blob: 277a36de11c17d0fff294b5852d8419593319f5e [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070015#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/mm.h>
Ingo Molnar5b3cc152017-02-02 20:43:54 +010018#include <linux/sched/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070026#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070027#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080039#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/notifier.h>
41#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080042#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070043#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080044#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080045#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070046#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070047#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070048#include <linux/oom.h>
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +000049#include <linux/pagevec.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070050#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070051#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080052#include <linux/dax.h>
Johannes Weinereb414682018-10-26 15:06:27 -070053#include <linux/psi.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070059#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Nick Piggin0f8053a2006-03-22 00:08:33 -080061#include "internal.h"
62
Mel Gorman33906bc2010-08-09 17:19:16 -070063#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080067 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
Johannes Weineree814fe2014-08-06 16:06:19 -070070 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070075
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070076 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080077 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080081
Johannes Weiner1276ad62017-02-24 14:56:11 -080082 /* Writepage batching in laptop mode; RECLAIM_WRITE */
Johannes Weineree814fe2014-08-06 16:06:19 -070083 unsigned int may_writepage:1;
84
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
87
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
90
Mel Gorman1c308442018-12-28 00:35:52 -080091 /* e.g. boosted watermark reclaim leaves slabs alone */
92 unsigned int may_shrinkslab:1;
93
Yisheng Xied6622f62017-05-03 14:53:57 -070094 /*
95 * Cgroups are not reclaimed below their configured memory.low,
96 * unless we threaten to OOM. If any cgroups are skipped due to
97 * memory.low and nothing was reclaimed, go back for memory.low.
98 */
99 unsigned int memcg_low_reclaim:1;
100 unsigned int memcg_low_skipped:1;
Johannes Weiner241994ed2015-02-11 15:26:06 -0800101
Johannes Weineree814fe2014-08-06 16:06:19 -0700102 unsigned int hibernation_mode:1;
103
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready:1;
106
Greg Thelenbb451fd2018-08-17 15:45:19 -0700107 /* Allocation order */
108 s8 order;
109
110 /* Scan (total_size >> priority) pages at once */
111 s8 priority;
112
113 /* The highest zone to isolate pages for reclaim from */
114 s8 reclaim_idx;
115
116 /* This context's GFP mask */
117 gfp_t gfp_mask;
118
Johannes Weineree814fe2014-08-06 16:06:19 -0700119 /* Incremented by the number of inactive pages that were scanned */
120 unsigned long nr_scanned;
121
122 /* Number of pages freed so far during a call to shrink_zones() */
123 unsigned long nr_reclaimed;
Andrey Ryabinind108c772018-04-10 16:27:59 -0700124
125 struct {
126 unsigned int dirty;
127 unsigned int unqueued_dirty;
128 unsigned int congested;
129 unsigned int writeback;
130 unsigned int immediate;
131 unsigned int file_taken;
132 unsigned int taken;
133 } nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134};
135
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136#ifdef ARCH_HAS_PREFETCH
137#define prefetch_prev_lru_page(_page, _base, _field) \
138 do { \
139 if ((_page)->lru.prev != _base) { \
140 struct page *prev; \
141 \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
144 } \
145 } while (0)
146#else
147#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148#endif
149
150#ifdef ARCH_HAS_PREFETCHW
151#define prefetchw_prev_lru_page(_page, _base, _field) \
152 do { \
153 if ((_page)->lru.prev != _base) { \
154 struct page *prev; \
155 \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
158 } \
159 } while (0)
160#else
161#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162#endif
163
164/*
165 * From 0 .. 100. Higher means more swappy.
166 */
167int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700168/*
169 * The total number of pages which are beyond the high watermark within all
170 * zones.
171 */
172unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173
174static LIST_HEAD(shrinker_list);
175static DECLARE_RWSEM(shrinker_rwsem);
176
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700177#ifdef CONFIG_MEMCG_KMEM
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700178
179/*
180 * We allow subsystems to populate their shrinker-related
181 * LRU lists before register_shrinker_prepared() is called
182 * for the shrinker, since we don't want to impose
183 * restrictions on their internal registration order.
184 * In this case shrink_slab_memcg() may find corresponding
185 * bit is set in the shrinkers map.
186 *
187 * This value is used by the function to detect registering
188 * shrinkers and to skip do_shrink_slab() calls for them.
189 */
190#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
191
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700192static DEFINE_IDR(shrinker_idr);
193static int shrinker_nr_max;
194
195static int prealloc_memcg_shrinker(struct shrinker *shrinker)
196{
197 int id, ret = -ENOMEM;
198
199 down_write(&shrinker_rwsem);
200 /* This may call shrinker, so it must use down_read_trylock() */
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700201 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700202 if (id < 0)
203 goto unlock;
204
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700205 if (id >= shrinker_nr_max) {
206 if (memcg_expand_shrinker_maps(id)) {
207 idr_remove(&shrinker_idr, id);
208 goto unlock;
209 }
210
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700211 shrinker_nr_max = id + 1;
Kirill Tkhai0a4465d2018-08-17 15:47:37 -0700212 }
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700213 shrinker->id = id;
214 ret = 0;
215unlock:
216 up_write(&shrinker_rwsem);
217 return ret;
218}
219
220static void unregister_memcg_shrinker(struct shrinker *shrinker)
221{
222 int id = shrinker->id;
223
224 BUG_ON(id < 0);
225
226 down_write(&shrinker_rwsem);
227 idr_remove(&shrinker_idr, id);
228 up_write(&shrinker_rwsem);
229}
230#else /* CONFIG_MEMCG_KMEM */
231static int prealloc_memcg_shrinker(struct shrinker *shrinker)
232{
233 return 0;
234}
235
236static void unregister_memcg_shrinker(struct shrinker *shrinker)
237{
238}
239#endif /* CONFIG_MEMCG_KMEM */
240
Andrew Mortonc255a452012-07-31 16:43:02 -0700241#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800242static bool global_reclaim(struct scan_control *sc)
243{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800244 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800245}
Tejun Heo97c93412015-05-22 18:23:36 -0400246
247/**
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
250 *
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
256 *
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
259 */
260static bool sane_reclaim(struct scan_control *sc)
261{
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
263
264 if (!memcg)
265 return true;
266#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400268 return true;
269#endif
270 return false;
271}
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -0700272
273static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
275 bool congested)
276{
277 struct mem_cgroup_per_node *mn;
278
279 if (!memcg)
280 return;
281
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
284}
285
286static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
288{
289 struct mem_cgroup_per_node *mn;
290
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
293
294}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800295#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800296static bool global_reclaim(struct scan_control *sc)
297{
298 return true;
299}
Tejun Heo97c93412015-05-22 18:23:36 -0400300
301static bool sane_reclaim(struct scan_control *sc)
302{
303 return true;
304}
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -0700305
306static inline void set_memcg_congestion(struct pglist_data *pgdat,
307 struct mem_cgroup *memcg, bool congested)
308{
309}
310
311static inline bool memcg_congested(struct pglist_data *pgdat,
312 struct mem_cgroup *memcg)
313{
314 return false;
315
316}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800317#endif
318
Mel Gorman5a1c84b2016-07-28 15:47:31 -0700319/*
320 * This misses isolated pages which are not accounted for to save counters.
321 * As the data only determines if reclaim or compaction continues, it is
322 * not expected that isolated pages will be a dominating factor.
323 */
324unsigned long zone_reclaimable_pages(struct zone *zone)
325{
326 unsigned long nr;
327
328 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
329 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
330 if (get_nr_swap_pages() > 0)
331 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
332 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
333
334 return nr;
335}
336
Michal Hockofd538802017-02-22 15:45:58 -0800337/**
338 * lruvec_lru_size - Returns the number of pages on the given LRU list.
339 * @lruvec: lru vector
340 * @lru: lru to use
341 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
342 */
343unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800344{
Michal Hockofd538802017-02-22 15:45:58 -0800345 unsigned long lru_size;
346 int zid;
347
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700348 if (!mem_cgroup_disabled())
Johannes Weiner205b20c2019-05-14 15:47:06 -0700349 lru_size = lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
Michal Hockofd538802017-02-22 15:45:58 -0800350 else
351 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800352
Michal Hockofd538802017-02-22 15:45:58 -0800353 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
354 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
355 unsigned long size;
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800356
Michal Hockofd538802017-02-22 15:45:58 -0800357 if (!managed_zone(zone))
358 continue;
Michal Hockob4536f0c82017-01-10 16:58:04 -0800359
Michal Hockofd538802017-02-22 15:45:58 -0800360 if (!mem_cgroup_disabled())
361 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
362 else
363 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
364 NR_ZONE_LRU_BASE + lru);
365 lru_size -= min(size, lru_size);
366 }
367
368 return lru_size;
369
Michal Hockob4536f0c82017-01-10 16:58:04 -0800370}
371
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000373 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 */
Tetsuo Handa8e049442018-04-04 19:53:07 +0900375int prealloc_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376{
Alexey Dobriyanb9726c22019-03-05 15:48:26 -0800377 unsigned int size = sizeof(*shrinker->nr_deferred);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000378
Glauber Costa1d3d4432013-08-28 10:18:04 +1000379 if (shrinker->flags & SHRINKER_NUMA_AWARE)
380 size *= nr_node_ids;
381
382 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
383 if (!shrinker->nr_deferred)
384 return -ENOMEM;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700385
386 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
387 if (prealloc_memcg_shrinker(shrinker))
388 goto free_deferred;
389 }
390
Tetsuo Handa8e049442018-04-04 19:53:07 +0900391 return 0;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700392
393free_deferred:
394 kfree(shrinker->nr_deferred);
395 shrinker->nr_deferred = NULL;
396 return -ENOMEM;
Tetsuo Handa8e049442018-04-04 19:53:07 +0900397}
Glauber Costa1d3d4432013-08-28 10:18:04 +1000398
Tetsuo Handa8e049442018-04-04 19:53:07 +0900399void free_prealloced_shrinker(struct shrinker *shrinker)
400{
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700401 if (!shrinker->nr_deferred)
402 return;
403
404 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
405 unregister_memcg_shrinker(shrinker);
406
Tetsuo Handa8e049442018-04-04 19:53:07 +0900407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
409}
410
411void register_shrinker_prepared(struct shrinker *shrinker)
412{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700413 down_write(&shrinker_rwsem);
414 list_add_tail(&shrinker->list, &shrinker_list);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700415#ifdef CONFIG_MEMCG_KMEM
Kirill Tkhai8df4a442018-08-21 21:51:49 -0700416 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
417 idr_replace(&shrinker_idr, shrinker, shrinker->id);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700418#endif
Rusty Russell8e1f9362007-07-17 04:03:17 -0700419 up_write(&shrinker_rwsem);
Tetsuo Handa8e049442018-04-04 19:53:07 +0900420}
421
422int register_shrinker(struct shrinker *shrinker)
423{
424 int err = prealloc_shrinker(shrinker);
425
426 if (err)
427 return err;
428 register_shrinker_prepared(shrinker);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000429 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700431EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
433/*
434 * Remove one
435 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700436void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Tetsuo Handabb422a72017-12-18 20:31:41 +0900438 if (!shrinker->nr_deferred)
439 return;
Kirill Tkhaib4c2b232018-08-17 15:47:29 -0700440 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
441 unregister_memcg_shrinker(shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 down_write(&shrinker_rwsem);
443 list_del(&shrinker->list);
444 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700445 kfree(shrinker->nr_deferred);
Tetsuo Handabb422a72017-12-18 20:31:41 +0900446 shrinker->nr_deferred = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700448EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
450#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000451
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800452static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
Josef Bacik9092c712018-01-31 16:16:26 -0800453 struct shrinker *shrinker, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000455 unsigned long freed = 0;
456 unsigned long long delta;
457 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700458 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000459 long nr;
460 long new_nr;
461 int nid = shrinkctl->nid;
462 long batch_size = shrinker->batch ? shrinker->batch
463 : SHRINK_BATCH;
Shaohua Li5f33a082016-12-12 16:41:50 -0800464 long scanned = 0, next_deferred;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000465
Kirill Tkhaiac7fb3a2018-08-17 15:48:30 -0700466 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
467 nid = 0;
468
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700469 freeable = shrinker->count_objects(shrinker, shrinkctl);
Kirill Tkhai9b996462018-08-17 15:48:21 -0700470 if (freeable == 0 || freeable == SHRINK_EMPTY)
471 return freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000472
473 /*
474 * copy the current shrinker scan count into a local variable
475 * and zero it so that other concurrent shrinker invocations
476 * don't also do this scanning work.
477 */
478 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
479
480 total_scan = nr;
Johannes Weiner4b85afb2018-10-26 15:06:42 -0700481 if (shrinker->seeks) {
482 delta = freeable >> priority;
483 delta *= 4;
484 do_div(delta, shrinker->seeks);
485 } else {
486 /*
487 * These objects don't require any IO to create. Trim
488 * them aggressively under memory pressure to keep
489 * them from causing refetches in the IO caches.
490 */
491 delta = freeable / 2;
492 }
Roman Gushchin172b06c32018-09-20 12:22:46 -0700493
Glauber Costa1d3d4432013-08-28 10:18:04 +1000494 total_scan += delta;
495 if (total_scan < 0) {
Sakari Ailusd75f7732019-03-25 21:32:28 +0200496 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000497 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700498 total_scan = freeable;
Shaohua Li5f33a082016-12-12 16:41:50 -0800499 next_deferred = nr;
500 } else
501 next_deferred = total_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000502
503 /*
504 * We need to avoid excessive windup on filesystem shrinkers
505 * due to large numbers of GFP_NOFS allocations causing the
506 * shrinkers to return -1 all the time. This results in a large
507 * nr being built up so when a shrink that can do some work
508 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700509 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000510 * memory.
511 *
512 * Hence only allow the shrinker to scan the entire cache when
513 * a large delta change is calculated directly.
514 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700515 if (delta < freeable / 4)
516 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000517
518 /*
519 * Avoid risking looping forever due to too large nr value:
520 * never try to free more than twice the estimate number of
521 * freeable entries.
522 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700523 if (total_scan > freeable * 2)
524 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000525
526 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Josef Bacik9092c712018-01-31 16:16:26 -0800527 freeable, delta, total_scan, priority);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000528
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800529 /*
530 * Normally, we should not scan less than batch_size objects in one
531 * pass to avoid too frequent shrinker calls, but if the slab has less
532 * than batch_size objects in total and we are really tight on memory,
533 * we will try to reclaim all available objects, otherwise we can end
534 * up failing allocations although there are plenty of reclaimable
535 * objects spread over several slabs with usage less than the
536 * batch_size.
537 *
538 * We detect the "tight on memory" situations by looking at the total
539 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700540 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800541 * scanning at high prio and therefore should try to reclaim as much as
542 * possible.
543 */
544 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700545 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000546 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800547 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000548
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800549 shrinkctl->nr_to_scan = nr_to_scan;
Chris Wilsond460acb2017-09-06 16:19:26 -0700550 shrinkctl->nr_scanned = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000551 ret = shrinker->scan_objects(shrinker, shrinkctl);
552 if (ret == SHRINK_STOP)
553 break;
554 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000555
Chris Wilsond460acb2017-09-06 16:19:26 -0700556 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
557 total_scan -= shrinkctl->nr_scanned;
558 scanned += shrinkctl->nr_scanned;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000559
560 cond_resched();
561 }
562
Shaohua Li5f33a082016-12-12 16:41:50 -0800563 if (next_deferred >= scanned)
564 next_deferred -= scanned;
565 else
566 next_deferred = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000567 /*
568 * move the unused scan count back into the shrinker in a
569 * manner that handles concurrent updates. If we exhausted the
570 * scan, there is no need to do an update.
571 */
Shaohua Li5f33a082016-12-12 16:41:50 -0800572 if (next_deferred > 0)
573 new_nr = atomic_long_add_return(next_deferred,
Glauber Costa1d3d4432013-08-28 10:18:04 +1000574 &shrinker->nr_deferred[nid]);
575 else
576 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
577
Dave Hansendf9024a2014-06-04 16:08:07 -0700578 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000579 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580}
581
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700582#ifdef CONFIG_MEMCG_KMEM
583static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
584 struct mem_cgroup *memcg, int priority)
585{
586 struct memcg_shrinker_map *map;
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700587 unsigned long ret, freed = 0;
588 int i;
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700589
590 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
591 return 0;
592
593 if (!down_read_trylock(&shrinker_rwsem))
594 return 0;
595
596 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
597 true);
598 if (unlikely(!map))
599 goto unlock;
600
601 for_each_set_bit(i, map->map, shrinker_nr_max) {
602 struct shrink_control sc = {
603 .gfp_mask = gfp_mask,
604 .nid = nid,
605 .memcg = memcg,
606 };
607 struct shrinker *shrinker;
608
609 shrinker = idr_find(&shrinker_idr, i);
Kirill Tkhai7e010df2018-08-17 15:48:34 -0700610 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
611 if (!shrinker)
612 clear_bit(i, map->map);
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700613 continue;
614 }
615
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700616 ret = do_shrink_slab(&sc, shrinker, priority);
Kirill Tkhaif90280d2018-08-17 15:48:25 -0700617 if (ret == SHRINK_EMPTY) {
618 clear_bit(i, map->map);
619 /*
620 * After the shrinker reported that it had no objects to
621 * free, but before we cleared the corresponding bit in
622 * the memcg shrinker map, a new object might have been
623 * added. To make sure, we have the bit set in this
624 * case, we invoke the shrinker one more time and reset
625 * the bit if it reports that it is not empty anymore.
626 * The memory barrier here pairs with the barrier in
627 * memcg_set_shrinker_bit():
628 *
629 * list_lru_add() shrink_slab_memcg()
630 * list_add_tail() clear_bit()
631 * <MB> <MB>
632 * set_bit() do_shrink_slab()
633 */
634 smp_mb__after_atomic();
635 ret = do_shrink_slab(&sc, shrinker, priority);
636 if (ret == SHRINK_EMPTY)
637 ret = 0;
638 else
639 memcg_set_shrinker_bit(memcg, nid, i);
640 }
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700641 freed += ret;
642
643 if (rwsem_is_contended(&shrinker_rwsem)) {
644 freed = freed ? : 1;
645 break;
646 }
647 }
648unlock:
649 up_read(&shrinker_rwsem);
650 return freed;
651}
652#else /* CONFIG_MEMCG_KMEM */
653static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
654 struct mem_cgroup *memcg, int priority)
655{
656 return 0;
657}
658#endif /* CONFIG_MEMCG_KMEM */
659
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800660/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800661 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800662 * @gfp_mask: allocation context
663 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800664 * @memcg: memory cgroup whose slab caches to target
Josef Bacik9092c712018-01-31 16:16:26 -0800665 * @priority: the reclaim priority
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800667 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800669 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
670 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 *
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700672 * @memcg specifies the memory cgroup to target. Unaware shrinkers
673 * are called only if it is the root cgroup.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800674 *
Josef Bacik9092c712018-01-31 16:16:26 -0800675 * @priority is sc->priority, we take the number of objects and >> by priority
676 * in order to get the scan target.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800678 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800680static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
681 struct mem_cgroup *memcg,
Josef Bacik9092c712018-01-31 16:16:26 -0800682 int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683{
Kirill Tkhaib8e57ef2018-10-05 15:52:10 -0700684 unsigned long ret, freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 struct shrinker *shrinker;
686
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700687 if (!mem_cgroup_is_root(memcg))
Kirill Tkhaib0dedc42018-08-17 15:48:14 -0700688 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800689
Tetsuo Handae830c632018-04-05 16:23:35 -0700690 if (!down_read_trylock(&shrinker_rwsem))
Minchan Kimf06590b2011-05-24 17:11:11 -0700691 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
693 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800694 struct shrink_control sc = {
695 .gfp_mask = gfp_mask,
696 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800697 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800698 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800699
Kirill Tkhai9b996462018-08-17 15:48:21 -0700700 ret = do_shrink_slab(&sc, shrinker, priority);
701 if (ret == SHRINK_EMPTY)
702 ret = 0;
703 freed += ret;
Minchan Kime4966122018-01-31 16:16:55 -0800704 /*
705 * Bail out if someone want to register a new shrinker to
706 * prevent the regsitration from being stalled for long periods
707 * by parallel ongoing shrinking.
708 */
709 if (rwsem_is_contended(&shrinker_rwsem)) {
710 freed = freed ? : 1;
711 break;
712 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800714
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700716out:
717 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000718 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719}
720
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800721void drop_slab_node(int nid)
722{
723 unsigned long freed;
724
725 do {
726 struct mem_cgroup *memcg = NULL;
727
728 freed = 0;
Vladimir Davydovaeed1d322018-08-17 15:48:17 -0700729 memcg = mem_cgroup_iter(NULL, NULL, NULL);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800730 do {
Josef Bacik9092c712018-01-31 16:16:26 -0800731 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800732 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
733 } while (freed > 10);
734}
735
736void drop_slab(void)
737{
738 int nid;
739
740 for_each_online_node(nid)
741 drop_slab_node(nid);
742}
743
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744static inline int is_page_cache_freeable(struct page *page)
745{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700746 /*
747 * A freeable page cache page is referenced only by the caller
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400748 * that isolated the page, the page cache and optional buffer
749 * heads at page->private.
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700750 */
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400751 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
Huang Yingbd4c82c22017-09-06 16:22:49 -0700752 HPAGE_PMD_NR : 1;
Matthew Wilcox67891ff2018-06-10 07:34:39 -0400753 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754}
755
Tejun Heo703c2702015-05-22 17:13:44 -0400756static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
Christoph Lameter930d9152006-01-08 01:00:47 -0800758 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400760 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400762 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 return 1;
764 return 0;
765}
766
767/*
768 * We detected a synchronous write error writing a page out. Probably
769 * -ENOSPC. We need to propagate that into the address_space for a subsequent
770 * fsync(), msync() or close().
771 *
772 * The tricky part is that after writepage we cannot touch the mapping: nothing
773 * prevents it from being freed up. But we have a ref on the page and once
774 * that page is locked, the mapping is pinned.
775 *
776 * We're allowed to run sleeping lock_page() here because we know the caller has
777 * __GFP_FS.
778 */
779static void handle_write_error(struct address_space *mapping,
780 struct page *page, int error)
781{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100782 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700783 if (page_mapping(page) == mapping)
784 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 unlock_page(page);
786}
787
Christoph Lameter04e62a22006-06-23 02:03:38 -0700788/* possible outcome of pageout() */
789typedef enum {
790 /* failed to write page out, page is locked */
791 PAGE_KEEP,
792 /* move page to the active list, page is locked */
793 PAGE_ACTIVATE,
794 /* page has been sent to the disk successfully, page is unlocked */
795 PAGE_SUCCESS,
796 /* page is clean and locked */
797 PAGE_CLEAN,
798} pageout_t;
799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800/*
Andrew Morton1742f192006-03-22 00:08:21 -0800801 * pageout is called by shrink_page_list() for each dirty page.
802 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700804static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700805 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806{
807 /*
808 * If the page is dirty, only perform writeback if that write
809 * will be non-blocking. To prevent this allocation from being
810 * stalled by pagecache activity. But note that there may be
811 * stalls if we need to run get_block(). We could test
812 * PagePrivate for that.
813 *
Al Viro81742022014-04-03 03:17:43 -0400814 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 * this page's queue, we can perform writeback even if that
816 * will block.
817 *
818 * If the page is swapcache, write it back even if that would
819 * block, for some throttling. This happens by accident, because
820 * swap_backing_dev_info is bust: it doesn't reflect the
821 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 */
823 if (!is_page_cache_freeable(page))
824 return PAGE_KEEP;
825 if (!mapping) {
826 /*
827 * Some data journaling orphaned pages can have
828 * page->mapping == NULL while being dirty with clean buffers.
829 */
David Howells266cf652009-04-03 16:42:36 +0100830 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 if (try_to_free_buffers(page)) {
832 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700833 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 return PAGE_CLEAN;
835 }
836 }
837 return PAGE_KEEP;
838 }
839 if (mapping->a_ops->writepage == NULL)
840 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400841 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 return PAGE_KEEP;
843
844 if (clear_page_dirty_for_io(page)) {
845 int res;
846 struct writeback_control wbc = {
847 .sync_mode = WB_SYNC_NONE,
848 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700849 .range_start = 0,
850 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 .for_reclaim = 1,
852 };
853
854 SetPageReclaim(page);
855 res = mapping->a_ops->writepage(page, &wbc);
856 if (res < 0)
857 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800858 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 ClearPageReclaim(page);
860 return PAGE_ACTIVATE;
861 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700862
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 if (!PageWriteback(page)) {
864 /* synchronous write or broken a_ops? */
865 ClearPageReclaim(page);
866 }
yalin wang3aa23852016-01-14 15:18:30 -0800867 trace_mm_vmscan_writepage(page);
Mel Gormanc4a25632016-07-28 15:46:23 -0700868 inc_node_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 return PAGE_SUCCESS;
870 }
871
872 return PAGE_CLEAN;
873}
874
Andrew Mortona649fd92006-10-17 00:09:36 -0700875/*
Nick Piggine2867812008-07-25 19:45:30 -0700876 * Same as remove_mapping, but if the page is removed from the mapping, it
877 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700878 */
Johannes Weinera5289102014-04-03 14:47:51 -0700879static int __remove_mapping(struct address_space *mapping, struct page *page,
880 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800881{
Greg Thelenc4843a72015-05-22 17:13:16 -0400882 unsigned long flags;
Huang Yingbd4c82c22017-09-06 16:22:49 -0700883 int refcount;
Greg Thelenc4843a72015-05-22 17:13:16 -0400884
Nick Piggin28e4d962006-09-25 23:31:23 -0700885 BUG_ON(!PageLocked(page));
886 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800887
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700888 xa_lock_irqsave(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800889 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700890 * The non racy check for a busy page.
891 *
892 * Must be careful with the order of the tests. When someone has
893 * a ref to the page, it may be possible that they dirty it then
894 * drop the reference. So if PageDirty is tested before page_count
895 * here, then the following race may occur:
896 *
897 * get_user_pages(&page);
898 * [user mapping goes away]
899 * write_to(page);
900 * !PageDirty(page) [good]
901 * SetPageDirty(page);
902 * put_page(page);
903 * !page_count(page) [good, discard it]
904 *
905 * [oops, our write_to data is lost]
906 *
907 * Reversing the order of the tests ensures such a situation cannot
908 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
Joonsoo Kim0139aa72016-05-19 17:10:49 -0700909 * load is not satisfied before that of page->_refcount.
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700910 *
911 * Note that if SetPageDirty is always performed via set_page_dirty,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700912 * and thus under the i_pages lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800913 */
Huang Yingbd4c82c22017-09-06 16:22:49 -0700914 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
915 refcount = 1 + HPAGE_PMD_NR;
916 else
917 refcount = 2;
918 if (!page_ref_freeze(page, refcount))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800919 goto cannot_free;
Jiang Biao1c4c3b92018-08-21 21:53:13 -0700920 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
Nick Piggine2867812008-07-25 19:45:30 -0700921 if (unlikely(PageDirty(page))) {
Huang Yingbd4c82c22017-09-06 16:22:49 -0700922 page_ref_unfreeze(page, refcount);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800923 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700924 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800925
926 if (PageSwapCache(page)) {
927 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700928 mem_cgroup_swapout(page, swap);
Matthew Wilcox4e17ec22017-11-29 08:32:39 -0500929 __delete_from_swap_cache(page, swap);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700930 xa_unlock_irqrestore(&mapping->i_pages, flags);
Minchan Kim75f6d6d2017-07-06 15:37:21 -0700931 put_swap_page(page, swap);
Nick Piggine2867812008-07-25 19:45:30 -0700932 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500933 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700934 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500935
936 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700937 /*
938 * Remember a shadow entry for reclaimed file cache in
939 * order to detect refaults, thus thrashing, later on.
940 *
941 * But don't store shadows in an address space that is
942 * already exiting. This is not just an optizimation,
943 * inode reclaim needs to empty out the radix tree or
944 * the nodes are lost. Don't plant shadows behind its
945 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800946 *
947 * We also don't store shadows for DAX mappings because the
948 * only page cache pages found in these are zero pages
949 * covering holes, and because we don't want to mix DAX
950 * exceptional entries and shadow exceptional entries in the
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700951 * same address_space.
Johannes Weinera5289102014-04-03 14:47:51 -0700952 */
953 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800954 !mapping_exiting(mapping) && !dax_mapping(mapping))
Andrey Ryabinina7ca12f2019-03-05 15:49:35 -0800955 shadow = workingset_eviction(page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700956 __delete_from_page_cache(page, shadow);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700957 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500958
959 if (freepage != NULL)
960 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800961 }
962
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800963 return 1;
964
965cannot_free:
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700966 xa_unlock_irqrestore(&mapping->i_pages, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800967 return 0;
968}
969
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970/*
Nick Piggine2867812008-07-25 19:45:30 -0700971 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
972 * someone else has a ref on the page, abort and return 0. If it was
973 * successfully detached, return 1. Assumes the caller has a single ref on
974 * this page.
975 */
976int remove_mapping(struct address_space *mapping, struct page *page)
977{
Johannes Weinera5289102014-04-03 14:47:51 -0700978 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700979 /*
980 * Unfreezing the refcount with 1 rather than 2 effectively
981 * drops the pagecache ref for us without requiring another
982 * atomic operation.
983 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700984 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700985 return 1;
986 }
987 return 0;
988}
989
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700990/**
991 * putback_lru_page - put previously isolated page onto appropriate LRU list
992 * @page: page to be put back to appropriate lru list
993 *
994 * Add previously isolated @page to appropriate LRU list.
995 * Page may still be unevictable for other reasons.
996 *
997 * lru_lock must not be held, interrupts must be enabled.
998 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700999void putback_lru_page(struct page *page)
1000{
Shakeel Butt9c4e6b12018-02-21 14:45:28 -08001001 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001002 put_page(page); /* drop ref from isolate */
1003}
1004
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001005enum page_references {
1006 PAGEREF_RECLAIM,
1007 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -08001008 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001009 PAGEREF_ACTIVATE,
1010};
1011
1012static enum page_references page_check_references(struct page *page,
1013 struct scan_control *sc)
1014{
Johannes Weiner645747462010-03-05 13:42:22 -08001015 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001016 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001017
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001018 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1019 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -08001020 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001021
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001022 /*
1023 * Mlock lost the isolation race with us. Let try_to_unmap()
1024 * move the page to the unevictable list.
1025 */
1026 if (vm_flags & VM_LOCKED)
1027 return PAGEREF_RECLAIM;
1028
Johannes Weiner645747462010-03-05 13:42:22 -08001029 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -07001030 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001031 return PAGEREF_ACTIVATE;
1032 /*
1033 * All mapped pages start out with page table
1034 * references from the instantiating fault, so we need
1035 * to look twice if a mapped file page is used more
1036 * than once.
1037 *
1038 * Mark it and spare it for another trip around the
1039 * inactive list. Another page table reference will
1040 * lead to its activation.
1041 *
1042 * Note: the mark is set for activated pages as well
1043 * so that recently deactivated but used pages are
1044 * quickly recovered.
1045 */
1046 SetPageReferenced(page);
1047
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -08001048 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -08001049 return PAGEREF_ACTIVATE;
1050
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -08001051 /*
1052 * Activate file-backed executable pages after first usage.
1053 */
1054 if (vm_flags & VM_EXEC)
1055 return PAGEREF_ACTIVATE;
1056
Johannes Weiner645747462010-03-05 13:42:22 -08001057 return PAGEREF_KEEP;
1058 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001059
1060 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -07001061 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -08001062 return PAGEREF_RECLAIM_CLEAN;
1063
1064 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001065}
1066
Mel Gormane2be15f2013-07-03 15:01:57 -07001067/* Check if a page is dirty or under writeback */
1068static void page_check_dirty_writeback(struct page *page,
1069 bool *dirty, bool *writeback)
1070{
Mel Gormanb4597222013-07-03 15:02:05 -07001071 struct address_space *mapping;
1072
Mel Gormane2be15f2013-07-03 15:01:57 -07001073 /*
1074 * Anonymous pages are not handled by flushers and must be written
1075 * from reclaim context. Do not stall reclaim based on them
1076 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001077 if (!page_is_file_cache(page) ||
1078 (PageAnon(page) && !PageSwapBacked(page))) {
Mel Gormane2be15f2013-07-03 15:01:57 -07001079 *dirty = false;
1080 *writeback = false;
1081 return;
1082 }
1083
1084 /* By default assume that the page flags are accurate */
1085 *dirty = PageDirty(page);
1086 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -07001087
1088 /* Verify dirty/writeback state if the filesystem supports it */
1089 if (!page_has_private(page))
1090 return;
1091
1092 mapping = page_mapping(page);
1093 if (mapping && mapping->a_ops->is_dirty_writeback)
1094 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -07001095}
1096
Nick Piggine2867812008-07-25 19:45:30 -07001097/*
Andrew Morton1742f192006-03-22 00:08:21 -08001098 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 */
Andrew Morton1742f192006-03-22 00:08:21 -08001100static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman599d0c92016-07-28 15:45:31 -07001101 struct pglist_data *pgdat,
Mel Gormanf84f6e22011-10-31 17:07:51 -07001102 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -07001103 enum ttu_flags ttu_flags,
Michal Hocko3c710c12017-02-22 15:44:27 -08001104 struct reclaim_stat *stat,
Minchan Kim02c6de82012-10-08 16:31:55 -07001105 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106{
1107 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001108 LIST_HEAD(free_pages);
Michal Hocko3c710c12017-02-22 15:44:27 -08001109 unsigned nr_reclaimed = 0;
Kirill Tkhai886cf192019-05-13 17:16:51 -07001110 unsigned pgactivate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
Kirill Tkhai060f0052019-03-05 15:48:15 -08001112 memset(stat, 0, sizeof(*stat));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 cond_resched();
1114
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 while (!list_empty(page_list)) {
1116 struct address_space *mapping;
1117 struct page *page;
1118 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -07001119 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -07001120 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
1122 cond_resched();
1123
1124 page = lru_to_page(page_list);
1125 list_del(&page->lru);
1126
Nick Piggin529ae9a2008-08-02 12:01:03 +02001127 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 goto keep;
1129
Sasha Levin309381fea2014-01-23 15:52:54 -08001130 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
1132 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -08001133
Hugh Dickins39b5f292012-10-08 16:33:18 -07001134 if (unlikely(!page_evictable(page)))
Minchan Kimad6b6702017-05-03 14:54:13 -07001135 goto activate_locked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001136
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001137 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -08001138 goto keep_locked;
1139
Andy Whitcroftc661b072007-08-22 14:01:26 -07001140 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1141 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1142
Mel Gorman283aba92013-07-03 15:01:51 -07001143 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07001144 * The number of dirty pages determines if a node is marked
Mel Gormane2be15f2013-07-03 15:01:57 -07001145 * reclaim_congested which affects wait_iff_congested. kswapd
1146 * will stall and start writing pages if the tail of the LRU
1147 * is all dirty unqueued pages.
1148 */
1149 page_check_dirty_writeback(page, &dirty, &writeback);
1150 if (dirty || writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001151 stat->nr_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001152
1153 if (dirty && !writeback)
Kirill Tkhai060f0052019-03-05 15:48:15 -08001154 stat->nr_unqueued_dirty++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001155
Mel Gormand04e8ac2013-07-03 15:02:03 -07001156 /*
1157 * Treat this page as congested if the underlying BDI is or if
1158 * pages are cycling through the LRU so quickly that the
1159 * pages marked for immediate reclaim are making it to the
1160 * end of the LRU a second time.
1161 */
Mel Gormane2be15f2013-07-03 15:01:57 -07001162 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -08001163 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -04001164 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -07001165 (writeback && PageReclaim(page)))
Kirill Tkhai060f0052019-03-05 15:48:15 -08001166 stat->nr_congested++;
Mel Gormane2be15f2013-07-03 15:01:57 -07001167
1168 /*
Mel Gorman283aba92013-07-03 15:01:51 -07001169 * If a page at the tail of the LRU is under writeback, there
1170 * are three cases to consider.
1171 *
1172 * 1) If reclaim is encountering an excessive number of pages
1173 * under writeback and this page is both under writeback and
1174 * PageReclaim then it indicates that pages are being queued
1175 * for IO but are being recycled through the LRU before the
1176 * IO can complete. Waiting on the page itself risks an
1177 * indefinite stall if it is impossible to writeback the
1178 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -07001179 * note that the LRU is being scanned too quickly and the
1180 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -07001181 *
Tejun Heo97c93412015-05-22 18:23:36 -04001182 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -07001183 * not marked for immediate reclaim, or the caller does not
1184 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1185 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -04001186 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -07001187 *
Michal Hockoecf5fc62015-08-04 14:36:58 -07001188 * Require may_enter_fs because we would wait on fs, which
1189 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -07001190 * enter reclaim, and deadlock if it waits on a page for
1191 * which it is needed to do the write (loop masks off
1192 * __GFP_IO|__GFP_FS for this reason); but more thought
1193 * would probably show more reasons.
1194 *
Hugh Dickins7fadc822015-09-08 15:03:46 -07001195 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -07001196 * PageReclaim. memcg does not have any dirty pages
1197 * throttling so we could easily OOM just because too many
1198 * pages are in writeback and there is nothing else to
1199 * reclaim. Wait for the writeback to complete.
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001200 *
1201 * In cases 1) and 2) we activate the pages to get them out of
1202 * the way while we continue scanning for clean pages on the
1203 * inactive list and refilling from the active list. The
1204 * observation here is that waiting for disk writes is more
1205 * expensive than potentially causing reloads down the line.
1206 * Since they're marked for immediate reclaim, they won't put
1207 * memory pressure on the cache working set any longer than it
1208 * takes to write them to disk.
Mel Gorman283aba92013-07-03 15:01:51 -07001209 */
Andy Whitcroftc661b072007-08-22 14:01:26 -07001210 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -07001211 /* Case 1 above */
1212 if (current_is_kswapd() &&
1213 PageReclaim(page) &&
Mel Gorman599d0c92016-07-28 15:45:31 -07001214 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
Kirill Tkhai060f0052019-03-05 15:48:15 -08001215 stat->nr_immediate++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001216 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001217
1218 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001219 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001220 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001221 /*
1222 * This is slightly racy - end_page_writeback()
1223 * might have just cleared PageReclaim, then
1224 * setting PageReclaim here end up interpreted
1225 * as PageReadahead - but that does not matter
1226 * enough to care. What we do want is for this
1227 * page to have PageReclaim set next time memcg
1228 * reclaim reaches the tests above, so it will
1229 * then wait_on_page_writeback() to avoid OOM;
1230 * and it's also appropriate in global reclaim.
1231 */
1232 SetPageReclaim(page);
Kirill Tkhai060f0052019-03-05 15:48:15 -08001233 stat->nr_writeback++;
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001234 goto activate_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001235
1236 /* Case 3 above */
1237 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001238 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001239 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001240 /* then go back and try same page again */
1241 list_add_tail(&page->lru, page_list);
1242 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001243 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
Minchan Kim02c6de82012-10-08 16:31:55 -07001246 if (!force_reclaim)
1247 references = page_check_references(page, sc);
1248
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001249 switch (references) {
1250 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001252 case PAGEREF_KEEP:
Kirill Tkhai060f0052019-03-05 15:48:15 -08001253 stat->nr_ref_keep++;
Johannes Weiner645747462010-03-05 13:42:22 -08001254 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001255 case PAGEREF_RECLAIM:
1256 case PAGEREF_RECLAIM_CLEAN:
1257 ; /* try to reclaim the page below */
1258 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 /*
1261 * Anonymous process memory has backing store?
1262 * Try to allocate it some swap space here.
Shaohua Li802a3a92017-05-03 14:52:32 -07001263 * Lazyfree page could be freed directly
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001265 if (PageAnon(page) && PageSwapBacked(page)) {
1266 if (!PageSwapCache(page)) {
1267 if (!(sc->gfp_mask & __GFP_IO))
1268 goto keep_locked;
1269 if (PageTransHuge(page)) {
1270 /* cannot split THP, skip it */
1271 if (!can_split_huge_page(page, NULL))
1272 goto activate_locked;
1273 /*
1274 * Split pages without a PMD map right
1275 * away. Chances are some or all of the
1276 * tail pages can be freed without IO.
1277 */
1278 if (!compound_mapcount(page) &&
1279 split_huge_page_to_list(page,
1280 page_list))
1281 goto activate_locked;
1282 }
1283 if (!add_to_swap(page)) {
1284 if (!PageTransHuge(page))
1285 goto activate_locked;
1286 /* Fallback to swap normal pages */
1287 if (split_huge_page_to_list(page,
1288 page_list))
1289 goto activate_locked;
Huang Yingfe490cc2017-09-06 16:22:52 -07001290#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1291 count_vm_event(THP_SWPOUT_FALLBACK);
1292#endif
Huang Yingbd4c82c22017-09-06 16:22:49 -07001293 if (!add_to_swap(page))
1294 goto activate_locked;
1295 }
Minchan Kim0f074652017-07-06 15:37:24 -07001296
Huang Yingbd4c82c22017-09-06 16:22:49 -07001297 may_enter_fs = 1;
1298
1299 /* Adding to swap updated mapping */
1300 mapping = page_mapping(page);
Minchan Kim0f074652017-07-06 15:37:24 -07001301 }
Kirill A. Shutemov7751b2d2016-07-26 15:25:56 -07001302 } else if (unlikely(PageTransHuge(page))) {
1303 /* Split file THP */
1304 if (split_huge_page_to_list(page, page_list))
1305 goto keep_locked;
Mel Gormane2be15f2013-07-03 15:01:57 -07001306 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307
1308 /*
1309 * The page is mapped into the page tables of one or more
1310 * processes. Try to unmap it here.
1311 */
Shaohua Li802a3a92017-05-03 14:52:32 -07001312 if (page_mapped(page)) {
Huang Yingbd4c82c22017-09-06 16:22:49 -07001313 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1314
1315 if (unlikely(PageTransHuge(page)))
1316 flags |= TTU_SPLIT_HUGE_PMD;
1317 if (!try_to_unmap(page, flags)) {
Kirill Tkhai060f0052019-03-05 15:48:15 -08001318 stat->nr_unmap_fail++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 }
1321 }
1322
1323 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001324 /*
Johannes Weiner4eda4822017-02-24 14:56:20 -08001325 * Only kswapd can writeback filesystem pages
1326 * to avoid risk of stack overflow. But avoid
1327 * injecting inefficient single-page IO into
1328 * flusher writeback as much as possible: only
1329 * write pages when we've encountered many
1330 * dirty pages, and when we've already scanned
1331 * the rest of the LRU for clean pages and see
1332 * the same dirty pages again (PageReclaim).
Mel Gormanee728862011-10-31 17:07:38 -07001333 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001334 if (page_is_file_cache(page) &&
Johannes Weiner4eda4822017-02-24 14:56:20 -08001335 (!current_is_kswapd() || !PageReclaim(page) ||
1336 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001337 /*
1338 * Immediately reclaim when written back.
1339 * Similar in principal to deactivate_page()
1340 * except we already have the page isolated
1341 * and know it's dirty
1342 */
Mel Gormanc4a25632016-07-28 15:46:23 -07001343 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001344 SetPageReclaim(page);
1345
Johannes Weinerc55e8d02017-02-24 14:56:23 -08001346 goto activate_locked;
Mel Gormanee728862011-10-31 17:07:38 -07001347 }
1348
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001349 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001351 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001353 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 goto keep_locked;
1355
Mel Gormand950c942015-09-04 15:47:35 -07001356 /*
1357 * Page is dirty. Flush the TLB if a writable entry
1358 * potentially exists to avoid CPU writes after IO
1359 * starts and then write it out here.
1360 */
1361 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001362 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 case PAGE_KEEP:
1364 goto keep_locked;
1365 case PAGE_ACTIVATE:
1366 goto activate_locked;
1367 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001368 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001369 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001370 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001372
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 /*
1374 * A synchronous write - probably a ramdisk. Go
1375 * ahead and try to reclaim the page.
1376 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001377 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 goto keep;
1379 if (PageDirty(page) || PageWriteback(page))
1380 goto keep_locked;
1381 mapping = page_mapping(page);
1382 case PAGE_CLEAN:
1383 ; /* try to free the page below */
1384 }
1385 }
1386
1387 /*
1388 * If the page has buffers, try to free the buffer mappings
1389 * associated with this page. If we succeed we try to free
1390 * the page as well.
1391 *
1392 * We do this even if the page is PageDirty().
1393 * try_to_release_page() does not perform I/O, but it is
1394 * possible for a page to have PageDirty set, but it is actually
1395 * clean (all its buffers are clean). This happens if the
1396 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001397 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 * try_to_release_page() will discover that cleanness and will
1399 * drop the buffers and mark the page clean - it can be freed.
1400 *
1401 * Rarely, pages can have buffers and no ->mapping. These are
1402 * the pages which were not successfully invalidated in
1403 * truncate_complete_page(). We try to drop those buffers here
1404 * and if that worked, and the page is no longer mapped into
1405 * process address space (page_count == 1) it can be freed.
1406 * Otherwise, leave the page on the LRU so it is swappable.
1407 */
David Howells266cf652009-04-03 16:42:36 +01001408 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 if (!try_to_release_page(page, sc->gfp_mask))
1410 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001411 if (!mapping && page_count(page) == 1) {
1412 unlock_page(page);
1413 if (put_page_testzero(page))
1414 goto free_it;
1415 else {
1416 /*
1417 * rare race with speculative reference.
1418 * the speculative reference will free
1419 * this page shortly, so we may
1420 * increment nr_reclaimed here (and
1421 * leave it off the LRU).
1422 */
1423 nr_reclaimed++;
1424 continue;
1425 }
1426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 }
1428
Shaohua Li802a3a92017-05-03 14:52:32 -07001429 if (PageAnon(page) && !PageSwapBacked(page)) {
1430 /* follow __remove_mapping for reference */
1431 if (!page_ref_freeze(page, 1))
1432 goto keep_locked;
1433 if (PageDirty(page)) {
1434 page_ref_unfreeze(page, 1);
1435 goto keep_locked;
1436 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437
Shaohua Li802a3a92017-05-03 14:52:32 -07001438 count_vm_event(PGLAZYFREED);
Roman Gushchin22621852017-07-06 15:40:25 -07001439 count_memcg_page_event(page, PGLAZYFREED);
Shaohua Li802a3a92017-05-03 14:52:32 -07001440 } else if (!mapping || !__remove_mapping(mapping, page, true))
1441 goto keep_locked;
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001442
1443 unlock_page(page);
Nick Piggine2867812008-07-25 19:45:30 -07001444free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001445 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001446
1447 /*
1448 * Is there need to periodically free_page_list? It would
1449 * appear not as the counts should be low
1450 */
Huang Yingbd4c82c22017-09-06 16:22:49 -07001451 if (unlikely(PageTransHuge(page))) {
1452 mem_cgroup_uncharge(page);
1453 (*get_compound_page_dtor(page))(page);
1454 } else
1455 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 continue;
1457
1458activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001459 /* Not a candidate for swapping, so reclaim swap space. */
Minchan Kimad6b6702017-05-03 14:54:13 -07001460 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1461 PageMlocked(page)))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001462 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001463 VM_BUG_ON_PAGE(PageActive(page), page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001464 if (!PageMlocked(page)) {
Kirill Tkhai886cf192019-05-13 17:16:51 -07001465 int type = page_is_file_cache(page);
Minchan Kimad6b6702017-05-03 14:54:13 -07001466 SetPageActive(page);
Kirill Tkhai886cf192019-05-13 17:16:51 -07001467 pgactivate++;
1468 stat->nr_activate[type] += hpage_nr_pages(page);
Roman Gushchin22621852017-07-06 15:40:25 -07001469 count_memcg_page_event(page, PGACTIVATE);
Minchan Kimad6b6702017-05-03 14:54:13 -07001470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471keep_locked:
1472 unlock_page(page);
1473keep:
1474 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001475 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001477
Johannes Weiner747db952014-08-08 14:19:24 -07001478 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001479 try_to_unmap_flush();
Mel Gorman2d4894b2017-11-15 17:37:59 -08001480 free_unref_page_list(&free_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001481
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 list_splice(&ret_pages, page_list);
Kirill Tkhai886cf192019-05-13 17:16:51 -07001483 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001484
Andrew Morton05ff5132006-03-22 00:08:20 -08001485 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486}
1487
Minchan Kim02c6de82012-10-08 16:31:55 -07001488unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1489 struct list_head *page_list)
1490{
1491 struct scan_control sc = {
1492 .gfp_mask = GFP_KERNEL,
1493 .priority = DEF_PRIORITY,
1494 .may_unmap = 1,
1495 };
Kirill Tkhai060f0052019-03-05 15:48:15 -08001496 struct reclaim_stat dummy_stat;
Michal Hocko3c710c12017-02-22 15:44:27 -08001497 unsigned long ret;
Minchan Kim02c6de82012-10-08 16:31:55 -07001498 struct page *page, *next;
1499 LIST_HEAD(clean_pages);
1500
1501 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001502 if (page_is_file_cache(page) && !PageDirty(page) &&
Minchan Kima58f2ce2019-06-13 15:56:15 -07001503 !__PageMovable(page) && !PageUnevictable(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001504 ClearPageActive(page);
1505 list_move(&page->lru, &clean_pages);
1506 }
1507 }
1508
Mel Gorman599d0c92016-07-28 15:45:31 -07001509 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
Kirill Tkhai060f0052019-03-05 15:48:15 -08001510 TTU_IGNORE_ACCESS, &dummy_stat, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001511 list_splice(&clean_pages, page_list);
Mel Gorman599d0c92016-07-28 15:45:31 -07001512 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001513 return ret;
1514}
1515
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001516/*
1517 * Attempt to remove the specified page from its LRU. Only take this page
1518 * if it is of the appropriate PageActive status. Pages which are being
1519 * freed elsewhere are also ignored.
1520 *
1521 * page: page to consider
1522 * mode: one of the LRU isolation modes defined above
1523 *
1524 * returns 0 on success, -ve errno on failure.
1525 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001526int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001527{
1528 int ret = -EINVAL;
1529
1530 /* Only take pages on the LRU. */
1531 if (!PageLRU(page))
1532 return ret;
1533
Minchan Kime46a2872012-10-08 16:33:48 -07001534 /* Compaction should not handle unevictable pages but CMA can do so */
1535 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001536 return ret;
1537
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001538 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001539
Mel Gormanc8244932012-01-12 17:19:38 -08001540 /*
1541 * To minimise LRU disruption, the caller can indicate that it only
1542 * wants to isolate pages it will be able to operate on without
1543 * blocking - clean pages for the most part.
1544 *
Mel Gormanc8244932012-01-12 17:19:38 -08001545 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1546 * that it is possible to migrate without blocking
1547 */
Johannes Weiner1276ad62017-02-24 14:56:11 -08001548 if (mode & ISOLATE_ASYNC_MIGRATE) {
Mel Gormanc8244932012-01-12 17:19:38 -08001549 /* All the caller can do on PageWriteback is block */
1550 if (PageWriteback(page))
1551 return ret;
1552
1553 if (PageDirty(page)) {
1554 struct address_space *mapping;
Mel Gorman69d763f2018-01-31 16:19:52 -08001555 bool migrate_dirty;
Mel Gormanc8244932012-01-12 17:19:38 -08001556
Mel Gormanc8244932012-01-12 17:19:38 -08001557 /*
1558 * Only pages without mappings or that have a
1559 * ->migratepage callback are possible to migrate
Mel Gorman69d763f2018-01-31 16:19:52 -08001560 * without blocking. However, we can be racing with
1561 * truncation so it's necessary to lock the page
1562 * to stabilise the mapping as truncation holds
1563 * the page lock until after the page is removed
1564 * from the page cache.
Mel Gormanc8244932012-01-12 17:19:38 -08001565 */
Mel Gorman69d763f2018-01-31 16:19:52 -08001566 if (!trylock_page(page))
1567 return ret;
1568
Mel Gormanc8244932012-01-12 17:19:38 -08001569 mapping = page_mapping(page);
Hugh Dickins145e1a72018-06-01 16:50:50 -07001570 migrate_dirty = !mapping || mapping->a_ops->migratepage;
Mel Gorman69d763f2018-01-31 16:19:52 -08001571 unlock_page(page);
1572 if (!migrate_dirty)
Mel Gormanc8244932012-01-12 17:19:38 -08001573 return ret;
1574 }
1575 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001576
Minchan Kimf80c0672011-10-31 17:06:55 -07001577 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1578 return ret;
1579
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001580 if (likely(get_page_unless_zero(page))) {
1581 /*
1582 * Be careful not to clear PageLRU until after we're
1583 * sure the page is not being freed elsewhere -- the
1584 * page release code relies on it.
1585 */
1586 ClearPageLRU(page);
1587 ret = 0;
1588 }
1589
1590 return ret;
1591}
1592
Mel Gorman7ee36a12016-07-28 15:47:17 -07001593
1594/*
1595 * Update LRU sizes after isolating pages. The LRU size updates must
1596 * be complete before mem_cgroup_update_lru_size due to a santity check.
1597 */
1598static __always_inline void update_lru_sizes(struct lruvec *lruvec,
Michal Hockob4536f0c82017-01-10 16:58:04 -08001599 enum lru_list lru, unsigned long *nr_zone_taken)
Mel Gorman7ee36a12016-07-28 15:47:17 -07001600{
Mel Gorman7ee36a12016-07-28 15:47:17 -07001601 int zid;
1602
Mel Gorman7ee36a12016-07-28 15:47:17 -07001603 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1604 if (!nr_zone_taken[zid])
1605 continue;
1606
1607 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001608#ifdef CONFIG_MEMCG
1609 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1610#endif
Mel Gorman7ee36a12016-07-28 15:47:17 -07001611 }
Mel Gorman7ee36a12016-07-28 15:47:17 -07001612
Mel Gorman7ee36a12016-07-28 15:47:17 -07001613}
1614
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001615/**
1616 * pgdat->lru_lock is heavily contended. Some of the functions that
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 * shrink the lists perform better by taking out a batch of pages
1618 * and working on them outside the LRU lock.
1619 *
1620 * For pagecache intensive workloads, this function is the hottest
1621 * spot in the kernel (apart from copy_*_user functions).
1622 *
1623 * Appropriate locks must be held before calling this function.
1624 *
Minchan Kim791b48b2017-05-12 15:47:06 -07001625 * @nr_to_scan: The number of eligible pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001626 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001628 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001629 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001630 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001631 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 *
1633 * returns how many pages were moved onto *@dst.
1634 */
Andrew Morton69e05942006-03-22 00:08:19 -08001635static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001636 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001637 unsigned long *nr_scanned, struct scan_control *sc,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001638 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001640 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001641 unsigned long nr_taken = 0;
Mel Gorman599d0c92016-07-28 15:45:31 -07001642 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001643 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
Johannes Weiner3db65812017-05-03 14:52:13 -07001644 unsigned long skipped = 0;
Minchan Kim791b48b2017-05-12 15:47:06 -07001645 unsigned long scan, total_scan, nr_pages;
Mel Gormanb2e18752016-07-28 15:45:37 -07001646 LIST_HEAD(pages_skipped);
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001647 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
Minchan Kim791b48b2017-05-12 15:47:06 -07001649 scan = 0;
1650 for (total_scan = 0;
1651 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1652 total_scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001653 struct page *page;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001654
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 page = lru_to_page(src);
1656 prefetchw_prev_lru_page(page, src, flags);
1657
Sasha Levin309381fea2014-01-23 15:52:54 -08001658 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001659
Mel Gormanb2e18752016-07-28 15:45:37 -07001660 if (page_zonenum(page) > sc->reclaim_idx) {
1661 list_move(&page->lru, &pages_skipped);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001662 nr_skipped[page_zonenum(page)]++;
Mel Gormanb2e18752016-07-28 15:45:37 -07001663 continue;
1664 }
1665
Minchan Kim791b48b2017-05-12 15:47:06 -07001666 /*
1667 * Do not count skipped pages because that makes the function
1668 * return with no isolated pages if the LRU mostly contains
1669 * ineligible pages. This causes the VM to not reclaim any
1670 * pages, triggering a premature OOM.
1671 */
1672 scan++;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001673 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001674 case 0:
Mel Gorman599d0c92016-07-28 15:45:31 -07001675 nr_pages = hpage_nr_pages(page);
1676 nr_taken += nr_pages;
1677 nr_zone_taken[page_zonenum(page)] += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001678 list_move(&page->lru, dst);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001679 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001680
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001681 case -EBUSY:
1682 /* else it is being freed elsewhere */
1683 list_move(&page->lru, src);
1684 continue;
1685
1686 default:
1687 BUG();
1688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 }
1690
Mel Gormanb2e18752016-07-28 15:45:37 -07001691 /*
1692 * Splice any skipped pages to the start of the LRU list. Note that
1693 * this disrupts the LRU order when reclaiming for lower zones but
1694 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1695 * scanning would soon rescan the same pages to skip and put the
1696 * system at risk of premature OOM.
1697 */
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001698 if (!list_empty(&pages_skipped)) {
1699 int zid;
1700
Johannes Weiner3db65812017-05-03 14:52:13 -07001701 list_splice(&pages_skipped, src);
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001702 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1703 if (!nr_skipped[zid])
1704 continue;
1705
1706 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
Michal Hocko1265e3a2017-02-22 15:44:21 -08001707 skipped += nr_skipped[zid];
Mel Gorman7cc30fc2016-07-28 15:46:59 -07001708 }
1709 }
Minchan Kim791b48b2017-05-12 15:47:06 -07001710 *nr_scanned = total_scan;
Michal Hocko1265e3a2017-02-22 15:44:21 -08001711 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
Minchan Kim791b48b2017-05-12 15:47:06 -07001712 total_scan, skipped, nr_taken, mode, lru);
Michal Hockob4536f0c82017-01-10 16:58:04 -08001713 update_lru_sizes(lruvec, lru, nr_zone_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 return nr_taken;
1715}
1716
Nick Piggin62695a82008-10-18 20:26:09 -07001717/**
1718 * isolate_lru_page - tries to isolate a page from its LRU list
1719 * @page: page to isolate from its LRU list
1720 *
1721 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1722 * vmstat statistic corresponding to whatever LRU list the page was on.
1723 *
1724 * Returns 0 if the page was removed from an LRU list.
1725 * Returns -EBUSY if the page was not on an LRU list.
1726 *
1727 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001728 * the active list, it will have PageActive set. If it was found on
1729 * the unevictable list, it will have the PageUnevictable bit set. That flag
1730 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001731 *
1732 * The vmstat statistic corresponding to the list on which the page was
1733 * found will be decremented.
1734 *
1735 * Restrictions:
Mike Rapoporta5d09be2018-02-06 15:42:19 -08001736 *
Nick Piggin62695a82008-10-18 20:26:09 -07001737 * (1) Must be called with an elevated refcount on the page. This is a
1738 * fundamentnal difference from isolate_lru_pages (which is called
1739 * without a stable reference).
1740 * (2) the lru_lock must not be held.
1741 * (3) interrupts must be enabled.
1742 */
1743int isolate_lru_page(struct page *page)
1744{
1745 int ret = -EBUSY;
1746
Sasha Levin309381fea2014-01-23 15:52:54 -08001747 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001748 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001749
Nick Piggin62695a82008-10-18 20:26:09 -07001750 if (PageLRU(page)) {
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001751 pg_data_t *pgdat = page_pgdat(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001752 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001753
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001754 spin_lock_irq(&pgdat->lru_lock);
1755 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001756 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001757 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001758 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001759 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001760 del_page_from_lru_list(page, lruvec, lru);
1761 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001762 }
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08001763 spin_unlock_irq(&pgdat->lru_lock);
Nick Piggin62695a82008-10-18 20:26:09 -07001764 }
1765 return ret;
1766}
1767
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001768/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001769 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1770 * then get resheduled. When there are massive number of tasks doing page
1771 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1772 * the LRU list will go small and be scanned faster than necessary, leading to
1773 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001774 */
Mel Gorman599d0c92016-07-28 15:45:31 -07001775static int too_many_isolated(struct pglist_data *pgdat, int file,
Rik van Riel35cd7812009-09-21 17:01:38 -07001776 struct scan_control *sc)
1777{
1778 unsigned long inactive, isolated;
1779
1780 if (current_is_kswapd())
1781 return 0;
1782
Tejun Heo97c93412015-05-22 18:23:36 -04001783 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001784 return 0;
1785
1786 if (file) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001787 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1788 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
Rik van Riel35cd7812009-09-21 17:01:38 -07001789 } else {
Mel Gorman599d0c92016-07-28 15:45:31 -07001790 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1791 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
Rik van Riel35cd7812009-09-21 17:01:38 -07001792 }
1793
Fengguang Wu3cf23842012-12-18 14:23:31 -08001794 /*
1795 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1796 * won't get blocked by normal direct-reclaimers, forming a circular
1797 * deadlock.
1798 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001799 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001800 inactive >>= 3;
1801
Rik van Riel35cd7812009-09-21 17:01:38 -07001802 return isolated > inactive;
1803}
1804
Kirill Tkhaia222f342019-05-13 17:17:00 -07001805/*
1806 * This moves pages from @list to corresponding LRU list.
1807 *
1808 * We move them the other way if the page is referenced by one or more
1809 * processes, from rmap.
1810 *
1811 * If the pages are mostly unmapped, the processing is fast and it is
1812 * appropriate to hold zone_lru_lock across the whole operation. But if
1813 * the pages are mapped, the processing is slow (page_referenced()) so we
1814 * should drop zone_lru_lock around each page. It's impossible to balance
1815 * this, so instead we remove the pages from the LRU while processing them.
1816 * It is safe to rely on PG_active against the non-LRU pages in here because
1817 * nobody will play with that bit on a non-LRU page.
1818 *
1819 * The downside is that we have to touch page->_refcount against each page.
1820 * But we had to alter page->flags anyway.
1821 *
1822 * Returns the number of pages moved to the given lruvec.
1823 */
1824
1825static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1826 struct list_head *list)
Mel Gorman66635622010-08-09 17:19:30 -07001827{
Mel Gorman599d0c92016-07-28 15:45:31 -07001828 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001829 int nr_pages, nr_moved = 0;
Hugh Dickins3f797682012-01-12 17:20:07 -08001830 LIST_HEAD(pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001831 struct page *page;
1832 enum lru_list lru;
Mel Gorman66635622010-08-09 17:19:30 -07001833
Kirill Tkhaia222f342019-05-13 17:17:00 -07001834 while (!list_empty(list)) {
1835 page = lru_to_page(list);
Sasha Levin309381fea2014-01-23 15:52:54 -08001836 VM_BUG_ON_PAGE(PageLRU(page), page);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001837 if (unlikely(!page_evictable(page))) {
Kirill Tkhaia222f342019-05-13 17:17:00 -07001838 list_del(&page->lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07001839 spin_unlock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001840 putback_lru_page(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001841 spin_lock_irq(&pgdat->lru_lock);
Mel Gorman66635622010-08-09 17:19:30 -07001842 continue;
1843 }
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);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001848
1849 nr_pages = hpage_nr_pages(page);
1850 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1851 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001852
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001853 if (put_page_testzero(page)) {
1854 __ClearPageLRU(page);
1855 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001856 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001857
1858 if (unlikely(PageCompound(page))) {
Mel Gorman599d0c92016-07-28 15:45:31 -07001859 spin_unlock_irq(&pgdat->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001860 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001861 (*get_compound_page_dtor(page))(page);
Mel Gorman599d0c92016-07-28 15:45:31 -07001862 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001863 } else
1864 list_add(&page->lru, &pages_to_free);
Kirill Tkhaia222f342019-05-13 17:17:00 -07001865 } else {
1866 nr_moved += nr_pages;
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 */
Kirill Tkhaia222f342019-05-13 17:17:00 -07001873 list_splice(&pages_to_free, list);
1874
1875 return nr_moved;
Mel Gorman66635622010-08-09 17:19:30 -07001876}
1877
1878/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001879 * If a kernel thread (such as nfsd for loop-back mounts) services
1880 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1881 * In that case we should only throttle if the backing device it is
1882 * writing to is congested. In other cases it is safe to throttle.
1883 */
1884static int current_may_throttle(void)
1885{
1886 return !(current->flags & PF_LESS_THROTTLE) ||
1887 current->backing_dev_info == NULL ||
1888 bdi_write_congested(current->backing_dev_info);
1889}
1890
1891/*
Mel Gormanb2e18752016-07-28 15:45:37 -07001892 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
Andrew Morton1742f192006-03-22 00:08:21 -08001893 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 */
Mel Gorman66635622010-08-09 17:19:30 -07001895static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001896shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001897 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898{
1899 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001900 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001901 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001902 unsigned long nr_taken;
Kirill Tkhai060f0052019-03-05 15:48:15 -08001903 struct reclaim_stat stat;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001904 int file = is_file_lru(lru);
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001905 enum vm_event_item item;
Mel Gorman599d0c92016-07-28 15:45:31 -07001906 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001907 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Michal Hockodb73ee02017-09-06 16:21:11 -07001908 bool stalled = false;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001909
Mel Gorman599d0c92016-07-28 15:45:31 -07001910 while (unlikely(too_many_isolated(pgdat, file, sc))) {
Michal Hockodb73ee02017-09-06 16:21:11 -07001911 if (stalled)
1912 return 0;
1913
1914 /* wait a bit for the reclaimer. */
1915 msleep(100);
1916 stalled = true;
Rik van Riel35cd7812009-09-21 17:01:38 -07001917
1918 /* We are about to die and free our memory. Return now. */
1919 if (fatal_signal_pending(current))
1920 return SWAP_CLUSTER_MAX;
1921 }
1922
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001924
Mel Gorman599d0c92016-07-28 15:45:31 -07001925 spin_lock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001927 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08001928 &nr_scanned, sc, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001929
Mel Gorman599d0c92016-07-28 15:45:31 -07001930 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07001931 reclaim_stat->recent_scanned[file] += nr_taken;
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001932
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001933 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1934 if (global_reclaim(sc))
1935 __count_vm_events(item, nr_scanned);
1936 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
Mel Gorman599d0c92016-07-28 15:45:31 -07001937 spin_unlock_irq(&pgdat->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001938
Hillf Dantond563c052012-03-21 16:34:02 -07001939 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001940 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001941
Shaohua Lia128ca72017-05-03 14:52:22 -07001942 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
Michal Hocko3c710c12017-02-22 15:44:27 -08001943 &stat, false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001944
Mel Gorman599d0c92016-07-28 15:45:31 -07001945 spin_lock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001946
Kirill Tkhaif46b7912019-05-13 17:22:33 -07001947 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1948 if (global_reclaim(sc))
1949 __count_vm_events(item, nr_reclaimed);
1950 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
Kirill Tkhaib17f18a2019-06-13 15:55:58 -07001951 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1952 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001953
Kirill Tkhaia222f342019-05-13 17:17:00 -07001954 move_pages_to_lru(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001955
Mel Gorman599d0c92016-07-28 15:45:31 -07001956 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001957
Mel Gorman599d0c92016-07-28 15:45:31 -07001958 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins3f797682012-01-12 17:20:07 -08001959
Johannes Weiner747db952014-08-08 14:19:24 -07001960 mem_cgroup_uncharge_list(&page_list);
Mel Gorman2d4894b2017-11-15 17:37:59 -08001961 free_unref_page_list(&page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001962
Mel Gorman92df3a72011-10-31 17:07:56 -07001963 /*
Andrey Ryabinin1c610d52018-03-22 16:17:42 -07001964 * If dirty pages are scanned that are not queued for IO, it
1965 * implies that flushers are not doing their job. This can
1966 * happen when memory pressure pushes dirty pages to the end of
1967 * the LRU before the dirty limits are breached and the dirty
1968 * data has expired. It can also happen when the proportion of
1969 * dirty pages grows not through writes but through memory
1970 * pressure reclaiming all the clean cache. And in some cases,
1971 * the flushers simply cannot keep up with the allocation
1972 * rate. Nudge the flusher threads in case they are asleep.
1973 */
1974 if (stat.nr_unqueued_dirty == nr_taken)
1975 wakeup_flusher_threads(WB_REASON_VMSCAN);
1976
Andrey Ryabinind108c772018-04-10 16:27:59 -07001977 sc->nr.dirty += stat.nr_dirty;
1978 sc->nr.congested += stat.nr_congested;
1979 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1980 sc->nr.writeback += stat.nr_writeback;
1981 sc->nr.immediate += stat.nr_immediate;
1982 sc->nr.taken += nr_taken;
1983 if (file)
1984 sc->nr.file_taken += nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001985
Mel Gorman599d0c92016-07-28 15:45:31 -07001986 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
Steven Rostedtd51d1e62018-04-10 16:28:07 -07001987 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001988 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989}
1990
Hugh Dickinsf6260122012-01-12 17:20:06 -08001991static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001992 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001993 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001994 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001996 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001997 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001998 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07002000 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002001 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002003 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Michal Hocko9d998b42017-02-22 15:44:18 -08002004 unsigned nr_deactivate, nr_activate;
2005 unsigned nr_rotated = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07002006 int file = is_file_lru(lru);
Mel Gorman599d0c92016-07-28 15:45:31 -07002007 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
2009 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07002010
Mel Gorman599d0c92016-07-28 15:45:31 -07002011 spin_lock_irq(&pgdat->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08002012
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07002013 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
Kirill Tkhaia9e7c392019-03-05 15:46:55 -08002014 &nr_scanned, sc, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002015
Mel Gorman599d0c92016-07-28 15:45:31 -07002016 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002017 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002018
Mel Gorman599d0c92016-07-28 15:45:31 -07002019 __count_vm_events(PGREFILL, nr_scanned);
Yafang Shao2fa26902019-05-13 17:23:02 -07002020 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
Hugh Dickins9d5e6a92016-05-19 17:12:38 -07002021
Mel Gorman599d0c92016-07-28 15:45:31 -07002022 spin_unlock_irq(&pgdat->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 while (!list_empty(&l_hold)) {
2025 cond_resched();
2026 page = lru_to_page(&l_hold);
2027 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07002028
Hugh Dickins39b5f292012-10-08 16:33:18 -07002029 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002030 putback_lru_page(page);
2031 continue;
2032 }
2033
Mel Gormancc715d92012-03-21 16:34:00 -07002034 if (unlikely(buffer_heads_over_limit)) {
2035 if (page_has_private(page) && trylock_page(page)) {
2036 if (page_has_private(page))
2037 try_to_release_page(page, 0);
2038 unlock_page(page);
2039 }
2040 }
2041
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07002042 if (page_referenced(page, 0, sc->target_mem_cgroup,
2043 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08002044 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07002045 /*
2046 * Identify referenced, file-backed active pages and
2047 * give them one more trip around the active list. So
2048 * that executable code get better chances to stay in
2049 * memory under moderate memory pressure. Anon pages
2050 * are not likely to be evicted by use-once streaming
2051 * IO, plus JVM can create lots of anon VM_EXEC pages,
2052 * so we ignore them here.
2053 */
Wu Fengguang41e20982009-10-26 16:49:53 -07002054 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07002055 list_add(&page->lru, &l_active);
2056 continue;
2057 }
2058 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07002059
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07002060 ClearPageActive(page); /* we are de-activating */
Johannes Weiner1899ad12018-10-26 15:06:04 -07002061 SetPageWorkingset(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 list_add(&page->lru, &l_inactive);
2063 }
2064
Andrew Mortonb5557492009-01-06 14:40:13 -08002065 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002066 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08002067 */
Mel Gorman599d0c92016-07-28 15:45:31 -07002068 spin_lock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002069 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07002070 * Count referenced pages from currently used mappings as rotated,
2071 * even though only some of them are actually re-activated. This
2072 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07002073 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07002074 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07002075 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07002076
Kirill Tkhaia222f342019-05-13 17:17:00 -07002077 nr_activate = move_pages_to_lru(lruvec, &l_active);
2078 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
Kirill Tkhaif372d892019-05-13 17:16:57 -07002079 /* Keep all free pages in l_active list */
2080 list_splice(&l_inactive, &l_active);
Kirill Tkhai9851ac12019-05-13 17:16:54 -07002081
2082 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2083 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2084
Mel Gorman599d0c92016-07-28 15:45:31 -07002085 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2086 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08002087
Kirill Tkhaif372d892019-05-13 17:16:57 -07002088 mem_cgroup_uncharge_list(&l_active);
2089 free_unref_page_list(&l_active);
Michal Hocko9d998b42017-02-22 15:44:18 -08002090 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2091 nr_deactivate, nr_rotated, sc->priority, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092}
2093
Rik van Riel59dc76b2016-05-20 16:56:31 -07002094/*
2095 * The inactive anon list should be small enough that the VM never has
2096 * to do too much work.
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002097 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002098 * The inactive file list should be small enough to leave most memory
2099 * to the established workingset on the scan-resistant active list,
2100 * but large enough to avoid thrashing the aggregate readahead window.
2101 *
2102 * Both inactive lists should also be large enough that each inactive
2103 * page has a chance to be referenced again before it is reclaimed.
2104 *
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002105 * If that fails and refaulting is observed, the inactive list grows.
2106 *
Rik van Riel59dc76b2016-05-20 16:56:31 -07002107 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
Andrey Ryabinin3a50d142017-11-15 17:34:15 -08002108 * on this LRU, maintained by the pageout code. An inactive_ratio
Rik van Riel59dc76b2016-05-20 16:56:31 -07002109 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2110 *
2111 * total target max
2112 * memory ratio inactive
2113 * -------------------------------------
2114 * 10MB 1 5MB
2115 * 100MB 1 50MB
2116 * 1GB 3 250MB
2117 * 10GB 10 0.9GB
2118 * 100GB 31 3GB
2119 * 1TB 101 10GB
2120 * 10TB 320 32GB
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002121 */
Mel Gormanf8d1a312016-07-28 15:47:34 -07002122static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
Kuo-Hsin Yang2c012a42019-07-11 20:52:04 -07002123 struct scan_control *sc, bool trace)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002124{
Michal Hockofd538802017-02-22 15:45:58 -08002125 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002126 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2127 enum lru_list inactive_lru = file * LRU_FILE;
2128 unsigned long inactive, active;
2129 unsigned long inactive_ratio;
2130 unsigned long refaults;
Rik van Riel59dc76b2016-05-20 16:56:31 -07002131 unsigned long gb;
2132
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002133 /*
2134 * If we don't have swap space, anonymous page deactivation
2135 * is pointless.
2136 */
Rik van Riel59dc76b2016-05-20 16:56:31 -07002137 if (!file && !total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08002138 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07002139
Michal Hockofd538802017-02-22 15:45:58 -08002140 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2141 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
Mel Gormanf8d1a312016-07-28 15:47:34 -07002142
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002143 /*
2144 * When refaults are being observed, it means a new workingset
2145 * is being established. Disable active list protection to get
2146 * rid of the stale workingset quickly.
2147 */
Johannes Weiner205b20c2019-05-14 15:47:06 -07002148 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
Kuo-Hsin Yang2c012a42019-07-11 20:52:04 -07002149 if (file && lruvec->refaults != refaults) {
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002150 inactive_ratio = 0;
2151 } else {
2152 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2153 if (gb)
2154 inactive_ratio = int_sqrt(10 * gb);
2155 else
2156 inactive_ratio = 1;
2157 }
2158
Kuo-Hsin Yang2c012a42019-07-11 20:52:04 -07002159 if (trace)
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002160 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2161 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2162 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2163 inactive_ratio, file);
Michal Hockofd538802017-02-22 15:45:58 -08002164
Rik van Riel59dc76b2016-05-20 16:56:31 -07002165 return inactive * inactive_ratio < active;
Rik van Rielb39415b2009-12-14 17:59:48 -08002166}
2167
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002168static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Johannes Weiner3b991202019-04-18 17:50:34 -07002169 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07002170{
Rik van Rielb39415b2009-12-14 17:59:48 -08002171 if (is_active_lru(lru)) {
Johannes Weiner3b991202019-04-18 17:50:34 -07002172 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002173 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07002174 return 0;
2175 }
2176
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002177 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07002178}
2179
Johannes Weiner9a265112013-02-22 16:32:17 -08002180enum scan_balance {
2181 SCAN_EQUAL,
2182 SCAN_FRACT,
2183 SCAN_ANON,
2184 SCAN_FILE,
2185};
2186
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002188 * Determine how aggressively the anon and file LRU lists should be
2189 * scanned. The relative value of each set of LRU lists is determined
2190 * by looking at the fraction of the pages scanned we did rotate back
2191 * onto the active list instead of evict.
2192 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08002193 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2194 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002195 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002196static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002197 struct scan_control *sc, unsigned long *nr,
2198 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002199{
Vladimir Davydov33377672016-01-20 15:02:59 -08002200 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002201 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08002202 u64 fraction[2];
2203 u64 denominator = 0; /* gcc */
Mel Gorman599d0c92016-07-28 15:45:31 -07002204 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08002205 unsigned long anon_prio, file_prio;
2206 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07002207 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08002208 unsigned long ap, fp;
2209 enum lru_list lru;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002210
2211 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002212 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002213 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002214 goto out;
2215 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002216
Johannes Weiner10316b32013-02-22 16:32:14 -08002217 /*
2218 * Global reclaim will swap to prevent OOM even with no
2219 * swappiness, but memcg users want to use this knob to
2220 * disable swapping for individual groups completely when
2221 * using the memory controller's swap limit feature would be
2222 * too expensive.
2223 */
Johannes Weiner02695172014-08-06 16:06:17 -07002224 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002225 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002226 goto out;
2227 }
2228
2229 /*
2230 * Do not apply any pressure balancing cleverness when the
2231 * system is close to OOM, scan both anon and file equally
2232 * (unless the swappiness setting disagrees with swapping).
2233 */
Johannes Weiner02695172014-08-06 16:06:17 -07002234 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002235 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002236 goto out;
2237 }
2238
Johannes Weiner11d16c22013-02-22 16:32:15 -08002239 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002240 * Prevent the reclaimer from falling into the cache trap: as
2241 * cache pages start out inactive, every cache fault will tip
2242 * the scan balance towards the file LRU. And as the file LRU
2243 * shrinks, so does the window for rotation from references.
2244 * This means we have a runaway feedback loop where a tiny
2245 * thrashing file LRU becomes infinitely more attractive than
2246 * anon pages. Try to detect this based on file LRU size.
2247 */
2248 if (global_reclaim(sc)) {
Mel Gorman599d0c92016-07-28 15:45:31 -07002249 unsigned long pgdatfile;
2250 unsigned long pgdatfree;
2251 int z;
2252 unsigned long total_high_wmark = 0;
Johannes Weiner62376252014-05-06 12:50:07 -07002253
Mel Gorman599d0c92016-07-28 15:45:31 -07002254 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2255 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2256 node_page_state(pgdat, NR_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002257
Mel Gorman599d0c92016-07-28 15:45:31 -07002258 for (z = 0; z < MAX_NR_ZONES; z++) {
2259 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002260 if (!managed_zone(zone))
Mel Gorman599d0c92016-07-28 15:45:31 -07002261 continue;
2262
2263 total_high_wmark += high_wmark_pages(zone);
2264 }
2265
2266 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
David Rientjes06226222017-07-10 15:47:20 -07002267 /*
2268 * Force SCAN_ANON if there are enough inactive
2269 * anonymous pages on the LRU in eligible zones.
2270 * Otherwise, the small LRU gets thrashed.
2271 */
Johannes Weiner3b991202019-04-18 17:50:34 -07002272 if (!inactive_list_is_low(lruvec, false, sc, false) &&
David Rientjes06226222017-07-10 15:47:20 -07002273 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2274 >> sc->priority) {
2275 scan_balance = SCAN_ANON;
2276 goto out;
2277 }
Johannes Weiner62376252014-05-06 12:50:07 -07002278 }
2279 }
2280
2281 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002282 * If there is enough inactive page cache, i.e. if the size of the
2283 * inactive list is greater than that of the active list *and* the
2284 * inactive list actually has some pages to scan on this priority, we
2285 * do not reclaim anything from the anonymous working set right now.
2286 * Without the second condition we could end up never scanning an
2287 * lruvec even if it has plenty of old anonymous pages unless the
2288 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002289 */
Johannes Weiner3b991202019-04-18 17:50:34 -07002290 if (!inactive_list_is_low(lruvec, true, sc, false) &&
Michal Hocko71ab6cf2017-02-22 15:46:01 -08002291 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002292 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002293 goto out;
2294 }
2295
Johannes Weiner9a265112013-02-22 16:32:17 -08002296 scan_balance = SCAN_FRACT;
2297
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002298 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002299 * With swappiness at 100, anonymous and file have the same priority.
2300 * This scanning priority is essentially the inverse of IO cost.
2301 */
Johannes Weiner02695172014-08-06 16:06:17 -07002302 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002303 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002304
2305 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002306 * OK, so we have swap space and a fair amount of page cache
2307 * pages. We use the recently rotated / recently scanned
2308 * ratios to determine how valuable each cache is.
2309 *
2310 * Because workloads change over time (and to avoid overflow)
2311 * we keep these statistics as a floating average, which ends
2312 * up weighing recent references more than old ones.
2313 *
2314 * anon in [0], file in [1]
2315 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002316
Michal Hockofd538802017-02-22 15:45:58 -08002317 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2318 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2319 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2320 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002321
Mel Gorman599d0c92016-07-28 15:45:31 -07002322 spin_lock_irq(&pgdat->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002323 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002324 reclaim_stat->recent_scanned[0] /= 2;
2325 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002326 }
2327
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002328 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002329 reclaim_stat->recent_scanned[1] /= 2;
2330 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002331 }
2332
2333 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002334 * The amount of pressure on anon vs file pages is inversely
2335 * proportional to the fraction of recently scanned pages on
2336 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002337 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002338 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002339 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002340
Satoru Moriyafe350042012-05-29 15:06:47 -07002341 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002342 fp /= reclaim_stat->recent_rotated[1] + 1;
Mel Gorman599d0c92016-07-28 15:45:31 -07002343 spin_unlock_irq(&pgdat->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002344
Shaohua Li76a33fc2010-05-24 14:32:36 -07002345 fraction[0] = ap;
2346 fraction[1] = fp;
2347 denominator = ap + fp + 1;
2348out:
Johannes Weiner688035f2017-05-03 14:52:07 -07002349 *lru_pages = 0;
2350 for_each_evictable_lru(lru) {
2351 int file = is_file_lru(lru);
2352 unsigned long size;
2353 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002354
Johannes Weiner688035f2017-05-03 14:52:07 -07002355 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2356 scan = size >> sc->priority;
2357 /*
2358 * If the cgroup's already been deleted, make sure to
2359 * scrape out the remaining cache.
2360 */
2361 if (!scan && !mem_cgroup_online(memcg))
2362 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002363
Johannes Weiner688035f2017-05-03 14:52:07 -07002364 switch (scan_balance) {
2365 case SCAN_EQUAL:
2366 /* Scan lists relative to size */
2367 break;
2368 case SCAN_FRACT:
Johannes Weiner9a265112013-02-22 16:32:17 -08002369 /*
Johannes Weiner688035f2017-05-03 14:52:07 -07002370 * Scan types proportional to swappiness and
2371 * their relative recent reclaim efficiency.
Roman Gushchin68600f62018-10-26 15:03:27 -07002372 * Make sure we don't miss the last page
2373 * because of a round-off error.
Johannes Weiner9a265112013-02-22 16:32:17 -08002374 */
Roman Gushchin68600f62018-10-26 15:03:27 -07002375 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2376 denominator);
Johannes Weiner688035f2017-05-03 14:52:07 -07002377 break;
2378 case SCAN_FILE:
2379 case SCAN_ANON:
2380 /* Scan one type exclusively */
2381 if ((scan_balance == SCAN_FILE) != file) {
2382 size = 0;
2383 scan = 0;
2384 }
2385 break;
2386 default:
2387 /* Look ma, no brain */
2388 BUG();
Johannes Weiner9a265112013-02-22 16:32:17 -08002389 }
Johannes Weiner688035f2017-05-03 14:52:07 -07002390
2391 *lru_pages += size;
2392 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002393 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002394}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002395
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002396/*
Mel Gormana9dd0a82016-07-28 15:46:02 -07002397 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002398 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002399static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
Vladimir Davydov33377672016-01-20 15:02:59 -08002400 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002401{
Mel Gormanef8f2322016-07-28 15:46:05 -07002402 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002403 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002404 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002405 unsigned long nr_to_scan;
2406 enum lru_list lru;
2407 unsigned long nr_reclaimed = 0;
2408 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2409 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002410 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002411
Vladimir Davydov33377672016-01-20 15:02:59 -08002412 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002413
Mel Gormane82e0562013-07-03 15:01:44 -07002414 /* Record the original scan target for proportional adjustments later */
2415 memcpy(targets, nr, sizeof(nr));
2416
Mel Gorman1a501902014-06-04 16:10:49 -07002417 /*
2418 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2419 * event that can occur when there is little memory pressure e.g.
2420 * multiple streaming readers/writers. Hence, we do not abort scanning
2421 * when the requested number of pages are reclaimed when scanning at
2422 * DEF_PRIORITY on the assumption that the fact we are direct
2423 * reclaiming implies that kswapd is not keeping up and it is best to
2424 * do a batch of work at once. For memcg reclaim one check is made to
2425 * abort proportional reclaim if either the file or anon lru has already
2426 * dropped to zero at the first pass.
2427 */
2428 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2429 sc->priority == DEF_PRIORITY);
2430
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002431 blk_start_plug(&plug);
2432 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2433 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002434 unsigned long nr_anon, nr_file, percentage;
2435 unsigned long nr_scanned;
2436
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002437 for_each_evictable_lru(lru) {
2438 if (nr[lru]) {
2439 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2440 nr[lru] -= nr_to_scan;
2441
2442 nr_reclaimed += shrink_list(lru, nr_to_scan,
Johannes Weiner3b991202019-04-18 17:50:34 -07002443 lruvec, sc);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002444 }
2445 }
Mel Gormane82e0562013-07-03 15:01:44 -07002446
Michal Hockobd041732016-12-02 17:26:48 -08002447 cond_resched();
2448
Mel Gormane82e0562013-07-03 15:01:44 -07002449 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2450 continue;
2451
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002452 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002453 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002454 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002455 * proportionally what was requested by get_scan_count(). We
2456 * stop reclaiming one LRU and reduce the amount scanning
2457 * proportional to the original scan target.
2458 */
2459 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2460 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2461
Mel Gorman1a501902014-06-04 16:10:49 -07002462 /*
2463 * It's just vindictive to attack the larger once the smaller
2464 * has gone to zero. And given the way we stop scanning the
2465 * smaller below, this makes sure that we only make one nudge
2466 * towards proportionality once we've got nr_to_reclaim.
2467 */
2468 if (!nr_file || !nr_anon)
2469 break;
2470
Mel Gormane82e0562013-07-03 15:01:44 -07002471 if (nr_file > nr_anon) {
2472 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2473 targets[LRU_ACTIVE_ANON] + 1;
2474 lru = LRU_BASE;
2475 percentage = nr_anon * 100 / scan_target;
2476 } else {
2477 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2478 targets[LRU_ACTIVE_FILE] + 1;
2479 lru = LRU_FILE;
2480 percentage = nr_file * 100 / scan_target;
2481 }
2482
2483 /* Stop scanning the smaller of the LRU */
2484 nr[lru] = 0;
2485 nr[lru + LRU_ACTIVE] = 0;
2486
2487 /*
2488 * Recalculate the other LRU scan count based on its original
2489 * scan target and the percentage scanning already complete
2490 */
2491 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2492 nr_scanned = targets[lru] - nr[lru];
2493 nr[lru] = targets[lru] * (100 - percentage) / 100;
2494 nr[lru] -= min(nr[lru], nr_scanned);
2495
2496 lru += LRU_ACTIVE;
2497 nr_scanned = targets[lru] - nr[lru];
2498 nr[lru] = targets[lru] * (100 - percentage) / 100;
2499 nr[lru] -= min(nr[lru], nr_scanned);
2500
2501 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002502 }
2503 blk_finish_plug(&plug);
2504 sc->nr_reclaimed += nr_reclaimed;
2505
2506 /*
2507 * Even if we did not try to evict anon pages at all, we want to
2508 * rebalance the anon lru active/inactive ratio.
2509 */
Johannes Weiner3b991202019-04-18 17:50:34 -07002510 if (inactive_list_is_low(lruvec, false, sc, true))
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002511 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2512 sc, LRU_ACTIVE_ANON);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002513}
2514
Mel Gorman23b9da52012-05-29 15:06:20 -07002515/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002516static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002517{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002518 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002519 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002520 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002521 return true;
2522
2523 return false;
2524}
2525
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002526/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002527 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2528 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2529 * true if more pages should be reclaimed such that when the page allocator
2530 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2531 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002532 */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002533static inline bool should_continue_reclaim(struct pglist_data *pgdat,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002534 unsigned long nr_reclaimed,
2535 unsigned long nr_scanned,
2536 struct scan_control *sc)
2537{
2538 unsigned long pages_for_compaction;
2539 unsigned long inactive_lru_pages;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002540 int z;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002541
2542 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002543 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002544 return false;
2545
Mel Gorman28765922011-02-25 14:44:20 -08002546 /* Consider stopping depending on scan and reclaim activity */
Michal Hockodcda9b02017-07-12 14:36:45 -07002547 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
Mel Gorman28765922011-02-25 14:44:20 -08002548 /*
Michal Hockodcda9b02017-07-12 14:36:45 -07002549 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
Mel Gorman28765922011-02-25 14:44:20 -08002550 * full LRU list has been scanned and we are still failing
2551 * to reclaim pages. This full LRU scan is potentially
Michal Hockodcda9b02017-07-12 14:36:45 -07002552 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
Mel Gorman28765922011-02-25 14:44:20 -08002553 */
2554 if (!nr_reclaimed && !nr_scanned)
2555 return false;
2556 } else {
2557 /*
Michal Hockodcda9b02017-07-12 14:36:45 -07002558 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
Mel Gorman28765922011-02-25 14:44:20 -08002559 * fail without consequence, stop if we failed to reclaim
2560 * any pages from the last SWAP_CLUSTER_MAX number of
2561 * pages that were scanned. This will return to the
2562 * caller faster at the risk reclaim/compaction and
2563 * the resulting allocation attempt fails
2564 */
2565 if (!nr_reclaimed)
2566 return false;
2567 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002568
2569 /*
2570 * If we have not reclaimed enough pages for compaction and the
2571 * inactive lists are large enough, continue reclaiming
2572 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07002573 pages_for_compaction = compact_gap(sc->order);
Mel Gormana9dd0a82016-07-28 15:46:02 -07002574 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002575 if (get_nr_swap_pages() > 0)
Mel Gormana9dd0a82016-07-28 15:46:02 -07002576 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002577 if (sc->nr_reclaimed < pages_for_compaction &&
2578 inactive_lru_pages > pages_for_compaction)
2579 return true;
2580
2581 /* If compaction would go ahead or the allocation would succeed, stop */
Mel Gormana9dd0a82016-07-28 15:46:02 -07002582 for (z = 0; z <= sc->reclaim_idx; z++) {
2583 struct zone *zone = &pgdat->node_zones[z];
Mel Gorman6aa303d2016-09-01 16:14:55 -07002584 if (!managed_zone(zone))
Mel Gormana9dd0a82016-07-28 15:46:02 -07002585 continue;
2586
2587 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
Vlastimil Babkacf378312016-10-07 16:57:41 -07002588 case COMPACT_SUCCESS:
Mel Gormana9dd0a82016-07-28 15:46:02 -07002589 case COMPACT_CONTINUE:
2590 return false;
2591 default:
2592 /* check next zone */
2593 ;
2594 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002595 }
Mel Gormana9dd0a82016-07-28 15:46:02 -07002596 return true;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002597}
2598
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002599static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2600{
2601 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2602 (memcg && memcg_congested(pgdat, memcg));
2603}
2604
Mel Gorman970a39a2016-07-28 15:46:35 -07002605static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002606{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002607 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002608 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002609 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002610
Johannes Weiner56600482012-01-12 17:17:59 -08002611 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002612 struct mem_cgroup *root = sc->target_mem_cgroup;
2613 struct mem_cgroup_reclaim_cookie reclaim = {
Mel Gormanef8f2322016-07-28 15:46:05 -07002614 .pgdat = pgdat,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002615 .priority = sc->priority,
2616 };
Mel Gormana9dd0a82016-07-28 15:46:02 -07002617 unsigned long node_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002618 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002619
Andrey Ryabinind108c772018-04-10 16:27:59 -07002620 memset(&sc->nr, 0, sizeof(sc->nr));
2621
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002622 nr_reclaimed = sc->nr_reclaimed;
2623 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002624
Andrew Morton694fbc02013-09-24 15:27:37 -07002625 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2626 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002627 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002628 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002629 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002630
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002631 switch (mem_cgroup_protected(root, memcg)) {
2632 case MEMCG_PROT_MIN:
2633 /*
2634 * Hard protection.
2635 * If there is no reclaimable memory, OOM.
2636 */
2637 continue;
2638 case MEMCG_PROT_LOW:
2639 /*
2640 * Soft protection.
2641 * Respect the protection only as long as
2642 * there is an unprotected supply
2643 * of reclaimable memory from other cgroups.
2644 */
Yisheng Xied6622f62017-05-03 14:53:57 -07002645 if (!sc->memcg_low_reclaim) {
2646 sc->memcg_low_skipped = 1;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002647 continue;
Yisheng Xied6622f62017-05-03 14:53:57 -07002648 }
Johannes Weinere27be242018-04-10 16:29:45 -07002649 memcg_memory_event(memcg, MEMCG_LOW);
Roman Gushchinbf8d5d52018-06-07 17:07:46 -07002650 break;
2651 case MEMCG_PROT_NONE:
2652 break;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002653 }
2654
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002655 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002656 scanned = sc->nr_scanned;
Mel Gormana9dd0a82016-07-28 15:46:02 -07002657 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2658 node_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002659
Mel Gorman1c308442018-12-28 00:35:52 -08002660 if (sc->may_shrinkslab) {
2661 shrink_slab(sc->gfp_mask, pgdat->node_id,
Vladimir Davydovaeed1d322018-08-17 15:48:17 -07002662 memcg, sc->priority);
Mel Gorman1c308442018-12-28 00:35:52 -08002663 }
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002664
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002665 /* Record the group's reclaim efficiency */
2666 vmpressure(sc->gfp_mask, memcg, false,
2667 sc->nr_scanned - scanned,
2668 sc->nr_reclaimed - reclaimed);
2669
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002670 /*
Yang Shi2bb0f342019-03-05 15:48:05 -08002671 * Kswapd have to scan all memory cgroups to fulfill
2672 * the overall scan target for the node.
Michal Hockoa394cb82013-02-22 16:32:30 -08002673 *
2674 * Limit reclaim, on the other hand, only cares about
2675 * nr_to_reclaim pages to be reclaimed and it will
2676 * retry with decreasing priority if one round over the
2677 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002678 */
Yang Shi2bb0f342019-03-05 15:48:05 -08002679 if (!current_is_kswapd() &&
Michal Hockoa394cb82013-02-22 16:32:30 -08002680 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002681 mem_cgroup_iter_break(root, memcg);
2682 break;
2683 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002684 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002685
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002686 if (reclaim_state) {
2687 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2688 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002689 }
2690
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002691 /* Record the subtree's reclaim efficiency */
2692 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002693 sc->nr_scanned - nr_scanned,
2694 sc->nr_reclaimed - nr_reclaimed);
2695
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002696 if (sc->nr_reclaimed - nr_reclaimed)
2697 reclaimable = true;
2698
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002699 if (current_is_kswapd()) {
2700 /*
2701 * If reclaim is isolating dirty pages under writeback,
2702 * it implies that the long-lived page allocation rate
2703 * is exceeding the page laundering rate. Either the
2704 * global limits are not being effective at throttling
2705 * processes due to the page distribution throughout
2706 * zones or there is heavy usage of a slow backing
2707 * device. The only option is to throttle from reclaim
2708 * context which is not ideal as there is no guarantee
2709 * the dirtying process is throttled in the same way
2710 * balance_dirty_pages() manages.
2711 *
2712 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2713 * count the number of pages under pages flagged for
2714 * immediate reclaim and stall if any are encountered
2715 * in the nr_immediate check below.
2716 */
2717 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2718 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002719
Andrey Ryabinind108c772018-04-10 16:27:59 -07002720 /*
2721 * Tag a node as congested if all the dirty pages
2722 * scanned were backed by a congested BDI and
2723 * wait_iff_congested will stall.
2724 */
2725 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2726 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2727
2728 /* Allow kswapd to start writing pages during reclaim.*/
2729 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2730 set_bit(PGDAT_DIRTY, &pgdat->flags);
2731
2732 /*
2733 * If kswapd scans pages marked marked for immediate
2734 * reclaim and under writeback (nr_immediate), it
2735 * implies that pages are cycling through the LRU
2736 * faster than they are written so also forcibly stall.
2737 */
2738 if (sc->nr.immediate)
2739 congestion_wait(BLK_RW_ASYNC, HZ/10);
2740 }
2741
2742 /*
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002743 * Legacy memcg will stall in page writeback so avoid forcibly
2744 * stalling in wait_iff_congested().
2745 */
2746 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2747 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2748 set_memcg_congestion(pgdat, root, true);
2749
2750 /*
Andrey Ryabinind108c772018-04-10 16:27:59 -07002751 * Stall direct reclaim for IO completions if underlying BDIs
2752 * and node is congested. Allow kswapd to continue until it
2753 * starts encountering unqueued dirty pages or cycling through
2754 * the LRU too quickly.
2755 */
2756 if (!sc->hibernation_mode && !current_is_kswapd() &&
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002757 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2758 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
Andrey Ryabinind108c772018-04-10 16:27:59 -07002759
Mel Gormana9dd0a82016-07-28 15:46:02 -07002760 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002761 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002762
Johannes Weinerc73322d2017-05-03 14:51:51 -07002763 /*
2764 * Kswapd gives up on balancing particular nodes after too
2765 * many failures to reclaim anything from them and goes to
2766 * sleep. On reclaim progress, reset the failure counter. A
2767 * successful direct reclaim run will revive a dormant kswapd.
2768 */
2769 if (reclaimable)
2770 pgdat->kswapd_failures = 0;
2771
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002772 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002773}
2774
Vlastimil Babka53853e22014-10-09 15:27:02 -07002775/*
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002776 * Returns true if compaction should go ahead for a costly-order request, or
2777 * the allocation would already succeed without compaction. Return false if we
2778 * should reclaim first.
Vlastimil Babka53853e22014-10-09 15:27:02 -07002779 */
Mel Gorman4f588332016-07-28 15:46:38 -07002780static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002781{
Mel Gorman31483b62016-07-28 15:45:46 -07002782 unsigned long watermark;
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002783 enum compact_result suitable;
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002784
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002785 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2786 if (suitable == COMPACT_SUCCESS)
2787 /* Allocation should succeed already. Don't reclaim. */
2788 return true;
2789 if (suitable == COMPACT_SKIPPED)
2790 /* Compaction cannot yet proceed. Do reclaim. */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002791 return false;
2792
Vlastimil Babkafdd4c6142016-10-07 16:58:03 -07002793 /*
2794 * Compaction is already possible, but it takes time to run and there
2795 * are potentially other callers using the pages just freed. So proceed
2796 * with reclaim to make a buffer of free pages available to give
2797 * compaction a reasonable chance of completing and allocating the page.
2798 * Note that we won't actually reclaim the whole buffer in one attempt
2799 * as the target watermark in should_continue_reclaim() is lower. But if
2800 * we are already above the high+gap watermark, don't reclaim at all.
2801 */
2802 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2803
2804 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002805}
2806
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807/*
2808 * This is the direct reclaim path, for page-allocating processes. We only
2809 * try to reclaim pages from zones which will satisfy the caller's allocation
2810 * request.
2811 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 * If a zone is deemed to be full of pinned pages then just give it a light
2813 * scan then give up on it.
2814 */
Michal Hocko0a0337e2016-05-20 16:57:00 -07002815static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816{
Mel Gormandd1a2392008-04-28 02:12:17 -07002817 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002818 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002819 unsigned long nr_soft_reclaimed;
2820 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002821 gfp_t orig_mask;
Mel Gorman79dafcd2016-07-28 15:45:53 -07002822 pg_data_t *last_pgdat = NULL;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002823
Mel Gormancc715d92012-03-21 16:34:00 -07002824 /*
2825 * If the number of buffer_heads in the machine exceeds the maximum
2826 * allowed level, force direct reclaim to scan the highmem zone as
2827 * highmem pages could be pinning lowmem pages storing buffer_heads
2828 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002829 orig_mask = sc->gfp_mask;
Mel Gormanb2e18752016-07-28 15:45:37 -07002830 if (buffer_heads_over_limit) {
Mel Gormancc715d92012-03-21 16:34:00 -07002831 sc->gfp_mask |= __GFP_HIGHMEM;
Mel Gorman4f588332016-07-28 15:46:38 -07002832 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
Mel Gormanb2e18752016-07-28 15:45:37 -07002833 }
Mel Gormancc715d92012-03-21 16:34:00 -07002834
Mel Gormand4debc62010-08-09 17:19:29 -07002835 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Mel Gormanb2e18752016-07-28 15:45:37 -07002836 sc->reclaim_idx, sc->nodemask) {
Mel Gormanb2e18752016-07-28 15:45:37 -07002837 /*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002838 * Take care memory controller reclaiming has small influence
2839 * to global LRU.
2840 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002841 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002842 if (!cpuset_zone_allowed(zone,
2843 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002844 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002845
Johannes Weiner0b064962014-08-06 16:06:12 -07002846 /*
2847 * If we already have plenty of memory free for
2848 * compaction in this zone, don't free any more.
2849 * Even though compaction is invoked for any
2850 * non-zero order, only frequent costly order
2851 * reclamation is disruptive enough to become a
2852 * noticeable problem, like transparent huge
2853 * page allocations.
2854 */
2855 if (IS_ENABLED(CONFIG_COMPACTION) &&
2856 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
Mel Gorman4f588332016-07-28 15:46:38 -07002857 compaction_ready(zone, sc)) {
Johannes Weiner0b064962014-08-06 16:06:12 -07002858 sc->compaction_ready = true;
2859 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002860 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002861
Andrew Morton0608f432013-09-24 15:27:41 -07002862 /*
Mel Gorman79dafcd2016-07-28 15:45:53 -07002863 * Shrink each node in the zonelist once. If the
2864 * zonelist is ordered by zone (not the default) then a
2865 * node may be shrunk multiple times but in that case
2866 * the user prefers lower zones being preserved.
2867 */
2868 if (zone->zone_pgdat == last_pgdat)
2869 continue;
2870
2871 /*
Andrew Morton0608f432013-09-24 15:27:41 -07002872 * This steals pages from memory cgroups over softlimit
2873 * and returns the number of reclaimed pages and
2874 * scanned pages. This works for global memory pressure
2875 * and balancing, not for a memcg's limit.
2876 */
2877 nr_soft_scanned = 0;
Mel Gormanef8f2322016-07-28 15:46:05 -07002878 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
Andrew Morton0608f432013-09-24 15:27:41 -07002879 sc->order, sc->gfp_mask,
2880 &nr_soft_scanned);
2881 sc->nr_reclaimed += nr_soft_reclaimed;
2882 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002883 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002884 }
Nick Piggin408d8542006-09-25 23:31:27 -07002885
Mel Gorman79dafcd2016-07-28 15:45:53 -07002886 /* See comment about same check for global reclaim above */
2887 if (zone->zone_pgdat == last_pgdat)
2888 continue;
2889 last_pgdat = zone->zone_pgdat;
Mel Gorman970a39a2016-07-28 15:46:35 -07002890 shrink_node(zone->zone_pgdat, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 }
Mel Gormane0c23272011-10-31 17:09:33 -07002892
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002893 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002894 * Restore to original mask to avoid the impact on the caller if we
2895 * promoted it to __GFP_HIGHMEM.
2896 */
2897 sc->gfp_mask = orig_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002899
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002900static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2901{
2902 struct mem_cgroup *memcg;
2903
2904 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2905 do {
2906 unsigned long refaults;
2907 struct lruvec *lruvec;
2908
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002909 lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner205b20c2019-05-14 15:47:06 -07002910 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002911 lruvec->refaults = refaults;
2912 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2913}
2914
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915/*
2916 * This is the main entry point to direct page reclaim.
2917 *
2918 * If a full scan of the inactive list fails to free enough memory then we
2919 * are "out of memory" and something needs to be killed.
2920 *
2921 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2922 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002923 * caller can't do much about. We kick the writeback threads and take explicit
2924 * naps in the hope that some of these pages can be written. But if the
2925 * allocating task holds filesystem locks which prevent writeout this might not
2926 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002927 *
2928 * returns: 0, if no pages reclaimed
2929 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 */
Mel Gormandac1d272008-04-28 02:12:12 -07002931static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002932 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002934 int initial_priority = sc->priority;
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002935 pg_data_t *last_pgdat;
2936 struct zoneref *z;
2937 struct zone *zone;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002938retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002939 delayacct_freepages_start();
2940
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002941 if (global_reclaim(sc))
Mel Gorman7cc30fc2016-07-28 15:46:59 -07002942 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002944 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002945 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2946 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002947 sc->nr_scanned = 0;
Michal Hocko0a0337e2016-05-20 16:57:00 -07002948 shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002949
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002950 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002951 break;
2952
2953 if (sc->compaction_ready)
2954 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955
2956 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002957 * If we're getting trouble reclaiming, start doing
2958 * writepage even in laptop mode.
2959 */
2960 if (sc->priority < DEF_PRIORITY - 2)
2961 sc->may_writepage = 1;
Johannes Weiner0b064962014-08-06 16:06:12 -07002962 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002963
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002964 last_pgdat = NULL;
2965 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2966 sc->nodemask) {
2967 if (zone->zone_pgdat == last_pgdat)
2968 continue;
2969 last_pgdat = zone->zone_pgdat;
2970 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
Andrey Ryabinine3c1ac52018-04-10 16:28:03 -07002971 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
Johannes Weiner2a2e4882017-05-03 14:55:03 -07002972 }
2973
Keika Kobayashi873b4772008-07-25 01:48:52 -07002974 delayacct_freepages_end();
2975
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002976 if (sc->nr_reclaimed)
2977 return sc->nr_reclaimed;
2978
Mel Gorman0cee34f2012-01-12 17:19:49 -08002979 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002980 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002981 return 1;
2982
Johannes Weiner241994ed2015-02-11 15:26:06 -08002983 /* Untapped cgroup reserves? Don't OOM, retry. */
Yisheng Xied6622f62017-05-03 14:53:57 -07002984 if (sc->memcg_low_skipped) {
Johannes Weiner241994ed2015-02-11 15:26:06 -08002985 sc->priority = initial_priority;
Yisheng Xied6622f62017-05-03 14:53:57 -07002986 sc->memcg_low_reclaim = 1;
2987 sc->memcg_low_skipped = 0;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002988 goto retry;
2989 }
2990
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002991 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992}
2993
Johannes Weinerc73322d2017-05-03 14:51:51 -07002994static bool allow_direct_reclaim(pg_data_t *pgdat)
Mel Gorman55150612012-07-31 16:44:35 -07002995{
2996 struct zone *zone;
2997 unsigned long pfmemalloc_reserve = 0;
2998 unsigned long free_pages = 0;
2999 int i;
3000 bool wmark_ok;
3001
Johannes Weinerc73322d2017-05-03 14:51:51 -07003002 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3003 return true;
3004
Mel Gorman55150612012-07-31 16:44:35 -07003005 for (i = 0; i <= ZONE_NORMAL; i++) {
3006 zone = &pgdat->node_zones[i];
Johannes Weinerd450abd82017-05-03 14:51:54 -07003007 if (!managed_zone(zone))
3008 continue;
3009
3010 if (!zone_reclaimable_pages(zone))
Mel Gorman675becc2014-06-04 16:07:35 -07003011 continue;
3012
Mel Gorman55150612012-07-31 16:44:35 -07003013 pfmemalloc_reserve += min_wmark_pages(zone);
3014 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3015 }
3016
Mel Gorman675becc2014-06-04 16:07:35 -07003017 /* If there are no reserves (unexpected config) then do not throttle */
3018 if (!pfmemalloc_reserve)
3019 return true;
3020
Mel Gorman55150612012-07-31 16:44:35 -07003021 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3022
3023 /* kswapd must be awake if processes are being throttled */
3024 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
Mel Gorman38087d92016-07-28 15:45:49 -07003025 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
Mel Gorman55150612012-07-31 16:44:35 -07003026 (enum zone_type)ZONE_NORMAL);
3027 wake_up_interruptible(&pgdat->kswapd_wait);
3028 }
3029
3030 return wmark_ok;
3031}
3032
3033/*
3034 * Throttle direct reclaimers if backing storage is backed by the network
3035 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3036 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08003037 * when the low watermark is reached.
3038 *
3039 * Returns true if a fatal signal was delivered during throttling. If this
3040 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07003041 */
Mel Gorman50694c22012-11-26 16:29:48 -08003042static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07003043 nodemask_t *nodemask)
3044{
Mel Gorman675becc2014-06-04 16:07:35 -07003045 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07003046 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07003047 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07003048
3049 /*
3050 * Kernel threads should not be throttled as they may be indirectly
3051 * responsible for cleaning pages necessary for reclaim to make forward
3052 * progress. kjournald for example may enter direct reclaim while
3053 * committing a transaction where throttling it could forcing other
3054 * processes to block on log_wait_commit().
3055 */
3056 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08003057 goto out;
3058
3059 /*
3060 * If a fatal signal is pending, this process should not throttle.
3061 * It should return quickly so it can exit and free its memory
3062 */
3063 if (fatal_signal_pending(current))
3064 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003065
Mel Gorman675becc2014-06-04 16:07:35 -07003066 /*
3067 * Check if the pfmemalloc reserves are ok by finding the first node
3068 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3069 * GFP_KERNEL will be required for allocating network buffers when
3070 * swapping over the network so ZONE_HIGHMEM is unusable.
3071 *
3072 * Throttling is based on the first usable node and throttled processes
3073 * wait on a queue until kswapd makes progress and wakes them. There
3074 * is an affinity then between processes waking up and where reclaim
3075 * progress has been made assuming the process wakes on the same node.
3076 * More importantly, processes running on remote nodes will not compete
3077 * for remote pfmemalloc reserves and processes on different nodes
3078 * should make reasonable progress.
3079 */
3080 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08003081 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07003082 if (zone_idx(zone) > ZONE_NORMAL)
3083 continue;
3084
3085 /* Throttle based on the first usable node */
3086 pgdat = zone->zone_pgdat;
Johannes Weinerc73322d2017-05-03 14:51:51 -07003087 if (allow_direct_reclaim(pgdat))
Mel Gorman675becc2014-06-04 16:07:35 -07003088 goto out;
3089 break;
3090 }
3091
3092 /* If no zone was usable by the allocation flags then do not throttle */
3093 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08003094 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07003095
Mel Gorman68243e72012-07-31 16:44:39 -07003096 /* Account for the throttling */
3097 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3098
Mel Gorman55150612012-07-31 16:44:35 -07003099 /*
3100 * If the caller cannot enter the filesystem, it's possible that it
3101 * is due to the caller holding an FS lock or performing a journal
3102 * transaction in the case of a filesystem like ext[3|4]. In this case,
3103 * it is not safe to block on pfmemalloc_wait as kswapd could be
3104 * blocked waiting on the same lock. Instead, throttle for up to a
3105 * second before continuing.
3106 */
3107 if (!(gfp_mask & __GFP_FS)) {
3108 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003109 allow_direct_reclaim(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08003110
3111 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07003112 }
3113
3114 /* Throttle until kswapd wakes the process */
3115 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
Johannes Weinerc73322d2017-05-03 14:51:51 -07003116 allow_direct_reclaim(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08003117
3118check_pending:
3119 if (fatal_signal_pending(current))
3120 return true;
3121
3122out:
3123 return false;
Mel Gorman55150612012-07-31 16:44:35 -07003124}
3125
Mel Gormandac1d272008-04-28 02:12:12 -07003126unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07003127 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08003128{
Mel Gorman33906bc2010-08-09 17:19:16 -07003129 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003130 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08003131 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003132 .gfp_mask = current_gfp_context(gfp_mask),
Mel Gormanb2e18752016-07-28 15:45:37 -07003133 .reclaim_idx = gfp_zone(gfp_mask),
Johannes Weineree814fe2014-08-06 16:06:19 -07003134 .order = order,
3135 .nodemask = nodemask,
3136 .priority = DEF_PRIORITY,
3137 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003138 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003139 .may_swap = 1,
Mel Gorman1c308442018-12-28 00:35:52 -08003140 .may_shrinkslab = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08003141 };
3142
Mel Gorman55150612012-07-31 16:44:35 -07003143 /*
Greg Thelenbb451fd2018-08-17 15:45:19 -07003144 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3145 * Confirm they are large enough for max values.
3146 */
3147 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3148 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3149 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3150
3151 /*
Mel Gorman50694c22012-11-26 16:29:48 -08003152 * Do not enter reclaim if fatal signal was delivered while throttled.
3153 * 1 is returned so that the page allocator does not OOM kill at this
3154 * point.
Mel Gorman55150612012-07-31 16:44:35 -07003155 */
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07003156 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07003157 return 1;
3158
Yafang Shao3481c372019-05-13 17:19:14 -07003159 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
Mel Gorman33906bc2010-08-09 17:19:16 -07003160
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003161 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07003162
3163 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3164
3165 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003166}
3167
Andrew Mortonc255a452012-07-31 16:43:02 -07003168#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08003169
Mel Gormana9dd0a82016-07-28 15:46:02 -07003170unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003171 gfp_t gfp_mask, bool noswap,
Mel Gormanef8f2322016-07-28 15:46:05 -07003172 pg_data_t *pgdat,
Ying Han0ae5e892011-05-26 16:25:25 -07003173 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07003174{
3175 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07003176 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07003177 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07003178 .may_writepage = !laptop_mode,
3179 .may_unmap = 1,
Mel Gormanb2e18752016-07-28 15:45:37 -07003180 .reclaim_idx = MAX_NR_ZONES - 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003181 .may_swap = !noswap,
Mel Gorman1c308442018-12-28 00:35:52 -08003182 .may_shrinkslab = 1,
Balbir Singh4e416952009-09-23 15:56:39 -07003183 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003184 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07003185
Balbir Singh4e416952009-09-23 15:56:39 -07003186 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3187 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003188
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003189 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
Yafang Shao3481c372019-05-13 17:19:14 -07003190 sc.gfp_mask);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003191
Balbir Singh4e416952009-09-23 15:56:39 -07003192 /*
3193 * NOTE: Although we can get the priority field, using it
3194 * here is not a good idea, since it limits the pages we can scan.
Mel Gormana9dd0a82016-07-28 15:46:02 -07003195 * if we don't reclaim here, the shrink_node from balance_pgdat
Balbir Singh4e416952009-09-23 15:56:39 -07003196 * will pick up pages from other mem cgroup's as well. We hack
3197 * the priority and make it zero.
3198 */
Mel Gormanef8f2322016-07-28 15:46:05 -07003199 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003200
3201 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3202
Ying Han0ae5e892011-05-26 16:25:25 -07003203 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07003204 return sc.nr_reclaimed;
3205}
3206
Johannes Weiner72835c82012-01-12 17:18:32 -08003207unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003208 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08003209 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003210 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08003211{
Balbir Singh4e416952009-09-23 15:56:39 -07003212 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003213 unsigned long nr_reclaimed;
Johannes Weinereb414682018-10-26 15:06:27 -07003214 unsigned long pflags;
Ying Han889976d2011-05-26 16:25:33 -07003215 int nid;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003216 unsigned int noreclaim_flag;
Balbir Singh66e17072008-02-07 00:13:56 -08003217 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003218 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Michal Hocko7dea19f2017-05-03 14:53:15 -07003219 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
Johannes Weineree814fe2014-08-06 16:06:19 -07003220 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
Mel Gormanb2e18752016-07-28 15:45:37 -07003221 .reclaim_idx = MAX_NR_ZONES - 1,
Johannes Weineree814fe2014-08-06 16:06:19 -07003222 .target_mem_cgroup = memcg,
3223 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08003224 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003225 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07003226 .may_swap = may_swap,
Mel Gorman1c308442018-12-28 00:35:52 -08003227 .may_shrinkslab = 1,
Ying Hana09ed5e2011-05-24 17:12:26 -07003228 };
Balbir Singh66e17072008-02-07 00:13:56 -08003229
Ying Han889976d2011-05-26 16:25:33 -07003230 /*
3231 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3232 * take care of from where we get pages. So the node where we start the
3233 * scan does not need to be the current node.
3234 */
Johannes Weiner72835c82012-01-12 17:18:32 -08003235 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07003236
Aneesh Kumar K.Vc9634cf2016-10-07 16:59:12 -07003237 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003238
Yafang Shao3481c372019-05-13 17:19:14 -07003239 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003240
Johannes Weinereb414682018-10-26 15:06:27 -07003241 psi_memstall_enter(&pflags);
Vlastimil Babka499118e2017-05-08 15:59:50 -07003242 noreclaim_flag = memalloc_noreclaim_save();
Johannes Weinereb414682018-10-26 15:06:27 -07003243
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003244 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Johannes Weinereb414682018-10-26 15:06:27 -07003245
Vlastimil Babka499118e2017-05-08 15:59:50 -07003246 memalloc_noreclaim_restore(noreclaim_flag);
Johannes Weinereb414682018-10-26 15:06:27 -07003247 psi_memstall_leave(&pflags);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07003248
3249 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3250
3251 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08003252}
3253#endif
3254
Mel Gorman1d82de62016-07-28 15:45:43 -07003255static void age_active_anon(struct pglist_data *pgdat,
Mel Gormanef8f2322016-07-28 15:46:05 -07003256 struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08003257{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003258 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08003259
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003260 if (!total_swap_pages)
3261 return;
3262
3263 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3264 do {
Mel Gormanef8f2322016-07-28 15:46:05 -07003265 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003266
Johannes Weiner3b991202019-04-18 17:50:34 -07003267 if (inactive_list_is_low(lruvec, false, sc, true))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07003268 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003269 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08003270
3271 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3272 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08003273}
3274
Mel Gorman1c308442018-12-28 00:35:52 -08003275static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3276{
3277 int i;
3278 struct zone *zone;
3279
3280 /*
3281 * Check for watermark boosts top-down as the higher zones
3282 * are more likely to be boosted. Both watermarks and boosts
3283 * should not be checked at the time time as reclaim would
3284 * start prematurely when there is no boosting and a lower
3285 * zone is balanced.
3286 */
3287 for (i = classzone_idx; i >= 0; i--) {
3288 zone = pgdat->node_zones + i;
3289 if (!managed_zone(zone))
3290 continue;
3291
3292 if (zone->watermark_boost)
3293 return true;
3294 }
3295
3296 return false;
3297}
3298
Mel Gormane716f2e2017-05-03 14:53:45 -07003299/*
3300 * Returns true if there is an eligible zone balanced for the request order
3301 * and classzone_idx
3302 */
3303static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08003304{
Mel Gormane716f2e2017-05-03 14:53:45 -07003305 int i;
3306 unsigned long mark = -1;
3307 struct zone *zone;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003308
Mel Gorman1c308442018-12-28 00:35:52 -08003309 /*
3310 * Check watermarks bottom-up as lower zones are more likely to
3311 * meet watermarks.
3312 */
Mel Gormane716f2e2017-05-03 14:53:45 -07003313 for (i = 0; i <= classzone_idx; i++) {
3314 zone = pgdat->node_zones + i;
Mel Gorman6256c6b2016-07-28 15:45:56 -07003315
Mel Gormane716f2e2017-05-03 14:53:45 -07003316 if (!managed_zone(zone))
3317 continue;
3318
3319 mark = high_wmark_pages(zone);
3320 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3321 return true;
3322 }
3323
3324 /*
3325 * If a node has no populated zone within classzone_idx, it does not
3326 * need balancing by definition. This can happen if a zone-restricted
3327 * allocation tries to wake a remote kswapd.
3328 */
3329 if (mark == -1)
3330 return true;
3331
3332 return false;
Johannes Weiner60cefed2012-11-29 13:54:23 -08003333}
3334
Mel Gorman631b6e02017-05-03 14:53:41 -07003335/* Clear pgdat state for congested, dirty or under writeback. */
3336static void clear_pgdat_congested(pg_data_t *pgdat)
3337{
3338 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3339 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3340 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3341}
3342
Mel Gorman1741c872011-01-13 15:46:21 -08003343/*
Mel Gorman55150612012-07-31 16:44:35 -07003344 * Prepare kswapd for sleeping. This verifies that there are no processes
3345 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3346 *
3347 * Returns true if kswapd is ready to sleep
3348 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003349static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003350{
Mel Gorman55150612012-07-31 16:44:35 -07003351 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003352 * The throttled processes are normally woken up in balance_pgdat() as
Johannes Weinerc73322d2017-05-03 14:51:51 -07003353 * soon as allow_direct_reclaim() is true. But there is a potential
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003354 * race between when kswapd checks the watermarks and a process gets
3355 * throttled. There is also a potential race if processes get
3356 * throttled, kswapd wakes, a large process exits thereby balancing the
3357 * zones, which causes kswapd to exit balance_pgdat() before reaching
3358 * the wake up checks. If kswapd is going to sleep, no process should
3359 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3360 * the wake up is premature, processes will wake kswapd and get
3361 * throttled again. The difference from wake ups in balance_pgdat() is
3362 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003363 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003364 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3365 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003366
Johannes Weinerc73322d2017-05-03 14:51:51 -07003367 /* Hopeless node, leave it to direct reclaim */
3368 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3369 return true;
3370
Mel Gormane716f2e2017-05-03 14:53:45 -07003371 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3372 clear_pgdat_congested(pgdat);
3373 return true;
Mel Gorman1d82de62016-07-28 15:45:43 -07003374 }
3375
Shantanu Goel333b0a42017-05-03 14:53:38 -07003376 return false;
Mel Gormanf50de2d2009-12-14 17:58:53 -08003377}
3378
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003380 * kswapd shrinks a node of pages that are at or below the highest usable
3381 * zone that is currently unbalanced.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003382 *
3383 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003384 * reclaim or if the lack of progress was due to pages under writeback.
3385 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003386 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003387static bool kswapd_shrink_node(pg_data_t *pgdat,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003388 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003389{
Mel Gorman1d82de62016-07-28 15:45:43 -07003390 struct zone *zone;
3391 int z;
Mel Gorman75485362013-07-03 15:01:42 -07003392
Mel Gorman1d82de62016-07-28 15:45:43 -07003393 /* Reclaim a number of pages proportional to the number of zones */
3394 sc->nr_to_reclaim = 0;
Mel Gorman970a39a2016-07-28 15:46:35 -07003395 for (z = 0; z <= sc->reclaim_idx; z++) {
Mel Gorman1d82de62016-07-28 15:45:43 -07003396 zone = pgdat->node_zones + z;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003397 if (!managed_zone(zone))
Mel Gorman1d82de62016-07-28 15:45:43 -07003398 continue;
Mel Gorman7c954f62013-07-03 15:01:54 -07003399
Mel Gorman1d82de62016-07-28 15:45:43 -07003400 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
Mel Gorman7c954f62013-07-03 15:01:54 -07003401 }
3402
Mel Gorman1d82de62016-07-28 15:45:43 -07003403 /*
3404 * Historically care was taken to put equal pressure on all zones but
3405 * now pressure is applied based on node LRU order.
3406 */
Mel Gorman970a39a2016-07-28 15:46:35 -07003407 shrink_node(pgdat, sc);
Mel Gorman1d82de62016-07-28 15:45:43 -07003408
3409 /*
3410 * Fragmentation may mean that the system cannot be rebalanced for
3411 * high-order allocations. If twice the allocation size has been
3412 * reclaimed then recheck watermarks only at order-0 to prevent
3413 * excessive reclaim. Assume that a process requested a high-order
3414 * can direct reclaim/compact.
3415 */
Vlastimil Babka9861a622016-10-07 16:57:53 -07003416 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
Mel Gorman1d82de62016-07-28 15:45:43 -07003417 sc->order = 0;
3418
Mel Gormanb8e83b92013-07-03 15:01:45 -07003419 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003420}
3421
3422/*
Mel Gorman1d82de62016-07-28 15:45:43 -07003423 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3424 * that are eligible for use by the caller until at least one zone is
3425 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 *
Mel Gorman1d82de62016-07-28 15:45:43 -07003427 * Returns the order kswapd finished reclaiming at.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 *
3429 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003430 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
Wei Yang8bb4e7a2019-03-05 15:46:22 -08003431 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
Mel Gorman1d82de62016-07-28 15:45:43 -07003432 * or lower is eligible for reclaim until at least one usable zone is
3433 * balanced.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003435static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437 int i;
Andrew Morton0608f432013-09-24 15:27:41 -07003438 unsigned long nr_soft_reclaimed;
3439 unsigned long nr_soft_scanned;
Johannes Weinereb414682018-10-26 15:06:27 -07003440 unsigned long pflags;
Mel Gorman1c308442018-12-28 00:35:52 -08003441 unsigned long nr_boost_reclaim;
3442 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3443 bool boosted;
Mel Gorman1d82de62016-07-28 15:45:43 -07003444 struct zone *zone;
Andrew Morton179e9632006-03-22 00:08:18 -08003445 struct scan_control sc = {
3446 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003447 .order = order,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003448 .may_unmap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003449 };
Omar Sandoval93781322018-06-07 17:07:02 -07003450
Johannes Weinereb414682018-10-26 15:06:27 -07003451 psi_memstall_enter(&pflags);
Omar Sandoval93781322018-06-07 17:07:02 -07003452 __fs_reclaim_acquire();
3453
Christoph Lameterf8891e52006-06-30 01:55:45 -07003454 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455
Mel Gorman1c308442018-12-28 00:35:52 -08003456 /*
3457 * Account for the reclaim boost. Note that the zone boost is left in
3458 * place so that parallel allocations that are near the watermark will
3459 * stall or direct reclaim until kswapd is finished.
3460 */
3461 nr_boost_reclaim = 0;
3462 for (i = 0; i <= classzone_idx; i++) {
3463 zone = pgdat->node_zones + i;
3464 if (!managed_zone(zone))
3465 continue;
3466
3467 nr_boost_reclaim += zone->watermark_boost;
3468 zone_boosts[i] = zone->watermark_boost;
3469 }
3470 boosted = nr_boost_reclaim;
3471
3472restart:
3473 sc.priority = DEF_PRIORITY;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003474 do {
Johannes Weinerc73322d2017-05-03 14:51:51 -07003475 unsigned long nr_reclaimed = sc.nr_reclaimed;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003476 bool raise_priority = true;
Mel Gorman1c308442018-12-28 00:35:52 -08003477 bool balanced;
Omar Sandoval93781322018-06-07 17:07:02 -07003478 bool ret;
Mel Gormanb8e83b92013-07-03 15:01:45 -07003479
Mel Gorman84c7a772016-07-28 15:46:44 -07003480 sc.reclaim_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481
Mel Gorman86c79f62016-07-28 15:45:59 -07003482 /*
Mel Gorman84c7a772016-07-28 15:46:44 -07003483 * If the number of buffer_heads exceeds the maximum allowed
3484 * then consider reclaiming from all zones. This has a dual
3485 * purpose -- on 64-bit systems it is expected that
3486 * buffer_heads are stripped during active rotation. On 32-bit
3487 * systems, highmem pages can pin lowmem memory and shrinking
3488 * buffers can relieve lowmem pressure. Reclaim may still not
3489 * go ahead if all eligible zones for the original allocation
3490 * request are balanced to avoid excessive reclaim from kswapd.
Mel Gorman86c79f62016-07-28 15:45:59 -07003491 */
3492 if (buffer_heads_over_limit) {
3493 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3494 zone = pgdat->node_zones + i;
Mel Gorman6aa303d2016-09-01 16:14:55 -07003495 if (!managed_zone(zone))
Mel Gorman86c79f62016-07-28 15:45:59 -07003496 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497
Mel Gorman970a39a2016-07-28 15:46:35 -07003498 sc.reclaim_idx = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003499 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003502
Mel Gorman86c79f62016-07-28 15:45:59 -07003503 /*
Mel Gorman1c308442018-12-28 00:35:52 -08003504 * If the pgdat is imbalanced then ignore boosting and preserve
3505 * the watermarks for a later time and restart. Note that the
3506 * zone watermarks will be still reset at the end of balancing
3507 * on the grounds that the normal reclaim should be enough to
3508 * re-evaluate if boosting is required when kswapd next wakes.
Mel Gorman86c79f62016-07-28 15:45:59 -07003509 */
Mel Gorman1c308442018-12-28 00:35:52 -08003510 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3511 if (!balanced && nr_boost_reclaim) {
3512 nr_boost_reclaim = 0;
3513 goto restart;
3514 }
3515
3516 /*
3517 * If boosting is not active then only reclaim if there are no
3518 * eligible zones. Note that sc.reclaim_idx is not used as
3519 * buffer_heads_over_limit may have adjusted it.
3520 */
3521 if (!nr_boost_reclaim && balanced)
Mel Gormane716f2e2017-05-03 14:53:45 -07003522 goto out;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003523
Mel Gorman1c308442018-12-28 00:35:52 -08003524 /* Limit the priority of boosting to avoid reclaim writeback */
3525 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3526 raise_priority = false;
3527
3528 /*
3529 * Do not writeback or swap pages for boosted reclaim. The
3530 * intent is to relieve pressure not issue sub-optimal IO
3531 * from reclaim context. If no pages are reclaimed, the
3532 * reclaim will be aborted.
3533 */
3534 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3535 sc.may_swap = !nr_boost_reclaim;
3536 sc.may_shrinkslab = !nr_boost_reclaim;
3537
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;
Mel Gorman1c308442018-12-28 00:35:52 -08003589 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3590
3591 /*
3592 * If reclaim made no progress for a boost, stop reclaim as
3593 * IO cannot be queued and it could be an infinite loop in
3594 * extreme circumstances.
3595 */
3596 if (nr_boost_reclaim && !nr_reclaimed)
3597 break;
3598
Johannes Weinerc73322d2017-05-03 14:51:51 -07003599 if (raise_priority || !nr_reclaimed)
Mel Gormanb8e83b92013-07-03 15:01:45 -07003600 sc.priority--;
Mel Gorman1d82de62016-07-28 15:45:43 -07003601 } while (sc.priority >= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602
Johannes Weinerc73322d2017-05-03 14:51:51 -07003603 if (!sc.nr_reclaimed)
3604 pgdat->kswapd_failures++;
3605
Mel Gormanb8e83b92013-07-03 15:01:45 -07003606out:
Mel Gorman1c308442018-12-28 00:35:52 -08003607 /* If reclaim was boosted, account for the reclaim done in this pass */
3608 if (boosted) {
3609 unsigned long flags;
3610
3611 for (i = 0; i <= classzone_idx; i++) {
3612 if (!zone_boosts[i])
3613 continue;
3614
3615 /* Increments are under the zone lock */
3616 zone = pgdat->node_zones + i;
3617 spin_lock_irqsave(&zone->lock, flags);
3618 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3619 spin_unlock_irqrestore(&zone->lock, flags);
3620 }
3621
3622 /*
3623 * As there is now likely space, wakeup kcompact to defragment
3624 * pageblocks.
3625 */
3626 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3627 }
3628
Johannes Weiner2a2e4882017-05-03 14:55:03 -07003629 snapshot_refaults(NULL, pgdat);
Omar Sandoval93781322018-06-07 17:07:02 -07003630 __fs_reclaim_release();
Johannes Weinereb414682018-10-26 15:06:27 -07003631 psi_memstall_leave(&pflags);
Mel Gorman0abdee22011-01-13 15:46:22 -08003632 /*
Mel Gorman1d82de62016-07-28 15:45:43 -07003633 * Return the order kswapd stopped reclaiming at as
3634 * prepare_kswapd_sleep() takes it into account. If another caller
3635 * entered the allocator slow path while kswapd was awake, order will
3636 * remain at the higher level.
Mel Gorman0abdee22011-01-13 15:46:22 -08003637 */
Mel Gorman1d82de62016-07-28 15:45:43 -07003638 return sc.order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639}
3640
Mel Gormane716f2e2017-05-03 14:53:45 -07003641/*
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003642 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3643 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3644 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3645 * after previous reclaim attempt (node is still unbalanced). In that case
3646 * return the zone index of the previous kswapd reclaim cycle.
Mel Gormane716f2e2017-05-03 14:53:45 -07003647 */
3648static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003649 enum zone_type prev_classzone_idx)
Mel Gormane716f2e2017-05-03 14:53:45 -07003650{
3651 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003652 return prev_classzone_idx;
3653 return pgdat->kswapd_classzone_idx;
Mel Gormane716f2e2017-05-03 14:53:45 -07003654}
3655
Mel Gorman38087d92016-07-28 15:45:49 -07003656static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3657 unsigned int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003658{
3659 long remaining = 0;
3660 DEFINE_WAIT(wait);
3661
3662 if (freezing(current) || kthread_should_stop())
3663 return;
3664
3665 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3666
Shantanu Goel333b0a42017-05-03 14:53:38 -07003667 /*
3668 * Try to sleep for a short interval. Note that kcompactd will only be
3669 * woken if it is possible to sleep for a short interval. This is
3670 * deliberate on the assumption that if reclaim cannot keep an
3671 * eligible zone balanced that it's also unlikely that compaction will
3672 * succeed.
3673 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003674 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003675 /*
3676 * Compaction records what page blocks it recently failed to
3677 * isolate pages from and skips them in the future scanning.
3678 * When kswapd is going to sleep, it is reasonable to assume
3679 * that pages and compaction may succeed so reset the cache.
3680 */
3681 reset_isolation_suitable(pgdat);
3682
3683 /*
3684 * We have freed the memory, now we should compact it to make
3685 * allocation of the requested order possible.
3686 */
Mel Gorman38087d92016-07-28 15:45:49 -07003687 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
Vlastimil Babkafd901c92016-04-28 16:18:49 -07003688
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003689 remaining = schedule_timeout(HZ/10);
Mel Gorman38087d92016-07-28 15:45:49 -07003690
3691 /*
3692 * If woken prematurely then reset kswapd_classzone_idx and
3693 * order. The values will either be from a wakeup request or
3694 * the previous request that slept prematurely.
3695 */
3696 if (remaining) {
Mel Gormane716f2e2017-05-03 14:53:45 -07003697 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003698 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3699 }
3700
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003701 finish_wait(&pgdat->kswapd_wait, &wait);
3702 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3703 }
3704
3705 /*
3706 * After a short sleep, check if it was a premature sleep. If not, then
3707 * go fully to sleep until explicitly woken up.
3708 */
Mel Gormand9f21d42016-07-28 15:46:41 -07003709 if (!remaining &&
3710 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003711 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3712
3713 /*
3714 * vmstat counters are not perfectly accurate and the estimated
3715 * value for counters such as NR_FREE_PAGES can deviate from the
3716 * true value by nr_online_cpus * threshold. To avoid the zone
3717 * watermarks being breached while under pressure, we reduce the
3718 * per-cpu vmstat threshold while kswapd is awake and restore
3719 * them before going back to sleep.
3720 */
3721 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003722
3723 if (!kthread_should_stop())
3724 schedule();
3725
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003726 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3727 } else {
3728 if (remaining)
3729 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3730 else
3731 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3732 }
3733 finish_wait(&pgdat->kswapd_wait, &wait);
3734}
3735
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736/*
3737 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003738 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739 *
3740 * This basically trickles out pages so that we have _some_
3741 * free memory available even if there is no other activity
3742 * that frees anything up. This is needed for things like routing
3743 * etc, where we otherwise might have all activity going on in
3744 * asynchronous contexts that cannot page things out.
3745 *
3746 * If there are applications that are active memory-allocators
3747 * (most normal use), this basically shouldn't matter.
3748 */
3749static int kswapd(void *p)
3750{
Mel Gormane716f2e2017-05-03 14:53:45 -07003751 unsigned int alloc_order, reclaim_order;
3752 unsigned int classzone_idx = MAX_NR_ZONES - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 pg_data_t *pgdat = (pg_data_t*)p;
3754 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003755
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 struct reclaim_state reclaim_state = {
3757 .reclaimed_slab = 0,
3758 };
Rusty Russella70f7302009-03-13 14:49:46 +10303759 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760
Rusty Russell174596a2009-01-01 10:12:29 +10303761 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003762 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 current->reclaim_state = &reclaim_state;
3764
3765 /*
3766 * Tell the memory management that we're a "memory allocator",
3767 * and that if we need more memory we should get access to it
3768 * regardless (see "__alloc_pages()"). "kswapd" should
3769 * never get caught in the normal page freeing logic.
3770 *
3771 * (Kswapd normally doesn't need memory anyway, but sometimes
3772 * you need a small amount of memory in order to be able to
3773 * page out something else, and this flag essentially protects
3774 * us from recursively trying to free more memory as we're
3775 * trying to free the first piece of memory in the first place).
3776 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003777 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003778 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779
Mel Gormane716f2e2017-05-03 14:53:45 -07003780 pgdat->kswapd_order = 0;
3781 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003783 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003784
Mel Gormane716f2e2017-05-03 14:53:45 -07003785 alloc_order = reclaim_order = pgdat->kswapd_order;
3786 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3787
Mel Gorman38087d92016-07-28 15:45:49 -07003788kswapd_try_sleep:
3789 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3790 classzone_idx);
Mel Gorman215ddd62011-07-08 15:39:40 -07003791
Mel Gorman38087d92016-07-28 15:45:49 -07003792 /* Read the new order and classzone_idx */
3793 alloc_order = reclaim_order = pgdat->kswapd_order;
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003794 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003795 pgdat->kswapd_order = 0;
Mel Gormane716f2e2017-05-03 14:53:45 -07003796 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797
David Rientjes8fe23e02009-12-14 17:58:33 -08003798 ret = try_to_freeze();
3799 if (kthread_should_stop())
3800 break;
3801
3802 /*
3803 * We can speed up thawing tasks if we don't call balance_pgdat
3804 * after returning from the refrigerator
3805 */
Mel Gorman38087d92016-07-28 15:45:49 -07003806 if (ret)
3807 continue;
Mel Gorman1d82de62016-07-28 15:45:43 -07003808
Mel Gorman38087d92016-07-28 15:45:49 -07003809 /*
3810 * Reclaim begins at the requested order but if a high-order
3811 * reclaim fails then kswapd falls back to reclaiming for
3812 * order-0. If that happens, kswapd will consider sleeping
3813 * for the order it finished reclaiming at (reclaim_order)
3814 * but kcompactd is woken to compact for the original
3815 * request (alloc_order).
3816 */
Mel Gormane5146b12016-07-28 15:46:47 -07003817 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3818 alloc_order);
Mel Gorman38087d92016-07-28 15:45:49 -07003819 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3820 if (reclaim_order < alloc_order)
3821 goto kswapd_try_sleep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003823
Johannes Weiner71abdc12014-06-06 14:35:35 -07003824 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003825 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003826
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827 return 0;
3828}
3829
3830/*
David Rientjes5ecd9d42018-04-05 16:25:16 -07003831 * A zone is low on free memory or too fragmented for high-order memory. If
3832 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3833 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3834 * has failed or is not needed, still wake up kcompactd if only compaction is
3835 * needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 */
David Rientjes5ecd9d42018-04-05 16:25:16 -07003837void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3838 enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839{
3840 pg_data_t *pgdat;
3841
Mel Gorman6aa303d2016-09-01 16:14:55 -07003842 if (!managed_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843 return;
3844
David Rientjes5ecd9d42018-04-05 16:25:16 -07003845 if (!cpuset_zone_allowed(zone, gfp_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003847 pgdat = zone->zone_pgdat;
Shakeel Buttdffcac2c2019-07-04 15:14:42 -07003848
3849 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3850 pgdat->kswapd_classzone_idx = classzone_idx;
3851 else
3852 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3853 classzone_idx);
Mel Gorman38087d92016-07-28 15:45:49 -07003854 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003855 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 return;
Mel Gormane1a55632016-07-28 15:46:26 -07003857
David Rientjes5ecd9d42018-04-05 16:25:16 -07003858 /* Hopeless node, leave it to direct reclaim if possible */
3859 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
Mel Gorman1c308442018-12-28 00:35:52 -08003860 (pgdat_balanced(pgdat, order, classzone_idx) &&
3861 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
David Rientjes5ecd9d42018-04-05 16:25:16 -07003862 /*
3863 * There may be plenty of free memory available, but it's too
3864 * fragmented for high-order allocations. Wake up kcompactd
3865 * and rely on compaction_suitable() to determine if it's
3866 * needed. If it fails, it will defer subsequent attempts to
3867 * ratelimit its work.
3868 */
3869 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3870 wakeup_kcompactd(pgdat, order, classzone_idx);
Johannes Weinerc73322d2017-05-03 14:51:51 -07003871 return;
David Rientjes5ecd9d42018-04-05 16:25:16 -07003872 }
Johannes Weinerc73322d2017-05-03 14:51:51 -07003873
David Rientjes5ecd9d42018-04-05 16:25:16 -07003874 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3875 gfp_flags);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003876 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877}
3878
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003879#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003881 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003882 * freed pages.
3883 *
3884 * Rather than trying to age LRUs the aim is to preserve the overall
3885 * LRU order by reclaiming preferentially
3886 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003888unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003890 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003891 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003892 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003893 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Mel Gormanb2e18752016-07-28 15:45:37 -07003894 .reclaim_idx = MAX_NR_ZONES - 1,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003895 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003896 .may_writepage = 1,
3897 .may_unmap = 1,
3898 .may_swap = 1,
3899 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003901 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003902 struct task_struct *p = current;
3903 unsigned long nr_reclaimed;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003904 unsigned int noreclaim_flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905
Peter Zijlstrad92a8cf2017-03-03 10:13:38 +01003906 fs_reclaim_acquire(sc.gfp_mask);
Omar Sandoval93781322018-06-07 17:07:02 -07003907 noreclaim_flag = memalloc_noreclaim_save();
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003908 reclaim_state.reclaimed_slab = 0;
3909 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003910
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003911 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003912
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003913 p->reclaim_state = NULL;
Vlastimil Babka499118e2017-05-08 15:59:50 -07003914 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07003915 fs_reclaim_release(sc.gfp_mask);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003916
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003917 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003919#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921/* It's optimal to keep kswapds on the same CPUs as their memory, but
3922 not required for correctness. So if the last cpu in a node goes
3923 away, we get changed to run anywhere: as the first one comes back,
3924 restore their cpu bindings. */
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003925static int kswapd_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003927 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003929 for_each_node_state(nid, N_MEMORY) {
3930 pg_data_t *pgdat = NODE_DATA(nid);
3931 const struct cpumask *mask;
Rusty Russella70f7302009-03-13 14:49:46 +10303932
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003933 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003934
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003935 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3936 /* One of our CPUs online: restore mask */
3937 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938 }
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003939 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941
Yasunori Goto3218ae12006-06-27 02:53:33 -07003942/*
3943 * This kswapd start function will be called by init and node-hot-add.
3944 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3945 */
3946int kswapd_run(int nid)
3947{
3948 pg_data_t *pgdat = NODE_DATA(nid);
3949 int ret = 0;
3950
3951 if (pgdat->kswapd)
3952 return 0;
3953
3954 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3955 if (IS_ERR(pgdat->kswapd)) {
3956 /* failure at boot is fatal */
Thomas Gleixnerc6202ad2017-05-16 20:42:46 +02003957 BUG_ON(system_state < SYSTEM_RUNNING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003958 pr_err("Failed to start kswapd on node %d\n", nid);
3959 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003960 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003961 }
3962 return ret;
3963}
3964
David Rientjes8fe23e02009-12-14 17:58:33 -08003965/*
Jiang Liud8adde12012-07-11 14:01:52 -07003966 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003967 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003968 */
3969void kswapd_stop(int nid)
3970{
3971 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3972
Jiang Liud8adde12012-07-11 14:01:52 -07003973 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003974 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003975 NODE_DATA(nid)->kswapd = NULL;
3976 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003977}
3978
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979static int __init kswapd_init(void)
3980{
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003981 int nid, ret;
Andrew Morton69e05942006-03-22 00:08:19 -08003982
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003984 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003985 kswapd_run(nid);
Sebastian Andrzej Siewior517bbed2016-11-03 15:50:03 +01003986 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3987 "mm/vmscan:online", kswapd_cpu_online,
3988 NULL);
3989 WARN_ON(ret < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 return 0;
3991}
3992
3993module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003994
3995#ifdef CONFIG_NUMA
3996/*
Mel Gormana5f5f912016-07-28 15:46:32 -07003997 * Node reclaim mode
Christoph Lameter9eeff232006-01-18 17:42:31 -08003998 *
Mel Gormana5f5f912016-07-28 15:46:32 -07003999 * If non-zero call node_reclaim when the number of free pages falls below
Christoph Lameter9eeff232006-01-18 17:42:31 -08004000 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004001 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004002int node_reclaim_mode __read_mostly;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004003
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08004004#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07004005#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08004006#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004007#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08004008
Christoph Lameter9eeff232006-01-18 17:42:31 -08004009/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004010 * Priority for NODE_RECLAIM. This determines the fraction of pages
Christoph Lametera92f7122006-02-01 03:05:32 -08004011 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4012 * a zone.
4013 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004014#define NODE_RECLAIM_PRIORITY 4
Christoph Lametera92f7122006-02-01 03:05:32 -08004015
Christoph Lameter9eeff232006-01-18 17:42:31 -08004016/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004017 * Percentage of pages in a zone that must be unmapped for node_reclaim to
Christoph Lameter96146342006-07-03 00:24:13 -07004018 * occur.
4019 */
4020int sysctl_min_unmapped_ratio = 1;
4021
4022/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07004023 * If the number of slab pages in a zone grows beyond this percentage then
4024 * slab reclaim needs to occur.
4025 */
4026int sysctl_min_slab_ratio = 5;
4027
Mel Gorman11fb9982016-07-28 15:46:20 -07004028static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07004029{
Mel Gorman11fb9982016-07-28 15:46:20 -07004030 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4031 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4032 node_page_state(pgdat, NR_ACTIVE_FILE);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004033
4034 /*
4035 * It's possible for there to be more file mapped pages than
4036 * accounted for by the pages on the file LRU lists because
4037 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4038 */
4039 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4040}
4041
4042/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Mel Gormana5f5f912016-07-28 15:46:32 -07004043static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
Mel Gorman90afa5d2009-06-16 15:33:20 -07004044{
Alexandru Moised031a152015-11-05 18:48:08 -08004045 unsigned long nr_pagecache_reclaimable;
4046 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07004047
4048 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004049 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07004050 * potentially reclaimable. Otherwise, we have to worry about
Mel Gorman11fb9982016-07-28 15:46:20 -07004051 * pages like swapcache and node_unmapped_file_pages() provides
Mel Gorman90afa5d2009-06-16 15:33:20 -07004052 * a better estimate
4053 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004054 if (node_reclaim_mode & RECLAIM_UNMAP)
4055 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004056 else
Mel Gormana5f5f912016-07-28 15:46:32 -07004057 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004058
4059 /* If we can't clean pages, remove dirty pages from consideration */
Mel Gormana5f5f912016-07-28 15:46:32 -07004060 if (!(node_reclaim_mode & RECLAIM_WRITE))
4061 delta += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman90afa5d2009-06-16 15:33:20 -07004062
4063 /* Watch for any possible underflows due to delta */
4064 if (unlikely(delta > nr_pagecache_reclaimable))
4065 delta = nr_pagecache_reclaimable;
4066
4067 return nr_pagecache_reclaimable - delta;
4068}
4069
Christoph Lameter0ff38492006-09-25 23:31:52 -07004070/*
Mel Gormana5f5f912016-07-28 15:46:32 -07004071 * Try to free up some pages from this node through reclaim.
Christoph Lameter9eeff232006-01-18 17:42:31 -08004072 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004073static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08004074{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08004075 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08004076 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004077 struct task_struct *p = current;
4078 struct reclaim_state reclaim_state;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004079 unsigned int noreclaim_flag;
Andrew Morton179e9632006-03-22 00:08:18 -08004080 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08004081 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004082 .gfp_mask = current_gfp_context(gfp_mask),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07004083 .order = order,
Mel Gormana5f5f912016-07-28 15:46:32 -07004084 .priority = NODE_RECLAIM_PRIORITY,
4085 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4086 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07004087 .may_swap = 1,
Nick Desaulniersf2f43e52017-07-06 15:36:50 -07004088 .reclaim_idx = gfp_zone(gfp_mask),
Andrew Morton179e9632006-03-22 00:08:18 -08004089 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08004090
Yafang Shao132bb8c2019-05-13 17:17:53 -07004091 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4092 sc.gfp_mask);
4093
Christoph Lameter9eeff232006-01-18 17:42:31 -08004094 cond_resched();
Omar Sandoval93781322018-06-07 17:07:02 -07004095 fs_reclaim_acquire(sc.gfp_mask);
Christoph Lameterd4f77962006-02-24 13:04:22 -08004096 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004097 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08004098 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07004099 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08004100 */
Vlastimil Babka499118e2017-05-08 15:59:50 -07004101 noreclaim_flag = memalloc_noreclaim_save();
4102 p->flags |= PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004103 reclaim_state.reclaimed_slab = 0;
4104 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08004105
Mel Gormana5f5f912016-07-28 15:46:32 -07004106 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07004107 /*
Andrey Ryabinin894befe2018-04-10 16:27:51 -07004108 * Free memory by calling shrink node with increasing
Christoph Lameter0ff38492006-09-25 23:31:52 -07004109 * priorities until we have enough memory freed.
4110 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07004111 do {
Mel Gorman970a39a2016-07-28 15:46:35 -07004112 shrink_node(pgdat, &sc);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07004113 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07004114 }
Christoph Lameterc84db232006-02-01 03:05:29 -08004115
Christoph Lameter9eeff232006-01-18 17:42:31 -08004116 p->reclaim_state = NULL;
Vlastimil Babka499118e2017-05-08 15:59:50 -07004117 current->flags &= ~PF_SWAPWRITE;
4118 memalloc_noreclaim_restore(noreclaim_flag);
Omar Sandoval93781322018-06-07 17:07:02 -07004119 fs_reclaim_release(sc.gfp_mask);
Yafang Shao132bb8c2019-05-13 17:17:53 -07004120
4121 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4122
Rik van Riela79311c2009-01-06 14:40:01 -08004123 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08004124}
Andrew Morton179e9632006-03-22 00:08:18 -08004125
Mel Gormana5f5f912016-07-28 15:46:32 -07004126int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
Andrew Morton179e9632006-03-22 00:08:18 -08004127{
David Rientjesd773ed62007-10-16 23:26:01 -07004128 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004129
4130 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004131 * Node reclaim reclaims unmapped file backed pages and
Christoph Lameter0ff38492006-09-25 23:31:52 -07004132 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07004133 *
Christoph Lameter96146342006-07-03 00:24:13 -07004134 * A small portion of unmapped file backed pages is needed for
4135 * file I/O otherwise pages read by file I/O will be immediately
Mel Gormana5f5f912016-07-28 15:46:32 -07004136 * thrown out if the node is overallocated. So we do not reclaim
4137 * if less than a specified percentage of the node is used by
Christoph Lameter96146342006-07-03 00:24:13 -07004138 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08004139 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004140 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
Johannes Weiner385386c2017-07-06 15:40:43 -07004141 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
Mel Gormana5f5f912016-07-28 15:46:32 -07004142 return NODE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08004143
4144 /*
David Rientjesd773ed62007-10-16 23:26:01 -07004145 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08004146 */
Mel Gormand0164ad2015-11-06 16:28:21 -08004147 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormana5f5f912016-07-28 15:46:32 -07004148 return NODE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08004149
4150 /*
Mel Gormana5f5f912016-07-28 15:46:32 -07004151 * Only run node reclaim on the local node or on nodes that do not
Andrew Morton179e9632006-03-22 00:08:18 -08004152 * have associated processors. This will favor the local processor
4153 * over remote processors and spread off node memory allocations
4154 * as wide as possible.
4155 */
Mel Gormana5f5f912016-07-28 15:46:32 -07004156 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4157 return NODE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07004158
Mel Gormana5f5f912016-07-28 15:46:32 -07004159 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4160 return NODE_RECLAIM_NOSCAN;
Mel Gormanfa5e0842009-06-16 15:33:22 -07004161
Mel Gormana5f5f912016-07-28 15:46:32 -07004162 ret = __node_reclaim(pgdat, gfp_mask, order);
4163 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07004164
Mel Gorman24cf725182009-06-16 15:33:23 -07004165 if (!ret)
4166 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4167
David Rientjesd773ed62007-10-16 23:26:01 -07004168 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08004169}
Christoph Lameter9eeff232006-01-18 17:42:31 -08004170#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004171
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004172/*
4173 * page_evictable - test whether a page is evictable
4174 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004175 *
4176 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07004177 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004178 *
4179 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07004180 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07004181 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07004182 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004183 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07004184int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004185{
Huang Yinge92bb4d2018-04-05 16:23:20 -07004186 int ret;
4187
4188 /* Prevent address_space of inode and swap cache from being freed */
4189 rcu_read_lock();
4190 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4191 rcu_read_unlock();
4192 return ret;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07004193}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004194
4195/**
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004196 * check_move_unevictable_pages - check pages for evictability and move to
4197 * appropriate zone lru list
4198 * @pvec: pagevec with lru pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004199 *
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004200 * Checks pages for evictability, if an evictable page is in the unevictable
4201 * lru list, moves it to the appropriate evictable lru list. This function
4202 * should be only used for lru pages.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004203 */
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004204void check_move_unevictable_pages(struct pagevec *pvec)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004205{
Johannes Weiner925b7672012-01-12 17:18:15 -08004206 struct lruvec *lruvec;
Mel Gorman785b99f2016-07-28 15:47:23 -07004207 struct pglist_data *pgdat = NULL;
Hugh Dickins24513262012-01-20 14:34:21 -08004208 int pgscanned = 0;
4209 int pgrescued = 0;
4210 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004211
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004212 for (i = 0; i < pvec->nr; i++) {
4213 struct page *page = pvec->pages[i];
Mel Gorman785b99f2016-07-28 15:47:23 -07004214 struct pglist_data *pagepgdat = page_pgdat(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07004215
Hugh Dickins24513262012-01-20 14:34:21 -08004216 pgscanned++;
Mel Gorman785b99f2016-07-28 15:47:23 -07004217 if (pagepgdat != pgdat) {
4218 if (pgdat)
4219 spin_unlock_irq(&pgdat->lru_lock);
4220 pgdat = pagepgdat;
4221 spin_lock_irq(&pgdat->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004222 }
Mel Gorman785b99f2016-07-28 15:47:23 -07004223 lruvec = mem_cgroup_page_lruvec(page, pgdat);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004224
Hugh Dickins24513262012-01-20 14:34:21 -08004225 if (!PageLRU(page) || !PageUnevictable(page))
4226 continue;
4227
Hugh Dickins39b5f292012-10-08 16:33:18 -07004228 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08004229 enum lru_list lru = page_lru_base_type(page);
4230
Sasha Levin309381fea2014-01-23 15:52:54 -08004231 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08004232 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07004233 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4234 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08004235 pgrescued++;
4236 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07004237 }
Hugh Dickins24513262012-01-20 14:34:21 -08004238
Mel Gorman785b99f2016-07-28 15:47:23 -07004239 if (pgdat) {
Hugh Dickins24513262012-01-20 14:34:21 -08004240 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4241 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
Mel Gorman785b99f2016-07-28 15:47:23 -07004242 spin_unlock_irq(&pgdat->lru_lock);
Hugh Dickins24513262012-01-20 14:34:21 -08004243 }
Hugh Dickins850465792012-01-20 14:34:19 -08004244}
Kuo-Hsin Yang64e3d122018-11-06 13:23:24 +00004245EXPORT_SYMBOL_GPL(check_move_unevictable_pages);