Greg Kroah-Hartman | b244131 | 2017-11-01 15:07:57 +0100 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2 | /* |
| 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 Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 15 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 16 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | #include <linux/mm.h> |
Ingo Molnar | 5b3cc15 | 2017-02-02 20:43:54 +0100 | [diff] [blame] | 18 | #include <linux/sched/mm.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 19 | #include <linux/module.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 20 | #include <linux/gfp.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 21 | #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 Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 26 | #include <linux/vmpressure.h> |
Andrew Morton | e129b5c | 2006-09-27 01:50:00 -0700 | [diff] [blame] | 27 | #include <linux/vmstat.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 28 | #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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 34 | #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 Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 39 | #include <linux/compaction.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 40 | #include <linux/notifier.h> |
| 41 | #include <linux/rwsem.h> |
Rafael J. Wysocki | 248a030 | 2006-03-22 00:09:04 -0800 | [diff] [blame] | 42 | #include <linux/delay.h> |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 43 | #include <linux/kthread.h> |
Nigel Cunningham | 7dfb710 | 2006-12-06 20:34:23 -0800 | [diff] [blame] | 44 | #include <linux/freezer.h> |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 45 | #include <linux/memcontrol.h> |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 46 | #include <linux/delayacct.h> |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 47 | #include <linux/sysctl.h> |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 48 | #include <linux/oom.h> |
Linus Torvalds | 268bb0c | 2011-05-20 12:50:29 -0700 | [diff] [blame] | 49 | #include <linux/prefetch.h> |
Mitchel Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 50 | #include <linux/printk.h> |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 51 | #include <linux/dax.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 52 | |
| 53 | #include <asm/tlbflush.h> |
| 54 | #include <asm/div64.h> |
| 55 | |
| 56 | #include <linux/swapops.h> |
Rafael Aquini | 117aad1 | 2013-09-30 13:45:16 -0700 | [diff] [blame] | 57 | #include <linux/balloon_compaction.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 58 | |
Nick Piggin | 0f8053a | 2006-03-22 00:08:33 -0800 | [diff] [blame] | 59 | #include "internal.h" |
| 60 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 61 | #define CREATE_TRACE_POINTS |
| 62 | #include <trace/events/vmscan.h> |
| 63 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 64 | struct scan_control { |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 65 | /* How many pages shrink_list() should reclaim */ |
| 66 | unsigned long nr_to_reclaim; |
| 67 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 68 | /* This context's GFP mask */ |
Al Viro | 6daa0e2 | 2005-10-21 03:18:50 -0400 | [diff] [blame] | 69 | gfp_t gfp_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 70 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 71 | /* Allocation order */ |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 72 | int order; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 73 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 74 | /* |
| 75 | * Nodemask of nodes allowed by the caller. If NULL, all nodes |
| 76 | * are scanned. |
| 77 | */ |
| 78 | nodemask_t *nodemask; |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 79 | |
KOSAKI Motohiro | 5f53e76 | 2010-05-24 14:32:37 -0700 | [diff] [blame] | 80 | /* |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 81 | * The memory cgroup that hit its limit and as a result is the |
| 82 | * primary target of this reclaim invocation. |
| 83 | */ |
| 84 | struct mem_cgroup *target_mem_cgroup; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 85 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 86 | /* Scan (total_size >> priority) pages at once */ |
| 87 | int priority; |
| 88 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 89 | /* The highest zone to isolate pages for reclaim from */ |
| 90 | enum zone_type reclaim_idx; |
| 91 | |
Johannes Weiner | 1276ad6 | 2017-02-24 14:56:11 -0800 | [diff] [blame] | 92 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 93 | unsigned int may_writepage:1; |
| 94 | |
| 95 | /* Can mapped pages be reclaimed? */ |
| 96 | unsigned int may_unmap:1; |
| 97 | |
| 98 | /* Can pages be swapped as part of reclaim? */ |
| 99 | unsigned int may_swap:1; |
| 100 | |
Yisheng Xie | d6622f6 | 2017-05-03 14:53:57 -0700 | [diff] [blame] | 101 | /* |
| 102 | * Cgroups are not reclaimed below their configured memory.low, |
| 103 | * unless we threaten to OOM. If any cgroups are skipped due to |
| 104 | * memory.low and nothing was reclaimed, go back for memory.low. |
| 105 | */ |
| 106 | unsigned int memcg_low_reclaim:1; |
| 107 | unsigned int memcg_low_skipped:1; |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 108 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 109 | unsigned int hibernation_mode:1; |
| 110 | |
| 111 | /* One of the zones is ready for compaction */ |
| 112 | unsigned int compaction_ready:1; |
| 113 | |
| 114 | /* Incremented by the number of inactive pages that were scanned */ |
| 115 | unsigned long nr_scanned; |
| 116 | |
| 117 | /* Number of pages freed so far during a call to shrink_zones() */ |
| 118 | unsigned long nr_reclaimed; |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 119 | |
| 120 | struct { |
| 121 | unsigned int dirty; |
| 122 | unsigned int unqueued_dirty; |
| 123 | unsigned int congested; |
| 124 | unsigned int writeback; |
| 125 | unsigned int immediate; |
| 126 | unsigned int file_taken; |
| 127 | unsigned int taken; |
| 128 | } nr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | }; |
| 130 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | #ifdef ARCH_HAS_PREFETCH |
| 132 | #define prefetch_prev_lru_page(_page, _base, _field) \ |
| 133 | do { \ |
| 134 | if ((_page)->lru.prev != _base) { \ |
| 135 | struct page *prev; \ |
| 136 | \ |
| 137 | prev = lru_to_page(&(_page->lru)); \ |
| 138 | prefetch(&prev->_field); \ |
| 139 | } \ |
| 140 | } while (0) |
| 141 | #else |
| 142 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) |
| 143 | #endif |
| 144 | |
| 145 | #ifdef ARCH_HAS_PREFETCHW |
| 146 | #define prefetchw_prev_lru_page(_page, _base, _field) \ |
| 147 | do { \ |
| 148 | if ((_page)->lru.prev != _base) { \ |
| 149 | struct page *prev; \ |
| 150 | \ |
| 151 | prev = lru_to_page(&(_page->lru)); \ |
| 152 | prefetchw(&prev->_field); \ |
| 153 | } \ |
| 154 | } while (0) |
| 155 | #else |
| 156 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) |
| 157 | #endif |
| 158 | |
| 159 | /* |
| 160 | * From 0 .. 100. Higher means more swappy. |
| 161 | */ |
| 162 | int vm_swappiness = 60; |
Wang Sheng-Hui | d0480be | 2014-08-06 16:07:07 -0700 | [diff] [blame] | 163 | /* |
| 164 | * The total number of pages which are beyond the high watermark within all |
| 165 | * zones. |
| 166 | */ |
| 167 | unsigned long vm_total_pages; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 168 | |
| 169 | static LIST_HEAD(shrinker_list); |
| 170 | static DECLARE_RWSEM(shrinker_rwsem); |
| 171 | |
Andrew Morton | c255a45 | 2012-07-31 16:43:02 -0700 | [diff] [blame] | 172 | #ifdef CONFIG_MEMCG |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 173 | static bool global_reclaim(struct scan_control *sc) |
| 174 | { |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 175 | return !sc->target_mem_cgroup; |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 176 | } |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 177 | |
| 178 | /** |
| 179 | * sane_reclaim - is the usual dirty throttling mechanism operational? |
| 180 | * @sc: scan_control in question |
| 181 | * |
| 182 | * The normal page dirty throttling mechanism in balance_dirty_pages() is |
| 183 | * completely broken with the legacy memcg and direct stalling in |
| 184 | * shrink_page_list() is used for throttling instead, which lacks all the |
| 185 | * niceties such as fairness, adaptive pausing, bandwidth proportional |
| 186 | * allocation and configurability. |
| 187 | * |
| 188 | * This function tests whether the vmscan currently in progress can assume |
| 189 | * that the normal dirty throttling mechanism is operational. |
| 190 | */ |
| 191 | static bool sane_reclaim(struct scan_control *sc) |
| 192 | { |
| 193 | struct mem_cgroup *memcg = sc->target_mem_cgroup; |
| 194 | |
| 195 | if (!memcg) |
| 196 | return true; |
| 197 | #ifdef CONFIG_CGROUP_WRITEBACK |
Linus Torvalds | 69234ac | 2015-11-05 14:51:32 -0800 | [diff] [blame] | 198 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 199 | return true; |
| 200 | #endif |
| 201 | return false; |
| 202 | } |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 203 | |
| 204 | static void set_memcg_congestion(pg_data_t *pgdat, |
| 205 | struct mem_cgroup *memcg, |
| 206 | bool congested) |
| 207 | { |
| 208 | struct mem_cgroup_per_node *mn; |
| 209 | |
| 210 | if (!memcg) |
| 211 | return; |
| 212 | |
| 213 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); |
| 214 | WRITE_ONCE(mn->congested, congested); |
| 215 | } |
| 216 | |
| 217 | static bool memcg_congested(pg_data_t *pgdat, |
| 218 | struct mem_cgroup *memcg) |
| 219 | { |
| 220 | struct mem_cgroup_per_node *mn; |
| 221 | |
| 222 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); |
| 223 | return READ_ONCE(mn->congested); |
| 224 | |
| 225 | } |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 226 | #else |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 227 | static bool global_reclaim(struct scan_control *sc) |
| 228 | { |
| 229 | return true; |
| 230 | } |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 231 | |
| 232 | static bool sane_reclaim(struct scan_control *sc) |
| 233 | { |
| 234 | return true; |
| 235 | } |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 236 | |
| 237 | static inline void set_memcg_congestion(struct pglist_data *pgdat, |
| 238 | struct mem_cgroup *memcg, bool congested) |
| 239 | { |
| 240 | } |
| 241 | |
| 242 | static inline bool memcg_congested(struct pglist_data *pgdat, |
| 243 | struct mem_cgroup *memcg) |
| 244 | { |
| 245 | return false; |
| 246 | |
| 247 | } |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 248 | #endif |
| 249 | |
Mel Gorman | 5a1c84b | 2016-07-28 15:47:31 -0700 | [diff] [blame] | 250 | /* |
| 251 | * This misses isolated pages which are not accounted for to save counters. |
| 252 | * As the data only determines if reclaim or compaction continues, it is |
| 253 | * not expected that isolated pages will be a dominating factor. |
| 254 | */ |
| 255 | unsigned long zone_reclaimable_pages(struct zone *zone) |
| 256 | { |
| 257 | unsigned long nr; |
| 258 | |
| 259 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + |
| 260 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); |
| 261 | if (get_nr_swap_pages() > 0) |
| 262 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
| 263 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); |
| 264 | |
| 265 | return nr; |
| 266 | } |
| 267 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 268 | /** |
| 269 | * lruvec_lru_size - Returns the number of pages on the given LRU list. |
| 270 | * @lruvec: lru vector |
| 271 | * @lru: lru to use |
| 272 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) |
| 273 | */ |
| 274 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) |
KOSAKI Motohiro | c9f299d | 2009-01-07 18:08:16 -0800 | [diff] [blame] | 275 | { |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 276 | unsigned long lru_size; |
| 277 | int zid; |
| 278 | |
Hugh Dickins | c3c787e | 2012-05-29 15:06:52 -0700 | [diff] [blame] | 279 | if (!mem_cgroup_disabled()) |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 280 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
| 281 | else |
| 282 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); |
KOSAKI Motohiro | a3d8e05 | 2009-01-07 18:08:19 -0800 | [diff] [blame] | 283 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 284 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
| 285 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
| 286 | unsigned long size; |
KOSAKI Motohiro | c9f299d | 2009-01-07 18:08:16 -0800 | [diff] [blame] | 287 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 288 | if (!managed_zone(zone)) |
| 289 | continue; |
Michal Hocko | b4536f0c8 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 290 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 291 | if (!mem_cgroup_disabled()) |
| 292 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
| 293 | else |
| 294 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], |
| 295 | NR_ZONE_LRU_BASE + lru); |
| 296 | lru_size -= min(size, lru_size); |
| 297 | } |
| 298 | |
| 299 | return lru_size; |
| 300 | |
Michal Hocko | b4536f0c8 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 301 | } |
| 302 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 303 | /* |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 304 | * Add a shrinker callback to be called from the vm. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 305 | */ |
Tetsuo Handa | 8e04944 | 2018-04-04 19:53:07 +0900 | [diff] [blame] | 306 | int prealloc_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 307 | { |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 308 | size_t size = sizeof(*shrinker->nr_deferred); |
| 309 | |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 310 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
| 311 | size *= nr_node_ids; |
| 312 | |
| 313 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); |
| 314 | if (!shrinker->nr_deferred) |
| 315 | return -ENOMEM; |
Tetsuo Handa | 8e04944 | 2018-04-04 19:53:07 +0900 | [diff] [blame] | 316 | return 0; |
| 317 | } |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 318 | |
Tetsuo Handa | 8e04944 | 2018-04-04 19:53:07 +0900 | [diff] [blame] | 319 | void free_prealloced_shrinker(struct shrinker *shrinker) |
| 320 | { |
| 321 | kfree(shrinker->nr_deferred); |
| 322 | shrinker->nr_deferred = NULL; |
| 323 | } |
| 324 | |
| 325 | void register_shrinker_prepared(struct shrinker *shrinker) |
| 326 | { |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 327 | down_write(&shrinker_rwsem); |
| 328 | list_add_tail(&shrinker->list, &shrinker_list); |
| 329 | up_write(&shrinker_rwsem); |
Tetsuo Handa | 8e04944 | 2018-04-04 19:53:07 +0900 | [diff] [blame] | 330 | } |
| 331 | |
| 332 | int register_shrinker(struct shrinker *shrinker) |
| 333 | { |
| 334 | int err = prealloc_shrinker(shrinker); |
| 335 | |
| 336 | if (err) |
| 337 | return err; |
| 338 | register_shrinker_prepared(shrinker); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 339 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 340 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 341 | EXPORT_SYMBOL(register_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 342 | |
| 343 | /* |
| 344 | * Remove one |
| 345 | */ |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 346 | void unregister_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 347 | { |
Tetsuo Handa | bb422a7 | 2017-12-18 20:31:41 +0900 | [diff] [blame] | 348 | if (!shrinker->nr_deferred) |
| 349 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 350 | down_write(&shrinker_rwsem); |
| 351 | list_del(&shrinker->list); |
| 352 | up_write(&shrinker_rwsem); |
Andrew Vagin | ae39332 | 2013-10-16 13:46:46 -0700 | [diff] [blame] | 353 | kfree(shrinker->nr_deferred); |
Tetsuo Handa | bb422a7 | 2017-12-18 20:31:41 +0900 | [diff] [blame] | 354 | shrinker->nr_deferred = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 355 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 356 | EXPORT_SYMBOL(unregister_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 357 | |
| 358 | #define SHRINK_BATCH 128 |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 359 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 360 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 361 | struct shrinker *shrinker, int priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 362 | { |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 363 | unsigned long freed = 0; |
| 364 | unsigned long long delta; |
| 365 | long total_scan; |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 366 | long freeable; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 367 | long nr; |
| 368 | long new_nr; |
| 369 | int nid = shrinkctl->nid; |
| 370 | long batch_size = shrinker->batch ? shrinker->batch |
| 371 | : SHRINK_BATCH; |
Shaohua Li | 5f33a08 | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 372 | long scanned = 0, next_deferred; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 373 | |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 374 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
| 375 | if (freeable == 0) |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 376 | return 0; |
| 377 | |
| 378 | /* |
| 379 | * copy the current shrinker scan count into a local variable |
| 380 | * and zero it so that other concurrent shrinker invocations |
| 381 | * don't also do this scanning work. |
| 382 | */ |
| 383 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); |
| 384 | |
| 385 | total_scan = nr; |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 386 | delta = freeable >> priority; |
| 387 | delta *= 4; |
| 388 | do_div(delta, shrinker->seeks); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 389 | total_scan += delta; |
| 390 | if (total_scan < 0) { |
Pintu Kumar | 8612c66 | 2014-12-10 15:42:58 -0800 | [diff] [blame] | 391 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 392 | shrinker->scan_objects, total_scan); |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 393 | total_scan = freeable; |
Shaohua Li | 5f33a08 | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 394 | next_deferred = nr; |
| 395 | } else |
| 396 | next_deferred = total_scan; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 397 | |
| 398 | /* |
| 399 | * We need to avoid excessive windup on filesystem shrinkers |
| 400 | * due to large numbers of GFP_NOFS allocations causing the |
| 401 | * shrinkers to return -1 all the time. This results in a large |
| 402 | * nr being built up so when a shrink that can do some work |
| 403 | * comes along it empties the entire cache due to nr >>> |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 404 | * freeable. This is bad for sustaining a working set in |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 405 | * memory. |
| 406 | * |
| 407 | * Hence only allow the shrinker to scan the entire cache when |
| 408 | * a large delta change is calculated directly. |
| 409 | */ |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 410 | if (delta < freeable / 4) |
| 411 | total_scan = min(total_scan, freeable / 2); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 412 | |
| 413 | /* |
| 414 | * Avoid risking looping forever due to too large nr value: |
| 415 | * never try to free more than twice the estimate number of |
| 416 | * freeable entries. |
| 417 | */ |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 418 | if (total_scan > freeable * 2) |
| 419 | total_scan = freeable * 2; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 420 | |
| 421 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 422 | freeable, delta, total_scan, priority); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 423 | |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 424 | /* |
| 425 | * Normally, we should not scan less than batch_size objects in one |
| 426 | * pass to avoid too frequent shrinker calls, but if the slab has less |
| 427 | * than batch_size objects in total and we are really tight on memory, |
| 428 | * we will try to reclaim all available objects, otherwise we can end |
| 429 | * up failing allocations although there are plenty of reclaimable |
| 430 | * objects spread over several slabs with usage less than the |
| 431 | * batch_size. |
| 432 | * |
| 433 | * We detect the "tight on memory" situations by looking at the total |
| 434 | * number of objects we want to scan (total_scan). If it is greater |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 435 | * than the total number of objects on slab (freeable), we must be |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 436 | * scanning at high prio and therefore should try to reclaim as much as |
| 437 | * possible. |
| 438 | */ |
| 439 | while (total_scan >= batch_size || |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 440 | total_scan >= freeable) { |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 441 | unsigned long ret; |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 442 | unsigned long nr_to_scan = min(batch_size, total_scan); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 443 | |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 444 | shrinkctl->nr_to_scan = nr_to_scan; |
Chris Wilson | d460acb | 2017-09-06 16:19:26 -0700 | [diff] [blame] | 445 | shrinkctl->nr_scanned = nr_to_scan; |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 446 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
| 447 | if (ret == SHRINK_STOP) |
| 448 | break; |
| 449 | freed += ret; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 450 | |
Chris Wilson | d460acb | 2017-09-06 16:19:26 -0700 | [diff] [blame] | 451 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
| 452 | total_scan -= shrinkctl->nr_scanned; |
| 453 | scanned += shrinkctl->nr_scanned; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 454 | |
| 455 | cond_resched(); |
| 456 | } |
| 457 | |
Shaohua Li | 5f33a08 | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 458 | if (next_deferred >= scanned) |
| 459 | next_deferred -= scanned; |
| 460 | else |
| 461 | next_deferred = 0; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 462 | /* |
| 463 | * move the unused scan count back into the shrinker in a |
| 464 | * manner that handles concurrent updates. If we exhausted the |
| 465 | * scan, there is no need to do an update. |
| 466 | */ |
Shaohua Li | 5f33a08 | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 467 | if (next_deferred > 0) |
| 468 | new_nr = atomic_long_add_return(next_deferred, |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 469 | &shrinker->nr_deferred[nid]); |
| 470 | else |
| 471 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); |
| 472 | |
Dave Hansen | df9024a | 2014-06-04 16:08:07 -0700 | [diff] [blame] | 473 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 474 | return freed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 475 | } |
| 476 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 477 | /** |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 478 | * shrink_slab - shrink slab caches |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 479 | * @gfp_mask: allocation context |
| 480 | * @nid: node whose slab caches to target |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 481 | * @memcg: memory cgroup whose slab caches to target |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 482 | * @priority: the reclaim priority |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 483 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 484 | * Call the shrink functions to age shrinkable caches. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 485 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 486 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
| 487 | * unaware shrinkers will receive a node id of 0 instead. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 488 | * |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 489 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
| 490 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 491 | * objects from the memory cgroup specified. Otherwise, only unaware |
| 492 | * shrinkers are called. |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 493 | * |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 494 | * @priority is sc->priority, we take the number of objects and >> by priority |
| 495 | * in order to get the scan target. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 496 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 497 | * Returns the number of reclaimed slab objects. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 498 | */ |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 499 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
| 500 | struct mem_cgroup *memcg, |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 501 | int priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 502 | { |
| 503 | struct shrinker *shrinker; |
Dave Chinner | 24f7c6b | 2013-08-28 10:17:56 +1000 | [diff] [blame] | 504 | unsigned long freed = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 505 | |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 506 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 507 | return 0; |
| 508 | |
Tetsuo Handa | e830c63 | 2018-04-05 16:23:35 -0700 | [diff] [blame] | 509 | if (!down_read_trylock(&shrinker_rwsem)) |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 510 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 511 | |
| 512 | list_for_each_entry(shrinker, &shrinker_list, list) { |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 513 | struct shrink_control sc = { |
| 514 | .gfp_mask = gfp_mask, |
| 515 | .nid = nid, |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 516 | .memcg = memcg, |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 517 | }; |
Vladimir Davydov | ec97097 | 2014-01-23 15:53:23 -0800 | [diff] [blame] | 518 | |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 519 | /* |
| 520 | * If kernel memory accounting is disabled, we ignore |
| 521 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers |
| 522 | * passing NULL for memcg. |
| 523 | */ |
| 524 | if (memcg_kmem_enabled() && |
| 525 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 526 | continue; |
| 527 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 528 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
| 529 | sc.nid = 0; |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 530 | |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 531 | freed += do_shrink_slab(&sc, shrinker, priority); |
Minchan Kim | e496612 | 2018-01-31 16:16:55 -0800 | [diff] [blame] | 532 | /* |
| 533 | * Bail out if someone want to register a new shrinker to |
| 534 | * prevent the regsitration from being stalled for long periods |
| 535 | * by parallel ongoing shrinking. |
| 536 | */ |
| 537 | if (rwsem_is_contended(&shrinker_rwsem)) { |
| 538 | freed = freed ? : 1; |
| 539 | break; |
| 540 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 541 | } |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 542 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 543 | up_read(&shrinker_rwsem); |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 544 | out: |
| 545 | cond_resched(); |
Dave Chinner | 24f7c6b | 2013-08-28 10:17:56 +1000 | [diff] [blame] | 546 | return freed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 547 | } |
| 548 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 549 | void drop_slab_node(int nid) |
| 550 | { |
| 551 | unsigned long freed; |
| 552 | |
| 553 | do { |
| 554 | struct mem_cgroup *memcg = NULL; |
| 555 | |
| 556 | freed = 0; |
| 557 | do { |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 558 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 559 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
| 560 | } while (freed > 10); |
| 561 | } |
| 562 | |
| 563 | void drop_slab(void) |
| 564 | { |
| 565 | int nid; |
| 566 | |
| 567 | for_each_online_node(nid) |
| 568 | drop_slab_node(nid); |
| 569 | } |
| 570 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 571 | static inline int is_page_cache_freeable(struct page *page) |
| 572 | { |
Johannes Weiner | ceddc3a | 2009-09-21 17:03:00 -0700 | [diff] [blame] | 573 | /* |
| 574 | * A freeable page cache page is referenced only by the caller |
| 575 | * that isolated the page, the page cache radix tree and |
| 576 | * optional buffer heads at page->private. |
| 577 | */ |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 578 | int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? |
| 579 | HPAGE_PMD_NR : 1; |
| 580 | return page_count(page) - page_has_private(page) == 1 + radix_pins; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 581 | } |
| 582 | |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 583 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 584 | { |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 585 | if (current->flags & PF_SWAPWRITE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 586 | return 1; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 587 | if (!inode_write_congested(inode)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 588 | return 1; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 589 | if (inode_to_bdi(inode) == current->backing_dev_info) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 590 | return 1; |
| 591 | return 0; |
| 592 | } |
| 593 | |
| 594 | /* |
| 595 | * We detected a synchronous write error writing a page out. Probably |
| 596 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
| 597 | * fsync(), msync() or close(). |
| 598 | * |
| 599 | * The tricky part is that after writepage we cannot touch the mapping: nothing |
| 600 | * prevents it from being freed up. But we have a ref on the page and once |
| 601 | * that page is locked, the mapping is pinned. |
| 602 | * |
| 603 | * We're allowed to run sleeping lock_page() here because we know the caller has |
| 604 | * __GFP_FS. |
| 605 | */ |
| 606 | static void handle_write_error(struct address_space *mapping, |
| 607 | struct page *page, int error) |
| 608 | { |
Jens Axboe | 7eaceac | 2011-03-10 08:52:07 +0100 | [diff] [blame] | 609 | lock_page(page); |
Guillaume Chazarain | 3e9f45b | 2007-05-08 00:23:25 -0700 | [diff] [blame] | 610 | if (page_mapping(page) == mapping) |
| 611 | mapping_set_error(mapping, error); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 612 | unlock_page(page); |
| 613 | } |
| 614 | |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 615 | /* possible outcome of pageout() */ |
| 616 | typedef enum { |
| 617 | /* failed to write page out, page is locked */ |
| 618 | PAGE_KEEP, |
| 619 | /* move page to the active list, page is locked */ |
| 620 | PAGE_ACTIVATE, |
| 621 | /* page has been sent to the disk successfully, page is unlocked */ |
| 622 | PAGE_SUCCESS, |
| 623 | /* page is clean and locked */ |
| 624 | PAGE_CLEAN, |
| 625 | } pageout_t; |
| 626 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 627 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 628 | * pageout is called by shrink_page_list() for each dirty page. |
| 629 | * Calls ->writepage(). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 630 | */ |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 631 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 632 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 633 | { |
| 634 | /* |
| 635 | * If the page is dirty, only perform writeback if that write |
| 636 | * will be non-blocking. To prevent this allocation from being |
| 637 | * stalled by pagecache activity. But note that there may be |
| 638 | * stalls if we need to run get_block(). We could test |
| 639 | * PagePrivate for that. |
| 640 | * |
Al Viro | 8174202 | 2014-04-03 03:17:43 -0400 | [diff] [blame] | 641 | * If this process is currently in __generic_file_write_iter() against |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 642 | * this page's queue, we can perform writeback even if that |
| 643 | * will block. |
| 644 | * |
| 645 | * If the page is swapcache, write it back even if that would |
| 646 | * block, for some throttling. This happens by accident, because |
| 647 | * swap_backing_dev_info is bust: it doesn't reflect the |
| 648 | * congestion state of the swapdevs. Easy to fix, if needed. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 649 | */ |
| 650 | if (!is_page_cache_freeable(page)) |
| 651 | return PAGE_KEEP; |
| 652 | if (!mapping) { |
| 653 | /* |
| 654 | * Some data journaling orphaned pages can have |
| 655 | * page->mapping == NULL while being dirty with clean buffers. |
| 656 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 657 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 658 | if (try_to_free_buffers(page)) { |
| 659 | ClearPageDirty(page); |
Mitchel Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 660 | pr_info("%s: orphaned page\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 661 | return PAGE_CLEAN; |
| 662 | } |
| 663 | } |
| 664 | return PAGE_KEEP; |
| 665 | } |
| 666 | if (mapping->a_ops->writepage == NULL) |
| 667 | return PAGE_ACTIVATE; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 668 | if (!may_write_to_inode(mapping->host, sc)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 669 | return PAGE_KEEP; |
| 670 | |
| 671 | if (clear_page_dirty_for_io(page)) { |
| 672 | int res; |
| 673 | struct writeback_control wbc = { |
| 674 | .sync_mode = WB_SYNC_NONE, |
| 675 | .nr_to_write = SWAP_CLUSTER_MAX, |
OGAWA Hirofumi | 111ebb6 | 2006-06-23 02:03:26 -0700 | [diff] [blame] | 676 | .range_start = 0, |
| 677 | .range_end = LLONG_MAX, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 678 | .for_reclaim = 1, |
| 679 | }; |
| 680 | |
| 681 | SetPageReclaim(page); |
| 682 | res = mapping->a_ops->writepage(page, &wbc); |
| 683 | if (res < 0) |
| 684 | handle_write_error(mapping, page, res); |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 685 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 686 | ClearPageReclaim(page); |
| 687 | return PAGE_ACTIVATE; |
| 688 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 689 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 690 | if (!PageWriteback(page)) { |
| 691 | /* synchronous write or broken a_ops? */ |
| 692 | ClearPageReclaim(page); |
| 693 | } |
yalin wang | 3aa2385 | 2016-01-14 15:18:30 -0800 | [diff] [blame] | 694 | trace_mm_vmscan_writepage(page); |
Mel Gorman | c4a2563 | 2016-07-28 15:46:23 -0700 | [diff] [blame] | 695 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 696 | return PAGE_SUCCESS; |
| 697 | } |
| 698 | |
| 699 | return PAGE_CLEAN; |
| 700 | } |
| 701 | |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 702 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 703 | * Same as remove_mapping, but if the page is removed from the mapping, it |
| 704 | * gets returned with a refcount of 0. |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 705 | */ |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 706 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
| 707 | bool reclaimed) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 708 | { |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 709 | unsigned long flags; |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 710 | int refcount; |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 711 | |
Nick Piggin | 28e4d96 | 2006-09-25 23:31:23 -0700 | [diff] [blame] | 712 | BUG_ON(!PageLocked(page)); |
| 713 | BUG_ON(mapping != page_mapping(page)); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 714 | |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 715 | xa_lock_irqsave(&mapping->i_pages, flags); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 716 | /* |
Nick Piggin | 0fd0e6b | 2006-09-27 01:50:02 -0700 | [diff] [blame] | 717 | * The non racy check for a busy page. |
| 718 | * |
| 719 | * Must be careful with the order of the tests. When someone has |
| 720 | * a ref to the page, it may be possible that they dirty it then |
| 721 | * drop the reference. So if PageDirty is tested before page_count |
| 722 | * here, then the following race may occur: |
| 723 | * |
| 724 | * get_user_pages(&page); |
| 725 | * [user mapping goes away] |
| 726 | * write_to(page); |
| 727 | * !PageDirty(page) [good] |
| 728 | * SetPageDirty(page); |
| 729 | * put_page(page); |
| 730 | * !page_count(page) [good, discard it] |
| 731 | * |
| 732 | * [oops, our write_to data is lost] |
| 733 | * |
| 734 | * Reversing the order of the tests ensures such a situation cannot |
| 735 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags |
Joonsoo Kim | 0139aa7 | 2016-05-19 17:10:49 -0700 | [diff] [blame] | 736 | * load is not satisfied before that of page->_refcount. |
Nick Piggin | 0fd0e6b | 2006-09-27 01:50:02 -0700 | [diff] [blame] | 737 | * |
| 738 | * Note that if SetPageDirty is always performed via set_page_dirty, |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 739 | * and thus under the i_pages lock, then this ordering is not required. |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 740 | */ |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 741 | if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) |
| 742 | refcount = 1 + HPAGE_PMD_NR; |
| 743 | else |
| 744 | refcount = 2; |
| 745 | if (!page_ref_freeze(page, refcount)) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 746 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 747 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
| 748 | if (unlikely(PageDirty(page))) { |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 749 | page_ref_unfreeze(page, refcount); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 750 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 751 | } |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 752 | |
| 753 | if (PageSwapCache(page)) { |
| 754 | swp_entry_t swap = { .val = page_private(page) }; |
Johannes Weiner | 0a31bc9 | 2014-08-08 14:19:22 -0700 | [diff] [blame] | 755 | mem_cgroup_swapout(page, swap); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 756 | __delete_from_swap_cache(page); |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 757 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
Minchan Kim | 75f6d6d | 2017-07-06 15:37:21 -0700 | [diff] [blame] | 758 | put_swap_page(page, swap); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 759 | } else { |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 760 | void (*freepage)(struct page *); |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 761 | void *shadow = NULL; |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 762 | |
| 763 | freepage = mapping->a_ops->freepage; |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 764 | /* |
| 765 | * Remember a shadow entry for reclaimed file cache in |
| 766 | * order to detect refaults, thus thrashing, later on. |
| 767 | * |
| 768 | * But don't store shadows in an address space that is |
| 769 | * already exiting. This is not just an optizimation, |
| 770 | * inode reclaim needs to empty out the radix tree or |
| 771 | * the nodes are lost. Don't plant shadows behind its |
| 772 | * back. |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 773 | * |
| 774 | * We also don't store shadows for DAX mappings because the |
| 775 | * only page cache pages found in these are zero pages |
| 776 | * covering holes, and because we don't want to mix DAX |
| 777 | * exceptional entries and shadow exceptional entries in the |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 778 | * same address_space. |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 779 | */ |
| 780 | if (reclaimed && page_is_file_cache(page) && |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 781 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 782 | shadow = workingset_eviction(mapping, page); |
Johannes Weiner | 62cccb8 | 2016-03-15 14:57:22 -0700 | [diff] [blame] | 783 | __delete_from_page_cache(page, shadow); |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 784 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 785 | |
| 786 | if (freepage != NULL) |
| 787 | freepage(page); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 788 | } |
| 789 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 790 | return 1; |
| 791 | |
| 792 | cannot_free: |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 793 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 794 | return 0; |
| 795 | } |
| 796 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 797 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 798 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if |
| 799 | * someone else has a ref on the page, abort and return 0. If it was |
| 800 | * successfully detached, return 1. Assumes the caller has a single ref on |
| 801 | * this page. |
| 802 | */ |
| 803 | int remove_mapping(struct address_space *mapping, struct page *page) |
| 804 | { |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 805 | if (__remove_mapping(mapping, page, false)) { |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 806 | /* |
| 807 | * Unfreezing the refcount with 1 rather than 2 effectively |
| 808 | * drops the pagecache ref for us without requiring another |
| 809 | * atomic operation. |
| 810 | */ |
Joonsoo Kim | fe896d1 | 2016-03-17 14:19:26 -0700 | [diff] [blame] | 811 | page_ref_unfreeze(page, 1); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 812 | return 1; |
| 813 | } |
| 814 | return 0; |
| 815 | } |
| 816 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 817 | /** |
| 818 | * putback_lru_page - put previously isolated page onto appropriate LRU list |
| 819 | * @page: page to be put back to appropriate lru list |
| 820 | * |
| 821 | * Add previously isolated @page to appropriate LRU list. |
| 822 | * Page may still be unevictable for other reasons. |
| 823 | * |
| 824 | * lru_lock must not be held, interrupts must be enabled. |
| 825 | */ |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 826 | void putback_lru_page(struct page *page) |
| 827 | { |
Shakeel Butt | 9c4e6b1 | 2018-02-21 14:45:28 -0800 | [diff] [blame] | 828 | lru_cache_add(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 829 | put_page(page); /* drop ref from isolate */ |
| 830 | } |
| 831 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 832 | enum page_references { |
| 833 | PAGEREF_RECLAIM, |
| 834 | PAGEREF_RECLAIM_CLEAN, |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 835 | PAGEREF_KEEP, |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 836 | PAGEREF_ACTIVATE, |
| 837 | }; |
| 838 | |
| 839 | static enum page_references page_check_references(struct page *page, |
| 840 | struct scan_control *sc) |
| 841 | { |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 842 | int referenced_ptes, referenced_page; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 843 | unsigned long vm_flags; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 844 | |
Johannes Weiner | c3ac9a8 | 2012-05-29 15:06:25 -0700 | [diff] [blame] | 845 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
| 846 | &vm_flags); |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 847 | referenced_page = TestClearPageReferenced(page); |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 848 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 849 | /* |
| 850 | * Mlock lost the isolation race with us. Let try_to_unmap() |
| 851 | * move the page to the unevictable list. |
| 852 | */ |
| 853 | if (vm_flags & VM_LOCKED) |
| 854 | return PAGEREF_RECLAIM; |
| 855 | |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 856 | if (referenced_ptes) { |
Michal Hocko | e4898273 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 857 | if (PageSwapBacked(page)) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 858 | return PAGEREF_ACTIVATE; |
| 859 | /* |
| 860 | * All mapped pages start out with page table |
| 861 | * references from the instantiating fault, so we need |
| 862 | * to look twice if a mapped file page is used more |
| 863 | * than once. |
| 864 | * |
| 865 | * Mark it and spare it for another trip around the |
| 866 | * inactive list. Another page table reference will |
| 867 | * lead to its activation. |
| 868 | * |
| 869 | * Note: the mark is set for activated pages as well |
| 870 | * so that recently deactivated but used pages are |
| 871 | * quickly recovered. |
| 872 | */ |
| 873 | SetPageReferenced(page); |
| 874 | |
Konstantin Khlebnikov | 34dbc67 | 2012-01-10 15:06:59 -0800 | [diff] [blame] | 875 | if (referenced_page || referenced_ptes > 1) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 876 | return PAGEREF_ACTIVATE; |
| 877 | |
Konstantin Khlebnikov | c909e99 | 2012-01-10 15:07:03 -0800 | [diff] [blame] | 878 | /* |
| 879 | * Activate file-backed executable pages after first usage. |
| 880 | */ |
| 881 | if (vm_flags & VM_EXEC) |
| 882 | return PAGEREF_ACTIVATE; |
| 883 | |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 884 | return PAGEREF_KEEP; |
| 885 | } |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 886 | |
| 887 | /* Reclaim if clean, defer dirty pages to writeback */ |
KOSAKI Motohiro | 2e30244 | 2010-10-26 14:21:46 -0700 | [diff] [blame] | 888 | if (referenced_page && !PageSwapBacked(page)) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 889 | return PAGEREF_RECLAIM_CLEAN; |
| 890 | |
| 891 | return PAGEREF_RECLAIM; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 892 | } |
| 893 | |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 894 | /* Check if a page is dirty or under writeback */ |
| 895 | static void page_check_dirty_writeback(struct page *page, |
| 896 | bool *dirty, bool *writeback) |
| 897 | { |
Mel Gorman | b459722 | 2013-07-03 15:02:05 -0700 | [diff] [blame] | 898 | struct address_space *mapping; |
| 899 | |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 900 | /* |
| 901 | * Anonymous pages are not handled by flushers and must be written |
| 902 | * from reclaim context. Do not stall reclaim based on them |
| 903 | */ |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 904 | if (!page_is_file_cache(page) || |
| 905 | (PageAnon(page) && !PageSwapBacked(page))) { |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 906 | *dirty = false; |
| 907 | *writeback = false; |
| 908 | return; |
| 909 | } |
| 910 | |
| 911 | /* By default assume that the page flags are accurate */ |
| 912 | *dirty = PageDirty(page); |
| 913 | *writeback = PageWriteback(page); |
Mel Gorman | b459722 | 2013-07-03 15:02:05 -0700 | [diff] [blame] | 914 | |
| 915 | /* Verify dirty/writeback state if the filesystem supports it */ |
| 916 | if (!page_has_private(page)) |
| 917 | return; |
| 918 | |
| 919 | mapping = page_mapping(page); |
| 920 | if (mapping && mapping->a_ops->is_dirty_writeback) |
| 921 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 922 | } |
| 923 | |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 924 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 925 | * shrink_page_list() returns the number of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 926 | */ |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 927 | static unsigned long shrink_page_list(struct list_head *page_list, |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 928 | struct pglist_data *pgdat, |
Mel Gorman | f84f6e2 | 2011-10-31 17:07:51 -0700 | [diff] [blame] | 929 | struct scan_control *sc, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 930 | enum ttu_flags ttu_flags, |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 931 | struct reclaim_stat *stat, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 932 | bool force_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 933 | { |
| 934 | LIST_HEAD(ret_pages); |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 935 | LIST_HEAD(free_pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 936 | int pgactivate = 0; |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 937 | unsigned nr_unqueued_dirty = 0; |
| 938 | unsigned nr_dirty = 0; |
| 939 | unsigned nr_congested = 0; |
| 940 | unsigned nr_reclaimed = 0; |
| 941 | unsigned nr_writeback = 0; |
| 942 | unsigned nr_immediate = 0; |
Michal Hocko | 5bccd16 | 2017-02-22 15:44:30 -0800 | [diff] [blame] | 943 | unsigned nr_ref_keep = 0; |
| 944 | unsigned nr_unmap_fail = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 945 | |
| 946 | cond_resched(); |
| 947 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 948 | while (!list_empty(page_list)) { |
| 949 | struct address_space *mapping; |
| 950 | struct page *page; |
| 951 | int may_enter_fs; |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 952 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 953 | bool dirty, writeback; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 954 | |
| 955 | cond_resched(); |
| 956 | |
| 957 | page = lru_to_page(page_list); |
| 958 | list_del(&page->lru); |
| 959 | |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 960 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 961 | goto keep; |
| 962 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 963 | VM_BUG_ON_PAGE(PageActive(page), page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 964 | |
| 965 | sc->nr_scanned++; |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 966 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 967 | if (unlikely(!page_evictable(page))) |
Minchan Kim | ad6b670 | 2017-05-03 14:54:13 -0700 | [diff] [blame] | 968 | goto activate_locked; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 969 | |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 970 | if (!sc->may_unmap && page_mapped(page)) |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 971 | goto keep_locked; |
| 972 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 973 | /* Double the slab pressure for mapped and swapcache pages */ |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 974 | if ((page_mapped(page) || PageSwapCache(page)) && |
| 975 | !(PageAnon(page) && !PageSwapBacked(page))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 976 | sc->nr_scanned++; |
| 977 | |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 978 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
| 979 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); |
| 980 | |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 981 | /* |
Andrey Ryabinin | 894befe | 2018-04-10 16:27:51 -0700 | [diff] [blame] | 982 | * The number of dirty pages determines if a node is marked |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 983 | * reclaim_congested which affects wait_iff_congested. kswapd |
| 984 | * will stall and start writing pages if the tail of the LRU |
| 985 | * is all dirty unqueued pages. |
| 986 | */ |
| 987 | page_check_dirty_writeback(page, &dirty, &writeback); |
| 988 | if (dirty || writeback) |
| 989 | nr_dirty++; |
| 990 | |
| 991 | if (dirty && !writeback) |
| 992 | nr_unqueued_dirty++; |
| 993 | |
Mel Gorman | d04e8ac | 2013-07-03 15:02:03 -0700 | [diff] [blame] | 994 | /* |
| 995 | * Treat this page as congested if the underlying BDI is or if |
| 996 | * pages are cycling through the LRU so quickly that the |
| 997 | * pages marked for immediate reclaim are making it to the |
| 998 | * end of the LRU a second time. |
| 999 | */ |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1000 | mapping = page_mapping(page); |
Jamie Liu | 1da58ee | 2014-12-10 15:43:20 -0800 | [diff] [blame] | 1001 | if (((dirty || writeback) && mapping && |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 1002 | inode_write_congested(mapping->host)) || |
Mel Gorman | d04e8ac | 2013-07-03 15:02:03 -0700 | [diff] [blame] | 1003 | (writeback && PageReclaim(page))) |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1004 | nr_congested++; |
| 1005 | |
| 1006 | /* |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1007 | * If a page at the tail of the LRU is under writeback, there |
| 1008 | * are three cases to consider. |
| 1009 | * |
| 1010 | * 1) If reclaim is encountering an excessive number of pages |
| 1011 | * under writeback and this page is both under writeback and |
| 1012 | * PageReclaim then it indicates that pages are being queued |
| 1013 | * for IO but are being recycled through the LRU before the |
| 1014 | * IO can complete. Waiting on the page itself risks an |
| 1015 | * indefinite stall if it is impossible to writeback the |
| 1016 | * page due to IO error or disconnected storage so instead |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1017 | * note that the LRU is being scanned too quickly and the |
| 1018 | * caller can stall after page list has been processed. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1019 | * |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1020 | * 2) Global or new memcg reclaim encounters a page that is |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1021 | * not marked for immediate reclaim, or the caller does not |
| 1022 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, |
| 1023 | * not to fs). In this case mark the page for immediate |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1024 | * reclaim and continue scanning. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1025 | * |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1026 | * Require may_enter_fs because we would wait on fs, which |
| 1027 | * may not have submitted IO yet. And the loop driver might |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1028 | * enter reclaim, and deadlock if it waits on a page for |
| 1029 | * which it is needed to do the write (loop masks off |
| 1030 | * __GFP_IO|__GFP_FS for this reason); but more thought |
| 1031 | * would probably show more reasons. |
| 1032 | * |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1033 | * 3) Legacy memcg encounters a page that is already marked |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1034 | * PageReclaim. memcg does not have any dirty pages |
| 1035 | * throttling so we could easily OOM just because too many |
| 1036 | * pages are in writeback and there is nothing else to |
| 1037 | * reclaim. Wait for the writeback to complete. |
Johannes Weiner | c55e8d0 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1038 | * |
| 1039 | * In cases 1) and 2) we activate the pages to get them out of |
| 1040 | * the way while we continue scanning for clean pages on the |
| 1041 | * inactive list and refilling from the active list. The |
| 1042 | * observation here is that waiting for disk writes is more |
| 1043 | * expensive than potentially causing reloads down the line. |
| 1044 | * Since they're marked for immediate reclaim, they won't put |
| 1045 | * memory pressure on the cache working set any longer than it |
| 1046 | * takes to write them to disk. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1047 | */ |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 1048 | if (PageWriteback(page)) { |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1049 | /* Case 1 above */ |
| 1050 | if (current_is_kswapd() && |
| 1051 | PageReclaim(page) && |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1052 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1053 | nr_immediate++; |
Johannes Weiner | c55e8d0 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1054 | goto activate_locked; |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1055 | |
| 1056 | /* Case 2 above */ |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1057 | } else if (sane_reclaim(sc) || |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1058 | !PageReclaim(page) || !may_enter_fs) { |
Hugh Dickins | c3b94f4 | 2012-07-31 16:45:59 -0700 | [diff] [blame] | 1059 | /* |
| 1060 | * This is slightly racy - end_page_writeback() |
| 1061 | * might have just cleared PageReclaim, then |
| 1062 | * setting PageReclaim here end up interpreted |
| 1063 | * as PageReadahead - but that does not matter |
| 1064 | * enough to care. What we do want is for this |
| 1065 | * page to have PageReclaim set next time memcg |
| 1066 | * reclaim reaches the tests above, so it will |
| 1067 | * then wait_on_page_writeback() to avoid OOM; |
| 1068 | * and it's also appropriate in global reclaim. |
| 1069 | */ |
| 1070 | SetPageReclaim(page); |
Michal Hocko | e62e384 | 2012-07-31 16:45:55 -0700 | [diff] [blame] | 1071 | nr_writeback++; |
Johannes Weiner | c55e8d0 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1072 | goto activate_locked; |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1073 | |
| 1074 | /* Case 3 above */ |
| 1075 | } else { |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1076 | unlock_page(page); |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1077 | wait_on_page_writeback(page); |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1078 | /* then go back and try same page again */ |
| 1079 | list_add_tail(&page->lru, page_list); |
| 1080 | continue; |
Michal Hocko | e62e384 | 2012-07-31 16:45:55 -0700 | [diff] [blame] | 1081 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 1082 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1083 | |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1084 | if (!force_reclaim) |
| 1085 | references = page_check_references(page, sc); |
| 1086 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1087 | switch (references) { |
| 1088 | case PAGEREF_ACTIVATE: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1089 | goto activate_locked; |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 1090 | case PAGEREF_KEEP: |
Michal Hocko | 5bccd16 | 2017-02-22 15:44:30 -0800 | [diff] [blame] | 1091 | nr_ref_keep++; |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 1092 | goto keep_locked; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1093 | case PAGEREF_RECLAIM: |
| 1094 | case PAGEREF_RECLAIM_CLEAN: |
| 1095 | ; /* try to reclaim the page below */ |
| 1096 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1097 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1098 | /* |
| 1099 | * Anonymous process memory has backing store? |
| 1100 | * Try to allocate it some swap space here. |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1101 | * Lazyfree page could be freed directly |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1102 | */ |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 1103 | if (PageAnon(page) && PageSwapBacked(page)) { |
| 1104 | if (!PageSwapCache(page)) { |
| 1105 | if (!(sc->gfp_mask & __GFP_IO)) |
| 1106 | goto keep_locked; |
| 1107 | if (PageTransHuge(page)) { |
| 1108 | /* cannot split THP, skip it */ |
| 1109 | if (!can_split_huge_page(page, NULL)) |
| 1110 | goto activate_locked; |
| 1111 | /* |
| 1112 | * Split pages without a PMD map right |
| 1113 | * away. Chances are some or all of the |
| 1114 | * tail pages can be freed without IO. |
| 1115 | */ |
| 1116 | if (!compound_mapcount(page) && |
| 1117 | split_huge_page_to_list(page, |
| 1118 | page_list)) |
| 1119 | goto activate_locked; |
| 1120 | } |
| 1121 | if (!add_to_swap(page)) { |
| 1122 | if (!PageTransHuge(page)) |
| 1123 | goto activate_locked; |
| 1124 | /* Fallback to swap normal pages */ |
| 1125 | if (split_huge_page_to_list(page, |
| 1126 | page_list)) |
| 1127 | goto activate_locked; |
Huang Ying | fe490cc | 2017-09-06 16:22:52 -0700 | [diff] [blame] | 1128 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 1129 | count_vm_event(THP_SWPOUT_FALLBACK); |
| 1130 | #endif |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 1131 | if (!add_to_swap(page)) |
| 1132 | goto activate_locked; |
| 1133 | } |
Minchan Kim | 0f07465 | 2017-07-06 15:37:24 -0700 | [diff] [blame] | 1134 | |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 1135 | may_enter_fs = 1; |
| 1136 | |
| 1137 | /* Adding to swap updated mapping */ |
| 1138 | mapping = page_mapping(page); |
Minchan Kim | 0f07465 | 2017-07-06 15:37:24 -0700 | [diff] [blame] | 1139 | } |
Kirill A. Shutemov | 7751b2d | 2016-07-26 15:25:56 -0700 | [diff] [blame] | 1140 | } else if (unlikely(PageTransHuge(page))) { |
| 1141 | /* Split file THP */ |
| 1142 | if (split_huge_page_to_list(page, page_list)) |
| 1143 | goto keep_locked; |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1144 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1145 | |
| 1146 | /* |
| 1147 | * The page is mapped into the page tables of one or more |
| 1148 | * processes. Try to unmap it here. |
| 1149 | */ |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1150 | if (page_mapped(page)) { |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 1151 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
| 1152 | |
| 1153 | if (unlikely(PageTransHuge(page))) |
| 1154 | flags |= TTU_SPLIT_HUGE_PMD; |
| 1155 | if (!try_to_unmap(page, flags)) { |
Michal Hocko | 5bccd16 | 2017-02-22 15:44:30 -0800 | [diff] [blame] | 1156 | nr_unmap_fail++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1157 | goto activate_locked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1158 | } |
| 1159 | } |
| 1160 | |
| 1161 | if (PageDirty(page)) { |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1162 | /* |
Johannes Weiner | 4eda482 | 2017-02-24 14:56:20 -0800 | [diff] [blame] | 1163 | * Only kswapd can writeback filesystem pages |
| 1164 | * to avoid risk of stack overflow. But avoid |
| 1165 | * injecting inefficient single-page IO into |
| 1166 | * flusher writeback as much as possible: only |
| 1167 | * write pages when we've encountered many |
| 1168 | * dirty pages, and when we've already scanned |
| 1169 | * the rest of the LRU for clean pages and see |
| 1170 | * the same dirty pages again (PageReclaim). |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1171 | */ |
Mel Gorman | f84f6e2 | 2011-10-31 17:07:51 -0700 | [diff] [blame] | 1172 | if (page_is_file_cache(page) && |
Johannes Weiner | 4eda482 | 2017-02-24 14:56:20 -0800 | [diff] [blame] | 1173 | (!current_is_kswapd() || !PageReclaim(page) || |
| 1174 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
Mel Gorman | 49ea7eb | 2011-10-31 17:07:59 -0700 | [diff] [blame] | 1175 | /* |
| 1176 | * Immediately reclaim when written back. |
| 1177 | * Similar in principal to deactivate_page() |
| 1178 | * except we already have the page isolated |
| 1179 | * and know it's dirty |
| 1180 | */ |
Mel Gorman | c4a2563 | 2016-07-28 15:46:23 -0700 | [diff] [blame] | 1181 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
Mel Gorman | 49ea7eb | 2011-10-31 17:07:59 -0700 | [diff] [blame] | 1182 | SetPageReclaim(page); |
| 1183 | |
Johannes Weiner | c55e8d0 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1184 | goto activate_locked; |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1185 | } |
| 1186 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1187 | if (references == PAGEREF_RECLAIM_CLEAN) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1188 | goto keep_locked; |
Andrew Morton | 4dd4b92 | 2008-03-24 12:29:52 -0700 | [diff] [blame] | 1189 | if (!may_enter_fs) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1190 | goto keep_locked; |
Christoph Lameter | 52a8363 | 2006-02-01 03:05:28 -0800 | [diff] [blame] | 1191 | if (!sc->may_writepage) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1192 | goto keep_locked; |
| 1193 | |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 1194 | /* |
| 1195 | * Page is dirty. Flush the TLB if a writable entry |
| 1196 | * potentially exists to avoid CPU writes after IO |
| 1197 | * starts and then write it out here. |
| 1198 | */ |
| 1199 | try_to_unmap_flush_dirty(); |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1200 | switch (pageout(page, mapping, sc)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1201 | case PAGE_KEEP: |
| 1202 | goto keep_locked; |
| 1203 | case PAGE_ACTIVATE: |
| 1204 | goto activate_locked; |
| 1205 | case PAGE_SUCCESS: |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1206 | if (PageWriteback(page)) |
Mel Gorman | 41ac199 | 2012-05-29 15:06:19 -0700 | [diff] [blame] | 1207 | goto keep; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1208 | if (PageDirty(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1209 | goto keep; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1210 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1211 | /* |
| 1212 | * A synchronous write - probably a ramdisk. Go |
| 1213 | * ahead and try to reclaim the page. |
| 1214 | */ |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 1215 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1216 | goto keep; |
| 1217 | if (PageDirty(page) || PageWriteback(page)) |
| 1218 | goto keep_locked; |
| 1219 | mapping = page_mapping(page); |
| 1220 | case PAGE_CLEAN: |
| 1221 | ; /* try to free the page below */ |
| 1222 | } |
| 1223 | } |
| 1224 | |
| 1225 | /* |
| 1226 | * If the page has buffers, try to free the buffer mappings |
| 1227 | * associated with this page. If we succeed we try to free |
| 1228 | * the page as well. |
| 1229 | * |
| 1230 | * We do this even if the page is PageDirty(). |
| 1231 | * try_to_release_page() does not perform I/O, but it is |
| 1232 | * possible for a page to have PageDirty set, but it is actually |
| 1233 | * clean (all its buffers are clean). This happens if the |
| 1234 | * buffers were written out directly, with submit_bh(). ext3 |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1235 | * will do this, as well as the blockdev mapping. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1236 | * try_to_release_page() will discover that cleanness and will |
| 1237 | * drop the buffers and mark the page clean - it can be freed. |
| 1238 | * |
| 1239 | * Rarely, pages can have buffers and no ->mapping. These are |
| 1240 | * the pages which were not successfully invalidated in |
| 1241 | * truncate_complete_page(). We try to drop those buffers here |
| 1242 | * and if that worked, and the page is no longer mapped into |
| 1243 | * process address space (page_count == 1) it can be freed. |
| 1244 | * Otherwise, leave the page on the LRU so it is swappable. |
| 1245 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 1246 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1247 | if (!try_to_release_page(page, sc->gfp_mask)) |
| 1248 | goto activate_locked; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 1249 | if (!mapping && page_count(page) == 1) { |
| 1250 | unlock_page(page); |
| 1251 | if (put_page_testzero(page)) |
| 1252 | goto free_it; |
| 1253 | else { |
| 1254 | /* |
| 1255 | * rare race with speculative reference. |
| 1256 | * the speculative reference will free |
| 1257 | * this page shortly, so we may |
| 1258 | * increment nr_reclaimed here (and |
| 1259 | * leave it off the LRU). |
| 1260 | */ |
| 1261 | nr_reclaimed++; |
| 1262 | continue; |
| 1263 | } |
| 1264 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1265 | } |
| 1266 | |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1267 | if (PageAnon(page) && !PageSwapBacked(page)) { |
| 1268 | /* follow __remove_mapping for reference */ |
| 1269 | if (!page_ref_freeze(page, 1)) |
| 1270 | goto keep_locked; |
| 1271 | if (PageDirty(page)) { |
| 1272 | page_ref_unfreeze(page, 1); |
| 1273 | goto keep_locked; |
| 1274 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1275 | |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1276 | count_vm_event(PGLAZYFREED); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1277 | count_memcg_page_event(page, PGLAZYFREED); |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1278 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
| 1279 | goto keep_locked; |
Nick Piggin | a978d6f | 2008-10-18 20:26:58 -0700 | [diff] [blame] | 1280 | /* |
| 1281 | * At this point, we have no other references and there is |
| 1282 | * no way to pick any more up (removed from LRU, removed |
| 1283 | * from pagecache). Can use non-atomic bitops now (and |
| 1284 | * we obviously don't have to worry about waking up a process |
| 1285 | * waiting on the page lock, because there are no references. |
| 1286 | */ |
Kirill A. Shutemov | 48c935a | 2016-01-15 16:51:24 -0800 | [diff] [blame] | 1287 | __ClearPageLocked(page); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 1288 | free_it: |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1289 | nr_reclaimed++; |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1290 | |
| 1291 | /* |
| 1292 | * Is there need to periodically free_page_list? It would |
| 1293 | * appear not as the counts should be low |
| 1294 | */ |
Huang Ying | bd4c82c2 | 2017-09-06 16:22:49 -0700 | [diff] [blame] | 1295 | if (unlikely(PageTransHuge(page))) { |
| 1296 | mem_cgroup_uncharge(page); |
| 1297 | (*get_compound_page_dtor(page))(page); |
| 1298 | } else |
| 1299 | list_add(&page->lru, &free_pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1300 | continue; |
| 1301 | |
| 1302 | activate_locked: |
Rik van Riel | 68a22394 | 2008-10-18 20:26:23 -0700 | [diff] [blame] | 1303 | /* Not a candidate for swapping, so reclaim swap space. */ |
Minchan Kim | ad6b670 | 2017-05-03 14:54:13 -0700 | [diff] [blame] | 1304 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
| 1305 | PageMlocked(page))) |
Hugh Dickins | a2c43ee | 2009-01-06 14:39:36 -0800 | [diff] [blame] | 1306 | try_to_free_swap(page); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1307 | VM_BUG_ON_PAGE(PageActive(page), page); |
Minchan Kim | ad6b670 | 2017-05-03 14:54:13 -0700 | [diff] [blame] | 1308 | if (!PageMlocked(page)) { |
| 1309 | SetPageActive(page); |
| 1310 | pgactivate++; |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1311 | count_memcg_page_event(page, PGACTIVATE); |
Minchan Kim | ad6b670 | 2017-05-03 14:54:13 -0700 | [diff] [blame] | 1312 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1313 | keep_locked: |
| 1314 | unlock_page(page); |
| 1315 | keep: |
| 1316 | list_add(&page->lru, &ret_pages); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1317 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1318 | } |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1319 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1320 | mem_cgroup_uncharge_list(&free_pages); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 1321 | try_to_unmap_flush(); |
Mel Gorman | 2d4894b | 2017-11-15 17:37:59 -0800 | [diff] [blame] | 1322 | free_unref_page_list(&free_pages); |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1323 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1324 | list_splice(&ret_pages, page_list); |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 1325 | count_vm_events(PGACTIVATE, pgactivate); |
Johannes Weiner | 0a31bc9 | 2014-08-08 14:19:22 -0700 | [diff] [blame] | 1326 | |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 1327 | if (stat) { |
| 1328 | stat->nr_dirty = nr_dirty; |
| 1329 | stat->nr_congested = nr_congested; |
| 1330 | stat->nr_unqueued_dirty = nr_unqueued_dirty; |
| 1331 | stat->nr_writeback = nr_writeback; |
| 1332 | stat->nr_immediate = nr_immediate; |
Michal Hocko | 5bccd16 | 2017-02-22 15:44:30 -0800 | [diff] [blame] | 1333 | stat->nr_activate = pgactivate; |
| 1334 | stat->nr_ref_keep = nr_ref_keep; |
| 1335 | stat->nr_unmap_fail = nr_unmap_fail; |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 1336 | } |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1337 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1338 | } |
| 1339 | |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1340 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
| 1341 | struct list_head *page_list) |
| 1342 | { |
| 1343 | struct scan_control sc = { |
| 1344 | .gfp_mask = GFP_KERNEL, |
| 1345 | .priority = DEF_PRIORITY, |
| 1346 | .may_unmap = 1, |
| 1347 | }; |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 1348 | unsigned long ret; |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1349 | struct page *page, *next; |
| 1350 | LIST_HEAD(clean_pages); |
| 1351 | |
| 1352 | list_for_each_entry_safe(page, next, page_list, lru) { |
Rafael Aquini | 117aad1 | 2013-09-30 13:45:16 -0700 | [diff] [blame] | 1353 | if (page_is_file_cache(page) && !PageDirty(page) && |
Minchan Kim | b1123ea6 | 2016-07-26 15:23:09 -0700 | [diff] [blame] | 1354 | !__PageMovable(page)) { |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1355 | ClearPageActive(page); |
| 1356 | list_move(&page->lru, &clean_pages); |
| 1357 | } |
| 1358 | } |
| 1359 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1360 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
Shaohua Li | a128ca7 | 2017-05-03 14:52:22 -0700 | [diff] [blame] | 1361 | TTU_IGNORE_ACCESS, NULL, true); |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1362 | list_splice(&clean_pages, page_list); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1363 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1364 | return ret; |
| 1365 | } |
| 1366 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1367 | /* |
| 1368 | * Attempt to remove the specified page from its LRU. Only take this page |
| 1369 | * if it is of the appropriate PageActive status. Pages which are being |
| 1370 | * freed elsewhere are also ignored. |
| 1371 | * |
| 1372 | * page: page to consider |
| 1373 | * mode: one of the LRU isolation modes defined above |
| 1374 | * |
| 1375 | * returns 0 on success, -ve errno on failure. |
| 1376 | */ |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1377 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1378 | { |
| 1379 | int ret = -EINVAL; |
| 1380 | |
| 1381 | /* Only take pages on the LRU. */ |
| 1382 | if (!PageLRU(page)) |
| 1383 | return ret; |
| 1384 | |
Minchan Kim | e46a287 | 2012-10-08 16:33:48 -0700 | [diff] [blame] | 1385 | /* Compaction should not handle unevictable pages but CMA can do so */ |
| 1386 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1387 | return ret; |
| 1388 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1389 | ret = -EBUSY; |
KAMEZAWA Hiroyuki | 08e552c | 2009-01-07 18:08:01 -0800 | [diff] [blame] | 1390 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1391 | /* |
| 1392 | * To minimise LRU disruption, the caller can indicate that it only |
| 1393 | * wants to isolate pages it will be able to operate on without |
| 1394 | * blocking - clean pages for the most part. |
| 1395 | * |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1396 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
| 1397 | * that it is possible to migrate without blocking |
| 1398 | */ |
Johannes Weiner | 1276ad6 | 2017-02-24 14:56:11 -0800 | [diff] [blame] | 1399 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1400 | /* All the caller can do on PageWriteback is block */ |
| 1401 | if (PageWriteback(page)) |
| 1402 | return ret; |
| 1403 | |
| 1404 | if (PageDirty(page)) { |
| 1405 | struct address_space *mapping; |
Mel Gorman | 69d763f | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1406 | bool migrate_dirty; |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1407 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1408 | /* |
| 1409 | * Only pages without mappings or that have a |
| 1410 | * ->migratepage callback are possible to migrate |
Mel Gorman | 69d763f | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1411 | * without blocking. However, we can be racing with |
| 1412 | * truncation so it's necessary to lock the page |
| 1413 | * to stabilise the mapping as truncation holds |
| 1414 | * the page lock until after the page is removed |
| 1415 | * from the page cache. |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1416 | */ |
Mel Gorman | 69d763f | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1417 | if (!trylock_page(page)) |
| 1418 | return ret; |
| 1419 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1420 | mapping = page_mapping(page); |
Hugh Dickins | 145e1a7 | 2018-06-01 16:50:50 -0700 | [diff] [blame^] | 1421 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
Mel Gorman | 69d763f | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1422 | unlock_page(page); |
| 1423 | if (!migrate_dirty) |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1424 | return ret; |
| 1425 | } |
| 1426 | } |
Minchan Kim | 39deaf8 | 2011-10-31 17:06:51 -0700 | [diff] [blame] | 1427 | |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1428 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
| 1429 | return ret; |
| 1430 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1431 | if (likely(get_page_unless_zero(page))) { |
| 1432 | /* |
| 1433 | * Be careful not to clear PageLRU until after we're |
| 1434 | * sure the page is not being freed elsewhere -- the |
| 1435 | * page release code relies on it. |
| 1436 | */ |
| 1437 | ClearPageLRU(page); |
| 1438 | ret = 0; |
| 1439 | } |
| 1440 | |
| 1441 | return ret; |
| 1442 | } |
| 1443 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1444 | |
| 1445 | /* |
| 1446 | * Update LRU sizes after isolating pages. The LRU size updates must |
| 1447 | * be complete before mem_cgroup_update_lru_size due to a santity check. |
| 1448 | */ |
| 1449 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, |
Michal Hocko | b4536f0c8 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1450 | enum lru_list lru, unsigned long *nr_zone_taken) |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1451 | { |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1452 | int zid; |
| 1453 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1454 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
| 1455 | if (!nr_zone_taken[zid]) |
| 1456 | continue; |
| 1457 | |
| 1458 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
Michal Hocko | b4536f0c8 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1459 | #ifdef CONFIG_MEMCG |
| 1460 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
| 1461 | #endif |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1462 | } |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1463 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1464 | } |
| 1465 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 1466 | /* |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1467 | * zone_lru_lock is heavily contended. Some of the functions that |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1468 | * shrink the lists perform better by taking out a batch of pages |
| 1469 | * and working on them outside the LRU lock. |
| 1470 | * |
| 1471 | * For pagecache intensive workloads, this function is the hottest |
| 1472 | * spot in the kernel (apart from copy_*_user functions). |
| 1473 | * |
| 1474 | * Appropriate locks must be held before calling this function. |
| 1475 | * |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1476 | * @nr_to_scan: The number of eligible pages to look through on the list. |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1477 | * @lruvec: The LRU vector to pull pages from. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1478 | * @dst: The temp list to put pages on to. |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 1479 | * @nr_scanned: The number of pages that were scanned. |
Rik van Riel | fe2c2a1 | 2012-03-21 16:33:51 -0700 | [diff] [blame] | 1480 | * @sc: The scan_control struct for this reclaim session |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1481 | * @mode: One of the LRU isolation modes |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1482 | * @lru: LRU list id for isolating |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1483 | * |
| 1484 | * returns how many pages were moved onto *@dst. |
| 1485 | */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1486 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1487 | struct lruvec *lruvec, struct list_head *dst, |
Rik van Riel | fe2c2a1 | 2012-03-21 16:33:51 -0700 | [diff] [blame] | 1488 | unsigned long *nr_scanned, struct scan_control *sc, |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1489 | isolate_mode_t mode, enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1490 | { |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1491 | struct list_head *src = &lruvec->lists[lru]; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1492 | unsigned long nr_taken = 0; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1493 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1494 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
Johannes Weiner | 3db6581 | 2017-05-03 14:52:13 -0700 | [diff] [blame] | 1495 | unsigned long skipped = 0; |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1496 | unsigned long scan, total_scan, nr_pages; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1497 | LIST_HEAD(pages_skipped); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1498 | |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1499 | scan = 0; |
| 1500 | for (total_scan = 0; |
| 1501 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); |
| 1502 | total_scan++) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1503 | struct page *page; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1504 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1505 | page = lru_to_page(src); |
| 1506 | prefetchw_prev_lru_page(page, src, flags); |
| 1507 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1508 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
Nick Piggin | 8d438f9 | 2006-03-22 00:07:59 -0800 | [diff] [blame] | 1509 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1510 | if (page_zonenum(page) > sc->reclaim_idx) { |
| 1511 | list_move(&page->lru, &pages_skipped); |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1512 | nr_skipped[page_zonenum(page)]++; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1513 | continue; |
| 1514 | } |
| 1515 | |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1516 | /* |
| 1517 | * Do not count skipped pages because that makes the function |
| 1518 | * return with no isolated pages if the LRU mostly contains |
| 1519 | * ineligible pages. This causes the VM to not reclaim any |
| 1520 | * pages, triggering a premature OOM. |
| 1521 | */ |
| 1522 | scan++; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1523 | switch (__isolate_lru_page(page, mode)) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1524 | case 0: |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1525 | nr_pages = hpage_nr_pages(page); |
| 1526 | nr_taken += nr_pages; |
| 1527 | nr_zone_taken[page_zonenum(page)] += nr_pages; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1528 | list_move(&page->lru, dst); |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1529 | break; |
Nick Piggin | 46453a6 | 2006-03-22 00:07:58 -0800 | [diff] [blame] | 1530 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1531 | case -EBUSY: |
| 1532 | /* else it is being freed elsewhere */ |
| 1533 | list_move(&page->lru, src); |
| 1534 | continue; |
| 1535 | |
| 1536 | default: |
| 1537 | BUG(); |
| 1538 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1539 | } |
| 1540 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1541 | /* |
| 1542 | * Splice any skipped pages to the start of the LRU list. Note that |
| 1543 | * this disrupts the LRU order when reclaiming for lower zones but |
| 1544 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX |
| 1545 | * scanning would soon rescan the same pages to skip and put the |
| 1546 | * system at risk of premature OOM. |
| 1547 | */ |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1548 | if (!list_empty(&pages_skipped)) { |
| 1549 | int zid; |
| 1550 | |
Johannes Weiner | 3db6581 | 2017-05-03 14:52:13 -0700 | [diff] [blame] | 1551 | list_splice(&pages_skipped, src); |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1552 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
| 1553 | if (!nr_skipped[zid]) |
| 1554 | continue; |
| 1555 | |
| 1556 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); |
Michal Hocko | 1265e3a | 2017-02-22 15:44:21 -0800 | [diff] [blame] | 1557 | skipped += nr_skipped[zid]; |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1558 | } |
| 1559 | } |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1560 | *nr_scanned = total_scan; |
Michal Hocko | 1265e3a | 2017-02-22 15:44:21 -0800 | [diff] [blame] | 1561 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
Minchan Kim | 791b48b | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1562 | total_scan, skipped, nr_taken, mode, lru); |
Michal Hocko | b4536f0c8 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1563 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1564 | return nr_taken; |
| 1565 | } |
| 1566 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1567 | /** |
| 1568 | * isolate_lru_page - tries to isolate a page from its LRU list |
| 1569 | * @page: page to isolate from its LRU list |
| 1570 | * |
| 1571 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the |
| 1572 | * vmstat statistic corresponding to whatever LRU list the page was on. |
| 1573 | * |
| 1574 | * Returns 0 if the page was removed from an LRU list. |
| 1575 | * Returns -EBUSY if the page was not on an LRU list. |
| 1576 | * |
| 1577 | * The returned page will have PageLRU() cleared. If it was found on |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1578 | * the active list, it will have PageActive set. If it was found on |
| 1579 | * the unevictable list, it will have the PageUnevictable bit set. That flag |
| 1580 | * may need to be cleared by the caller before letting the page go. |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1581 | * |
| 1582 | * The vmstat statistic corresponding to the list on which the page was |
| 1583 | * found will be decremented. |
| 1584 | * |
| 1585 | * Restrictions: |
Mike Rapoport | a5d09be | 2018-02-06 15:42:19 -0800 | [diff] [blame] | 1586 | * |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1587 | * (1) Must be called with an elevated refcount on the page. This is a |
| 1588 | * fundamentnal difference from isolate_lru_pages (which is called |
| 1589 | * without a stable reference). |
| 1590 | * (2) the lru_lock must not be held. |
| 1591 | * (3) interrupts must be enabled. |
| 1592 | */ |
| 1593 | int isolate_lru_page(struct page *page) |
| 1594 | { |
| 1595 | int ret = -EBUSY; |
| 1596 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1597 | VM_BUG_ON_PAGE(!page_count(page), page); |
Kirill A. Shutemov | cf2a82e | 2016-02-05 15:36:36 -0800 | [diff] [blame] | 1598 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1599 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1600 | if (PageLRU(page)) { |
| 1601 | struct zone *zone = page_zone(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1602 | struct lruvec *lruvec; |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1603 | |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1604 | spin_lock_irq(zone_lru_lock(zone)); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1605 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1606 | if (PageLRU(page)) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1607 | int lru = page_lru(page); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1608 | get_page(page); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1609 | ClearPageLRU(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1610 | del_page_from_lru_list(page, lruvec, lru); |
| 1611 | ret = 0; |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1612 | } |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1613 | spin_unlock_irq(zone_lru_lock(zone)); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1614 | } |
| 1615 | return ret; |
| 1616 | } |
| 1617 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1618 | /* |
Fengguang Wu | d37dd5d | 2012-12-18 14:23:28 -0800 | [diff] [blame] | 1619 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
| 1620 | * then get resheduled. When there are massive number of tasks doing page |
| 1621 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
| 1622 | * the LRU list will go small and be scanned faster than necessary, leading to |
| 1623 | * unnecessary swapping, thrashing and OOM. |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1624 | */ |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1625 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1626 | struct scan_control *sc) |
| 1627 | { |
| 1628 | unsigned long inactive, isolated; |
| 1629 | |
| 1630 | if (current_is_kswapd()) |
| 1631 | return 0; |
| 1632 | |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1633 | if (!sane_reclaim(sc)) |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1634 | return 0; |
| 1635 | |
| 1636 | if (file) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1637 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
| 1638 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1639 | } else { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1640 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
| 1641 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1642 | } |
| 1643 | |
Fengguang Wu | 3cf2384 | 2012-12-18 14:23:31 -0800 | [diff] [blame] | 1644 | /* |
| 1645 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they |
| 1646 | * won't get blocked by normal direct-reclaimers, forming a circular |
| 1647 | * deadlock. |
| 1648 | */ |
Mel Gorman | d0164ad | 2015-11-06 16:28:21 -0800 | [diff] [blame] | 1649 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
Fengguang Wu | 3cf2384 | 2012-12-18 14:23:31 -0800 | [diff] [blame] | 1650 | inactive >>= 3; |
| 1651 | |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1652 | return isolated > inactive; |
| 1653 | } |
| 1654 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1655 | static noinline_for_stack void |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1656 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1657 | { |
Konstantin Khlebnikov | 27ac81d | 2012-05-29 15:07:00 -0700 | [diff] [blame] | 1658 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1659 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1660 | LIST_HEAD(pages_to_free); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1661 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1662 | /* |
| 1663 | * Put back any unfreeable pages. |
| 1664 | */ |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1665 | while (!list_empty(page_list)) { |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1666 | struct page *page = lru_to_page(page_list); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1667 | int lru; |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1668 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1669 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1670 | list_del(&page->lru); |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 1671 | if (unlikely(!page_evictable(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1672 | spin_unlock_irq(&pgdat->lru_lock); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1673 | putback_lru_page(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1674 | spin_lock_irq(&pgdat->lru_lock); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1675 | continue; |
| 1676 | } |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1677 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1678 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1679 | |
Linus Torvalds | 7a60857 | 2011-01-17 14:42:19 -0800 | [diff] [blame] | 1680 | SetPageLRU(page); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1681 | lru = page_lru(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1682 | add_page_to_lru_list(page, lruvec, lru); |
| 1683 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1684 | if (is_active_lru(lru)) { |
| 1685 | int file = is_file_lru(lru); |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1686 | int numpages = hpage_nr_pages(page); |
| 1687 | reclaim_stat->recent_rotated[file] += numpages; |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1688 | } |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1689 | if (put_page_testzero(page)) { |
| 1690 | __ClearPageLRU(page); |
| 1691 | __ClearPageActive(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1692 | del_page_from_lru_list(page, lruvec, lru); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1693 | |
| 1694 | if (unlikely(PageCompound(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1695 | spin_unlock_irq(&pgdat->lru_lock); |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1696 | mem_cgroup_uncharge(page); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1697 | (*get_compound_page_dtor(page))(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1698 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1699 | } else |
| 1700 | list_add(&page->lru, &pages_to_free); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1701 | } |
| 1702 | } |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1703 | |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1704 | /* |
| 1705 | * To save our caller's stack, now use input list for pages to free. |
| 1706 | */ |
| 1707 | list_splice(&pages_to_free, page_list); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1708 | } |
| 1709 | |
| 1710 | /* |
NeilBrown | 399ba0b | 2014-06-04 16:07:42 -0700 | [diff] [blame] | 1711 | * If a kernel thread (such as nfsd for loop-back mounts) services |
| 1712 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. |
| 1713 | * In that case we should only throttle if the backing device it is |
| 1714 | * writing to is congested. In other cases it is safe to throttle. |
| 1715 | */ |
| 1716 | static int current_may_throttle(void) |
| 1717 | { |
| 1718 | return !(current->flags & PF_LESS_THROTTLE) || |
| 1719 | current->backing_dev_info == NULL || |
| 1720 | bdi_write_congested(current->backing_dev_info); |
| 1721 | } |
| 1722 | |
| 1723 | /* |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1724 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 1725 | * of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1726 | */ |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1727 | static noinline_for_stack unsigned long |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1728 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 1729 | struct scan_control *sc, enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1730 | { |
| 1731 | LIST_HEAD(page_list); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1732 | unsigned long nr_scanned; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1733 | unsigned long nr_reclaimed = 0; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1734 | unsigned long nr_taken; |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 1735 | struct reclaim_stat stat = {}; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1736 | isolate_mode_t isolate_mode = 0; |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1737 | int file = is_file_lru(lru); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1738 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1739 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Michal Hocko | db73ee0 | 2017-09-06 16:21:11 -0700 | [diff] [blame] | 1740 | bool stalled = false; |
KOSAKI Motohiro | 78dc583 | 2009-06-16 15:31:40 -0700 | [diff] [blame] | 1741 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1742 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
Michal Hocko | db73ee0 | 2017-09-06 16:21:11 -0700 | [diff] [blame] | 1743 | if (stalled) |
| 1744 | return 0; |
| 1745 | |
| 1746 | /* wait a bit for the reclaimer. */ |
| 1747 | msleep(100); |
| 1748 | stalled = true; |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1749 | |
| 1750 | /* We are about to die and free our memory. Return now. */ |
| 1751 | if (fatal_signal_pending(current)) |
| 1752 | return SWAP_CLUSTER_MAX; |
| 1753 | } |
| 1754 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1755 | lru_add_drain(); |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1756 | |
| 1757 | if (!sc->may_unmap) |
Hillf Danton | 6131728 | 2012-03-21 16:33:48 -0700 | [diff] [blame] | 1758 | isolate_mode |= ISOLATE_UNMAPPED; |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1759 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1760 | spin_lock_irq(&pgdat->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1761 | |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1762 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
| 1763 | &nr_scanned, sc, isolate_mode, lru); |
Konstantin Khlebnikov | 95d918f | 2012-05-29 15:06:59 -0700 | [diff] [blame] | 1764 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1765 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 1766 | reclaim_stat->recent_scanned[file] += nr_taken; |
Konstantin Khlebnikov | 95d918f | 2012-05-29 15:06:59 -0700 | [diff] [blame] | 1767 | |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1768 | if (current_is_kswapd()) { |
| 1769 | if (global_reclaim(sc)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1770 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1771 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, |
| 1772 | nr_scanned); |
| 1773 | } else { |
| 1774 | if (global_reclaim(sc)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1775 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1776 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, |
| 1777 | nr_scanned); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1778 | } |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1779 | spin_unlock_irq(&pgdat->lru_lock); |
KOSAKI Motohiro | a731286 | 2009-09-21 17:01:37 -0700 | [diff] [blame] | 1780 | |
Hillf Danton | d563c05 | 2012-03-21 16:34:02 -0700 | [diff] [blame] | 1781 | if (nr_taken == 0) |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1782 | return 0; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1783 | |
Shaohua Li | a128ca7 | 2017-05-03 14:52:22 -0700 | [diff] [blame] | 1784 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
Michal Hocko | 3c710c1 | 2017-02-22 15:44:27 -0800 | [diff] [blame] | 1785 | &stat, false); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1786 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1787 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1788 | |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1789 | if (current_is_kswapd()) { |
| 1790 | if (global_reclaim(sc)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1791 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1792 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, |
| 1793 | nr_reclaimed); |
| 1794 | } else { |
| 1795 | if (global_reclaim(sc)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1796 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1797 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, |
| 1798 | nr_reclaimed); |
Ying Han | 904249a | 2012-04-25 16:01:48 -0700 | [diff] [blame] | 1799 | } |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1800 | |
Konstantin Khlebnikov | 27ac81d | 2012-05-29 15:07:00 -0700 | [diff] [blame] | 1801 | putback_inactive_pages(lruvec, &page_list); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1802 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1803 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1804 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1805 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1806 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1807 | mem_cgroup_uncharge_list(&page_list); |
Mel Gorman | 2d4894b | 2017-11-15 17:37:59 -0800 | [diff] [blame] | 1808 | free_unref_page_list(&page_list); |
Mel Gorman | e11da5b | 2010-10-26 14:21:40 -0700 | [diff] [blame] | 1809 | |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1810 | /* |
Andrey Ryabinin | 1c610d5 | 2018-03-22 16:17:42 -0700 | [diff] [blame] | 1811 | * If dirty pages are scanned that are not queued for IO, it |
| 1812 | * implies that flushers are not doing their job. This can |
| 1813 | * happen when memory pressure pushes dirty pages to the end of |
| 1814 | * the LRU before the dirty limits are breached and the dirty |
| 1815 | * data has expired. It can also happen when the proportion of |
| 1816 | * dirty pages grows not through writes but through memory |
| 1817 | * pressure reclaiming all the clean cache. And in some cases, |
| 1818 | * the flushers simply cannot keep up with the allocation |
| 1819 | * rate. Nudge the flusher threads in case they are asleep. |
| 1820 | */ |
| 1821 | if (stat.nr_unqueued_dirty == nr_taken) |
| 1822 | wakeup_flusher_threads(WB_REASON_VMSCAN); |
| 1823 | |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 1824 | sc->nr.dirty += stat.nr_dirty; |
| 1825 | sc->nr.congested += stat.nr_congested; |
| 1826 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; |
| 1827 | sc->nr.writeback += stat.nr_writeback; |
| 1828 | sc->nr.immediate += stat.nr_immediate; |
| 1829 | sc->nr.taken += nr_taken; |
| 1830 | if (file) |
| 1831 | sc->nr.file_taken += nr_taken; |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1832 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1833 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
Steven Rostedt | d51d1e6 | 2018-04-10 16:28:07 -0700 | [diff] [blame] | 1834 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1835 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1836 | } |
| 1837 | |
Martin Bligh | 3bb1a852 | 2006-10-28 10:38:24 -0700 | [diff] [blame] | 1838 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1839 | * This moves pages from the active list to the inactive list. |
| 1840 | * |
| 1841 | * We move them the other way if the page is referenced by one or more |
| 1842 | * processes, from rmap. |
| 1843 | * |
| 1844 | * If the pages are mostly unmapped, the processing is fast and it is |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1845 | * appropriate to hold zone_lru_lock across the whole operation. But if |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1846 | * the pages are mapped, the processing is slow (page_referenced()) so we |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1847 | * should drop zone_lru_lock around each page. It's impossible to balance |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1848 | * this, so instead we remove the pages from the LRU while processing them. |
| 1849 | * It is safe to rely on PG_active against the non-LRU pages in here because |
| 1850 | * nobody will play with that bit on a non-LRU page. |
| 1851 | * |
Joonsoo Kim | 0139aa7 | 2016-05-19 17:10:49 -0700 | [diff] [blame] | 1852 | * The downside is that we have to touch page->_refcount against each page. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1853 | * But we had to alter page->flags anyway. |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1854 | * |
| 1855 | * Returns the number of pages moved to the given lru. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1856 | */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1857 | |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1858 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1859 | struct list_head *list, |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1860 | struct list_head *pages_to_free, |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1861 | enum lru_list lru) |
| 1862 | { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1863 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1864 | struct page *page; |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1865 | int nr_pages; |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1866 | int nr_moved = 0; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1867 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1868 | while (!list_empty(list)) { |
| 1869 | page = lru_to_page(list); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1870 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1871 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1872 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1873 | SetPageLRU(page); |
| 1874 | |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1875 | nr_pages = hpage_nr_pages(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1876 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 1877 | list_move(&page->lru, &lruvec->lists[lru]); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1878 | |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1879 | if (put_page_testzero(page)) { |
| 1880 | __ClearPageLRU(page); |
| 1881 | __ClearPageActive(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1882 | del_page_from_lru_list(page, lruvec, lru); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1883 | |
| 1884 | if (unlikely(PageCompound(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1885 | spin_unlock_irq(&pgdat->lru_lock); |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1886 | mem_cgroup_uncharge(page); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1887 | (*get_compound_page_dtor(page))(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1888 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1889 | } else |
| 1890 | list_add(&page->lru, pages_to_free); |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1891 | } else { |
| 1892 | nr_moved += nr_pages; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1893 | } |
| 1894 | } |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 1895 | |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1896 | if (!is_active_lru(lru)) { |
Michal Hocko | f095890 | 2017-02-22 15:45:55 -0800 | [diff] [blame] | 1897 | __count_vm_events(PGDEACTIVATE, nr_moved); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1898 | count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, |
| 1899 | nr_moved); |
| 1900 | } |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1901 | |
| 1902 | return nr_moved; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1903 | } |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1904 | |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 1905 | static void shrink_active_list(unsigned long nr_to_scan, |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1906 | struct lruvec *lruvec, |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 1907 | struct scan_control *sc, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 1908 | enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1909 | { |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 1910 | unsigned long nr_taken; |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 1911 | unsigned long nr_scanned; |
Wu Fengguang | 6fe6b7e | 2009-06-16 15:33:05 -0700 | [diff] [blame] | 1912 | unsigned long vm_flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1913 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1914 | LIST_HEAD(l_active); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 1915 | LIST_HEAD(l_inactive); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1916 | struct page *page; |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1917 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1918 | unsigned nr_deactivate, nr_activate; |
| 1919 | unsigned nr_rotated = 0; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1920 | isolate_mode_t isolate_mode = 0; |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1921 | int file = is_file_lru(lru); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1922 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1923 | |
| 1924 | lru_add_drain(); |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1925 | |
| 1926 | if (!sc->may_unmap) |
Hillf Danton | 6131728 | 2012-03-21 16:33:48 -0700 | [diff] [blame] | 1927 | isolate_mode |= ISOLATE_UNMAPPED; |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1928 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1929 | spin_lock_irq(&pgdat->lru_lock); |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 1930 | |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1931 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
| 1932 | &nr_scanned, sc, isolate_mode, lru); |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 1933 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1934 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 1935 | reclaim_stat->recent_scanned[file] += nr_taken; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1936 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1937 | __count_vm_events(PGREFILL, nr_scanned); |
Roman Gushchin | 2262185 | 2017-07-06 15:40:25 -0700 | [diff] [blame] | 1938 | count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 1939 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1940 | spin_unlock_irq(&pgdat->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1941 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1942 | while (!list_empty(&l_hold)) { |
| 1943 | cond_resched(); |
| 1944 | page = lru_to_page(&l_hold); |
| 1945 | list_del(&page->lru); |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1946 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 1947 | if (unlikely(!page_evictable(page))) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1948 | putback_lru_page(page); |
| 1949 | continue; |
| 1950 | } |
| 1951 | |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1952 | if (unlikely(buffer_heads_over_limit)) { |
| 1953 | if (page_has_private(page) && trylock_page(page)) { |
| 1954 | if (page_has_private(page)) |
| 1955 | try_to_release_page(page, 0); |
| 1956 | unlock_page(page); |
| 1957 | } |
| 1958 | } |
| 1959 | |
Johannes Weiner | c3ac9a8 | 2012-05-29 15:06:25 -0700 | [diff] [blame] | 1960 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
| 1961 | &vm_flags)) { |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1962 | nr_rotated += hpage_nr_pages(page); |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1963 | /* |
| 1964 | * Identify referenced, file-backed active pages and |
| 1965 | * give them one more trip around the active list. So |
| 1966 | * that executable code get better chances to stay in |
| 1967 | * memory under moderate memory pressure. Anon pages |
| 1968 | * are not likely to be evicted by use-once streaming |
| 1969 | * IO, plus JVM can create lots of anon VM_EXEC pages, |
| 1970 | * so we ignore them here. |
| 1971 | */ |
Wu Fengguang | 41e2098 | 2009-10-26 16:49:53 -0700 | [diff] [blame] | 1972 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1973 | list_add(&page->lru, &l_active); |
| 1974 | continue; |
| 1975 | } |
| 1976 | } |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1977 | |
KOSAKI Motohiro | 5205e56 | 2009-09-21 17:01:44 -0700 | [diff] [blame] | 1978 | ClearPageActive(page); /* we are de-activating */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1979 | list_add(&page->lru, &l_inactive); |
| 1980 | } |
| 1981 | |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 1982 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1983 | * Move pages back to the lru list. |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 1984 | */ |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1985 | spin_lock_irq(&pgdat->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1986 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1987 | * Count referenced pages from currently used mappings as rotated, |
| 1988 | * even though only some of them are actually re-activated. This |
| 1989 | * helps balance scan pressure between file and anonymous pages in |
Jerome Marchand | 7c0db9e | 2014-08-06 16:08:01 -0700 | [diff] [blame] | 1990 | * get_scan_count. |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1991 | */ |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 1992 | reclaim_stat->recent_rotated[file] += nr_rotated; |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 1993 | |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 1994 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
| 1995 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1996 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
| 1997 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1998 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1999 | mem_cgroup_uncharge_list(&l_hold); |
Mel Gorman | 2d4894b | 2017-11-15 17:37:59 -0800 | [diff] [blame] | 2000 | free_unref_page_list(&l_hold); |
Michal Hocko | 9d998b4 | 2017-02-22 15:44:18 -0800 | [diff] [blame] | 2001 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
| 2002 | nr_deactivate, nr_rotated, sc->priority, file); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2003 | } |
| 2004 | |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2005 | /* |
| 2006 | * The inactive anon list should be small enough that the VM never has |
| 2007 | * to do too much work. |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2008 | * |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2009 | * The inactive file list should be small enough to leave most memory |
| 2010 | * to the established workingset on the scan-resistant active list, |
| 2011 | * but large enough to avoid thrashing the aggregate readahead window. |
| 2012 | * |
| 2013 | * Both inactive lists should also be large enough that each inactive |
| 2014 | * page has a chance to be referenced again before it is reclaimed. |
| 2015 | * |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2016 | * If that fails and refaulting is observed, the inactive list grows. |
| 2017 | * |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2018 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
Andrey Ryabinin | 3a50d14 | 2017-11-15 17:34:15 -0800 | [diff] [blame] | 2019 | * on this LRU, maintained by the pageout code. An inactive_ratio |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2020 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
| 2021 | * |
| 2022 | * total target max |
| 2023 | * memory ratio inactive |
| 2024 | * ------------------------------------- |
| 2025 | * 10MB 1 5MB |
| 2026 | * 100MB 1 50MB |
| 2027 | * 1GB 3 250MB |
| 2028 | * 10GB 10 0.9GB |
| 2029 | * 100GB 31 3GB |
| 2030 | * 1TB 101 10GB |
| 2031 | * 10TB 320 32GB |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2032 | */ |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2033 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2034 | struct mem_cgroup *memcg, |
| 2035 | struct scan_control *sc, bool actual_reclaim) |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2036 | { |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2037 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2038 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
| 2039 | enum lru_list inactive_lru = file * LRU_FILE; |
| 2040 | unsigned long inactive, active; |
| 2041 | unsigned long inactive_ratio; |
| 2042 | unsigned long refaults; |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2043 | unsigned long gb; |
| 2044 | |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 2045 | /* |
| 2046 | * If we don't have swap space, anonymous page deactivation |
| 2047 | * is pointless. |
| 2048 | */ |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2049 | if (!file && !total_swap_pages) |
Yaowei Bai | 42e2e45 | 2015-11-05 18:47:36 -0800 | [diff] [blame] | 2050 | return false; |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 2051 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2052 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
| 2053 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2054 | |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2055 | if (memcg) |
Johannes Weiner | ccda7f4 | 2017-05-03 14:55:16 -0700 | [diff] [blame] | 2056 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2057 | else |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2058 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2059 | |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2060 | /* |
| 2061 | * When refaults are being observed, it means a new workingset |
| 2062 | * is being established. Disable active list protection to get |
| 2063 | * rid of the stale workingset quickly. |
| 2064 | */ |
| 2065 | if (file && actual_reclaim && lruvec->refaults != refaults) { |
| 2066 | inactive_ratio = 0; |
| 2067 | } else { |
| 2068 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
| 2069 | if (gb) |
| 2070 | inactive_ratio = int_sqrt(10 * gb); |
| 2071 | else |
| 2072 | inactive_ratio = 1; |
| 2073 | } |
| 2074 | |
| 2075 | if (actual_reclaim) |
| 2076 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, |
| 2077 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, |
| 2078 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, |
| 2079 | inactive_ratio, file); |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2080 | |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2081 | return inactive * inactive_ratio < active; |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2082 | } |
| 2083 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2084 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2085 | struct lruvec *lruvec, struct mem_cgroup *memcg, |
| 2086 | struct scan_control *sc) |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2087 | { |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2088 | if (is_active_lru(lru)) { |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2089 | if (inactive_list_is_low(lruvec, is_file_lru(lru), |
| 2090 | memcg, sc, true)) |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2091 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2092 | return 0; |
| 2093 | } |
| 2094 | |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2095 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2096 | } |
| 2097 | |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2098 | enum scan_balance { |
| 2099 | SCAN_EQUAL, |
| 2100 | SCAN_FRACT, |
| 2101 | SCAN_ANON, |
| 2102 | SCAN_FILE, |
| 2103 | }; |
| 2104 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2105 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2106 | * Determine how aggressively the anon and file LRU lists should be |
| 2107 | * scanned. The relative value of each set of LRU lists is determined |
| 2108 | * by looking at the fraction of the pages scanned we did rotate back |
| 2109 | * onto the active list instead of evict. |
| 2110 | * |
Wanpeng Li | be7bd59 | 2012-06-14 20:41:02 +0800 | [diff] [blame] | 2111 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
| 2112 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2113 | */ |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2114 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2115 | struct scan_control *sc, unsigned long *nr, |
| 2116 | unsigned long *lru_pages) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2117 | { |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2118 | int swappiness = mem_cgroup_swappiness(memcg); |
Konstantin Khlebnikov | 9012637 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2119 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2120 | u64 fraction[2]; |
| 2121 | u64 denominator = 0; /* gcc */ |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2122 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2123 | unsigned long anon_prio, file_prio; |
| 2124 | enum scan_balance scan_balance; |
Johannes Weiner | 0bf1457 | 2014-04-08 16:04:10 -0700 | [diff] [blame] | 2125 | unsigned long anon, file; |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2126 | unsigned long ap, fp; |
| 2127 | enum lru_list lru; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2128 | |
| 2129 | /* If we have no swap space, do not bother scanning anon pages. */ |
Vladimir Davydov | d8b3843 | 2016-01-20 15:03:07 -0800 | [diff] [blame] | 2130 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2131 | scan_balance = SCAN_FILE; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2132 | goto out; |
| 2133 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2134 | |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2135 | /* |
| 2136 | * Global reclaim will swap to prevent OOM even with no |
| 2137 | * swappiness, but memcg users want to use this knob to |
| 2138 | * disable swapping for individual groups completely when |
| 2139 | * using the memory controller's swap limit feature would be |
| 2140 | * too expensive. |
| 2141 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2142 | if (!global_reclaim(sc) && !swappiness) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2143 | scan_balance = SCAN_FILE; |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2144 | goto out; |
| 2145 | } |
| 2146 | |
| 2147 | /* |
| 2148 | * Do not apply any pressure balancing cleverness when the |
| 2149 | * system is close to OOM, scan both anon and file equally |
| 2150 | * (unless the swappiness setting disagrees with swapping). |
| 2151 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2152 | if (!sc->priority && swappiness) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2153 | scan_balance = SCAN_EQUAL; |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2154 | goto out; |
| 2155 | } |
| 2156 | |
Johannes Weiner | 11d16c2 | 2013-02-22 16:32:15 -0800 | [diff] [blame] | 2157 | /* |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2158 | * Prevent the reclaimer from falling into the cache trap: as |
| 2159 | * cache pages start out inactive, every cache fault will tip |
| 2160 | * the scan balance towards the file LRU. And as the file LRU |
| 2161 | * shrinks, so does the window for rotation from references. |
| 2162 | * This means we have a runaway feedback loop where a tiny |
| 2163 | * thrashing file LRU becomes infinitely more attractive than |
| 2164 | * anon pages. Try to detect this based on file LRU size. |
| 2165 | */ |
| 2166 | if (global_reclaim(sc)) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2167 | unsigned long pgdatfile; |
| 2168 | unsigned long pgdatfree; |
| 2169 | int z; |
| 2170 | unsigned long total_high_wmark = 0; |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2171 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2172 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
| 2173 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + |
| 2174 | node_page_state(pgdat, NR_INACTIVE_FILE); |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2175 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2176 | for (z = 0; z < MAX_NR_ZONES; z++) { |
| 2177 | struct zone *zone = &pgdat->node_zones[z]; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 2178 | if (!managed_zone(zone)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2179 | continue; |
| 2180 | |
| 2181 | total_high_wmark += high_wmark_pages(zone); |
| 2182 | } |
| 2183 | |
| 2184 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
David Rientjes | 0622622 | 2017-07-10 15:47:20 -0700 | [diff] [blame] | 2185 | /* |
| 2186 | * Force SCAN_ANON if there are enough inactive |
| 2187 | * anonymous pages on the LRU in eligible zones. |
| 2188 | * Otherwise, the small LRU gets thrashed. |
| 2189 | */ |
| 2190 | if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && |
| 2191 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) |
| 2192 | >> sc->priority) { |
| 2193 | scan_balance = SCAN_ANON; |
| 2194 | goto out; |
| 2195 | } |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2196 | } |
| 2197 | } |
| 2198 | |
| 2199 | /* |
Vladimir Davydov | 316bda0 | 2016-01-14 15:19:38 -0800 | [diff] [blame] | 2200 | * If there is enough inactive page cache, i.e. if the size of the |
| 2201 | * inactive list is greater than that of the active list *and* the |
| 2202 | * inactive list actually has some pages to scan on this priority, we |
| 2203 | * do not reclaim anything from the anonymous working set right now. |
| 2204 | * Without the second condition we could end up never scanning an |
| 2205 | * lruvec even if it has plenty of old anonymous pages unless the |
| 2206 | * system is under heavy pressure. |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2207 | */ |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2208 | if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && |
Michal Hocko | 71ab6cf | 2017-02-22 15:46:01 -0800 | [diff] [blame] | 2209 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2210 | scan_balance = SCAN_FILE; |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2211 | goto out; |
| 2212 | } |
| 2213 | |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2214 | scan_balance = SCAN_FRACT; |
| 2215 | |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2216 | /* |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 2217 | * With swappiness at 100, anonymous and file have the same priority. |
| 2218 | * This scanning priority is essentially the inverse of IO cost. |
| 2219 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2220 | anon_prio = swappiness; |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2221 | file_prio = 200 - anon_prio; |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 2222 | |
| 2223 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2224 | * OK, so we have swap space and a fair amount of page cache |
| 2225 | * pages. We use the recently rotated / recently scanned |
| 2226 | * ratios to determine how valuable each cache is. |
| 2227 | * |
| 2228 | * Because workloads change over time (and to avoid overflow) |
| 2229 | * we keep these statistics as a floating average, which ends |
| 2230 | * up weighing recent references more than old ones. |
| 2231 | * |
| 2232 | * anon in [0], file in [1] |
| 2233 | */ |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2234 | |
Michal Hocko | fd53880 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2235 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
| 2236 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); |
| 2237 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + |
| 2238 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2239 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2240 | spin_lock_irq(&pgdat->lru_lock); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2241 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2242 | reclaim_stat->recent_scanned[0] /= 2; |
| 2243 | reclaim_stat->recent_rotated[0] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2244 | } |
| 2245 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2246 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2247 | reclaim_stat->recent_scanned[1] /= 2; |
| 2248 | reclaim_stat->recent_rotated[1] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2249 | } |
| 2250 | |
| 2251 | /* |
Rik van Riel | 00d8089 | 2008-11-19 15:36:44 -0800 | [diff] [blame] | 2252 | * The amount of pressure on anon vs file pages is inversely |
| 2253 | * proportional to the fraction of recently scanned pages on |
| 2254 | * each list that were recently referenced and in active use. |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2255 | */ |
Satoru Moriya | fe35004 | 2012-05-29 15:06:47 -0700 | [diff] [blame] | 2256 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2257 | ap /= reclaim_stat->recent_rotated[0] + 1; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2258 | |
Satoru Moriya | fe35004 | 2012-05-29 15:06:47 -0700 | [diff] [blame] | 2259 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2260 | fp /= reclaim_stat->recent_rotated[1] + 1; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2261 | spin_unlock_irq(&pgdat->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2262 | |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2263 | fraction[0] = ap; |
| 2264 | fraction[1] = fp; |
| 2265 | denominator = ap + fp + 1; |
| 2266 | out: |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2267 | *lru_pages = 0; |
| 2268 | for_each_evictable_lru(lru) { |
| 2269 | int file = is_file_lru(lru); |
| 2270 | unsigned long size; |
| 2271 | unsigned long scan; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2272 | |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2273 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
| 2274 | scan = size >> sc->priority; |
| 2275 | /* |
| 2276 | * If the cgroup's already been deleted, make sure to |
| 2277 | * scrape out the remaining cache. |
| 2278 | */ |
| 2279 | if (!scan && !mem_cgroup_online(memcg)) |
| 2280 | scan = min(size, SWAP_CLUSTER_MAX); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2281 | |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2282 | switch (scan_balance) { |
| 2283 | case SCAN_EQUAL: |
| 2284 | /* Scan lists relative to size */ |
| 2285 | break; |
| 2286 | case SCAN_FRACT: |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2287 | /* |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2288 | * Scan types proportional to swappiness and |
| 2289 | * their relative recent reclaim efficiency. |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2290 | */ |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2291 | scan = div64_u64(scan * fraction[file], |
| 2292 | denominator); |
| 2293 | break; |
| 2294 | case SCAN_FILE: |
| 2295 | case SCAN_ANON: |
| 2296 | /* Scan one type exclusively */ |
| 2297 | if ((scan_balance == SCAN_FILE) != file) { |
| 2298 | size = 0; |
| 2299 | scan = 0; |
| 2300 | } |
| 2301 | break; |
| 2302 | default: |
| 2303 | /* Look ma, no brain */ |
| 2304 | BUG(); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2305 | } |
Johannes Weiner | 688035f | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2306 | |
| 2307 | *lru_pages += size; |
| 2308 | nr[lru] = scan; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2309 | } |
Wu Fengguang | 6e08a36 | 2009-06-16 15:32:29 -0700 | [diff] [blame] | 2310 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2311 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2312 | /* |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2313 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2314 | */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2315 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2316 | struct scan_control *sc, unsigned long *lru_pages) |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2317 | { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2318 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2319 | unsigned long nr[NR_LRU_LISTS]; |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2320 | unsigned long targets[NR_LRU_LISTS]; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2321 | unsigned long nr_to_scan; |
| 2322 | enum lru_list lru; |
| 2323 | unsigned long nr_reclaimed = 0; |
| 2324 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
| 2325 | struct blk_plug plug; |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2326 | bool scan_adjusted; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2327 | |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2328 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2329 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2330 | /* Record the original scan target for proportional adjustments later */ |
| 2331 | memcpy(targets, nr, sizeof(nr)); |
| 2332 | |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2333 | /* |
| 2334 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal |
| 2335 | * event that can occur when there is little memory pressure e.g. |
| 2336 | * multiple streaming readers/writers. Hence, we do not abort scanning |
| 2337 | * when the requested number of pages are reclaimed when scanning at |
| 2338 | * DEF_PRIORITY on the assumption that the fact we are direct |
| 2339 | * reclaiming implies that kswapd is not keeping up and it is best to |
| 2340 | * do a batch of work at once. For memcg reclaim one check is made to |
| 2341 | * abort proportional reclaim if either the file or anon lru has already |
| 2342 | * dropped to zero at the first pass. |
| 2343 | */ |
| 2344 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && |
| 2345 | sc->priority == DEF_PRIORITY); |
| 2346 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2347 | blk_start_plug(&plug); |
| 2348 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
| 2349 | nr[LRU_INACTIVE_FILE]) { |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2350 | unsigned long nr_anon, nr_file, percentage; |
| 2351 | unsigned long nr_scanned; |
| 2352 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2353 | for_each_evictable_lru(lru) { |
| 2354 | if (nr[lru]) { |
| 2355 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); |
| 2356 | nr[lru] -= nr_to_scan; |
| 2357 | |
| 2358 | nr_reclaimed += shrink_list(lru, nr_to_scan, |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2359 | lruvec, memcg, sc); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2360 | } |
| 2361 | } |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2362 | |
Michal Hocko | bd04173 | 2016-12-02 17:26:48 -0800 | [diff] [blame] | 2363 | cond_resched(); |
| 2364 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2365 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
| 2366 | continue; |
| 2367 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2368 | /* |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2369 | * For kswapd and memcg, reclaim at least the number of pages |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2370 | * requested. Ensure that the anon and file LRUs are scanned |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2371 | * proportionally what was requested by get_scan_count(). We |
| 2372 | * stop reclaiming one LRU and reduce the amount scanning |
| 2373 | * proportional to the original scan target. |
| 2374 | */ |
| 2375 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; |
| 2376 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; |
| 2377 | |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2378 | /* |
| 2379 | * It's just vindictive to attack the larger once the smaller |
| 2380 | * has gone to zero. And given the way we stop scanning the |
| 2381 | * smaller below, this makes sure that we only make one nudge |
| 2382 | * towards proportionality once we've got nr_to_reclaim. |
| 2383 | */ |
| 2384 | if (!nr_file || !nr_anon) |
| 2385 | break; |
| 2386 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2387 | if (nr_file > nr_anon) { |
| 2388 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + |
| 2389 | targets[LRU_ACTIVE_ANON] + 1; |
| 2390 | lru = LRU_BASE; |
| 2391 | percentage = nr_anon * 100 / scan_target; |
| 2392 | } else { |
| 2393 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + |
| 2394 | targets[LRU_ACTIVE_FILE] + 1; |
| 2395 | lru = LRU_FILE; |
| 2396 | percentage = nr_file * 100 / scan_target; |
| 2397 | } |
| 2398 | |
| 2399 | /* Stop scanning the smaller of the LRU */ |
| 2400 | nr[lru] = 0; |
| 2401 | nr[lru + LRU_ACTIVE] = 0; |
| 2402 | |
| 2403 | /* |
| 2404 | * Recalculate the other LRU scan count based on its original |
| 2405 | * scan target and the percentage scanning already complete |
| 2406 | */ |
| 2407 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; |
| 2408 | nr_scanned = targets[lru] - nr[lru]; |
| 2409 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
| 2410 | nr[lru] -= min(nr[lru], nr_scanned); |
| 2411 | |
| 2412 | lru += LRU_ACTIVE; |
| 2413 | nr_scanned = targets[lru] - nr[lru]; |
| 2414 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
| 2415 | nr[lru] -= min(nr[lru], nr_scanned); |
| 2416 | |
| 2417 | scan_adjusted = true; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2418 | } |
| 2419 | blk_finish_plug(&plug); |
| 2420 | sc->nr_reclaimed += nr_reclaimed; |
| 2421 | |
| 2422 | /* |
| 2423 | * Even if we did not try to evict anon pages at all, we want to |
| 2424 | * rebalance the anon lru active/inactive ratio. |
| 2425 | */ |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2426 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2427 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
| 2428 | sc, LRU_ACTIVE_ANON); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2429 | } |
| 2430 | |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2431 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2432 | static bool in_reclaim_compaction(struct scan_control *sc) |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2433 | { |
Kirill A. Shutemov | d84da3f | 2012-12-11 16:00:31 -0800 | [diff] [blame] | 2434 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2435 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2436 | sc->priority < DEF_PRIORITY - 2)) |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2437 | return true; |
| 2438 | |
| 2439 | return false; |
| 2440 | } |
| 2441 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2442 | /* |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2443 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
| 2444 | * order-0 pages before compacting the zone. should_continue_reclaim() returns |
| 2445 | * true if more pages should be reclaimed such that when the page allocator |
| 2446 | * calls try_to_compact_zone() that it will have enough free pages to succeed. |
| 2447 | * It will give up earlier than that if there is difficulty reclaiming pages. |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2448 | */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2449 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2450 | unsigned long nr_reclaimed, |
| 2451 | unsigned long nr_scanned, |
| 2452 | struct scan_control *sc) |
| 2453 | { |
| 2454 | unsigned long pages_for_compaction; |
| 2455 | unsigned long inactive_lru_pages; |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2456 | int z; |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2457 | |
| 2458 | /* If not in reclaim/compaction mode, stop */ |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2459 | if (!in_reclaim_compaction(sc)) |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2460 | return false; |
| 2461 | |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2462 | /* Consider stopping depending on scan and reclaim activity */ |
Michal Hocko | dcda9b0 | 2017-07-12 14:36:45 -0700 | [diff] [blame] | 2463 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2464 | /* |
Michal Hocko | dcda9b0 | 2017-07-12 14:36:45 -0700 | [diff] [blame] | 2465 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2466 | * full LRU list has been scanned and we are still failing |
| 2467 | * to reclaim pages. This full LRU scan is potentially |
Michal Hocko | dcda9b0 | 2017-07-12 14:36:45 -0700 | [diff] [blame] | 2468 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2469 | */ |
| 2470 | if (!nr_reclaimed && !nr_scanned) |
| 2471 | return false; |
| 2472 | } else { |
| 2473 | /* |
Michal Hocko | dcda9b0 | 2017-07-12 14:36:45 -0700 | [diff] [blame] | 2474 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2475 | * fail without consequence, stop if we failed to reclaim |
| 2476 | * any pages from the last SWAP_CLUSTER_MAX number of |
| 2477 | * pages that were scanned. This will return to the |
| 2478 | * caller faster at the risk reclaim/compaction and |
| 2479 | * the resulting allocation attempt fails |
| 2480 | */ |
| 2481 | if (!nr_reclaimed) |
| 2482 | return false; |
| 2483 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2484 | |
| 2485 | /* |
| 2486 | * If we have not reclaimed enough pages for compaction and the |
| 2487 | * inactive lists are large enough, continue reclaiming |
| 2488 | */ |
Vlastimil Babka | 9861a62 | 2016-10-07 16:57:53 -0700 | [diff] [blame] | 2489 | pages_for_compaction = compact_gap(sc->order); |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2490 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
Shaohua Li | ec8acf2 | 2013-02-22 16:34:38 -0800 | [diff] [blame] | 2491 | if (get_nr_swap_pages() > 0) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2492 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2493 | if (sc->nr_reclaimed < pages_for_compaction && |
| 2494 | inactive_lru_pages > pages_for_compaction) |
| 2495 | return true; |
| 2496 | |
| 2497 | /* If compaction would go ahead or the allocation would succeed, stop */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2498 | for (z = 0; z <= sc->reclaim_idx; z++) { |
| 2499 | struct zone *zone = &pgdat->node_zones[z]; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 2500 | if (!managed_zone(zone)) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2501 | continue; |
| 2502 | |
| 2503 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { |
Vlastimil Babka | cf37831 | 2016-10-07 16:57:41 -0700 | [diff] [blame] | 2504 | case COMPACT_SUCCESS: |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2505 | case COMPACT_CONTINUE: |
| 2506 | return false; |
| 2507 | default: |
| 2508 | /* check next zone */ |
| 2509 | ; |
| 2510 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2511 | } |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2512 | return true; |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2513 | } |
| 2514 | |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 2515 | static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg) |
| 2516 | { |
| 2517 | return test_bit(PGDAT_CONGESTED, &pgdat->flags) || |
| 2518 | (memcg && memcg_congested(pgdat, memcg)); |
| 2519 | } |
| 2520 | |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 2521 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2522 | { |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2523 | struct reclaim_state *reclaim_state = current->reclaim_state; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2524 | unsigned long nr_reclaimed, nr_scanned; |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2525 | bool reclaimable = false; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2526 | |
Johannes Weiner | 5660048 | 2012-01-12 17:17:59 -0800 | [diff] [blame] | 2527 | do { |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2528 | struct mem_cgroup *root = sc->target_mem_cgroup; |
| 2529 | struct mem_cgroup_reclaim_cookie reclaim = { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2530 | .pgdat = pgdat, |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2531 | .priority = sc->priority, |
| 2532 | }; |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2533 | unsigned long node_lru_pages = 0; |
Andrew Morton | 694fbc0 | 2013-09-24 15:27:37 -0700 | [diff] [blame] | 2534 | struct mem_cgroup *memcg; |
Johannes Weiner | 5660048 | 2012-01-12 17:17:59 -0800 | [diff] [blame] | 2535 | |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 2536 | memset(&sc->nr, 0, sizeof(sc->nr)); |
| 2537 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2538 | nr_reclaimed = sc->nr_reclaimed; |
| 2539 | nr_scanned = sc->nr_scanned; |
Konstantin Khlebnikov | f9be23d | 2012-05-29 15:07:02 -0700 | [diff] [blame] | 2540 | |
Andrew Morton | 694fbc0 | 2013-09-24 15:27:37 -0700 | [diff] [blame] | 2541 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
| 2542 | do { |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2543 | unsigned long lru_pages; |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2544 | unsigned long reclaimed; |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2545 | unsigned long scanned; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2546 | |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2547 | if (mem_cgroup_low(root, memcg)) { |
Yisheng Xie | d6622f6 | 2017-05-03 14:53:57 -0700 | [diff] [blame] | 2548 | if (!sc->memcg_low_reclaim) { |
| 2549 | sc->memcg_low_skipped = 1; |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2550 | continue; |
Yisheng Xie | d6622f6 | 2017-05-03 14:53:57 -0700 | [diff] [blame] | 2551 | } |
Johannes Weiner | e27be24 | 2018-04-10 16:29:45 -0700 | [diff] [blame] | 2552 | memcg_memory_event(memcg, MEMCG_LOW); |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2553 | } |
| 2554 | |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2555 | reclaimed = sc->nr_reclaimed; |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2556 | scanned = sc->nr_scanned; |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2557 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
| 2558 | node_lru_pages += lru_pages; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2559 | |
Vladimir Davydov | b5afba2 | 2016-08-02 14:03:04 -0700 | [diff] [blame] | 2560 | if (memcg) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2561 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 2562 | memcg, sc->priority); |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2563 | |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2564 | /* Record the group's reclaim efficiency */ |
| 2565 | vmpressure(sc->gfp_mask, memcg, false, |
| 2566 | sc->nr_scanned - scanned, |
| 2567 | sc->nr_reclaimed - reclaimed); |
| 2568 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2569 | /* |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2570 | * Direct reclaim and kswapd have to scan all memory |
| 2571 | * cgroups to fulfill the overall scan target for the |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2572 | * node. |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2573 | * |
| 2574 | * Limit reclaim, on the other hand, only cares about |
| 2575 | * nr_to_reclaim pages to be reclaimed and it will |
| 2576 | * retry with decreasing priority if one round over the |
| 2577 | * whole hierarchy is not sufficient. |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2578 | */ |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2579 | if (!global_reclaim(sc) && |
| 2580 | sc->nr_reclaimed >= sc->nr_to_reclaim) { |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2581 | mem_cgroup_iter_break(root, memcg); |
| 2582 | break; |
| 2583 | } |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2584 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 2585 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2586 | if (global_reclaim(sc)) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2587 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
Josef Bacik | 9092c71 | 2018-01-31 16:16:26 -0800 | [diff] [blame] | 2588 | sc->priority); |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2589 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2590 | if (reclaim_state) { |
| 2591 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
| 2592 | reclaim_state->reclaimed_slab = 0; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2593 | } |
| 2594 | |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2595 | /* Record the subtree's reclaim efficiency */ |
| 2596 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 2597 | sc->nr_scanned - nr_scanned, |
| 2598 | sc->nr_reclaimed - nr_reclaimed); |
| 2599 | |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2600 | if (sc->nr_reclaimed - nr_reclaimed) |
| 2601 | reclaimable = true; |
| 2602 | |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 2603 | if (current_is_kswapd()) { |
| 2604 | /* |
| 2605 | * If reclaim is isolating dirty pages under writeback, |
| 2606 | * it implies that the long-lived page allocation rate |
| 2607 | * is exceeding the page laundering rate. Either the |
| 2608 | * global limits are not being effective at throttling |
| 2609 | * processes due to the page distribution throughout |
| 2610 | * zones or there is heavy usage of a slow backing |
| 2611 | * device. The only option is to throttle from reclaim |
| 2612 | * context which is not ideal as there is no guarantee |
| 2613 | * the dirtying process is throttled in the same way |
| 2614 | * balance_dirty_pages() manages. |
| 2615 | * |
| 2616 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will |
| 2617 | * count the number of pages under pages flagged for |
| 2618 | * immediate reclaim and stall if any are encountered |
| 2619 | * in the nr_immediate check below. |
| 2620 | */ |
| 2621 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) |
| 2622 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 2623 | |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 2624 | /* |
| 2625 | * Tag a node as congested if all the dirty pages |
| 2626 | * scanned were backed by a congested BDI and |
| 2627 | * wait_iff_congested will stall. |
| 2628 | */ |
| 2629 | if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
| 2630 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
| 2631 | |
| 2632 | /* Allow kswapd to start writing pages during reclaim.*/ |
| 2633 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) |
| 2634 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
| 2635 | |
| 2636 | /* |
| 2637 | * If kswapd scans pages marked marked for immediate |
| 2638 | * reclaim and under writeback (nr_immediate), it |
| 2639 | * implies that pages are cycling through the LRU |
| 2640 | * faster than they are written so also forcibly stall. |
| 2641 | */ |
| 2642 | if (sc->nr.immediate) |
| 2643 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
| 2644 | } |
| 2645 | |
| 2646 | /* |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 2647 | * Legacy memcg will stall in page writeback so avoid forcibly |
| 2648 | * stalling in wait_iff_congested(). |
| 2649 | */ |
| 2650 | if (!global_reclaim(sc) && sane_reclaim(sc) && |
| 2651 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
| 2652 | set_memcg_congestion(pgdat, root, true); |
| 2653 | |
| 2654 | /* |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 2655 | * Stall direct reclaim for IO completions if underlying BDIs |
| 2656 | * and node is congested. Allow kswapd to continue until it |
| 2657 | * starts encountering unqueued dirty pages or cycling through |
| 2658 | * the LRU too quickly. |
| 2659 | */ |
| 2660 | if (!sc->hibernation_mode && !current_is_kswapd() && |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 2661 | current_may_throttle() && pgdat_memcg_congested(pgdat, root)) |
| 2662 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); |
Andrey Ryabinin | d108c77 | 2018-04-10 16:27:59 -0700 | [diff] [blame] | 2663 | |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2664 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2665 | sc->nr_scanned - nr_scanned, sc)); |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2666 | |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2667 | /* |
| 2668 | * Kswapd gives up on balancing particular nodes after too |
| 2669 | * many failures to reclaim anything from them and goes to |
| 2670 | * sleep. On reclaim progress, reset the failure counter. A |
| 2671 | * successful direct reclaim run will revive a dormant kswapd. |
| 2672 | */ |
| 2673 | if (reclaimable) |
| 2674 | pgdat->kswapd_failures = 0; |
| 2675 | |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2676 | return reclaimable; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2677 | } |
| 2678 | |
Vlastimil Babka | 53853e2 | 2014-10-09 15:27:02 -0700 | [diff] [blame] | 2679 | /* |
Vlastimil Babka | fdd4c614 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2680 | * Returns true if compaction should go ahead for a costly-order request, or |
| 2681 | * the allocation would already succeed without compaction. Return false if we |
| 2682 | * should reclaim first. |
Vlastimil Babka | 53853e2 | 2014-10-09 15:27:02 -0700 | [diff] [blame] | 2683 | */ |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2684 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2685 | { |
Mel Gorman | 31483b6 | 2016-07-28 15:45:46 -0700 | [diff] [blame] | 2686 | unsigned long watermark; |
Vlastimil Babka | fdd4c614 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2687 | enum compact_result suitable; |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2688 | |
Vlastimil Babka | fdd4c614 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2689 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
| 2690 | if (suitable == COMPACT_SUCCESS) |
| 2691 | /* Allocation should succeed already. Don't reclaim. */ |
| 2692 | return true; |
| 2693 | if (suitable == COMPACT_SKIPPED) |
| 2694 | /* Compaction cannot yet proceed. Do reclaim. */ |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2695 | return false; |
| 2696 | |
Vlastimil Babka | fdd4c614 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2697 | /* |
| 2698 | * Compaction is already possible, but it takes time to run and there |
| 2699 | * are potentially other callers using the pages just freed. So proceed |
| 2700 | * with reclaim to make a buffer of free pages available to give |
| 2701 | * compaction a reasonable chance of completing and allocating the page. |
| 2702 | * Note that we won't actually reclaim the whole buffer in one attempt |
| 2703 | * as the target watermark in should_continue_reclaim() is lower. But if |
| 2704 | * we are already above the high+gap watermark, don't reclaim at all. |
| 2705 | */ |
| 2706 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
| 2707 | |
| 2708 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2709 | } |
| 2710 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2711 | /* |
| 2712 | * This is the direct reclaim path, for page-allocating processes. We only |
| 2713 | * try to reclaim pages from zones which will satisfy the caller's allocation |
| 2714 | * request. |
| 2715 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2716 | * If a zone is deemed to be full of pinned pages then just give it a light |
| 2717 | * scan then give up on it. |
| 2718 | */ |
Michal Hocko | 0a0337e | 2016-05-20 16:57:00 -0700 | [diff] [blame] | 2719 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2720 | { |
Mel Gorman | dd1a239 | 2008-04-28 02:12:17 -0700 | [diff] [blame] | 2721 | struct zoneref *z; |
Mel Gorman | 54a6eb5 | 2008-04-28 02:12:16 -0700 | [diff] [blame] | 2722 | struct zone *zone; |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2723 | unsigned long nr_soft_reclaimed; |
| 2724 | unsigned long nr_soft_scanned; |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2725 | gfp_t orig_mask; |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2726 | pg_data_t *last_pgdat = NULL; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2727 | |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2728 | /* |
| 2729 | * If the number of buffer_heads in the machine exceeds the maximum |
| 2730 | * allowed level, force direct reclaim to scan the highmem zone as |
| 2731 | * highmem pages could be pinning lowmem pages storing buffer_heads |
| 2732 | */ |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2733 | orig_mask = sc->gfp_mask; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2734 | if (buffer_heads_over_limit) { |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2735 | sc->gfp_mask |= __GFP_HIGHMEM; |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2736 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2737 | } |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2738 | |
Mel Gorman | d4debc6 | 2010-08-09 17:19:29 -0700 | [diff] [blame] | 2739 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2740 | sc->reclaim_idx, sc->nodemask) { |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2741 | /* |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2742 | * Take care memory controller reclaiming has small influence |
| 2743 | * to global LRU. |
| 2744 | */ |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 2745 | if (global_reclaim(sc)) { |
Vladimir Davydov | 344736f | 2014-10-20 15:50:30 +0400 | [diff] [blame] | 2746 | if (!cpuset_zone_allowed(zone, |
| 2747 | GFP_KERNEL | __GFP_HARDWALL)) |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2748 | continue; |
Vladimir Davydov | 65ec02c | 2014-04-03 14:47:20 -0700 | [diff] [blame] | 2749 | |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2750 | /* |
| 2751 | * If we already have plenty of memory free for |
| 2752 | * compaction in this zone, don't free any more. |
| 2753 | * Even though compaction is invoked for any |
| 2754 | * non-zero order, only frequent costly order |
| 2755 | * reclamation is disruptive enough to become a |
| 2756 | * noticeable problem, like transparent huge |
| 2757 | * page allocations. |
| 2758 | */ |
| 2759 | if (IS_ENABLED(CONFIG_COMPACTION) && |
| 2760 | sc->order > PAGE_ALLOC_COSTLY_ORDER && |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2761 | compaction_ready(zone, sc)) { |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2762 | sc->compaction_ready = true; |
| 2763 | continue; |
Rik van Riel | e0887c1 | 2011-10-31 17:09:31 -0700 | [diff] [blame] | 2764 | } |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2765 | |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2766 | /* |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2767 | * Shrink each node in the zonelist once. If the |
| 2768 | * zonelist is ordered by zone (not the default) then a |
| 2769 | * node may be shrunk multiple times but in that case |
| 2770 | * the user prefers lower zones being preserved. |
| 2771 | */ |
| 2772 | if (zone->zone_pgdat == last_pgdat) |
| 2773 | continue; |
| 2774 | |
| 2775 | /* |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2776 | * This steals pages from memory cgroups over softlimit |
| 2777 | * and returns the number of reclaimed pages and |
| 2778 | * scanned pages. This works for global memory pressure |
| 2779 | * and balancing, not for a memcg's limit. |
| 2780 | */ |
| 2781 | nr_soft_scanned = 0; |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2782 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2783 | sc->order, sc->gfp_mask, |
| 2784 | &nr_soft_scanned); |
| 2785 | sc->nr_reclaimed += nr_soft_reclaimed; |
| 2786 | sc->nr_scanned += nr_soft_scanned; |
KAMEZAWA Hiroyuki | ac34a1a | 2011-06-27 16:18:12 -0700 | [diff] [blame] | 2787 | /* need some check for avoid more shrink_zone() */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2788 | } |
Nick Piggin | 408d854 | 2006-09-25 23:31:27 -0700 | [diff] [blame] | 2789 | |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2790 | /* See comment about same check for global reclaim above */ |
| 2791 | if (zone->zone_pgdat == last_pgdat) |
| 2792 | continue; |
| 2793 | last_pgdat = zone->zone_pgdat; |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 2794 | shrink_node(zone->zone_pgdat, sc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2795 | } |
Mel Gorman | e0c2327 | 2011-10-31 17:09:33 -0700 | [diff] [blame] | 2796 | |
Vladimir Davydov | 65ec02c | 2014-04-03 14:47:20 -0700 | [diff] [blame] | 2797 | /* |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2798 | * Restore to original mask to avoid the impact on the caller if we |
| 2799 | * promoted it to __GFP_HIGHMEM. |
| 2800 | */ |
| 2801 | sc->gfp_mask = orig_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2802 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2803 | |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2804 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
| 2805 | { |
| 2806 | struct mem_cgroup *memcg; |
| 2807 | |
| 2808 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); |
| 2809 | do { |
| 2810 | unsigned long refaults; |
| 2811 | struct lruvec *lruvec; |
| 2812 | |
| 2813 | if (memcg) |
Johannes Weiner | ccda7f4 | 2017-05-03 14:55:16 -0700 | [diff] [blame] | 2814 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2815 | else |
| 2816 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); |
| 2817 | |
| 2818 | lruvec = mem_cgroup_lruvec(pgdat, memcg); |
| 2819 | lruvec->refaults = refaults; |
| 2820 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); |
| 2821 | } |
| 2822 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2823 | /* |
| 2824 | * This is the main entry point to direct page reclaim. |
| 2825 | * |
| 2826 | * If a full scan of the inactive list fails to free enough memory then we |
| 2827 | * are "out of memory" and something needs to be killed. |
| 2828 | * |
| 2829 | * If the caller is !__GFP_FS then the probability of a failure is reasonably |
| 2830 | * high - the zone may be full of dirty or under-writeback pages, which this |
Jens Axboe | 5b0830c | 2009-09-23 19:37:09 +0200 | [diff] [blame] | 2831 | * caller can't do much about. We kick the writeback threads and take explicit |
| 2832 | * naps in the hope that some of these pages can be written. But if the |
| 2833 | * allocating task holds filesystem locks which prevent writeout this might not |
| 2834 | * work, and the allocation attempt will fail. |
Nishanth Aravamudan | a41f24e | 2008-04-29 00:58:25 -0700 | [diff] [blame] | 2835 | * |
| 2836 | * returns: 0, if no pages reclaimed |
| 2837 | * else, the number of pages reclaimed |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2838 | */ |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 2839 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 2840 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2841 | { |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2842 | int initial_priority = sc->priority; |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2843 | pg_data_t *last_pgdat; |
| 2844 | struct zoneref *z; |
| 2845 | struct zone *zone; |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2846 | retry: |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2847 | delayacct_freepages_start(); |
| 2848 | |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 2849 | if (global_reclaim(sc)) |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 2850 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2851 | |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2852 | do { |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 2853 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
| 2854 | sc->priority); |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2855 | sc->nr_scanned = 0; |
Michal Hocko | 0a0337e | 2016-05-20 16:57:00 -0700 | [diff] [blame] | 2856 | shrink_zones(zonelist, sc); |
Mel Gorman | e0c2327 | 2011-10-31 17:09:33 -0700 | [diff] [blame] | 2857 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2858 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2859 | break; |
| 2860 | |
| 2861 | if (sc->compaction_ready) |
| 2862 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2863 | |
| 2864 | /* |
Minchan Kim | 0e50ce3 | 2013-02-22 16:35:37 -0800 | [diff] [blame] | 2865 | * If we're getting trouble reclaiming, start doing |
| 2866 | * writepage even in laptop mode. |
| 2867 | */ |
| 2868 | if (sc->priority < DEF_PRIORITY - 2) |
| 2869 | sc->may_writepage = 1; |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2870 | } while (--sc->priority >= 0); |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2871 | |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2872 | last_pgdat = NULL; |
| 2873 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, |
| 2874 | sc->nodemask) { |
| 2875 | if (zone->zone_pgdat == last_pgdat) |
| 2876 | continue; |
| 2877 | last_pgdat = zone->zone_pgdat; |
| 2878 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
Andrey Ryabinin | e3c1ac5 | 2018-04-10 16:28:03 -0700 | [diff] [blame] | 2879 | set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false); |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 2880 | } |
| 2881 | |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2882 | delayacct_freepages_end(); |
| 2883 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2884 | if (sc->nr_reclaimed) |
| 2885 | return sc->nr_reclaimed; |
| 2886 | |
Mel Gorman | 0cee34f | 2012-01-12 17:19:49 -0800 | [diff] [blame] | 2887 | /* Aborted reclaim to try compaction? don't OOM, then */ |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2888 | if (sc->compaction_ready) |
Mel Gorman | 7335084 | 2012-01-12 17:19:33 -0800 | [diff] [blame] | 2889 | return 1; |
| 2890 | |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2891 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
Yisheng Xie | d6622f6 | 2017-05-03 14:53:57 -0700 | [diff] [blame] | 2892 | if (sc->memcg_low_skipped) { |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2893 | sc->priority = initial_priority; |
Yisheng Xie | d6622f6 | 2017-05-03 14:53:57 -0700 | [diff] [blame] | 2894 | sc->memcg_low_reclaim = 1; |
| 2895 | sc->memcg_low_skipped = 0; |
Johannes Weiner | 241994ed | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2896 | goto retry; |
| 2897 | } |
| 2898 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2899 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2900 | } |
| 2901 | |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2902 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2903 | { |
| 2904 | struct zone *zone; |
| 2905 | unsigned long pfmemalloc_reserve = 0; |
| 2906 | unsigned long free_pages = 0; |
| 2907 | int i; |
| 2908 | bool wmark_ok; |
| 2909 | |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2910 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
| 2911 | return true; |
| 2912 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2913 | for (i = 0; i <= ZONE_NORMAL; i++) { |
| 2914 | zone = &pgdat->node_zones[i]; |
Johannes Weiner | d450abd | 2017-05-03 14:51:54 -0700 | [diff] [blame] | 2915 | if (!managed_zone(zone)) |
| 2916 | continue; |
| 2917 | |
| 2918 | if (!zone_reclaimable_pages(zone)) |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2919 | continue; |
| 2920 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2921 | pfmemalloc_reserve += min_wmark_pages(zone); |
| 2922 | free_pages += zone_page_state(zone, NR_FREE_PAGES); |
| 2923 | } |
| 2924 | |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2925 | /* If there are no reserves (unexpected config) then do not throttle */ |
| 2926 | if (!pfmemalloc_reserve) |
| 2927 | return true; |
| 2928 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2929 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
| 2930 | |
| 2931 | /* kswapd must be awake if processes are being throttled */ |
| 2932 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 2933 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2934 | (enum zone_type)ZONE_NORMAL); |
| 2935 | wake_up_interruptible(&pgdat->kswapd_wait); |
| 2936 | } |
| 2937 | |
| 2938 | return wmark_ok; |
| 2939 | } |
| 2940 | |
| 2941 | /* |
| 2942 | * Throttle direct reclaimers if backing storage is backed by the network |
| 2943 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously |
| 2944 | * depleted. kswapd will continue to make progress and wake the processes |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2945 | * when the low watermark is reached. |
| 2946 | * |
| 2947 | * Returns true if a fatal signal was delivered during throttling. If this |
| 2948 | * happens, the page allocator should not consider triggering the OOM killer. |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2949 | */ |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2950 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2951 | nodemask_t *nodemask) |
| 2952 | { |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2953 | struct zoneref *z; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2954 | struct zone *zone; |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2955 | pg_data_t *pgdat = NULL; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2956 | |
| 2957 | /* |
| 2958 | * Kernel threads should not be throttled as they may be indirectly |
| 2959 | * responsible for cleaning pages necessary for reclaim to make forward |
| 2960 | * progress. kjournald for example may enter direct reclaim while |
| 2961 | * committing a transaction where throttling it could forcing other |
| 2962 | * processes to block on log_wait_commit(). |
| 2963 | */ |
| 2964 | if (current->flags & PF_KTHREAD) |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2965 | goto out; |
| 2966 | |
| 2967 | /* |
| 2968 | * If a fatal signal is pending, this process should not throttle. |
| 2969 | * It should return quickly so it can exit and free its memory |
| 2970 | */ |
| 2971 | if (fatal_signal_pending(current)) |
| 2972 | goto out; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2973 | |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2974 | /* |
| 2975 | * Check if the pfmemalloc reserves are ok by finding the first node |
| 2976 | * with a usable ZONE_NORMAL or lower zone. The expectation is that |
| 2977 | * GFP_KERNEL will be required for allocating network buffers when |
| 2978 | * swapping over the network so ZONE_HIGHMEM is unusable. |
| 2979 | * |
| 2980 | * Throttling is based on the first usable node and throttled processes |
| 2981 | * wait on a queue until kswapd makes progress and wakes them. There |
| 2982 | * is an affinity then between processes waking up and where reclaim |
| 2983 | * progress has been made assuming the process wakes on the same node. |
| 2984 | * More importantly, processes running on remote nodes will not compete |
| 2985 | * for remote pfmemalloc reserves and processes on different nodes |
| 2986 | * should make reasonable progress. |
| 2987 | */ |
| 2988 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
Michael S. Tsirkin | 17636fa | 2015-01-26 12:58:41 -0800 | [diff] [blame] | 2989 | gfp_zone(gfp_mask), nodemask) { |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2990 | if (zone_idx(zone) > ZONE_NORMAL) |
| 2991 | continue; |
| 2992 | |
| 2993 | /* Throttle based on the first usable node */ |
| 2994 | pgdat = zone->zone_pgdat; |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2995 | if (allow_direct_reclaim(pgdat)) |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2996 | goto out; |
| 2997 | break; |
| 2998 | } |
| 2999 | |
| 3000 | /* If no zone was usable by the allocation flags then do not throttle */ |
| 3001 | if (!pgdat) |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3002 | goto out; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3003 | |
Mel Gorman | 68243e7 | 2012-07-31 16:44:39 -0700 | [diff] [blame] | 3004 | /* Account for the throttling */ |
| 3005 | count_vm_event(PGSCAN_DIRECT_THROTTLE); |
| 3006 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3007 | /* |
| 3008 | * If the caller cannot enter the filesystem, it's possible that it |
| 3009 | * is due to the caller holding an FS lock or performing a journal |
| 3010 | * transaction in the case of a filesystem like ext[3|4]. In this case, |
| 3011 | * it is not safe to block on pfmemalloc_wait as kswapd could be |
| 3012 | * blocked waiting on the same lock. Instead, throttle for up to a |
| 3013 | * second before continuing. |
| 3014 | */ |
| 3015 | if (!(gfp_mask & __GFP_FS)) { |
| 3016 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3017 | allow_direct_reclaim(pgdat), HZ); |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3018 | |
| 3019 | goto check_pending; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3020 | } |
| 3021 | |
| 3022 | /* Throttle until kswapd wakes the process */ |
| 3023 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3024 | allow_direct_reclaim(pgdat)); |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3025 | |
| 3026 | check_pending: |
| 3027 | if (fatal_signal_pending(current)) |
| 3028 | return true; |
| 3029 | |
| 3030 | out: |
| 3031 | return false; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3032 | } |
| 3033 | |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 3034 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 3035 | gfp_t gfp_mask, nodemask_t *nodemask) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3036 | { |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3037 | unsigned long nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3038 | struct scan_control sc = { |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 3039 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Nick Desaulniers | f2f43e5 | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3040 | .gfp_mask = current_gfp_context(gfp_mask), |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3041 | .reclaim_idx = gfp_zone(gfp_mask), |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3042 | .order = order, |
| 3043 | .nodemask = nodemask, |
| 3044 | .priority = DEF_PRIORITY, |
| 3045 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3046 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 3047 | .may_swap = 1, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3048 | }; |
| 3049 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3050 | /* |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3051 | * Do not enter reclaim if fatal signal was delivered while throttled. |
| 3052 | * 1 is returned so that the page allocator does not OOM kill at this |
| 3053 | * point. |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3054 | */ |
Nick Desaulniers | f2f43e5 | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3055 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3056 | return 1; |
| 3057 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3058 | trace_mm_vmscan_direct_reclaim_begin(order, |
| 3059 | sc.may_writepage, |
Nick Desaulniers | f2f43e5 | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3060 | sc.gfp_mask, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3061 | sc.reclaim_idx); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3062 | |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3063 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3064 | |
| 3065 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); |
| 3066 | |
| 3067 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3068 | } |
| 3069 | |
Andrew Morton | c255a45 | 2012-07-31 16:43:02 -0700 | [diff] [blame] | 3070 | #ifdef CONFIG_MEMCG |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3071 | |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 3072 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3073 | gfp_t gfp_mask, bool noswap, |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3074 | pg_data_t *pgdat, |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3075 | unsigned long *nr_scanned) |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3076 | { |
| 3077 | struct scan_control sc = { |
KOSAKI Motohiro | b8f5c56 | 2010-08-10 18:03:02 -0700 | [diff] [blame] | 3078 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3079 | .target_mem_cgroup = memcg, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3080 | .may_writepage = !laptop_mode, |
| 3081 | .may_unmap = 1, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3082 | .reclaim_idx = MAX_NR_ZONES - 1, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3083 | .may_swap = !noswap, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3084 | }; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 3085 | unsigned long lru_pages; |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3086 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3087 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
| 3088 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3089 | |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3090 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3091 | sc.may_writepage, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3092 | sc.gfp_mask, |
| 3093 | sc.reclaim_idx); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3094 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3095 | /* |
| 3096 | * NOTE: Although we can get the priority field, using it |
| 3097 | * here is not a good idea, since it limits the pages we can scan. |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 3098 | * if we don't reclaim here, the shrink_node from balance_pgdat |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3099 | * will pick up pages from other mem cgroup's as well. We hack |
| 3100 | * the priority and make it zero. |
| 3101 | */ |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3102 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3103 | |
| 3104 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); |
| 3105 | |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3106 | *nr_scanned = sc.nr_scanned; |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3107 | return sc.nr_reclaimed; |
| 3108 | } |
| 3109 | |
Johannes Weiner | 72835c8 | 2012-01-12 17:18:32 -0800 | [diff] [blame] | 3110 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3111 | unsigned long nr_pages, |
KOSAKI Motohiro | a7885eb | 2009-01-07 18:08:24 -0800 | [diff] [blame] | 3112 | gfp_t gfp_mask, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3113 | bool may_swap) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3114 | { |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3115 | struct zonelist *zonelist; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3116 | unsigned long nr_reclaimed; |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3117 | int nid; |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3118 | unsigned int noreclaim_flag; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3119 | struct scan_control sc = { |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3120 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
Michal Hocko | 7dea19f | 2017-05-03 14:53:15 -0700 | [diff] [blame] | 3121 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3122 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3123 | .reclaim_idx = MAX_NR_ZONES - 1, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3124 | .target_mem_cgroup = memcg, |
| 3125 | .priority = DEF_PRIORITY, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3126 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3127 | .may_unmap = 1, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3128 | .may_swap = may_swap, |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 3129 | }; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3130 | |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3131 | /* |
| 3132 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't |
| 3133 | * take care of from where we get pages. So the node where we start the |
| 3134 | * scan does not need to be the current node. |
| 3135 | */ |
Johannes Weiner | 72835c8 | 2012-01-12 17:18:32 -0800 | [diff] [blame] | 3136 | nid = mem_cgroup_select_victim_node(memcg); |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3137 | |
Aneesh Kumar K.V | c9634cf | 2016-10-07 16:59:12 -0700 | [diff] [blame] | 3138 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3139 | |
| 3140 | trace_mm_vmscan_memcg_reclaim_begin(0, |
| 3141 | sc.may_writepage, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3142 | sc.gfp_mask, |
| 3143 | sc.reclaim_idx); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3144 | |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3145 | noreclaim_flag = memalloc_noreclaim_save(); |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3146 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3147 | memalloc_noreclaim_restore(noreclaim_flag); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3148 | |
| 3149 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
| 3150 | |
| 3151 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3152 | } |
| 3153 | #endif |
| 3154 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3155 | static void age_active_anon(struct pglist_data *pgdat, |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3156 | struct scan_control *sc) |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3157 | { |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3158 | struct mem_cgroup *memcg; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3159 | |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3160 | if (!total_swap_pages) |
| 3161 | return; |
| 3162 | |
| 3163 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
| 3164 | do { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3165 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3166 | |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 3167 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 3168 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3169 | sc, LRU_ACTIVE_ANON); |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3170 | |
| 3171 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
| 3172 | } while (memcg); |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3173 | } |
| 3174 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3175 | /* |
| 3176 | * Returns true if there is an eligible zone balanced for the request order |
| 3177 | * and classzone_idx |
| 3178 | */ |
| 3179 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
Johannes Weiner | 60cefed | 2012-11-29 13:54:23 -0800 | [diff] [blame] | 3180 | { |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3181 | int i; |
| 3182 | unsigned long mark = -1; |
| 3183 | struct zone *zone; |
Johannes Weiner | 60cefed | 2012-11-29 13:54:23 -0800 | [diff] [blame] | 3184 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3185 | for (i = 0; i <= classzone_idx; i++) { |
| 3186 | zone = pgdat->node_zones + i; |
Mel Gorman | 6256c6b | 2016-07-28 15:45:56 -0700 | [diff] [blame] | 3187 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3188 | if (!managed_zone(zone)) |
| 3189 | continue; |
| 3190 | |
| 3191 | mark = high_wmark_pages(zone); |
| 3192 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) |
| 3193 | return true; |
| 3194 | } |
| 3195 | |
| 3196 | /* |
| 3197 | * If a node has no populated zone within classzone_idx, it does not |
| 3198 | * need balancing by definition. This can happen if a zone-restricted |
| 3199 | * allocation tries to wake a remote kswapd. |
| 3200 | */ |
| 3201 | if (mark == -1) |
| 3202 | return true; |
| 3203 | |
| 3204 | return false; |
Johannes Weiner | 60cefed | 2012-11-29 13:54:23 -0800 | [diff] [blame] | 3205 | } |
| 3206 | |
Mel Gorman | 631b6e0 | 2017-05-03 14:53:41 -0700 | [diff] [blame] | 3207 | /* Clear pgdat state for congested, dirty or under writeback. */ |
| 3208 | static void clear_pgdat_congested(pg_data_t *pgdat) |
| 3209 | { |
| 3210 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); |
| 3211 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
| 3212 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); |
| 3213 | } |
| 3214 | |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 3215 | /* |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3216 | * Prepare kswapd for sleeping. This verifies that there are no processes |
| 3217 | * waiting in throttle_direct_reclaim() and that watermarks have been met. |
| 3218 | * |
| 3219 | * Returns true if kswapd is ready to sleep |
| 3220 | */ |
Mel Gorman | d9f21d4 | 2016-07-28 15:46:41 -0700 | [diff] [blame] | 3221 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 3222 | { |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3223 | /* |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3224 | * The throttled processes are normally woken up in balance_pgdat() as |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3225 | * soon as allow_direct_reclaim() is true. But there is a potential |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3226 | * race between when kswapd checks the watermarks and a process gets |
| 3227 | * throttled. There is also a potential race if processes get |
| 3228 | * throttled, kswapd wakes, a large process exits thereby balancing the |
| 3229 | * zones, which causes kswapd to exit balance_pgdat() before reaching |
| 3230 | * the wake up checks. If kswapd is going to sleep, no process should |
| 3231 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If |
| 3232 | * the wake up is premature, processes will wake kswapd and get |
| 3233 | * throttled again. The difference from wake ups in balance_pgdat() is |
| 3234 | * that here we are under prepare_to_wait(). |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3235 | */ |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3236 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
| 3237 | wake_up_all(&pgdat->pfmemalloc_wait); |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 3238 | |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3239 | /* Hopeless node, leave it to direct reclaim */ |
| 3240 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
| 3241 | return true; |
| 3242 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3243 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
| 3244 | clear_pgdat_congested(pgdat); |
| 3245 | return true; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3246 | } |
| 3247 | |
Shantanu Goel | 333b0a4 | 2017-05-03 14:53:38 -0700 | [diff] [blame] | 3248 | return false; |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 3249 | } |
| 3250 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3251 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3252 | * kswapd shrinks a node of pages that are at or below the highest usable |
| 3253 | * zone that is currently unbalanced. |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3254 | * |
| 3255 | * Returns true if kswapd scanned at least the requested number of pages to |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 3256 | * reclaim or if the lack of progress was due to pages under writeback. |
| 3257 | * This is used to determine if the scanning priority needs to be raised. |
Mel Gorman | 7548536 | 2013-07-03 15:01:42 -0700 | [diff] [blame] | 3258 | */ |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3259 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
Vlastimil Babka | accf624 | 2016-03-17 14:18:15 -0700 | [diff] [blame] | 3260 | struct scan_control *sc) |
Mel Gorman | 7548536 | 2013-07-03 15:01:42 -0700 | [diff] [blame] | 3261 | { |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3262 | struct zone *zone; |
| 3263 | int z; |
Mel Gorman | 7548536 | 2013-07-03 15:01:42 -0700 | [diff] [blame] | 3264 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3265 | /* Reclaim a number of pages proportional to the number of zones */ |
| 3266 | sc->nr_to_reclaim = 0; |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3267 | for (z = 0; z <= sc->reclaim_idx; z++) { |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3268 | zone = pgdat->node_zones + z; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 3269 | if (!managed_zone(zone)) |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3270 | continue; |
Mel Gorman | 7c954f6 | 2013-07-03 15:01:54 -0700 | [diff] [blame] | 3271 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3272 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
Mel Gorman | 7c954f6 | 2013-07-03 15:01:54 -0700 | [diff] [blame] | 3273 | } |
| 3274 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3275 | /* |
| 3276 | * Historically care was taken to put equal pressure on all zones but |
| 3277 | * now pressure is applied based on node LRU order. |
| 3278 | */ |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3279 | shrink_node(pgdat, sc); |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3280 | |
| 3281 | /* |
| 3282 | * Fragmentation may mean that the system cannot be rebalanced for |
| 3283 | * high-order allocations. If twice the allocation size has been |
| 3284 | * reclaimed then recheck watermarks only at order-0 to prevent |
| 3285 | * excessive reclaim. Assume that a process requested a high-order |
| 3286 | * can direct reclaim/compact. |
| 3287 | */ |
Vlastimil Babka | 9861a62 | 2016-10-07 16:57:53 -0700 | [diff] [blame] | 3288 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3289 | sc->order = 0; |
| 3290 | |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3291 | return sc->nr_scanned >= sc->nr_to_reclaim; |
Mel Gorman | 7548536 | 2013-07-03 15:01:42 -0700 | [diff] [blame] | 3292 | } |
| 3293 | |
| 3294 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3295 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
| 3296 | * that are eligible for use by the caller until at least one zone is |
| 3297 | * balanced. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3298 | * |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3299 | * Returns the order kswapd finished reclaiming at. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3300 | * |
| 3301 | * kswapd scans the zones in the highmem->normal->dma direction. It skips |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 3302 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3303 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
| 3304 | * or lower is eligible for reclaim until at least one usable zone is |
| 3305 | * balanced. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3306 | */ |
Vlastimil Babka | accf624 | 2016-03-17 14:18:15 -0700 | [diff] [blame] | 3307 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3308 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3309 | int i; |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 3310 | unsigned long nr_soft_reclaimed; |
| 3311 | unsigned long nr_soft_scanned; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3312 | struct zone *zone; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3313 | struct scan_control sc = { |
| 3314 | .gfp_mask = GFP_KERNEL, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3315 | .order = order, |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3316 | .priority = DEF_PRIORITY, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3317 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3318 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 3319 | .may_swap = 1, |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3320 | }; |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 3321 | count_vm_event(PAGEOUTRUN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3322 | |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3323 | do { |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3324 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3325 | bool raise_priority = true; |
| 3326 | |
Mel Gorman | 84c7a77 | 2016-07-28 15:46:44 -0700 | [diff] [blame] | 3327 | sc.reclaim_idx = classzone_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3328 | |
Mel Gorman | 86c79f6 | 2016-07-28 15:45:59 -0700 | [diff] [blame] | 3329 | /* |
Mel Gorman | 84c7a77 | 2016-07-28 15:46:44 -0700 | [diff] [blame] | 3330 | * If the number of buffer_heads exceeds the maximum allowed |
| 3331 | * then consider reclaiming from all zones. This has a dual |
| 3332 | * purpose -- on 64-bit systems it is expected that |
| 3333 | * buffer_heads are stripped during active rotation. On 32-bit |
| 3334 | * systems, highmem pages can pin lowmem memory and shrinking |
| 3335 | * buffers can relieve lowmem pressure. Reclaim may still not |
| 3336 | * go ahead if all eligible zones for the original allocation |
| 3337 | * request are balanced to avoid excessive reclaim from kswapd. |
Mel Gorman | 86c79f6 | 2016-07-28 15:45:59 -0700 | [diff] [blame] | 3338 | */ |
| 3339 | if (buffer_heads_over_limit) { |
| 3340 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { |
| 3341 | zone = pgdat->node_zones + i; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 3342 | if (!managed_zone(zone)) |
Mel Gorman | 86c79f6 | 2016-07-28 15:45:59 -0700 | [diff] [blame] | 3343 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3344 | |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3345 | sc.reclaim_idx = i; |
Andrew Morton | e1dbeda | 2006-12-06 20:32:01 -0800 | [diff] [blame] | 3346 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3347 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3348 | } |
Zlatko Calusic | dafcb73 | 2013-02-22 16:32:34 -0800 | [diff] [blame] | 3349 | |
Mel Gorman | 86c79f6 | 2016-07-28 15:45:59 -0700 | [diff] [blame] | 3350 | /* |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3351 | * Only reclaim if there are no eligible zones. Note that |
| 3352 | * sc.reclaim_idx is not used as buffer_heads_over_limit may |
| 3353 | * have adjusted it. |
Mel Gorman | 86c79f6 | 2016-07-28 15:45:59 -0700 | [diff] [blame] | 3354 | */ |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3355 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) |
| 3356 | goto out; |
Andrew Morton | e1dbeda | 2006-12-06 20:32:01 -0800 | [diff] [blame] | 3357 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3358 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3359 | * Do some background aging of the anon list, to give |
| 3360 | * pages a chance to be referenced before reclaiming. All |
| 3361 | * pages are rotated regardless of classzone as this is |
| 3362 | * about consistent aging. |
| 3363 | */ |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3364 | age_active_anon(pgdat, &sc); |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3365 | |
| 3366 | /* |
Mel Gorman | b7ea3c4 | 2013-07-03 15:01:53 -0700 | [diff] [blame] | 3367 | * If we're getting trouble reclaiming, start doing writepage |
| 3368 | * even in laptop mode. |
| 3369 | */ |
Johannes Weiner | 047d72c | 2017-05-03 14:51:57 -0700 | [diff] [blame] | 3370 | if (sc.priority < DEF_PRIORITY - 2) |
Mel Gorman | b7ea3c4 | 2013-07-03 15:01:53 -0700 | [diff] [blame] | 3371 | sc.may_writepage = 1; |
| 3372 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3373 | /* Call soft limit reclaim before calling shrink_node. */ |
| 3374 | sc.nr_scanned = 0; |
| 3375 | nr_soft_scanned = 0; |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3376 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3377 | sc.gfp_mask, &nr_soft_scanned); |
| 3378 | sc.nr_reclaimed += nr_soft_reclaimed; |
| 3379 | |
Mel Gorman | b7ea3c4 | 2013-07-03 15:01:53 -0700 | [diff] [blame] | 3380 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3381 | * There should be no need to raise the scanning priority if |
| 3382 | * enough pages are already being scanned that that high |
| 3383 | * watermark would be met at 100% efficiency. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3384 | */ |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3385 | if (kswapd_shrink_node(pgdat, &sc)) |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3386 | raise_priority = false; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3387 | |
| 3388 | /* |
| 3389 | * If the low watermark is met there is no need for processes |
| 3390 | * to be throttled on pfmemalloc_wait as they should not be |
| 3391 | * able to safely make forward progress. Wake them |
| 3392 | */ |
| 3393 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3394 | allow_direct_reclaim(pgdat)) |
Vlastimil Babka | cfc5115 | 2015-02-11 15:25:12 -0800 | [diff] [blame] | 3395 | wake_up_all(&pgdat->pfmemalloc_wait); |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3396 | |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3397 | /* Check if kswapd should be suspending */ |
| 3398 | if (try_to_freeze() || kthread_should_stop()) |
| 3399 | break; |
| 3400 | |
| 3401 | /* |
| 3402 | * Raise priority if scanning rate is too low or there was no |
| 3403 | * progress in reclaiming pages |
| 3404 | */ |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3405 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
| 3406 | if (raise_priority || !nr_reclaimed) |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3407 | sc.priority--; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3408 | } while (sc.priority >= 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3409 | |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3410 | if (!sc.nr_reclaimed) |
| 3411 | pgdat->kswapd_failures++; |
| 3412 | |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3413 | out: |
Johannes Weiner | 2a2e488 | 2017-05-03 14:55:03 -0700 | [diff] [blame] | 3414 | snapshot_refaults(NULL, pgdat); |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 3415 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3416 | * Return the order kswapd stopped reclaiming at as |
| 3417 | * prepare_kswapd_sleep() takes it into account. If another caller |
| 3418 | * entered the allocator slow path while kswapd was awake, order will |
| 3419 | * remain at the higher level. |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 3420 | */ |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3421 | return sc.order; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3422 | } |
| 3423 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3424 | /* |
| 3425 | * pgdat->kswapd_classzone_idx is the highest zone index that a recent |
| 3426 | * allocation request woke kswapd for. When kswapd has not woken recently, |
| 3427 | * the value is MAX_NR_ZONES which is not a valid index. This compares a |
| 3428 | * given classzone and returns it or the highest classzone index kswapd |
| 3429 | * was recently woke for. |
| 3430 | */ |
| 3431 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, |
| 3432 | enum zone_type classzone_idx) |
| 3433 | { |
| 3434 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) |
| 3435 | return classzone_idx; |
| 3436 | |
| 3437 | return max(pgdat->kswapd_classzone_idx, classzone_idx); |
| 3438 | } |
| 3439 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3440 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
| 3441 | unsigned int classzone_idx) |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3442 | { |
| 3443 | long remaining = 0; |
| 3444 | DEFINE_WAIT(wait); |
| 3445 | |
| 3446 | if (freezing(current) || kthread_should_stop()) |
| 3447 | return; |
| 3448 | |
| 3449 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
| 3450 | |
Shantanu Goel | 333b0a4 | 2017-05-03 14:53:38 -0700 | [diff] [blame] | 3451 | /* |
| 3452 | * Try to sleep for a short interval. Note that kcompactd will only be |
| 3453 | * woken if it is possible to sleep for a short interval. This is |
| 3454 | * deliberate on the assumption that if reclaim cannot keep an |
| 3455 | * eligible zone balanced that it's also unlikely that compaction will |
| 3456 | * succeed. |
| 3457 | */ |
Mel Gorman | d9f21d4 | 2016-07-28 15:46:41 -0700 | [diff] [blame] | 3458 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
Vlastimil Babka | fd901c9 | 2016-04-28 16:18:49 -0700 | [diff] [blame] | 3459 | /* |
| 3460 | * Compaction records what page blocks it recently failed to |
| 3461 | * isolate pages from and skips them in the future scanning. |
| 3462 | * When kswapd is going to sleep, it is reasonable to assume |
| 3463 | * that pages and compaction may succeed so reset the cache. |
| 3464 | */ |
| 3465 | reset_isolation_suitable(pgdat); |
| 3466 | |
| 3467 | /* |
| 3468 | * We have freed the memory, now we should compact it to make |
| 3469 | * allocation of the requested order possible. |
| 3470 | */ |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3471 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
Vlastimil Babka | fd901c9 | 2016-04-28 16:18:49 -0700 | [diff] [blame] | 3472 | |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3473 | remaining = schedule_timeout(HZ/10); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3474 | |
| 3475 | /* |
| 3476 | * If woken prematurely then reset kswapd_classzone_idx and |
| 3477 | * order. The values will either be from a wakeup request or |
| 3478 | * the previous request that slept prematurely. |
| 3479 | */ |
| 3480 | if (remaining) { |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3481 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3482 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
| 3483 | } |
| 3484 | |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3485 | finish_wait(&pgdat->kswapd_wait, &wait); |
| 3486 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
| 3487 | } |
| 3488 | |
| 3489 | /* |
| 3490 | * After a short sleep, check if it was a premature sleep. If not, then |
| 3491 | * go fully to sleep until explicitly woken up. |
| 3492 | */ |
Mel Gorman | d9f21d4 | 2016-07-28 15:46:41 -0700 | [diff] [blame] | 3493 | if (!remaining && |
| 3494 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3495 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
| 3496 | |
| 3497 | /* |
| 3498 | * vmstat counters are not perfectly accurate and the estimated |
| 3499 | * value for counters such as NR_FREE_PAGES can deviate from the |
| 3500 | * true value by nr_online_cpus * threshold. To avoid the zone |
| 3501 | * watermarks being breached while under pressure, we reduce the |
| 3502 | * per-cpu vmstat threshold while kswapd is awake and restore |
| 3503 | * them before going back to sleep. |
| 3504 | */ |
| 3505 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); |
Aaditya Kumar | 1c7e7f6 | 2012-07-17 15:48:07 -0700 | [diff] [blame] | 3506 | |
| 3507 | if (!kthread_should_stop()) |
| 3508 | schedule(); |
| 3509 | |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3510 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
| 3511 | } else { |
| 3512 | if (remaining) |
| 3513 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); |
| 3514 | else |
| 3515 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); |
| 3516 | } |
| 3517 | finish_wait(&pgdat->kswapd_wait, &wait); |
| 3518 | } |
| 3519 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3520 | /* |
| 3521 | * The background pageout daemon, started as a kernel thread |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 3522 | * from the init process. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3523 | * |
| 3524 | * This basically trickles out pages so that we have _some_ |
| 3525 | * free memory available even if there is no other activity |
| 3526 | * that frees anything up. This is needed for things like routing |
| 3527 | * etc, where we otherwise might have all activity going on in |
| 3528 | * asynchronous contexts that cannot page things out. |
| 3529 | * |
| 3530 | * If there are applications that are active memory-allocators |
| 3531 | * (most normal use), this basically shouldn't matter. |
| 3532 | */ |
| 3533 | static int kswapd(void *p) |
| 3534 | { |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3535 | unsigned int alloc_order, reclaim_order; |
| 3536 | unsigned int classzone_idx = MAX_NR_ZONES - 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3537 | pg_data_t *pgdat = (pg_data_t*)p; |
| 3538 | struct task_struct *tsk = current; |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 3539 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3540 | struct reclaim_state reclaim_state = { |
| 3541 | .reclaimed_slab = 0, |
| 3542 | }; |
Rusty Russell | a70f730 | 2009-03-13 14:49:46 +1030 | [diff] [blame] | 3543 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3544 | |
Rusty Russell | 174596a | 2009-01-01 10:12:29 +1030 | [diff] [blame] | 3545 | if (!cpumask_empty(cpumask)) |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3546 | set_cpus_allowed_ptr(tsk, cpumask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3547 | current->reclaim_state = &reclaim_state; |
| 3548 | |
| 3549 | /* |
| 3550 | * Tell the memory management that we're a "memory allocator", |
| 3551 | * and that if we need more memory we should get access to it |
| 3552 | * regardless (see "__alloc_pages()"). "kswapd" should |
| 3553 | * never get caught in the normal page freeing logic. |
| 3554 | * |
| 3555 | * (Kswapd normally doesn't need memory anyway, but sometimes |
| 3556 | * you need a small amount of memory in order to be able to |
| 3557 | * page out something else, and this flag essentially protects |
| 3558 | * us from recursively trying to free more memory as we're |
| 3559 | * trying to free the first piece of memory in the first place). |
| 3560 | */ |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 3561 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
Rafael J. Wysocki | 8314418 | 2007-07-17 04:03:35 -0700 | [diff] [blame] | 3562 | set_freezable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3563 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3564 | pgdat->kswapd_order = 0; |
| 3565 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3566 | for ( ; ; ) { |
Jeff Liu | 6f6313d | 2012-12-11 16:02:48 -0800 | [diff] [blame] | 3567 | bool ret; |
Christoph Lameter | 3e1d1d2 | 2005-06-24 23:13:50 -0700 | [diff] [blame] | 3568 | |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3569 | alloc_order = reclaim_order = pgdat->kswapd_order; |
| 3570 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
| 3571 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3572 | kswapd_try_sleep: |
| 3573 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, |
| 3574 | classzone_idx); |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 3575 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3576 | /* Read the new order and classzone_idx */ |
| 3577 | alloc_order = reclaim_order = pgdat->kswapd_order; |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3578 | classzone_idx = kswapd_classzone_idx(pgdat, 0); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3579 | pgdat->kswapd_order = 0; |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3580 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3581 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3582 | ret = try_to_freeze(); |
| 3583 | if (kthread_should_stop()) |
| 3584 | break; |
| 3585 | |
| 3586 | /* |
| 3587 | * We can speed up thawing tasks if we don't call balance_pgdat |
| 3588 | * after returning from the refrigerator |
| 3589 | */ |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3590 | if (ret) |
| 3591 | continue; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3592 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3593 | /* |
| 3594 | * Reclaim begins at the requested order but if a high-order |
| 3595 | * reclaim fails then kswapd falls back to reclaiming for |
| 3596 | * order-0. If that happens, kswapd will consider sleeping |
| 3597 | * for the order it finished reclaiming at (reclaim_order) |
| 3598 | * but kcompactd is woken to compact for the original |
| 3599 | * request (alloc_order). |
| 3600 | */ |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3601 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
| 3602 | alloc_order); |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3603 | fs_reclaim_acquire(GFP_KERNEL); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3604 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3605 | fs_reclaim_release(GFP_KERNEL); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3606 | if (reclaim_order < alloc_order) |
| 3607 | goto kswapd_try_sleep; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3608 | } |
Takamori Yamaguchi | b0a8cc5 | 2012-11-08 15:53:39 -0800 | [diff] [blame] | 3609 | |
Johannes Weiner | 71abdc1 | 2014-06-06 14:35:35 -0700 | [diff] [blame] | 3610 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
Takamori Yamaguchi | b0a8cc5 | 2012-11-08 15:53:39 -0800 | [diff] [blame] | 3611 | current->reclaim_state = NULL; |
Johannes Weiner | 71abdc1 | 2014-06-06 14:35:35 -0700 | [diff] [blame] | 3612 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3613 | return 0; |
| 3614 | } |
| 3615 | |
| 3616 | /* |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3617 | * A zone is low on free memory or too fragmented for high-order memory. If |
| 3618 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's |
| 3619 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim |
| 3620 | * has failed or is not needed, still wake up kcompactd if only compaction is |
| 3621 | * needed. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3622 | */ |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3623 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
| 3624 | enum zone_type classzone_idx) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3625 | { |
| 3626 | pg_data_t *pgdat; |
| 3627 | |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 3628 | if (!managed_zone(zone)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3629 | return; |
| 3630 | |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3631 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3632 | return; |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 3633 | pgdat = zone->zone_pgdat; |
Mel Gorman | e716f2e | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3634 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, |
| 3635 | classzone_idx); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3636 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 3637 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3638 | return; |
Mel Gorman | e1a5563 | 2016-07-28 15:46:26 -0700 | [diff] [blame] | 3639 | |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3640 | /* Hopeless node, leave it to direct reclaim if possible */ |
| 3641 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || |
| 3642 | pgdat_balanced(pgdat, order, classzone_idx)) { |
| 3643 | /* |
| 3644 | * There may be plenty of free memory available, but it's too |
| 3645 | * fragmented for high-order allocations. Wake up kcompactd |
| 3646 | * and rely on compaction_suitable() to determine if it's |
| 3647 | * needed. If it fails, it will defer subsequent attempts to |
| 3648 | * ratelimit its work. |
| 3649 | */ |
| 3650 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) |
| 3651 | wakeup_kcompactd(pgdat, order, classzone_idx); |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3652 | return; |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3653 | } |
Johannes Weiner | c73322d | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3654 | |
David Rientjes | 5ecd9d4 | 2018-04-05 16:25:16 -0700 | [diff] [blame] | 3655 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, |
| 3656 | gfp_flags); |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 3657 | wake_up_interruptible(&pgdat->kswapd_wait); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3658 | } |
| 3659 | |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 3660 | #ifdef CONFIG_HIBERNATION |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3661 | /* |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3662 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3663 | * freed pages. |
| 3664 | * |
| 3665 | * Rather than trying to age LRUs the aim is to preserve the overall |
| 3666 | * LRU order by reclaiming preferentially |
| 3667 | * inactive > active > active referenced > active mapped |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3668 | */ |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3669 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3670 | { |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3671 | struct reclaim_state reclaim_state; |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3672 | struct scan_control sc = { |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3673 | .nr_to_reclaim = nr_to_reclaim, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3674 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3675 | .reclaim_idx = MAX_NR_ZONES - 1, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3676 | .priority = DEF_PRIORITY, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3677 | .may_writepage = 1, |
| 3678 | .may_unmap = 1, |
| 3679 | .may_swap = 1, |
| 3680 | .hibernation_mode = 1, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3681 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 3682 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3683 | struct task_struct *p = current; |
| 3684 | unsigned long nr_reclaimed; |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3685 | unsigned int noreclaim_flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3686 | |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3687 | noreclaim_flag = memalloc_noreclaim_save(); |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3688 | fs_reclaim_acquire(sc.gfp_mask); |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3689 | reclaim_state.reclaimed_slab = 0; |
| 3690 | p->reclaim_state = &reclaim_state; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3691 | |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3692 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3693 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3694 | p->reclaim_state = NULL; |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3695 | fs_reclaim_release(sc.gfp_mask); |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3696 | memalloc_noreclaim_restore(noreclaim_flag); |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3697 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3698 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3699 | } |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 3700 | #endif /* CONFIG_HIBERNATION */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3701 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3702 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
| 3703 | not required for correctness. So if the last cpu in a node goes |
| 3704 | away, we get changed to run anywhere: as the first one comes back, |
| 3705 | restore their cpu bindings. */ |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3706 | static int kswapd_cpu_online(unsigned int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3707 | { |
Yasunori Goto | 58c0a4a | 2007-10-16 01:25:40 -0700 | [diff] [blame] | 3708 | int nid; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3709 | |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3710 | for_each_node_state(nid, N_MEMORY) { |
| 3711 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3712 | const struct cpumask *mask; |
Rusty Russell | a70f730 | 2009-03-13 14:49:46 +1030 | [diff] [blame] | 3713 | |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3714 | mask = cpumask_of_node(pgdat->node_id); |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3715 | |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3716 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
| 3717 | /* One of our CPUs online: restore mask */ |
| 3718 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3719 | } |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3720 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3721 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3722 | |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3723 | /* |
| 3724 | * This kswapd start function will be called by init and node-hot-add. |
| 3725 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. |
| 3726 | */ |
| 3727 | int kswapd_run(int nid) |
| 3728 | { |
| 3729 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3730 | int ret = 0; |
| 3731 | |
| 3732 | if (pgdat->kswapd) |
| 3733 | return 0; |
| 3734 | |
| 3735 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); |
| 3736 | if (IS_ERR(pgdat->kswapd)) { |
| 3737 | /* failure at boot is fatal */ |
Thomas Gleixner | c6202ad | 2017-05-16 20:42:46 +0200 | [diff] [blame] | 3738 | BUG_ON(system_state < SYSTEM_RUNNING); |
Gavin Shan | d5dc0ad | 2012-10-08 16:29:27 -0700 | [diff] [blame] | 3739 | pr_err("Failed to start kswapd on node %d\n", nid); |
| 3740 | ret = PTR_ERR(pgdat->kswapd); |
Xishi Qiu | d72515b | 2013-04-17 15:58:34 -0700 | [diff] [blame] | 3741 | pgdat->kswapd = NULL; |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3742 | } |
| 3743 | return ret; |
| 3744 | } |
| 3745 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3746 | /* |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3747 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
Vladimir Davydov | bfc8c90 | 2014-06-04 16:07:18 -0700 | [diff] [blame] | 3748 | * hold mem_hotplug_begin/end(). |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3749 | */ |
| 3750 | void kswapd_stop(int nid) |
| 3751 | { |
| 3752 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; |
| 3753 | |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3754 | if (kswapd) { |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3755 | kthread_stop(kswapd); |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3756 | NODE_DATA(nid)->kswapd = NULL; |
| 3757 | } |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3758 | } |
| 3759 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3760 | static int __init kswapd_init(void) |
| 3761 | { |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3762 | int nid, ret; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3763 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3764 | swap_setup(); |
Lai Jiangshan | 48fb2e2 | 2012-12-12 13:51:43 -0800 | [diff] [blame] | 3765 | for_each_node_state(nid, N_MEMORY) |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3766 | kswapd_run(nid); |
Sebastian Andrzej Siewior | 517bbed | 2016-11-03 15:50:03 +0100 | [diff] [blame] | 3767 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
| 3768 | "mm/vmscan:online", kswapd_cpu_online, |
| 3769 | NULL); |
| 3770 | WARN_ON(ret < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3771 | return 0; |
| 3772 | } |
| 3773 | |
| 3774 | module_init(kswapd_init) |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3775 | |
| 3776 | #ifdef CONFIG_NUMA |
| 3777 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3778 | * Node reclaim mode |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3779 | * |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3780 | * If non-zero call node_reclaim when the number of free pages falls below |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3781 | * the watermarks. |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3782 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3783 | int node_reclaim_mode __read_mostly; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3784 | |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3785 | #define RECLAIM_OFF 0 |
Fernando Luis Vazquez Cao | 7d03431 | 2008-07-29 22:33:41 -0700 | [diff] [blame] | 3786 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3787 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3788 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3789 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3790 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3791 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
Christoph Lameter | a92f712 | 2006-02-01 03:05:32 -0800 | [diff] [blame] | 3792 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
| 3793 | * a zone. |
| 3794 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3795 | #define NODE_RECLAIM_PRIORITY 4 |
Christoph Lameter | a92f712 | 2006-02-01 03:05:32 -0800 | [diff] [blame] | 3796 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3797 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3798 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3799 | * occur. |
| 3800 | */ |
| 3801 | int sysctl_min_unmapped_ratio = 1; |
| 3802 | |
| 3803 | /* |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3804 | * If the number of slab pages in a zone grows beyond this percentage then |
| 3805 | * slab reclaim needs to occur. |
| 3806 | */ |
| 3807 | int sysctl_min_slab_ratio = 5; |
| 3808 | |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3809 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3810 | { |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3811 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
| 3812 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + |
| 3813 | node_page_state(pgdat, NR_ACTIVE_FILE); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3814 | |
| 3815 | /* |
| 3816 | * It's possible for there to be more file mapped pages than |
| 3817 | * accounted for by the pages on the file LRU lists because |
| 3818 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED |
| 3819 | */ |
| 3820 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; |
| 3821 | } |
| 3822 | |
| 3823 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3824 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3825 | { |
Alexandru Moise | d031a15 | 2015-11-05 18:48:08 -0800 | [diff] [blame] | 3826 | unsigned long nr_pagecache_reclaimable; |
| 3827 | unsigned long delta = 0; |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3828 | |
| 3829 | /* |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3830 | * If RECLAIM_UNMAP is set, then all file pages are considered |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3831 | * potentially reclaimable. Otherwise, we have to worry about |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3832 | * pages like swapcache and node_unmapped_file_pages() provides |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3833 | * a better estimate |
| 3834 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3835 | if (node_reclaim_mode & RECLAIM_UNMAP) |
| 3836 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3837 | else |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3838 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3839 | |
| 3840 | /* If we can't clean pages, remove dirty pages from consideration */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3841 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
| 3842 | delta += node_page_state(pgdat, NR_FILE_DIRTY); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3843 | |
| 3844 | /* Watch for any possible underflows due to delta */ |
| 3845 | if (unlikely(delta > nr_pagecache_reclaimable)) |
| 3846 | delta = nr_pagecache_reclaimable; |
| 3847 | |
| 3848 | return nr_pagecache_reclaimable - delta; |
| 3849 | } |
| 3850 | |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3851 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3852 | * Try to free up some pages from this node through reclaim. |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3853 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3854 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3855 | { |
Christoph Lameter | 7fb2d46 | 2006-03-22 00:08:22 -0800 | [diff] [blame] | 3856 | /* Minimum pages needed in order to stay on node */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3857 | const unsigned long nr_pages = 1 << order; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3858 | struct task_struct *p = current; |
| 3859 | struct reclaim_state reclaim_state; |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3860 | unsigned int noreclaim_flag; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3861 | struct scan_control sc = { |
Andrew Morton | 62b726c | 2013-02-22 16:32:24 -0800 | [diff] [blame] | 3862 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
Nick Desaulniers | f2f43e5 | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3863 | .gfp_mask = current_gfp_context(gfp_mask), |
Johannes Weiner | bd2f619 | 2009-03-31 15:19:38 -0700 | [diff] [blame] | 3864 | .order = order, |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3865 | .priority = NODE_RECLAIM_PRIORITY, |
| 3866 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), |
| 3867 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3868 | .may_swap = 1, |
Nick Desaulniers | f2f43e5 | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3869 | .reclaim_idx = gfp_zone(gfp_mask), |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3870 | }; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3871 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3872 | cond_resched(); |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3873 | /* |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3874 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3875 | * and we also need to be able to write out pages for RECLAIM_WRITE |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3876 | * and RECLAIM_UNMAP. |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3877 | */ |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3878 | noreclaim_flag = memalloc_noreclaim_save(); |
| 3879 | p->flags |= PF_SWAPWRITE; |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3880 | fs_reclaim_acquire(sc.gfp_mask); |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3881 | reclaim_state.reclaimed_slab = 0; |
| 3882 | p->reclaim_state = &reclaim_state; |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3883 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3884 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3885 | /* |
Andrey Ryabinin | 894befe | 2018-04-10 16:27:51 -0700 | [diff] [blame] | 3886 | * Free memory by calling shrink node with increasing |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3887 | * priorities until we have enough memory freed. |
| 3888 | */ |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3889 | do { |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3890 | shrink_node(pgdat, &sc); |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3891 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3892 | } |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3893 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3894 | p->reclaim_state = NULL; |
Peter Zijlstra | d92a8cf | 2017-03-03 10:13:38 +0100 | [diff] [blame] | 3895 | fs_reclaim_release(gfp_mask); |
Vlastimil Babka | 499118e | 2017-05-08 15:59:50 -0700 | [diff] [blame] | 3896 | current->flags &= ~PF_SWAPWRITE; |
| 3897 | memalloc_noreclaim_restore(noreclaim_flag); |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 3898 | return sc.nr_reclaimed >= nr_pages; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3899 | } |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3900 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3901 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3902 | { |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3903 | int ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3904 | |
| 3905 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3906 | * Node reclaim reclaims unmapped file backed pages and |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3907 | * slab pages if we are over the defined limits. |
Christoph Lameter | 34aa133 | 2006-06-30 01:55:37 -0700 | [diff] [blame] | 3908 | * |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3909 | * A small portion of unmapped file backed pages is needed for |
| 3910 | * file I/O otherwise pages read by file I/O will be immediately |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3911 | * thrown out if the node is overallocated. So we do not reclaim |
| 3912 | * if less than a specified percentage of the node is used by |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3913 | * unmapped file backed pages. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3914 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3915 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
Johannes Weiner | 385386c | 2017-07-06 15:40:43 -0700 | [diff] [blame] | 3916 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3917 | return NODE_RECLAIM_FULL; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3918 | |
| 3919 | /* |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3920 | * Do not scan if the allocation should not be delayed. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3921 | */ |
Mel Gorman | d0164ad | 2015-11-06 16:28:21 -0800 | [diff] [blame] | 3922 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3923 | return NODE_RECLAIM_NOSCAN; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3924 | |
| 3925 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3926 | * Only run node reclaim on the local node or on nodes that do not |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3927 | * have associated processors. This will favor the local processor |
| 3928 | * over remote processors and spread off node memory allocations |
| 3929 | * as wide as possible. |
| 3930 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3931 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
| 3932 | return NODE_RECLAIM_NOSCAN; |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3933 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3934 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
| 3935 | return NODE_RECLAIM_NOSCAN; |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3936 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3937 | ret = __node_reclaim(pgdat, gfp_mask, order); |
| 3938 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3939 | |
Mel Gorman | 24cf72518 | 2009-06-16 15:33:23 -0700 | [diff] [blame] | 3940 | if (!ret) |
| 3941 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); |
| 3942 | |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3943 | return ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3944 | } |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3945 | #endif |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3946 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3947 | /* |
| 3948 | * page_evictable - test whether a page is evictable |
| 3949 | * @page: the page to test |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3950 | * |
| 3951 | * Test whether page is evictable--i.e., should be placed on active/inactive |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 3952 | * lists vs unevictable list. |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3953 | * |
| 3954 | * Reasons page might not be evictable: |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3955 | * (1) page's mapping marked unevictable |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 3956 | * (2) page is part of an mlocked VMA |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3957 | * |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3958 | */ |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 3959 | int page_evictable(struct page *page) |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3960 | { |
Huang Ying | e92bb4d | 2018-04-05 16:23:20 -0700 | [diff] [blame] | 3961 | int ret; |
| 3962 | |
| 3963 | /* Prevent address_space of inode and swap cache from being freed */ |
| 3964 | rcu_read_lock(); |
| 3965 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
| 3966 | rcu_read_unlock(); |
| 3967 | return ret; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3968 | } |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3969 | |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 3970 | #ifdef CONFIG_SHMEM |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3971 | /** |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3972 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
| 3973 | * @pages: array of pages to check |
| 3974 | * @nr_pages: number of pages to check |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3975 | * |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3976 | * Checks pages for evictability and moves them to the appropriate lru list. |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 3977 | * |
| 3978 | * This function is only used for SysV IPC SHM_UNLOCK. |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3979 | */ |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3980 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3981 | { |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 3982 | struct lruvec *lruvec; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3983 | struct pglist_data *pgdat = NULL; |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3984 | int pgscanned = 0; |
| 3985 | int pgrescued = 0; |
| 3986 | int i; |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3987 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3988 | for (i = 0; i < nr_pages; i++) { |
| 3989 | struct page *page = pages[i]; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3990 | struct pglist_data *pagepgdat = page_pgdat(page); |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3991 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3992 | pgscanned++; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3993 | if (pagepgdat != pgdat) { |
| 3994 | if (pgdat) |
| 3995 | spin_unlock_irq(&pgdat->lru_lock); |
| 3996 | pgdat = pagepgdat; |
| 3997 | spin_lock_irq(&pgdat->lru_lock); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3998 | } |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3999 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 4000 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4001 | if (!PageLRU(page) || !PageUnevictable(page)) |
| 4002 | continue; |
| 4003 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 4004 | if (page_evictable(page)) { |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4005 | enum lru_list lru = page_lru_base_type(page); |
| 4006 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 4007 | VM_BUG_ON_PAGE(PageActive(page), page); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4008 | ClearPageUnevictable(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 4009 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
| 4010 | add_page_to_lru_list(page, lruvec, lru); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4011 | pgrescued++; |
| 4012 | } |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 4013 | } |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4014 | |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 4015 | if (pgdat) { |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4016 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
| 4017 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 4018 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4019 | } |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 4020 | } |
| 4021 | #endif /* CONFIG_SHMEM */ |