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 | |
| 14 | #include <linux/mm.h> |
| 15 | #include <linux/module.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 16 | #include <linux/gfp.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | #include <linux/kernel_stat.h> |
| 18 | #include <linux/swap.h> |
| 19 | #include <linux/pagemap.h> |
| 20 | #include <linux/init.h> |
| 21 | #include <linux/highmem.h> |
Andrew Morton | e129b5c | 2006-09-27 01:50:00 -0700 | [diff] [blame] | 22 | #include <linux/vmstat.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 23 | #include <linux/file.h> |
| 24 | #include <linux/writeback.h> |
| 25 | #include <linux/blkdev.h> |
| 26 | #include <linux/buffer_head.h> /* for try_to_release_page(), |
| 27 | buffer_heads_over_limit */ |
| 28 | #include <linux/mm_inline.h> |
| 29 | #include <linux/pagevec.h> |
| 30 | #include <linux/backing-dev.h> |
| 31 | #include <linux/rmap.h> |
| 32 | #include <linux/topology.h> |
| 33 | #include <linux/cpu.h> |
| 34 | #include <linux/cpuset.h> |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 35 | #include <linux/compaction.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 36 | #include <linux/notifier.h> |
| 37 | #include <linux/rwsem.h> |
Rafael J. Wysocki | 248a030 | 2006-03-22 00:09:04 -0800 | [diff] [blame] | 38 | #include <linux/delay.h> |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 39 | #include <linux/kthread.h> |
Nigel Cunningham | 7dfb710 | 2006-12-06 20:34:23 -0800 | [diff] [blame] | 40 | #include <linux/freezer.h> |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 41 | #include <linux/memcontrol.h> |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 42 | #include <linux/delayacct.h> |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 43 | #include <linux/sysctl.h> |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 44 | #include <linux/oom.h> |
Linus Torvalds | 268bb0c | 2011-05-20 12:50:29 -0700 | [diff] [blame] | 45 | #include <linux/prefetch.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 46 | |
| 47 | #include <asm/tlbflush.h> |
| 48 | #include <asm/div64.h> |
| 49 | |
| 50 | #include <linux/swapops.h> |
| 51 | |
Nick Piggin | 0f8053a | 2006-03-22 00:08:33 -0800 | [diff] [blame] | 52 | #include "internal.h" |
| 53 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 54 | #define CREATE_TRACE_POINTS |
| 55 | #include <trace/events/vmscan.h> |
| 56 | |
Mel Gorman | ee64fc9 | 2011-01-13 15:45:55 -0800 | [diff] [blame] | 57 | /* |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 58 | * reclaim_mode determines how the inactive list is shrunk |
| 59 | * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages |
| 60 | * RECLAIM_MODE_ASYNC: Do not block |
| 61 | * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback |
| 62 | * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference |
Mel Gorman | ee64fc9 | 2011-01-13 15:45:55 -0800 | [diff] [blame] | 63 | * page from the LRU and reclaim all pages within a |
| 64 | * naturally aligned range |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 65 | * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 66 | * order-0 pages and then compact the zone |
Mel Gorman | ee64fc9 | 2011-01-13 15:45:55 -0800 | [diff] [blame] | 67 | */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 68 | typedef unsigned __bitwise__ reclaim_mode_t; |
| 69 | #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u) |
| 70 | #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u) |
| 71 | #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u) |
| 72 | #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u) |
| 73 | #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u) |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 74 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 75 | struct scan_control { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 76 | /* Incremented by the number of inactive pages that were scanned */ |
| 77 | unsigned long nr_scanned; |
| 78 | |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 79 | /* Number of pages freed so far during a call to shrink_zones() */ |
| 80 | unsigned long nr_reclaimed; |
| 81 | |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 82 | /* How many pages shrink_list() should reclaim */ |
| 83 | unsigned long nr_to_reclaim; |
| 84 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 85 | unsigned long hibernation_mode; |
| 86 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 87 | /* This context's GFP mask */ |
Al Viro | 6daa0e2 | 2005-10-21 03:18:50 -0400 | [diff] [blame] | 88 | gfp_t gfp_mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | |
| 90 | int may_writepage; |
| 91 | |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 92 | /* Can mapped pages be reclaimed? */ |
| 93 | int may_unmap; |
Christoph Lameter | f1fd106 | 2006-01-18 17:42:30 -0800 | [diff] [blame] | 94 | |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 95 | /* Can pages be swapped as part of reclaim? */ |
| 96 | int may_swap; |
| 97 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 98 | int order; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 99 | |
KOSAKI Motohiro | 5f53e76 | 2010-05-24 14:32:37 -0700 | [diff] [blame] | 100 | /* |
Nikanth Karthikesan | 415b54e | 2010-08-17 15:39:09 +0530 | [diff] [blame] | 101 | * Intend to reclaim enough continuous memory rather than reclaim |
| 102 | * enough amount of memory. i.e, mode for high order allocation. |
KOSAKI Motohiro | 5f53e76 | 2010-05-24 14:32:37 -0700 | [diff] [blame] | 103 | */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 104 | reclaim_mode_t reclaim_mode; |
KOSAKI Motohiro | 5f53e76 | 2010-05-24 14:32:37 -0700 | [diff] [blame] | 105 | |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 106 | /* Which cgroup do we reclaim from */ |
| 107 | struct mem_cgroup *mem_cgroup; |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 108 | struct memcg_scanrecord *memcg_record; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 109 | |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 110 | /* |
| 111 | * Nodemask of nodes allowed by the caller. If NULL, all nodes |
| 112 | * are scanned. |
| 113 | */ |
| 114 | nodemask_t *nodemask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 115 | }; |
| 116 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 117 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
| 118 | |
| 119 | #ifdef ARCH_HAS_PREFETCH |
| 120 | #define prefetch_prev_lru_page(_page, _base, _field) \ |
| 121 | do { \ |
| 122 | if ((_page)->lru.prev != _base) { \ |
| 123 | struct page *prev; \ |
| 124 | \ |
| 125 | prev = lru_to_page(&(_page->lru)); \ |
| 126 | prefetch(&prev->_field); \ |
| 127 | } \ |
| 128 | } while (0) |
| 129 | #else |
| 130 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) |
| 131 | #endif |
| 132 | |
| 133 | #ifdef ARCH_HAS_PREFETCHW |
| 134 | #define prefetchw_prev_lru_page(_page, _base, _field) \ |
| 135 | do { \ |
| 136 | if ((_page)->lru.prev != _base) { \ |
| 137 | struct page *prev; \ |
| 138 | \ |
| 139 | prev = lru_to_page(&(_page->lru)); \ |
| 140 | prefetchw(&prev->_field); \ |
| 141 | } \ |
| 142 | } while (0) |
| 143 | #else |
| 144 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) |
| 145 | #endif |
| 146 | |
| 147 | /* |
| 148 | * From 0 .. 100. Higher means more swappy. |
| 149 | */ |
| 150 | int vm_swappiness = 60; |
Andrew Morton | bd1e22b | 2006-06-23 02:03:47 -0700 | [diff] [blame] | 151 | long vm_total_pages; /* The total number of pages which the VM controls */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 152 | |
| 153 | static LIST_HEAD(shrinker_list); |
| 154 | static DECLARE_RWSEM(shrinker_rwsem); |
| 155 | |
Balbir Singh | 00f0b82 | 2008-03-04 14:28:39 -0800 | [diff] [blame] | 156 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 157 | #define scanning_global_lru(sc) (!(sc)->mem_cgroup) |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 158 | #else |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 159 | #define scanning_global_lru(sc) (1) |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 160 | #endif |
| 161 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 162 | static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, |
| 163 | struct scan_control *sc) |
| 164 | { |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 165 | if (!scanning_global_lru(sc)) |
KOSAKI Motohiro | 3e2f41f | 2009-01-07 18:08:20 -0800 | [diff] [blame] | 166 | return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); |
| 167 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 168 | return &zone->reclaim_stat; |
| 169 | } |
| 170 | |
Vincent Li | 0b21767 | 2009-09-21 17:03:09 -0700 | [diff] [blame] | 171 | static unsigned long zone_nr_lru_pages(struct zone *zone, |
| 172 | struct scan_control *sc, enum lru_list lru) |
KOSAKI Motohiro | c9f299d | 2009-01-07 18:08:16 -0800 | [diff] [blame] | 173 | { |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 174 | if (!scanning_global_lru(sc)) |
KAMEZAWA Hiroyuki | bb2a0de | 2011-07-26 16:08:22 -0700 | [diff] [blame] | 175 | return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, |
| 176 | zone_to_nid(zone), zone_idx(zone), BIT(lru)); |
KOSAKI Motohiro | a3d8e05 | 2009-01-07 18:08:19 -0800 | [diff] [blame] | 177 | |
KOSAKI Motohiro | c9f299d | 2009-01-07 18:08:16 -0800 | [diff] [blame] | 178 | return zone_page_state(zone, NR_LRU_BASE + lru); |
| 179 | } |
| 180 | |
| 181 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 182 | /* |
| 183 | * Add a shrinker callback to be called from the vm |
| 184 | */ |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 185 | void register_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 186 | { |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 187 | shrinker->nr = 0; |
| 188 | down_write(&shrinker_rwsem); |
| 189 | list_add_tail(&shrinker->list, &shrinker_list); |
| 190 | up_write(&shrinker_rwsem); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 191 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 192 | EXPORT_SYMBOL(register_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 193 | |
| 194 | /* |
| 195 | * Remove one |
| 196 | */ |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 197 | void unregister_shrinker(struct shrinker *shrinker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 198 | { |
| 199 | down_write(&shrinker_rwsem); |
| 200 | list_del(&shrinker->list); |
| 201 | up_write(&shrinker_rwsem); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 202 | } |
Rusty Russell | 8e1f936 | 2007-07-17 04:03:17 -0700 | [diff] [blame] | 203 | EXPORT_SYMBOL(unregister_shrinker); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 204 | |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 205 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
| 206 | struct shrink_control *sc, |
| 207 | unsigned long nr_to_scan) |
| 208 | { |
| 209 | sc->nr_to_scan = nr_to_scan; |
| 210 | return (*shrinker->shrink)(shrinker, sc); |
| 211 | } |
| 212 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 213 | #define SHRINK_BATCH 128 |
| 214 | /* |
| 215 | * Call the shrink functions to age shrinkable caches |
| 216 | * |
| 217 | * Here we assume it costs one seek to replace a lru page and that it also |
| 218 | * takes a seek to recreate a cache object. With this in mind we age equal |
| 219 | * percentages of the lru and ageable caches. This should balance the seeks |
| 220 | * generated by these structures. |
| 221 | * |
Simon Arlott | 183ff22 | 2007-10-20 01:27:18 +0200 | [diff] [blame] | 222 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 223 | * slab to avoid swapping. |
| 224 | * |
| 225 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. |
| 226 | * |
| 227 | * `lru_pages' represents the number of on-LRU pages in all the zones which |
| 228 | * are eligible for the caller's allocation attempt. It is used for balancing |
| 229 | * slab reclaim versus page reclaim. |
akpm@osdl.org | b15e090 | 2005-06-21 17:14:35 -0700 | [diff] [blame] | 230 | * |
| 231 | * Returns the number of slab objects which we shrunk. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 232 | */ |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 233 | unsigned long shrink_slab(struct shrink_control *shrink, |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 234 | unsigned long nr_pages_scanned, |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 235 | unsigned long lru_pages) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | { |
| 237 | struct shrinker *shrinker; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 238 | unsigned long ret = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 240 | if (nr_pages_scanned == 0) |
| 241 | nr_pages_scanned = SWAP_CLUSTER_MAX; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 242 | |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 243 | if (!down_read_trylock(&shrinker_rwsem)) { |
| 244 | /* Assume we'll be able to shrink next time */ |
| 245 | ret = 1; |
| 246 | goto out; |
| 247 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 248 | |
| 249 | list_for_each_entry(shrinker, &shrinker_list, list) { |
| 250 | unsigned long long delta; |
| 251 | unsigned long total_scan; |
Dave Chinner | 7f8275d | 2010-07-19 14:56:17 +1000 | [diff] [blame] | 252 | unsigned long max_pass; |
Dave Chinner | 0957607 | 2011-07-08 14:14:34 +1000 | [diff] [blame] | 253 | int shrink_ret = 0; |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 254 | long nr; |
| 255 | long new_nr; |
Dave Chinner | e9299f5 | 2011-07-08 14:14:37 +1000 | [diff] [blame] | 256 | long batch_size = shrinker->batch ? shrinker->batch |
| 257 | : SHRINK_BATCH; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 258 | |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 259 | /* |
| 260 | * copy the current shrinker scan count into a local variable |
| 261 | * and zero it so that other concurrent shrinker invocations |
| 262 | * don't also do this scanning work. |
| 263 | */ |
| 264 | do { |
| 265 | nr = shrinker->nr; |
| 266 | } while (cmpxchg(&shrinker->nr, nr, 0) != nr); |
| 267 | |
| 268 | total_scan = nr; |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 269 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
| 270 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
Andrea Arcangeli | ea164d7 | 2005-11-28 13:44:15 -0800 | [diff] [blame] | 271 | delta *= max_pass; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 272 | do_div(delta, lru_pages + 1); |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 273 | total_scan += delta; |
| 274 | if (total_scan < 0) { |
David Rientjes | 88c3bd7 | 2009-03-31 15:23:29 -0700 | [diff] [blame] | 275 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
| 276 | "delete nr=%ld\n", |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 277 | shrinker->shrink, total_scan); |
| 278 | total_scan = max_pass; |
Andrea Arcangeli | ea164d7 | 2005-11-28 13:44:15 -0800 | [diff] [blame] | 279 | } |
| 280 | |
| 281 | /* |
Dave Chinner | 3567b59 | 2011-07-08 14:14:36 +1000 | [diff] [blame] | 282 | * We need to avoid excessive windup on filesystem shrinkers |
| 283 | * due to large numbers of GFP_NOFS allocations causing the |
| 284 | * shrinkers to return -1 all the time. This results in a large |
| 285 | * nr being built up so when a shrink that can do some work |
| 286 | * comes along it empties the entire cache due to nr >>> |
| 287 | * max_pass. This is bad for sustaining a working set in |
| 288 | * memory. |
| 289 | * |
| 290 | * Hence only allow the shrinker to scan the entire cache when |
| 291 | * a large delta change is calculated directly. |
| 292 | */ |
| 293 | if (delta < max_pass / 4) |
| 294 | total_scan = min(total_scan, max_pass / 2); |
| 295 | |
| 296 | /* |
Andrea Arcangeli | ea164d7 | 2005-11-28 13:44:15 -0800 | [diff] [blame] | 297 | * Avoid risking looping forever due to too large nr value: |
| 298 | * never try to free more than twice the estimate number of |
| 299 | * freeable entries. |
| 300 | */ |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 301 | if (total_scan > max_pass * 2) |
| 302 | total_scan = max_pass * 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 303 | |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 304 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
Dave Chinner | 0957607 | 2011-07-08 14:14:34 +1000 | [diff] [blame] | 305 | nr_pages_scanned, lru_pages, |
| 306 | max_pass, delta, total_scan); |
| 307 | |
Dave Chinner | e9299f5 | 2011-07-08 14:14:37 +1000 | [diff] [blame] | 308 | while (total_scan >= batch_size) { |
akpm@osdl.org | b15e090 | 2005-06-21 17:14:35 -0700 | [diff] [blame] | 309 | int nr_before; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 310 | |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 311 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
| 312 | shrink_ret = do_shrinker_shrink(shrinker, shrink, |
Dave Chinner | e9299f5 | 2011-07-08 14:14:37 +1000 | [diff] [blame] | 313 | batch_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 314 | if (shrink_ret == -1) |
| 315 | break; |
akpm@osdl.org | b15e090 | 2005-06-21 17:14:35 -0700 | [diff] [blame] | 316 | if (shrink_ret < nr_before) |
| 317 | ret += nr_before - shrink_ret; |
Dave Chinner | e9299f5 | 2011-07-08 14:14:37 +1000 | [diff] [blame] | 318 | count_vm_events(SLABS_SCANNED, batch_size); |
| 319 | total_scan -= batch_size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 320 | |
| 321 | cond_resched(); |
| 322 | } |
| 323 | |
Dave Chinner | acf92b4 | 2011-07-08 14:14:35 +1000 | [diff] [blame] | 324 | /* |
| 325 | * move the unused scan count back into the shrinker in a |
| 326 | * manner that handles concurrent updates. If we exhausted the |
| 327 | * scan, there is no need to do an update. |
| 328 | */ |
| 329 | do { |
| 330 | nr = shrinker->nr; |
| 331 | new_nr = total_scan + nr; |
| 332 | if (total_scan <= 0) |
| 333 | break; |
| 334 | } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr); |
| 335 | |
| 336 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 337 | } |
| 338 | up_read(&shrinker_rwsem); |
Minchan Kim | f06590b | 2011-05-24 17:11:11 -0700 | [diff] [blame] | 339 | out: |
| 340 | cond_resched(); |
akpm@osdl.org | b15e090 | 2005-06-21 17:14:35 -0700 | [diff] [blame] | 341 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 342 | } |
| 343 | |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 344 | static void set_reclaim_mode(int priority, struct scan_control *sc, |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 345 | bool sync) |
| 346 | { |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 347 | reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 348 | |
| 349 | /* |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 350 | * Initially assume we are entering either lumpy reclaim or |
| 351 | * reclaim/compaction.Depending on the order, we will either set the |
| 352 | * sync mode or just reclaim order-0 pages later. |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 353 | */ |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 354 | if (COMPACTION_BUILD) |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 355 | sc->reclaim_mode = RECLAIM_MODE_COMPACTION; |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 356 | else |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 357 | sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 358 | |
| 359 | /* |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 360 | * Avoid using lumpy reclaim or reclaim/compaction if possible by |
| 361 | * restricting when its set to either costly allocations or when |
| 362 | * under memory pressure |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 363 | */ |
| 364 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 365 | sc->reclaim_mode |= syncmode; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 366 | else if (sc->order && priority < DEF_PRIORITY - 2) |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 367 | sc->reclaim_mode |= syncmode; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 368 | else |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 369 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 370 | } |
| 371 | |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 372 | static void reset_reclaim_mode(struct scan_control *sc) |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 373 | { |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 374 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 375 | } |
| 376 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 377 | static inline int is_page_cache_freeable(struct page *page) |
| 378 | { |
Johannes Weiner | ceddc3a | 2009-09-21 17:03:00 -0700 | [diff] [blame] | 379 | /* |
| 380 | * A freeable page cache page is referenced only by the caller |
| 381 | * that isolated the page, the page cache radix tree and |
| 382 | * optional buffer heads at page->private. |
| 383 | */ |
Johannes Weiner | edcf474 | 2009-09-21 17:02:59 -0700 | [diff] [blame] | 384 | return page_count(page) - page_has_private(page) == 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 385 | } |
| 386 | |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 387 | static int may_write_to_queue(struct backing_dev_info *bdi, |
| 388 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 389 | { |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 390 | if (current->flags & PF_SWAPWRITE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 391 | return 1; |
| 392 | if (!bdi_write_congested(bdi)) |
| 393 | return 1; |
| 394 | if (bdi == current->backing_dev_info) |
| 395 | return 1; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 396 | |
| 397 | /* lumpy reclaim for hugepage often need a lot of write */ |
| 398 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) |
| 399 | return 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 400 | return 0; |
| 401 | } |
| 402 | |
| 403 | /* |
| 404 | * We detected a synchronous write error writing a page out. Probably |
| 405 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
| 406 | * fsync(), msync() or close(). |
| 407 | * |
| 408 | * The tricky part is that after writepage we cannot touch the mapping: nothing |
| 409 | * prevents it from being freed up. But we have a ref on the page and once |
| 410 | * that page is locked, the mapping is pinned. |
| 411 | * |
| 412 | * We're allowed to run sleeping lock_page() here because we know the caller has |
| 413 | * __GFP_FS. |
| 414 | */ |
| 415 | static void handle_write_error(struct address_space *mapping, |
| 416 | struct page *page, int error) |
| 417 | { |
Jens Axboe | 7eaceac | 2011-03-10 08:52:07 +0100 | [diff] [blame] | 418 | lock_page(page); |
Guillaume Chazarain | 3e9f45b | 2007-05-08 00:23:25 -0700 | [diff] [blame] | 419 | if (page_mapping(page) == mapping) |
| 420 | mapping_set_error(mapping, error); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 421 | unlock_page(page); |
| 422 | } |
| 423 | |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 424 | /* possible outcome of pageout() */ |
| 425 | typedef enum { |
| 426 | /* failed to write page out, page is locked */ |
| 427 | PAGE_KEEP, |
| 428 | /* move page to the active list, page is locked */ |
| 429 | PAGE_ACTIVATE, |
| 430 | /* page has been sent to the disk successfully, page is unlocked */ |
| 431 | PAGE_SUCCESS, |
| 432 | /* page is clean and locked */ |
| 433 | PAGE_CLEAN, |
| 434 | } pageout_t; |
| 435 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 436 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 437 | * pageout is called by shrink_page_list() for each dirty page. |
| 438 | * Calls ->writepage(). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 439 | */ |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 440 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 441 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 442 | { |
| 443 | /* |
| 444 | * If the page is dirty, only perform writeback if that write |
| 445 | * will be non-blocking. To prevent this allocation from being |
| 446 | * stalled by pagecache activity. But note that there may be |
| 447 | * stalls if we need to run get_block(). We could test |
| 448 | * PagePrivate for that. |
| 449 | * |
Vincent Li | 6aceb53 | 2009-12-14 17:58:49 -0800 | [diff] [blame] | 450 | * If this process is currently in __generic_file_aio_write() against |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 451 | * this page's queue, we can perform writeback even if that |
| 452 | * will block. |
| 453 | * |
| 454 | * If the page is swapcache, write it back even if that would |
| 455 | * block, for some throttling. This happens by accident, because |
| 456 | * swap_backing_dev_info is bust: it doesn't reflect the |
| 457 | * congestion state of the swapdevs. Easy to fix, if needed. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 458 | */ |
| 459 | if (!is_page_cache_freeable(page)) |
| 460 | return PAGE_KEEP; |
| 461 | if (!mapping) { |
| 462 | /* |
| 463 | * Some data journaling orphaned pages can have |
| 464 | * page->mapping == NULL while being dirty with clean buffers. |
| 465 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 466 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 467 | if (try_to_free_buffers(page)) { |
| 468 | ClearPageDirty(page); |
Harvey Harrison | d40cee2 | 2008-04-30 00:55:07 -0700 | [diff] [blame] | 469 | printk("%s: orphaned page\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 470 | return PAGE_CLEAN; |
| 471 | } |
| 472 | } |
| 473 | return PAGE_KEEP; |
| 474 | } |
| 475 | if (mapping->a_ops->writepage == NULL) |
| 476 | return PAGE_ACTIVATE; |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 477 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 478 | return PAGE_KEEP; |
| 479 | |
| 480 | if (clear_page_dirty_for_io(page)) { |
| 481 | int res; |
| 482 | struct writeback_control wbc = { |
| 483 | .sync_mode = WB_SYNC_NONE, |
| 484 | .nr_to_write = SWAP_CLUSTER_MAX, |
OGAWA Hirofumi | 111ebb6 | 2006-06-23 02:03:26 -0700 | [diff] [blame] | 485 | .range_start = 0, |
| 486 | .range_end = LLONG_MAX, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 487 | .for_reclaim = 1, |
| 488 | }; |
| 489 | |
| 490 | SetPageReclaim(page); |
| 491 | res = mapping->a_ops->writepage(page, &wbc); |
| 492 | if (res < 0) |
| 493 | handle_write_error(mapping, page, res); |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 494 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 495 | ClearPageReclaim(page); |
| 496 | return PAGE_ACTIVATE; |
| 497 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 498 | |
| 499 | /* |
| 500 | * Wait on writeback if requested to. This happens when |
| 501 | * direct reclaiming a large contiguous area and the |
| 502 | * first attempt to free a range of pages fails. |
| 503 | */ |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 504 | if (PageWriteback(page) && |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 505 | (sc->reclaim_mode & RECLAIM_MODE_SYNC)) |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 506 | wait_on_page_writeback(page); |
| 507 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 508 | if (!PageWriteback(page)) { |
| 509 | /* synchronous write or broken a_ops? */ |
| 510 | ClearPageReclaim(page); |
| 511 | } |
Mel Gorman | 755f022 | 2010-08-09 17:19:18 -0700 | [diff] [blame] | 512 | trace_mm_vmscan_writepage(page, |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 513 | trace_reclaim_flags(page, sc->reclaim_mode)); |
Andrew Morton | e129b5c | 2006-09-27 01:50:00 -0700 | [diff] [blame] | 514 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 515 | return PAGE_SUCCESS; |
| 516 | } |
| 517 | |
| 518 | return PAGE_CLEAN; |
| 519 | } |
| 520 | |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 521 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 522 | * Same as remove_mapping, but if the page is removed from the mapping, it |
| 523 | * gets returned with a refcount of 0. |
Andrew Morton | a649fd9 | 2006-10-17 00:09:36 -0700 | [diff] [blame] | 524 | */ |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 525 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 526 | { |
Nick Piggin | 28e4d96 | 2006-09-25 23:31:23 -0700 | [diff] [blame] | 527 | BUG_ON(!PageLocked(page)); |
| 528 | BUG_ON(mapping != page_mapping(page)); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 529 | |
Nick Piggin | 19fd623 | 2008-07-25 19:45:32 -0700 | [diff] [blame] | 530 | spin_lock_irq(&mapping->tree_lock); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 531 | /* |
Nick Piggin | 0fd0e6b | 2006-09-27 01:50:02 -0700 | [diff] [blame] | 532 | * The non racy check for a busy page. |
| 533 | * |
| 534 | * Must be careful with the order of the tests. When someone has |
| 535 | * a ref to the page, it may be possible that they dirty it then |
| 536 | * drop the reference. So if PageDirty is tested before page_count |
| 537 | * here, then the following race may occur: |
| 538 | * |
| 539 | * get_user_pages(&page); |
| 540 | * [user mapping goes away] |
| 541 | * write_to(page); |
| 542 | * !PageDirty(page) [good] |
| 543 | * SetPageDirty(page); |
| 544 | * put_page(page); |
| 545 | * !page_count(page) [good, discard it] |
| 546 | * |
| 547 | * [oops, our write_to data is lost] |
| 548 | * |
| 549 | * Reversing the order of the tests ensures such a situation cannot |
| 550 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags |
| 551 | * load is not satisfied before that of page->_count. |
| 552 | * |
| 553 | * Note that if SetPageDirty is always performed via set_page_dirty, |
| 554 | * and thus under tree_lock, then this ordering is not required. |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 555 | */ |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 556 | if (!page_freeze_refs(page, 2)) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 557 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 558 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
| 559 | if (unlikely(PageDirty(page))) { |
| 560 | page_unfreeze_refs(page, 2); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 561 | goto cannot_free; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 562 | } |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 563 | |
| 564 | if (PageSwapCache(page)) { |
| 565 | swp_entry_t swap = { .val = page_private(page) }; |
| 566 | __delete_from_swap_cache(page); |
Nick Piggin | 19fd623 | 2008-07-25 19:45:32 -0700 | [diff] [blame] | 567 | spin_unlock_irq(&mapping->tree_lock); |
KAMEZAWA Hiroyuki | cb4b86b | 2009-06-16 15:32:52 -0700 | [diff] [blame] | 568 | swapcache_free(swap, page); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 569 | } else { |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 570 | void (*freepage)(struct page *); |
| 571 | |
| 572 | freepage = mapping->a_ops->freepage; |
| 573 | |
Minchan Kim | e64a782 | 2011-03-22 16:32:44 -0700 | [diff] [blame] | 574 | __delete_from_page_cache(page); |
Nick Piggin | 19fd623 | 2008-07-25 19:45:32 -0700 | [diff] [blame] | 575 | spin_unlock_irq(&mapping->tree_lock); |
Daisuke Nishimura | e767e05 | 2009-05-28 14:34:28 -0700 | [diff] [blame] | 576 | mem_cgroup_uncharge_cache_page(page); |
Linus Torvalds | 6072d13 | 2010-12-01 13:35:19 -0500 | [diff] [blame] | 577 | |
| 578 | if (freepage != NULL) |
| 579 | freepage(page); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 580 | } |
| 581 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 582 | return 1; |
| 583 | |
| 584 | cannot_free: |
Nick Piggin | 19fd623 | 2008-07-25 19:45:32 -0700 | [diff] [blame] | 585 | spin_unlock_irq(&mapping->tree_lock); |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 586 | return 0; |
| 587 | } |
| 588 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 589 | /* |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 590 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if |
| 591 | * someone else has a ref on the page, abort and return 0. If it was |
| 592 | * successfully detached, return 1. Assumes the caller has a single ref on |
| 593 | * this page. |
| 594 | */ |
| 595 | int remove_mapping(struct address_space *mapping, struct page *page) |
| 596 | { |
| 597 | if (__remove_mapping(mapping, page)) { |
| 598 | /* |
| 599 | * Unfreezing the refcount with 1 rather than 2 effectively |
| 600 | * drops the pagecache ref for us without requiring another |
| 601 | * atomic operation. |
| 602 | */ |
| 603 | page_unfreeze_refs(page, 1); |
| 604 | return 1; |
| 605 | } |
| 606 | return 0; |
| 607 | } |
| 608 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 609 | /** |
| 610 | * putback_lru_page - put previously isolated page onto appropriate LRU list |
| 611 | * @page: page to be put back to appropriate lru list |
| 612 | * |
| 613 | * Add previously isolated @page to appropriate LRU list. |
| 614 | * Page may still be unevictable for other reasons. |
| 615 | * |
| 616 | * lru_lock must not be held, interrupts must be enabled. |
| 617 | */ |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 618 | void putback_lru_page(struct page *page) |
| 619 | { |
| 620 | int lru; |
| 621 | int active = !!TestClearPageActive(page); |
Lee Schermerhorn | bbfd28e | 2008-10-18 20:26:40 -0700 | [diff] [blame] | 622 | int was_unevictable = PageUnevictable(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 623 | |
| 624 | VM_BUG_ON(PageLRU(page)); |
| 625 | |
| 626 | redo: |
| 627 | ClearPageUnevictable(page); |
| 628 | |
| 629 | if (page_evictable(page, NULL)) { |
| 630 | /* |
| 631 | * For evictable pages, we can use the cache. |
| 632 | * In event of a race, worst case is we end up with an |
| 633 | * unevictable page on [in]active list. |
| 634 | * We know how to handle that. |
| 635 | */ |
Johannes Weiner | 401a8e1 | 2009-09-21 17:02:58 -0700 | [diff] [blame] | 636 | lru = active + page_lru_base_type(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 637 | lru_cache_add_lru(page, lru); |
| 638 | } else { |
| 639 | /* |
| 640 | * Put unevictable pages directly on zone's unevictable |
| 641 | * list. |
| 642 | */ |
| 643 | lru = LRU_UNEVICTABLE; |
| 644 | add_page_to_unevictable_list(page); |
Johannes Weiner | 6a7b954 | 2009-10-26 16:50:00 -0700 | [diff] [blame] | 645 | /* |
| 646 | * When racing with an mlock clearing (page is |
| 647 | * unlocked), make sure that if the other thread does |
| 648 | * not observe our setting of PG_lru and fails |
| 649 | * isolation, we see PG_mlocked cleared below and move |
| 650 | * the page back to the evictable list. |
| 651 | * |
| 652 | * The other side is TestClearPageMlocked(). |
| 653 | */ |
| 654 | smp_mb(); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 655 | } |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 656 | |
| 657 | /* |
| 658 | * page's status can change while we move it among lru. If an evictable |
| 659 | * page is on unevictable list, it never be freed. To avoid that, |
| 660 | * check after we added it to the list, again. |
| 661 | */ |
| 662 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { |
| 663 | if (!isolate_lru_page(page)) { |
| 664 | put_page(page); |
| 665 | goto redo; |
| 666 | } |
| 667 | /* This means someone else dropped this page from LRU |
| 668 | * So, it will be freed or putback to LRU again. There is |
| 669 | * nothing to do here. |
| 670 | */ |
| 671 | } |
| 672 | |
Lee Schermerhorn | bbfd28e | 2008-10-18 20:26:40 -0700 | [diff] [blame] | 673 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
| 674 | count_vm_event(UNEVICTABLE_PGRESCUED); |
| 675 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) |
| 676 | count_vm_event(UNEVICTABLE_PGCULLED); |
| 677 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 678 | put_page(page); /* drop ref from isolate */ |
| 679 | } |
| 680 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 681 | enum page_references { |
| 682 | PAGEREF_RECLAIM, |
| 683 | PAGEREF_RECLAIM_CLEAN, |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 684 | PAGEREF_KEEP, |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 685 | PAGEREF_ACTIVATE, |
| 686 | }; |
| 687 | |
| 688 | static enum page_references page_check_references(struct page *page, |
| 689 | struct scan_control *sc) |
| 690 | { |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 691 | int referenced_ptes, referenced_page; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 692 | unsigned long vm_flags; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 693 | |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 694 | referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags); |
| 695 | referenced_page = TestClearPageReferenced(page); |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 696 | |
| 697 | /* Lumpy reclaim - ignore references */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 698 | if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 699 | return PAGEREF_RECLAIM; |
| 700 | |
| 701 | /* |
| 702 | * Mlock lost the isolation race with us. Let try_to_unmap() |
| 703 | * move the page to the unevictable list. |
| 704 | */ |
| 705 | if (vm_flags & VM_LOCKED) |
| 706 | return PAGEREF_RECLAIM; |
| 707 | |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 708 | if (referenced_ptes) { |
| 709 | if (PageAnon(page)) |
| 710 | return PAGEREF_ACTIVATE; |
| 711 | /* |
| 712 | * All mapped pages start out with page table |
| 713 | * references from the instantiating fault, so we need |
| 714 | * to look twice if a mapped file page is used more |
| 715 | * than once. |
| 716 | * |
| 717 | * Mark it and spare it for another trip around the |
| 718 | * inactive list. Another page table reference will |
| 719 | * lead to its activation. |
| 720 | * |
| 721 | * Note: the mark is set for activated pages as well |
| 722 | * so that recently deactivated but used pages are |
| 723 | * quickly recovered. |
| 724 | */ |
| 725 | SetPageReferenced(page); |
| 726 | |
| 727 | if (referenced_page) |
| 728 | return PAGEREF_ACTIVATE; |
| 729 | |
| 730 | return PAGEREF_KEEP; |
| 731 | } |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 732 | |
| 733 | /* Reclaim if clean, defer dirty pages to writeback */ |
KOSAKI Motohiro | 2e30244 | 2010-10-26 14:21:46 -0700 | [diff] [blame] | 734 | if (referenced_page && !PageSwapBacked(page)) |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 735 | return PAGEREF_RECLAIM_CLEAN; |
| 736 | |
| 737 | return PAGEREF_RECLAIM; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 738 | } |
| 739 | |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 740 | static noinline_for_stack void free_page_list(struct list_head *free_pages) |
| 741 | { |
| 742 | struct pagevec freed_pvec; |
| 743 | struct page *page, *tmp; |
| 744 | |
| 745 | pagevec_init(&freed_pvec, 1); |
| 746 | |
| 747 | list_for_each_entry_safe(page, tmp, free_pages, lru) { |
| 748 | list_del(&page->lru); |
| 749 | if (!pagevec_add(&freed_pvec, page)) { |
| 750 | __pagevec_free(&freed_pvec); |
| 751 | pagevec_reinit(&freed_pvec); |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | pagevec_free(&freed_pvec); |
| 756 | } |
| 757 | |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 758 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 759 | * shrink_page_list() returns the number of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 760 | */ |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 761 | static unsigned long shrink_page_list(struct list_head *page_list, |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 762 | struct zone *zone, |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 763 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 764 | { |
| 765 | LIST_HEAD(ret_pages); |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 766 | LIST_HEAD(free_pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 767 | int pgactivate = 0; |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 768 | unsigned long nr_dirty = 0; |
| 769 | unsigned long nr_congested = 0; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 770 | unsigned long nr_reclaimed = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 771 | |
| 772 | cond_resched(); |
| 773 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 774 | while (!list_empty(page_list)) { |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 775 | enum page_references references; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 776 | struct address_space *mapping; |
| 777 | struct page *page; |
| 778 | int may_enter_fs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 779 | |
| 780 | cond_resched(); |
| 781 | |
| 782 | page = lru_to_page(page_list); |
| 783 | list_del(&page->lru); |
| 784 | |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 785 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 786 | goto keep; |
| 787 | |
Nick Piggin | 725d704 | 2006-09-25 23:30:55 -0700 | [diff] [blame] | 788 | VM_BUG_ON(PageActive(page)); |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 789 | VM_BUG_ON(page_zone(page) != zone); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 790 | |
| 791 | sc->nr_scanned++; |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 792 | |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 793 | if (unlikely(!page_evictable(page, NULL))) |
| 794 | goto cull_mlocked; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 795 | |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 796 | if (!sc->may_unmap && page_mapped(page)) |
Christoph Lameter | 80e4342 | 2006-02-11 17:55:53 -0800 | [diff] [blame] | 797 | goto keep_locked; |
| 798 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 799 | /* Double the slab pressure for mapped and swapcache pages */ |
| 800 | if (page_mapped(page) || PageSwapCache(page)) |
| 801 | sc->nr_scanned++; |
| 802 | |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 803 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
| 804 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); |
| 805 | |
| 806 | if (PageWriteback(page)) { |
| 807 | /* |
| 808 | * Synchronous reclaim is performed in two passes, |
| 809 | * first an asynchronous pass over the list to |
| 810 | * start parallel writeback, and a second synchronous |
| 811 | * pass to wait for the IO to complete. Wait here |
| 812 | * for any page for which writeback has already |
| 813 | * started. |
| 814 | */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 815 | if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) && |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 816 | may_enter_fs) |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 817 | wait_on_page_writeback(page); |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 818 | else { |
| 819 | unlock_page(page); |
| 820 | goto keep_lumpy; |
| 821 | } |
Andy Whitcroft | c661b07 | 2007-08-22 14:01:26 -0700 | [diff] [blame] | 822 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 823 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 824 | references = page_check_references(page, sc); |
| 825 | switch (references) { |
| 826 | case PAGEREF_ACTIVATE: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 827 | goto activate_locked; |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 828 | case PAGEREF_KEEP: |
| 829 | goto keep_locked; |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 830 | case PAGEREF_RECLAIM: |
| 831 | case PAGEREF_RECLAIM_CLEAN: |
| 832 | ; /* try to reclaim the page below */ |
| 833 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 834 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 835 | /* |
| 836 | * Anonymous process memory has backing store? |
| 837 | * Try to allocate it some swap space here. |
| 838 | */ |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 839 | if (PageAnon(page) && !PageSwapCache(page)) { |
Hugh Dickins | 63eb6b9 | 2008-11-19 15:36:37 -0800 | [diff] [blame] | 840 | if (!(sc->gfp_mask & __GFP_IO)) |
| 841 | goto keep_locked; |
Hugh Dickins | ac47b00 | 2009-01-06 14:39:39 -0800 | [diff] [blame] | 842 | if (!add_to_swap(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 843 | goto activate_locked; |
Hugh Dickins | 63eb6b9 | 2008-11-19 15:36:37 -0800 | [diff] [blame] | 844 | may_enter_fs = 1; |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 845 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 846 | |
| 847 | mapping = page_mapping(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | |
| 849 | /* |
| 850 | * The page is mapped into the page tables of one or more |
| 851 | * processes. Try to unmap it here. |
| 852 | */ |
| 853 | if (page_mapped(page) && mapping) { |
Andi Kleen | 14fa31b | 2009-09-16 11:50:10 +0200 | [diff] [blame] | 854 | switch (try_to_unmap(page, TTU_UNMAP)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 855 | case SWAP_FAIL: |
| 856 | goto activate_locked; |
| 857 | case SWAP_AGAIN: |
| 858 | goto keep_locked; |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 859 | case SWAP_MLOCK: |
| 860 | goto cull_mlocked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 861 | case SWAP_SUCCESS: |
| 862 | ; /* try to free the page below */ |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | if (PageDirty(page)) { |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 867 | nr_dirty++; |
| 868 | |
Johannes Weiner | dfc8d63 | 2010-03-05 13:42:19 -0800 | [diff] [blame] | 869 | if (references == PAGEREF_RECLAIM_CLEAN) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 870 | goto keep_locked; |
Andrew Morton | 4dd4b92 | 2008-03-24 12:29:52 -0700 | [diff] [blame] | 871 | if (!may_enter_fs) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 872 | goto keep_locked; |
Christoph Lameter | 52a8363 | 2006-02-01 03:05:28 -0800 | [diff] [blame] | 873 | if (!sc->may_writepage) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 874 | goto keep_locked; |
| 875 | |
| 876 | /* Page is dirty, try to write it out here */ |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 877 | switch (pageout(page, mapping, sc)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 878 | case PAGE_KEEP: |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 879 | nr_congested++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 880 | goto keep_locked; |
| 881 | case PAGE_ACTIVATE: |
| 882 | goto activate_locked; |
| 883 | case PAGE_SUCCESS: |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 884 | if (PageWriteback(page)) |
| 885 | goto keep_lumpy; |
| 886 | if (PageDirty(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 887 | goto keep; |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 888 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 889 | /* |
| 890 | * A synchronous write - probably a ramdisk. Go |
| 891 | * ahead and try to reclaim the page. |
| 892 | */ |
Nick Piggin | 529ae9a | 2008-08-02 12:01:03 +0200 | [diff] [blame] | 893 | if (!trylock_page(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 894 | goto keep; |
| 895 | if (PageDirty(page) || PageWriteback(page)) |
| 896 | goto keep_locked; |
| 897 | mapping = page_mapping(page); |
| 898 | case PAGE_CLEAN: |
| 899 | ; /* try to free the page below */ |
| 900 | } |
| 901 | } |
| 902 | |
| 903 | /* |
| 904 | * If the page has buffers, try to free the buffer mappings |
| 905 | * associated with this page. If we succeed we try to free |
| 906 | * the page as well. |
| 907 | * |
| 908 | * We do this even if the page is PageDirty(). |
| 909 | * try_to_release_page() does not perform I/O, but it is |
| 910 | * possible for a page to have PageDirty set, but it is actually |
| 911 | * clean (all its buffers are clean). This happens if the |
| 912 | * buffers were written out directly, with submit_bh(). ext3 |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 913 | * will do this, as well as the blockdev mapping. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 914 | * try_to_release_page() will discover that cleanness and will |
| 915 | * drop the buffers and mark the page clean - it can be freed. |
| 916 | * |
| 917 | * Rarely, pages can have buffers and no ->mapping. These are |
| 918 | * the pages which were not successfully invalidated in |
| 919 | * truncate_complete_page(). We try to drop those buffers here |
| 920 | * and if that worked, and the page is no longer mapped into |
| 921 | * process address space (page_count == 1) it can be freed. |
| 922 | * Otherwise, leave the page on the LRU so it is swappable. |
| 923 | */ |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 924 | if (page_has_private(page)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 925 | if (!try_to_release_page(page, sc->gfp_mask)) |
| 926 | goto activate_locked; |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 927 | if (!mapping && page_count(page) == 1) { |
| 928 | unlock_page(page); |
| 929 | if (put_page_testzero(page)) |
| 930 | goto free_it; |
| 931 | else { |
| 932 | /* |
| 933 | * rare race with speculative reference. |
| 934 | * the speculative reference will free |
| 935 | * this page shortly, so we may |
| 936 | * increment nr_reclaimed here (and |
| 937 | * leave it off the LRU). |
| 938 | */ |
| 939 | nr_reclaimed++; |
| 940 | continue; |
| 941 | } |
| 942 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 943 | } |
| 944 | |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 945 | if (!mapping || !__remove_mapping(mapping, page)) |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 946 | goto keep_locked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 947 | |
Nick Piggin | a978d6f | 2008-10-18 20:26:58 -0700 | [diff] [blame] | 948 | /* |
| 949 | * At this point, we have no other references and there is |
| 950 | * no way to pick any more up (removed from LRU, removed |
| 951 | * from pagecache). Can use non-atomic bitops now (and |
| 952 | * we obviously don't have to worry about waking up a process |
| 953 | * waiting on the page lock, because there are no references. |
| 954 | */ |
| 955 | __clear_page_locked(page); |
Nick Piggin | e286781 | 2008-07-25 19:45:30 -0700 | [diff] [blame] | 956 | free_it: |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 957 | nr_reclaimed++; |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 958 | |
| 959 | /* |
| 960 | * Is there need to periodically free_page_list? It would |
| 961 | * appear not as the counts should be low |
| 962 | */ |
| 963 | list_add(&page->lru, &free_pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 964 | continue; |
| 965 | |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 966 | cull_mlocked: |
Hugh Dickins | 63d6c5a | 2009-01-06 14:39:38 -0800 | [diff] [blame] | 967 | if (PageSwapCache(page)) |
| 968 | try_to_free_swap(page); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 969 | unlock_page(page); |
| 970 | putback_lru_page(page); |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 971 | reset_reclaim_mode(sc); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 972 | continue; |
| 973 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 974 | activate_locked: |
Rik van Riel | 68a22394 | 2008-10-18 20:26:23 -0700 | [diff] [blame] | 975 | /* Not a candidate for swapping, so reclaim swap space. */ |
| 976 | if (PageSwapCache(page) && vm_swap_full()) |
Hugh Dickins | a2c43ee | 2009-01-06 14:39:36 -0800 | [diff] [blame] | 977 | try_to_free_swap(page); |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 978 | VM_BUG_ON(PageActive(page)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 979 | SetPageActive(page); |
| 980 | pgactivate++; |
| 981 | keep_locked: |
| 982 | unlock_page(page); |
| 983 | keep: |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 984 | reset_reclaim_mode(sc); |
KOSAKI Motohiro | 7d3579e | 2010-10-26 14:21:42 -0700 | [diff] [blame] | 985 | keep_lumpy: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 986 | list_add(&page->lru, &ret_pages); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 987 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 988 | } |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 989 | |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 990 | /* |
| 991 | * Tag a zone as congested if all the dirty pages encountered were |
| 992 | * backed by a congested BDI. In this case, reclaimers should just |
| 993 | * back off and wait for congestion to clear because further reclaim |
| 994 | * will encounter the same problem |
| 995 | */ |
KAMEZAWA Hiroyuki | d6c438b | 2011-05-17 15:44:10 -0700 | [diff] [blame] | 996 | if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc)) |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 997 | zone_set_flag(zone, ZONE_CONGESTED); |
| 998 | |
Mel Gorman | abe4c3b | 2010-08-09 17:19:31 -0700 | [diff] [blame] | 999 | free_page_list(&free_pages); |
| 1000 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1001 | list_splice(&ret_pages, page_list); |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 1002 | count_vm_events(PGACTIVATE, pgactivate); |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1003 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1004 | } |
| 1005 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1006 | /* |
| 1007 | * Attempt to remove the specified page from its LRU. Only take this page |
| 1008 | * if it is of the appropriate PageActive status. Pages which are being |
| 1009 | * freed elsewhere are also ignored. |
| 1010 | * |
| 1011 | * page: page to consider |
| 1012 | * mode: one of the LRU isolation modes defined above |
| 1013 | * |
| 1014 | * returns 0 on success, -ve errno on failure. |
| 1015 | */ |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1016 | int __isolate_lru_page(struct page *page, int mode, int file) |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1017 | { |
| 1018 | int ret = -EINVAL; |
| 1019 | |
| 1020 | /* Only take pages on the LRU. */ |
| 1021 | if (!PageLRU(page)) |
| 1022 | return ret; |
| 1023 | |
| 1024 | /* |
| 1025 | * When checking the active state, we need to be sure we are |
| 1026 | * dealing with comparible boolean values. Take the logical not |
| 1027 | * of each. |
| 1028 | */ |
| 1029 | if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) |
| 1030 | return ret; |
| 1031 | |
Johannes Weiner | 6c0b135 | 2009-09-21 17:02:59 -0700 | [diff] [blame] | 1032 | if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1033 | return ret; |
| 1034 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1035 | /* |
| 1036 | * When this function is being called for lumpy reclaim, we |
| 1037 | * initially look into all LRU pages, active, inactive and |
| 1038 | * unevictable; only give shrink_page_list evictable pages. |
| 1039 | */ |
| 1040 | if (PageUnevictable(page)) |
| 1041 | return ret; |
| 1042 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1043 | ret = -EBUSY; |
KAMEZAWA Hiroyuki | 08e552c | 2009-01-07 18:08:01 -0800 | [diff] [blame] | 1044 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1045 | if (likely(get_page_unless_zero(page))) { |
| 1046 | /* |
| 1047 | * Be careful not to clear PageLRU until after we're |
| 1048 | * sure the page is not being freed elsewhere -- the |
| 1049 | * page release code relies on it. |
| 1050 | */ |
| 1051 | ClearPageLRU(page); |
| 1052 | ret = 0; |
| 1053 | } |
| 1054 | |
| 1055 | return ret; |
| 1056 | } |
| 1057 | |
Christoph Lameter | 49d2e9c | 2006-01-08 01:00:48 -0800 | [diff] [blame] | 1058 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1059 | * zone->lru_lock is heavily contended. Some of the functions that |
| 1060 | * shrink the lists perform better by taking out a batch of pages |
| 1061 | * and working on them outside the LRU lock. |
| 1062 | * |
| 1063 | * For pagecache intensive workloads, this function is the hottest |
| 1064 | * spot in the kernel (apart from copy_*_user functions). |
| 1065 | * |
| 1066 | * Appropriate locks must be held before calling this function. |
| 1067 | * |
| 1068 | * @nr_to_scan: The number of pages to look through on the list. |
| 1069 | * @src: The LRU list to pull pages off. |
| 1070 | * @dst: The temp list to put pages on to. |
| 1071 | * @scanned: The number of pages that were scanned. |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1072 | * @order: The caller's attempted allocation order |
| 1073 | * @mode: One of the LRU isolation modes |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1074 | * @file: True [1] if isolating file [!anon] pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1075 | * |
| 1076 | * returns how many pages were moved onto *@dst. |
| 1077 | */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1078 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
| 1079 | struct list_head *src, struct list_head *dst, |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1080 | unsigned long *scanned, int order, int mode, int file) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1081 | { |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1082 | unsigned long nr_taken = 0; |
Mel Gorman | a8a94d1 | 2010-08-09 17:19:17 -0700 | [diff] [blame] | 1083 | unsigned long nr_lumpy_taken = 0; |
| 1084 | unsigned long nr_lumpy_dirty = 0; |
| 1085 | unsigned long nr_lumpy_failed = 0; |
Wu Fengguang | c9b02d9 | 2006-03-22 00:08:23 -0800 | [diff] [blame] | 1086 | unsigned long scan; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1087 | |
Wu Fengguang | c9b02d9 | 2006-03-22 00:08:23 -0800 | [diff] [blame] | 1088 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1089 | struct page *page; |
| 1090 | unsigned long pfn; |
| 1091 | unsigned long end_pfn; |
| 1092 | unsigned long page_pfn; |
| 1093 | int zone_id; |
| 1094 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1095 | page = lru_to_page(src); |
| 1096 | prefetchw_prev_lru_page(page, src, flags); |
| 1097 | |
Nick Piggin | 725d704 | 2006-09-25 23:30:55 -0700 | [diff] [blame] | 1098 | VM_BUG_ON(!PageLRU(page)); |
Nick Piggin | 8d438f9 | 2006-03-22 00:07:59 -0800 | [diff] [blame] | 1099 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1100 | switch (__isolate_lru_page(page, mode, file)) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1101 | case 0: |
| 1102 | list_move(&page->lru, dst); |
KAMEZAWA Hiroyuki | 2ffebca | 2009-06-17 16:27:21 -0700 | [diff] [blame] | 1103 | mem_cgroup_del_lru(page); |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1104 | nr_taken += hpage_nr_pages(page); |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1105 | break; |
Nick Piggin | 46453a6 | 2006-03-22 00:07:58 -0800 | [diff] [blame] | 1106 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1107 | case -EBUSY: |
| 1108 | /* else it is being freed elsewhere */ |
| 1109 | list_move(&page->lru, src); |
KAMEZAWA Hiroyuki | 2ffebca | 2009-06-17 16:27:21 -0700 | [diff] [blame] | 1110 | mem_cgroup_rotate_lru_list(page, page_lru(page)); |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1111 | continue; |
| 1112 | |
| 1113 | default: |
| 1114 | BUG(); |
| 1115 | } |
| 1116 | |
| 1117 | if (!order) |
| 1118 | continue; |
| 1119 | |
| 1120 | /* |
| 1121 | * Attempt to take all pages in the order aligned region |
| 1122 | * surrounding the tag page. Only take those pages of |
| 1123 | * the same active state as that tag page. We may safely |
| 1124 | * round the target page pfn down to the requested order |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1125 | * as the mem_map is guaranteed valid out to MAX_ORDER, |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1126 | * where that page is in a different zone we will detect |
| 1127 | * it from its zone id and abort this block scan. |
| 1128 | */ |
| 1129 | zone_id = page_zone_id(page); |
| 1130 | page_pfn = page_to_pfn(page); |
| 1131 | pfn = page_pfn & ~((1 << order) - 1); |
| 1132 | end_pfn = pfn + (1 << order); |
| 1133 | for (; pfn < end_pfn; pfn++) { |
| 1134 | struct page *cursor_page; |
| 1135 | |
| 1136 | /* The target page is in the block, ignore it. */ |
| 1137 | if (unlikely(pfn == page_pfn)) |
| 1138 | continue; |
| 1139 | |
| 1140 | /* Avoid holes within the zone. */ |
| 1141 | if (unlikely(!pfn_valid_within(pfn))) |
| 1142 | break; |
| 1143 | |
| 1144 | cursor_page = pfn_to_page(pfn); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1145 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1146 | /* Check that we have not crossed a zone boundary. */ |
| 1147 | if (unlikely(page_zone_id(cursor_page) != zone_id)) |
KOSAKI Motohiro | 08fc468 | 2010-10-26 14:21:44 -0700 | [diff] [blame] | 1148 | break; |
Minchan Kim | de2e756 | 2009-09-21 17:01:43 -0700 | [diff] [blame] | 1149 | |
| 1150 | /* |
| 1151 | * If we don't have enough swap space, reclaiming of |
| 1152 | * anon page which don't already have a swap slot is |
| 1153 | * pointless. |
| 1154 | */ |
| 1155 | if (nr_swap_pages <= 0 && PageAnon(cursor_page) && |
KOSAKI Motohiro | 08fc468 | 2010-10-26 14:21:44 -0700 | [diff] [blame] | 1156 | !PageSwapCache(cursor_page)) |
| 1157 | break; |
Minchan Kim | de2e756 | 2009-09-21 17:01:43 -0700 | [diff] [blame] | 1158 | |
KAMEZAWA Hiroyuki | ee993b1 | 2009-06-16 15:33:24 -0700 | [diff] [blame] | 1159 | if (__isolate_lru_page(cursor_page, mode, file) == 0) { |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1160 | list_move(&cursor_page->lru, dst); |
KAMEZAWA Hiroyuki | cb4cbcf | 2009-06-23 08:57:55 +0900 | [diff] [blame] | 1161 | mem_cgroup_del_lru(cursor_page); |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1162 | nr_taken += hpage_nr_pages(page); |
Mel Gorman | a8a94d1 | 2010-08-09 17:19:17 -0700 | [diff] [blame] | 1163 | nr_lumpy_taken++; |
| 1164 | if (PageDirty(cursor_page)) |
| 1165 | nr_lumpy_dirty++; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1166 | scan++; |
Mel Gorman | a8a94d1 | 2010-08-09 17:19:17 -0700 | [diff] [blame] | 1167 | } else { |
Andrea Arcangeli | d179e84 | 2011-06-15 15:08:51 -0700 | [diff] [blame] | 1168 | /* |
| 1169 | * Check if the page is freed already. |
| 1170 | * |
| 1171 | * We can't use page_count() as that |
| 1172 | * requires compound_head and we don't |
| 1173 | * have a pin on the page here. If a |
| 1174 | * page is tail, we may or may not |
| 1175 | * have isolated the head, so assume |
| 1176 | * it's not free, it'd be tricky to |
| 1177 | * track the head status without a |
| 1178 | * page pin. |
| 1179 | */ |
| 1180 | if (!PageTail(cursor_page) && |
| 1181 | !atomic_read(&cursor_page->_count)) |
KOSAKI Motohiro | 08fc468 | 2010-10-26 14:21:44 -0700 | [diff] [blame] | 1182 | continue; |
| 1183 | break; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1184 | } |
| 1185 | } |
KOSAKI Motohiro | 08fc468 | 2010-10-26 14:21:44 -0700 | [diff] [blame] | 1186 | |
| 1187 | /* If we break out of the loop above, lumpy reclaim failed */ |
| 1188 | if (pfn < end_pfn) |
| 1189 | nr_lumpy_failed++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1190 | } |
| 1191 | |
| 1192 | *scanned = scan; |
Mel Gorman | a8a94d1 | 2010-08-09 17:19:17 -0700 | [diff] [blame] | 1193 | |
| 1194 | trace_mm_vmscan_lru_isolate(order, |
| 1195 | nr_to_scan, scan, |
| 1196 | nr_taken, |
| 1197 | nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed, |
| 1198 | mode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1199 | return nr_taken; |
| 1200 | } |
| 1201 | |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 1202 | static unsigned long isolate_pages_global(unsigned long nr, |
| 1203 | struct list_head *dst, |
| 1204 | unsigned long *scanned, int order, |
| 1205 | int mode, struct zone *z, |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1206 | int active, int file) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 1207 | { |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1208 | int lru = LRU_BASE; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 1209 | if (active) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1210 | lru += LRU_ACTIVE; |
| 1211 | if (file) |
| 1212 | lru += LRU_FILE; |
| 1213 | return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 1214 | mode, file); |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 1215 | } |
| 1216 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1217 | /* |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1218 | * clear_active_flags() is a helper for shrink_active_list(), clearing |
| 1219 | * any active bits from the pages in the list. |
| 1220 | */ |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1221 | static unsigned long clear_active_flags(struct list_head *page_list, |
| 1222 | unsigned int *count) |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1223 | { |
| 1224 | int nr_active = 0; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1225 | int lru; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1226 | struct page *page; |
| 1227 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1228 | list_for_each_entry(page, page_list, lru) { |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1229 | int numpages = hpage_nr_pages(page); |
Johannes Weiner | 401a8e1 | 2009-09-21 17:02:58 -0700 | [diff] [blame] | 1230 | lru = page_lru_base_type(page); |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1231 | if (PageActive(page)) { |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1232 | lru += LRU_ACTIVE; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1233 | ClearPageActive(page); |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1234 | nr_active += numpages; |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1235 | } |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1236 | if (count) |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1237 | count[lru] += numpages; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1238 | } |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1239 | |
| 1240 | return nr_active; |
| 1241 | } |
| 1242 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1243 | /** |
| 1244 | * isolate_lru_page - tries to isolate a page from its LRU list |
| 1245 | * @page: page to isolate from its LRU list |
| 1246 | * |
| 1247 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the |
| 1248 | * vmstat statistic corresponding to whatever LRU list the page was on. |
| 1249 | * |
| 1250 | * Returns 0 if the page was removed from an LRU list. |
| 1251 | * Returns -EBUSY if the page was not on an LRU list. |
| 1252 | * |
| 1253 | * The returned page will have PageLRU() cleared. If it was found on |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1254 | * the active list, it will have PageActive set. If it was found on |
| 1255 | * the unevictable list, it will have the PageUnevictable bit set. That flag |
| 1256 | * 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] | 1257 | * |
| 1258 | * The vmstat statistic corresponding to the list on which the page was |
| 1259 | * found will be decremented. |
| 1260 | * |
| 1261 | * Restrictions: |
| 1262 | * (1) Must be called with an elevated refcount on the page. This is a |
| 1263 | * fundamentnal difference from isolate_lru_pages (which is called |
| 1264 | * without a stable reference). |
| 1265 | * (2) the lru_lock must not be held. |
| 1266 | * (3) interrupts must be enabled. |
| 1267 | */ |
| 1268 | int isolate_lru_page(struct page *page) |
| 1269 | { |
| 1270 | int ret = -EBUSY; |
| 1271 | |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1272 | VM_BUG_ON(!page_count(page)); |
| 1273 | |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1274 | if (PageLRU(page)) { |
| 1275 | struct zone *zone = page_zone(page); |
| 1276 | |
| 1277 | spin_lock_irq(&zone->lru_lock); |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1278 | if (PageLRU(page)) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1279 | int lru = page_lru(page); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1280 | ret = 0; |
Konstantin Khlebnikov | 0c91731 | 2011-05-24 17:12:21 -0700 | [diff] [blame] | 1281 | get_page(page); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1282 | ClearPageLRU(page); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1283 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1284 | del_page_from_lru_list(zone, page, lru); |
Nick Piggin | 62695a8 | 2008-10-18 20:26:09 -0700 | [diff] [blame] | 1285 | } |
| 1286 | spin_unlock_irq(&zone->lru_lock); |
| 1287 | } |
| 1288 | return ret; |
| 1289 | } |
| 1290 | |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 1291 | /* |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1292 | * Are there way too many processes in the direct reclaim path already? |
| 1293 | */ |
| 1294 | static int too_many_isolated(struct zone *zone, int file, |
| 1295 | struct scan_control *sc) |
| 1296 | { |
| 1297 | unsigned long inactive, isolated; |
| 1298 | |
| 1299 | if (current_is_kswapd()) |
| 1300 | return 0; |
| 1301 | |
| 1302 | if (!scanning_global_lru(sc)) |
| 1303 | return 0; |
| 1304 | |
| 1305 | if (file) { |
| 1306 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); |
| 1307 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); |
| 1308 | } else { |
| 1309 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); |
| 1310 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); |
| 1311 | } |
| 1312 | |
| 1313 | return isolated > inactive; |
| 1314 | } |
| 1315 | |
| 1316 | /* |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1317 | * TODO: Try merging with migrations version of putback_lru_pages |
| 1318 | */ |
| 1319 | static noinline_for_stack void |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1320 | putback_lru_pages(struct zone *zone, struct scan_control *sc, |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1321 | unsigned long nr_anon, unsigned long nr_file, |
| 1322 | struct list_head *page_list) |
| 1323 | { |
| 1324 | struct page *page; |
| 1325 | struct pagevec pvec; |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1326 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1327 | |
| 1328 | pagevec_init(&pvec, 1); |
| 1329 | |
| 1330 | /* |
| 1331 | * Put back any unfreeable pages. |
| 1332 | */ |
| 1333 | spin_lock(&zone->lru_lock); |
| 1334 | while (!list_empty(page_list)) { |
| 1335 | int lru; |
| 1336 | page = lru_to_page(page_list); |
| 1337 | VM_BUG_ON(PageLRU(page)); |
| 1338 | list_del(&page->lru); |
| 1339 | if (unlikely(!page_evictable(page, NULL))) { |
| 1340 | spin_unlock_irq(&zone->lru_lock); |
| 1341 | putback_lru_page(page); |
| 1342 | spin_lock_irq(&zone->lru_lock); |
| 1343 | continue; |
| 1344 | } |
Linus Torvalds | 7a60857 | 2011-01-17 14:42:19 -0800 | [diff] [blame] | 1345 | SetPageLRU(page); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1346 | lru = page_lru(page); |
Linus Torvalds | 7a60857 | 2011-01-17 14:42:19 -0800 | [diff] [blame] | 1347 | add_page_to_lru_list(zone, page, lru); |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1348 | if (is_active_lru(lru)) { |
| 1349 | int file = is_file_lru(lru); |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1350 | int numpages = hpage_nr_pages(page); |
| 1351 | reclaim_stat->recent_rotated[file] += numpages; |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 1352 | if (!scanning_global_lru(sc)) |
| 1353 | sc->memcg_record->nr_rotated[file] += numpages; |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1354 | } |
| 1355 | if (!pagevec_add(&pvec, page)) { |
| 1356 | spin_unlock_irq(&zone->lru_lock); |
| 1357 | __pagevec_release(&pvec); |
| 1358 | spin_lock_irq(&zone->lru_lock); |
| 1359 | } |
| 1360 | } |
| 1361 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); |
| 1362 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); |
| 1363 | |
| 1364 | spin_unlock_irq(&zone->lru_lock); |
| 1365 | pagevec_release(&pvec); |
| 1366 | } |
| 1367 | |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1368 | static noinline_for_stack void update_isolated_counts(struct zone *zone, |
| 1369 | struct scan_control *sc, |
| 1370 | unsigned long *nr_anon, |
| 1371 | unsigned long *nr_file, |
| 1372 | struct list_head *isolated_list) |
| 1373 | { |
| 1374 | unsigned long nr_active; |
| 1375 | unsigned int count[NR_LRU_LISTS] = { 0, }; |
| 1376 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
| 1377 | |
| 1378 | nr_active = clear_active_flags(isolated_list, count); |
| 1379 | __count_vm_events(PGDEACTIVATE, nr_active); |
| 1380 | |
| 1381 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, |
| 1382 | -count[LRU_ACTIVE_FILE]); |
| 1383 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, |
| 1384 | -count[LRU_INACTIVE_FILE]); |
| 1385 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, |
| 1386 | -count[LRU_ACTIVE_ANON]); |
| 1387 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, |
| 1388 | -count[LRU_INACTIVE_ANON]); |
| 1389 | |
| 1390 | *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; |
| 1391 | *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; |
| 1392 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); |
| 1393 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); |
| 1394 | |
| 1395 | reclaim_stat->recent_scanned[0] += *nr_anon; |
| 1396 | reclaim_stat->recent_scanned[1] += *nr_file; |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 1397 | if (!scanning_global_lru(sc)) { |
| 1398 | sc->memcg_record->nr_scanned[0] += *nr_anon; |
| 1399 | sc->memcg_record->nr_scanned[1] += *nr_file; |
| 1400 | } |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1401 | } |
| 1402 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1403 | /* |
Wu Fengguang | e31f3698 | 2010-08-09 17:20:01 -0700 | [diff] [blame] | 1404 | * Returns true if the caller should wait to clean dirty/writeback pages. |
| 1405 | * |
| 1406 | * If we are direct reclaiming for contiguous pages and we do not reclaim |
| 1407 | * everything in the list, try again and wait for writeback IO to complete. |
| 1408 | * This will stall high-order allocations noticeably. Only do that when really |
| 1409 | * need to free the pages under high memory pressure. |
| 1410 | */ |
| 1411 | static inline bool should_reclaim_stall(unsigned long nr_taken, |
| 1412 | unsigned long nr_freed, |
| 1413 | int priority, |
| 1414 | struct scan_control *sc) |
| 1415 | { |
| 1416 | int lumpy_stall_priority; |
| 1417 | |
| 1418 | /* kswapd should not stall on sync IO */ |
| 1419 | if (current_is_kswapd()) |
| 1420 | return false; |
| 1421 | |
| 1422 | /* Only stall on lumpy reclaim */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1423 | if (sc->reclaim_mode & RECLAIM_MODE_SINGLE) |
Wu Fengguang | e31f3698 | 2010-08-09 17:20:01 -0700 | [diff] [blame] | 1424 | return false; |
| 1425 | |
| 1426 | /* If we have relaimed everything on the isolated list, no stall */ |
| 1427 | if (nr_freed == nr_taken) |
| 1428 | return false; |
| 1429 | |
| 1430 | /* |
| 1431 | * For high-order allocations, there are two stall thresholds. |
| 1432 | * High-cost allocations stall immediately where as lower |
| 1433 | * order allocations such as stacks require the scanning |
| 1434 | * priority to be much higher before stalling. |
| 1435 | */ |
| 1436 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) |
| 1437 | lumpy_stall_priority = DEF_PRIORITY; |
| 1438 | else |
| 1439 | lumpy_stall_priority = DEF_PRIORITY / 3; |
| 1440 | |
| 1441 | return priority <= lumpy_stall_priority; |
| 1442 | } |
| 1443 | |
| 1444 | /* |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 1445 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
| 1446 | * of reclaimed pages |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1447 | */ |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1448 | static noinline_for_stack unsigned long |
| 1449 | shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone, |
| 1450 | struct scan_control *sc, int priority, int file) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1451 | { |
| 1452 | LIST_HEAD(page_list); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1453 | unsigned long nr_scanned; |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1454 | unsigned long nr_reclaimed = 0; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1455 | unsigned long nr_taken; |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1456 | unsigned long nr_anon; |
| 1457 | unsigned long nr_file; |
KOSAKI Motohiro | 78dc583 | 2009-06-16 15:31:40 -0700 | [diff] [blame] | 1458 | |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1459 | while (unlikely(too_many_isolated(zone, file, sc))) { |
KOSAKI Motohiro | 58355c7 | 2009-10-26 16:49:35 -0700 | [diff] [blame] | 1460 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
Rik van Riel | 35cd781 | 2009-09-21 17:01:38 -0700 | [diff] [blame] | 1461 | |
| 1462 | /* We are about to die and free our memory. Return now. */ |
| 1463 | if (fatal_signal_pending(current)) |
| 1464 | return SWAP_CLUSTER_MAX; |
| 1465 | } |
| 1466 | |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1467 | set_reclaim_mode(priority, sc, false); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1468 | lru_add_drain(); |
| 1469 | spin_lock_irq(&zone->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1470 | |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1471 | if (scanning_global_lru(sc)) { |
| 1472 | nr_taken = isolate_pages_global(nr_to_scan, |
| 1473 | &page_list, &nr_scanned, sc->order, |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1474 | sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ? |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 1475 | ISOLATE_BOTH : ISOLATE_INACTIVE, |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1476 | zone, 0, file); |
| 1477 | zone->pages_scanned += nr_scanned; |
KOSAKI Motohiro | b35ea17 | 2009-09-21 17:01:36 -0700 | [diff] [blame] | 1478 | if (current_is_kswapd()) |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1479 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, |
| 1480 | nr_scanned); |
| 1481 | else |
| 1482 | __count_zone_vm_events(PGSCAN_DIRECT, zone, |
| 1483 | nr_scanned); |
| 1484 | } else { |
| 1485 | nr_taken = mem_cgroup_isolate_pages(nr_to_scan, |
| 1486 | &page_list, &nr_scanned, sc->order, |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1487 | sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ? |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 1488 | ISOLATE_BOTH : ISOLATE_INACTIVE, |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1489 | zone, sc->mem_cgroup, |
| 1490 | 0, file); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1491 | /* |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1492 | * mem_cgroup_isolate_pages() keeps track of |
| 1493 | * scanned pages on its own. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1494 | */ |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1495 | } |
KOSAKI Motohiro | a731286 | 2009-09-21 17:01:37 -0700 | [diff] [blame] | 1496 | |
Mel Gorman | 6663562 | 2010-08-09 17:19:30 -0700 | [diff] [blame] | 1497 | if (nr_taken == 0) { |
| 1498 | spin_unlock_irq(&zone->lru_lock); |
| 1499 | return 0; |
| 1500 | } |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1501 | |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1502 | update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1503 | |
| 1504 | spin_unlock_irq(&zone->lru_lock); |
| 1505 | |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 1506 | nr_reclaimed = shrink_page_list(&page_list, zone, sc); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1507 | |
Wu Fengguang | e31f3698 | 2010-08-09 17:20:01 -0700 | [diff] [blame] | 1508 | /* Check if we should syncronously wait for writeback */ |
| 1509 | if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) { |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1510 | set_reclaim_mode(priority, sc, true); |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 1511 | nr_reclaimed += shrink_page_list(&page_list, zone, sc); |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1512 | } |
| 1513 | |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 1514 | if (!scanning_global_lru(sc)) |
| 1515 | sc->memcg_record->nr_freed[file] += nr_reclaimed; |
| 1516 | |
KOSAKI Motohiro | e247dbc | 2010-08-09 17:19:28 -0700 | [diff] [blame] | 1517 | local_irq_disable(); |
| 1518 | if (current_is_kswapd()) |
| 1519 | __count_vm_events(KSWAPD_STEAL, nr_reclaimed); |
| 1520 | __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed); |
| 1521 | |
Mel Gorman | 1489fa1 | 2010-08-09 17:19:33 -0700 | [diff] [blame] | 1522 | putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list); |
Mel Gorman | e11da5b | 2010-10-26 14:21:40 -0700 | [diff] [blame] | 1523 | |
| 1524 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
| 1525 | zone_idx(zone), |
| 1526 | nr_scanned, nr_reclaimed, |
| 1527 | priority, |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1528 | trace_shrink_flags(file, sc->reclaim_mode)); |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1529 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1530 | } |
| 1531 | |
Martin Bligh | 3bb1a85 | 2006-10-28 10:38:24 -0700 | [diff] [blame] | 1532 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1533 | * This moves pages from the active list to the inactive list. |
| 1534 | * |
| 1535 | * We move them the other way if the page is referenced by one or more |
| 1536 | * processes, from rmap. |
| 1537 | * |
| 1538 | * If the pages are mostly unmapped, the processing is fast and it is |
| 1539 | * appropriate to hold zone->lru_lock across the whole operation. But if |
| 1540 | * the pages are mapped, the processing is slow (page_referenced()) so we |
| 1541 | * should drop zone->lru_lock around each page. It's impossible to balance |
| 1542 | * this, so instead we remove the pages from the LRU while processing them. |
| 1543 | * It is safe to rely on PG_active against the non-LRU pages in here because |
| 1544 | * nobody will play with that bit on a non-LRU page. |
| 1545 | * |
| 1546 | * The downside is that we have to touch page->_count against each page. |
| 1547 | * But we had to alter page->flags anyway. |
| 1548 | */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1549 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1550 | static void move_active_pages_to_lru(struct zone *zone, |
| 1551 | struct list_head *list, |
| 1552 | enum lru_list lru) |
| 1553 | { |
| 1554 | unsigned long pgmoved = 0; |
| 1555 | struct pagevec pvec; |
| 1556 | struct page *page; |
| 1557 | |
| 1558 | pagevec_init(&pvec, 1); |
| 1559 | |
| 1560 | while (!list_empty(list)) { |
| 1561 | page = lru_to_page(list); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1562 | |
| 1563 | VM_BUG_ON(PageLRU(page)); |
| 1564 | SetPageLRU(page); |
| 1565 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1566 | list_move(&page->lru, &zone->lru[lru].list); |
| 1567 | mem_cgroup_add_lru_list(page, lru); |
Rik van Riel | 2c888cf | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1568 | pgmoved += hpage_nr_pages(page); |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1569 | |
| 1570 | if (!pagevec_add(&pvec, page) || list_empty(list)) { |
| 1571 | spin_unlock_irq(&zone->lru_lock); |
| 1572 | if (buffer_heads_over_limit) |
| 1573 | pagevec_strip(&pvec); |
| 1574 | __pagevec_release(&pvec); |
| 1575 | spin_lock_irq(&zone->lru_lock); |
| 1576 | } |
| 1577 | } |
| 1578 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); |
| 1579 | if (!is_active_lru(lru)) |
| 1580 | __count_vm_events(PGDEACTIVATE, pgmoved); |
| 1581 | } |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1582 | |
Andrew Morton | 1742f19 | 2006-03-22 00:08:21 -0800 | [diff] [blame] | 1583 | static void shrink_active_list(unsigned long nr_pages, struct zone *zone, |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1584 | struct scan_control *sc, int priority, int file) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1585 | { |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 1586 | unsigned long nr_taken; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 1587 | unsigned long pgscanned; |
Wu Fengguang | 6fe6b7e | 2009-06-16 15:33:05 -0700 | [diff] [blame] | 1588 | unsigned long vm_flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1589 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1590 | LIST_HEAD(l_active); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 1591 | LIST_HEAD(l_inactive); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1592 | struct page *page; |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1593 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 1594 | unsigned long nr_rotated = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1595 | |
| 1596 | lru_add_drain(); |
| 1597 | spin_lock_irq(&zone->lru_lock); |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 1598 | if (scanning_global_lru(sc)) { |
Johannes Weiner | 8b25c6d | 2010-05-24 14:32:40 -0700 | [diff] [blame] | 1599 | nr_taken = isolate_pages_global(nr_pages, &l_hold, |
| 1600 | &pgscanned, sc->order, |
| 1601 | ISOLATE_ACTIVE, zone, |
| 1602 | 1, file); |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1603 | zone->pages_scanned += pgscanned; |
Johannes Weiner | 8b25c6d | 2010-05-24 14:32:40 -0700 | [diff] [blame] | 1604 | } else { |
| 1605 | nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold, |
| 1606 | &pgscanned, sc->order, |
| 1607 | ISOLATE_ACTIVE, zone, |
| 1608 | sc->mem_cgroup, 1, file); |
| 1609 | /* |
| 1610 | * mem_cgroup_isolate_pages() keeps track of |
| 1611 | * scanned pages on its own. |
| 1612 | */ |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1613 | } |
Johannes Weiner | 8b25c6d | 2010-05-24 14:32:40 -0700 | [diff] [blame] | 1614 | |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 1615 | reclaim_stat->recent_scanned[file] += nr_taken; |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 1616 | if (!scanning_global_lru(sc)) |
| 1617 | sc->memcg_record->nr_scanned[file] += nr_taken; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 1618 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1619 | __count_zone_vm_events(PGREFILL, zone, pgscanned); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1620 | if (file) |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 1621 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1622 | else |
KOSAKI Motohiro | 44c241f | 2009-09-21 17:01:35 -0700 | [diff] [blame] | 1623 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); |
KOSAKI Motohiro | a731286 | 2009-09-21 17:01:37 -0700 | [diff] [blame] | 1624 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1625 | spin_unlock_irq(&zone->lru_lock); |
| 1626 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1627 | while (!list_empty(&l_hold)) { |
| 1628 | cond_resched(); |
| 1629 | page = lru_to_page(&l_hold); |
| 1630 | list_del(&page->lru); |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1631 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 1632 | if (unlikely(!page_evictable(page, NULL))) { |
| 1633 | putback_lru_page(page); |
| 1634 | continue; |
| 1635 | } |
| 1636 | |
Johannes Weiner | 6457474 | 2010-03-05 13:42:22 -0800 | [diff] [blame] | 1637 | if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) { |
Rik van Riel | 9992af1 | 2011-01-13 15:47:13 -0800 | [diff] [blame] | 1638 | nr_rotated += hpage_nr_pages(page); |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1639 | /* |
| 1640 | * Identify referenced, file-backed active pages and |
| 1641 | * give them one more trip around the active list. So |
| 1642 | * that executable code get better chances to stay in |
| 1643 | * memory under moderate memory pressure. Anon pages |
| 1644 | * are not likely to be evicted by use-once streaming |
| 1645 | * IO, plus JVM can create lots of anon VM_EXEC pages, |
| 1646 | * so we ignore them here. |
| 1647 | */ |
Wu Fengguang | 41e2098 | 2009-10-26 16:49:53 -0700 | [diff] [blame] | 1648 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1649 | list_add(&page->lru, &l_active); |
| 1650 | continue; |
| 1651 | } |
| 1652 | } |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1653 | |
KOSAKI Motohiro | 5205e56 | 2009-09-21 17:01:44 -0700 | [diff] [blame] | 1654 | ClearPageActive(page); /* we are de-activating */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1655 | list_add(&page->lru, &l_inactive); |
| 1656 | } |
| 1657 | |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 1658 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1659 | * Move pages back to the lru list. |
Andrew Morton | b555749 | 2009-01-06 14:40:13 -0800 | [diff] [blame] | 1660 | */ |
Johannes Weiner | 2a1dc50 | 2008-12-01 03:00:35 +0100 | [diff] [blame] | 1661 | spin_lock_irq(&zone->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1662 | /* |
Wu Fengguang | 8cab475 | 2009-06-16 15:33:12 -0700 | [diff] [blame] | 1663 | * Count referenced pages from currently used mappings as rotated, |
| 1664 | * even though only some of them are actually re-activated. This |
| 1665 | * helps balance scan pressure between file and anonymous pages in |
| 1666 | * get_scan_ratio. |
Rik van Riel | 7e9cd48 | 2008-10-18 20:26:35 -0700 | [diff] [blame] | 1667 | */ |
Johannes Weiner | b7c46d1 | 2009-09-21 17:02:56 -0700 | [diff] [blame] | 1668 | reclaim_stat->recent_rotated[file] += nr_rotated; |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 1669 | if (!scanning_global_lru(sc)) |
| 1670 | sc->memcg_record->nr_rotated[file] += nr_rotated; |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 1671 | |
Wu Fengguang | 3eb4140 | 2009-06-16 15:33:13 -0700 | [diff] [blame] | 1672 | move_active_pages_to_lru(zone, &l_active, |
| 1673 | LRU_ACTIVE + file * LRU_FILE); |
| 1674 | move_active_pages_to_lru(zone, &l_inactive, |
| 1675 | LRU_BASE + file * LRU_FILE); |
KOSAKI Motohiro | a731286 | 2009-09-21 17:01:37 -0700 | [diff] [blame] | 1676 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 1677 | spin_unlock_irq(&zone->lru_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1678 | } |
| 1679 | |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 1680 | #ifdef CONFIG_SWAP |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 1681 | static int inactive_anon_is_low_global(struct zone *zone) |
KOSAKI Motohiro | f89eb90 | 2009-01-07 18:08:14 -0800 | [diff] [blame] | 1682 | { |
| 1683 | unsigned long active, inactive; |
| 1684 | |
| 1685 | active = zone_page_state(zone, NR_ACTIVE_ANON); |
| 1686 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); |
| 1687 | |
| 1688 | if (inactive * zone->inactive_ratio < active) |
| 1689 | return 1; |
| 1690 | |
| 1691 | return 0; |
| 1692 | } |
| 1693 | |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 1694 | /** |
| 1695 | * inactive_anon_is_low - check if anonymous pages need to be deactivated |
| 1696 | * @zone: zone to check |
| 1697 | * @sc: scan control of this context |
| 1698 | * |
| 1699 | * Returns true if the zone does not have enough inactive anon pages, |
| 1700 | * meaning some active anon pages need to be deactivated. |
| 1701 | */ |
| 1702 | static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) |
| 1703 | { |
| 1704 | int low; |
| 1705 | |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 1706 | /* |
| 1707 | * If we don't have swap space, anonymous page deactivation |
| 1708 | * is pointless. |
| 1709 | */ |
| 1710 | if (!total_swap_pages) |
| 1711 | return 0; |
| 1712 | |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 1713 | if (scanning_global_lru(sc)) |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 1714 | low = inactive_anon_is_low_global(zone); |
| 1715 | else |
KOSAKI Motohiro | c772be9 | 2009-01-07 18:08:25 -0800 | [diff] [blame] | 1716 | low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 1717 | return low; |
| 1718 | } |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 1719 | #else |
| 1720 | static inline int inactive_anon_is_low(struct zone *zone, |
| 1721 | struct scan_control *sc) |
| 1722 | { |
| 1723 | return 0; |
| 1724 | } |
| 1725 | #endif |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 1726 | |
Rik van Riel | 56e49d2 | 2009-06-16 15:32:28 -0700 | [diff] [blame] | 1727 | static int inactive_file_is_low_global(struct zone *zone) |
| 1728 | { |
| 1729 | unsigned long active, inactive; |
| 1730 | |
| 1731 | active = zone_page_state(zone, NR_ACTIVE_FILE); |
| 1732 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); |
| 1733 | |
| 1734 | return (active > inactive); |
| 1735 | } |
| 1736 | |
| 1737 | /** |
| 1738 | * inactive_file_is_low - check if file pages need to be deactivated |
| 1739 | * @zone: zone to check |
| 1740 | * @sc: scan control of this context |
| 1741 | * |
| 1742 | * When the system is doing streaming IO, memory pressure here |
| 1743 | * ensures that active file pages get deactivated, until more |
| 1744 | * than half of the file pages are on the inactive list. |
| 1745 | * |
| 1746 | * Once we get to that situation, protect the system's working |
| 1747 | * set from being evicted by disabling active file page aging. |
| 1748 | * |
| 1749 | * This uses a different ratio than the anonymous pages, because |
| 1750 | * the page cache uses a use-once replacement algorithm. |
| 1751 | */ |
| 1752 | static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) |
| 1753 | { |
| 1754 | int low; |
| 1755 | |
| 1756 | if (scanning_global_lru(sc)) |
| 1757 | low = inactive_file_is_low_global(zone); |
| 1758 | else |
| 1759 | low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); |
| 1760 | return low; |
| 1761 | } |
| 1762 | |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 1763 | static int inactive_list_is_low(struct zone *zone, struct scan_control *sc, |
| 1764 | int file) |
| 1765 | { |
| 1766 | if (file) |
| 1767 | return inactive_file_is_low(zone, sc); |
| 1768 | else |
| 1769 | return inactive_anon_is_low(zone, sc); |
| 1770 | } |
| 1771 | |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1772 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 1773 | struct zone *zone, struct scan_control *sc, int priority) |
| 1774 | { |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1775 | int file = is_file_lru(lru); |
| 1776 | |
Rik van Riel | b39415b | 2009-12-14 17:59:48 -0800 | [diff] [blame] | 1777 | if (is_active_lru(lru)) { |
| 1778 | if (inactive_list_is_low(zone, sc, file)) |
| 1779 | shrink_active_list(nr_to_scan, zone, sc, priority, file); |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 1780 | return 0; |
| 1781 | } |
| 1782 | |
Rik van Riel | 33c120e | 2008-10-18 20:26:36 -0700 | [diff] [blame] | 1783 | return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 1784 | } |
| 1785 | |
KAMEZAWA Hiroyuki | 1f4c025 | 2011-07-26 16:08:21 -0700 | [diff] [blame] | 1786 | static int vmscan_swappiness(struct scan_control *sc) |
| 1787 | { |
| 1788 | if (scanning_global_lru(sc)) |
| 1789 | return vm_swappiness; |
| 1790 | return mem_cgroup_swappiness(sc->mem_cgroup); |
| 1791 | } |
| 1792 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1793 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1794 | * Determine how aggressively the anon and file LRU lists should be |
| 1795 | * scanned. The relative value of each set of LRU lists is determined |
| 1796 | * by looking at the fraction of the pages scanned we did rotate back |
| 1797 | * onto the active list instead of evict. |
| 1798 | * |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1799 | * nr[0] = anon pages to scan; nr[1] = file pages to scan |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1800 | */ |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1801 | static void get_scan_count(struct zone *zone, struct scan_control *sc, |
| 1802 | unsigned long *nr, int priority) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1803 | { |
| 1804 | unsigned long anon, file, free; |
| 1805 | unsigned long anon_prio, file_prio; |
| 1806 | unsigned long ap, fp; |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1807 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1808 | u64 fraction[2], denominator; |
| 1809 | enum lru_list l; |
| 1810 | int noswap = 0; |
KAMEZAWA Hiroyuki | 246e87a | 2011-05-26 16:25:34 -0700 | [diff] [blame] | 1811 | int force_scan = 0; |
KAMEZAWA Hiroyuki | 4508378 | 2011-07-26 16:08:24 -0700 | [diff] [blame] | 1812 | unsigned long nr_force_scan[2]; |
KAMEZAWA Hiroyuki | 246e87a | 2011-05-26 16:25:34 -0700 | [diff] [blame] | 1813 | |
| 1814 | |
| 1815 | anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) + |
| 1816 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON); |
| 1817 | file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) + |
| 1818 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); |
| 1819 | |
| 1820 | if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) { |
| 1821 | /* kswapd does zone balancing and need to scan this zone */ |
| 1822 | if (scanning_global_lru(sc) && current_is_kswapd()) |
| 1823 | force_scan = 1; |
| 1824 | /* memcg may have small limit and need to avoid priority drop */ |
| 1825 | if (!scanning_global_lru(sc)) |
| 1826 | force_scan = 1; |
| 1827 | } |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1828 | |
| 1829 | /* If we have no swap space, do not bother scanning anon pages. */ |
| 1830 | if (!sc->may_swap || (nr_swap_pages <= 0)) { |
| 1831 | noswap = 1; |
| 1832 | fraction[0] = 0; |
| 1833 | fraction[1] = 1; |
| 1834 | denominator = 1; |
KAMEZAWA Hiroyuki | 4508378 | 2011-07-26 16:08:24 -0700 | [diff] [blame] | 1835 | nr_force_scan[0] = 0; |
| 1836 | nr_force_scan[1] = SWAP_CLUSTER_MAX; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1837 | goto out; |
| 1838 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1839 | |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 1840 | if (scanning_global_lru(sc)) { |
KOSAKI Motohiro | eeee9a8 | 2009-01-07 18:08:17 -0800 | [diff] [blame] | 1841 | free = zone_page_state(zone, NR_FREE_PAGES); |
| 1842 | /* If we have very few page cache pages, |
| 1843 | force-scan anon pages. */ |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 1844 | if (unlikely(file + free <= high_wmark_pages(zone))) { |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1845 | fraction[0] = 1; |
| 1846 | fraction[1] = 0; |
| 1847 | denominator = 1; |
KAMEZAWA Hiroyuki | 4508378 | 2011-07-26 16:08:24 -0700 | [diff] [blame] | 1848 | nr_force_scan[0] = SWAP_CLUSTER_MAX; |
| 1849 | nr_force_scan[1] = 0; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1850 | goto out; |
KOSAKI Motohiro | eeee9a8 | 2009-01-07 18:08:17 -0800 | [diff] [blame] | 1851 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1852 | } |
| 1853 | |
| 1854 | /* |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 1855 | * With swappiness at 100, anonymous and file have the same priority. |
| 1856 | * This scanning priority is essentially the inverse of IO cost. |
| 1857 | */ |
KAMEZAWA Hiroyuki | 1f4c025 | 2011-07-26 16:08:21 -0700 | [diff] [blame] | 1858 | anon_prio = vmscan_swappiness(sc); |
| 1859 | file_prio = 200 - vmscan_swappiness(sc); |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 1860 | |
| 1861 | /* |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1862 | * OK, so we have swap space and a fair amount of page cache |
| 1863 | * pages. We use the recently rotated / recently scanned |
| 1864 | * ratios to determine how valuable each cache is. |
| 1865 | * |
| 1866 | * Because workloads change over time (and to avoid overflow) |
| 1867 | * we keep these statistics as a floating average, which ends |
| 1868 | * up weighing recent references more than old ones. |
| 1869 | * |
| 1870 | * anon in [0], file in [1] |
| 1871 | */ |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 1872 | spin_lock_irq(&zone->lru_lock); |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1873 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1874 | reclaim_stat->recent_scanned[0] /= 2; |
| 1875 | reclaim_stat->recent_rotated[0] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1876 | } |
| 1877 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1878 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1879 | reclaim_stat->recent_scanned[1] /= 2; |
| 1880 | reclaim_stat->recent_rotated[1] /= 2; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1881 | } |
| 1882 | |
| 1883 | /* |
Rik van Riel | 00d8089 | 2008-11-19 15:36:44 -0800 | [diff] [blame] | 1884 | * The amount of pressure on anon vs file pages is inversely |
| 1885 | * proportional to the fraction of recently scanned pages on |
| 1886 | * each list that were recently referenced and in active use. |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1887 | */ |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1888 | ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); |
| 1889 | ap /= reclaim_stat->recent_rotated[0] + 1; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1890 | |
KOSAKI Motohiro | 6e90157 | 2009-01-07 18:08:15 -0800 | [diff] [blame] | 1891 | fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); |
| 1892 | fp /= reclaim_stat->recent_rotated[1] + 1; |
KOSAKI Motohiro | 58c37f6 | 2010-08-09 17:19:51 -0700 | [diff] [blame] | 1893 | spin_unlock_irq(&zone->lru_lock); |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1894 | |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1895 | fraction[0] = ap; |
| 1896 | fraction[1] = fp; |
| 1897 | denominator = ap + fp + 1; |
KAMEZAWA Hiroyuki | 4508378 | 2011-07-26 16:08:24 -0700 | [diff] [blame] | 1898 | if (force_scan) { |
| 1899 | unsigned long scan = SWAP_CLUSTER_MAX; |
| 1900 | nr_force_scan[0] = div64_u64(scan * ap, denominator); |
| 1901 | nr_force_scan[1] = div64_u64(scan * fp, denominator); |
| 1902 | } |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1903 | out: |
| 1904 | for_each_evictable_lru(l) { |
| 1905 | int file = is_file_lru(l); |
| 1906 | unsigned long scan; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1907 | |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1908 | scan = zone_nr_lru_pages(zone, sc, l); |
| 1909 | if (priority || noswap) { |
| 1910 | scan >>= priority; |
| 1911 | scan = div64_u64(scan * fraction[file], denominator); |
| 1912 | } |
KAMEZAWA Hiroyuki | 246e87a | 2011-05-26 16:25:34 -0700 | [diff] [blame] | 1913 | |
| 1914 | /* |
| 1915 | * If zone is small or memcg is small, nr[l] can be 0. |
| 1916 | * This results no-scan on this priority and priority drop down. |
| 1917 | * For global direct reclaim, it can visit next zone and tend |
| 1918 | * not to have problems. For global kswapd, it's for zone |
| 1919 | * balancing and it need to scan a small amounts. When using |
| 1920 | * memcg, priority drop can cause big latency. So, it's better |
| 1921 | * to scan small amount. See may_noscan above. |
| 1922 | */ |
KAMEZAWA Hiroyuki | 4508378 | 2011-07-26 16:08:24 -0700 | [diff] [blame] | 1923 | if (!scan && force_scan) |
| 1924 | scan = nr_force_scan[file]; |
KAMEZAWA Hiroyuki | 246e87a | 2011-05-26 16:25:34 -0700 | [diff] [blame] | 1925 | nr[l] = scan; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 1926 | } |
Wu Fengguang | 6e08a36 | 2009-06-16 15:32:29 -0700 | [diff] [blame] | 1927 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 1928 | |
| 1929 | /* |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 1930 | * Reclaim/compaction depends on a number of pages being freed. To avoid |
| 1931 | * disruption to the system, a small number of order-0 pages continue to be |
| 1932 | * rotated and reclaimed in the normal fashion. However, by the time we get |
| 1933 | * back to the allocator and call try_to_compact_zone(), we ensure that |
| 1934 | * there are enough free pages for it to be likely successful |
| 1935 | */ |
| 1936 | static inline bool should_continue_reclaim(struct zone *zone, |
| 1937 | unsigned long nr_reclaimed, |
| 1938 | unsigned long nr_scanned, |
| 1939 | struct scan_control *sc) |
| 1940 | { |
| 1941 | unsigned long pages_for_compaction; |
| 1942 | unsigned long inactive_lru_pages; |
| 1943 | |
| 1944 | /* If not in reclaim/compaction mode, stop */ |
Mel Gorman | f3a310b | 2011-01-13 15:46:00 -0800 | [diff] [blame] | 1945 | if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 1946 | return false; |
| 1947 | |
Mel Gorman | 2876592 | 2011-02-25 14:44:20 -0800 | [diff] [blame] | 1948 | /* Consider stopping depending on scan and reclaim activity */ |
| 1949 | if (sc->gfp_mask & __GFP_REPEAT) { |
| 1950 | /* |
| 1951 | * For __GFP_REPEAT allocations, stop reclaiming if the |
| 1952 | * full LRU list has been scanned and we are still failing |
| 1953 | * to reclaim pages. This full LRU scan is potentially |
| 1954 | * expensive but a __GFP_REPEAT caller really wants to succeed |
| 1955 | */ |
| 1956 | if (!nr_reclaimed && !nr_scanned) |
| 1957 | return false; |
| 1958 | } else { |
| 1959 | /* |
| 1960 | * For non-__GFP_REPEAT allocations which can presumably |
| 1961 | * fail without consequence, stop if we failed to reclaim |
| 1962 | * any pages from the last SWAP_CLUSTER_MAX number of |
| 1963 | * pages that were scanned. This will return to the |
| 1964 | * caller faster at the risk reclaim/compaction and |
| 1965 | * the resulting allocation attempt fails |
| 1966 | */ |
| 1967 | if (!nr_reclaimed) |
| 1968 | return false; |
| 1969 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 1970 | |
| 1971 | /* |
| 1972 | * If we have not reclaimed enough pages for compaction and the |
| 1973 | * inactive lists are large enough, continue reclaiming |
| 1974 | */ |
| 1975 | pages_for_compaction = (2UL << sc->order); |
| 1976 | inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) + |
| 1977 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); |
| 1978 | if (sc->nr_reclaimed < pages_for_compaction && |
| 1979 | inactive_lru_pages > pages_for_compaction) |
| 1980 | return true; |
| 1981 | |
| 1982 | /* If compaction would go ahead or the allocation would succeed, stop */ |
| 1983 | switch (compaction_suitable(zone, sc->order)) { |
| 1984 | case COMPACT_PARTIAL: |
| 1985 | case COMPACT_CONTINUE: |
| 1986 | return false; |
| 1987 | default: |
| 1988 | return true; |
| 1989 | } |
| 1990 | } |
| 1991 | |
| 1992 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1993 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. |
| 1994 | */ |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 1995 | static void shrink_zone(int priority, struct zone *zone, |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 1996 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1997 | { |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 1998 | unsigned long nr[NR_LRU_LISTS]; |
Christoph Lameter | 8695949 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 1999 | unsigned long nr_to_scan; |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2000 | enum lru_list l; |
Johannes Weiner | f0fdc5e | 2011-02-10 15:01:34 -0800 | [diff] [blame] | 2001 | unsigned long nr_reclaimed, nr_scanned; |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2002 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2003 | |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2004 | restart: |
| 2005 | nr_reclaimed = 0; |
Johannes Weiner | f0fdc5e | 2011-02-10 15:01:34 -0800 | [diff] [blame] | 2006 | nr_scanned = sc->nr_scanned; |
Shaohua Li | 76a33fc | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 2007 | get_scan_count(zone, sc, nr, priority); |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2008 | |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2009 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
| 2010 | nr[LRU_INACTIVE_FILE]) { |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 2011 | for_each_evictable_lru(l) { |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2012 | if (nr[l]) { |
KOSAKI Motohiro | ece74b2 | 2009-12-14 17:59:14 -0800 | [diff] [blame] | 2013 | nr_to_scan = min_t(unsigned long, |
| 2014 | nr[l], SWAP_CLUSTER_MAX); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2015 | nr[l] -= nr_to_scan; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2016 | |
KOSAKI Motohiro | 01dbe5c | 2009-01-06 14:40:02 -0800 | [diff] [blame] | 2017 | nr_reclaimed += shrink_list(l, nr_to_scan, |
| 2018 | zone, sc, priority); |
Christoph Lameter | b69408e | 2008-10-18 20:26:14 -0700 | [diff] [blame] | 2019 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2020 | } |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2021 | /* |
| 2022 | * On large memory systems, scan >> priority can become |
| 2023 | * really large. This is fine for the starting priority; |
| 2024 | * we want to put equal scanning pressure on each zone. |
| 2025 | * However, if the VM has a harder time of freeing pages, |
| 2026 | * with multiple processes reclaiming pages, the total |
| 2027 | * freeing target can get unreasonably large. |
| 2028 | */ |
KOSAKI Motohiro | 338fde9 | 2009-12-14 17:59:15 -0800 | [diff] [blame] | 2029 | if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY) |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2030 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2031 | } |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2032 | sc->nr_reclaimed += nr_reclaimed; |
KOSAKI Motohiro | 01dbe5c | 2009-01-06 14:40:02 -0800 | [diff] [blame] | 2033 | |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2034 | /* |
| 2035 | * Even if we did not try to evict anon pages at all, we want to |
| 2036 | * rebalance the anon lru active/inactive ratio. |
| 2037 | */ |
Minchan Kim | 74e3f3c | 2010-10-26 14:21:31 -0700 | [diff] [blame] | 2038 | if (inactive_anon_is_low(zone, sc)) |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2039 | shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); |
| 2040 | |
Mel Gorman | 3e7d344 | 2011-01-13 15:45:56 -0800 | [diff] [blame] | 2041 | /* reclaim/compaction might need reclaim to continue */ |
| 2042 | if (should_continue_reclaim(zone, nr_reclaimed, |
| 2043 | sc->nr_scanned - nr_scanned, sc)) |
| 2044 | goto restart; |
| 2045 | |
Andrew Morton | 232ea4d | 2007-02-28 20:13:21 -0800 | [diff] [blame] | 2046 | throttle_vm_writeout(sc->gfp_mask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2047 | } |
| 2048 | |
| 2049 | /* |
| 2050 | * This is the direct reclaim path, for page-allocating processes. We only |
| 2051 | * try to reclaim pages from zones which will satisfy the caller's allocation |
| 2052 | * request. |
| 2053 | * |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 2054 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
| 2055 | * Because: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2056 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
| 2057 | * allocation or |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 2058 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
| 2059 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' |
| 2060 | * zone defense algorithm. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2061 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2062 | * If a zone is deemed to be full of pinned pages then just give it a light |
| 2063 | * scan then give up on it. |
| 2064 | */ |
KAMEZAWA Hiroyuki | ac34a1a | 2011-06-27 16:18:12 -0700 | [diff] [blame] | 2065 | static void shrink_zones(int priority, struct zonelist *zonelist, |
Andrew Morton | 05ff513 | 2006-03-22 00:08:20 -0800 | [diff] [blame] | 2066 | struct scan_control *sc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2067 | { |
Mel Gorman | dd1a239 | 2008-04-28 02:12:17 -0700 | [diff] [blame] | 2068 | struct zoneref *z; |
Mel Gorman | 54a6eb5 | 2008-04-28 02:12:16 -0700 | [diff] [blame] | 2069 | struct zone *zone; |
Ying Han | d149e3b | 2011-05-26 16:25:27 -0700 | [diff] [blame] | 2070 | unsigned long nr_soft_reclaimed; |
| 2071 | unsigned long nr_soft_scanned; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2072 | |
Mel Gorman | d4debc6 | 2010-08-09 17:19:29 -0700 | [diff] [blame] | 2073 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 2074 | gfp_zone(sc->gfp_mask), sc->nodemask) { |
Con Kolivas | f3fe651 | 2006-01-06 00:11:15 -0800 | [diff] [blame] | 2075 | if (!populated_zone(zone)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2076 | continue; |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2077 | /* |
| 2078 | * Take care memory controller reclaiming has small influence |
| 2079 | * to global LRU. |
| 2080 | */ |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 2081 | if (scanning_global_lru(sc)) { |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2082 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
| 2083 | continue; |
KOSAKI Motohiro | 93e4a89 | 2010-03-05 13:41:55 -0800 | [diff] [blame] | 2084 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2085 | continue; /* Let kswapd poll it */ |
KAMEZAWA Hiroyuki | ac34a1a | 2011-06-27 16:18:12 -0700 | [diff] [blame] | 2086 | /* |
| 2087 | * This steals pages from memory cgroups over softlimit |
| 2088 | * and returns the number of reclaimed pages and |
| 2089 | * scanned pages. This works for global memory pressure |
| 2090 | * and balancing, not for a memcg's limit. |
| 2091 | */ |
| 2092 | nr_soft_scanned = 0; |
| 2093 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
| 2094 | sc->order, sc->gfp_mask, |
| 2095 | &nr_soft_scanned); |
| 2096 | sc->nr_reclaimed += nr_soft_reclaimed; |
| 2097 | sc->nr_scanned += nr_soft_scanned; |
| 2098 | /* need some check for avoid more shrink_zone() */ |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2099 | } |
Nick Piggin | 408d854 | 2006-09-25 23:31:27 -0700 | [diff] [blame] | 2100 | |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2101 | shrink_zone(priority, zone, sc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2102 | } |
Minchan Kim | d190836 | 2010-09-22 13:05:01 -0700 | [diff] [blame] | 2103 | } |
| 2104 | |
| 2105 | static bool zone_reclaimable(struct zone *zone) |
| 2106 | { |
| 2107 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; |
| 2108 | } |
| 2109 | |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 2110 | /* All zones in zonelist are unreclaimable? */ |
Minchan Kim | d190836 | 2010-09-22 13:05:01 -0700 | [diff] [blame] | 2111 | static bool all_unreclaimable(struct zonelist *zonelist, |
| 2112 | struct scan_control *sc) |
| 2113 | { |
| 2114 | struct zoneref *z; |
| 2115 | struct zone *zone; |
Minchan Kim | d190836 | 2010-09-22 13:05:01 -0700 | [diff] [blame] | 2116 | |
| 2117 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 2118 | gfp_zone(sc->gfp_mask), sc->nodemask) { |
| 2119 | if (!populated_zone(zone)) |
| 2120 | continue; |
| 2121 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
| 2122 | continue; |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 2123 | if (!zone->all_unreclaimable) |
| 2124 | return false; |
Minchan Kim | d190836 | 2010-09-22 13:05:01 -0700 | [diff] [blame] | 2125 | } |
| 2126 | |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 2127 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2128 | } |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2129 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2130 | /* |
| 2131 | * This is the main entry point to direct page reclaim. |
| 2132 | * |
| 2133 | * If a full scan of the inactive list fails to free enough memory then we |
| 2134 | * are "out of memory" and something needs to be killed. |
| 2135 | * |
| 2136 | * If the caller is !__GFP_FS then the probability of a failure is reasonably |
| 2137 | * 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] | 2138 | * caller can't do much about. We kick the writeback threads and take explicit |
| 2139 | * naps in the hope that some of these pages can be written. But if the |
| 2140 | * allocating task holds filesystem locks which prevent writeout this might not |
| 2141 | * work, and the allocation attempt will fail. |
Nishanth Aravamudan | a41f24e | 2008-04-29 00:58:25 -0700 | [diff] [blame] | 2142 | * |
| 2143 | * returns: 0, if no pages reclaimed |
| 2144 | * else, the number of pages reclaimed |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2145 | */ |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 2146 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2147 | struct scan_control *sc, |
| 2148 | struct shrink_control *shrink) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2149 | { |
| 2150 | int priority; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 2151 | unsigned long total_scanned = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2152 | struct reclaim_state *reclaim_state = current->reclaim_state; |
Mel Gorman | dd1a239 | 2008-04-28 02:12:17 -0700 | [diff] [blame] | 2153 | struct zoneref *z; |
Mel Gorman | 54a6eb5 | 2008-04-28 02:12:16 -0700 | [diff] [blame] | 2154 | struct zone *zone; |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2155 | unsigned long writeback_threshold; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2156 | |
Miao Xie | c0ff745 | 2010-05-24 14:32:08 -0700 | [diff] [blame] | 2157 | get_mems_allowed(); |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2158 | delayacct_freepages_start(); |
| 2159 | |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 2160 | if (scanning_global_lru(sc)) |
KAMEZAWA Hiroyuki | 1cfb419 | 2008-02-07 00:14:37 -0800 | [diff] [blame] | 2161 | count_vm_event(ALLOCSTALL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2162 | |
| 2163 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2164 | sc->nr_scanned = 0; |
Rik van Riel | f7b7fd8 | 2005-11-28 13:44:07 -0800 | [diff] [blame] | 2165 | if (!priority) |
KOSAKI Motohiro | a433658 | 2011-06-15 15:08:13 -0700 | [diff] [blame] | 2166 | disable_swap_token(sc->mem_cgroup); |
KAMEZAWA Hiroyuki | ac34a1a | 2011-06-27 16:18:12 -0700 | [diff] [blame] | 2167 | shrink_zones(priority, zonelist, sc); |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2168 | /* |
| 2169 | * Don't shrink slabs when reclaiming memory from |
| 2170 | * over limit cgroups |
| 2171 | */ |
KAMEZAWA Hiroyuki | e72e2bd | 2009-01-07 18:08:23 -0800 | [diff] [blame] | 2172 | if (scanning_global_lru(sc)) { |
KOSAKI Motohiro | c6a8a8c | 2010-08-09 17:19:14 -0700 | [diff] [blame] | 2173 | unsigned long lru_pages = 0; |
Mel Gorman | d4debc6 | 2010-08-09 17:19:29 -0700 | [diff] [blame] | 2174 | for_each_zone_zonelist(zone, z, zonelist, |
| 2175 | gfp_zone(sc->gfp_mask)) { |
KOSAKI Motohiro | c6a8a8c | 2010-08-09 17:19:14 -0700 | [diff] [blame] | 2176 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
| 2177 | continue; |
| 2178 | |
| 2179 | lru_pages += zone_reclaimable_pages(zone); |
| 2180 | } |
| 2181 | |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 2182 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 2183 | if (reclaim_state) { |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2184 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
KAMEZAWA Hiroyuki | 91a4547 | 2008-02-07 00:14:29 -0800 | [diff] [blame] | 2185 | reclaim_state->reclaimed_slab = 0; |
| 2186 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2187 | } |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2188 | total_scanned += sc->nr_scanned; |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2189 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2190 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2191 | |
| 2192 | /* |
| 2193 | * Try to write back as many pages as we just scanned. This |
| 2194 | * tends to cause slow streaming writers to write data to the |
| 2195 | * disk smoothly, at the dirtying rate, which is nice. But |
| 2196 | * that's undesirable in laptop mode, where we *want* lumpy |
| 2197 | * writeout. So in laptop mode, write out the whole world. |
| 2198 | */ |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2199 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
| 2200 | if (total_scanned > writeback_threshold) { |
Jens Axboe | 03ba378 | 2009-09-09 09:08:54 +0200 | [diff] [blame] | 2201 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2202 | sc->may_writepage = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2203 | } |
| 2204 | |
| 2205 | /* Take a nap, wait for some writeback to complete */ |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2206 | if (!sc->hibernation_mode && sc->nr_scanned && |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 2207 | priority < DEF_PRIORITY - 2) { |
| 2208 | struct zone *preferred_zone; |
| 2209 | |
| 2210 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), |
David Rientjes | f33261d | 2011-01-25 15:07:20 -0800 | [diff] [blame] | 2211 | &cpuset_current_mems_allowed, |
| 2212 | &preferred_zone); |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 2213 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
| 2214 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2215 | } |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2216 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2217 | out: |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2218 | delayacct_freepages_end(); |
Miao Xie | c0ff745 | 2010-05-24 14:32:08 -0700 | [diff] [blame] | 2219 | put_mems_allowed(); |
Keika Kobayashi | 873b477 | 2008-07-25 01:48:52 -0700 | [diff] [blame] | 2220 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2221 | if (sc->nr_reclaimed) |
| 2222 | return sc->nr_reclaimed; |
| 2223 | |
KOSAKI Motohiro | 929bea7 | 2011-04-14 15:22:12 -0700 | [diff] [blame] | 2224 | /* |
| 2225 | * As hibernation is going on, kswapd is freezed so that it can't mark |
| 2226 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable |
| 2227 | * check. |
| 2228 | */ |
| 2229 | if (oom_killer_disabled) |
| 2230 | return 0; |
| 2231 | |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2232 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
Minchan Kim | d190836 | 2010-09-22 13:05:01 -0700 | [diff] [blame] | 2233 | if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc)) |
KOSAKI Motohiro | bb21c7c | 2010-06-04 14:15:05 -0700 | [diff] [blame] | 2234 | return 1; |
| 2235 | |
| 2236 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2237 | } |
| 2238 | |
Mel Gorman | dac1d27 | 2008-04-28 02:12:12 -0700 | [diff] [blame] | 2239 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 2240 | gfp_t gfp_mask, nodemask_t *nodemask) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2241 | { |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 2242 | unsigned long nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2243 | struct scan_control sc = { |
| 2244 | .gfp_mask = gfp_mask, |
| 2245 | .may_writepage = !laptop_mode, |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2246 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 2247 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 2248 | .may_swap = 1, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2249 | .order = order, |
| 2250 | .mem_cgroup = NULL, |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 2251 | .nodemask = nodemask, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2252 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2253 | struct shrink_control shrink = { |
| 2254 | .gfp_mask = sc.gfp_mask, |
| 2255 | }; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2256 | |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 2257 | trace_mm_vmscan_direct_reclaim_begin(order, |
| 2258 | sc.may_writepage, |
| 2259 | gfp_mask); |
| 2260 | |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2261 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 2262 | |
| 2263 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); |
| 2264 | |
| 2265 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2266 | } |
| 2267 | |
Balbir Singh | 00f0b82 | 2008-03-04 14:28:39 -0800 | [diff] [blame] | 2268 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2269 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2270 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem, |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2271 | gfp_t gfp_mask, bool noswap, |
| 2272 | struct zone *zone, |
| 2273 | struct memcg_scanrecord *rec, |
| 2274 | unsigned long *scanned) |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2275 | { |
| 2276 | struct scan_control sc = { |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 2277 | .nr_scanned = 0, |
KOSAKI Motohiro | b8f5c56 | 2010-08-10 18:03:02 -0700 | [diff] [blame] | 2278 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2279 | .may_writepage = !laptop_mode, |
| 2280 | .may_unmap = 1, |
| 2281 | .may_swap = !noswap, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2282 | .order = 0, |
| 2283 | .mem_cgroup = mem, |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2284 | .memcg_record = rec, |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2285 | }; |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2286 | ktime_t start, end; |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 2287 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2288 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
| 2289 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 2290 | |
| 2291 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(0, |
| 2292 | sc.may_writepage, |
| 2293 | sc.gfp_mask); |
| 2294 | |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2295 | start = ktime_get(); |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2296 | /* |
| 2297 | * NOTE: Although we can get the priority field, using it |
| 2298 | * here is not a good idea, since it limits the pages we can scan. |
| 2299 | * if we don't reclaim here, the shrink_zone from balance_pgdat |
| 2300 | * will pick up pages from other mem cgroup's as well. We hack |
| 2301 | * the priority and make it zero. |
| 2302 | */ |
| 2303 | shrink_zone(0, zone, &sc); |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2304 | end = ktime_get(); |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2305 | |
| 2306 | if (rec) |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2307 | rec->elapsed += ktime_to_ns(ktime_sub(end, start)); |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2308 | *scanned = sc.nr_scanned; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 2309 | |
| 2310 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); |
| 2311 | |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2312 | return sc.nr_reclaimed; |
| 2313 | } |
| 2314 | |
Balbir Singh | e1a1cd5 | 2008-02-07 00:14:02 -0800 | [diff] [blame] | 2315 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, |
KOSAKI Motohiro | a7885eb | 2009-01-07 18:08:24 -0800 | [diff] [blame] | 2316 | gfp_t gfp_mask, |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2317 | bool noswap, |
| 2318 | struct memcg_scanrecord *rec) |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2319 | { |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2320 | struct zonelist *zonelist; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 2321 | unsigned long nr_reclaimed; |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2322 | ktime_t start, end; |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 2323 | int nid; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2324 | struct scan_control sc = { |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2325 | .may_writepage = !laptop_mode, |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 2326 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 2327 | .may_swap = !noswap, |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2328 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2329 | .order = 0, |
| 2330 | .mem_cgroup = mem_cont, |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2331 | .memcg_record = rec, |
KAMEZAWA Hiroyuki | 327c0e9 | 2009-03-31 15:23:31 -0700 | [diff] [blame] | 2332 | .nodemask = NULL, /* we don't care the placement */ |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2333 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
| 2334 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
| 2335 | }; |
| 2336 | struct shrink_control shrink = { |
| 2337 | .gfp_mask = sc.gfp_mask, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2338 | }; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2339 | |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2340 | start = ktime_get(); |
Ying Han | 889976d | 2011-05-26 16:25:33 -0700 | [diff] [blame] | 2341 | /* |
| 2342 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't |
| 2343 | * take care of from where we get pages. So the node where we start the |
| 2344 | * scan does not need to be the current node. |
| 2345 | */ |
| 2346 | nid = mem_cgroup_select_victim_node(mem_cont); |
| 2347 | |
| 2348 | zonelist = NODE_DATA(nid)->node_zonelists; |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 2349 | |
| 2350 | trace_mm_vmscan_memcg_reclaim_begin(0, |
| 2351 | sc.may_writepage, |
| 2352 | sc.gfp_mask); |
| 2353 | |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2354 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2355 | end = ktime_get(); |
KAMEZAWA Hiroyuki | 82f9d48 | 2011-07-26 16:08:26 -0700 | [diff] [blame] | 2356 | if (rec) |
Shaohua Li | f51bdd2 | 2011-08-25 15:59:10 -0700 | [diff] [blame] | 2357 | rec->elapsed += ktime_to_ns(ktime_sub(end, start)); |
KOSAKI Motohiro | bdce6d9 | 2010-08-09 17:19:56 -0700 | [diff] [blame] | 2358 | |
| 2359 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
| 2360 | |
| 2361 | return nr_reclaimed; |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2362 | } |
| 2363 | #endif |
| 2364 | |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2365 | /* |
| 2366 | * pgdat_balanced is used when checking if a node is balanced for high-order |
| 2367 | * allocations. Only zones that meet watermarks and are in a zone allowed |
| 2368 | * by the callers classzone_idx are added to balanced_pages. The total of |
| 2369 | * balanced pages must be at least 25% of the zones allowed by classzone_idx |
| 2370 | * for the node to be considered balanced. Forcing all zones to be balanced |
| 2371 | * for high orders can cause excessive reclaim when there are imbalanced zones. |
| 2372 | * The choice of 25% is due to |
| 2373 | * o a 16M DMA zone that is balanced will not balance a zone on any |
| 2374 | * reasonable sized machine |
| 2375 | * o On all other machines, the top zone must be at least a reasonable |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 2376 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2377 | * would need to be at least 256M for it to be balance a whole node. |
| 2378 | * Similarly, on x86-64 the Normal zone would need to be at least 1G |
| 2379 | * to balance a node on its own. These seemed like reasonable ratios. |
| 2380 | */ |
| 2381 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, |
| 2382 | int classzone_idx) |
| 2383 | { |
| 2384 | unsigned long present_pages = 0; |
| 2385 | int i; |
| 2386 | |
| 2387 | for (i = 0; i <= classzone_idx; i++) |
| 2388 | present_pages += pgdat->node_zones[i].present_pages; |
| 2389 | |
Shaohua Li | 4746efd | 2011-07-19 08:49:26 -0700 | [diff] [blame] | 2390 | /* A special case here: if zone has no page, we think it's balanced */ |
| 2391 | return balanced_pages >= (present_pages >> 2); |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2392 | } |
| 2393 | |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2394 | /* is kswapd sleeping prematurely? */ |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2395 | static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, |
| 2396 | int classzone_idx) |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2397 | { |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2398 | int i; |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2399 | unsigned long balanced = 0; |
| 2400 | bool all_zones_ok = true; |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2401 | |
| 2402 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
| 2403 | if (remaining) |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2404 | return true; |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2405 | |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 2406 | /* Check the watermark levels */ |
Mel Gorman | 08951e5 | 2011-07-08 15:39:36 -0700 | [diff] [blame] | 2407 | for (i = 0; i <= classzone_idx; i++) { |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2408 | struct zone *zone = pgdat->node_zones + i; |
| 2409 | |
| 2410 | if (!populated_zone(zone)) |
| 2411 | continue; |
| 2412 | |
Mel Gorman | 355b09c | 2011-01-13 15:46:24 -0800 | [diff] [blame] | 2413 | /* |
| 2414 | * balance_pgdat() skips over all_unreclaimable after |
| 2415 | * DEF_PRIORITY. Effectively, it considers them balanced so |
| 2416 | * they must be considered balanced here as well if kswapd |
| 2417 | * is to sleep |
| 2418 | */ |
| 2419 | if (zone->all_unreclaimable) { |
| 2420 | balanced += zone->present_pages; |
KOSAKI Motohiro | de3fab3 | 2010-01-15 17:01:25 -0800 | [diff] [blame] | 2421 | continue; |
Mel Gorman | 355b09c | 2011-01-13 15:46:24 -0800 | [diff] [blame] | 2422 | } |
KOSAKI Motohiro | de3fab3 | 2010-01-15 17:01:25 -0800 | [diff] [blame] | 2423 | |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2424 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), |
Mel Gorman | da175d0 | 2011-07-08 15:39:39 -0700 | [diff] [blame] | 2425 | i, 0)) |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2426 | all_zones_ok = false; |
| 2427 | else |
| 2428 | balanced += zone->present_pages; |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2429 | } |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2430 | |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2431 | /* |
| 2432 | * For high-order requests, the balanced zones must contain at least |
| 2433 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones |
| 2434 | * must be balanced |
| 2435 | */ |
| 2436 | if (order) |
Johannes Weiner | afc7e32 | 2011-05-24 17:11:09 -0700 | [diff] [blame] | 2437 | return !pgdat_balanced(pgdat, balanced, classzone_idx); |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2438 | else |
| 2439 | return !all_zones_ok; |
Mel Gorman | f50de2d | 2009-12-14 17:58:53 -0800 | [diff] [blame] | 2440 | } |
| 2441 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2442 | /* |
| 2443 | * For kswapd, balance_pgdat() will work across all this node's zones until |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 2444 | * they are all at high_wmark_pages(zone). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2445 | * |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 2446 | * Returns the final order kswapd was reclaiming at |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2447 | * |
| 2448 | * There is special handling here for zones which are full of pinned pages. |
| 2449 | * This can happen if the pages are all mlocked, or if they are all used by |
| 2450 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. |
| 2451 | * What we do is to detect the case where all pages in the zone have been |
| 2452 | * scanned twice and there has been zero successful reclaim. Mark the zone as |
| 2453 | * dead and from now on, only perform a short scan. Basically we're polling |
| 2454 | * the zone for when the problem goes away. |
| 2455 | * |
| 2456 | * kswapd scans the zones in the highmem->normal->dma direction. It skips |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 2457 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
| 2458 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the |
| 2459 | * lower zones regardless of the number of free pages in the lower zones. This |
| 2460 | * interoperates with the page allocator fallback scheme to ensure that aging |
| 2461 | * of pages is balanced across the zones. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2462 | */ |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2463 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2464 | int *classzone_idx) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2465 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2466 | int all_zones_ok; |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2467 | unsigned long balanced; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2468 | int priority; |
| 2469 | int i; |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2470 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 2471 | unsigned long total_scanned; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2472 | struct reclaim_state *reclaim_state = current->reclaim_state; |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 2473 | unsigned long nr_soft_reclaimed; |
| 2474 | unsigned long nr_soft_scanned; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 2475 | struct scan_control sc = { |
| 2476 | .gfp_mask = GFP_KERNEL, |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 2477 | .may_unmap = 1, |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 2478 | .may_swap = 1, |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 2479 | /* |
| 2480 | * kswapd doesn't want to be bailed out while reclaim. because |
| 2481 | * we want to put equal scanning pressure on each zone. |
| 2482 | */ |
| 2483 | .nr_to_reclaim = ULONG_MAX, |
Andy Whitcroft | 5ad333e | 2007-07-17 04:03:16 -0700 | [diff] [blame] | 2484 | .order = order, |
Balbir Singh | 66e1707 | 2008-02-07 00:13:56 -0800 | [diff] [blame] | 2485 | .mem_cgroup = NULL, |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 2486 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2487 | struct shrink_control shrink = { |
| 2488 | .gfp_mask = sc.gfp_mask, |
| 2489 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2490 | loop_again: |
| 2491 | total_scanned = 0; |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2492 | sc.nr_reclaimed = 0; |
Christoph Lameter | c0bbbc7 | 2006-06-11 15:22:26 -0700 | [diff] [blame] | 2493 | sc.may_writepage = !laptop_mode; |
Christoph Lameter | f8891e5 | 2006-06-30 01:55:45 -0700 | [diff] [blame] | 2494 | count_vm_event(PAGEOUTRUN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2495 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2496 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2497 | unsigned long lru_pages = 0; |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2498 | int has_under_min_watermark_zone = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2499 | |
Rik van Riel | f7b7fd8 | 2005-11-28 13:44:07 -0800 | [diff] [blame] | 2500 | /* The swap token gets in the way of swapout... */ |
| 2501 | if (!priority) |
KOSAKI Motohiro | a433658 | 2011-06-15 15:08:13 -0700 | [diff] [blame] | 2502 | disable_swap_token(NULL); |
Rik van Riel | f7b7fd8 | 2005-11-28 13:44:07 -0800 | [diff] [blame] | 2503 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2504 | all_zones_ok = 1; |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2505 | balanced = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2506 | |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2507 | /* |
| 2508 | * Scan in the highmem->dma direction for the highest |
| 2509 | * zone which needs scanning |
| 2510 | */ |
| 2511 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { |
| 2512 | struct zone *zone = pgdat->node_zones + i; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2513 | |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2514 | if (!populated_zone(zone)) |
| 2515 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2516 | |
KOSAKI Motohiro | 93e4a89 | 2010-03-05 13:41:55 -0800 | [diff] [blame] | 2517 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2518 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2519 | |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2520 | /* |
| 2521 | * Do some background aging of the anon list, to give |
| 2522 | * pages a chance to be referenced before reclaiming. |
| 2523 | */ |
KOSAKI Motohiro | 14797e2 | 2009-01-07 18:08:18 -0800 | [diff] [blame] | 2524 | if (inactive_anon_is_low(zone, &sc)) |
Rik van Riel | 556adec | 2008-10-18 20:26:34 -0700 | [diff] [blame] | 2525 | shrink_active_list(SWAP_CLUSTER_MAX, zone, |
| 2526 | &sc, priority, 0); |
| 2527 | |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2528 | if (!zone_watermark_ok_safe(zone, order, |
Mel Gorman | 4185896 | 2009-06-16 15:32:12 -0700 | [diff] [blame] | 2529 | high_wmark_pages(zone), 0, 0)) { |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2530 | end_zone = i; |
Andrew Morton | e1dbeda | 2006-12-06 20:32:01 -0800 | [diff] [blame] | 2531 | break; |
Shaohua Li | 439423f | 2011-08-25 15:59:12 -0700 | [diff] [blame] | 2532 | } else { |
| 2533 | /* If balanced, clear the congested flag */ |
| 2534 | zone_clear_flag(zone, ZONE_CONGESTED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2535 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2536 | } |
Andrew Morton | e1dbeda | 2006-12-06 20:32:01 -0800 | [diff] [blame] | 2537 | if (i < 0) |
| 2538 | goto out; |
| 2539 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2540 | for (i = 0; i <= end_zone; i++) { |
| 2541 | struct zone *zone = pgdat->node_zones + i; |
| 2542 | |
Wu Fengguang | adea02a | 2009-09-21 17:01:42 -0700 | [diff] [blame] | 2543 | lru_pages += zone_reclaimable_pages(zone); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2544 | } |
| 2545 | |
| 2546 | /* |
| 2547 | * Now scan the zone in the dma->highmem direction, stopping |
| 2548 | * at the last zone which needs scanning. |
| 2549 | * |
| 2550 | * We do this because the page allocator works in the opposite |
| 2551 | * direction. This prevents the page allocator from allocating |
| 2552 | * pages behind kswapd's direction of progress, which would |
| 2553 | * cause too much scanning of the lower zones. |
| 2554 | */ |
| 2555 | for (i = 0; i <= end_zone; i++) { |
| 2556 | struct zone *zone = pgdat->node_zones + i; |
akpm@osdl.org | b15e090 | 2005-06-21 17:14:35 -0700 | [diff] [blame] | 2557 | int nr_slab; |
Mel Gorman | 8afdcec | 2011-03-22 16:33:04 -0700 | [diff] [blame] | 2558 | unsigned long balance_gap; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2559 | |
Con Kolivas | f3fe651 | 2006-01-06 00:11:15 -0800 | [diff] [blame] | 2560 | if (!populated_zone(zone)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2561 | continue; |
| 2562 | |
KOSAKI Motohiro | 93e4a89 | 2010-03-05 13:41:55 -0800 | [diff] [blame] | 2563 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2564 | continue; |
| 2565 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2566 | sc.nr_scanned = 0; |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2567 | |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 2568 | nr_soft_scanned = 0; |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2569 | /* |
| 2570 | * Call soft limit reclaim before calling shrink_zone. |
Balbir Singh | 4e41695 | 2009-09-23 15:56:39 -0700 | [diff] [blame] | 2571 | */ |
Ying Han | 0ae5e89 | 2011-05-26 16:25:25 -0700 | [diff] [blame] | 2572 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
| 2573 | order, sc.gfp_mask, |
| 2574 | &nr_soft_scanned); |
| 2575 | sc.nr_reclaimed += nr_soft_reclaimed; |
| 2576 | total_scanned += nr_soft_scanned; |
KOSAKI Motohiro | 00918b6 | 2010-08-10 18:03:05 -0700 | [diff] [blame] | 2577 | |
Rik van Riel | 32a4330 | 2007-10-16 01:24:50 -0700 | [diff] [blame] | 2578 | /* |
Mel Gorman | 8afdcec | 2011-03-22 16:33:04 -0700 | [diff] [blame] | 2579 | * We put equal pressure on every zone, unless |
| 2580 | * one zone has way too many pages free |
| 2581 | * already. The "too many pages" is defined |
| 2582 | * as the high wmark plus a "gap" where the |
| 2583 | * gap is either the low watermark or 1% |
| 2584 | * of the zone, whichever is smaller. |
Rik van Riel | 32a4330 | 2007-10-16 01:24:50 -0700 | [diff] [blame] | 2585 | */ |
Mel Gorman | 8afdcec | 2011-03-22 16:33:04 -0700 | [diff] [blame] | 2586 | balance_gap = min(low_wmark_pages(zone), |
| 2587 | (zone->present_pages + |
| 2588 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / |
| 2589 | KSWAPD_ZONE_BALANCE_GAP_RATIO); |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2590 | if (!zone_watermark_ok_safe(zone, order, |
Mel Gorman | 8afdcec | 2011-03-22 16:33:04 -0700 | [diff] [blame] | 2591 | high_wmark_pages(zone) + balance_gap, |
Mel Gorman | d7868da | 2011-07-08 15:39:38 -0700 | [diff] [blame] | 2592 | end_zone, 0)) { |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2593 | shrink_zone(priority, zone, &sc); |
Andrea Arcangeli | 5a03b05 | 2011-01-13 15:47:11 -0800 | [diff] [blame] | 2594 | |
Mel Gorman | d7868da | 2011-07-08 15:39:38 -0700 | [diff] [blame] | 2595 | reclaim_state->reclaimed_slab = 0; |
| 2596 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); |
| 2597 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; |
| 2598 | total_scanned += sc.nr_scanned; |
| 2599 | |
| 2600 | if (nr_slab == 0 && !zone_reclaimable(zone)) |
| 2601 | zone->all_unreclaimable = 1; |
| 2602 | } |
| 2603 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2604 | /* |
| 2605 | * If we've done a decent amount of scanning and |
| 2606 | * the reclaim ratio is low, start doing writepage |
| 2607 | * even in laptop mode |
| 2608 | */ |
| 2609 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2610 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2611 | sc.may_writepage = 1; |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2612 | |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2613 | if (zone->all_unreclaimable) { |
| 2614 | if (end_zone && end_zone == i) |
| 2615 | end_zone--; |
Mel Gorman | d7868da | 2011-07-08 15:39:38 -0700 | [diff] [blame] | 2616 | continue; |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2617 | } |
Mel Gorman | d7868da | 2011-07-08 15:39:38 -0700 | [diff] [blame] | 2618 | |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2619 | if (!zone_watermark_ok_safe(zone, order, |
Minchan Kim | 45973d7 | 2010-03-05 13:41:45 -0800 | [diff] [blame] | 2620 | high_wmark_pages(zone), end_zone, 0)) { |
| 2621 | all_zones_ok = 0; |
| 2622 | /* |
| 2623 | * We are still under min water mark. This |
| 2624 | * means that we have a GFP_ATOMIC allocation |
| 2625 | * failure risk. Hurry up! |
| 2626 | */ |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2627 | if (!zone_watermark_ok_safe(zone, order, |
Minchan Kim | 45973d7 | 2010-03-05 13:41:45 -0800 | [diff] [blame] | 2628 | min_wmark_pages(zone), end_zone, 0)) |
| 2629 | has_under_min_watermark_zone = 1; |
Mel Gorman | 0e093d99 | 2010-10-26 14:21:45 -0700 | [diff] [blame] | 2630 | } else { |
| 2631 | /* |
| 2632 | * If a zone reaches its high watermark, |
| 2633 | * consider it to be no longer congested. It's |
| 2634 | * possible there are dirty pages backed by |
| 2635 | * congested BDIs but as pressure is relieved, |
| 2636 | * spectulatively avoid congestion waits |
| 2637 | */ |
| 2638 | zone_clear_flag(zone, ZONE_CONGESTED); |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2639 | if (i <= *classzone_idx) |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2640 | balanced += zone->present_pages; |
Minchan Kim | 45973d7 | 2010-03-05 13:41:45 -0800 | [diff] [blame] | 2641 | } |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2642 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2643 | } |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2644 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2645 | break; /* kswapd: all done */ |
| 2646 | /* |
| 2647 | * OK, kswapd is getting into trouble. Take a nap, then take |
| 2648 | * another pass across the zones. |
| 2649 | */ |
KOSAKI Motohiro | bb3ab59 | 2009-12-14 17:58:55 -0800 | [diff] [blame] | 2650 | if (total_scanned && (priority < DEF_PRIORITY - 2)) { |
| 2651 | if (has_under_min_watermark_zone) |
| 2652 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); |
| 2653 | else |
| 2654 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
| 2655 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2656 | |
| 2657 | /* |
| 2658 | * We do this so kswapd doesn't build up large priorities for |
| 2659 | * example when it is freeing in parallel with allocators. It |
| 2660 | * matches the direct reclaim path behaviour in terms of impact |
| 2661 | * on zone->*_priority. |
| 2662 | */ |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 2663 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2664 | break; |
| 2665 | } |
| 2666 | out: |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2667 | |
| 2668 | /* |
| 2669 | * order-0: All zones must meet high watermark for a balanced node |
Mel Gorman | 1741c87 | 2011-01-13 15:46:21 -0800 | [diff] [blame] | 2670 | * high-order: Balanced zones must make up at least 25% of the node |
| 2671 | * for the node to be balanced |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2672 | */ |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2673 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2674 | cond_resched(); |
Rafael J. Wysocki | 8357376 | 2006-12-06 20:34:18 -0800 | [diff] [blame] | 2675 | |
| 2676 | try_to_freeze(); |
| 2677 | |
KOSAKI Motohiro | 73ce02e | 2009-01-06 14:40:33 -0800 | [diff] [blame] | 2678 | /* |
| 2679 | * Fragmentation may mean that the system cannot be |
| 2680 | * rebalanced for high-order allocations in all zones. |
| 2681 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, |
| 2682 | * it means the zones have been fully scanned and are still |
| 2683 | * not balanced. For high-order allocations, there is |
| 2684 | * little point trying all over again as kswapd may |
| 2685 | * infinite loop. |
| 2686 | * |
| 2687 | * Instead, recheck all watermarks at order-0 as they |
| 2688 | * are the most important. If watermarks are ok, kswapd will go |
| 2689 | * back to sleep. High-order users can still perform direct |
| 2690 | * reclaim if they wish. |
| 2691 | */ |
| 2692 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) |
| 2693 | order = sc.order = 0; |
| 2694 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2695 | goto loop_again; |
| 2696 | } |
| 2697 | |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2698 | /* |
| 2699 | * If kswapd was reclaiming at a higher order, it has the option of |
| 2700 | * sleeping without all zones being balanced. Before it does, it must |
| 2701 | * ensure that the watermarks for order-0 on *all* zones are met and |
| 2702 | * that the congestion flags are cleared. The congestion flag must |
| 2703 | * be cleared as kswapd is the only mechanism that clears the flag |
| 2704 | * and it is potentially going to sleep here. |
| 2705 | */ |
| 2706 | if (order) { |
| 2707 | for (i = 0; i <= end_zone; i++) { |
| 2708 | struct zone *zone = pgdat->node_zones + i; |
| 2709 | |
| 2710 | if (!populated_zone(zone)) |
| 2711 | continue; |
| 2712 | |
| 2713 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
| 2714 | continue; |
| 2715 | |
| 2716 | /* Confirm the zone is balanced for order-0 */ |
| 2717 | if (!zone_watermark_ok(zone, 0, |
| 2718 | high_wmark_pages(zone), 0, 0)) { |
| 2719 | order = sc.order = 0; |
| 2720 | goto loop_again; |
| 2721 | } |
| 2722 | |
| 2723 | /* If balanced, clear the congested flag */ |
| 2724 | zone_clear_flag(zone, ZONE_CONGESTED); |
| 2725 | } |
| 2726 | } |
| 2727 | |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 2728 | /* |
| 2729 | * Return the order we were reclaiming at so sleeping_prematurely() |
| 2730 | * makes a decision on the order we were last reclaiming at. However, |
| 2731 | * if another caller entered the allocator slow path while kswapd |
| 2732 | * was awake, order will remain at the higher level |
| 2733 | */ |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2734 | *classzone_idx = end_zone; |
Mel Gorman | 0abdee2 | 2011-01-13 15:46:22 -0800 | [diff] [blame] | 2735 | return order; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2736 | } |
| 2737 | |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2738 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 2739 | { |
| 2740 | long remaining = 0; |
| 2741 | DEFINE_WAIT(wait); |
| 2742 | |
| 2743 | if (freezing(current) || kthread_should_stop()) |
| 2744 | return; |
| 2745 | |
| 2746 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
| 2747 | |
| 2748 | /* Try to sleep for a short interval */ |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2749 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 2750 | remaining = schedule_timeout(HZ/10); |
| 2751 | finish_wait(&pgdat->kswapd_wait, &wait); |
| 2752 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
| 2753 | } |
| 2754 | |
| 2755 | /* |
| 2756 | * After a short sleep, check if it was a premature sleep. If not, then |
| 2757 | * go fully to sleep until explicitly woken up. |
| 2758 | */ |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2759 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 2760 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
| 2761 | |
| 2762 | /* |
| 2763 | * vmstat counters are not perfectly accurate and the estimated |
| 2764 | * value for counters such as NR_FREE_PAGES can deviate from the |
| 2765 | * true value by nr_online_cpus * threshold. To avoid the zone |
| 2766 | * watermarks being breached while under pressure, we reduce the |
| 2767 | * per-cpu vmstat threshold while kswapd is awake and restore |
| 2768 | * them before going back to sleep. |
| 2769 | */ |
| 2770 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); |
| 2771 | schedule(); |
| 2772 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
| 2773 | } else { |
| 2774 | if (remaining) |
| 2775 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); |
| 2776 | else |
| 2777 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); |
| 2778 | } |
| 2779 | finish_wait(&pgdat->kswapd_wait, &wait); |
| 2780 | } |
| 2781 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2782 | /* |
| 2783 | * The background pageout daemon, started as a kernel thread |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2784 | * from the init process. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2785 | * |
| 2786 | * This basically trickles out pages so that we have _some_ |
| 2787 | * free memory available even if there is no other activity |
| 2788 | * that frees anything up. This is needed for things like routing |
| 2789 | * etc, where we otherwise might have all activity going on in |
| 2790 | * asynchronous contexts that cannot page things out. |
| 2791 | * |
| 2792 | * If there are applications that are active memory-allocators |
| 2793 | * (most normal use), this basically shouldn't matter. |
| 2794 | */ |
| 2795 | static int kswapd(void *p) |
| 2796 | { |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2797 | unsigned long order, new_order; |
| 2798 | int classzone_idx, new_classzone_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2799 | pg_data_t *pgdat = (pg_data_t*)p; |
| 2800 | struct task_struct *tsk = current; |
KOSAKI Motohiro | f0bc0a6 | 2011-01-13 15:45:50 -0800 | [diff] [blame] | 2801 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2802 | struct reclaim_state reclaim_state = { |
| 2803 | .reclaimed_slab = 0, |
| 2804 | }; |
Rusty Russell | a70f730 | 2009-03-13 14:49:46 +1030 | [diff] [blame] | 2805 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2806 | |
Nick Piggin | cf40bd1 | 2009-01-21 08:12:39 +0100 | [diff] [blame] | 2807 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
| 2808 | |
Rusty Russell | 174596a | 2009-01-01 10:12:29 +1030 | [diff] [blame] | 2809 | if (!cpumask_empty(cpumask)) |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 2810 | set_cpus_allowed_ptr(tsk, cpumask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2811 | current->reclaim_state = &reclaim_state; |
| 2812 | |
| 2813 | /* |
| 2814 | * Tell the memory management that we're a "memory allocator", |
| 2815 | * and that if we need more memory we should get access to it |
| 2816 | * regardless (see "__alloc_pages()"). "kswapd" should |
| 2817 | * never get caught in the normal page freeing logic. |
| 2818 | * |
| 2819 | * (Kswapd normally doesn't need memory anyway, but sometimes |
| 2820 | * you need a small amount of memory in order to be able to |
| 2821 | * page out something else, and this flag essentially protects |
| 2822 | * us from recursively trying to free more memory as we're |
| 2823 | * trying to free the first piece of memory in the first place). |
| 2824 | */ |
Christoph Lameter | 930d915 | 2006-01-08 01:00:47 -0800 | [diff] [blame] | 2825 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
Rafael J. Wysocki | 8314418 | 2007-07-17 04:03:35 -0700 | [diff] [blame] | 2826 | set_freezable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2827 | |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2828 | order = new_order = 0; |
| 2829 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2830 | for ( ; ; ) { |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 2831 | int ret; |
Christoph Lameter | 3e1d1d2 | 2005-06-24 23:13:50 -0700 | [diff] [blame] | 2832 | |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2833 | /* |
| 2834 | * If the last balance_pgdat was unsuccessful it's unlikely a |
| 2835 | * new request of a similar or harder type will succeed soon |
| 2836 | * so consider going to sleep on the basis we reclaimed at |
| 2837 | */ |
| 2838 | if (classzone_idx >= new_classzone_idx && order == new_order) { |
| 2839 | new_order = pgdat->kswapd_max_order; |
| 2840 | new_classzone_idx = pgdat->classzone_idx; |
| 2841 | pgdat->kswapd_max_order = 0; |
| 2842 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
| 2843 | } |
| 2844 | |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2845 | if (order < new_order || classzone_idx > new_classzone_idx) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2846 | /* |
| 2847 | * Don't sleep if someone wants a larger 'order' |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2848 | * allocation or has tigher zone constraints |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2849 | */ |
| 2850 | order = new_order; |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2851 | classzone_idx = new_classzone_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2852 | } else { |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2853 | kswapd_try_to_sleep(pgdat, order, classzone_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2854 | order = pgdat->kswapd_max_order; |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2855 | classzone_idx = pgdat->classzone_idx; |
Mel Gorman | 4d40502 | 2011-01-13 15:46:23 -0800 | [diff] [blame] | 2856 | pgdat->kswapd_max_order = 0; |
Mel Gorman | 215ddd6 | 2011-07-08 15:39:40 -0700 | [diff] [blame] | 2857 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2858 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2859 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 2860 | ret = try_to_freeze(); |
| 2861 | if (kthread_should_stop()) |
| 2862 | break; |
| 2863 | |
| 2864 | /* |
| 2865 | * We can speed up thawing tasks if we don't call balance_pgdat |
| 2866 | * after returning from the refrigerator |
| 2867 | */ |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 2868 | if (!ret) { |
| 2869 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); |
Mel Gorman | dc83edd | 2011-01-13 15:46:26 -0800 | [diff] [blame] | 2870 | order = balance_pgdat(pgdat, order, &classzone_idx); |
Mel Gorman | 33906bc | 2010-08-09 17:19:16 -0700 | [diff] [blame] | 2871 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2872 | } |
| 2873 | return 0; |
| 2874 | } |
| 2875 | |
| 2876 | /* |
| 2877 | * A zone is low on free memory, so wake its kswapd task to service it. |
| 2878 | */ |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2879 | 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] | 2880 | { |
| 2881 | pg_data_t *pgdat; |
| 2882 | |
Con Kolivas | f3fe651 | 2006-01-06 00:11:15 -0800 | [diff] [blame] | 2883 | if (!populated_zone(zone)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2884 | return; |
| 2885 | |
Paul Jackson | 02a0e53 | 2006-12-13 00:34:25 -0800 | [diff] [blame] | 2886 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2887 | return; |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2888 | pgdat = zone->zone_pgdat; |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2889 | if (pgdat->kswapd_max_order < order) { |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2890 | pgdat->kswapd_max_order = order; |
Mel Gorman | 9950474 | 2011-01-13 15:46:20 -0800 | [diff] [blame] | 2891 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
| 2892 | } |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 2893 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2894 | return; |
Mel Gorman | 88f5acf | 2011-01-13 15:45:41 -0800 | [diff] [blame] | 2895 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
| 2896 | return; |
| 2897 | |
| 2898 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); |
Con Kolivas | 8d0986e | 2005-09-13 01:25:07 -0700 | [diff] [blame] | 2899 | wake_up_interruptible(&pgdat->kswapd_wait); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2900 | } |
| 2901 | |
Wu Fengguang | adea02a | 2009-09-21 17:01:42 -0700 | [diff] [blame] | 2902 | /* |
| 2903 | * The reclaimable count would be mostly accurate. |
| 2904 | * The less reclaimable pages may be |
| 2905 | * - mlocked pages, which will be moved to unevictable list when encountered |
| 2906 | * - mapped pages, which may require several travels to be reclaimed |
| 2907 | * - dirty pages, which is not "instantly" reclaimable |
| 2908 | */ |
| 2909 | unsigned long global_reclaimable_pages(void) |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2910 | { |
Wu Fengguang | adea02a | 2009-09-21 17:01:42 -0700 | [diff] [blame] | 2911 | int nr; |
| 2912 | |
| 2913 | nr = global_page_state(NR_ACTIVE_FILE) + |
| 2914 | global_page_state(NR_INACTIVE_FILE); |
| 2915 | |
| 2916 | if (nr_swap_pages > 0) |
| 2917 | nr += global_page_state(NR_ACTIVE_ANON) + |
| 2918 | global_page_state(NR_INACTIVE_ANON); |
| 2919 | |
| 2920 | return nr; |
| 2921 | } |
| 2922 | |
| 2923 | unsigned long zone_reclaimable_pages(struct zone *zone) |
| 2924 | { |
| 2925 | int nr; |
| 2926 | |
| 2927 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + |
| 2928 | zone_page_state(zone, NR_INACTIVE_FILE); |
| 2929 | |
| 2930 | if (nr_swap_pages > 0) |
| 2931 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + |
| 2932 | zone_page_state(zone, NR_INACTIVE_ANON); |
| 2933 | |
| 2934 | return nr; |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 2935 | } |
| 2936 | |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 2937 | #ifdef CONFIG_HIBERNATION |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2938 | /* |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2939 | * 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] | 2940 | * freed pages. |
| 2941 | * |
| 2942 | * Rather than trying to age LRUs the aim is to preserve the overall |
| 2943 | * LRU order by reclaiming preferentially |
| 2944 | * inactive > active > active referenced > active mapped |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2945 | */ |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2946 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2947 | { |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2948 | struct reclaim_state reclaim_state; |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2949 | struct scan_control sc = { |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2950 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
| 2951 | .may_swap = 1, |
| 2952 | .may_unmap = 1, |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2953 | .may_writepage = 1, |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2954 | .nr_to_reclaim = nr_to_reclaim, |
| 2955 | .hibernation_mode = 1, |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2956 | .order = 0, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2957 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2958 | struct shrink_control shrink = { |
| 2959 | .gfp_mask = sc.gfp_mask, |
| 2960 | }; |
| 2961 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2962 | struct task_struct *p = current; |
| 2963 | unsigned long nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2964 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2965 | p->flags |= PF_MEMALLOC; |
| 2966 | lockdep_set_current_reclaim_state(sc.gfp_mask); |
| 2967 | reclaim_state.reclaimed_slab = 0; |
| 2968 | p->reclaim_state = &reclaim_state; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 2969 | |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 2970 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2971 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2972 | p->reclaim_state = NULL; |
| 2973 | lockdep_clear_current_reclaim_state(); |
| 2974 | p->flags &= ~PF_MEMALLOC; |
Rafael J. Wysocki | d6277db | 2006-06-23 02:03:18 -0700 | [diff] [blame] | 2975 | |
KOSAKI Motohiro | 7b51755 | 2009-12-14 17:59:12 -0800 | [diff] [blame] | 2976 | return nr_reclaimed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2977 | } |
Rafael J. Wysocki | c6f37f1 | 2009-05-24 22:16:31 +0200 | [diff] [blame] | 2978 | #endif /* CONFIG_HIBERNATION */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2979 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2980 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
| 2981 | not required for correctness. So if the last cpu in a node goes |
| 2982 | away, we get changed to run anywhere: as the first one comes back, |
| 2983 | restore their cpu bindings. */ |
Chandra Seetharaman | 9c7b216 | 2006-06-27 02:54:07 -0700 | [diff] [blame] | 2984 | static int __devinit cpu_callback(struct notifier_block *nfb, |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 2985 | unsigned long action, void *hcpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2986 | { |
Yasunori Goto | 58c0a4a | 2007-10-16 01:25:40 -0700 | [diff] [blame] | 2987 | int nid; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2988 | |
Rafael J. Wysocki | 8bb7844 | 2007-05-09 02:35:10 -0700 | [diff] [blame] | 2989 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
Yasunori Goto | 58c0a4a | 2007-10-16 01:25:40 -0700 | [diff] [blame] | 2990 | for_each_node_state(nid, N_HIGH_MEMORY) { |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 2991 | pg_data_t *pgdat = NODE_DATA(nid); |
Rusty Russell | a70f730 | 2009-03-13 14:49:46 +1030 | [diff] [blame] | 2992 | const struct cpumask *mask; |
| 2993 | |
| 2994 | mask = cpumask_of_node(pgdat->node_id); |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 2995 | |
Rusty Russell | 3e59794 | 2009-01-01 10:12:24 +1030 | [diff] [blame] | 2996 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2997 | /* One of our CPUs online: restore mask */ |
Mike Travis | c5f59f0 | 2008-04-04 18:11:10 -0700 | [diff] [blame] | 2998 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2999 | } |
| 3000 | } |
| 3001 | return NOTIFY_OK; |
| 3002 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3004 | /* |
| 3005 | * This kswapd start function will be called by init and node-hot-add. |
| 3006 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. |
| 3007 | */ |
| 3008 | int kswapd_run(int nid) |
| 3009 | { |
| 3010 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3011 | int ret = 0; |
| 3012 | |
| 3013 | if (pgdat->kswapd) |
| 3014 | return 0; |
| 3015 | |
| 3016 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); |
| 3017 | if (IS_ERR(pgdat->kswapd)) { |
| 3018 | /* failure at boot is fatal */ |
| 3019 | BUG_ON(system_state == SYSTEM_BOOTING); |
| 3020 | printk("Failed to start kswapd on node %d\n",nid); |
| 3021 | ret = -1; |
| 3022 | } |
| 3023 | return ret; |
| 3024 | } |
| 3025 | |
David Rientjes | 8fe23e0 | 2009-12-14 17:58:33 -0800 | [diff] [blame] | 3026 | /* |
| 3027 | * Called by memory hotplug when all memory in a node is offlined. |
| 3028 | */ |
| 3029 | void kswapd_stop(int nid) |
| 3030 | { |
| 3031 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; |
| 3032 | |
| 3033 | if (kswapd) |
| 3034 | kthread_stop(kswapd); |
| 3035 | } |
| 3036 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3037 | static int __init kswapd_init(void) |
| 3038 | { |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3039 | int nid; |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3040 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3041 | swap_setup(); |
Christoph Lameter | 9422ffb | 2007-10-16 01:25:31 -0700 | [diff] [blame] | 3042 | for_each_node_state(nid, N_HIGH_MEMORY) |
Yasunori Goto | 3218ae1 | 2006-06-27 02:53:33 -0700 | [diff] [blame] | 3043 | kswapd_run(nid); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3044 | hotcpu_notifier(cpu_callback, 0); |
| 3045 | return 0; |
| 3046 | } |
| 3047 | |
| 3048 | module_init(kswapd_init) |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3049 | |
| 3050 | #ifdef CONFIG_NUMA |
| 3051 | /* |
| 3052 | * Zone reclaim mode |
| 3053 | * |
| 3054 | * If non-zero call zone_reclaim when the number of free pages falls below |
| 3055 | * the watermarks. |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3056 | */ |
| 3057 | int zone_reclaim_mode __read_mostly; |
| 3058 | |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3059 | #define RECLAIM_OFF 0 |
Fernando Luis Vazquez Cao | 7d03431 | 2008-07-29 22:33:41 -0700 | [diff] [blame] | 3060 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
Christoph Lameter | 1b2ffb7 | 2006-02-01 03:05:34 -0800 | [diff] [blame] | 3061 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
| 3062 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ |
| 3063 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3064 | /* |
Christoph Lameter | a92f712 | 2006-02-01 03:05:32 -0800 | [diff] [blame] | 3065 | * Priority for ZONE_RECLAIM. This determines the fraction of pages |
| 3066 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
| 3067 | * a zone. |
| 3068 | */ |
| 3069 | #define ZONE_RECLAIM_PRIORITY 4 |
| 3070 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3071 | /* |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3072 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to |
| 3073 | * occur. |
| 3074 | */ |
| 3075 | int sysctl_min_unmapped_ratio = 1; |
| 3076 | |
| 3077 | /* |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3078 | * If the number of slab pages in a zone grows beyond this percentage then |
| 3079 | * slab reclaim needs to occur. |
| 3080 | */ |
| 3081 | int sysctl_min_slab_ratio = 5; |
| 3082 | |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3083 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
| 3084 | { |
| 3085 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); |
| 3086 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + |
| 3087 | zone_page_state(zone, NR_ACTIVE_FILE); |
| 3088 | |
| 3089 | /* |
| 3090 | * It's possible for there to be more file mapped pages than |
| 3091 | * accounted for by the pages on the file LRU lists because |
| 3092 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED |
| 3093 | */ |
| 3094 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; |
| 3095 | } |
| 3096 | |
| 3097 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ |
| 3098 | static long zone_pagecache_reclaimable(struct zone *zone) |
| 3099 | { |
| 3100 | long nr_pagecache_reclaimable; |
| 3101 | long delta = 0; |
| 3102 | |
| 3103 | /* |
| 3104 | * If RECLAIM_SWAP is set, then all file pages are considered |
| 3105 | * potentially reclaimable. Otherwise, we have to worry about |
| 3106 | * pages like swapcache and zone_unmapped_file_pages() provides |
| 3107 | * a better estimate |
| 3108 | */ |
| 3109 | if (zone_reclaim_mode & RECLAIM_SWAP) |
| 3110 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); |
| 3111 | else |
| 3112 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); |
| 3113 | |
| 3114 | /* If we can't clean pages, remove dirty pages from consideration */ |
| 3115 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) |
| 3116 | delta += zone_page_state(zone, NR_FILE_DIRTY); |
| 3117 | |
| 3118 | /* Watch for any possible underflows due to delta */ |
| 3119 | if (unlikely(delta > nr_pagecache_reclaimable)) |
| 3120 | delta = nr_pagecache_reclaimable; |
| 3121 | |
| 3122 | return nr_pagecache_reclaimable - delta; |
| 3123 | } |
| 3124 | |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3125 | /* |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3126 | * Try to free up some pages from this zone through reclaim. |
| 3127 | */ |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3128 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3129 | { |
Christoph Lameter | 7fb2d46 | 2006-03-22 00:08:22 -0800 | [diff] [blame] | 3130 | /* Minimum pages needed in order to stay on node */ |
Andrew Morton | 69e0594 | 2006-03-22 00:08:19 -0800 | [diff] [blame] | 3131 | const unsigned long nr_pages = 1 << order; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3132 | struct task_struct *p = current; |
| 3133 | struct reclaim_state reclaim_state; |
Christoph Lameter | 8695949 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3134 | int priority; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3135 | struct scan_control sc = { |
| 3136 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), |
Johannes Weiner | a6dc60f | 2009-03-31 15:19:30 -0700 | [diff] [blame] | 3137 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
KOSAKI Motohiro | 2e2e425 | 2009-04-21 12:24:57 -0700 | [diff] [blame] | 3138 | .may_swap = 1, |
KOSAKI Motohiro | 22fba33 | 2009-12-14 17:59:10 -0800 | [diff] [blame] | 3139 | .nr_to_reclaim = max_t(unsigned long, nr_pages, |
| 3140 | SWAP_CLUSTER_MAX), |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3141 | .gfp_mask = gfp_mask, |
Johannes Weiner | bd2f619 | 2009-03-31 15:19:38 -0700 | [diff] [blame] | 3142 | .order = order, |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3143 | }; |
Ying Han | a09ed5e | 2011-05-24 17:12:26 -0700 | [diff] [blame] | 3144 | struct shrink_control shrink = { |
| 3145 | .gfp_mask = sc.gfp_mask, |
| 3146 | }; |
KOSAKI Motohiro | 1574804 | 2010-08-09 17:19:50 -0700 | [diff] [blame] | 3147 | unsigned long nr_slab_pages0, nr_slab_pages1; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3148 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3149 | cond_resched(); |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3150 | /* |
| 3151 | * We need to be able to allocate from the reserves for RECLAIM_SWAP |
| 3152 | * and we also need to be able to write out pages for RECLAIM_WRITE |
| 3153 | * and RECLAIM_SWAP. |
| 3154 | */ |
| 3155 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; |
KOSAKI Motohiro | 76ca542 | 2010-03-05 13:41:47 -0800 | [diff] [blame] | 3156 | lockdep_set_current_reclaim_state(gfp_mask); |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3157 | reclaim_state.reclaimed_slab = 0; |
| 3158 | p->reclaim_state = &reclaim_state; |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3159 | |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3160 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3161 | /* |
| 3162 | * Free memory by calling shrink zone with increasing |
| 3163 | * priorities until we have enough memory freed. |
| 3164 | */ |
| 3165 | priority = ZONE_RECLAIM_PRIORITY; |
| 3166 | do { |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 3167 | shrink_zone(priority, zone, &sc); |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3168 | priority--; |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 3169 | } while (priority >= 0 && sc.nr_reclaimed < nr_pages); |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3170 | } |
Christoph Lameter | c84db23 | 2006-02-01 03:05:29 -0800 | [diff] [blame] | 3171 | |
KOSAKI Motohiro | 1574804 | 2010-08-09 17:19:50 -0700 | [diff] [blame] | 3172 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
| 3173 | if (nr_slab_pages0 > zone->min_slab_pages) { |
Christoph Lameter | 2a16e3f | 2006-02-01 03:05:35 -0800 | [diff] [blame] | 3174 | /* |
Christoph Lameter | 7fb2d46 | 2006-03-22 00:08:22 -0800 | [diff] [blame] | 3175 | * shrink_slab() does not currently allow us to determine how |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3176 | * many pages were freed in this zone. So we take the current |
| 3177 | * number of slab pages and shake the slab until it is reduced |
| 3178 | * by the same nr_pages that we used for reclaiming unmapped |
| 3179 | * pages. |
Christoph Lameter | 2a16e3f | 2006-02-01 03:05:35 -0800 | [diff] [blame] | 3180 | * |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3181 | * Note that shrink_slab will free memory on all zones and may |
| 3182 | * take a long time. |
Christoph Lameter | 2a16e3f | 2006-02-01 03:05:35 -0800 | [diff] [blame] | 3183 | */ |
KOSAKI Motohiro | 4dc4b3d | 2010-08-09 17:19:54 -0700 | [diff] [blame] | 3184 | for (;;) { |
| 3185 | unsigned long lru_pages = zone_reclaimable_pages(zone); |
| 3186 | |
| 3187 | /* No reclaimable slab or very low memory pressure */ |
Ying Han | 1495f23 | 2011-05-24 17:12:27 -0700 | [diff] [blame] | 3188 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
KOSAKI Motohiro | 4dc4b3d | 2010-08-09 17:19:54 -0700 | [diff] [blame] | 3189 | break; |
| 3190 | |
| 3191 | /* Freed enough memory */ |
| 3192 | nr_slab_pages1 = zone_page_state(zone, |
| 3193 | NR_SLAB_RECLAIMABLE); |
| 3194 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) |
| 3195 | break; |
| 3196 | } |
Christoph Lameter | 83e33a4 | 2006-09-25 23:31:53 -0700 | [diff] [blame] | 3197 | |
| 3198 | /* |
| 3199 | * Update nr_reclaimed by the number of slab pages we |
| 3200 | * reclaimed from this zone. |
| 3201 | */ |
KOSAKI Motohiro | 1574804 | 2010-08-09 17:19:50 -0700 | [diff] [blame] | 3202 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
| 3203 | if (nr_slab_pages1 < nr_slab_pages0) |
| 3204 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; |
Christoph Lameter | 2a16e3f | 2006-02-01 03:05:35 -0800 | [diff] [blame] | 3205 | } |
| 3206 | |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3207 | p->reclaim_state = NULL; |
Christoph Lameter | d4f7796 | 2006-02-24 13:04:22 -0800 | [diff] [blame] | 3208 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
KOSAKI Motohiro | 76ca542 | 2010-03-05 13:41:47 -0800 | [diff] [blame] | 3209 | lockdep_clear_current_reclaim_state(); |
Rik van Riel | a79311c | 2009-01-06 14:40:01 -0800 | [diff] [blame] | 3210 | return sc.nr_reclaimed >= nr_pages; |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3211 | } |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3212 | |
| 3213 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
| 3214 | { |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3215 | int node_id; |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3216 | int ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3217 | |
| 3218 | /* |
Christoph Lameter | 0ff3849 | 2006-09-25 23:31:52 -0700 | [diff] [blame] | 3219 | * Zone reclaim reclaims unmapped file backed pages and |
| 3220 | * slab pages if we are over the defined limits. |
Christoph Lameter | 34aa133 | 2006-06-30 01:55:37 -0700 | [diff] [blame] | 3221 | * |
Christoph Lameter | 9614634 | 2006-07-03 00:24:13 -0700 | [diff] [blame] | 3222 | * A small portion of unmapped file backed pages is needed for |
| 3223 | * file I/O otherwise pages read by file I/O will be immediately |
| 3224 | * thrown out if the zone is overallocated. So we do not reclaim |
| 3225 | * if less than a specified percentage of the zone is used by |
| 3226 | * unmapped file backed pages. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3227 | */ |
Mel Gorman | 90afa5d | 2009-06-16 15:33:20 -0700 | [diff] [blame] | 3228 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
| 3229 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3230 | return ZONE_RECLAIM_FULL; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3231 | |
KOSAKI Motohiro | 93e4a89 | 2010-03-05 13:41:55 -0800 | [diff] [blame] | 3232 | if (zone->all_unreclaimable) |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3233 | return ZONE_RECLAIM_FULL; |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3234 | |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3235 | /* |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3236 | * Do not scan if the allocation should not be delayed. |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3237 | */ |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3238 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3239 | return ZONE_RECLAIM_NOSCAN; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3240 | |
| 3241 | /* |
| 3242 | * Only run zone reclaim on the local zone or on zones that do not |
| 3243 | * have associated processors. This will favor the local processor |
| 3244 | * over remote processors and spread off node memory allocations |
| 3245 | * as wide as possible. |
| 3246 | */ |
Christoph Lameter | 89fa302 | 2006-09-25 23:31:55 -0700 | [diff] [blame] | 3247 | node_id = zone_to_nid(zone); |
Christoph Lameter | 37c0708 | 2007-10-16 01:25:36 -0700 | [diff] [blame] | 3248 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3249 | return ZONE_RECLAIM_NOSCAN; |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3250 | |
| 3251 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) |
Mel Gorman | fa5e084 | 2009-06-16 15:33:22 -0700 | [diff] [blame] | 3252 | return ZONE_RECLAIM_NOSCAN; |
| 3253 | |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3254 | ret = __zone_reclaim(zone, gfp_mask, order); |
| 3255 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); |
| 3256 | |
Mel Gorman | 24cf72518 | 2009-06-16 15:33:23 -0700 | [diff] [blame] | 3257 | if (!ret) |
| 3258 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); |
| 3259 | |
David Rientjes | d773ed6 | 2007-10-16 23:26:01 -0700 | [diff] [blame] | 3260 | return ret; |
Andrew Morton | 179e963 | 2006-03-22 00:08:18 -0800 | [diff] [blame] | 3261 | } |
Christoph Lameter | 9eeff23 | 2006-01-18 17:42:31 -0800 | [diff] [blame] | 3262 | #endif |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3263 | |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3264 | /* |
| 3265 | * page_evictable - test whether a page is evictable |
| 3266 | * @page: the page to test |
| 3267 | * @vma: the VMA in which the page is or will be mapped, may be NULL |
| 3268 | * |
| 3269 | * Test whether page is evictable--i.e., should be placed on active/inactive |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 3270 | * lists vs unevictable list. The vma argument is !NULL when called from the |
| 3271 | * fault path to determine how to instantate a new page. |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3272 | * |
| 3273 | * Reasons page might not be evictable: |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3274 | * (1) page's mapping marked unevictable |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 3275 | * (2) page is part of an mlocked VMA |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3276 | * |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3277 | */ |
| 3278 | int page_evictable(struct page *page, struct vm_area_struct *vma) |
| 3279 | { |
| 3280 | |
Lee Schermerhorn | ba9ddf4 | 2008-10-18 20:26:42 -0700 | [diff] [blame] | 3281 | if (mapping_unevictable(page_mapping(page))) |
| 3282 | return 0; |
| 3283 | |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 3284 | if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) |
| 3285 | return 0; |
Lee Schermerhorn | 894bc31 | 2008-10-18 20:26:39 -0700 | [diff] [blame] | 3286 | |
| 3287 | return 1; |
| 3288 | } |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3289 | |
| 3290 | /** |
| 3291 | * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list |
| 3292 | * @page: page to check evictability and move to appropriate lru list |
| 3293 | * @zone: zone page is in |
| 3294 | * |
| 3295 | * Checks a page for evictability and moves the page to the appropriate |
| 3296 | * zone lru list. |
| 3297 | * |
| 3298 | * Restrictions: zone->lru_lock must be held, page must be on LRU and must |
| 3299 | * have PageUnevictable set. |
| 3300 | */ |
| 3301 | static void check_move_unevictable_page(struct page *page, struct zone *zone) |
| 3302 | { |
| 3303 | VM_BUG_ON(PageActive(page)); |
| 3304 | |
| 3305 | retry: |
| 3306 | ClearPageUnevictable(page); |
| 3307 | if (page_evictable(page, NULL)) { |
Johannes Weiner | 401a8e1 | 2009-09-21 17:02:58 -0700 | [diff] [blame] | 3308 | enum lru_list l = page_lru_base_type(page); |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3309 | |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3310 | __dec_zone_state(zone, NR_UNEVICTABLE); |
| 3311 | list_move(&page->lru, &zone->lru[l].list); |
KAMEZAWA Hiroyuki | 08e552c | 2009-01-07 18:08:01 -0800 | [diff] [blame] | 3312 | mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3313 | __inc_zone_state(zone, NR_INACTIVE_ANON + l); |
| 3314 | __count_vm_event(UNEVICTABLE_PGRESCUED); |
| 3315 | } else { |
| 3316 | /* |
| 3317 | * rotate unevictable list |
| 3318 | */ |
| 3319 | SetPageUnevictable(page); |
| 3320 | list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); |
KAMEZAWA Hiroyuki | 08e552c | 2009-01-07 18:08:01 -0800 | [diff] [blame] | 3321 | mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); |
Lee Schermerhorn | 89e004ea | 2008-10-18 20:26:43 -0700 | [diff] [blame] | 3322 | if (page_evictable(page, NULL)) |
| 3323 | goto retry; |
| 3324 | } |
| 3325 | } |
| 3326 | |
| 3327 | /** |
| 3328 | * scan_mapping_unevictable_pages - scan an address space for evictable pages |
| 3329 | * @mapping: struct address_space to scan for evictable pages |
| 3330 | * |
| 3331 | * Scan all pages in mapping. Check unevictable pages for |
| 3332 | * evictability and move them to the appropriate zone lru list. |
| 3333 | */ |
| 3334 | void scan_mapping_unevictable_pages(struct address_space *mapping) |
| 3335 | { |
| 3336 | pgoff_t next = 0; |
| 3337 | pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> |
| 3338 | PAGE_CACHE_SHIFT; |
| 3339 | struct zone *zone; |
| 3340 | struct pagevec pvec; |
| 3341 | |
| 3342 | if (mapping->nrpages == 0) |
| 3343 | return; |
| 3344 | |
| 3345 | pagevec_init(&pvec, 0); |
| 3346 | while (next < end && |
| 3347 | pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { |
| 3348 | int i; |
| 3349 | int pg_scanned = 0; |
| 3350 | |
| 3351 | zone = NULL; |
| 3352 | |
| 3353 | for (i = 0; i < pagevec_count(&pvec); i++) { |
| 3354 | struct page *page = pvec.pages[i]; |
| 3355 | pgoff_t page_index = page->index; |
| 3356 | struct zone *pagezone = page_zone(page); |
| 3357 | |
| 3358 | pg_scanned++; |
| 3359 | if (page_index > next) |
| 3360 | next = page_index; |
| 3361 | next++; |
| 3362 | |
| 3363 | if (pagezone != zone) { |
| 3364 | if (zone) |
| 3365 | spin_unlock_irq(&zone->lru_lock); |
| 3366 | zone = pagezone; |
| 3367 | spin_lock_irq(&zone->lru_lock); |
| 3368 | } |
| 3369 | |
| 3370 | if (PageLRU(page) && PageUnevictable(page)) |
| 3371 | check_move_unevictable_page(page, zone); |
| 3372 | } |
| 3373 | if (zone) |
| 3374 | spin_unlock_irq(&zone->lru_lock); |
| 3375 | pagevec_release(&pvec); |
| 3376 | |
| 3377 | count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); |
| 3378 | } |
| 3379 | |
| 3380 | } |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3381 | |
| 3382 | /** |
| 3383 | * scan_zone_unevictable_pages - check unevictable list for evictable pages |
| 3384 | * @zone - zone of which to scan the unevictable list |
| 3385 | * |
| 3386 | * Scan @zone's unevictable LRU lists to check for pages that have become |
| 3387 | * evictable. Move those that have to @zone's inactive list where they |
| 3388 | * become candidates for reclaim, unless shrink_inactive_zone() decides |
| 3389 | * to reactivate them. Pages that are still unevictable are rotated |
| 3390 | * back onto @zone's unevictable list. |
| 3391 | */ |
| 3392 | #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ |
KOSAKI Motohiro | 14b90b2 | 2009-01-06 14:39:45 -0800 | [diff] [blame] | 3393 | static void scan_zone_unevictable_pages(struct zone *zone) |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3394 | { |
| 3395 | struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; |
| 3396 | unsigned long scan; |
| 3397 | unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); |
| 3398 | |
| 3399 | while (nr_to_scan > 0) { |
| 3400 | unsigned long batch_size = min(nr_to_scan, |
| 3401 | SCAN_UNEVICTABLE_BATCH_SIZE); |
| 3402 | |
| 3403 | spin_lock_irq(&zone->lru_lock); |
| 3404 | for (scan = 0; scan < batch_size; scan++) { |
| 3405 | struct page *page = lru_to_page(l_unevictable); |
| 3406 | |
| 3407 | if (!trylock_page(page)) |
| 3408 | continue; |
| 3409 | |
| 3410 | prefetchw_prev_lru_page(page, l_unevictable, flags); |
| 3411 | |
| 3412 | if (likely(PageLRU(page) && PageUnevictable(page))) |
| 3413 | check_move_unevictable_page(page, zone); |
| 3414 | |
| 3415 | unlock_page(page); |
| 3416 | } |
| 3417 | spin_unlock_irq(&zone->lru_lock); |
| 3418 | |
| 3419 | nr_to_scan -= batch_size; |
| 3420 | } |
| 3421 | } |
| 3422 | |
| 3423 | |
| 3424 | /** |
| 3425 | * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages |
| 3426 | * |
| 3427 | * A really big hammer: scan all zones' unevictable LRU lists to check for |
| 3428 | * pages that have become evictable. Move those back to the zones' |
| 3429 | * inactive list where they become candidates for reclaim. |
| 3430 | * This occurs when, e.g., we have unswappable pages on the unevictable lists, |
| 3431 | * and we add swap to the system. As such, it runs in the context of a task |
| 3432 | * that has possibly/probably made some previously unevictable pages |
| 3433 | * evictable. |
| 3434 | */ |
KOSAKI Motohiro | ff30153 | 2009-01-06 14:39:44 -0800 | [diff] [blame] | 3435 | static void scan_all_zones_unevictable_pages(void) |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3436 | { |
| 3437 | struct zone *zone; |
| 3438 | |
| 3439 | for_each_zone(zone) { |
| 3440 | scan_zone_unevictable_pages(zone); |
| 3441 | } |
| 3442 | } |
| 3443 | |
| 3444 | /* |
| 3445 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of |
| 3446 | * all nodes' unevictable lists for evictable pages |
| 3447 | */ |
| 3448 | unsigned long scan_unevictable_pages; |
| 3449 | |
| 3450 | int scan_unevictable_handler(struct ctl_table *table, int write, |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 3451 | void __user *buffer, |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3452 | size_t *length, loff_t *ppos) |
| 3453 | { |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 3454 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3455 | |
| 3456 | if (write && *(unsigned long *)table->data) |
| 3457 | scan_all_zones_unevictable_pages(); |
| 3458 | |
| 3459 | scan_unevictable_pages = 0; |
| 3460 | return 0; |
| 3461 | } |
| 3462 | |
Thadeu Lima de Souza Cascardo | e4455ab | 2010-10-26 14:21:28 -0700 | [diff] [blame] | 3463 | #ifdef CONFIG_NUMA |
Lee Schermerhorn | af936a1 | 2008-10-18 20:26:53 -0700 | [diff] [blame] | 3464 | /* |
| 3465 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of |
| 3466 | * a specified node's per zone unevictable lists for evictable pages. |
| 3467 | */ |
| 3468 | |
| 3469 | static ssize_t read_scan_unevictable_node(struct sys_device *dev, |
| 3470 | struct sysdev_attribute *attr, |
| 3471 | char *buf) |
| 3472 | { |
| 3473 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
| 3474 | } |
| 3475 | |
| 3476 | static ssize_t write_scan_unevictable_node(struct sys_device *dev, |
| 3477 | struct sysdev_attribute *attr, |
| 3478 | const char *buf, size_t count) |
| 3479 | { |
| 3480 | struct zone *node_zones = NODE_DATA(dev->id)->node_zones; |
| 3481 | struct zone *zone; |
| 3482 | unsigned long res; |
| 3483 | unsigned long req = strict_strtoul(buf, 10, &res); |
| 3484 | |
| 3485 | if (!req) |
| 3486 | return 1; /* zero is no-op */ |
| 3487 | |
| 3488 | for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { |
| 3489 | if (!populated_zone(zone)) |
| 3490 | continue; |
| 3491 | scan_zone_unevictable_pages(zone); |
| 3492 | } |
| 3493 | return 1; |
| 3494 | } |
| 3495 | |
| 3496 | |
| 3497 | static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
| 3498 | read_scan_unevictable_node, |
| 3499 | write_scan_unevictable_node); |
| 3500 | |
| 3501 | int scan_unevictable_register_node(struct node *node) |
| 3502 | { |
| 3503 | return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); |
| 3504 | } |
| 3505 | |
| 3506 | void scan_unevictable_unregister_node(struct node *node) |
| 3507 | { |
| 3508 | sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); |
| 3509 | } |
Thadeu Lima de Souza Cascardo | e4455ab | 2010-10-26 14:21:28 -0700 | [diff] [blame] | 3510 | #endif |