try to evict anon pages at all, we want to
2585 * rebalance the anon lru active/inactive ratio.
2586 */
Johannes Weiner6536de82019-04-18 17:50:34 -07002587 if (inactive_list_is_low(lruvec, false, sc, true))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002588 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2589 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002590}
2591
Mel Gorman23b9da52012-05-29 15:06:20 -07002592/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002593static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002594{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002595 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002596 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002597 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002598 return true;
2599
2600 return false;
2601}
2602
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002603/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002604 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2605 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2606 * true if more pages should be reclaimed such that when the page allocator
2607 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2608 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002609 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002610static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002611 unsigned long nr_reclaimed,
2612 unsigned long nr_scanned,
2613 struct scan_control *sc)
2614{
2615 unsigned long pages_for_compaction;
2616 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002617 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002618
2619 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002620 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002621 return false;
2622
Mel Gorman28765922011-02-25 14:44:20 -08002623 /* Consider stopping depending on scan and reclaim activity */
Michal Hockodcda9b02017-07-12 14:36:45 -07002624 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
Mel Gorman28765922011-02-25 14:44:20 -08002625 /*
Michal Hockodcda9b02017-07-12 14:36:45 -07002626 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
Mel Gorman28765922011-02-25 14:44:20 -08002627 * full LRU list has been scanned and we are still failing
2628 * to reclaim pages. This full LRU scan is potentially
Michal Hockodcda9b02017-07-12 14:36:45 -07002629 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
Mel Gorman28765922011-02-25 14:44:20 -08002630 */
2631 if (!nr_reclaimed && !nr_scanned)
2632 return false;
2633 } else {
2634 /*
Michal Hockodcda9b02017-07-12 14:36:45 -07002635 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
Mel Gorman28765922011-02-25 14:44:20 -08002636 * fail without consequence, stop if we failed to reclaim
2637 * any pages from the last SWAP_CLUSTER_MAX number of
2638 * pages that were scanned. This will return to the
2639 * caller faster at the risk reclaim/compaction and
2640 * the resulting allocation attempt fails
2641 */
2642 if (!nr_reclaimed)
2643 return false;
2644 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002645
2646 /*
2647 * If we have not reclaimed enough pages for compaction and the
2648 * inactive lists are large enough, continue reclaiming
2649 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002650 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002651 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002652 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002653 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002654 if (sc->nr_reclaimed < pages_for_compaction &&
2655 inactive_lru_pages > pages_for_compaction)
2656 return true;
2657
2658 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002659 for (z = 0; z <= sc->reclaim_idx; z++) {
2660 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002661 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002662 continue;
2663
2664 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002665 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002666 case COMPACT_CONTINUE:
2667 return false;
2668 default:
2669 /* check next zone */
2670 ;
2671 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002672 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002673 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002674}
2675
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002676static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2677{
2678 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2679 (memcg && memcg_congested(pgdat, memcg));
2680}
2681
Mel Gorman970a39a2016-07-28 15:46:35 -07002682static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002683{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002684 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002685 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002686 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002687
Johannes Weiner56600482012-01-12 17:17:59 -08002688 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002689 struct mem_cgroup *root = sc->target_mem_cgroup;
2690 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002691 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002692 .priority = sc->priority,
2693 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002694 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002695 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002696
Andrey Ryabinind108c772018-04-10 16:27:59 -07002697 memset(&sc->nr, 0, sizeof(sc->nr));
2698
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002699 nr_reclaimed = sc->nr_reclaimed;
2700 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002701
Andrew Morton694fbc02013-09-24 15:27:37 -07002702 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2703 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002704 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002705 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002706 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002707
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002708 switch (mem_cgroup_protected(root, memcg)) {
2709 case MEMCG_PROT_MIN:
2710 /*
2711 * Hard protection.
2712 * If there is no reclaimable memory, OOM.
2713 */
2714 continue;
2715 case MEMCG_PROT_LOW:
2716 /*
2717 * Soft protection.
2718 * Respect the protection only as long as
2719 * there is an unprotected supply
2720 * of reclaimable memory from other cgroups.
2721 */
Yisheng Xied6622f62017-05-03 14:53:57 -07002722 if (!sc->memcg_low_reclaim) {
2723 sc->memcg_low_skipped = 1;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002724 continue;
Yisheng Xied6622f62017-05-03 14:53:57 -07002725 }
Johannes Weinere27be242018-04-10 16:29:45 -07002726 memcg_memory_event(memcg, MEMCG_LOW);
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002727 break;
2728 case MEMCG_PROT_NONE:
2729 break;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002730 }
2731
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002732 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002733 scanned = sc->nr_scanned;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002734 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2735 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002736
Vladimir Davydovaeed1d322018-08-17 15:48:17 -07002737 shrink_slab(sc->gfp_mask, pgdat->node_id,
2738 memcg, sc->priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002739
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002740 /* Record the group's reclaim efficiency */
2741 vmpressure(sc->gfp_mask, memcg, false,
2742 sc->nr_scanned - scanned,
2743 sc->nr_reclaimed - reclaimed);
2744
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002745 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002746 * Direct reclaim and kswapd have to scan all memory
2747 * cgroups to fulfill the overall scan target for the
Mel Gormana9dd0a82016-07-28 15:46:02 -07002748 * node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002749 *
2750 * Limit reclaim, on the other hand, only cares about
2751 * nr_to_reclaim pages to be reclaimed and it will
2752 * retry with decreasing priority if one round over the
2753 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002754 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002755 if (!global_reclaim(sc) &&
2756 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002757 mem_cgroup_iter_break(root, memcg);
2758 break;
2759 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002760 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002761
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002762 if (reclaim_state) {
2763 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2764 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002765 }
2766
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002767 /* Record the subtree's reclaim efficiency */
2768 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002769 sc->nr_scanned - nr_scanned,
2770 sc->nr_reclaimed - nr_reclaimed);
2771
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002772 if (sc->nr_reclaimed - nr_reclaimed)
2773 reclaimable = true;
2774
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002775 if (current_is_kswapd()) {
2776 /*
2777 * If reclaim is isolating dirty pages under writeback,
2778 * it implies that the long-lived page allocation rate
2779 * is exceeding the page laundering rate. Either the
2780 * global limits are not being effective at throttling
2781 * processes due to the page distribution throughout
2782 * zones or there is heavy usage of a slow backing
2783 * device. The only option is to throttle from reclaim
2784 * context which is not ideal as there is no guarantee
2785 * the dirtying process is throttled in the same way
2786 * balance_dirty_pages() manages.
2787 *
2788 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2789 * count the number of pages under pages flagged for
2790 * immediate reclaim and stall if any are encountered
2791 * in the nr_immediate check below.
2792 */
2793 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2794 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002795
Andrey Ryabinind108c772018-04-10 16:27:59 -07002796 /*
2797 * Tag a node as congested if all the dirty pages
2798 * scanned were backed by a congested BDI and
2799 * wait_iff_congested will stall.
2800 */
2801 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2802 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2803
2804 /* Allow kswapd to start writing pages during reclaim.*/
2805 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2806 set_bit(PGDAT_DIRTY, &pgdat->flags);
2807
2808 /*
2809 * If kswapd scans pages marked marked for immediate
2810 * reclaim and under writeback (nr_immediate), it
2811 * implies that pages are cycling through the LRU
2812 * faster than they are written so also forcibly stall.
2813 */
2814 if (sc->nr.immediate)
2815 congestion_wait(BLK_RW_ASYNC, HZ/10);
2816 }
2817
2818 /*
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002819 * Legacy memcg will stall in page writeback so avoid forcibly
2820 * stalling in wait_iff_congested().
2821 */
2822 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2823 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2824 set_memcg_congestion(pgdat, root, true);
2825
2826 /*
Andrey Ryabinind108c772018-04-10 16:27:59 -07002827 * Stall direct reclaim for IO completions if underlying BDIs
2828 * and node is congested. Allow kswapd to continue until it
2829 * starts encountering unqueued dirty pages or cycling through
2830 * the LRU too quickly.
2831 */
2832 if (!sc->hibernation_mode && !current_is_kswapd() &&
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002833 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2834 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002835
Mel Gormana9dd0a82016-07-28 15:46:02 -07002836 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002837 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002838
Johannes Weinerc73322d2017-05-03 14:51:51 -07002839 /*
2840 * Kswapd gives up on balancing particular nodes after too
2841 * many failures to reclaim anything from them and goes to
2842 * sleep. On reclaim progress, reset the failure counter. A
2843 * successful direct reclaim run will revive a dormant kswapd.
2844 */
2845 if (reclaimable)
2846 pgdat->kswapd_failures = 0;
2847
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002848 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002849}
2850
Vlastimil Babka53853e22014-10-09 15:27:02 -07002851/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002852 * Returns true if compaction should go ahead for a costly-order request, or
2853 * the allocation would already succeed without compaction. Return false if we
2854 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002855 */
Mel Gorman4f588332016-07-28 15:46:38 -07002856static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002857{
Mel Gorman31483b62016-07-28 15:45:46 -07002858 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002859 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002860
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002861 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2862 if (suitable == COMPACT_SUCCESS)
2863 /* Allocation should succeed already. Don't reclaim. */
2864 return true;
2865 if (suitable == COMPACT_SKIPPED)
2866 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002867 return false;
2868
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002869 /*
2870 * Compaction is already possible, but it takes time to run and there
2871 * are potentially other callers using the pages just freed. So proceed
2872 * with reclaim to make a buffer of free pages available to give
2873 * compaction a reasonable chance of completing and allocating the page.
2874 * Note that we won't actually reclaim the whole buffer in one attempt
2875 * as the target watermark in should_continue_reclaim() is lower. But if
2876 * we are already above the high+gap watermark, don't reclaim at all.
2877 */
2878 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2879
2880 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002881}
2882
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883/*
2884 * This is the direct reclaim path, for page-allocating processes. We only
2885 * try to reclaim pages from zones which will satisfy the caller's allocation
2886 * request.
2887 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 * If a zone is deemed to be full of pinned pages then just give it a light
2889 * scan then give up on it.
2890 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002891static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892{
Mel Gormandd1a2392008-04-28 02:12:17 -07002893 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002894 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002895 unsigned long nr_soft_reclaimed;
2896 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002897 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002898 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002899
Mel Gormancc715d92012-03-21 16:34:00 -07002900 /*
2901 * If the number of buffer_heads in the machine exceeds the maximum
2902 * allowed level, force direct reclaim to scan the highmem zone as
2903 * highmem pages could be pinning lowmem pages storing buffer_heads
2904 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002905 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002906 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002907 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002908 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002909 }
Mel Gormancc715d92012-03-21 16:34:00 -07002910
Mel Gormand4debc62010-08-09 17:19:29 -07002911 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002912 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002913 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002914 * Take care memory controller reclaiming has small influence
2915 * to global LRU.
2916 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002917 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002918 if (!cpuset_zone_allowed(zone,
2919 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002920 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002921
Johannes Weiner0b064962014-08-06 16:06:12 -07002922 /*
2923 * If we already have plenty of memory free for
2924 * compaction in this zone, don't free any more.
2925 * Even though compaction is invoked for any
2926 * non-zero order, only frequent costly order
2927 * reclamation is disruptive enough to become a
2928 * noticeable problem, like transparent huge
2929 * page allocations.
2930 */
2931 if (IS_ENABLED(CONFIG_COMPACTION) &&
2932 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002933 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002934 sc->compaction_ready = true;
2935 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002936 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002937
Andrew Morton0608f432013-09-24 15:27:41 -07002938 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002939 * Shrink each node in the zonelist once. If the
2940 * zonelist is ordered by zone (not the default) then a
2941 * node may be shrunk multiple times but in that case
2942 * the user prefers lower zones being preserved.
2943 */
2944 if (zone->zone_pgdat == last_pgdat)
2945 continue;
2946
2947 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002948 * This steals pages from memory cgroups over softlimit
2949 * and returns the number of reclaimed pages and
2950 * scanned pages. This works for global memory pressure
2951 * and balancing, not for a memcg's limit.
2952 */
2953 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002954 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002955 sc->order, sc->gfp_mask,
2956 &nr_soft_scanned);
2957 sc->nr_reclaimed += nr_soft_reclaimed;
2958 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002959 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002960 }
Nick Piggin408d8542006-09-25 23:31:27 -07002961
Mel Gorman79dafcd2016-07-28 15:45:53 -07002962 /* See comment about same check for global reclaim above */
2963 if (zone->zone_pgdat == last_pgdat)
2964 continue;
2965 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002966 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 }
Mel Gormane0c23272011-10-31 17:09:33 -07002968
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002969 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002970 * Restore to original mask to avoid the impact on the caller if we
2971 * promoted it to __GFP_HIGHMEM.
2972 */
2973 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002975
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002976static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2977{
2978 struct mem_cgroup *memcg;
2979
2980 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2981 do {
2982 unsigned long refaults;
2983 struct lruvec *lruvec;
2984
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002985 lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner6536de82019-04-18 17:50:34 -07002986 refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE);
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002987 lruvec->refaults = refaults;
2988 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2989}
2990
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991/*
2992 * This is the main entry point to direct page reclaim.
2993 *
2994 * If a full scan of the inactive list fails to free enough memory then we
2995 * are "out of memory" and something needs to be killed.
2996 *
2997 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2998 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002999 * caller can't do much about. We kick the writeback threads and take explicit
3000 * naps in the hope that some of these pages can be written. But if the
3001 * allocating task holds filesystem locks which prevent writeout this might not
3002 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07003003 *
3004 * returns: 0, if no pages reclaimed
3005 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 */
Mel Gormandac1d272008-04-28 02:12:12 -07003007static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003008 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009{
Johannes Weiner241994ed2015-02-11 15:26:06 -08003010 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003011 pg_data_t *last_pgdat;
3012 struct zoneref *z;
3013 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003014retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07003015 delayacct_freepages_start();
3016
Johannes Weiner89b5fae2012-01-12 17:17:50 -08003017 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07003018 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003020 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07003021 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3022 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08003023 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07003024 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07003025
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003026 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07003027 break;
3028
3029 if (sc->compaction_ready)
3030 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031
3032 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08003033 * If we're getting trouble reclaiming, start doing
3034 * writepage even in laptop mode.
3035 */
3036 if (sc->priority < DEF_PRIORITY - 2)
3037 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07003038 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003039
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003040 last_pgdat = NULL;
3041 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3042 sc->nodemask) {
3043 if (zone->zone_pgdat == last_pgdat)
3044 continue;
3045 last_pgdat = zone->zone_pgdat;
3046 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07003047 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003048 }
3049
Keika Kobayashi873b4772008-07-25 01:48:52 -07003050 delayacct_freepages_end();
3051
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003052 if (sc->nr_reclaimed)
3053 return sc->nr_reclaimed;
3054
Mel Gorman0cee34f2012-01-12 17:19:49 -08003055 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07003056 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08003057 return 1;
3058
Johannes Weiner241994ed2015-02-11 15:26:06 -08003059 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07003060 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08003061 sc->priority = initial_priority;
Yisheng Xied6622f62017-05-03 14:53:57 -07003062 sc->memcg_low_reclaim = 1;
3063 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08003064 goto retry;
3065 }
3066
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07003067 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068}
3069
Johannes Weinerc73322d2017-05-03 14:51:51 -07003070static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07003071{
3072 struct zone *zone;
3073 unsigned long pfmemalloc_reserve = 0;
3074 unsigned long free_pages = 0;
3075 int i;
3076 bool wmark_ok;
3077
Johannes Weinerc73322d2017-05-03 14:51:51 -07003078 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3079 return true;
3080
Mel Gorman55150612012-07-31 16:44:35 -07003081 for (i = 0; i <= ZONE_NORMAL; i++) {
3082 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd2017-05-03 14:51:54 -07003083 if (!managed_zone(zone))
3084 continue;
3085
3086 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07003087 continue;
3088
Mel Gorman55150612012-07-31 16:44:35 -07003089 pfmemalloc_reserve += min_wmark_pages(zone);
3090 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3091 }
3092
Mel Gorman675becc2014-06-04 16:07:35 -07003093 /* If there are no reserves (unexpected config) then do not throttle */
3094 if (!pfmemalloc_reserve)
3095 return true;
3096
Mel Gorman55150612012-07-31 16:44:35 -07003097 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3098
3099 /* kswapd must be awake if processes are being throttled */
3100 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07003101 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07003102 (enum zone_type)ZONE_NORMAL);
3103 wake_up_interruptible(&pgdat->kswapd_wait);
3104 }
3105
3106 return wmark_ok;
3107}
3108
3109/*
3110 * Throttle direct reclaimers if backing storage is backed by the network
3111 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3112 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08003113 * when the low watermark is reached.
3114 *
3115 * Returns true if a fatal signal was delivered during throttling. If this
3116 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07003117 */
Mel Gorman50694c22012-11-26 16:29:48 -08003118static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07003119 nodemask_t *nodemask)
3120{
Mel Gorman675becc2014-06-04 16:07:35 -07003121 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07003122 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07003123 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07003124
3125 /*
3126 * Kernel threads should not be throttled as they may be indirectly
3127 * responsible for cleaning pages necessary for reclaim to make forward
3128 * progress. kjournald for example may enter direct reclaim while
3129 * committing a transaction where throttling it could forcing other
3130 * processes to block on log_wait_commit().
3131 */
3132 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08003133 goto out;
3134
3135 /*
3136 * If a fatal signal is pending, this process should not throttle.
3137 * It should return quickly so it can exit and free its memory
3138 */
3139 if (fatal_signal_pending(current))
3140 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003141
Mel Gorman675becc2014-06-04 16:07:35 -07003142 /*
3143 * Check if the pfmemalloc reserves are ok by finding the first node
3144 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3145 * GFP_KERNEL will be required for allocating network buffers when
3146 * swapping over the network so ZONE_HIGHMEM is unusable.
3147 *
3148 * Throttling is based on the first usable node and throttled processes
3149 * wait on a queue until kswapd makes progress and wakes them. There
3150 * is an affinity then between processes waking up and where reclaim
3151 * progress has been made assuming the process wakes on the same node.
3152 * More importantly, processes running on remote nodes will not compete
3153 * for remote pfmemalloc reserves and processes on different nodes
3154 * should make reasonable progress.
3155 */
3156 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08003157 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07003158 if (zone_idx(zone) > ZONE_NORMAL)
3159 continue;
3160
3161 /* Throttle based on the first usable node */
3162 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07003163 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07003164 goto out;
3165 break;
3166 }
3167
3168 /* If no zone was usable by the allocation flags then do not throttle */
3169 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08003170 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003171
Mel Gorman68243e72012-07-31 16:44:39 -07003172 /* Account for the throttling */
3173 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3174
Mel Gorman55150612012-07-31 16:44:35 -07003175 /*
3176 * If the caller cannot enter the filesystem, it's possible that it
3177 * is due to the caller holding an FS lock or performing a journal
3178 * transaction in the case of a filesystem like ext[3|4]. In this case,
3179 * it is not safe to block on pfmemalloc_wait as kswapd could be
3180 * blocked waiting on the same lock. Instead, throttle for up to a
3181 * second before continuing.
3182 */
3183 if (!(gfp_mask & __GFP_FS)) {
3184 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003185 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003186
3187 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003188 }
3189
3190 /* Throttle until kswapd wakes the process */
3191 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003192 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003193
3194check_pending:
3195 if (fatal_signal_pending(current))
3196 return true;
3197
3198out:
3199 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003200}
3201
Mel Gormandac1d272008-04-28 02:12:12 -07003202unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003203 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003204{
Mel Gorman33906bc2010-08-09 17:19:16 -07003205 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003206 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003207 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003208 .gfp_mask = current_gfp_context(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07003209 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003210 .order = order,
3211 .nodemask = nodemask,
3212 .priority = DEF_PRIORITY,
3213 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003214 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003215 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003216 };
3217
Mel Gorman55150612012-07-31 16:44:35 -07003218 /*
Greg Thelenbb451fd2018-08-17 15:45:19 -07003219 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3220 * Confirm they are large enough for max values.
3221 */
3222 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3223 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3224 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3225
3226 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003227 * Do not enter reclaim if fatal signal was delivered while throttled.
3228 * 1 is returned so that the page allocator does not OOM kill at this
3229 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003230 */
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003231 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003232 return 1;
3233
Mel Gorman33906bc2010-08-09 17:19:16 -07003234 trace_mm_vmscan_direct_reclaim_begin(order,
3235 sc.may_writepage,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003236 sc.gfp_mask,
Mel Gormane5146b12016-07-28 15:46:47 -07003237 sc.reclaim_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003238
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003239 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003240
3241 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3242
3243 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003244}
3245
Andrew Mortonc255a452012-07-31 16:43:02 -07003246#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003247
Mel Gormana9dd0a82016-07-28 15:46:02 -07003248unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003249 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003250 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003251 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003252{
3253 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003254 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003255 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003256 .may_writepage = !laptop_mode,
3257 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003258 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003259 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07003260 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003261 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003262
Balbir Singh4e416952009-09-23 15:56:39 -07003263 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3264 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003265
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003266 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003267 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003268 sc.gfp_mask,
3269 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003270
Balbir Singh4e416952009-09-23 15:56:39 -07003271 /*
3272 * NOTE: Although we can get the priority field, using it
3273 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003274 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003275 * will pick up pages from other mem cgroup's as well. We hack
3276 * the priority and make it zero.
3277 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003278 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003279
3280 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3281
Ying Han0ae5e892011-05-26 16:25:25 -07003282 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003283 return sc.nr_reclaimed;
3284}
3285
Johannes Weiner72835c82012-01-12 17:18:32 -08003286unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003287 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003288 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003289 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003290{
Balbir Singh4e416952009-09-23 15:56:39 -07003291 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003292 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07003293 int nid;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003294 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08003295 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003296 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003297 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07003298 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003299 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003300 .target_mem_cgroup = memcg,
3301 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003302 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003303 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003304 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07003305 };
Balbir Singh66e17072008-02-07 00:13:56 -08003306
Ying Han889976d2011-05-26 16:25:33 -07003307 /*
3308 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3309 * take care of from where we get pages. So the node where we start the
3310 * scan does not need to be the current node.
3311 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003312 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003313
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003314 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003315
3316 trace_mm_vmscan_memcg_reclaim_begin(0,
3317 sc.may_writepage,
Mel Gormane5146b12016-07-28 15:46:47 -07003318 sc.gfp_mask,
3319 sc.reclaim_idx);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003320
Vlastimil Babka499118e2017-05-08 15:59:50 -07003321 noreclaim_flag = memalloc_noreclaim_save();
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003322 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003323 memalloc_noreclaim_restore(noreclaim_flag);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003324
3325 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3326
3327 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003328}
3329#endif
3330
Mel Gorman1d82de62016-07-28 15:45:43 -07003331static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003332 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003333{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003334 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003335
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003336 if (!total_swap_pages)
3337 return;
3338
3339 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3340 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003341 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003342
Johannes Weiner6536de82019-04-18 17:50:34 -07003343 if (inactive_list_is_low(lruvec, false, sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003344 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003345 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003346
3347 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3348 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003349}
3350
Mel Gormane716f2e2017-05-03 14:53:45 -07003351/*
3352 * Returns true if there is an eligible zone balanced for the request order
3353 * and classzone_idx
3354 */
3355static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003356{
Mel Gormane716f2e2017-05-03 14:53:45 -07003357 int i;
3358 unsigned long mark = -1;
3359 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003360
Mel Gormane716f2e2017-05-03 14:53:45 -07003361 for (i = 0; i <= classzone_idx; i++) {
3362 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07003363
Mel Gormane716f2e2017-05-03 14:53:45 -07003364 if (!managed_zone(zone))
3365 continue;
3366
3367 mark = high_wmark_pages(zone);
3368 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3369 return true;
3370 }
3371
3372 /*
3373 * If a node has no populated zone within classzone_idx, it does not
3374 * need balancing by definition. This can happen if a zone-restricted
3375 * allocation tries to wake a remote kswapd.
3376 */
3377 if (mark == -1)
3378 return true;
3379
3380 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003381}
3382
Mel Gorman631b6e02017-05-03 14:53:41 -07003383/* Clear pgdat state for congested, dirty or under writeback. */
3384static void clear_pgdat_congested(pg_data_t *pgdat)
3385{
3386 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3387 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3388 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3389}
3390
Mel Gorman1741c872011-01-13 15:46:21 -08003391/*
Mel Gorman55150612012-07-31 16:44:35 -07003392 * Prepare kswapd for sleeping. This verifies that there are no processes
3393 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3394 *
3395 * Returns true if kswapd is ready to sleep
3396 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003397static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003398{
Mel Gorman55150612012-07-31 16:44:35 -07003399 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003400 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07003401 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003402 * race between when kswapd checks the watermarks and a process gets
3403 * throttled. There is also a potential race if processes get
3404 * throttled, kswapd wakes, a large process exits thereby balancing the
3405 * zones, which causes kswapd to exit balance_pgdat() before reaching
3406 * the wake up checks. If kswapd is going to sleep, no process should
3407 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3408 * the wake up is premature, processes will wake kswapd and get
3409 * throttled again. The difference from wake ups in balance_pgdat() is
3410 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003411 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003412 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3413 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003414
Johannes Weinerc73322d2017-05-03 14:51:51 -07003415 /* Hopeless node, leave it to direct reclaim */
3416 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3417 return true;
3418
Mel Gormane716f2e2017-05-03 14:53:45 -07003419 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3420 clear_pgdat_congested(pgdat);
3421 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003422 }
3423
Shantanu Goel333b0a42017-05-03 14:53:38 -07003424 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003425}
3426
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003428 * kswapd shrinks a node of pages that are at or below the highest usable
3429 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003430 *
3431 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003432 * reclaim or if the lack of progress was due to pages under writeback.
3433 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003434 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003435static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003436 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003437{
Mel Gorman1d82de62016-07-28 15:45:43 -07003438 struct zone *zone;
3439 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003440
Mel Gorman1d82de62016-07-28 15:45:43 -07003441 /* Reclaim a number of pages proportional to the number of zones */
3442 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003443 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003444 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003445 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003446 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003447
Mel Gorman1d82de62016-07-28 15:45:43 -07003448 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003449 }
3450
Mel Gorman1d82de62016-07-28 15:45:43 -07003451 /*
3452 * Historically care was taken to put equal pressure on all zones but
3453 * now pressure is applied based on node LRU order.
3454 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003455 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003456
3457 /*
3458 * Fragmentation may mean that the system cannot be rebalanced for
3459 * high-order allocations. If twice the allocation size has been
3460 * reclaimed then recheck watermarks only at order-0 to prevent
3461 * excessive reclaim. Assume that a process requested a high-order
3462 * can direct reclaim/compact.
3463 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003464 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003465 sc->order = 0;
3466
Mel Gormanb8e83b92013-07-03 15:01:45 -07003467 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003468}
3469
3470/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003471 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3472 * that are eligible for use by the caller until at least one zone is
3473 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003475 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 *
3477 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003478 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Mel Gorman1d82de62016-07-28 15:45:43 -07003479 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3480 * or lower is eligible for reclaim until at least one usable zone is
3481 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003483static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003486 unsigned long nr_soft_reclaimed;
3487 unsigned long nr_soft_scanned;
Mel Gorman1d82de62016-07-28 15:45:43 -07003488 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003489 struct scan_control sc = {
3490 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003491 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003492 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003493 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003494 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003495 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003496 };
Omar Sandoval93781322018-06-07 17:07:02 -07003497
3498 __fs_reclaim_acquire();
3499
Christoph Lameterf8891e52006-06-30 01:55:45 -07003500 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003502 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07003503 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003504 bool raise_priority = true;
Omar Sandoval93781322018-06-07 17:07:02 -07003505 bool ret;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003506
Mel Gorman84c7a772016-07-28 15:46:44 -07003507 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508
Mel Gorman86c79f62016-07-28 15:45:59 -07003509 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003510 * If the number of buffer_heads exceeds the maximum allowed
3511 * then consider reclaiming from all zones. This has a dual
3512 * purpose -- on 64-bit systems it is expected that
3513 * buffer_heads are stripped during active rotation. On 32-bit
3514 * systems, highmem pages can pin lowmem memory and shrinking
3515 * buffers can relieve lowmem pressure. Reclaim may still not
3516 * go ahead if all eligible zones for the original allocation
3517 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003518 */
3519 if (buffer_heads_over_limit) {
3520 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3521 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003522 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003523 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
Mel Gorman970a39a2016-07-28 15:46:35 -07003525 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003526 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003529
Mel Gorman86c79f62016-07-28 15:45:59 -07003530 /*
Mel Gormane716f2e2017-05-03 14:53:45 -07003531 * Only reclaim if there are no eligible zones. Note that
3532 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3533 * have adjusted it.
Mel Gorman86c79f62016-07-28 15:45:59 -07003534 */
Mel Gormane716f2e2017-05-03 14:53:45 -07003535 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3536 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003537
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003539 * Do some background aging of the anon list, to give
3540 * pages a chance to be referenced before reclaiming. All
3541 * pages are rotated regardless of classzone as this is
3542 * about consistent aging.
3543 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003544 age_active_anon(pgdat, &sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003545
3546 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003547 * If we're getting trouble reclaiming, start doing writepage
3548 * even in laptop mode.
3549 */
Johannes Weiner047d72c2017-05-03 14:51:57 -07003550 if (sc.priority < DEF_PRIORITY - 2)
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003551 sc.may_writepage = 1;
3552
Mel Gorman1d82de62016-07-28 15:45:43 -07003553 /* Call soft limit reclaim before calling shrink_node. */
3554 sc.nr_scanned = 0;
3555 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07003556 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
Mel Gorman1d82de62016-07-28 15:45:43 -07003557 sc.gfp_mask, &nr_soft_scanned);
3558 sc.nr_reclaimed += nr_soft_reclaimed;
3559
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003560 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003561 * There should be no need to raise the scanning priority if
3562 * enough pages are already being scanned that that high
3563 * watermark would be met at 100% efficiency.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003565 if (kswapd_shrink_node(pgdat, &sc))
Mel Gorman1d82de62016-07-28 15:45:43 -07003566 raise_priority = false;
Mel Gorman55150612012-07-31 16:44:35 -07003567
3568 /*
3569 * If the low watermark is met there is no need for processes
3570 * to be throttled on pfmemalloc_wait as they should not be
3571 * able to safely make forward progress. Wake them
3572 */
3573 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
Johannes Weinerc73322d2017-05-03 14:51:51 -07003574 allow_direct_reclaim(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003575 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003576
Mel Gormanb8e83b92013-07-03 15:01:45 -07003577 /* Check if kswapd should be suspending */
Omar Sandoval93781322018-06-07 17:07:02 -07003578 __fs_reclaim_release();
3579 ret = try_to_freeze();
3580 __fs_reclaim_acquire();
3581 if (ret || kthread_should_stop())
Mel Gormanb8e83b92013-07-03 15:01:45 -07003582 break;
3583
3584 /*
3585 * Raise priority if scanning rate is too low or there was no
3586 * progress in reclaiming pages
3587 */
Johannes Weinerc73322d2017-05-03 14:51:51 -07003588 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3589 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003590 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003591 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592
Johannes Weinerc73322d2017-05-03 14:51:51 -07003593 if (!sc.nr_reclaimed)
3594 pgdat->kswapd_failures++;
3595
Mel Gormanb8e83b92013-07-03 15:01:45 -07003596out:
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003597 snapshot_refaults(NULL, pgdat);
Omar Sandoval93781322018-06-07 17:07:02 -07003598 __fs_reclaim_release();
Mel Gorman0abdee22011-01-13 15:46:22 -08003599 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003600 * Return the order kswapd stopped reclaiming at as
3601 * prepare_kswapd_sleep() takes it into account. If another caller
3602 * entered the allocator slow path while kswapd was awake, order will
3603 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003604 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003605 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606}
3607
Mel Gormane716f2e2017-05-03 14:53:45 -07003608/*
Shakeel Butt27ce6c22019-07-04 15:14:42 -07003609 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3610 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3611 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3612 * after previous reclaim attempt (node is still unbalanced). In that case
3613 * return the zone index of the previous kswapd reclaim cycle.
Mel Gormane716f2e2017-05-03 14:53:45 -07003614 */
3615static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
Shakeel Butt27ce6c22019-07-04 15:14:42 -07003616 enum zone_type prev_classzone_idx)
Mel Gormane716f2e2017-05-03 14:53:45 -07003617{
3618 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
Shakeel Butt27ce6c22019-07-04 15:14:42 -07003619 return prev_classzone_idx;
3620 return pgdat->kswapd_classzone_idx;
Mel Gormane716f2e2017-05-03 14:53:45 -07003621}
3622
Mel Gorman38087d92016-07-28 15:45:49 -07003623static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3624 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003625{
3626 long remaining = 0;
3627 DEFINE_WAIT(wait);
3628
3629 if (freezing(current) || kthread_should_stop())
3630 return;
3631
3632 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3633
Shantanu Goel333b0a42017-05-03 14:53:38 -07003634 /*
3635 * Try to sleep for a short interval. Note that kcompactd will only be
3636 * woken if it is possible to sleep for a short interval. This is
3637 * deliberate on the assumption that if reclaim cannot keep an
3638 * eligible zone balanced that it's also unlikely that compaction will
3639 * succeed.
3640 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003641 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003642 /*
3643 * Compaction records what page blocks it recently failed to
3644 * isolate pages from and skips them in the future scanning.
3645 * When kswapd is going to sleep, it is reasonable to assume
3646 * that pages and compaction may succeed so reset the cache.
3647 */
3648 reset_isolation_suitable(pgdat);
3649
3650 /*
3651 * We have freed the memory, now we should compact it to make
3652 * allocation of the requested order possible.
3653 */
Mel Gorman38087d92016-07-28 15:45:49 -07003654 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003655
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003656 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003657
3658 /*
3659 * If woken prematurely then reset kswapd_classzone_idx and
3660 * order. The values will either be from a wakeup request or
3661 * the previous request that slept prematurely.
3662 */
3663 if (remaining) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003664 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003665 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3666 }
3667
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003668 finish_wait(&pgdat->kswapd_wait, &wait);
3669 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3670 }
3671
3672 /*
3673 * After a short sleep, check if it was a premature sleep. If not, then
3674 * go fully to sleep until explicitly woken up.
3675 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003676 if (!remaining &&
3677 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003678 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3679
3680 /*
3681 * vmstat counters are not perfectly accurate and the estimated
3682 * value for counters such as NR_FREE_PAGES can deviate from the
3683 * true value by nr_online_cpus * threshold. To avoid the zone
3684 * watermarks being breached while under pressure, we reduce the
3685 * per-cpu vmstat threshold while kswapd is awake and restore
3686 * them before going back to sleep.
3687 */
3688 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003689
3690 if (!kthread_should_stop())
3691 schedule();
3692
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003693 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3694 } else {
3695 if (remaining)
3696 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3697 else
3698 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3699 }
3700 finish_wait(&pgdat->kswapd_wait, &wait);
3701}
3702
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703/*
3704 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003705 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 *
3707 * This basically trickles out pages so that we have _some_
3708 * free memory available even if there is no other activity
3709 * that frees anything up. This is needed for things like routing
3710 * etc, where we otherwise might have all activity going on in
3711 * asynchronous contexts that cannot page things out.
3712 *
3713 * If there are applications that are active memory-allocators
3714 * (most normal use), this basically shouldn't matter.
3715 */
3716static int kswapd(void *p)
3717{
Mel Gormane716f2e2017-05-03 14:53:45 -07003718 unsigned int alloc_order, reclaim_order;
3719 unsigned int classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720 pg_data_t *pgdat = (pg_data_t*)p;
3721 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003722
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723 struct reclaim_state reclaim_state = {
3724 .reclaimed_slab = 0,
3725 };
Rusty Russella70f7302009-03-13 14:49:46 +10303726 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727
Rusty Russell174596a2009-01-01 10:12:29 +10303728 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003729 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 current->reclaim_state = &reclaim_state;
3731
3732 /*
3733 * Tell the memory management that we're a "memory allocator",
3734 * and that if we need more memory we should get access to it
3735 * regardless (see "__alloc_pages()"). "kswapd" should
3736 * never get caught in the normal page freeing logic.
3737 *
3738 * (Kswapd normally doesn't need memory anyway, but sometimes
3739 * you need a small amount of memory in order to be able to
3740 * page out something else, and this flag essentially protects
3741 * us from recursively trying to free more memory as we're
3742 * trying to free the first piece of memory in the first place).
3743 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003744 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003745 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746
Mel Gormane716f2e2017-05-03 14:53:45 -07003747 pgdat->kswapd_order = 0;
3748 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003750 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003751
Mel Gormane716f2e2017-05-03 14:53:45 -07003752 alloc_order = reclaim_order = pgdat->kswapd_order;
3753 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3754
Mel Gorman38087d92016-07-28 15:45:49 -07003755kswapd_try_sleep:
3756 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3757 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003758
Mel Gorman38087d92016-07-28 15:45:49 -07003759 /* Read the new order and classzone_idx */
3760 alloc_order = reclaim_order = pgdat->kswapd_order;
Shakeel Butt27ce6c22019-07-04 15:14:42 -07003761 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003762 pgdat->kswapd_order = 0;
Mel Gormane716f2e2017-05-03 14:53:45 -07003763 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764
David Rientjes8fe23e02009-12-14 17:58:33 -08003765 ret = try_to_freeze();
3766 if (kthread_should_stop())
3767 break;
3768
3769 /*
3770 * We can speed up thawing tasks if we don't call balance_pgdat
3771 * after returning from the refrigerator
3772 */
Mel Gorman38087d92016-07-28 15:45:49 -07003773 if (ret)
3774 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003775
Mel Gorman38087d92016-07-28 15:45:49 -07003776 /*
3777 * Reclaim begins at the requested order but if a high-order
3778 * reclaim fails then kswapd falls back to reclaiming for
3779 * order-0. If that happens, kswapd will consider sleeping
3780 * for the order it finished reclaiming at (reclaim_order)
3781 * but kcompactd is woken to compact for the original
3782 * request (alloc_order).
3783 */
Mel Gormane5146b12016-07-28 15:46:47 -07003784 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3785 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003786 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3787 if (reclaim_order < alloc_order)
3788 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003790
Johannes Weiner71abdc12014-06-06 14:35:35 -07003791 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003792 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003793
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 return 0;
3795}
3796
3797/*
David Rientjes5ecd9d42018-04-05 16:25:16 -07003798 * A zone is low on free memory or too fragmented for high-order memory. If
3799 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3800 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3801 * has failed or is not needed, still wake up kcompactd if only compaction is
3802 * needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 */
David Rientjes5ecd9d42018-04-05 16:25:16 -07003804void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3805 enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806{
3807 pg_data_t *pgdat;
3808
Mel Gorman6aa303d2016-09-01 16:14:55 -07003809 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810 return;
3811
David Rientjes5ecd9d42018-04-05 16:25:16 -07003812 if (!cpuset_zone_allowed(zone, gfp_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003814 pgdat = zone->zone_pgdat;
Shakeel Butt27ce6c22019-07-04 15:14:42 -07003815
3816 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3817 pgdat->kswapd_classzone_idx = classzone_idx;
3818 else
3819 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3820 classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003821 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003822 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003824
David Rientjes5ecd9d42018-04-05 16:25:16 -07003825 /* Hopeless node, leave it to direct reclaim if possible */
3826 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3827 pgdat_balanced(pgdat, order, classzone_idx)) {
3828 /*
3829 * There may be plenty of free memory available, but it's too
3830 * fragmented for high-order allocations. Wake up kcompactd
3831 * and rely on compaction_suitable() to determine if it's
3832 * needed. If it fails, it will defer subsequent attempts to
3833 * ratelimit its work.
3834 */
3835 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3836 wakeup_kcompactd(pgdat, order, classzone_idx);
Johannes Weinerc73322d2017-05-03 14:51:51 -07003837 return;
David Rientjes5ecd9d42018-04-05 16:25:16 -07003838 }
Johannes Weinerc73322d2017-05-03 14:51:51 -07003839
David Rientjes5ecd9d42018-04-05 16:25:16 -07003840 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3841 gfp_flags);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003842 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843}
3844
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003845#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003847 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003848 * freed pages.
3849 *
3850 * Rather than trying to age LRUs the aim is to preserve the overall
3851 * LRU order by reclaiming preferentially
3852 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003854unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003856 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003857 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003858 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003859 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003860 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003861 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003862 .may_writepage = 1,
3863 .may_unmap = 1,
3864 .may_swap = 1,
3865 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003867 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003868 struct task_struct *p = current;
3869 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003870 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871
Peter Zijlstrad92a8cf2017-03-03 10:13:38 +01003872 fs_reclaim_acquire(sc.gfp_mask);
Omar Sandoval93781322018-06-07 17:07:02 -07003873 noreclaim_flag = memalloc_noreclaim_save();
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003874 reclaim_state.reclaimed_slab = 0;
3875 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003876
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003877 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003878
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003879 p->reclaim_state = NULL;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003880 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07003881 fs_reclaim_release(sc.gfp_mask);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003882
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003883 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003885#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887/* It's optimal to keep kswapds on the same CPUs as their memory, but
3888 not required for correctness. So if the last cpu in a node goes
3889 away, we get changed to run anywhere: as the first one comes back,
3890 restore their cpu bindings. */
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003891static int kswapd_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003893 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003895 for_each_node_state(nid, N_MEMORY) {
3896 pg_data_t *pgdat = NODE_DATA(nid);
3897 const struct cpumask *mask;
Rusty Russella70f7302009-03-13 14:49:46 +10303898
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003899 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003900
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003901 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3902 /* One of our CPUs online: restore mask */
3903 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 }
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003905 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907
Yasunori Goto3218ae12006-06-27 02:53:33 -07003908/*
3909 * This kswapd start function will be called by init and node-hot-add.
3910 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3911 */
3912int kswapd_run(int nid)
3913{
3914 pg_data_t *pgdat = NODE_DATA(nid);
3915 int ret = 0;
3916
3917 if (pgdat->kswapd)
3918 return 0;
3919
3920 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3921 if (IS_ERR(pgdat->kswapd)) {
3922 /* failure at boot is fatal */
Thomas Gleixnerc6202ad2017-05-16 20:42:46 +02003923 BUG_ON(system_state < SYSTEM_RUNNING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003924 pr_err("Failed to start kswapd on node %d\n", nid);
3925 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003926 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003927 }
3928 return ret;
3929}
3930
David Rientjes8fe23e02009-12-14 17:58:33 -08003931/*
Jiang Liud8adde12012-07-11 14:01:52 -07003932 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003933 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003934 */
3935void kswapd_stop(int nid)
3936{
3937 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3938
Jiang Liud8adde12012-07-11 14:01:52 -07003939 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003940 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003941 NODE_DATA(nid)->kswapd = NULL;
3942 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003943}
3944
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945static int __init kswapd_init(void)
3946{
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003947 int nid, ret;
Andrew Morton69e05942006-03-22 00:08:19 -08003948
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003950 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003951 kswapd_run(nid);
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003952 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3953 "mm/vmscan:online", kswapd_cpu_online,
3954 NULL);
3955 WARN_ON(ret < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 return 0;
3957}
3958
3959module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003960
3961#ifdef CONFIG_NUMA
3962/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003963 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003964 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003965 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08003966 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003967 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003968int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003969
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003970#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003971#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003972#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003973#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003974
Christoph Lameter9eeff232006-01-18 17:42:31 -08003975/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003976 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08003977 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3978 * a zone.
3979 */
Mel Gormana5f5f912016-07-28 15:46:32 -07003980#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08003981
Christoph Lameter9eeff232006-01-18 17:42:31 -08003982/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003983 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07003984 * occur.
3985 */
3986int sysctl_min_unmapped_ratio = 1;
3987
3988/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003989 * If the number of slab pages in a zone grows beyond this percentage then
3990 * slab reclaim needs to occur.
3991 */
3992int sysctl_min_slab_ratio = 5;
3993
Mel Gorman11fb9982016-07-28 15:46:20 -07003994static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003995{
Mel Gorman11fb9982016-07-28 15:46:20 -07003996 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3997 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3998 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07003999
4000 /*
4001 * It's possible for there to be more file mapped pages than
4002 * accounted for by the pages on the file LRU lists because
4003 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4004 */
4005 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4006}
4007
4008/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07004009static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07004010{
Alexandru Moised031a152015-11-05 18:48:08 -08004011 unsigned long nr_pagecache_reclaimable;
4012 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07004013
4014 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004015 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07004016 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07004017 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07004018 * a better estimate
4019 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004020 if (node_reclaim_mode & RECLAIM_UNMAP)
4021 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004022 else
Mel Gormana5f5f912016-07-28 15:46:32 -07004023 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004024
4025 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07004026 if (!(node_reclaim_mode & RECLAIM_WRITE))
4027 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004028
4029 /* Watch for any possible underflows due to delta */
4030 if (unlikely(delta > nr_pagecache_reclaimable))
4031 delta = nr_pagecache_reclaimable;
4032
4033 return nr_pagecache_reclaimable - delta;
4034}
4035
Christoph Lameter0ff38492006-09-25 23:31:52 -07004036/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004037 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004038 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004039static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004040{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08004041 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08004042 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004043 struct task_struct *p = current;
4044 struct reclaim_state reclaim_state;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004045 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08004046 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08004047 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004048 .gfp_mask = current_gfp_context(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07004049 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07004050 .priority = NODE_RECLAIM_PRIORITY,
4051 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4052 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07004053 .may_swap = 1,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004054 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08004055 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08004056
Christoph Lameter9eeff232006-01-18 17:42:31 -08004057 cond_resched();
Omar Sandoval93781322018-06-07 17:07:02 -07004058 fs_reclaim_acquire(sc.gfp_mask);
Christoph Lameterd4f77962006-02-24 13:04:22 -08004059 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004060 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08004061 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004062 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08004063 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07004064 noreclaim_flag = memalloc_noreclaim_save();
4065 p->flags |= PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004066 reclaim_state.reclaimed_slab = 0;
4067 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08004068
Mel Gormana5f5f912016-07-28 15:46:32 -07004069 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07004070 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07004071 * Free memory by calling shrink node with increasing
Christoph Lameter0ff38492006-09-25 23:31:52 -07004072 * priorities until we have enough memory freed.
4073 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07004074 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07004075 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004076 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07004077 }
Christoph Lameterc84db232006-02-01 03:05:29 -08004078
Christoph Lameter9eeff232006-01-18 17:42:31 -08004079 p->reclaim_state = NULL;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004080 current->flags &= ~PF_SWAPWRITE;
4081 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004082 fs_reclaim_release(sc.gfp_mask);
Rik van Riela79311c2009-01-06 14:40:01 -08004083 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004084}
Andrew Morton179e9632006-03-22 00:08:18 -08004085
Mel Gormana5f5f912016-07-28 15:46:32 -07004086int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08004087{
David Rientjesd773ed62007-10-16 23:26:01 -07004088 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004089
4090 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004091 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07004092 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07004093 *
Christoph Lameter96146342006-07-03 00:24:13 -07004094 * A small portion of unmapped file backed pages is needed for
4095 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07004096 * thrown out if the node is overallocated. So we do not reclaim
4097 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07004098 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08004099 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004100 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
Johannes Weiner385386c2017-07-06 15:40:43 -07004101 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
Mel Gormana5f5f912016-07-28 15:46:32 -07004102 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08004103
4104 /*
David Rientjesd773ed62007-10-16 23:26:01 -07004105 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08004106 */
Mel Gormand0164ad2015-11-06 16:28:21 -08004107 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07004108 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08004109
4110 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004111 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08004112 * have associated processors. This will favor the local processor
4113 * over remote processors and spread off node memory allocations
4114 * as wide as possible.
4115 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004116 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4117 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07004118
Mel Gormana5f5f912016-07-28 15:46:32 -07004119 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4120 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07004121
Mel Gormana5f5f912016-07-28 15:46:32 -07004122 ret = __node_reclaim(pgdat, gfp_mask, order);
4123 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07004124
Mel Gorman24cf725182009-06-16 15:33:23 -07004125 if (!ret)
4126 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4127
David Rientjesd773ed62007-10-16 23:26:01 -07004128 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004129}
Christoph Lameter9eeff232006-01-18 17:42:31 -08004130#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004131
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004132/*
4133 * page_evictable - test whether a page is evictable
4134 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004135 *
4136 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07004137 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004138 *
4139 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07004140 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07004141 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07004142 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004143 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07004144int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004145{
Huang Yinge92bb4d2018-04-05 16:23:20 -07004146 int ret;
4147
4148 /* Prevent address_space of inode and swap cache from being freed */
4149 rcu_read_lock();
4150 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4151 rcu_read_unlock();
4152 return ret;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004153}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004154
Hugh Dickins850465792012-01-20 14:34:19 -08004155#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004156/**
Hugh Dickins24513262012-01-20 14:34:21 -08004157 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
4158 * @pages: array of pages to check
4159 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004160 *
Hugh Dickins24513262012-01-20 14:34:21 -08004161 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08004162 *
4163 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004164 */
Hugh Dickins24513262012-01-20 14:34:21 -08004165void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004166{
Johannes Weiner925b7672012-01-12 17:18:15 -08004167 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07004168 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08004169 int pgscanned = 0;
4170 int pgrescued = 0;
4171 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004172
Hugh Dickins24513262012-01-20 14:34:21 -08004173 for (i = 0; i < nr_pages; i++) {
4174 struct page *page = pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07004175 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07004176
Hugh Dickins24513262012-01-20 14:34:21 -08004177 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07004178 if (pagepgdat != pgdat) {
4179 if (pgdat)
4180 spin_unlock_irq(&pgdat->lru_lock);
4181 pgdat = pagepgdat;
4182 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004183 }
Mel Gorman785b99f2016-07-28 15:47:23 -07004184 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004185
Hugh Dickins24513262012-01-20 14:34:21 -08004186 if (!PageLRU(page) || !PageUnevictable(page))
4187 continue;
4188
Hugh Dickins39b5f292012-10-08 16:33:18 -07004189 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08004190 enum lru_list lru = page_lru_base_type(page);
4191
Sasha Levin309381fea2014-01-23 15:52:54 -08004192 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08004193 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07004194 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4195 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08004196 pgrescued++;
4197 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004198 }
Hugh Dickins24513262012-01-20 14:34:21 -08004199
Mel Gorman785b99f2016-07-28 15:47:23 -07004200 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08004201 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4202 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07004203 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08004204 }
Hugh Dickins850465792012-01-20 14:34:19 -08004205}
4206#endif /* CONFIG_SHMEM */