Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * linux/mm/vmscan.c |
| 3 | * |
| 4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
| 5 | * |
| 6 | * Swap reorganised 29.12.95, Stephen Tweedie. |
| 7 | * kswapd added: 7.1.96 sct |
| 8 | * Removed kswapd_ctl limits, and swap out as many pages as needed |
| 9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. |
| 10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). |
| 11 | * Multiqueue VM started 5.8.00, Rik van Riel. |
| 12 | */ |
| 13 | |
Mitchel Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 15 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 16 | #include <linux/mm.h> |
| 17 | #include <linux/module.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 18 | #include <linux/gfp.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 19 | #include <linux/kernel_stat.h> |
| 20 | #include <linux/swap.h> |
| 21 | #include <linux/pagemap.h> |
| 22 | #include <linux/init.h> |
| 23 | #include <linux/highmem.h> |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 24 | #include <linux/vmpressure.h> |
Andrew Morton | e129b5c | 2006-09-27 01:50:00 -0700 | [diff] [blame] | 25 | #include <linux/vmstat.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 26 | #include <linux/file.h> |
| 27 | #include <linux/writeback.h> |
| 28 | #include <linux/blkdev.h> |
| 29 | #include <linux/buffer_head.h> /* for try_to_release_page(), |
| 30 | buffer_heads_over_limit */ |
| 31 | #include <linux/mm_inline.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 32 | #include <linux/backing-dev.h> |
| 33 | #include <linux/rmap.h> |
| 34 | #include <linux/topology.h> |
| 35 | #include <linux/cpu.h> |
| 36 | #include <linux/cpuset.h> |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 37 | #include <linux/compaction.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 38 | #include <linux/notifier.h> |
| 39 | #include <linux/rwsem.h> |
Rafael J. Wysocki | 248a030 | 2006-03-22 00:09:04 -0800 | [diff] [blame] | 40 | #include <linux/delay.h> |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 41 | #include <linux/kthread.h> |
Nigel Cunningham | 7dfb710 | 2006-12-06 20:34:23 -0800 | [diff] [blame] | 42 | #include <linux/freezer.h> |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 43 | #include <linux/memcontrol.h> |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 44 | #include <linux/delayacct.h> |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 45 | #include <linux/sysctl.h> |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 46 | #include <linux/oom.h> |
Linus Torvalds | 268bb0c | 2011-05-20 12:50:29 -0700 | [diff] [blame] | 47 | #include <linux/prefetch.h> |
Mitchel Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 48 | #include <linux/printk.h> |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 49 | #include <linux/dax.h> |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 50 | #include <linux/psi.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 51 | |
| 52 | #include <asm/tlbflush.h> |
| 53 | #include <asm/div64.h> |
| 54 | |
| 55 | #include <linux/swapops.h> |
Rafael Aquini | 117aad1 | 2013-09-30 13:45:16 -0700 | [diff] [blame] | 56 | #include <linux/balloon_compaction.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 57 | |
Nick Piggin | 0f8053a | 2006-03-22 00:08:33 -0800 | [diff] [blame] | 58 | #include "internal.h" |
| 59 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 60 | #define CREATE_TRACE_POINTS |
| 61 | #include <trace/events/vmscan.h> |
| 62 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 63 | struct scan_control { |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 64 | /* How many pages shrink_list() should reclaim */ |
| 65 | unsigned long nr_to_reclaim; |
| 66 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 67 | /* This context's GFP mask */ |
Al Viro | 6daa0e2 | 2005-10-21 03:18:50 -0400 | [diff] [blame] | 68 | gfp_t gfp_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 69 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 70 | /* Allocation order */ |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 71 | int order; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 72 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 73 | /* |
| 74 | * Nodemask of nodes allowed by the caller. If NULL, all nodes |
| 75 | * are scanned. |
| 76 | */ |
| 77 | nodemask_t *nodemask; |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 78 | |
KOSAKI Motohiro | 5f53e76 | 2010-05-24 14:32:37 -0700 | [diff] [blame] | 79 | /* |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 80 | * The memory cgroup that hit its limit and as a result is the |
| 81 | * primary target of this reclaim invocation. |
| 82 | */ |
| 83 | struct mem_cgroup *target_mem_cgroup; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 84 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 85 | /* Scan (total_size >> priority) pages at once */ |
| 86 | int priority; |
| 87 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 88 | /* The highest zone to isolate pages for reclaim from */ |
| 89 | enum zone_type reclaim_idx; |
| 90 | |
Johannes Weiner | 14d6ecf | 2017-02-24 14:56:11 -0800 | [diff] [blame] | 91 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 92 | unsigned int may_writepage:1; |
| 93 | |
| 94 | /* Can mapped pages be reclaimed? */ |
| 95 | unsigned int may_unmap:1; |
| 96 | |
| 97 | /* Can pages be swapped as part of reclaim? */ |
| 98 | unsigned int may_swap:1; |
| 99 | |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 100 | /* Can cgroups be reclaimed below their normal consumption range? */ |
| 101 | unsigned int may_thrash:1; |
| 102 | |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 103 | unsigned int hibernation_mode:1; |
| 104 | |
| 105 | /* One of the zones is ready for compaction */ |
| 106 | unsigned int compaction_ready:1; |
| 107 | |
| 108 | /* Incremented by the number of inactive pages that were scanned */ |
| 109 | unsigned long nr_scanned; |
| 110 | |
| 111 | /* Number of pages freed so far during a call to shrink_zones() */ |
| 112 | unsigned long nr_reclaimed; |
Minchan Kim | cd256bf | 2013-05-09 16:21:27 +0900 | [diff] [blame] | 113 | |
| 114 | /* |
| 115 | * Reclaim pages from a vma. If the page is shared by other tasks |
| 116 | * it is zapped from a vma without reclaim so it ends up remaining |
| 117 | * on memory until last task zap it. |
| 118 | */ |
| 119 | struct vm_area_struct *target_vma; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 120 | }; |
| 121 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 122 | #ifdef ARCH_HAS_PREFETCH |
| 123 | #define prefetch_prev_lru_page(_page, _base, _field) \ |
| 124 | do { \ |
| 125 | if ((_page)->lru.prev != _base) { \ |
| 126 | struct page *prev; \ |
| 127 | \ |
| 128 | prev = lru_to_page(&(_page->lru)); \ |
| 129 | prefetch(&prev->_field); \ |
| 130 | } \ |
| 131 | } while (0) |
| 132 | #else |
| 133 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) |
| 134 | #endif |
| 135 | |
| 136 | #ifdef ARCH_HAS_PREFETCHW |
| 137 | #define prefetchw_prev_lru_page(_page, _base, _field) \ |
| 138 | do { \ |
| 139 | if ((_page)->lru.prev != _base) { \ |
| 140 | struct page *prev; \ |
| 141 | \ |
| 142 | prev = lru_to_page(&(_page->lru)); \ |
| 143 | prefetchw(&prev->_field); \ |
| 144 | } \ |
| 145 | } while (0) |
| 146 | #else |
| 147 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) |
| 148 | #endif |
| 149 | |
| 150 | /* |
| 151 | * From 0 .. 100. Higher means more swappy. |
| 152 | */ |
| 153 | int vm_swappiness = 60; |
Wang Sheng-Hui | d0480be | 2014-08-06 16:07:07 -0700 | [diff] [blame] | 154 | /* |
| 155 | * The total number of pages which are beyond the high watermark within all |
| 156 | * zones. |
| 157 | */ |
| 158 | unsigned long vm_total_pages; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 159 | |
| 160 | static LIST_HEAD(shrinker_list); |
| 161 | static DECLARE_RWSEM(shrinker_rwsem); |
| 162 | |
Andrew Morton | c255a45 | 2012-07-31 16:43:02 -0700 | [diff] [blame] | 163 | #ifdef CONFIG_MEMCG |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 164 | static bool global_reclaim(struct scan_control *sc) |
| 165 | { |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 166 | return !sc->target_mem_cgroup; |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 167 | } |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 168 | |
| 169 | /** |
| 170 | * sane_reclaim - is the usual dirty throttling mechanism operational? |
| 171 | * @sc: scan_control in question |
| 172 | * |
| 173 | * The normal page dirty throttling mechanism in balance_dirty_pages() is |
| 174 | * completely broken with the legacy memcg and direct stalling in |
| 175 | * shrink_page_list() is used for throttling instead, which lacks all the |
| 176 | * niceties such as fairness, adaptive pausing, bandwidth proportional |
| 177 | * allocation and configurability. |
| 178 | * |
| 179 | * This function tests whether the vmscan currently in progress can assume |
| 180 | * that the normal dirty throttling mechanism is operational. |
| 181 | */ |
| 182 | static bool sane_reclaim(struct scan_control *sc) |
| 183 | { |
| 184 | struct mem_cgroup *memcg = sc->target_mem_cgroup; |
| 185 | |
| 186 | if (!memcg) |
| 187 | return true; |
| 188 | #ifdef CONFIG_CGROUP_WRITEBACK |
Linus Torvalds | 69234ac | 2015-11-05 14:51:32 -0800 | [diff] [blame] | 189 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 190 | return true; |
| 191 | #endif |
| 192 | return false; |
| 193 | } |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 194 | #else |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 195 | static bool global_reclaim(struct scan_control *sc) |
| 196 | { |
| 197 | return true; |
| 198 | } |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 199 | |
| 200 | static bool sane_reclaim(struct scan_control *sc) |
| 201 | { |
| 202 | return true; |
| 203 | } |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 204 | #endif |
| 205 | |
Mel Gorman | 5a1c84b | 2016-07-28 15:47:31 -0700 | [diff] [blame] | 206 | /* |
| 207 | * This misses isolated pages which are not accounted for to save counters. |
| 208 | * As the data only determines if reclaim or compaction continues, it is |
| 209 | * not expected that isolated pages will be a dominating factor. |
| 210 | */ |
| 211 | unsigned long zone_reclaimable_pages(struct zone *zone) |
| 212 | { |
| 213 | unsigned long nr; |
| 214 | |
| 215 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + |
| 216 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); |
Vinayak Menon | 3771b88 | 2018-02-09 18:24:58 +0530 | [diff] [blame] | 217 | if (get_nr_swap_pages() > 0 |
| 218 | || IS_ENABLED(CONFIG_ANDROID_LOW_MEMORY_KILLER)) |
Mel Gorman | 5a1c84b | 2016-07-28 15:47:31 -0700 | [diff] [blame] | 219 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
| 220 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); |
| 221 | |
| 222 | return nr; |
| 223 | } |
| 224 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 225 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
Lisa Du | 6e543d5 | 2013-09-11 14:22:36 -0700 | [diff] [blame] | 226 | { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 227 | unsigned long nr; |
| 228 | |
| 229 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + |
| 230 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + |
| 231 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); |
| 232 | |
| 233 | if (get_nr_swap_pages() > 0) |
| 234 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
| 235 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + |
| 236 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); |
| 237 | |
| 238 | return nr; |
| 239 | } |
| 240 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 241 | /** |
| 242 | * lruvec_lru_size - Returns the number of pages on the given LRU list. |
| 243 | * @lruvec: lru vector |
| 244 | * @lru: lru to use |
| 245 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) |
| 246 | */ |
| 247 | 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] | 248 | { |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 249 | unsigned long lru_size; |
| 250 | int zid; |
| 251 | |
Hugh Dickins | c3c787e | 2012-05-29 15:06:52 -0700 | [diff] [blame] | 252 | if (!mem_cgroup_disabled()) |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 253 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
| 254 | else |
| 255 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); |
KOSAKI Motohiro | a3d8e05 | 2009-01-07 18:08:19 -0800 | [diff] [blame] | 256 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 257 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
| 258 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
| 259 | unsigned long size; |
KOSAKI Motohiro | c9f299d | 2009-01-07 18:08:16 -0800 | [diff] [blame] | 260 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 261 | if (!managed_zone(zone)) |
| 262 | continue; |
Michal Hocko | 07fc957 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 263 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 264 | if (!mem_cgroup_disabled()) |
| 265 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
| 266 | else |
| 267 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], |
| 268 | NR_ZONE_LRU_BASE + lru); |
| 269 | lru_size -= min(size, lru_size); |
| 270 | } |
| 271 | |
| 272 | return lru_size; |
| 273 | |
Michal Hocko | 07fc957 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 274 | } |
| 275 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 276 | /* |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 277 | * Add a shrinker callback to be called from the vm. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | */ |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 279 | int register_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 280 | { |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 281 | size_t size = sizeof(*shrinker->nr_deferred); |
| 282 | |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 283 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
| 284 | size *= nr_node_ids; |
| 285 | |
| 286 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); |
| 287 | if (!shrinker->nr_deferred) |
| 288 | return -ENOMEM; |
| 289 | |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 290 | down_write(&shrinker_rwsem); |
| 291 | list_add_tail(&shrinker->list, &shrinker_list); |
| 292 | up_write(&shrinker_rwsem); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 293 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 294 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 295 | EXPORT_SYMBOL(register_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 296 | |
| 297 | /* |
| 298 | * Remove one |
| 299 | */ |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 300 | void unregister_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | { |
Tetsuo Handa | 274ee93 | 2017-12-18 20:31:41 +0900 | [diff] [blame] | 302 | if (!shrinker->nr_deferred) |
| 303 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 304 | down_write(&shrinker_rwsem); |
| 305 | list_del(&shrinker->list); |
| 306 | up_write(&shrinker_rwsem); |
Andrew Vagin | ae39332 | 2013-10-16 13:46:46 -0700 | [diff] [blame] | 307 | kfree(shrinker->nr_deferred); |
Tetsuo Handa | 274ee93 | 2017-12-18 20:31:41 +0900 | [diff] [blame] | 308 | shrinker->nr_deferred = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 309 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 310 | EXPORT_SYMBOL(unregister_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 311 | |
| 312 | #define SHRINK_BATCH 128 |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 313 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 314 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
| 315 | struct shrinker *shrinker, |
| 316 | unsigned long nr_scanned, |
| 317 | unsigned long nr_eligible) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 318 | { |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 319 | unsigned long freed = 0; |
| 320 | unsigned long long delta; |
| 321 | long total_scan; |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 322 | long freeable; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 323 | long nr; |
| 324 | long new_nr; |
| 325 | int nid = shrinkctl->nid; |
| 326 | long batch_size = shrinker->batch ? shrinker->batch |
| 327 | : SHRINK_BATCH; |
Shaohua Li | 04597be | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 328 | long scanned = 0, next_deferred; |
Liam Mark | a708a02 | 2014-11-25 14:43:42 -0800 | [diff] [blame] | 329 | long min_cache_size = batch_size; |
| 330 | |
| 331 | if (current_is_kswapd()) |
| 332 | min_cache_size = 0; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 333 | |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 334 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
| 335 | if (freeable == 0) |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 336 | return 0; |
| 337 | |
| 338 | /* |
| 339 | * copy the current shrinker scan count into a local variable |
| 340 | * and zero it so that other concurrent shrinker invocations |
| 341 | * don't also do this scanning work. |
| 342 | */ |
| 343 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); |
| 344 | |
| 345 | total_scan = nr; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 346 | delta = (4 * nr_scanned) / shrinker->seeks; |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 347 | delta *= freeable; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 348 | do_div(delta, nr_eligible + 1); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 349 | total_scan += delta; |
| 350 | if (total_scan < 0) { |
Pintu Kumar | 8612c66 | 2014-12-10 15:42:58 -0800 | [diff] [blame] | 351 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 352 | shrinker->scan_objects, total_scan); |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 353 | total_scan = freeable; |
Shaohua Li | 04597be | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 354 | next_deferred = nr; |
| 355 | } else |
| 356 | next_deferred = total_scan; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 357 | |
| 358 | /* |
| 359 | * We need to avoid excessive windup on filesystem shrinkers |
| 360 | * due to large numbers of GFP_NOFS allocations causing the |
| 361 | * shrinkers to return -1 all the time. This results in a large |
| 362 | * nr being built up so when a shrink that can do some work |
| 363 | * comes along it empties the entire cache due to nr >>> |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 364 | * freeable. This is bad for sustaining a working set in |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 365 | * memory. |
| 366 | * |
| 367 | * Hence only allow the shrinker to scan the entire cache when |
| 368 | * a large delta change is calculated directly. |
| 369 | */ |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 370 | if (delta < freeable / 4) |
| 371 | total_scan = min(total_scan, freeable / 2); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 372 | |
| 373 | /* |
| 374 | * Avoid risking looping forever due to too large nr value: |
| 375 | * never try to free more than twice the estimate number of |
| 376 | * freeable entries. |
| 377 | */ |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 378 | if (total_scan > freeable * 2) |
| 379 | total_scan = freeable * 2; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 380 | |
| 381 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 382 | nr_scanned, nr_eligible, |
| 383 | freeable, delta, total_scan); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 384 | |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 385 | /* |
| 386 | * Normally, we should not scan less than batch_size objects in one |
| 387 | * pass to avoid too frequent shrinker calls, but if the slab has less |
| 388 | * than batch_size objects in total and we are really tight on memory, |
| 389 | * we will try to reclaim all available objects, otherwise we can end |
| 390 | * up failing allocations although there are plenty of reclaimable |
| 391 | * objects spread over several slabs with usage less than the |
| 392 | * batch_size. |
| 393 | * |
| 394 | * We detect the "tight on memory" situations by looking at the total |
| 395 | * 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] | 396 | * than the total number of objects on slab (freeable), we must be |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 397 | * scanning at high prio and therefore should try to reclaim as much as |
| 398 | * possible. |
| 399 | */ |
Liam Mark | a708a02 | 2014-11-25 14:43:42 -0800 | [diff] [blame] | 400 | while (total_scan > min_cache_size || |
Vladimir Davydov | d5bc5fd | 2014-04-03 14:47:32 -0700 | [diff] [blame] | 401 | total_scan >= freeable) { |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 402 | unsigned long ret; |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 403 | unsigned long nr_to_scan = min(batch_size, total_scan); |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 404 | |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 405 | shrinkctl->nr_to_scan = nr_to_scan; |
Dave Chinner | a0b0213 | 2013-08-28 10:18:16 +1000 | [diff] [blame] | 406 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
| 407 | if (ret == SHRINK_STOP) |
| 408 | break; |
| 409 | freed += ret; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 410 | |
Vladimir Davydov | 0b1fb40 | 2014-01-23 15:53:22 -0800 | [diff] [blame] | 411 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
| 412 | total_scan -= nr_to_scan; |
Shaohua Li | 04597be | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 413 | scanned += nr_to_scan; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 414 | |
| 415 | cond_resched(); |
| 416 | } |
| 417 | |
Shaohua Li | 04597be | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 418 | if (next_deferred >= scanned) |
| 419 | next_deferred -= scanned; |
| 420 | else |
| 421 | next_deferred = 0; |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 422 | /* |
| 423 | * move the unused scan count back into the shrinker in a |
| 424 | * manner that handles concurrent updates. If we exhausted the |
| 425 | * scan, there is no need to do an update. |
| 426 | */ |
Shaohua Li | 04597be | 2016-12-12 16:41:50 -0800 | [diff] [blame] | 427 | if (next_deferred > 0) |
| 428 | new_nr = atomic_long_add_return(next_deferred, |
Glauber Costa | 1d3d443 | 2013-08-28 10:18:04 +1000 | [diff] [blame] | 429 | &shrinker->nr_deferred[nid]); |
| 430 | else |
| 431 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); |
| 432 | |
Dave Hansen | df9024a | 2014-06-04 16:08:07 -0700 | [diff] [blame] | 433 | 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] | 434 | return freed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 435 | } |
| 436 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 437 | /** |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 438 | * shrink_slab - shrink slab caches |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 439 | * @gfp_mask: allocation context |
| 440 | * @nid: node whose slab caches to target |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 441 | * @memcg: memory cgroup whose slab caches to target |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 442 | * @nr_scanned: pressure numerator |
| 443 | * @nr_eligible: pressure denominator |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 444 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 445 | * Call the shrink functions to age shrinkable caches. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 446 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 447 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
| 448 | * unaware shrinkers will receive a node id of 0 instead. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 449 | * |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 450 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
| 451 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 452 | * objects from the memory cgroup specified. Otherwise, only unaware |
| 453 | * shrinkers are called. |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 454 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 455 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
| 456 | * the available objects should be scanned. Page reclaim for example |
| 457 | * passes the number of pages scanned and the number of pages on the |
| 458 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned |
| 459 | * when it encountered mapped pages. The ratio is further biased by |
| 460 | * the ->seeks setting of the shrink function, which indicates the |
| 461 | * cost to recreate an object relative to that of an LRU page. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 462 | * |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 463 | * Returns the number of reclaimed slab objects. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 464 | */ |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 465 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
| 466 | struct mem_cgroup *memcg, |
| 467 | unsigned long nr_scanned, |
| 468 | unsigned long nr_eligible) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 469 | { |
| 470 | struct shrinker *shrinker; |
Dave Chinner | 24f7c6b | 2013-08-28 10:17:56 +1000 | [diff] [blame] | 471 | unsigned long freed = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 472 | |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 473 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 474 | return 0; |
| 475 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 476 | if (nr_scanned == 0) |
| 477 | nr_scanned = SWAP_CLUSTER_MAX; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 478 | |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 479 | if (!down_read_trylock(&shrinker_rwsem)) { |
Dave Chinner | 24f7c6b | 2013-08-28 10:17:56 +1000 | [diff] [blame] | 480 | /* |
| 481 | * If we would return 0, our callers would understand that we |
| 482 | * have nothing else to shrink and give up trying. By returning |
| 483 | * 1 we keep it going and assume we'll be able to shrink next |
| 484 | * time. |
| 485 | */ |
| 486 | freed = 1; |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 487 | goto out; |
| 488 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 489 | |
| 490 | list_for_each_entry(shrinker, &shrinker_list, list) { |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 491 | struct shrink_control sc = { |
| 492 | .gfp_mask = gfp_mask, |
| 493 | .nid = nid, |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 494 | .memcg = memcg, |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 495 | }; |
Vladimir Davydov | ec97097 | 2014-01-23 15:53:23 -0800 | [diff] [blame] | 496 | |
Vladimir Davydov | 0fc9f58 | 2016-03-17 14:18:30 -0700 | [diff] [blame] | 497 | /* |
| 498 | * If kernel memory accounting is disabled, we ignore |
| 499 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers |
| 500 | * passing NULL for memcg. |
| 501 | */ |
| 502 | if (memcg_kmem_enabled() && |
| 503 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 504 | continue; |
| 505 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 506 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
| 507 | sc.nid = 0; |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 508 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 509 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 510 | } |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 511 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 512 | up_read(&shrinker_rwsem); |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 513 | out: |
| 514 | cond_resched(); |
Dave Chinner | 24f7c6b | 2013-08-28 10:17:56 +1000 | [diff] [blame] | 515 | return freed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 516 | } |
| 517 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 518 | void drop_slab_node(int nid) |
| 519 | { |
| 520 | unsigned long freed; |
| 521 | |
| 522 | do { |
| 523 | struct mem_cgroup *memcg = NULL; |
| 524 | |
| 525 | freed = 0; |
| 526 | do { |
| 527 | freed += shrink_slab(GFP_KERNEL, nid, memcg, |
| 528 | 1000, 1000); |
| 529 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
| 530 | } while (freed > 10); |
| 531 | } |
| 532 | |
| 533 | void drop_slab(void) |
| 534 | { |
| 535 | int nid; |
| 536 | |
| 537 | for_each_online_node(nid) |
| 538 | drop_slab_node(nid); |
| 539 | } |
| 540 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 541 | static inline int is_page_cache_freeable(struct page *page) |
| 542 | { |
Johannes Weiner | ceddc3a | 2009-09-21 17:03:00 -0700 | [diff] [blame] | 543 | /* |
| 544 | * A freeable page cache page is referenced only by the caller |
| 545 | * that isolated the page, the page cache radix tree and |
| 546 | * optional buffer heads at page->private. |
| 547 | */ |
Johannes Weiner | edcf474 | 2009-09-21 17:02:59 -0700 | [diff] [blame] | 548 | return page_count(page) - page_has_private(page) == 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 549 | } |
| 550 | |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 551 | 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] | 552 | { |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 553 | if (current->flags & PF_SWAPWRITE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 554 | return 1; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 555 | if (!inode_write_congested(inode)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 556 | return 1; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 557 | if (inode_to_bdi(inode) == current->backing_dev_info) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 558 | return 1; |
| 559 | return 0; |
| 560 | } |
| 561 | |
| 562 | /* |
| 563 | * We detected a synchronous write error writing a page out. Probably |
| 564 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
| 565 | * fsync(), msync() or close(). |
| 566 | * |
| 567 | * The tricky part is that after writepage we cannot touch the mapping: nothing |
| 568 | * prevents it from being freed up. But we have a ref on the page and once |
| 569 | * that page is locked, the mapping is pinned. |
| 570 | * |
| 571 | * We're allowed to run sleeping lock_page() here because we know the caller has |
| 572 | * __GFP_FS. |
| 573 | */ |
| 574 | static void handle_write_error(struct address_space *mapping, |
| 575 | struct page *page, int error) |
| 576 | { |
Jens Axboe | 7eaceac | 2011-03-10 08:52:07 +0100 | [diff] [blame] | 577 | lock_page(page); |
Guillaume Chazarain | 3e9f45b | 2007-05-08 00:23:25 -0700 | [diff] [blame] | 578 | if (page_mapping(page) == mapping) |
| 579 | mapping_set_error(mapping, error); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 580 | unlock_page(page); |
| 581 | } |
| 582 | |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 583 | /* possible outcome of pageout() */ |
| 584 | typedef enum { |
| 585 | /* failed to write page out, page is locked */ |
| 586 | PAGE_KEEP, |
| 587 | /* move page to the active list, page is locked */ |
| 588 | PAGE_ACTIVATE, |
| 589 | /* page has been sent to the disk successfully, page is unlocked */ |
| 590 | PAGE_SUCCESS, |
| 591 | /* page is clean and locked */ |
| 592 | PAGE_CLEAN, |
| 593 | } pageout_t; |
| 594 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 595 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 596 | * pageout is called by shrink_page_list() for each dirty page. |
| 597 | * Calls ->writepage(). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 598 | */ |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 599 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 600 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 601 | { |
| 602 | /* |
| 603 | * If the page is dirty, only perform writeback if that write |
| 604 | * will be non-blocking. To prevent this allocation from being |
| 605 | * stalled by pagecache activity. But note that there may be |
| 606 | * stalls if we need to run get_block(). We could test |
| 607 | * PagePrivate for that. |
| 608 | * |
Al Viro | 8174202 | 2014-04-03 03:17:43 -0400 | [diff] [blame] | 609 | * If this process is currently in __generic_file_write_iter() against |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | * this page's queue, we can perform writeback even if that |
| 611 | * will block. |
| 612 | * |
| 613 | * If the page is swapcache, write it back even if that would |
| 614 | * block, for some throttling. This happens by accident, because |
| 615 | * swap_backing_dev_info is bust: it doesn't reflect the |
| 616 | * congestion state of the swapdevs. Easy to fix, if needed. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 617 | */ |
| 618 | if (!is_page_cache_freeable(page)) |
| 619 | return PAGE_KEEP; |
| 620 | if (!mapping) { |
| 621 | /* |
| 622 | * Some data journaling orphaned pages can have |
| 623 | * page->mapping == NULL while being dirty with clean buffers. |
| 624 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 625 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 626 | if (try_to_free_buffers(page)) { |
| 627 | ClearPageDirty(page); |
Mitchel Humpherys | b1de0d1 | 2014-06-06 14:38:30 -0700 | [diff] [blame] | 628 | pr_info("%s: orphaned page\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 629 | return PAGE_CLEAN; |
| 630 | } |
| 631 | } |
| 632 | return PAGE_KEEP; |
| 633 | } |
| 634 | if (mapping->a_ops->writepage == NULL) |
| 635 | return PAGE_ACTIVATE; |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 636 | if (!may_write_to_inode(mapping->host, sc)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 637 | return PAGE_KEEP; |
| 638 | |
| 639 | if (clear_page_dirty_for_io(page)) { |
| 640 | int res; |
| 641 | struct writeback_control wbc = { |
| 642 | .sync_mode = WB_SYNC_NONE, |
| 643 | .nr_to_write = SWAP_CLUSTER_MAX, |
OGAWA Hirofumi | 111ebb6 | 2006-06-23 02:03:26 -0700 | [diff] [blame] | 644 | .range_start = 0, |
| 645 | .range_end = LLONG_MAX, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 646 | .for_reclaim = 1, |
| 647 | }; |
| 648 | |
| 649 | SetPageReclaim(page); |
| 650 | res = mapping->a_ops->writepage(page, &wbc); |
| 651 | if (res < 0) |
| 652 | handle_write_error(mapping, page, res); |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 653 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 654 | ClearPageReclaim(page); |
| 655 | return PAGE_ACTIVATE; |
| 656 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 657 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 658 | if (!PageWriteback(page)) { |
| 659 | /* synchronous write or broken a_ops? */ |
| 660 | ClearPageReclaim(page); |
| 661 | } |
yalin wang | 3aa2385 | 2016-01-14 15:18:30 -0800 | [diff] [blame] | 662 | trace_mm_vmscan_writepage(page); |
Mel Gorman | c4a2563 | 2016-07-28 15:46:23 -0700 | [diff] [blame] | 663 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 664 | return PAGE_SUCCESS; |
| 665 | } |
| 666 | |
| 667 | return PAGE_CLEAN; |
| 668 | } |
| 669 | |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 670 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 671 | * Same as remove_mapping, but if the page is removed from the mapping, it |
| 672 | * gets returned with a refcount of 0. |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 673 | */ |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 674 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
| 675 | bool reclaimed) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 676 | { |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 677 | unsigned long flags; |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 678 | |
Nick Piggin | 28e4d96 | 2006-09-25 23:31:23 -0700 | [diff] [blame] | 679 | BUG_ON(!PageLocked(page)); |
| 680 | BUG_ON(mapping != page_mapping(page)); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 681 | |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 682 | spin_lock_irqsave(&mapping->tree_lock, flags); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 683 | /* |
Nick Piggin | 0fd0e6b | 2006-09-27 01:50:02 -0700 | [diff] [blame] | 684 | * The non racy check for a busy page. |
| 685 | * |
| 686 | * Must be careful with the order of the tests. When someone has |
| 687 | * a ref to the page, it may be possible that they dirty it then |
| 688 | * drop the reference. So if PageDirty is tested before page_count |
| 689 | * here, then the following race may occur: |
| 690 | * |
| 691 | * get_user_pages(&page); |
| 692 | * [user mapping goes away] |
| 693 | * write_to(page); |
| 694 | * !PageDirty(page) [good] |
| 695 | * SetPageDirty(page); |
| 696 | * put_page(page); |
| 697 | * !page_count(page) [good, discard it] |
| 698 | * |
| 699 | * [oops, our write_to data is lost] |
| 700 | * |
| 701 | * Reversing the order of the tests ensures such a situation cannot |
| 702 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags |
Joonsoo Kim | 0139aa7 | 2016-05-19 17:10:49 -0700 | [diff] [blame] | 703 | * load is not satisfied before that of page->_refcount. |
Nick Piggin | 0fd0e6b | 2006-09-27 01:50:02 -0700 | [diff] [blame] | 704 | * |
| 705 | * Note that if SetPageDirty is always performed via set_page_dirty, |
| 706 | * and thus under tree_lock, then this ordering is not required. |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 707 | */ |
Joonsoo Kim | fe896d1 | 2016-03-17 14:19:26 -0700 | [diff] [blame] | 708 | if (!page_ref_freeze(page, 2)) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 709 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 710 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
| 711 | if (unlikely(PageDirty(page))) { |
Joonsoo Kim | fe896d1 | 2016-03-17 14:19:26 -0700 | [diff] [blame] | 712 | page_ref_unfreeze(page, 2); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 713 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 714 | } |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 715 | |
| 716 | if (PageSwapCache(page)) { |
| 717 | swp_entry_t swap = { .val = page_private(page) }; |
Johannes Weiner | 0a31bc9 | 2014-08-08 14:19:22 -0700 | [diff] [blame] | 718 | mem_cgroup_swapout(page, swap); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 719 | __delete_from_swap_cache(page); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 720 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
Johannes Weiner | 0a31bc9 | 2014-08-08 14:19:22 -0700 | [diff] [blame] | 721 | swapcache_free(swap); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 722 | } else { |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 723 | void (*freepage)(struct page *); |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 724 | void *shadow = NULL; |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 725 | |
| 726 | freepage = mapping->a_ops->freepage; |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 727 | /* |
| 728 | * Remember a shadow entry for reclaimed file cache in |
| 729 | * order to detect refaults, thus thrashing, later on. |
| 730 | * |
| 731 | * But don't store shadows in an address space that is |
| 732 | * already exiting. This is not just an optizimation, |
| 733 | * inode reclaim needs to empty out the radix tree or |
| 734 | * the nodes are lost. Don't plant shadows behind its |
| 735 | * back. |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 736 | * |
| 737 | * We also don't store shadows for DAX mappings because the |
| 738 | * only page cache pages found in these are zero pages |
| 739 | * covering holes, and because we don't want to mix DAX |
| 740 | * exceptional entries and shadow exceptional entries in the |
| 741 | * same page_tree. |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 742 | */ |
| 743 | if (reclaimed && page_is_file_cache(page) && |
Ross Zwisler | f9fe48b | 2016-01-22 15:10:40 -0800 | [diff] [blame] | 744 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 745 | shadow = workingset_eviction(mapping, page); |
Johannes Weiner | 62cccb8 | 2016-03-15 14:57:22 -0700 | [diff] [blame] | 746 | __delete_from_page_cache(page, shadow); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 747 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 748 | |
| 749 | if (freepage != NULL) |
| 750 | freepage(page); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 751 | } |
| 752 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 753 | return 1; |
| 754 | |
| 755 | cannot_free: |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 756 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 757 | return 0; |
| 758 | } |
| 759 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 760 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 761 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if |
| 762 | * someone else has a ref on the page, abort and return 0. If it was |
| 763 | * successfully detached, return 1. Assumes the caller has a single ref on |
| 764 | * this page. |
| 765 | */ |
| 766 | int remove_mapping(struct address_space *mapping, struct page *page) |
| 767 | { |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 768 | if (__remove_mapping(mapping, page, false)) { |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 769 | /* |
| 770 | * Unfreezing the refcount with 1 rather than 2 effectively |
| 771 | * drops the pagecache ref for us without requiring another |
| 772 | * atomic operation. |
| 773 | */ |
Joonsoo Kim | fe896d1 | 2016-03-17 14:19:26 -0700 | [diff] [blame] | 774 | page_ref_unfreeze(page, 1); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 775 | return 1; |
| 776 | } |
| 777 | return 0; |
| 778 | } |
| 779 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 780 | /** |
| 781 | * putback_lru_page - put previously isolated page onto appropriate LRU list |
| 782 | * @page: page to be put back to appropriate lru list |
| 783 | * |
| 784 | * Add previously isolated @page to appropriate LRU list. |
| 785 | * Page may still be unevictable for other reasons. |
| 786 | * |
| 787 | * lru_lock must not be held, interrupts must be enabled. |
| 788 | */ |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 789 | void putback_lru_page(struct page *page) |
| 790 | { |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 791 | bool is_unevictable; |
Lee Schermerhorn | bbfd28e | 2008-10-18 20:26:40 -0700 | [diff] [blame] | 792 | int was_unevictable = PageUnevictable(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 793 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 794 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 795 | |
| 796 | redo: |
| 797 | ClearPageUnevictable(page); |
| 798 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 799 | if (page_evictable(page)) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 800 | /* |
| 801 | * For evictable pages, we can use the cache. |
| 802 | * In event of a race, worst case is we end up with an |
| 803 | * unevictable page on [in]active list. |
| 804 | * We know how to handle that. |
| 805 | */ |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 806 | is_unevictable = false; |
Mel Gorman | c53954a | 2013-07-03 15:02:34 -0700 | [diff] [blame] | 807 | lru_cache_add(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 808 | } else { |
| 809 | /* |
| 810 | * Put unevictable pages directly on zone's unevictable |
| 811 | * list. |
| 812 | */ |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 813 | is_unevictable = true; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 814 | add_page_to_unevictable_list(page); |
Johannes Weiner | 6a7b954 | 2009-10-26 16:50:00 -0700 | [diff] [blame] | 815 | /* |
Minchan Kim | 21ee9f3 | 2011-10-31 17:09:28 -0700 | [diff] [blame] | 816 | * When racing with an mlock or AS_UNEVICTABLE clearing |
| 817 | * (page is unlocked) make sure that if the other thread |
| 818 | * does not observe our setting of PG_lru and fails |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 819 | * isolation/check_move_unevictable_pages, |
Minchan Kim | 21ee9f3 | 2011-10-31 17:09:28 -0700 | [diff] [blame] | 820 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
Johannes Weiner | 6a7b954 | 2009-10-26 16:50:00 -0700 | [diff] [blame] | 821 | * the page back to the evictable list. |
| 822 | * |
Minchan Kim | 21ee9f3 | 2011-10-31 17:09:28 -0700 | [diff] [blame] | 823 | * The other side is TestClearPageMlocked() or shmem_lock(). |
Johannes Weiner | 6a7b954 | 2009-10-26 16:50:00 -0700 | [diff] [blame] | 824 | */ |
| 825 | smp_mb(); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 826 | } |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 827 | |
| 828 | /* |
| 829 | * page's status can change while we move it among lru. If an evictable |
| 830 | * page is on unevictable list, it never be freed. To avoid that, |
| 831 | * check after we added it to the list, again. |
| 832 | */ |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 833 | if (is_unevictable && page_evictable(page)) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 834 | if (!isolate_lru_page(page)) { |
| 835 | put_page(page); |
| 836 | goto redo; |
| 837 | } |
| 838 | /* This means someone else dropped this page from LRU |
| 839 | * So, it will be freed or putback to LRU again. There is |
| 840 | * nothing to do here. |
| 841 | */ |
| 842 | } |
| 843 | |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 844 | if (was_unevictable && !is_unevictable) |
Lee Schermerhorn | bbfd28e | 2008-10-18 20:26:40 -0700 | [diff] [blame] | 845 | count_vm_event(UNEVICTABLE_PGRESCUED); |
Vlastimil Babka | 0ec3b74 | 2013-09-11 14:22:26 -0700 | [diff] [blame] | 846 | else if (!was_unevictable && is_unevictable) |
Lee Schermerhorn | bbfd28e | 2008-10-18 20:26:40 -0700 | [diff] [blame] | 847 | count_vm_event(UNEVICTABLE_PGCULLED); |
| 848 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 849 | put_page(page); /* drop ref from isolate */ |
| 850 | } |
| 851 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 852 | enum page_references { |
| 853 | PAGEREF_RECLAIM, |
| 854 | PAGEREF_RECLAIM_CLEAN, |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 855 | PAGEREF_KEEP, |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 856 | PAGEREF_ACTIVATE, |
| 857 | }; |
| 858 | |
| 859 | static enum page_references page_check_references(struct page *page, |
| 860 | struct scan_control *sc) |
| 861 | { |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 862 | int referenced_ptes, referenced_page; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 863 | unsigned long vm_flags; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 864 | |
Johannes Weiner | c3ac9a8 | 2012-05-29 15:06:25 -0700 | [diff] [blame] | 865 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
| 866 | &vm_flags); |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 867 | referenced_page = TestClearPageReferenced(page); |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 868 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 869 | /* |
| 870 | * Mlock lost the isolation race with us. Let try_to_unmap() |
| 871 | * move the page to the unevictable list. |
| 872 | */ |
| 873 | if (vm_flags & VM_LOCKED) |
| 874 | return PAGEREF_RECLAIM; |
| 875 | |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 876 | if (referenced_ptes) { |
Michal Hocko | e4898273 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 877 | if (PageSwapBacked(page)) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 878 | return PAGEREF_ACTIVATE; |
| 879 | /* |
| 880 | * All mapped pages start out with page table |
| 881 | * references from the instantiating fault, so we need |
| 882 | * to look twice if a mapped file page is used more |
| 883 | * than once. |
| 884 | * |
| 885 | * Mark it and spare it for another trip around the |
| 886 | * inactive list. Another page table reference will |
| 887 | * lead to its activation. |
| 888 | * |
| 889 | * Note: the mark is set for activated pages as well |
| 890 | * so that recently deactivated but used pages are |
| 891 | * quickly recovered. |
| 892 | */ |
| 893 | SetPageReferenced(page); |
| 894 | |
Konstantin Khlebnikov | 34dbc67 | 2012-01-10 15:06:59 -0800 | [diff] [blame] | 895 | if (referenced_page || referenced_ptes > 1) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 896 | return PAGEREF_ACTIVATE; |
| 897 | |
Konstantin Khlebnikov | c909e99 | 2012-01-10 15:07:03 -0800 | [diff] [blame] | 898 | /* |
| 899 | * Activate file-backed executable pages after first usage. |
| 900 | */ |
| 901 | if (vm_flags & VM_EXEC) |
| 902 | return PAGEREF_ACTIVATE; |
| 903 | |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 904 | return PAGEREF_KEEP; |
| 905 | } |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 906 | |
| 907 | /* Reclaim if clean, defer dirty pages to writeback */ |
KOSAKI Motohiro | 2e30244 | 2010-10-26 14:21:46 -0700 | [diff] [blame] | 908 | if (referenced_page && !PageSwapBacked(page)) |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 909 | return PAGEREF_RECLAIM_CLEAN; |
| 910 | |
| 911 | return PAGEREF_RECLAIM; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 912 | } |
| 913 | |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 914 | /* Check if a page is dirty or under writeback */ |
| 915 | static void page_check_dirty_writeback(struct page *page, |
| 916 | bool *dirty, bool *writeback) |
| 917 | { |
Mel Gorman | b459722 | 2013-07-03 15:02:05 -0700 | [diff] [blame] | 918 | struct address_space *mapping; |
| 919 | |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 920 | /* |
| 921 | * Anonymous pages are not handled by flushers and must be written |
| 922 | * from reclaim context. Do not stall reclaim based on them |
| 923 | */ |
| 924 | if (!page_is_file_cache(page)) { |
| 925 | *dirty = false; |
| 926 | *writeback = false; |
| 927 | return; |
| 928 | } |
| 929 | |
| 930 | /* By default assume that the page flags are accurate */ |
| 931 | *dirty = PageDirty(page); |
| 932 | *writeback = PageWriteback(page); |
Mel Gorman | b459722 | 2013-07-03 15:02:05 -0700 | [diff] [blame] | 933 | |
| 934 | /* Verify dirty/writeback state if the filesystem supports it */ |
| 935 | if (!page_has_private(page)) |
| 936 | return; |
| 937 | |
| 938 | mapping = page_mapping(page); |
| 939 | if (mapping && mapping->a_ops->is_dirty_writeback) |
| 940 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 941 | } |
| 942 | |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 943 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 944 | * shrink_page_list() returns the number of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 945 | */ |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 946 | static unsigned long shrink_page_list(struct list_head *page_list, |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 947 | struct pglist_data *pgdat, |
Mel Gorman | f84f6e2 | 2011-10-31 17:07:51 -0700 | [diff] [blame] | 948 | struct scan_control *sc, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 949 | enum ttu_flags ttu_flags, |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 950 | unsigned long *ret_nr_dirty, |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 951 | unsigned long *ret_nr_unqueued_dirty, |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 952 | unsigned long *ret_nr_congested, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 953 | unsigned long *ret_nr_writeback, |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 954 | unsigned long *ret_nr_immediate, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 955 | bool force_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 956 | { |
| 957 | LIST_HEAD(ret_pages); |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 958 | LIST_HEAD(free_pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 959 | int pgactivate = 0; |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 960 | unsigned long nr_unqueued_dirty = 0; |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 961 | unsigned long nr_dirty = 0; |
| 962 | unsigned long nr_congested = 0; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 963 | unsigned long nr_reclaimed = 0; |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 964 | unsigned long nr_writeback = 0; |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 965 | unsigned long nr_immediate = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 966 | |
| 967 | cond_resched(); |
| 968 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 969 | while (!list_empty(page_list)) { |
| 970 | struct address_space *mapping; |
| 971 | struct page *page; |
| 972 | int may_enter_fs; |
Minchan Kim | 13e7b91 | 2015-01-09 18:36:55 +0530 | [diff] [blame] | 973 | enum page_references references = PAGEREF_RECLAIM; |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 974 | bool dirty, writeback; |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 975 | bool lazyfree = false; |
| 976 | int ret = SWAP_SUCCESS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 977 | |
| 978 | cond_resched(); |
| 979 | |
| 980 | page = lru_to_page(page_list); |
| 981 | list_del(&page->lru); |
| 982 | |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 983 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 984 | goto keep; |
| 985 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 986 | VM_BUG_ON_PAGE(PageActive(page), page); |
Minchan Kim | 07e6c8d | 2013-05-09 16:21:25 +0900 | [diff] [blame] | 987 | if (pgdat) |
| 988 | VM_BUG_ON_PAGE(page_pgdat(page) != pgdat, page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 989 | |
| 990 | sc->nr_scanned++; |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 991 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 992 | if (unlikely(!page_evictable(page))) |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 993 | goto cull_mlocked; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 994 | |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 995 | if (!sc->may_unmap && page_mapped(page)) |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 996 | goto keep_locked; |
| 997 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 998 | /* Double the slab pressure for mapped and swapcache pages */ |
| 999 | if (page_mapped(page) || PageSwapCache(page)) |
| 1000 | sc->nr_scanned++; |
| 1001 | |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 1002 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
| 1003 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); |
| 1004 | |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1005 | /* |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1006 | * The number of dirty pages determines if a zone is marked |
| 1007 | * reclaim_congested which affects wait_iff_congested. kswapd |
| 1008 | * will stall and start writing pages if the tail of the LRU |
| 1009 | * is all dirty unqueued pages. |
| 1010 | */ |
| 1011 | page_check_dirty_writeback(page, &dirty, &writeback); |
| 1012 | if (dirty || writeback) |
| 1013 | nr_dirty++; |
| 1014 | |
| 1015 | if (dirty && !writeback) |
| 1016 | nr_unqueued_dirty++; |
| 1017 | |
Mel Gorman | d04e8ac | 2013-07-03 15:02:03 -0700 | [diff] [blame] | 1018 | /* |
| 1019 | * Treat this page as congested if the underlying BDI is or if |
| 1020 | * pages are cycling through the LRU so quickly that the |
| 1021 | * pages marked for immediate reclaim are making it to the |
| 1022 | * end of the LRU a second time. |
| 1023 | */ |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1024 | mapping = page_mapping(page); |
Jamie Liu | 1da58ee | 2014-12-10 15:43:20 -0800 | [diff] [blame] | 1025 | if (((dirty || writeback) && mapping && |
Tejun Heo | 703c270 | 2015-05-22 17:13:44 -0400 | [diff] [blame] | 1026 | inode_write_congested(mapping->host)) || |
Mel Gorman | d04e8ac | 2013-07-03 15:02:03 -0700 | [diff] [blame] | 1027 | (writeback && PageReclaim(page))) |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1028 | nr_congested++; |
| 1029 | |
| 1030 | /* |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1031 | * If a page at the tail of the LRU is under writeback, there |
| 1032 | * are three cases to consider. |
| 1033 | * |
| 1034 | * 1) If reclaim is encountering an excessive number of pages |
| 1035 | * under writeback and this page is both under writeback and |
| 1036 | * PageReclaim then it indicates that pages are being queued |
| 1037 | * for IO but are being recycled through the LRU before the |
| 1038 | * IO can complete. Waiting on the page itself risks an |
| 1039 | * indefinite stall if it is impossible to writeback the |
| 1040 | * page due to IO error or disconnected storage so instead |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1041 | * note that the LRU is being scanned too quickly and the |
| 1042 | * caller can stall after page list has been processed. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1043 | * |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1044 | * 2) Global or new memcg reclaim encounters a page that is |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1045 | * not marked for immediate reclaim, or the caller does not |
| 1046 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, |
| 1047 | * not to fs). In this case mark the page for immediate |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1048 | * reclaim and continue scanning. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1049 | * |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1050 | * Require may_enter_fs because we would wait on fs, which |
| 1051 | * may not have submitted IO yet. And the loop driver might |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1052 | * enter reclaim, and deadlock if it waits on a page for |
| 1053 | * which it is needed to do the write (loop masks off |
| 1054 | * __GFP_IO|__GFP_FS for this reason); but more thought |
| 1055 | * would probably show more reasons. |
| 1056 | * |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1057 | * 3) Legacy memcg encounters a page that is already marked |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1058 | * PageReclaim. memcg does not have any dirty pages |
| 1059 | * throttling so we could easily OOM just because too many |
| 1060 | * pages are in writeback and there is nothing else to |
| 1061 | * reclaim. Wait for the writeback to complete. |
Johannes Weiner | 6b25f41 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1062 | * |
| 1063 | * In cases 1) and 2) we activate the pages to get them out of |
| 1064 | * the way while we continue scanning for clean pages on the |
| 1065 | * inactive list and refilling from the active list. The |
| 1066 | * observation here is that waiting for disk writes is more |
| 1067 | * expensive than potentially causing reloads down the line. |
| 1068 | * Since they're marked for immediate reclaim, they won't put |
| 1069 | * memory pressure on the cache working set any longer than it |
| 1070 | * takes to write them to disk. |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1071 | */ |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 1072 | if (PageWriteback(page)) { |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1073 | /* Case 1 above */ |
| 1074 | if (current_is_kswapd() && |
| 1075 | PageReclaim(page) && |
Minchan Kim | 07e6c8d | 2013-05-09 16:21:25 +0900 | [diff] [blame] | 1076 | (pgdat && test_bit(PGDAT_WRITEBACK, &pgdat->flags))) { |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1077 | nr_immediate++; |
Johannes Weiner | 6b25f41 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1078 | goto activate_locked; |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1079 | |
| 1080 | /* Case 2 above */ |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1081 | } else if (sane_reclaim(sc) || |
Michal Hocko | ecf5fc6 | 2015-08-04 14:36:58 -0700 | [diff] [blame] | 1082 | !PageReclaim(page) || !may_enter_fs) { |
Hugh Dickins | c3b94f4 | 2012-07-31 16:45:59 -0700 | [diff] [blame] | 1083 | /* |
| 1084 | * This is slightly racy - end_page_writeback() |
| 1085 | * might have just cleared PageReclaim, then |
| 1086 | * setting PageReclaim here end up interpreted |
| 1087 | * as PageReadahead - but that does not matter |
| 1088 | * enough to care. What we do want is for this |
| 1089 | * page to have PageReclaim set next time memcg |
| 1090 | * reclaim reaches the tests above, so it will |
| 1091 | * then wait_on_page_writeback() to avoid OOM; |
| 1092 | * and it's also appropriate in global reclaim. |
| 1093 | */ |
| 1094 | SetPageReclaim(page); |
Michal Hocko | e62e384 | 2012-07-31 16:45:55 -0700 | [diff] [blame] | 1095 | nr_writeback++; |
Johannes Weiner | 6b25f41 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1096 | goto activate_locked; |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1097 | |
| 1098 | /* Case 3 above */ |
| 1099 | } else { |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1100 | unlock_page(page); |
Mel Gorman | 283aba9 | 2013-07-03 15:01:51 -0700 | [diff] [blame] | 1101 | wait_on_page_writeback(page); |
Hugh Dickins | 7fadc82 | 2015-09-08 15:03:46 -0700 | [diff] [blame] | 1102 | /* then go back and try same page again */ |
| 1103 | list_add_tail(&page->lru, page_list); |
| 1104 | continue; |
Michal Hocko | e62e384 | 2012-07-31 16:45:55 -0700 | [diff] [blame] | 1105 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 1106 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1107 | |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1108 | if (!force_reclaim) |
| 1109 | references = page_check_references(page, sc); |
| 1110 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1111 | switch (references) { |
| 1112 | case PAGEREF_ACTIVATE: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1113 | goto activate_locked; |
Johannes Weiner | 64574746 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 1114 | case PAGEREF_KEEP: |
| 1115 | goto keep_locked; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1116 | case PAGEREF_RECLAIM: |
| 1117 | case PAGEREF_RECLAIM_CLEAN: |
| 1118 | ; /* try to reclaim the page below */ |
| 1119 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1120 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1121 | /* |
| 1122 | * Anonymous process memory has backing store? |
| 1123 | * Try to allocate it some swap space here. |
| 1124 | */ |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 1125 | if (PageAnon(page) && !PageSwapCache(page)) { |
Hugh Dickins | 63eb6b9 | 2008-11-19 15:36:37 -0800 | [diff] [blame] | 1126 | if (!(sc->gfp_mask & __GFP_IO)) |
| 1127 | goto keep_locked; |
Shaohua Li | 5bc7b8a | 2013-04-29 15:08:36 -0700 | [diff] [blame] | 1128 | if (!add_to_swap(page, page_list)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1129 | goto activate_locked; |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1130 | lazyfree = true; |
Hugh Dickins | 63eb6b9 | 2008-11-19 15:36:37 -0800 | [diff] [blame] | 1131 | may_enter_fs = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1132 | |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1133 | /* Adding to swap updated mapping */ |
| 1134 | mapping = page_mapping(page); |
Kirill A. Shutemov | 7751b2d | 2016-07-26 15:25:56 -0700 | [diff] [blame] | 1135 | } else if (unlikely(PageTransHuge(page))) { |
| 1136 | /* Split file THP */ |
| 1137 | if (split_huge_page_to_list(page, page_list)) |
| 1138 | goto keep_locked; |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1139 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1140 | |
Kirill A. Shutemov | 7751b2d | 2016-07-26 15:25:56 -0700 | [diff] [blame] | 1141 | VM_BUG_ON_PAGE(PageTransHuge(page), page); |
| 1142 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1143 | /* |
| 1144 | * The page is mapped into the page tables of one or more |
| 1145 | * processes. Try to unmap it here. |
| 1146 | */ |
| 1147 | if (page_mapped(page) && mapping) { |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1148 | switch (ret = try_to_unmap(page, lazyfree ? |
| 1149 | (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) : |
Minchan Kim | cd256bf | 2013-05-09 16:21:27 +0900 | [diff] [blame] | 1150 | (ttu_flags | TTU_BATCH_FLUSH), |
| 1151 | sc->target_vma)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1152 | case SWAP_FAIL: |
| 1153 | goto activate_locked; |
| 1154 | case SWAP_AGAIN: |
| 1155 | goto keep_locked; |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 1156 | case SWAP_MLOCK: |
| 1157 | goto cull_mlocked; |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1158 | case SWAP_LZFREE: |
| 1159 | goto lazyfree; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1160 | case SWAP_SUCCESS: |
| 1161 | ; /* try to free the page below */ |
| 1162 | } |
| 1163 | } |
| 1164 | |
| 1165 | if (PageDirty(page)) { |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1166 | /* |
Johannes Weiner | 319400b | 2017-02-24 14:56:20 -0800 | [diff] [blame] | 1167 | * Only kswapd can writeback filesystem pages |
| 1168 | * to avoid risk of stack overflow. But avoid |
| 1169 | * injecting inefficient single-page IO into |
| 1170 | * flusher writeback as much as possible: only |
| 1171 | * write pages when we've encountered many |
| 1172 | * dirty pages, and when we've already scanned |
| 1173 | * the rest of the LRU for clean pages and see |
| 1174 | * the same dirty pages again (PageReclaim). |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1175 | */ |
Mel Gorman | f84f6e2 | 2011-10-31 17:07:51 -0700 | [diff] [blame] | 1176 | if (page_is_file_cache(page) && |
Johannes Weiner | 319400b | 2017-02-24 14:56:20 -0800 | [diff] [blame] | 1177 | (!current_is_kswapd() || !PageReclaim(page) || |
| 1178 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
Mel Gorman | 49ea7eb | 2011-10-31 17:07:59 -0700 | [diff] [blame] | 1179 | /* |
| 1180 | * Immediately reclaim when written back. |
| 1181 | * Similar in principal to deactivate_page() |
| 1182 | * except we already have the page isolated |
| 1183 | * and know it's dirty |
| 1184 | */ |
Mel Gorman | c4a2563 | 2016-07-28 15:46:23 -0700 | [diff] [blame] | 1185 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
Mel Gorman | 49ea7eb | 2011-10-31 17:07:59 -0700 | [diff] [blame] | 1186 | SetPageReclaim(page); |
| 1187 | |
Johannes Weiner | 6b25f41 | 2017-02-24 14:56:23 -0800 | [diff] [blame] | 1188 | goto activate_locked; |
Mel Gorman | ee72886 | 2011-10-31 17:07:38 -0700 | [diff] [blame] | 1189 | } |
| 1190 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 1191 | if (references == PAGEREF_RECLAIM_CLEAN) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1192 | goto keep_locked; |
Andrew Morton | 4dd4b92 | 2008-03-24 12:29:52 -0700 | [diff] [blame] | 1193 | if (!may_enter_fs) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1194 | goto keep_locked; |
Christoph Lameter | 52a8363 | 2006-02-01 03:05:28 -0800 | [diff] [blame] | 1195 | if (!sc->may_writepage) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1196 | goto keep_locked; |
| 1197 | |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 1198 | /* |
| 1199 | * Page is dirty. Flush the TLB if a writable entry |
| 1200 | * potentially exists to avoid CPU writes after IO |
| 1201 | * starts and then write it out here. |
| 1202 | */ |
| 1203 | try_to_unmap_flush_dirty(); |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1204 | switch (pageout(page, mapping, sc)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1205 | case PAGE_KEEP: |
| 1206 | goto keep_locked; |
| 1207 | case PAGE_ACTIVATE: |
| 1208 | goto activate_locked; |
| 1209 | case PAGE_SUCCESS: |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1210 | if (PageWriteback(page)) |
Mel Gorman | 41ac199 | 2012-05-29 15:06:19 -0700 | [diff] [blame] | 1211 | goto keep; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1212 | if (PageDirty(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1213 | goto keep; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 1214 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1215 | /* |
| 1216 | * A synchronous write - probably a ramdisk. Go |
| 1217 | * ahead and try to reclaim the page. |
| 1218 | */ |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 1219 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1220 | goto keep; |
| 1221 | if (PageDirty(page) || PageWriteback(page)) |
| 1222 | goto keep_locked; |
| 1223 | mapping = page_mapping(page); |
| 1224 | case PAGE_CLEAN: |
| 1225 | ; /* try to free the page below */ |
| 1226 | } |
| 1227 | } |
| 1228 | |
| 1229 | /* |
| 1230 | * If the page has buffers, try to free the buffer mappings |
| 1231 | * associated with this page. If we succeed we try to free |
| 1232 | * the page as well. |
| 1233 | * |
| 1234 | * We do this even if the page is PageDirty(). |
| 1235 | * try_to_release_page() does not perform I/O, but it is |
| 1236 | * possible for a page to have PageDirty set, but it is actually |
| 1237 | * clean (all its buffers are clean). This happens if the |
| 1238 | * buffers were written out directly, with submit_bh(). ext3 |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1239 | * will do this, as well as the blockdev mapping. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1240 | * try_to_release_page() will discover that cleanness and will |
| 1241 | * drop the buffers and mark the page clean - it can be freed. |
| 1242 | * |
| 1243 | * Rarely, pages can have buffers and no ->mapping. These are |
| 1244 | * the pages which were not successfully invalidated in |
| 1245 | * truncate_complete_page(). We try to drop those buffers here |
| 1246 | * and if that worked, and the page is no longer mapped into |
| 1247 | * process address space (page_count == 1) it can be freed. |
| 1248 | * Otherwise, leave the page on the LRU so it is swappable. |
| 1249 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 1250 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1251 | if (!try_to_release_page(page, sc->gfp_mask)) |
| 1252 | goto activate_locked; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 1253 | if (!mapping && page_count(page) == 1) { |
| 1254 | unlock_page(page); |
| 1255 | if (put_page_testzero(page)) |
| 1256 | goto free_it; |
| 1257 | else { |
| 1258 | /* |
| 1259 | * rare race with speculative reference. |
| 1260 | * the speculative reference will free |
| 1261 | * this page shortly, so we may |
| 1262 | * increment nr_reclaimed here (and |
| 1263 | * leave it off the LRU). |
| 1264 | */ |
| 1265 | nr_reclaimed++; |
| 1266 | continue; |
| 1267 | } |
| 1268 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1269 | } |
| 1270 | |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1271 | lazyfree: |
Johannes Weiner | a528910 | 2014-04-03 14:47:51 -0700 | [diff] [blame] | 1272 | if (!mapping || !__remove_mapping(mapping, page, true)) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 1273 | goto keep_locked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1274 | |
Nick Piggin | a978d6f | 2008-10-18 20:26:58 -0700 | [diff] [blame] | 1275 | /* |
| 1276 | * At this point, we have no other references and there is |
| 1277 | * no way to pick any more up (removed from LRU, removed |
| 1278 | * from pagecache). Can use non-atomic bitops now (and |
| 1279 | * we obviously don't have to worry about waking up a process |
| 1280 | * waiting on the page lock, because there are no references. |
| 1281 | */ |
Kirill A. Shutemov | 48c935a | 2016-01-15 16:51:24 -0800 | [diff] [blame] | 1282 | __ClearPageLocked(page); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 1283 | free_it: |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1284 | if (ret == SWAP_LZFREE) |
| 1285 | count_vm_event(PGLAZYFREED); |
| 1286 | |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1287 | nr_reclaimed++; |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1288 | |
| 1289 | /* |
| 1290 | * Is there need to periodically free_page_list? It would |
| 1291 | * appear not as the counts should be low |
| 1292 | */ |
| 1293 | list_add(&page->lru, &free_pages); |
Minchan Kim | 74ac882 | 2013-05-09 16:21:26 +0900 | [diff] [blame] | 1294 | /* |
| 1295 | * If pagelist are from multiple zones, we should decrease |
| 1296 | * NR_ISOLATED_ANON + x on freed pages in here. |
| 1297 | */ |
| 1298 | if (!pgdat) |
| 1299 | dec_node_page_state(page, NR_ISOLATED_ANON + |
| 1300 | page_is_file_cache(page)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1301 | continue; |
| 1302 | |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 1303 | cull_mlocked: |
Hugh Dickins | 63d6c5a | 2009-01-06 14:39:38 -0800 | [diff] [blame] | 1304 | if (PageSwapCache(page)) |
| 1305 | try_to_free_swap(page); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 1306 | unlock_page(page); |
Jaewon Kim | c54839a | 2015-09-08 15:02:21 -0700 | [diff] [blame] | 1307 | list_add(&page->lru, &ret_pages); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 1308 | continue; |
| 1309 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1310 | activate_locked: |
Rik van Riel | 68a22394 | 2008-10-18 20:26:23 -0700 | [diff] [blame] | 1311 | /* Not a candidate for swapping, so reclaim swap space. */ |
Vladimir Davydov | 5ccc5ab | 2016-01-20 15:03:10 -0800 | [diff] [blame] | 1312 | if (PageSwapCache(page) && mem_cgroup_swap_full(page)) |
Hugh Dickins | a2c43ee | 2009-01-06 14:39:36 -0800 | [diff] [blame] | 1313 | try_to_free_swap(page); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1314 | VM_BUG_ON_PAGE(PageActive(page), page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1315 | SetPageActive(page); |
| 1316 | pgactivate++; |
| 1317 | keep_locked: |
| 1318 | unlock_page(page); |
| 1319 | keep: |
| 1320 | list_add(&page->lru, &ret_pages); |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1321 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1322 | } |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1323 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1324 | mem_cgroup_uncharge_list(&free_pages); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 1325 | try_to_unmap_flush(); |
Mel Gorman | b745bc8 | 2014-06-04 16:10:22 -0700 | [diff] [blame] | 1326 | free_hot_cold_page_list(&free_pages, true); |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 1327 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1328 | list_splice(&ret_pages, page_list); |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 1329 | count_vm_events(PGACTIVATE, pgactivate); |
Johannes Weiner | 0a31bc9 | 2014-08-08 14:19:22 -0700 | [diff] [blame] | 1330 | |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1331 | *ret_nr_dirty += nr_dirty; |
| 1332 | *ret_nr_congested += nr_congested; |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 1333 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1334 | *ret_nr_writeback += nr_writeback; |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1335 | *ret_nr_immediate += nr_immediate; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1336 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1337 | } |
| 1338 | |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1339 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
| 1340 | struct list_head *page_list) |
| 1341 | { |
| 1342 | struct scan_control sc = { |
| 1343 | .gfp_mask = GFP_KERNEL, |
| 1344 | .priority = DEF_PRIORITY, |
| 1345 | .may_unmap = 1, |
Minchan Kim | 13e7b91 | 2015-01-09 18:36:55 +0530 | [diff] [blame] | 1346 | /* Doesn't allow to write out dirty page */ |
| 1347 | .may_writepage = 0, |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1348 | }; |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1349 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1350 | struct page *page, *next; |
| 1351 | LIST_HEAD(clean_pages); |
| 1352 | |
| 1353 | list_for_each_entry_safe(page, next, page_list, lru) { |
Rafael Aquini | 117aad1 | 2013-09-30 13:45:16 -0700 | [diff] [blame] | 1354 | if (page_is_file_cache(page) && !PageDirty(page) && |
Minchan Kim | b1123ea6 | 2016-07-26 15:23:09 -0700 | [diff] [blame] | 1355 | !__PageMovable(page)) { |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1356 | ClearPageActive(page); |
| 1357 | list_move(&page->lru, &clean_pages); |
| 1358 | } |
| 1359 | } |
| 1360 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1361 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1362 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
| 1363 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1364 | list_splice(&clean_pages, page_list); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1365 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
Minchan Kim | 02c6de8 | 2012-10-08 16:31:55 -0700 | [diff] [blame] | 1366 | return ret; |
| 1367 | } |
| 1368 | |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1369 | #ifdef CONFIG_PROCESS_RECLAIM |
Minchan Kim | cd256bf | 2013-05-09 16:21:27 +0900 | [diff] [blame] | 1370 | unsigned long reclaim_pages_from_list(struct list_head *page_list, |
| 1371 | struct vm_area_struct *vma) |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1372 | { |
| 1373 | struct scan_control sc = { |
| 1374 | .gfp_mask = GFP_KERNEL, |
| 1375 | .priority = DEF_PRIORITY, |
| 1376 | .may_writepage = 1, |
| 1377 | .may_unmap = 1, |
| 1378 | .may_swap = 1, |
Minchan Kim | cd256bf | 2013-05-09 16:21:27 +0900 | [diff] [blame] | 1379 | .target_vma = vma, |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1380 | }; |
| 1381 | |
Minchan Kim | 74ac882 | 2013-05-09 16:21:26 +0900 | [diff] [blame] | 1382 | unsigned long nr_reclaimed; |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1383 | struct page *page; |
| 1384 | unsigned long dummy1, dummy2, dummy3, dummy4, dummy5; |
Minchan Kim | 74ac882 | 2013-05-09 16:21:26 +0900 | [diff] [blame] | 1385 | |
| 1386 | list_for_each_entry(page, page_list, lru) |
| 1387 | ClearPageActive(page); |
| 1388 | |
| 1389 | nr_reclaimed = shrink_page_list(page_list, NULL, &sc, |
| 1390 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
| 1391 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1392 | |
| 1393 | while (!list_empty(page_list)) { |
| 1394 | page = lru_to_page(page_list); |
| 1395 | list_del(&page->lru); |
Minchan Kim | ff92689 | 2013-05-09 16:21:24 +0900 | [diff] [blame] | 1396 | dec_node_page_state(page, NR_ISOLATED_ANON + |
| 1397 | page_is_file_cache(page)); |
| 1398 | putback_lru_page(page); |
| 1399 | } |
| 1400 | |
| 1401 | return nr_reclaimed; |
| 1402 | } |
| 1403 | #endif |
| 1404 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1405 | /* |
| 1406 | * Attempt to remove the specified page from its LRU. Only take this page |
| 1407 | * if it is of the appropriate PageActive status. Pages which are being |
| 1408 | * freed elsewhere are also ignored. |
| 1409 | * |
| 1410 | * page: page to consider |
| 1411 | * mode: one of the LRU isolation modes defined above |
| 1412 | * |
| 1413 | * returns 0 on success, -ve errno on failure. |
| 1414 | */ |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1415 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1416 | { |
| 1417 | int ret = -EINVAL; |
| 1418 | |
| 1419 | /* Only take pages on the LRU. */ |
| 1420 | if (!PageLRU(page)) |
| 1421 | return ret; |
| 1422 | |
Minchan Kim | e46a287 | 2012-10-08 16:33:48 -0700 | [diff] [blame] | 1423 | /* Compaction should not handle unevictable pages but CMA can do so */ |
| 1424 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1425 | return ret; |
| 1426 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1427 | ret = -EBUSY; |
KAMEZAWA Hiroyuki | 08e552c | 2009-01-07 18:08:01 -0800 | [diff] [blame] | 1428 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1429 | /* |
| 1430 | * To minimise LRU disruption, the caller can indicate that it only |
| 1431 | * wants to isolate pages it will be able to operate on without |
| 1432 | * blocking - clean pages for the most part. |
| 1433 | * |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1434 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
| 1435 | * that it is possible to migrate without blocking |
| 1436 | */ |
Johannes Weiner | 14d6ecf | 2017-02-24 14:56:11 -0800 | [diff] [blame] | 1437 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1438 | /* All the caller can do on PageWriteback is block */ |
| 1439 | if (PageWriteback(page)) |
| 1440 | return ret; |
| 1441 | |
| 1442 | if (PageDirty(page)) { |
| 1443 | struct address_space *mapping; |
Mel Gorman | ab88b8a | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1444 | bool migrate_dirty; |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1445 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1446 | /* |
| 1447 | * Only pages without mappings or that have a |
| 1448 | * ->migratepage callback are possible to migrate |
Mel Gorman | ab88b8a | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1449 | * without blocking. However, we can be racing with |
| 1450 | * truncation so it's necessary to lock the page |
| 1451 | * to stabilise the mapping as truncation holds |
| 1452 | * the page lock until after the page is removed |
| 1453 | * from the page cache. |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1454 | */ |
Mel Gorman | ab88b8a | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1455 | if (!trylock_page(page)) |
| 1456 | return ret; |
| 1457 | |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1458 | mapping = page_mapping(page); |
Hugh Dickins | 93960f9 | 2018-06-01 16:50:50 -0700 | [diff] [blame] | 1459 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
Mel Gorman | ab88b8a | 2018-01-31 16:19:52 -0800 | [diff] [blame] | 1460 | unlock_page(page); |
| 1461 | if (!migrate_dirty) |
Mel Gorman | c824493 | 2012-01-12 17:19:38 -0800 | [diff] [blame] | 1462 | return ret; |
| 1463 | } |
| 1464 | } |
Minchan Kim | 39deaf8 | 2011-10-31 17:06:51 -0700 | [diff] [blame] | 1465 | |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1466 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
| 1467 | return ret; |
| 1468 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1469 | if (likely(get_page_unless_zero(page))) { |
| 1470 | /* |
| 1471 | * Be careful not to clear PageLRU until after we're |
| 1472 | * sure the page is not being freed elsewhere -- the |
| 1473 | * page release code relies on it. |
| 1474 | */ |
| 1475 | ClearPageLRU(page); |
| 1476 | ret = 0; |
| 1477 | } |
| 1478 | |
| 1479 | return ret; |
| 1480 | } |
| 1481 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1482 | |
| 1483 | /* |
| 1484 | * Update LRU sizes after isolating pages. The LRU size updates must |
| 1485 | * be complete before mem_cgroup_update_lru_size due to a santity check. |
| 1486 | */ |
| 1487 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, |
Michal Hocko | 07fc957 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1488 | enum lru_list lru, unsigned long *nr_zone_taken) |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1489 | { |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1490 | int zid; |
| 1491 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1492 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
| 1493 | if (!nr_zone_taken[zid]) |
| 1494 | continue; |
| 1495 | |
| 1496 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
Michal Hocko | 07fc957 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1497 | #ifdef CONFIG_MEMCG |
| 1498 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
| 1499 | #endif |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1500 | } |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1501 | |
Mel Gorman | 7ee36a1 | 2016-07-28 15:47:17 -0700 | [diff] [blame] | 1502 | } |
| 1503 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 1504 | /* |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1505 | * zone_lru_lock is heavily contended. Some of the functions that |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1506 | * shrink the lists perform better by taking out a batch of pages |
| 1507 | * and working on them outside the LRU lock. |
| 1508 | * |
| 1509 | * For pagecache intensive workloads, this function is the hottest |
| 1510 | * spot in the kernel (apart from copy_*_user functions). |
| 1511 | * |
| 1512 | * Appropriate locks must be held before calling this function. |
| 1513 | * |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1514 | * @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] | 1515 | * @lruvec: The LRU vector to pull pages from. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1516 | * @dst: The temp list to put pages on to. |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 1517 | * @nr_scanned: The number of pages that were scanned. |
Rik van Riel | fe2c2a1 | 2012-03-21 16:33:51 -0700 | [diff] [blame] | 1518 | * @sc: The scan_control struct for this reclaim session |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1519 | * @mode: One of the LRU isolation modes |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1520 | * @lru: LRU list id for isolating |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1521 | * |
| 1522 | * returns how many pages were moved onto *@dst. |
| 1523 | */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1524 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1525 | struct lruvec *lruvec, struct list_head *dst, |
Rik van Riel | fe2c2a1 | 2012-03-21 16:33:51 -0700 | [diff] [blame] | 1526 | unsigned long *nr_scanned, struct scan_control *sc, |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1527 | isolate_mode_t mode, enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1528 | { |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1529 | struct list_head *src = &lruvec->lists[lru]; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1530 | unsigned long nr_taken = 0; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1531 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1532 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
Johannes Weiner | 061d6da | 2017-05-03 14:52:13 -0700 | [diff] [blame] | 1533 | unsigned long skipped = 0; |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1534 | unsigned long scan, total_scan, nr_pages; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1535 | LIST_HEAD(pages_skipped); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1536 | |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1537 | scan = 0; |
| 1538 | for (total_scan = 0; |
| 1539 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); |
| 1540 | total_scan++) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1541 | struct page *page; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1542 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1543 | page = lru_to_page(src); |
| 1544 | prefetchw_prev_lru_page(page, src, flags); |
| 1545 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1546 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
Nick Piggin | 8d438f9 | 2006-03-22 00:07:59 -0800 | [diff] [blame] | 1547 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1548 | if (page_zonenum(page) > sc->reclaim_idx) { |
| 1549 | list_move(&page->lru, &pages_skipped); |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1550 | nr_skipped[page_zonenum(page)]++; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1551 | continue; |
| 1552 | } |
| 1553 | |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1554 | /* |
| 1555 | * Do not count skipped pages because that makes the function |
| 1556 | * return with no isolated pages if the LRU mostly contains |
| 1557 | * ineligible pages. This causes the VM to not reclaim any |
| 1558 | * pages, triggering a premature OOM. |
| 1559 | */ |
| 1560 | scan++; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1561 | switch (__isolate_lru_page(page, mode)) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1562 | case 0: |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1563 | nr_pages = hpage_nr_pages(page); |
| 1564 | nr_taken += nr_pages; |
| 1565 | nr_zone_taken[page_zonenum(page)] += nr_pages; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1566 | list_move(&page->lru, dst); |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1567 | break; |
Nick Piggin | 46453a6 | 2006-03-22 00:07:58 -0800 | [diff] [blame] | 1568 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1569 | case -EBUSY: |
| 1570 | /* else it is being freed elsewhere */ |
| 1571 | list_move(&page->lru, src); |
| 1572 | continue; |
| 1573 | |
| 1574 | default: |
| 1575 | BUG(); |
| 1576 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1577 | } |
| 1578 | |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1579 | /* |
| 1580 | * Splice any skipped pages to the start of the LRU list. Note that |
| 1581 | * this disrupts the LRU order when reclaiming for lower zones but |
| 1582 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX |
| 1583 | * scanning would soon rescan the same pages to skip and put the |
| 1584 | * system at risk of premature OOM. |
| 1585 | */ |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1586 | if (!list_empty(&pages_skipped)) { |
| 1587 | int zid; |
| 1588 | |
Johannes Weiner | 061d6da | 2017-05-03 14:52:13 -0700 | [diff] [blame] | 1589 | list_splice(&pages_skipped, src); |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1590 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
| 1591 | if (!nr_skipped[zid]) |
| 1592 | continue; |
| 1593 | |
| 1594 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); |
Michal Hocko | 4d7658d | 2017-02-22 15:44:21 -0800 | [diff] [blame] | 1595 | skipped += nr_skipped[zid]; |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 1596 | } |
| 1597 | } |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1598 | *nr_scanned = total_scan; |
Michal Hocko | 4d7658d | 2017-02-22 15:44:21 -0800 | [diff] [blame] | 1599 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
Minchan Kim | 470e5b6 | 2017-05-12 15:47:06 -0700 | [diff] [blame] | 1600 | total_scan, skipped, nr_taken, mode, |
| 1601 | is_file_lru(lru)); |
Michal Hocko | 07fc957 | 2017-01-10 16:58:04 -0800 | [diff] [blame] | 1602 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1603 | return nr_taken; |
| 1604 | } |
| 1605 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1606 | /** |
| 1607 | * isolate_lru_page - tries to isolate a page from its LRU list |
| 1608 | * @page: page to isolate from its LRU list |
| 1609 | * |
| 1610 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the |
| 1611 | * vmstat statistic corresponding to whatever LRU list the page was on. |
| 1612 | * |
| 1613 | * Returns 0 if the page was removed from an LRU list. |
| 1614 | * Returns -EBUSY if the page was not on an LRU list. |
| 1615 | * |
| 1616 | * The returned page will have PageLRU() cleared. If it was found on |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1617 | * the active list, it will have PageActive set. If it was found on |
| 1618 | * the unevictable list, it will have the PageUnevictable bit set. That flag |
| 1619 | * 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] | 1620 | * |
| 1621 | * The vmstat statistic corresponding to the list on which the page was |
| 1622 | * found will be decremented. |
| 1623 | * |
| 1624 | * Restrictions: |
| 1625 | * (1) Must be called with an elevated refcount on the page. This is a |
| 1626 | * fundamentnal difference from isolate_lru_pages (which is called |
| 1627 | * without a stable reference). |
| 1628 | * (2) the lru_lock must not be held. |
| 1629 | * (3) interrupts must be enabled. |
| 1630 | */ |
| 1631 | int isolate_lru_page(struct page *page) |
| 1632 | { |
| 1633 | int ret = -EBUSY; |
| 1634 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1635 | VM_BUG_ON_PAGE(!page_count(page), page); |
Kirill A. Shutemov | cf2a82e | 2016-02-05 15:36:36 -0800 | [diff] [blame] | 1636 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1637 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1638 | if (PageLRU(page)) { |
| 1639 | struct zone *zone = page_zone(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1640 | struct lruvec *lruvec; |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1641 | |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1642 | spin_lock_irq(zone_lru_lock(zone)); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1643 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1644 | if (PageLRU(page)) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1645 | int lru = page_lru(page); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1646 | get_page(page); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1647 | ClearPageLRU(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1648 | del_page_from_lru_list(page, lruvec, lru); |
| 1649 | ret = 0; |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1650 | } |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1651 | spin_unlock_irq(zone_lru_lock(zone)); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1652 | } |
| 1653 | return ret; |
| 1654 | } |
| 1655 | |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1656 | static int __too_many_isolated(struct pglist_data *pgdat, int file, |
| 1657 | struct scan_control *sc, int safe) |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1658 | { |
| 1659 | unsigned long inactive, isolated; |
| 1660 | |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1661 | if (file) { |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1662 | if (safe) { |
| 1663 | inactive = node_page_state_snapshot(pgdat, |
| 1664 | NR_INACTIVE_FILE); |
| 1665 | isolated = node_page_state_snapshot(pgdat, |
| 1666 | NR_ISOLATED_FILE); |
| 1667 | } else { |
| 1668 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
| 1669 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); |
| 1670 | } |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1671 | } else { |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1672 | if (safe) { |
| 1673 | inactive = node_page_state_snapshot(pgdat, |
| 1674 | NR_INACTIVE_ANON); |
| 1675 | isolated = node_page_state_snapshot(pgdat, |
| 1676 | NR_ISOLATED_ANON); |
| 1677 | } else { |
| 1678 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
| 1679 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); |
| 1680 | } |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1681 | } |
| 1682 | |
Fengguang Wu | 3cf2384 | 2012-12-18 14:23:31 -0800 | [diff] [blame] | 1683 | /* |
| 1684 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they |
| 1685 | * won't get blocked by normal direct-reclaimers, forming a circular |
| 1686 | * deadlock. |
| 1687 | */ |
Mel Gorman | d0164ad | 2015-11-06 16:28:21 -0800 | [diff] [blame] | 1688 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
Fengguang Wu | 3cf2384 | 2012-12-18 14:23:31 -0800 | [diff] [blame] | 1689 | inactive >>= 3; |
| 1690 | |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1691 | return isolated > inactive; |
| 1692 | } |
| 1693 | |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1694 | /* |
| 1695 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
| 1696 | * then get resheduled. When there are massive number of tasks doing page |
| 1697 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
| 1698 | * the LRU list will go small and be scanned faster than necessary, leading to |
| 1699 | * unnecessary swapping, thrashing and OOM. |
| 1700 | */ |
| 1701 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
| 1702 | struct scan_control *sc, int safe) |
| 1703 | { |
| 1704 | if (current_is_kswapd()) |
| 1705 | return 0; |
| 1706 | |
| 1707 | if (!sane_reclaim(sc)) |
| 1708 | return 0; |
| 1709 | |
| 1710 | if (unlikely(__too_many_isolated(pgdat, file, sc, 0))) { |
| 1711 | if (safe) |
| 1712 | return __too_many_isolated(pgdat, file, sc, safe); |
| 1713 | else |
| 1714 | return 1; |
| 1715 | } |
| 1716 | |
| 1717 | return 0; |
| 1718 | } |
| 1719 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1720 | static noinline_for_stack void |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1721 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1722 | { |
Konstantin Khlebnikov | 27ac81d | 2012-05-29 15:07:00 -0700 | [diff] [blame] | 1723 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1724 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1725 | LIST_HEAD(pages_to_free); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1726 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1727 | /* |
| 1728 | * Put back any unfreeable pages. |
| 1729 | */ |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1730 | while (!list_empty(page_list)) { |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1731 | struct page *page = lru_to_page(page_list); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1732 | int lru; |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1733 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1734 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1735 | list_del(&page->lru); |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 1736 | if (unlikely(!page_evictable(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1737 | spin_unlock_irq(&pgdat->lru_lock); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1738 | putback_lru_page(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1739 | spin_lock_irq(&pgdat->lru_lock); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1740 | continue; |
| 1741 | } |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1742 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1743 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1744 | |
Linus Torvalds | 7a60857 | 2011-01-17 14:42:19 -0800 | [diff] [blame] | 1745 | SetPageLRU(page); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1746 | lru = page_lru(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1747 | add_page_to_lru_list(page, lruvec, lru); |
| 1748 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1749 | if (is_active_lru(lru)) { |
| 1750 | int file = is_file_lru(lru); |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1751 | int numpages = hpage_nr_pages(page); |
| 1752 | reclaim_stat->recent_rotated[file] += numpages; |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1753 | } |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1754 | if (put_page_testzero(page)) { |
| 1755 | __ClearPageLRU(page); |
| 1756 | __ClearPageActive(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1757 | del_page_from_lru_list(page, lruvec, lru); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1758 | |
| 1759 | if (unlikely(PageCompound(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1760 | spin_unlock_irq(&pgdat->lru_lock); |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1761 | mem_cgroup_uncharge(page); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1762 | (*get_compound_page_dtor(page))(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1763 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1764 | } else |
| 1765 | list_add(&page->lru, &pages_to_free); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1766 | } |
| 1767 | } |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1768 | |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1769 | /* |
| 1770 | * To save our caller's stack, now use input list for pages to free. |
| 1771 | */ |
| 1772 | list_splice(&pages_to_free, page_list); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1773 | } |
| 1774 | |
| 1775 | /* |
NeilBrown | 399ba0b | 2014-06-04 16:07:42 -0700 | [diff] [blame] | 1776 | * If a kernel thread (such as nfsd for loop-back mounts) services |
| 1777 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. |
| 1778 | * In that case we should only throttle if the backing device it is |
| 1779 | * writing to is congested. In other cases it is safe to throttle. |
| 1780 | */ |
| 1781 | static int current_may_throttle(void) |
| 1782 | { |
| 1783 | return !(current->flags & PF_LESS_THROTTLE) || |
| 1784 | current->backing_dev_info == NULL || |
| 1785 | bdi_write_congested(current->backing_dev_info); |
| 1786 | } |
| 1787 | |
Minchan Kim | 91dcade | 2016-07-28 15:47:40 -0700 | [diff] [blame] | 1788 | static bool inactive_reclaimable_pages(struct lruvec *lruvec, |
| 1789 | struct scan_control *sc, enum lru_list lru) |
| 1790 | { |
| 1791 | int zid; |
| 1792 | struct zone *zone; |
| 1793 | int file = is_file_lru(lru); |
| 1794 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
| 1795 | |
| 1796 | if (!global_reclaim(sc)) |
| 1797 | return true; |
| 1798 | |
| 1799 | for (zid = sc->reclaim_idx; zid >= 0; zid--) { |
| 1800 | zone = &pgdat->node_zones[zid]; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 1801 | if (!managed_zone(zone)) |
Minchan Kim | 91dcade | 2016-07-28 15:47:40 -0700 | [diff] [blame] | 1802 | continue; |
| 1803 | |
| 1804 | if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE + |
| 1805 | LRU_FILE * file) >= SWAP_CLUSTER_MAX) |
| 1806 | return true; |
| 1807 | } |
| 1808 | |
| 1809 | return false; |
| 1810 | } |
| 1811 | |
NeilBrown | 399ba0b | 2014-06-04 16:07:42 -0700 | [diff] [blame] | 1812 | /* |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 1813 | * 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] | 1814 | * of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1815 | */ |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1816 | static noinline_for_stack unsigned long |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1817 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 1818 | struct scan_control *sc, enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1819 | { |
| 1820 | LIST_HEAD(page_list); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1821 | unsigned long nr_scanned; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1822 | unsigned long nr_reclaimed = 0; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1823 | unsigned long nr_taken; |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1824 | unsigned long nr_dirty = 0; |
| 1825 | unsigned long nr_congested = 0; |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1826 | unsigned long nr_unqueued_dirty = 0; |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1827 | unsigned long nr_writeback = 0; |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1828 | unsigned long nr_immediate = 0; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 1829 | isolate_mode_t isolate_mode = 0; |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 1830 | int file = is_file_lru(lru); |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1831 | int safe = 0; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1832 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 1833 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
KOSAKI Motohiro | 78dc583 | 2009-06-16 15:31:40 -0700 | [diff] [blame] | 1834 | |
Minchan Kim | 91dcade | 2016-07-28 15:47:40 -0700 | [diff] [blame] | 1835 | if (!inactive_reclaimable_pages(lruvec, sc, lru)) |
| 1836 | return 0; |
| 1837 | |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1838 | while (unlikely(too_many_isolated(pgdat, file, sc, safe))) { |
KOSAKI Motohiro | 58355c7 | 2009-10-26 16:49:35 -0700 | [diff] [blame] | 1839 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1840 | |
| 1841 | /* We are about to die and free our memory. Return now. */ |
| 1842 | if (fatal_signal_pending(current)) |
| 1843 | return SWAP_CLUSTER_MAX; |
Vinayak Menon | dd68ed2 | 2014-12-26 19:29:41 +0530 | [diff] [blame] | 1844 | |
| 1845 | safe = 1; |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1846 | } |
| 1847 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1848 | lru_add_drain(); |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1849 | |
| 1850 | if (!sc->may_unmap) |
Hillf Danton | 6131728 | 2012-03-21 16:33:48 -0700 | [diff] [blame] | 1851 | isolate_mode |= ISOLATE_UNMAPPED; |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 1852 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1853 | spin_lock_irq(&pgdat->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1854 | |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 1855 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
| 1856 | &nr_scanned, sc, isolate_mode, lru); |
Konstantin Khlebnikov | 95d918f | 2012-05-29 15:06:59 -0700 | [diff] [blame] | 1857 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1858 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 1859 | reclaim_stat->recent_scanned[file] += nr_taken; |
Konstantin Khlebnikov | 95d918f | 2012-05-29 15:06:59 -0700 | [diff] [blame] | 1860 | |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 1861 | if (global_reclaim(sc)) { |
KOSAKI Motohiro | b35ea17 | 2009-09-21 17:01:36 -0700 | [diff] [blame] | 1862 | if (current_is_kswapd()) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1863 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1864 | else |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1865 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1866 | } |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1867 | spin_unlock_irq(&pgdat->lru_lock); |
KOSAKI Motohiro | a731286 | 2009-09-21 17:01:37 -0700 | [diff] [blame] | 1868 | |
Hillf Danton | d563c05 | 2012-03-21 16:34:02 -0700 | [diff] [blame] | 1869 | if (nr_taken == 0) |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1870 | return 0; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1871 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1872 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP, |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1873 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
| 1874 | &nr_writeback, &nr_immediate, |
| 1875 | false); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1876 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1877 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1878 | |
Ying Han | 904249a | 2012-04-25 16:01:48 -0700 | [diff] [blame] | 1879 | if (global_reclaim(sc)) { |
| 1880 | if (current_is_kswapd()) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1881 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
Ying Han | 904249a | 2012-04-25 16:01:48 -0700 | [diff] [blame] | 1882 | else |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1883 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
Ying Han | 904249a | 2012-04-25 16:01:48 -0700 | [diff] [blame] | 1884 | } |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1885 | |
Konstantin Khlebnikov | 27ac81d | 2012-05-29 15:07:00 -0700 | [diff] [blame] | 1886 | putback_inactive_pages(lruvec, &page_list); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1887 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1888 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1889 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1890 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 3f79768 | 2012-01-12 17:20:07 -0800 | [diff] [blame] | 1891 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 1892 | mem_cgroup_uncharge_list(&page_list); |
Mel Gorman | b745bc8 | 2014-06-04 16:10:22 -0700 | [diff] [blame] | 1893 | free_hot_cold_page_list(&page_list, true); |
Mel Gorman | e11da5b | 2010-10-26 14:21:40 -0700 | [diff] [blame] | 1894 | |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1895 | /* |
| 1896 | * If reclaim is isolating dirty pages under writeback, it implies |
| 1897 | * that the long-lived page allocation rate is exceeding the page |
| 1898 | * laundering rate. Either the global limits are not being effective |
| 1899 | * at throttling processes due to the page distribution throughout |
| 1900 | * zones or there is heavy usage of a slow backing device. The |
| 1901 | * only option is to throttle from reclaim context which is not ideal |
| 1902 | * as there is no guarantee the dirtying process is throttled in the |
| 1903 | * same way balance_dirty_pages() manages. |
| 1904 | * |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1905 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
| 1906 | * of pages under pages flagged for immediate reclaim and stall if any |
| 1907 | * are encountered in the nr_immediate check below. |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1908 | */ |
Mel Gorman | 918fc71 | 2013-07-08 16:00:25 -0700 | [diff] [blame] | 1909 | if (nr_writeback && nr_writeback == nr_taken) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1910 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
Mel Gorman | 92df3a7 | 2011-10-31 17:07:56 -0700 | [diff] [blame] | 1911 | |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 1912 | /* |
Andrey Ryabinin | 82ed834 | 2018-03-22 16:17:42 -0700 | [diff] [blame] | 1913 | * If dirty pages are scanned that are not queued for IO, it |
| 1914 | * implies that flushers are not doing their job. This can |
| 1915 | * happen when memory pressure pushes dirty pages to the end of |
| 1916 | * the LRU before the dirty limits are breached and the dirty |
| 1917 | * data has expired. It can also happen when the proportion of |
| 1918 | * dirty pages grows not through writes but through memory |
| 1919 | * pressure reclaiming all the clean cache. And in some cases, |
| 1920 | * the flushers simply cannot keep up with the allocation |
| 1921 | * rate. Nudge the flusher threads in case they are asleep. |
| 1922 | */ |
| 1923 | if (nr_unqueued_dirty == nr_taken) |
| 1924 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); |
| 1925 | |
| 1926 | /* |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1927 | * Legacy memcg will stall in page writeback so avoid forcibly |
| 1928 | * stalling here. |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 1929 | */ |
Tejun Heo | 97c9341 | 2015-05-22 18:23:36 -0400 | [diff] [blame] | 1930 | if (sane_reclaim(sc)) { |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1931 | /* |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1932 | * Tag a zone as congested if all the dirty pages scanned were |
| 1933 | * backed by a congested BDI and wait_iff_congested will stall. |
| 1934 | */ |
| 1935 | if (nr_dirty && nr_dirty == nr_congested) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1936 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1937 | |
Andrey Ryabinin | 82ed834 | 2018-03-22 16:17:42 -0700 | [diff] [blame] | 1938 | /* Allow kswapd to start writing pages during reclaim. */ |
| 1939 | if (nr_unqueued_dirty == nr_taken) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1940 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1941 | |
| 1942 | /* |
Linus Torvalds | b738d76 | 2014-06-08 14:17:00 -0700 | [diff] [blame] | 1943 | * If kswapd scans pages marked marked for immediate |
| 1944 | * reclaim and under writeback (nr_immediate), it implies |
| 1945 | * that pages are cycling through the LRU faster than |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1946 | * they are written so also forcibly stall. |
| 1947 | */ |
Linus Torvalds | b738d76 | 2014-06-08 14:17:00 -0700 | [diff] [blame] | 1948 | if (nr_immediate && current_may_throttle()) |
Mel Gorman | b1a6f21 | 2013-07-03 15:01:58 -0700 | [diff] [blame] | 1949 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
Mel Gorman | e2be15f | 2013-07-03 15:01:57 -0700 | [diff] [blame] | 1950 | } |
Mel Gorman | d43006d | 2013-07-03 15:01:50 -0700 | [diff] [blame] | 1951 | |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1952 | /* |
| 1953 | * Stall direct reclaim for IO completions if underlying BDIs or zone |
| 1954 | * is congested. Allow kswapd to continue until it starts encountering |
| 1955 | * unqueued dirty pages or cycling through the LRU too quickly. |
| 1956 | */ |
NeilBrown | 399ba0b | 2014-06-04 16:07:42 -0700 | [diff] [blame] | 1957 | if (!sc->hibernation_mode && !current_is_kswapd() && |
| 1958 | current_may_throttle()) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1959 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
Mel Gorman | 8e95028 | 2013-07-03 15:02:02 -0700 | [diff] [blame] | 1960 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1961 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
| 1962 | nr_scanned, nr_reclaimed, |
yalin wang | ba5e957 | 2016-01-14 15:18:48 -0800 | [diff] [blame] | 1963 | sc->priority, file); |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1964 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1965 | } |
| 1966 | |
Martin Bligh | 3bb1a852 | 2006-10-28 10:38:24 -0700 | [diff] [blame] | 1967 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1968 | * This moves pages from the active list to the inactive list. |
| 1969 | * |
| 1970 | * We move them the other way if the page is referenced by one or more |
| 1971 | * processes, from rmap. |
| 1972 | * |
| 1973 | * 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] | 1974 | * appropriate to hold zone_lru_lock across the whole operation. But if |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1975 | * the pages are mapped, the processing is slow (page_referenced()) so we |
Mel Gorman | a52633d | 2016-07-28 15:45:28 -0700 | [diff] [blame] | 1976 | * 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] | 1977 | * this, so instead we remove the pages from the LRU while processing them. |
| 1978 | * It is safe to rely on PG_active against the non-LRU pages in here because |
| 1979 | * nobody will play with that bit on a non-LRU page. |
| 1980 | * |
Joonsoo Kim | 0139aa7 | 2016-05-19 17:10:49 -0700 | [diff] [blame] | 1981 | * 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] | 1982 | * But we had to alter page->flags anyway. |
| 1983 | */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1984 | |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1985 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1986 | struct list_head *list, |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 1987 | struct list_head *pages_to_free, |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1988 | enum lru_list lru) |
| 1989 | { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1990 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1991 | unsigned long pgmoved = 0; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1992 | struct page *page; |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 1993 | int nr_pages; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1994 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1995 | while (!list_empty(list)) { |
| 1996 | page = lru_to_page(list); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 1997 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1998 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1999 | VM_BUG_ON_PAGE(PageLRU(page), page); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 2000 | SetPageLRU(page); |
| 2001 | |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2002 | nr_pages = hpage_nr_pages(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2003 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 2004 | list_move(&page->lru, &lruvec->lists[lru]); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2005 | pgmoved += nr_pages; |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 2006 | |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 2007 | if (put_page_testzero(page)) { |
| 2008 | __ClearPageLRU(page); |
| 2009 | __ClearPageActive(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2010 | del_page_from_lru_list(page, lruvec, lru); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 2011 | |
| 2012 | if (unlikely(PageCompound(page))) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2013 | spin_unlock_irq(&pgdat->lru_lock); |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 2014 | mem_cgroup_uncharge(page); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 2015 | (*get_compound_page_dtor(page))(page); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2016 | spin_lock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 2017 | } else |
| 2018 | list_add(&page->lru, pages_to_free); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 2019 | } |
| 2020 | } |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 2021 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 2022 | if (!is_active_lru(lru)) |
| 2023 | __count_vm_events(PGDEACTIVATE, pgmoved); |
| 2024 | } |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2025 | |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 2026 | static void shrink_active_list(unsigned long nr_to_scan, |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2027 | struct lruvec *lruvec, |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2028 | struct scan_control *sc, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2029 | enum lru_list lru) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2030 | { |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 2031 | unsigned long nr_taken; |
Hugh Dickins | f626012 | 2012-01-12 17:20:06 -0800 | [diff] [blame] | 2032 | unsigned long nr_scanned; |
Wu Fengguang | 6fe6b7e | 2009-06-16 15:33:05 -0700 | [diff] [blame] | 2033 | unsigned long vm_flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2034 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 2035 | LIST_HEAD(l_active); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2036 | LIST_HEAD(l_inactive); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2037 | struct page *page; |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2038 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 2039 | unsigned long nr_rotated = 0; |
Konstantin Khlebnikov | f3fd4a6 | 2012-05-29 15:06:54 -0700 | [diff] [blame] | 2040 | isolate_mode_t isolate_mode = 0; |
Konstantin Khlebnikov | 3cb9945 | 2012-05-29 15:06:53 -0700 | [diff] [blame] | 2041 | int file = is_file_lru(lru); |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2042 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2043 | |
| 2044 | lru_add_drain(); |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 2045 | |
| 2046 | if (!sc->may_unmap) |
Hillf Danton | 6131728 | 2012-03-21 16:33:48 -0700 | [diff] [blame] | 2047 | isolate_mode |= ISOLATE_UNMAPPED; |
Minchan Kim | f80c067 | 2011-10-31 17:06:55 -0700 | [diff] [blame] | 2048 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2049 | spin_lock_irq(&pgdat->lru_lock); |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 2050 | |
Konstantin Khlebnikov | 5dc3597 | 2012-05-29 15:06:58 -0700 | [diff] [blame] | 2051 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
| 2052 | &nr_scanned, sc, isolate_mode, lru); |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 2053 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2054 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 2055 | reclaim_stat->recent_scanned[file] += nr_taken; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2056 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2057 | __count_vm_events(PGREFILL, nr_scanned); |
Hugh Dickins | 9d5e6a9 | 2016-05-19 17:12:38 -0700 | [diff] [blame] | 2058 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2059 | spin_unlock_irq(&pgdat->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2060 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2061 | while (!list_empty(&l_hold)) { |
| 2062 | cond_resched(); |
| 2063 | page = lru_to_page(&l_hold); |
| 2064 | list_del(&page->lru); |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 2065 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 2066 | if (unlikely(!page_evictable(page))) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 2067 | putback_lru_page(page); |
| 2068 | continue; |
| 2069 | } |
| 2070 | |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2071 | if (unlikely(buffer_heads_over_limit)) { |
| 2072 | if (page_has_private(page) && trylock_page(page)) { |
| 2073 | if (page_has_private(page)) |
| 2074 | try_to_release_page(page, 0); |
| 2075 | unlock_page(page); |
| 2076 | } |
| 2077 | } |
| 2078 | |
Johannes Weiner | c3ac9a8 | 2012-05-29 15:06:25 -0700 | [diff] [blame] | 2079 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
| 2080 | &vm_flags)) { |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 2081 | nr_rotated += hpage_nr_pages(page); |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 2082 | /* |
| 2083 | * Identify referenced, file-backed active pages and |
| 2084 | * give them one more trip around the active list. So |
| 2085 | * that executable code get better chances to stay in |
| 2086 | * memory under moderate memory pressure. Anon pages |
| 2087 | * are not likely to be evicted by use-once streaming |
| 2088 | * IO, plus JVM can create lots of anon VM_EXEC pages, |
| 2089 | * so we ignore them here. |
| 2090 | */ |
Wu Fengguang | 41e2098 | 2009-10-26 16:49:53 -0700 | [diff] [blame] | 2091 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 2092 | list_add(&page->lru, &l_active); |
| 2093 | continue; |
| 2094 | } |
| 2095 | } |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 2096 | |
KOSAKI Motohiro | 5205e56 | 2009-09-21 17:01:44 -0700 | [diff] [blame] | 2097 | ClearPageActive(page); /* we are de-activating */ |
Johannes Weiner | a238332 | 2018-10-26 15:06:04 -0700 | [diff] [blame] | 2098 | SetPageWorkingset(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2099 | list_add(&page->lru, &l_inactive); |
| 2100 | } |
| 2101 | |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 2102 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 2103 | * Move pages back to the lru list. |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 2104 | */ |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2105 | spin_lock_irq(&pgdat->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2106 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 2107 | * Count referenced pages from currently used mappings as rotated, |
| 2108 | * even though only some of them are actually re-activated. This |
| 2109 | * helps balance scan pressure between file and anonymous pages in |
Jerome Marchand | 7c0db9e | 2014-08-06 16:08:01 -0700 | [diff] [blame] | 2110 | * get_scan_count. |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 2111 | */ |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 2112 | reclaim_stat->recent_rotated[file] += nr_rotated; |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2113 | |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2114 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
| 2115 | 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] | 2116 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
| 2117 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2bcf887 | 2012-01-12 17:19:56 -0800 | [diff] [blame] | 2118 | |
Johannes Weiner | 747db95 | 2014-08-08 14:19:24 -0700 | [diff] [blame] | 2119 | mem_cgroup_uncharge_list(&l_hold); |
Mel Gorman | b745bc8 | 2014-06-04 16:10:22 -0700 | [diff] [blame] | 2120 | free_hot_cold_page_list(&l_hold, true); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2121 | } |
| 2122 | |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2123 | /* |
| 2124 | * The inactive anon list should be small enough that the VM never has |
| 2125 | * to do too much work. |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2126 | * |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2127 | * The inactive file list should be small enough to leave most memory |
| 2128 | * to the established workingset on the scan-resistant active list, |
| 2129 | * but large enough to avoid thrashing the aggregate readahead window. |
| 2130 | * |
| 2131 | * Both inactive lists should also be large enough that each inactive |
| 2132 | * page has a chance to be referenced again before it is reclaimed. |
| 2133 | * |
| 2134 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
| 2135 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio |
| 2136 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
| 2137 | * |
| 2138 | * total target max |
| 2139 | * memory ratio inactive |
| 2140 | * ------------------------------------- |
| 2141 | * 10MB 1 5MB |
| 2142 | * 100MB 1 50MB |
| 2143 | * 1GB 3 250MB |
| 2144 | * 10GB 10 0.9GB |
| 2145 | * 100GB 31 3GB |
| 2146 | * 1TB 101 10GB |
| 2147 | * 10TB 320 32GB |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2148 | */ |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2149 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
| 2150 | struct scan_control *sc) |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2151 | { |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2152 | unsigned long inactive_ratio; |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2153 | unsigned long inactive, active; |
| 2154 | enum lru_list inactive_lru = file * LRU_FILE; |
| 2155 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2156 | unsigned long gb; |
| 2157 | |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 2158 | /* |
| 2159 | * If we don't have swap space, anonymous page deactivation |
| 2160 | * is pointless. |
| 2161 | */ |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2162 | if (!file && !total_swap_pages) |
Yaowei Bai | 42e2e45 | 2015-11-05 18:47:36 -0800 | [diff] [blame] | 2163 | return false; |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 2164 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2165 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
| 2166 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2167 | |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2168 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
| 2169 | if (gb) |
| 2170 | inactive_ratio = int_sqrt(10 * gb); |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2171 | else |
Rik van Riel | 59dc76b | 2016-05-20 16:56:31 -0700 | [diff] [blame] | 2172 | inactive_ratio = 1; |
| 2173 | |
| 2174 | return inactive * inactive_ratio < active; |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2175 | } |
| 2176 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2177 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2178 | struct lruvec *lruvec, struct scan_control *sc) |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2179 | { |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 2180 | if (is_active_lru(lru)) { |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2181 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc)) |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2182 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2183 | return 0; |
| 2184 | } |
| 2185 | |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2186 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2187 | } |
| 2188 | |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2189 | enum scan_balance { |
| 2190 | SCAN_EQUAL, |
| 2191 | SCAN_FRACT, |
| 2192 | SCAN_ANON, |
| 2193 | SCAN_FILE, |
| 2194 | }; |
| 2195 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2196 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2197 | * Determine how aggressively the anon and file LRU lists should be |
| 2198 | * scanned. The relative value of each set of LRU lists is determined |
| 2199 | * by looking at the fraction of the pages scanned we did rotate back |
| 2200 | * onto the active list instead of evict. |
| 2201 | * |
Wanpeng Li | be7bd59 | 2012-06-14 20:41:02 +0800 | [diff] [blame] | 2202 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
| 2203 | * 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] | 2204 | */ |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2205 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2206 | struct scan_control *sc, unsigned long *nr, |
| 2207 | unsigned long *lru_pages) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2208 | { |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2209 | int swappiness = mem_cgroup_swappiness(memcg); |
Konstantin Khlebnikov | 9012637 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 2210 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2211 | u64 fraction[2]; |
| 2212 | u64 denominator = 0; /* gcc */ |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2213 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2214 | unsigned long anon_prio, file_prio; |
| 2215 | enum scan_balance scan_balance; |
Johannes Weiner | 0bf1457 | 2014-04-08 16:04:10 -0700 | [diff] [blame] | 2216 | unsigned long anon, file; |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2217 | unsigned long ap, fp; |
| 2218 | enum lru_list lru; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2219 | |
| 2220 | /* If we have no swap space, do not bother scanning anon pages. */ |
Vladimir Davydov | d8b3843 | 2016-01-20 15:03:07 -0800 | [diff] [blame] | 2221 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2222 | scan_balance = SCAN_FILE; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2223 | goto out; |
| 2224 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2225 | |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2226 | /* |
| 2227 | * Global reclaim will swap to prevent OOM even with no |
| 2228 | * swappiness, but memcg users want to use this knob to |
| 2229 | * disable swapping for individual groups completely when |
| 2230 | * using the memory controller's swap limit feature would be |
| 2231 | * too expensive. |
| 2232 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2233 | if (!global_reclaim(sc) && !swappiness) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2234 | scan_balance = SCAN_FILE; |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2235 | goto out; |
| 2236 | } |
| 2237 | |
| 2238 | /* |
| 2239 | * Do not apply any pressure balancing cleverness when the |
| 2240 | * system is close to OOM, scan both anon and file equally |
| 2241 | * (unless the swappiness setting disagrees with swapping). |
| 2242 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2243 | if (!sc->priority && swappiness) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2244 | scan_balance = SCAN_EQUAL; |
Johannes Weiner | 10316b3 | 2013-02-22 16:32:14 -0800 | [diff] [blame] | 2245 | goto out; |
| 2246 | } |
| 2247 | |
Johannes Weiner | 11d16c2 | 2013-02-22 16:32:15 -0800 | [diff] [blame] | 2248 | /* |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2249 | * Prevent the reclaimer from falling into the cache trap: as |
| 2250 | * cache pages start out inactive, every cache fault will tip |
| 2251 | * the scan balance towards the file LRU. And as the file LRU |
| 2252 | * shrinks, so does the window for rotation from references. |
| 2253 | * This means we have a runaway feedback loop where a tiny |
| 2254 | * thrashing file LRU becomes infinitely more attractive than |
| 2255 | * anon pages. Try to detect this based on file LRU size. |
| 2256 | */ |
| 2257 | if (global_reclaim(sc)) { |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2258 | unsigned long pgdatfile; |
| 2259 | unsigned long pgdatfree; |
| 2260 | int z; |
| 2261 | unsigned long total_high_wmark = 0; |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2262 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2263 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
| 2264 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + |
| 2265 | node_page_state(pgdat, NR_INACTIVE_FILE); |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2266 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2267 | for (z = 0; z < MAX_NR_ZONES; z++) { |
| 2268 | struct zone *zone = &pgdat->node_zones[z]; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 2269 | if (!managed_zone(zone)) |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2270 | continue; |
| 2271 | |
| 2272 | total_high_wmark += high_wmark_pages(zone); |
| 2273 | } |
| 2274 | |
| 2275 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
Johannes Weiner | 6237625 | 2014-05-06 12:50:07 -0700 | [diff] [blame] | 2276 | scan_balance = SCAN_ANON; |
| 2277 | goto out; |
| 2278 | } |
| 2279 | } |
| 2280 | |
| 2281 | /* |
Vladimir Davydov | 316bda0 | 2016-01-14 15:19:38 -0800 | [diff] [blame] | 2282 | * If there is enough inactive page cache, i.e. if the size of the |
| 2283 | * inactive list is greater than that of the active list *and* the |
| 2284 | * inactive list actually has some pages to scan on this priority, we |
| 2285 | * do not reclaim anything from the anonymous working set right now. |
| 2286 | * Without the second condition we could end up never scanning an |
| 2287 | * lruvec even if it has plenty of old anonymous pages unless the |
| 2288 | * system is under heavy pressure. |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2289 | */ |
Patrick Daly | 7231212 | 2017-03-14 18:30:41 -0700 | [diff] [blame] | 2290 | if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) && |
| 2291 | !inactive_list_is_low(lruvec, true, sc) && |
Michal Hocko | 521e92b | 2017-02-22 15:46:01 -0800 | [diff] [blame] | 2292 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2293 | scan_balance = SCAN_FILE; |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2294 | goto out; |
| 2295 | } |
| 2296 | |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2297 | scan_balance = SCAN_FRACT; |
| 2298 | |
Johannes Weiner | 7c5bd70 | 2013-02-22 16:32:10 -0800 | [diff] [blame] | 2299 | /* |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 2300 | * With swappiness at 100, anonymous and file have the same priority. |
| 2301 | * This scanning priority is essentially the inverse of IO cost. |
| 2302 | */ |
Johannes Weiner | 0269517 | 2014-08-06 16:06:17 -0700 | [diff] [blame] | 2303 | anon_prio = swappiness; |
Hugh Dickins | 75b00af | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 2304 | file_prio = 200 - anon_prio; |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 2305 | |
| 2306 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2307 | * OK, so we have swap space and a fair amount of page cache |
| 2308 | * pages. We use the recently rotated / recently scanned |
| 2309 | * ratios to determine how valuable each cache is. |
| 2310 | * |
| 2311 | * Because workloads change over time (and to avoid overflow) |
| 2312 | * we keep these statistics as a floating average, which ends |
| 2313 | * up weighing recent references more than old ones. |
| 2314 | * |
| 2315 | * anon in [0], file in [1] |
| 2316 | */ |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2317 | |
Michal Hocko | 7105313 | 2017-02-22 15:45:58 -0800 | [diff] [blame] | 2318 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
| 2319 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); |
| 2320 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + |
| 2321 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); |
Jerome Marchand | 2ab051e | 2014-08-06 16:08:03 -0700 | [diff] [blame] | 2322 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2323 | spin_lock_irq(&pgdat->lru_lock); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2324 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2325 | reclaim_stat->recent_scanned[0] /= 2; |
| 2326 | reclaim_stat->recent_rotated[0] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2327 | } |
| 2328 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2329 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2330 | reclaim_stat->recent_scanned[1] /= 2; |
| 2331 | reclaim_stat->recent_rotated[1] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2332 | } |
| 2333 | |
| 2334 | /* |
Rik van Riel | 00d8089 | 2008-11-19 15:36:44 -0800 | [diff] [blame] | 2335 | * The amount of pressure on anon vs file pages is inversely |
| 2336 | * proportional to the fraction of recently scanned pages on |
| 2337 | * each list that were recently referenced and in active use. |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2338 | */ |
Satoru Moriya | fe35004 | 2012-05-29 15:06:47 -0700 | [diff] [blame] | 2339 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2340 | ap /= reclaim_stat->recent_rotated[0] + 1; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2341 | |
Satoru Moriya | fe35004 | 2012-05-29 15:06:47 -0700 | [diff] [blame] | 2342 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 2343 | fp /= reclaim_stat->recent_rotated[1] + 1; |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 2344 | spin_unlock_irq(&pgdat->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2345 | |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2346 | fraction[0] = ap; |
| 2347 | fraction[1] = fp; |
| 2348 | denominator = ap + fp + 1; |
| 2349 | out: |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2350 | *lru_pages = 0; |
| 2351 | for_each_evictable_lru(lru) { |
| 2352 | int file = is_file_lru(lru); |
| 2353 | unsigned long size; |
| 2354 | unsigned long scan; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2355 | |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2356 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
| 2357 | scan = size >> sc->priority; |
| 2358 | /* |
| 2359 | * If the cgroup's already been deleted, make sure to |
| 2360 | * scrape out the remaining cache. |
| 2361 | */ |
| 2362 | if (!scan && !mem_cgroup_online(memcg)) |
| 2363 | scan = min(size, SWAP_CLUSTER_MAX); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2364 | |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2365 | switch (scan_balance) { |
| 2366 | case SCAN_EQUAL: |
| 2367 | /* Scan lists relative to size */ |
| 2368 | break; |
| 2369 | case SCAN_FRACT: |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2370 | /* |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2371 | * Scan types proportional to swappiness and |
| 2372 | * their relative recent reclaim efficiency. |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2373 | */ |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2374 | scan = div64_u64(scan * fraction[file], |
| 2375 | denominator); |
| 2376 | break; |
| 2377 | case SCAN_FILE: |
| 2378 | case SCAN_ANON: |
| 2379 | /* Scan one type exclusively */ |
| 2380 | if ((scan_balance == SCAN_FILE) != file) { |
| 2381 | size = 0; |
| 2382 | scan = 0; |
| 2383 | } |
| 2384 | break; |
| 2385 | default: |
| 2386 | /* Look ma, no brain */ |
| 2387 | BUG(); |
Johannes Weiner | 9a26511 | 2013-02-22 16:32:17 -0800 | [diff] [blame] | 2388 | } |
Johannes Weiner | c383d08 | 2017-05-03 14:52:07 -0700 | [diff] [blame] | 2389 | |
| 2390 | *lru_pages += size; |
| 2391 | nr[lru] = scan; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2392 | } |
Wu Fengguang | 6e08a36 | 2009-06-16 15:32:29 -0700 | [diff] [blame] | 2393 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2394 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2395 | /* |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2396 | * 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] | 2397 | */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2398 | 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] | 2399 | struct scan_control *sc, unsigned long *lru_pages) |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2400 | { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2401 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2402 | unsigned long nr[NR_LRU_LISTS]; |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2403 | unsigned long targets[NR_LRU_LISTS]; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2404 | unsigned long nr_to_scan; |
| 2405 | enum lru_list lru; |
| 2406 | unsigned long nr_reclaimed = 0; |
| 2407 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
| 2408 | struct blk_plug plug; |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2409 | bool scan_adjusted; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2410 | |
Vladimir Davydov | 3337767 | 2016-01-20 15:02:59 -0800 | [diff] [blame] | 2411 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2412 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2413 | /* Record the original scan target for proportional adjustments later */ |
| 2414 | memcpy(targets, nr, sizeof(nr)); |
| 2415 | |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2416 | /* |
| 2417 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal |
| 2418 | * event that can occur when there is little memory pressure e.g. |
| 2419 | * multiple streaming readers/writers. Hence, we do not abort scanning |
| 2420 | * when the requested number of pages are reclaimed when scanning at |
| 2421 | * DEF_PRIORITY on the assumption that the fact we are direct |
| 2422 | * reclaiming implies that kswapd is not keeping up and it is best to |
| 2423 | * do a batch of work at once. For memcg reclaim one check is made to |
| 2424 | * abort proportional reclaim if either the file or anon lru has already |
| 2425 | * dropped to zero at the first pass. |
| 2426 | */ |
| 2427 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && |
| 2428 | sc->priority == DEF_PRIORITY); |
| 2429 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2430 | blk_start_plug(&plug); |
| 2431 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
| 2432 | nr[LRU_INACTIVE_FILE]) { |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2433 | unsigned long nr_anon, nr_file, percentage; |
| 2434 | unsigned long nr_scanned; |
| 2435 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2436 | for_each_evictable_lru(lru) { |
| 2437 | if (nr[lru]) { |
| 2438 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); |
| 2439 | nr[lru] -= nr_to_scan; |
| 2440 | |
| 2441 | nr_reclaimed += shrink_list(lru, nr_to_scan, |
| 2442 | lruvec, sc); |
| 2443 | } |
| 2444 | } |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2445 | |
Michal Hocko | bd04173 | 2016-12-02 17:26:48 -0800 | [diff] [blame] | 2446 | cond_resched(); |
| 2447 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2448 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
| 2449 | continue; |
| 2450 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2451 | /* |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2452 | * For kswapd and memcg, reclaim at least the number of pages |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2453 | * requested. Ensure that the anon and file LRUs are scanned |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2454 | * proportionally what was requested by get_scan_count(). We |
| 2455 | * stop reclaiming one LRU and reduce the amount scanning |
| 2456 | * proportional to the original scan target. |
| 2457 | */ |
| 2458 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; |
| 2459 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; |
| 2460 | |
Mel Gorman | 1a50190 | 2014-06-04 16:10:49 -0700 | [diff] [blame] | 2461 | /* |
| 2462 | * It's just vindictive to attack the larger once the smaller |
| 2463 | * has gone to zero. And given the way we stop scanning the |
| 2464 | * smaller below, this makes sure that we only make one nudge |
| 2465 | * towards proportionality once we've got nr_to_reclaim. |
| 2466 | */ |
| 2467 | if (!nr_file || !nr_anon) |
| 2468 | break; |
| 2469 | |
Mel Gorman | e82e056 | 2013-07-03 15:01:44 -0700 | [diff] [blame] | 2470 | if (nr_file > nr_anon) { |
| 2471 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + |
| 2472 | targets[LRU_ACTIVE_ANON] + 1; |
| 2473 | lru = LRU_BASE; |
| 2474 | percentage = nr_anon * 100 / scan_target; |
| 2475 | } else { |
| 2476 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + |
| 2477 | targets[LRU_ACTIVE_FILE] + 1; |
| 2478 | lru = LRU_FILE; |
| 2479 | percentage = nr_file * 100 / scan_target; |
| 2480 | } |
| 2481 | |
| 2482 | /* Stop scanning the smaller of the LRU */ |
| 2483 | nr[lru] = 0; |
| 2484 | nr[lru + LRU_ACTIVE] = 0; |
| 2485 | |
| 2486 | /* |
| 2487 | * Recalculate the other LRU scan count based on its original |
| 2488 | * scan target and the percentage scanning already complete |
| 2489 | */ |
| 2490 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; |
| 2491 | nr_scanned = targets[lru] - nr[lru]; |
| 2492 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
| 2493 | nr[lru] -= min(nr[lru], nr_scanned); |
| 2494 | |
| 2495 | lru += LRU_ACTIVE; |
| 2496 | nr_scanned = targets[lru] - nr[lru]; |
| 2497 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
| 2498 | nr[lru] -= min(nr[lru], nr_scanned); |
| 2499 | |
| 2500 | scan_adjusted = true; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2501 | } |
| 2502 | blk_finish_plug(&plug); |
| 2503 | sc->nr_reclaimed += nr_reclaimed; |
| 2504 | |
| 2505 | /* |
| 2506 | * Even if we did not try to evict anon pages at all, we want to |
| 2507 | * rebalance the anon lru active/inactive ratio. |
| 2508 | */ |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 2509 | if (inactive_list_is_low(lruvec, false, sc)) |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2510 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
| 2511 | sc, LRU_ACTIVE_ANON); |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2512 | } |
| 2513 | |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2514 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2515 | static bool in_reclaim_compaction(struct scan_control *sc) |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2516 | { |
Kirill A. Shutemov | d84da3f | 2012-12-11 16:00:31 -0800 | [diff] [blame] | 2517 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2518 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2519 | sc->priority < DEF_PRIORITY - 2)) |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2520 | return true; |
| 2521 | |
| 2522 | return false; |
| 2523 | } |
| 2524 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2525 | /* |
Mel Gorman | 23b9da5 | 2012-05-29 15:06:20 -0700 | [diff] [blame] | 2526 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
| 2527 | * order-0 pages before compacting the zone. should_continue_reclaim() returns |
| 2528 | * true if more pages should be reclaimed such that when the page allocator |
| 2529 | * calls try_to_compact_zone() that it will have enough free pages to succeed. |
| 2530 | * 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] | 2531 | */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2532 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2533 | unsigned long nr_reclaimed, |
| 2534 | unsigned long nr_scanned, |
| 2535 | struct scan_control *sc) |
| 2536 | { |
| 2537 | unsigned long pages_for_compaction; |
| 2538 | unsigned long inactive_lru_pages; |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2539 | int z; |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2540 | |
| 2541 | /* If not in reclaim/compaction mode, stop */ |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2542 | if (!in_reclaim_compaction(sc)) |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2543 | return false; |
| 2544 | |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 2545 | /* Consider stopping depending on scan and reclaim activity */ |
| 2546 | if (sc->gfp_mask & __GFP_REPEAT) { |
| 2547 | /* |
| 2548 | * For __GFP_REPEAT allocations, stop reclaiming if the |
| 2549 | * full LRU list has been scanned and we are still failing |
| 2550 | * to reclaim pages. This full LRU scan is potentially |
| 2551 | * expensive but a __GFP_REPEAT caller really wants to succeed |
| 2552 | */ |
| 2553 | if (!nr_reclaimed && !nr_scanned) |
| 2554 | return false; |
| 2555 | } else { |
| 2556 | /* |
| 2557 | * For non-__GFP_REPEAT allocations which can presumably |
| 2558 | * fail without consequence, stop if we failed to reclaim |
| 2559 | * any pages from the last SWAP_CLUSTER_MAX number of |
| 2560 | * pages that were scanned. This will return to the |
| 2561 | * caller faster at the risk reclaim/compaction and |
| 2562 | * the resulting allocation attempt fails |
| 2563 | */ |
| 2564 | if (!nr_reclaimed) |
| 2565 | return false; |
| 2566 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2567 | |
| 2568 | /* |
| 2569 | * If we have not reclaimed enough pages for compaction and the |
| 2570 | * inactive lists are large enough, continue reclaiming |
| 2571 | */ |
Vlastimil Babka | 9861a62 | 2016-10-07 16:57:53 -0700 | [diff] [blame] | 2572 | pages_for_compaction = compact_gap(sc->order); |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2573 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
Shaohua Li | ec8acf2 | 2013-02-22 16:34:38 -0800 | [diff] [blame] | 2574 | if (get_nr_swap_pages() > 0) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2575 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2576 | if (sc->nr_reclaimed < pages_for_compaction && |
| 2577 | inactive_lru_pages > pages_for_compaction) |
| 2578 | return true; |
| 2579 | |
| 2580 | /* If compaction would go ahead or the allocation would succeed, stop */ |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2581 | for (z = 0; z <= sc->reclaim_idx; z++) { |
| 2582 | struct zone *zone = &pgdat->node_zones[z]; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 2583 | if (!managed_zone(zone)) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2584 | continue; |
| 2585 | |
| 2586 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { |
Vlastimil Babka | cf37831 | 2016-10-07 16:57:41 -0700 | [diff] [blame] | 2587 | case COMPACT_SUCCESS: |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2588 | case COMPACT_CONTINUE: |
| 2589 | return false; |
| 2590 | default: |
| 2591 | /* check next zone */ |
| 2592 | ; |
| 2593 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2594 | } |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2595 | return true; |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2596 | } |
| 2597 | |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 2598 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2599 | { |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2600 | struct reclaim_state *reclaim_state = current->reclaim_state; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2601 | unsigned long nr_reclaimed, nr_scanned; |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2602 | bool reclaimable = false; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2603 | |
Johannes Weiner | 5660048 | 2012-01-12 17:17:59 -0800 | [diff] [blame] | 2604 | do { |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2605 | struct mem_cgroup *root = sc->target_mem_cgroup; |
| 2606 | struct mem_cgroup_reclaim_cookie reclaim = { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2607 | .pgdat = pgdat, |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2608 | .priority = sc->priority, |
| 2609 | }; |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2610 | unsigned long node_lru_pages = 0; |
Andrew Morton | 694fbc0 | 2013-09-24 15:27:37 -0700 | [diff] [blame] | 2611 | struct mem_cgroup *memcg; |
Johannes Weiner | 5660048 | 2012-01-12 17:17:59 -0800 | [diff] [blame] | 2612 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2613 | nr_reclaimed = sc->nr_reclaimed; |
| 2614 | nr_scanned = sc->nr_scanned; |
Konstantin Khlebnikov | f9be23d | 2012-05-29 15:07:02 -0700 | [diff] [blame] | 2615 | |
Andrew Morton | 694fbc0 | 2013-09-24 15:27:37 -0700 | [diff] [blame] | 2616 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
| 2617 | do { |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2618 | unsigned long lru_pages; |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2619 | unsigned long reclaimed; |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2620 | unsigned long scanned; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2621 | |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2622 | if (mem_cgroup_low(root, memcg)) { |
| 2623 | if (!sc->may_thrash) |
| 2624 | continue; |
| 2625 | mem_cgroup_events(memcg, MEMCG_LOW, 1); |
| 2626 | } |
| 2627 | |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2628 | reclaimed = sc->nr_reclaimed; |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2629 | scanned = sc->nr_scanned; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2630 | |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2631 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
| 2632 | node_lru_pages += lru_pages; |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2633 | |
Vladimir Davydov | b5afba2 | 2016-08-02 14:03:04 -0700 | [diff] [blame] | 2634 | if (memcg) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2635 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2636 | memcg, sc->nr_scanned - scanned, |
| 2637 | lru_pages); |
| 2638 | |
Johannes Weiner | 8e8ae64 | 2016-01-14 15:21:32 -0800 | [diff] [blame] | 2639 | /* Record the group's reclaim efficiency */ |
| 2640 | vmpressure(sc->gfp_mask, memcg, false, |
| 2641 | sc->nr_scanned - scanned, |
| 2642 | sc->nr_reclaimed - reclaimed); |
| 2643 | |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2644 | /* |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2645 | * Direct reclaim and kswapd have to scan all memory |
| 2646 | * cgroups to fulfill the overall scan target for the |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2647 | * node. |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2648 | * |
| 2649 | * Limit reclaim, on the other hand, only cares about |
| 2650 | * nr_to_reclaim pages to be reclaimed and it will |
| 2651 | * retry with decreasing priority if one round over the |
| 2652 | * whole hierarchy is not sufficient. |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2653 | */ |
Michal Hocko | a394cb8 | 2013-02-22 16:32:30 -0800 | [diff] [blame] | 2654 | if (!global_reclaim(sc) && |
| 2655 | sc->nr_reclaimed >= sc->nr_to_reclaim) { |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2656 | mem_cgroup_iter_break(root, memcg); |
| 2657 | break; |
| 2658 | } |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2659 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 2660 | |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2661 | /* |
| 2662 | * Shrink the slab caches in the same proportion that |
| 2663 | * the eligible LRU pages were scanned. |
| 2664 | */ |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2665 | if (global_reclaim(sc)) |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2666 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2667 | sc->nr_scanned - nr_scanned, |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2668 | node_lru_pages); |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2669 | |
Vinayak Menon | 57e6f44 | 2017-02-19 20:00:31 +1100 | [diff] [blame] | 2670 | /* |
| 2671 | * Record the subtree's reclaim efficiency. The reclaimed |
| 2672 | * pages from slab is excluded here because the corresponding |
| 2673 | * scanned pages is not accounted. Moreover, freeing a page |
| 2674 | * by slab shrinking depends on each slab's object population, |
| 2675 | * making the cost model (i.e. scan:free) different from that |
| 2676 | * of LRU. |
| 2677 | */ |
| 2678 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
| 2679 | sc->nr_scanned - nr_scanned, |
| 2680 | sc->nr_reclaimed - nr_reclaimed); |
| 2681 | |
Vladimir Davydov | cb731d6 | 2015-02-12 14:58:54 -0800 | [diff] [blame] | 2682 | if (reclaim_state) { |
| 2683 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
| 2684 | reclaim_state->reclaimed_slab = 0; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 2685 | } |
| 2686 | |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2687 | if (sc->nr_reclaimed - nr_reclaimed) |
| 2688 | reclaimable = true; |
| 2689 | |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 2690 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
Johannes Weiner | 9b4f98c | 2013-02-22 16:32:19 -0800 | [diff] [blame] | 2691 | sc->nr_scanned - nr_scanned, sc)); |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2692 | |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2693 | /* |
| 2694 | * Kswapd gives up on balancing particular nodes after too |
| 2695 | * many failures to reclaim anything from them and goes to |
| 2696 | * sleep. On reclaim progress, reset the failure counter. A |
| 2697 | * successful direct reclaim run will revive a dormant kswapd. |
| 2698 | */ |
| 2699 | if (reclaimable) |
| 2700 | pgdat->kswapd_failures = 0; |
| 2701 | |
Johannes Weiner | 2344d7e | 2014-08-06 16:06:15 -0700 | [diff] [blame] | 2702 | return reclaimable; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 2703 | } |
| 2704 | |
Vlastimil Babka | 53853e2 | 2014-10-09 15:27:02 -0700 | [diff] [blame] | 2705 | /* |
Vlastimil Babka | fdd4c61 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2706 | * Returns true if compaction should go ahead for a costly-order request, or |
| 2707 | * the allocation would already succeed without compaction. Return false if we |
| 2708 | * should reclaim first. |
Vlastimil Babka | 53853e2 | 2014-10-09 15:27:02 -0700 | [diff] [blame] | 2709 | */ |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2710 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2711 | { |
Mel Gorman | 31483b6 | 2016-07-28 15:45:46 -0700 | [diff] [blame] | 2712 | unsigned long watermark; |
Vlastimil Babka | fdd4c61 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2713 | enum compact_result suitable; |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2714 | |
Vlastimil Babka | fdd4c61 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2715 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
| 2716 | if (suitable == COMPACT_SUCCESS) |
| 2717 | /* Allocation should succeed already. Don't reclaim. */ |
| 2718 | return true; |
| 2719 | if (suitable == COMPACT_SKIPPED) |
| 2720 | /* Compaction cannot yet proceed. Do reclaim. */ |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2721 | return false; |
| 2722 | |
Vlastimil Babka | fdd4c61 | 2016-10-07 16:58:03 -0700 | [diff] [blame] | 2723 | /* |
| 2724 | * Compaction is already possible, but it takes time to run and there |
| 2725 | * are potentially other callers using the pages just freed. So proceed |
| 2726 | * with reclaim to make a buffer of free pages available to give |
| 2727 | * compaction a reasonable chance of completing and allocating the page. |
| 2728 | * Note that we won't actually reclaim the whole buffer in one attempt |
| 2729 | * as the target watermark in should_continue_reclaim() is lower. But if |
| 2730 | * we are already above the high+gap watermark, don't reclaim at all. |
| 2731 | */ |
| 2732 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
| 2733 | |
| 2734 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
Mel Gorman | fe4b1b2 | 2012-01-12 17:19:45 -0800 | [diff] [blame] | 2735 | } |
| 2736 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2737 | /* |
| 2738 | * This is the direct reclaim path, for page-allocating processes. We only |
| 2739 | * try to reclaim pages from zones which will satisfy the caller's allocation |
| 2740 | * request. |
| 2741 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2742 | * If a zone is deemed to be full of pinned pages then just give it a light |
| 2743 | * scan then give up on it. |
| 2744 | */ |
Michal Hocko | 0a0337e | 2016-05-20 16:57:00 -0700 | [diff] [blame] | 2745 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2746 | { |
Mel Gorman | dd1a239 | 2008-04-28 02:12:17 -0700 | [diff] [blame] | 2747 | struct zoneref *z; |
Mel Gorman | 54a6eb5 | 2008-04-28 02:12:16 -0700 | [diff] [blame] | 2748 | struct zone *zone; |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2749 | unsigned long nr_soft_reclaimed; |
| 2750 | unsigned long nr_soft_scanned; |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2751 | gfp_t orig_mask; |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2752 | pg_data_t *last_pgdat = NULL; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2753 | |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2754 | /* |
| 2755 | * If the number of buffer_heads in the machine exceeds the maximum |
| 2756 | * allowed level, force direct reclaim to scan the highmem zone as |
| 2757 | * highmem pages could be pinning lowmem pages storing buffer_heads |
| 2758 | */ |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2759 | orig_mask = sc->gfp_mask; |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2760 | if (buffer_heads_over_limit) { |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2761 | sc->gfp_mask |= __GFP_HIGHMEM; |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2762 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2763 | } |
Mel Gorman | cc715d9 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 2764 | |
Mel Gorman | d4debc6 | 2010-08-09 17:19:29 -0700 | [diff] [blame] | 2765 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2766 | sc->reclaim_idx, sc->nodemask) { |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 2767 | /* |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2768 | * Take care memory controller reclaiming has small influence |
| 2769 | * to global LRU. |
| 2770 | */ |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 2771 | if (global_reclaim(sc)) { |
Vladimir Davydov | 344736f | 2014-10-20 15:50:30 +0400 | [diff] [blame] | 2772 | if (!cpuset_zone_allowed(zone, |
| 2773 | GFP_KERNEL | __GFP_HARDWALL)) |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2774 | continue; |
Vladimir Davydov | 65ec02c | 2014-04-03 14:47:20 -0700 | [diff] [blame] | 2775 | |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2776 | /* |
| 2777 | * If we already have plenty of memory free for |
| 2778 | * compaction in this zone, don't free any more. |
| 2779 | * Even though compaction is invoked for any |
| 2780 | * non-zero order, only frequent costly order |
| 2781 | * reclamation is disruptive enough to become a |
| 2782 | * noticeable problem, like transparent huge |
| 2783 | * page allocations. |
| 2784 | */ |
| 2785 | if (IS_ENABLED(CONFIG_COMPACTION) && |
| 2786 | sc->order > PAGE_ALLOC_COSTLY_ORDER && |
Mel Gorman | 4f58833 | 2016-07-28 15:46:38 -0700 | [diff] [blame] | 2787 | compaction_ready(zone, sc)) { |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2788 | sc->compaction_ready = true; |
| 2789 | continue; |
Rik van Riel | e0887c1 | 2011-10-31 17:09:31 -0700 | [diff] [blame] | 2790 | } |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2791 | |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2792 | /* |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2793 | * Shrink each node in the zonelist once. If the |
| 2794 | * zonelist is ordered by zone (not the default) then a |
| 2795 | * node may be shrunk multiple times but in that case |
| 2796 | * the user prefers lower zones being preserved. |
| 2797 | */ |
| 2798 | if (zone->zone_pgdat == last_pgdat) |
| 2799 | continue; |
| 2800 | |
| 2801 | /* |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2802 | * This steals pages from memory cgroups over softlimit |
| 2803 | * and returns the number of reclaimed pages and |
| 2804 | * scanned pages. This works for global memory pressure |
| 2805 | * and balancing, not for a memcg's limit. |
| 2806 | */ |
| 2807 | nr_soft_scanned = 0; |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 2808 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 2809 | sc->order, sc->gfp_mask, |
| 2810 | &nr_soft_scanned); |
| 2811 | sc->nr_reclaimed += nr_soft_reclaimed; |
| 2812 | sc->nr_scanned += nr_soft_scanned; |
KAMEZAWA Hiroyuki | ac34a1a | 2011-06-27 16:18:12 -0700 | [diff] [blame] | 2813 | /* need some check for avoid more shrink_zone() */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2814 | } |
Nick Piggin | 408d854 | 2006-09-25 23:31:27 -0700 | [diff] [blame] | 2815 | |
Mel Gorman | 79dafcd | 2016-07-28 15:45:53 -0700 | [diff] [blame] | 2816 | /* See comment about same check for global reclaim above */ |
| 2817 | if (zone->zone_pgdat == last_pgdat) |
| 2818 | continue; |
| 2819 | last_pgdat = zone->zone_pgdat; |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 2820 | shrink_node(zone->zone_pgdat, sc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2821 | } |
Mel Gorman | e0c2327 | 2011-10-31 17:09:33 -0700 | [diff] [blame] | 2822 | |
Vladimir Davydov | 65ec02c | 2014-04-03 14:47:20 -0700 | [diff] [blame] | 2823 | /* |
Weijie Yang | 619d0d76 | 2014-04-07 15:36:59 -0700 | [diff] [blame] | 2824 | * Restore to original mask to avoid the impact on the caller if we |
| 2825 | * promoted it to __GFP_HIGHMEM. |
| 2826 | */ |
| 2827 | sc->gfp_mask = orig_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2828 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2829 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2830 | /* |
| 2831 | * This is the main entry point to direct page reclaim. |
| 2832 | * |
| 2833 | * If a full scan of the inactive list fails to free enough memory then we |
| 2834 | * are "out of memory" and something needs to be killed. |
| 2835 | * |
| 2836 | * If the caller is !__GFP_FS then the probability of a failure is reasonably |
| 2837 | * 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] | 2838 | * caller can't do much about. We kick the writeback threads and take explicit |
| 2839 | * naps in the hope that some of these pages can be written. But if the |
| 2840 | * allocating task holds filesystem locks which prevent writeout this might not |
| 2841 | * work, and the allocation attempt will fail. |
Nishanth Aravamudan | a41f24e | 2008-04-29 00:58:25 -0700 | [diff] [blame] | 2842 | * |
| 2843 | * returns: 0, if no pages reclaimed |
| 2844 | * else, the number of pages reclaimed |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2845 | */ |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 2846 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 2847 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2848 | { |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2849 | int initial_priority = sc->priority; |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2850 | retry: |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2851 | delayacct_freepages_start(); |
| 2852 | |
Johannes Weiner | 89b5fae | 2012-01-12 17:17:50 -0800 | [diff] [blame] | 2853 | if (global_reclaim(sc)) |
Mel Gorman | 7cc30fc | 2016-07-28 15:46:59 -0700 | [diff] [blame] | 2854 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2855 | |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 2856 | do { |
Anton Vorontsov | 70ddf63 | 2013-04-29 15:08:31 -0700 | [diff] [blame] | 2857 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
| 2858 | sc->priority); |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2859 | sc->nr_scanned = 0; |
Michal Hocko | 0a0337e | 2016-05-20 16:57:00 -0700 | [diff] [blame] | 2860 | shrink_zones(zonelist, sc); |
Mel Gorman | e0c2327 | 2011-10-31 17:09:33 -0700 | [diff] [blame] | 2861 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2862 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2863 | break; |
| 2864 | |
| 2865 | if (sc->compaction_ready) |
| 2866 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2867 | |
| 2868 | /* |
Minchan Kim | 0e50ce3 | 2013-02-22 16:35:37 -0800 | [diff] [blame] | 2869 | * If we're getting trouble reclaiming, start doing |
| 2870 | * writepage even in laptop mode. |
| 2871 | */ |
| 2872 | if (sc->priority < DEF_PRIORITY - 2) |
| 2873 | sc->may_writepage = 1; |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2874 | } while (--sc->priority >= 0); |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2875 | |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2876 | delayacct_freepages_end(); |
| 2877 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2878 | if (sc->nr_reclaimed) |
| 2879 | return sc->nr_reclaimed; |
| 2880 | |
Mel Gorman | 0cee34f | 2012-01-12 17:19:49 -0800 | [diff] [blame] | 2881 | /* Aborted reclaim to try compaction? don't OOM, then */ |
Johannes Weiner | 0b06496 | 2014-08-06 16:06:12 -0700 | [diff] [blame] | 2882 | if (sc->compaction_ready) |
Mel Gorman | 7335084 | 2012-01-12 17:19:33 -0800 | [diff] [blame] | 2883 | return 1; |
| 2884 | |
Johannes Weiner | 241994e | 2015-02-11 15:26:06 -0800 | [diff] [blame] | 2885 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
| 2886 | if (!sc->may_thrash) { |
| 2887 | sc->priority = initial_priority; |
| 2888 | sc->may_thrash = 1; |
| 2889 | goto retry; |
| 2890 | } |
| 2891 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2892 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2893 | } |
| 2894 | |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2895 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2896 | { |
| 2897 | struct zone *zone; |
| 2898 | unsigned long pfmemalloc_reserve = 0; |
| 2899 | unsigned long free_pages = 0; |
| 2900 | int i; |
| 2901 | bool wmark_ok; |
| 2902 | |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2903 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
| 2904 | return true; |
| 2905 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2906 | for (i = 0; i <= ZONE_NORMAL; i++) { |
| 2907 | zone = &pgdat->node_zones[i]; |
Johannes Weiner | cbe5956 | 2017-05-03 14:51:54 -0700 | [diff] [blame] | 2908 | if (!managed_zone(zone)) |
| 2909 | continue; |
| 2910 | |
| 2911 | if (!zone_reclaimable_pages(zone)) |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2912 | continue; |
| 2913 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2914 | pfmemalloc_reserve += min_wmark_pages(zone); |
| 2915 | free_pages += zone_page_state(zone, NR_FREE_PAGES); |
| 2916 | } |
| 2917 | |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2918 | /* If there are no reserves (unexpected config) then do not throttle */ |
| 2919 | if (!pfmemalloc_reserve) |
| 2920 | return true; |
| 2921 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2922 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
| 2923 | |
| 2924 | /* kswapd must be awake if processes are being throttled */ |
| 2925 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 2926 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2927 | (enum zone_type)ZONE_NORMAL); |
| 2928 | wake_up_interruptible(&pgdat->kswapd_wait); |
| 2929 | } |
| 2930 | |
| 2931 | return wmark_ok; |
| 2932 | } |
| 2933 | |
| 2934 | /* |
| 2935 | * Throttle direct reclaimers if backing storage is backed by the network |
| 2936 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously |
| 2937 | * depleted. kswapd will continue to make progress and wake the processes |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2938 | * when the low watermark is reached. |
| 2939 | * |
| 2940 | * Returns true if a fatal signal was delivered during throttling. If this |
| 2941 | * happens, the page allocator should not consider triggering the OOM killer. |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2942 | */ |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2943 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2944 | nodemask_t *nodemask) |
| 2945 | { |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2946 | struct zoneref *z; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2947 | struct zone *zone; |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2948 | pg_data_t *pgdat = NULL; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2949 | |
| 2950 | /* |
| 2951 | * Kernel threads should not be throttled as they may be indirectly |
| 2952 | * responsible for cleaning pages necessary for reclaim to make forward |
| 2953 | * progress. kjournald for example may enter direct reclaim while |
| 2954 | * committing a transaction where throttling it could forcing other |
| 2955 | * processes to block on log_wait_commit(). |
| 2956 | */ |
| 2957 | if (current->flags & PF_KTHREAD) |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2958 | goto out; |
| 2959 | |
| 2960 | /* |
| 2961 | * If a fatal signal is pending, this process should not throttle. |
| 2962 | * It should return quickly so it can exit and free its memory |
| 2963 | */ |
| 2964 | if (fatal_signal_pending(current)) |
| 2965 | goto out; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2966 | |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2967 | /* |
| 2968 | * Check if the pfmemalloc reserves are ok by finding the first node |
| 2969 | * with a usable ZONE_NORMAL or lower zone. The expectation is that |
| 2970 | * GFP_KERNEL will be required for allocating network buffers when |
| 2971 | * swapping over the network so ZONE_HIGHMEM is unusable. |
| 2972 | * |
| 2973 | * Throttling is based on the first usable node and throttled processes |
| 2974 | * wait on a queue until kswapd makes progress and wakes them. There |
| 2975 | * is an affinity then between processes waking up and where reclaim |
| 2976 | * progress has been made assuming the process wakes on the same node. |
| 2977 | * More importantly, processes running on remote nodes will not compete |
| 2978 | * for remote pfmemalloc reserves and processes on different nodes |
| 2979 | * should make reasonable progress. |
| 2980 | */ |
| 2981 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
Michael S. Tsirkin | 17636fa | 2015-01-26 12:58:41 -0800 | [diff] [blame] | 2982 | gfp_zone(gfp_mask), nodemask) { |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2983 | if (zone_idx(zone) > ZONE_NORMAL) |
| 2984 | continue; |
| 2985 | |
| 2986 | /* Throttle based on the first usable node */ |
| 2987 | pgdat = zone->zone_pgdat; |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 2988 | if (allow_direct_reclaim(pgdat)) |
Mel Gorman | 675becc | 2014-06-04 16:07:35 -0700 | [diff] [blame] | 2989 | goto out; |
| 2990 | break; |
| 2991 | } |
| 2992 | |
| 2993 | /* If no zone was usable by the allocation flags then do not throttle */ |
| 2994 | if (!pgdat) |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 2995 | goto out; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 2996 | |
Mel Gorman | 68243e7 | 2012-07-31 16:44:39 -0700 | [diff] [blame] | 2997 | /* Account for the throttling */ |
| 2998 | count_vm_event(PGSCAN_DIRECT_THROTTLE); |
| 2999 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3000 | /* |
| 3001 | * If the caller cannot enter the filesystem, it's possible that it |
| 3002 | * is due to the caller holding an FS lock or performing a journal |
| 3003 | * transaction in the case of a filesystem like ext[3|4]. In this case, |
| 3004 | * it is not safe to block on pfmemalloc_wait as kswapd could be |
| 3005 | * blocked waiting on the same lock. Instead, throttle for up to a |
| 3006 | * second before continuing. |
| 3007 | */ |
| 3008 | if (!(gfp_mask & __GFP_FS)) { |
| 3009 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3010 | allow_direct_reclaim(pgdat), HZ); |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3011 | |
| 3012 | goto check_pending; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3013 | } |
| 3014 | |
| 3015 | /* Throttle until kswapd wakes the process */ |
| 3016 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3017 | allow_direct_reclaim(pgdat)); |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3018 | |
| 3019 | check_pending: |
| 3020 | if (fatal_signal_pending(current)) |
| 3021 | return true; |
| 3022 | |
| 3023 | out: |
| 3024 | return false; |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3025 | } |
| 3026 | |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 3027 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 3028 | gfp_t gfp_mask, nodemask_t *nodemask) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3029 | { |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3030 | unsigned long nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3031 | struct scan_control sc = { |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 3032 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3033 | .gfp_mask = memalloc_noio_flags(gfp_mask), |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3034 | .reclaim_idx = gfp_zone(gfp_mask), |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3035 | .order = order, |
| 3036 | .nodemask = nodemask, |
| 3037 | .priority = DEF_PRIORITY, |
| 3038 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3039 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 3040 | .may_swap = 1, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3041 | }; |
| 3042 | |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3043 | /* |
Mel Gorman | 50694c2 | 2012-11-26 16:29:48 -0800 | [diff] [blame] | 3044 | * Do not enter reclaim if fatal signal was delivered while throttled. |
| 3045 | * 1 is returned so that the page allocator does not OOM kill at this |
| 3046 | * point. |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3047 | */ |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3048 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3049 | return 1; |
| 3050 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3051 | trace_mm_vmscan_direct_reclaim_begin(order, |
| 3052 | sc.may_writepage, |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3053 | sc.gfp_mask, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3054 | sc.reclaim_idx); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3055 | |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3056 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 3057 | |
| 3058 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); |
| 3059 | |
| 3060 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3061 | } |
| 3062 | |
Andrew Morton | c255a45 | 2012-07-31 16:43:02 -0700 | [diff] [blame] | 3063 | #ifdef CONFIG_MEMCG |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3064 | |
Mel Gorman | a9dd0a8 | 2016-07-28 15:46:02 -0700 | [diff] [blame] | 3065 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3066 | gfp_t gfp_mask, bool noswap, |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3067 | pg_data_t *pgdat, |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3068 | unsigned long *nr_scanned) |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3069 | { |
| 3070 | struct scan_control sc = { |
KOSAKI Motohiro | b8f5c56 | 2010-08-10 18:03:02 -0700 | [diff] [blame] | 3071 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3072 | .target_mem_cgroup = memcg, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3073 | .may_writepage = !laptop_mode, |
| 3074 | .may_unmap = 1, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3075 | .reclaim_idx = MAX_NR_ZONES - 1, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3076 | .may_swap = !noswap, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3077 | }; |
Johannes Weiner | 6b4f779 | 2014-12-12 16:56:13 -0800 | [diff] [blame] | 3078 | unsigned long lru_pages; |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3079 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3080 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
| 3081 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3082 | |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3083 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3084 | sc.may_writepage, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3085 | sc.gfp_mask, |
| 3086 | sc.reclaim_idx); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3087 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3088 | /* |
| 3089 | * NOTE: Although we can get the priority field, using it |
| 3090 | * 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] | 3091 | * if we don't reclaim here, the shrink_node from balance_pgdat |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3092 | * will pick up pages from other mem cgroup's as well. We hack |
| 3093 | * the priority and make it zero. |
| 3094 | */ |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3095 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3096 | |
| 3097 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); |
| 3098 | |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 3099 | *nr_scanned = sc.nr_scanned; |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3100 | return sc.nr_reclaimed; |
| 3101 | } |
| 3102 | |
Johannes Weiner | 72835c8 | 2012-01-12 17:18:32 -0800 | [diff] [blame] | 3103 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3104 | unsigned long nr_pages, |
KOSAKI Motohiro | a7885eb | 2009-01-07 18:08:24 -0800 | [diff] [blame] | 3105 | gfp_t gfp_mask, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3106 | bool may_swap) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3107 | { |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 3108 | struct zonelist *zonelist; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3109 | unsigned long nr_reclaimed; |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3110 | unsigned long pflags; |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3111 | int nid; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3112 | struct scan_control sc = { |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3113 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3114 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
| 3115 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3116 | .reclaim_idx = MAX_NR_ZONES - 1, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3117 | .target_mem_cgroup = memcg, |
| 3118 | .priority = DEF_PRIORITY, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3119 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3120 | .may_unmap = 1, |
Johannes Weiner | b70a2a2 | 2014-10-09 15:28:56 -0700 | [diff] [blame] | 3121 | .may_swap = may_swap, |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 3122 | }; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3123 | |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3124 | /* |
| 3125 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't |
| 3126 | * take care of from where we get pages. So the node where we start the |
| 3127 | * scan does not need to be the current node. |
| 3128 | */ |
Johannes Weiner | 72835c8 | 2012-01-12 17:18:32 -0800 | [diff] [blame] | 3129 | nid = mem_cgroup_select_victim_node(memcg); |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 3130 | |
Aneesh Kumar K.V | c9634cf | 2016-10-07 16:59:12 -0700 | [diff] [blame] | 3131 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3132 | |
| 3133 | trace_mm_vmscan_memcg_reclaim_begin(0, |
| 3134 | sc.may_writepage, |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3135 | sc.gfp_mask, |
| 3136 | sc.reclaim_idx); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3137 | |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3138 | psi_memstall_enter(&pflags); |
Johannes Weiner | 89a2848 | 2016-10-27 17:46:56 -0700 | [diff] [blame] | 3139 | current->flags |= PF_MEMALLOC; |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3140 | |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3141 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3142 | |
Johannes Weiner | 89a2848 | 2016-10-27 17:46:56 -0700 | [diff] [blame] | 3143 | current->flags &= ~PF_MEMALLOC; |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3144 | psi_memstall_leave(&pflags); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 3145 | |
| 3146 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
| 3147 | |
| 3148 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 3149 | } |
| 3150 | #endif |
| 3151 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3152 | static void age_active_anon(struct pglist_data *pgdat, |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3153 | struct scan_control *sc) |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3154 | { |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3155 | struct mem_cgroup *memcg; |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3156 | |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3157 | if (!total_swap_pages) |
| 3158 | return; |
| 3159 | |
| 3160 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
| 3161 | do { |
Mel Gorman | ef8f232 | 2016-07-28 15:46:05 -0700 | [diff] [blame] | 3162 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3163 | |
Mel Gorman | f8d1a31 | 2016-07-28 15:47:34 -0700 | [diff] [blame] | 3164 | if (inactive_list_is_low(lruvec, false, sc)) |
Konstantin Khlebnikov | 1a93be0 | 2012-05-29 15:07:01 -0700 | [diff] [blame] | 3165 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3166 | sc, LRU_ACTIVE_ANON); |
Johannes Weiner | b95a2f2 | 2012-01-12 17:18:06 -0800 | [diff] [blame] | 3167 | |
| 3168 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
| 3169 | } while (memcg); |
Johannes Weiner | f16015f | 2012-01-12 17:17:52 -0800 | [diff] [blame] | 3170 | } |
| 3171 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3172 | /* |
| 3173 | * Returns true if there is an eligible zone balanced for the request order |
| 3174 | * and classzone_idx |
| 3175 | */ |
| 3176 | 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] | 3177 | { |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3178 | int i; |
| 3179 | unsigned long mark = -1; |
| 3180 | struct zone *zone; |
Johannes Weiner | 60cefed | 2012-11-29 13:54:23 -0800 | [diff] [blame] | 3181 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3182 | for (i = 0; i <= classzone_idx; i++) { |
| 3183 | zone = pgdat->node_zones + i; |
Mel Gorman | 6256c6b4 | 2016-07-28 15:45:56 -0700 | [diff] [blame] | 3184 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3185 | if (!managed_zone(zone)) |
| 3186 | continue; |
| 3187 | |
| 3188 | mark = high_wmark_pages(zone); |
| 3189 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) |
| 3190 | return true; |
| 3191 | } |
| 3192 | |
| 3193 | /* |
| 3194 | * If a node has no populated zone within classzone_idx, it does not |
| 3195 | * need balancing by definition. This can happen if a zone-restricted |
| 3196 | * allocation tries to wake a remote kswapd. |
| 3197 | */ |
| 3198 | if (mark == -1) |
| 3199 | return true; |
| 3200 | |
| 3201 | return false; |
Johannes Weiner | 60cefed | 2012-11-29 13:54:23 -0800 | [diff] [blame] | 3202 | } |
| 3203 | |
Mel Gorman | 62cd588 | 2017-05-03 14:53:41 -0700 | [diff] [blame] | 3204 | /* Clear pgdat state for congested, dirty or under writeback. */ |
| 3205 | static void clear_pgdat_congested(pg_data_t *pgdat) |
| 3206 | { |
| 3207 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); |
| 3208 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
| 3209 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); |
| 3210 | } |
| 3211 | |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 3212 | /* |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3213 | * Prepare kswapd for sleeping. This verifies that there are no processes |
| 3214 | * waiting in throttle_direct_reclaim() and that watermarks have been met. |
| 3215 | * |
| 3216 | * Returns true if kswapd is ready to sleep |
| 3217 | */ |
Mel Gorman | d9f21d4 | 2016-07-28 15:46:41 -0700 | [diff] [blame] | 3218 | 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] | 3219 | { |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3220 | /* |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3221 | * The throttled processes are normally woken up in balance_pgdat() as |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3222 | * soon as allow_direct_reclaim() is true. But there is a potential |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3223 | * race between when kswapd checks the watermarks and a process gets |
| 3224 | * throttled. There is also a potential race if processes get |
| 3225 | * throttled, kswapd wakes, a large process exits thereby balancing the |
| 3226 | * zones, which causes kswapd to exit balance_pgdat() before reaching |
| 3227 | * the wake up checks. If kswapd is going to sleep, no process should |
| 3228 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If |
| 3229 | * the wake up is premature, processes will wake kswapd and get |
| 3230 | * throttled again. The difference from wake ups in balance_pgdat() is |
| 3231 | * that here we are under prepare_to_wait(). |
Mel Gorman | 5515061 | 2012-07-31 16:44:35 -0700 | [diff] [blame] | 3232 | */ |
Vlastimil Babka | 9e5e366 | 2015-01-08 14:32:40 -0800 | [diff] [blame] | 3233 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
| 3234 | wake_up_all(&pgdat->pfmemalloc_wait); |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 3235 | |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3236 | /* Hopeless node, leave it to direct reclaim */ |
| 3237 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
| 3238 | return true; |
| 3239 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3240 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
| 3241 | clear_pgdat_congested(pgdat); |
| 3242 | return true; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3243 | } |
| 3244 | |
Shantanu Goel | eeb5a6a | 2017-05-03 14:53:38 -0700 | [diff] [blame] | 3245 | return false; |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 3246 | } |
| 3247 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3248 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3249 | * kswapd shrinks a node of pages that are at or below the highest usable |
| 3250 | * zone that is currently unbalanced. |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3251 | * |
| 3252 | * 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] | 3253 | * reclaim or if the lack of progress was due to pages under writeback. |
| 3254 | * 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] | 3255 | */ |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3256 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
Vlastimil Babka | accf624 | 2016-03-17 14:18:15 -0700 | [diff] [blame] | 3257 | struct scan_control *sc) |
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 | struct zone *zone; |
| 3260 | int z; |
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 | /* Reclaim a number of pages proportional to the number of zones */ |
| 3263 | sc->nr_to_reclaim = 0; |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3264 | for (z = 0; z <= sc->reclaim_idx; z++) { |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3265 | zone = pgdat->node_zones + z; |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 3266 | if (!managed_zone(zone)) |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3267 | continue; |
Mel Gorman | 7c954f6 | 2013-07-03 15:01:54 -0700 | [diff] [blame] | 3268 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3269 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
Mel Gorman | 7c954f6 | 2013-07-03 15:01:54 -0700 | [diff] [blame] | 3270 | } |
| 3271 | |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3272 | /* |
| 3273 | * Historically care was taken to put equal pressure on all zones but |
| 3274 | * now pressure is applied based on node LRU order. |
| 3275 | */ |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3276 | shrink_node(pgdat, sc); |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3277 | |
| 3278 | /* |
| 3279 | * Fragmentation may mean that the system cannot be rebalanced for |
| 3280 | * high-order allocations. If twice the allocation size has been |
| 3281 | * reclaimed then recheck watermarks only at order-0 to prevent |
| 3282 | * excessive reclaim. Assume that a process requested a high-order |
| 3283 | * can direct reclaim/compact. |
| 3284 | */ |
Vlastimil Babka | 9861a62 | 2016-10-07 16:57:53 -0700 | [diff] [blame] | 3285 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3286 | sc->order = 0; |
| 3287 | |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3288 | return sc->nr_scanned >= sc->nr_to_reclaim; |
Mel Gorman | 7548536 | 2013-07-03 15:01:42 -0700 | [diff] [blame] | 3289 | } |
| 3290 | |
| 3291 | /* |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3292 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
| 3293 | * that are eligible for use by the caller until at least one zone is |
| 3294 | * balanced. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3295 | * |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3296 | * Returns the order kswapd finished reclaiming at. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3297 | * |
| 3298 | * kswapd scans the zones in the highmem->normal->dma direction. It skips |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 3299 | * 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] | 3300 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
| 3301 | * or lower is eligible for reclaim until at least one usable zone is |
| 3302 | * balanced. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3303 | */ |
Vlastimil Babka | accf624 | 2016-03-17 14:18:15 -0700 | [diff] [blame] | 3304 | 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] | 3305 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3306 | int i; |
Andrew Morton | 0608f43 | 2013-09-24 15:27:41 -0700 | [diff] [blame] | 3307 | unsigned long nr_soft_reclaimed; |
| 3308 | unsigned long nr_soft_scanned; |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3309 | unsigned long pflags; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3310 | struct zone *zone; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3311 | struct scan_control sc = { |
| 3312 | .gfp_mask = GFP_KERNEL, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3313 | .order = order, |
Mel Gorman | b8e83b9 | 2013-07-03 15:01:45 -0700 | [diff] [blame] | 3314 | .priority = DEF_PRIORITY, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3315 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f8 | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3316 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 3317 | .may_swap = 1, |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3318 | }; |
Johannes Weiner | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3319 | |
| 3320 | psi_memstall_enter(&pflags); |
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 | c40db80 | 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 | 7fcc202 | 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 | 7fcc202 | 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 | 9da1b5e | 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 | c40db80 | 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 | c40db80 | 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 | c40db80 | 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 | 3df0e59 | 2018-10-26 15:06:27 -0700 | [diff] [blame] | 3414 | psi_memstall_leave(&pflags); |
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 | 7fcc202 | 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 | eeb5a6a | 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 | 7fcc202 | 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 | 7fcc202 | 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 | |
Nick Piggin | cf40bd1 | 2009-01-21 08:12:39 +0100 | [diff] [blame] | 3545 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
| 3546 | |
Rusty Russell | 174596a | 2009-01-01 10:12:29 +1030 | [diff] [blame] | 3547 | if (!cpumask_empty(cpumask)) |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3548 | set_cpus_allowed_ptr(tsk, cpumask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3549 | current->reclaim_state = &reclaim_state; |
| 3550 | |
| 3551 | /* |
| 3552 | * Tell the memory management that we're a "memory allocator", |
| 3553 | * and that if we need more memory we should get access to it |
| 3554 | * regardless (see "__alloc_pages()"). "kswapd" should |
| 3555 | * never get caught in the normal page freeing logic. |
| 3556 | * |
| 3557 | * (Kswapd normally doesn't need memory anyway, but sometimes |
| 3558 | * you need a small amount of memory in order to be able to |
| 3559 | * page out something else, and this flag essentially protects |
| 3560 | * us from recursively trying to free more memory as we're |
| 3561 | * trying to free the first piece of memory in the first place). |
| 3562 | */ |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 3563 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
Rafael J. Wysocki | 8314418 | 2007-07-17 04:03:35 -0700 | [diff] [blame] | 3564 | set_freezable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3565 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3566 | pgdat->kswapd_order = 0; |
| 3567 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3568 | for ( ; ; ) { |
Jeff Liu | 6f6313d | 2012-12-11 16:02:48 -0800 | [diff] [blame] | 3569 | bool ret; |
Christoph Lameter | 3e1d1d2 | 2005-06-24 23:13:50 -0700 | [diff] [blame] | 3570 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3571 | alloc_order = reclaim_order = pgdat->kswapd_order; |
| 3572 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
| 3573 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3574 | kswapd_try_sleep: |
| 3575 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, |
| 3576 | classzone_idx); |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 3577 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3578 | /* Read the new order and classzone_idx */ |
| 3579 | alloc_order = reclaim_order = pgdat->kswapd_order; |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3580 | classzone_idx = kswapd_classzone_idx(pgdat, 0); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3581 | pgdat->kswapd_order = 0; |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3582 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3583 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3584 | ret = try_to_freeze(); |
| 3585 | if (kthread_should_stop()) |
| 3586 | break; |
| 3587 | |
| 3588 | /* |
| 3589 | * We can speed up thawing tasks if we don't call balance_pgdat |
| 3590 | * after returning from the refrigerator |
| 3591 | */ |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3592 | if (ret) |
| 3593 | continue; |
Mel Gorman | 1d82de6 | 2016-07-28 15:45:43 -0700 | [diff] [blame] | 3594 | |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3595 | /* |
| 3596 | * Reclaim begins at the requested order but if a high-order |
| 3597 | * reclaim fails then kswapd falls back to reclaiming for |
| 3598 | * order-0. If that happens, kswapd will consider sleeping |
| 3599 | * for the order it finished reclaiming at (reclaim_order) |
| 3600 | * but kcompactd is woken to compact for the original |
| 3601 | * request (alloc_order). |
| 3602 | */ |
Mel Gorman | e5146b1 | 2016-07-28 15:46:47 -0700 | [diff] [blame] | 3603 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
| 3604 | alloc_order); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3605 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
| 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 | lockdep_clear_current_reclaim_state(); |
| 3613 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3614 | return 0; |
| 3615 | } |
| 3616 | |
| 3617 | /* |
| 3618 | * A zone is low on free memory, so wake its kswapd task to service it. |
| 3619 | */ |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 3620 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3621 | { |
| 3622 | pg_data_t *pgdat; |
| 3623 | |
Mel Gorman | 6aa303d | 2016-09-01 16:14:55 -0700 | [diff] [blame] | 3624 | if (!managed_zone(zone)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3625 | return; |
| 3626 | |
Vladimir Davydov | 344736f | 2014-10-20 15:50:30 +0400 | [diff] [blame] | 3627 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3628 | return; |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 3629 | pgdat = zone->zone_pgdat; |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3630 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, |
| 3631 | classzone_idx); |
Mel Gorman | 38087d9 | 2016-07-28 15:45:49 -0700 | [diff] [blame] | 3632 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 3633 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3634 | return; |
Mel Gorman | e1a5563 | 2016-07-28 15:46:26 -0700 | [diff] [blame] | 3635 | |
Johannes Weiner | c40db80 | 2017-05-03 14:51:51 -0700 | [diff] [blame] | 3636 | /* Hopeless node, leave it to direct reclaim */ |
| 3637 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
| 3638 | return; |
| 3639 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3640 | if (pgdat_balanced(pgdat, order, classzone_idx)) |
| 3641 | return; |
Mel Gorman | e1a5563 | 2016-07-28 15:46:26 -0700 | [diff] [blame] | 3642 | |
Mel Gorman | 7fcc202 | 2017-05-03 14:53:45 -0700 | [diff] [blame] | 3643 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 3644 | wake_up_interruptible(&pgdat->kswapd_wait); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3645 | } |
| 3646 | |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 3647 | #ifdef CONFIG_HIBERNATION |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3648 | /* |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3649 | * 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] | 3650 | * freed pages. |
| 3651 | * |
| 3652 | * Rather than trying to age LRUs the aim is to preserve the overall |
| 3653 | * LRU order by reclaiming preferentially |
| 3654 | * inactive > active > active referenced > active mapped |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3655 | */ |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3656 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3657 | { |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3658 | struct reclaim_state reclaim_state; |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3659 | struct scan_control sc = { |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3660 | .nr_to_reclaim = nr_to_reclaim, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3661 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
Mel Gorman | b2e1875 | 2016-07-28 15:45:37 -0700 | [diff] [blame] | 3662 | .reclaim_idx = MAX_NR_ZONES - 1, |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3663 | .priority = DEF_PRIORITY, |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3664 | .may_writepage = 1, |
| 3665 | .may_unmap = 1, |
| 3666 | .may_swap = 1, |
| 3667 | .hibernation_mode = 1, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3668 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 3669 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3670 | struct task_struct *p = current; |
| 3671 | unsigned long nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3672 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3673 | p->flags |= PF_MEMALLOC; |
| 3674 | lockdep_set_current_reclaim_state(sc.gfp_mask); |
| 3675 | reclaim_state.reclaimed_slab = 0; |
| 3676 | p->reclaim_state = &reclaim_state; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3677 | |
Vladimir Davydov | 3115cd9 | 2014-04-03 14:47:22 -0700 | [diff] [blame] | 3678 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3679 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3680 | p->reclaim_state = NULL; |
| 3681 | lockdep_clear_current_reclaim_state(); |
| 3682 | p->flags &= ~PF_MEMALLOC; |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 3683 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 3684 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3685 | } |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 3686 | #endif /* CONFIG_HIBERNATION */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3687 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3688 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
| 3689 | not required for correctness. So if the last cpu in a node goes |
| 3690 | away, we get changed to run anywhere: as the first one comes back, |
| 3691 | restore their cpu bindings. */ |
Greg Kroah-Hartman | fcb35a9 | 2012-12-21 15:01:06 -0800 | [diff] [blame] | 3692 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
| 3693 | void *hcpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3694 | { |
Yasunori Goto | 58c0a4a | 2007-10-16 01:25:40 -0700 | [diff] [blame] | 3695 | int nid; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3696 | |
Rafael J. Wysocki | 8bb7844 | 2007-05-09 02:35:10 -0700 | [diff] [blame] | 3697 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
Lai Jiangshan | 48fb2e2 | 2012-12-12 13:51:43 -0800 | [diff] [blame] | 3698 | for_each_node_state(nid, N_MEMORY) { |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3699 | pg_data_t *pgdat = NODE_DATA(nid); |
Rusty Russell | a70f730 | 2009-03-13 14:49:46 +1030 | [diff] [blame] | 3700 | const struct cpumask *mask; |
| 3701 | |
| 3702 | mask = cpumask_of_node(pgdat->node_id); |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3703 | |
Rusty Russell | 3e59794 | 2009-01-01 10:12:24 +1030 | [diff] [blame] | 3704 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3705 | /* One of our CPUs online: restore mask */ |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 3706 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3707 | } |
| 3708 | } |
| 3709 | return NOTIFY_OK; |
| 3710 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3711 | |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3712 | /* |
| 3713 | * This kswapd start function will be called by init and node-hot-add. |
| 3714 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. |
| 3715 | */ |
| 3716 | int kswapd_run(int nid) |
| 3717 | { |
| 3718 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3719 | int ret = 0; |
| 3720 | |
| 3721 | if (pgdat->kswapd) |
| 3722 | return 0; |
| 3723 | |
| 3724 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); |
| 3725 | if (IS_ERR(pgdat->kswapd)) { |
| 3726 | /* failure at boot is fatal */ |
| 3727 | BUG_ON(system_state == SYSTEM_BOOTING); |
Gavin Shan | d5dc0ad | 2012-10-08 16:29:27 -0700 | [diff] [blame] | 3728 | pr_err("Failed to start kswapd on node %d\n", nid); |
| 3729 | ret = PTR_ERR(pgdat->kswapd); |
Xishi Qiu | d72515b | 2013-04-17 15:58:34 -0700 | [diff] [blame] | 3730 | pgdat->kswapd = NULL; |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3731 | } |
| 3732 | return ret; |
| 3733 | } |
| 3734 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3735 | /* |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3736 | * 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] | 3737 | * hold mem_hotplug_begin/end(). |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3738 | */ |
| 3739 | void kswapd_stop(int nid) |
| 3740 | { |
| 3741 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; |
| 3742 | |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3743 | if (kswapd) { |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3744 | kthread_stop(kswapd); |
Jiang Liu | d8adde1 | 2012-07-11 14:01:52 -0700 | [diff] [blame] | 3745 | NODE_DATA(nid)->kswapd = NULL; |
| 3746 | } |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3747 | } |
| 3748 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3749 | static int __init kswapd_init(void) |
| 3750 | { |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3751 | int nid; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3752 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3753 | swap_setup(); |
Lai Jiangshan | 48fb2e2 | 2012-12-12 13:51:43 -0800 | [diff] [blame] | 3754 | for_each_node_state(nid, N_MEMORY) |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3755 | kswapd_run(nid); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3756 | hotcpu_notifier(cpu_callback, 0); |
| 3757 | return 0; |
| 3758 | } |
| 3759 | |
| 3760 | module_init(kswapd_init) |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3761 | |
| 3762 | #ifdef CONFIG_NUMA |
| 3763 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3764 | * Node reclaim mode |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3765 | * |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3766 | * 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] | 3767 | * the watermarks. |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3768 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3769 | int node_reclaim_mode __read_mostly; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3770 | |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3771 | #define RECLAIM_OFF 0 |
Fernando Luis Vazquez Cao | 7d03431 | 2008-07-29 22:33:41 -0700 | [diff] [blame] | 3772 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3773 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3774 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3775 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3776 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3777 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
Christoph Lameter | a92f712 | 2006-02-01 03:05:32 -0800 | [diff] [blame] | 3778 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
| 3779 | * a zone. |
| 3780 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3781 | #define NODE_RECLAIM_PRIORITY 4 |
Christoph Lameter | a92f712 | 2006-02-01 03:05:32 -0800 | [diff] [blame] | 3782 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3783 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3784 | * 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] | 3785 | * occur. |
| 3786 | */ |
| 3787 | int sysctl_min_unmapped_ratio = 1; |
| 3788 | |
| 3789 | /* |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3790 | * If the number of slab pages in a zone grows beyond this percentage then |
| 3791 | * slab reclaim needs to occur. |
| 3792 | */ |
| 3793 | int sysctl_min_slab_ratio = 5; |
| 3794 | |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3795 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3796 | { |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3797 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
| 3798 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + |
| 3799 | node_page_state(pgdat, NR_ACTIVE_FILE); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3800 | |
| 3801 | /* |
| 3802 | * It's possible for there to be more file mapped pages than |
| 3803 | * accounted for by the pages on the file LRU lists because |
| 3804 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED |
| 3805 | */ |
| 3806 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; |
| 3807 | } |
| 3808 | |
| 3809 | /* 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] | 3810 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3811 | { |
Alexandru Moise | d031a15 | 2015-11-05 18:48:08 -0800 | [diff] [blame] | 3812 | unsigned long nr_pagecache_reclaimable; |
| 3813 | unsigned long delta = 0; |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3814 | |
| 3815 | /* |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3816 | * If RECLAIM_UNMAP is set, then all file pages are considered |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3817 | * potentially reclaimable. Otherwise, we have to worry about |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 3818 | * pages like swapcache and node_unmapped_file_pages() provides |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3819 | * a better estimate |
| 3820 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3821 | if (node_reclaim_mode & RECLAIM_UNMAP) |
| 3822 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3823 | else |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3824 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3825 | |
| 3826 | /* If we can't clean pages, remove dirty pages from consideration */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3827 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
| 3828 | delta += node_page_state(pgdat, NR_FILE_DIRTY); |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3829 | |
| 3830 | /* Watch for any possible underflows due to delta */ |
| 3831 | if (unlikely(delta > nr_pagecache_reclaimable)) |
| 3832 | delta = nr_pagecache_reclaimable; |
| 3833 | |
| 3834 | return nr_pagecache_reclaimable - delta; |
| 3835 | } |
| 3836 | |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3837 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3838 | * Try to free up some pages from this node through reclaim. |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3839 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3840 | 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] | 3841 | { |
Christoph Lameter | 7fb2d46 | 2006-03-22 00:08:22 -0800 | [diff] [blame] | 3842 | /* Minimum pages needed in order to stay on node */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3843 | const unsigned long nr_pages = 1 << order; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3844 | struct task_struct *p = current; |
| 3845 | struct reclaim_state reclaim_state; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3846 | struct scan_control sc = { |
Andrew Morton | 62b726c | 2013-02-22 16:32:24 -0800 | [diff] [blame] | 3847 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3848 | .gfp_mask = memalloc_noio_flags(gfp_mask), |
Johannes Weiner | bd2f619 | 2009-03-31 15:19:38 -0700 | [diff] [blame] | 3849 | .order = order, |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3850 | .priority = NODE_RECLAIM_PRIORITY, |
| 3851 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), |
| 3852 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), |
Johannes Weiner | ee814fe | 2014-08-06 16:06:19 -0700 | [diff] [blame] | 3853 | .may_swap = 1, |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3854 | .reclaim_idx = gfp_zone(gfp_mask), |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3855 | }; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3856 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3857 | cond_resched(); |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3858 | /* |
Zhihui Zhang | 95bbc0c | 2015-06-24 16:56:42 -0700 | [diff] [blame] | 3859 | * 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] | 3860 | * 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] | 3861 | * and RECLAIM_UNMAP. |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3862 | */ |
| 3863 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; |
Nick Desaulniers | c2a3e4f | 2017-07-06 15:36:50 -0700 | [diff] [blame] | 3864 | lockdep_set_current_reclaim_state(sc.gfp_mask); |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3865 | reclaim_state.reclaimed_slab = 0; |
| 3866 | p->reclaim_state = &reclaim_state; |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3867 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3868 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3869 | /* |
| 3870 | * Free memory by calling shrink zone with increasing |
| 3871 | * priorities until we have enough memory freed. |
| 3872 | */ |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3873 | do { |
Mel Gorman | 970a39a | 2016-07-28 15:46:35 -0700 | [diff] [blame] | 3874 | shrink_node(pgdat, &sc); |
Konstantin Khlebnikov | 9e3b2f8 | 2012-05-29 15:06:57 -0700 | [diff] [blame] | 3875 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3876 | } |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3877 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3878 | p->reclaim_state = NULL; |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3879 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
KOSAKI Motohiro | 76ca542 | 2010-03-05 13:41:47 -0800 | [diff] [blame] | 3880 | lockdep_clear_current_reclaim_state(); |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 3881 | return sc.nr_reclaimed >= nr_pages; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3882 | } |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3883 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3884 | 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] | 3885 | { |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3886 | int ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3887 | |
| 3888 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3889 | * Node reclaim reclaims unmapped file backed pages and |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3890 | * slab pages if we are over the defined limits. |
Christoph Lameter | 34aa133 | 2006-06-30 01:55:37 -0700 | [diff] [blame] | 3891 | * |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3892 | * A small portion of unmapped file backed pages is needed for |
| 3893 | * 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] | 3894 | * thrown out if the node is overallocated. So we do not reclaim |
| 3895 | * if less than a specified percentage of the node is used by |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3896 | * unmapped file backed pages. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3897 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3898 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
| 3899 | sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
| 3900 | return NODE_RECLAIM_FULL; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3901 | |
| 3902 | /* |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3903 | * Do not scan if the allocation should not be delayed. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3904 | */ |
Mel Gorman | d0164ad | 2015-11-06 16:28:21 -0800 | [diff] [blame] | 3905 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3906 | return NODE_RECLAIM_NOSCAN; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3907 | |
| 3908 | /* |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3909 | * 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] | 3910 | * have associated processors. This will favor the local processor |
| 3911 | * over remote processors and spread off node memory allocations |
| 3912 | * as wide as possible. |
| 3913 | */ |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3914 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
| 3915 | return NODE_RECLAIM_NOSCAN; |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3916 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3917 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
| 3918 | return NODE_RECLAIM_NOSCAN; |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3919 | |
Mel Gorman | a5f5f91 | 2016-07-28 15:46:32 -0700 | [diff] [blame] | 3920 | ret = __node_reclaim(pgdat, gfp_mask, order); |
| 3921 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3922 | |
Mel Gorman | 24cf72518 | 2009-06-16 15:33:23 -0700 | [diff] [blame] | 3923 | if (!ret) |
| 3924 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); |
| 3925 | |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3926 | return ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3927 | } |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3928 | #endif |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3929 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3930 | /* |
| 3931 | * page_evictable - test whether a page is evictable |
| 3932 | * @page: the page to test |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3933 | * |
| 3934 | * 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] | 3935 | * lists vs unevictable list. |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3936 | * |
| 3937 | * Reasons page might not be evictable: |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3938 | * (1) page's mapping marked unevictable |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 3939 | * (2) page is part of an mlocked VMA |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3940 | * |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3941 | */ |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 3942 | int page_evictable(struct page *page) |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3943 | { |
Huang Ying | d7f4e94 | 2018-04-05 16:23:20 -0700 | [diff] [blame] | 3944 | int ret; |
| 3945 | |
| 3946 | /* Prevent address_space of inode and swap cache from being freed */ |
| 3947 | rcu_read_lock(); |
| 3948 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
| 3949 | rcu_read_unlock(); |
| 3950 | return ret; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3951 | } |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3952 | |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 3953 | #ifdef CONFIG_SHMEM |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3954 | /** |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3955 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
| 3956 | * @pages: array of pages to check |
| 3957 | * @nr_pages: number of pages to check |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3958 | * |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3959 | * Checks pages for evictability and moves them to the appropriate lru list. |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 3960 | * |
| 3961 | * This function is only used for SysV IPC SHM_UNLOCK. |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3962 | */ |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3963 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3964 | { |
Johannes Weiner | 925b767 | 2012-01-12 17:18:15 -0800 | [diff] [blame] | 3965 | struct lruvec *lruvec; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3966 | struct pglist_data *pgdat = NULL; |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3967 | int pgscanned = 0; |
| 3968 | int pgrescued = 0; |
| 3969 | int i; |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3970 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3971 | for (i = 0; i < nr_pages; i++) { |
| 3972 | struct page *page = pages[i]; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3973 | struct pglist_data *pagepgdat = page_pgdat(page); |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3974 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3975 | pgscanned++; |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3976 | if (pagepgdat != pgdat) { |
| 3977 | if (pgdat) |
| 3978 | spin_unlock_irq(&pgdat->lru_lock); |
| 3979 | pgdat = pagepgdat; |
| 3980 | spin_lock_irq(&pgdat->lru_lock); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3981 | } |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3982 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3983 | |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3984 | if (!PageLRU(page) || !PageUnevictable(page)) |
| 3985 | continue; |
| 3986 | |
Hugh Dickins | 39b5f29 | 2012-10-08 16:33:18 -0700 | [diff] [blame] | 3987 | if (page_evictable(page)) { |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3988 | enum lru_list lru = page_lru_base_type(page); |
| 3989 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 3990 | VM_BUG_ON_PAGE(PageActive(page), page); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3991 | ClearPageUnevictable(page); |
Hugh Dickins | fa9add6 | 2012-05-29 15:07:09 -0700 | [diff] [blame] | 3992 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
| 3993 | add_page_to_lru_list(page, lruvec, lru); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3994 | pgrescued++; |
| 3995 | } |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3996 | } |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3997 | |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 3998 | if (pgdat) { |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 3999 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
| 4000 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); |
Mel Gorman | 785b99f | 2016-07-28 15:47:23 -0700 | [diff] [blame] | 4001 | spin_unlock_irq(&pgdat->lru_lock); |
Hugh Dickins | 2451326 | 2012-01-20 14:34:21 -0800 | [diff] [blame] | 4002 | } |
Hugh Dickins | 85046579 | 2012-01-20 14:34:19 -0800 | [diff] [blame] | 4003 | } |
| 4004 | #endif /* CONFIG_SHMEM */ |