drivers/staging/android/lowmemorykiller.c

   1 /* drivers/misc/lowmemorykiller.c
   2  *
   3  * The lowmemorykiller driver lets user-space specify a set of memory thresholds
   4  * where processes with a range of oom_score_adj values will get killed. Specify
   5  * the minimum oom_score_adj values in
   6  * /sys/module/lowmemorykiller/parameters/adj and the number of free pages in
   7  * /sys/module/lowmemorykiller/parameters/minfree. Both files take a comma
   8  * separated list of numbers in ascending order.
   9  *
  10  * For example, write "0,8" to /sys/module/lowmemorykiller/parameters/adj and
  11  * "1024,4096" to /sys/module/lowmemorykiller/parameters/minfree to kill
  12  * processes with a oom_score_adj value of 8 or higher when the free memory
  13  * drops below 4096 pages and kill processes with a oom_score_adj value of 0 or
  14  * higher when the free memory drops below 1024 pages.
  15  *
  16  * The driver considers memory used for caches to be free, but if a large
  17  * percentage of the cached memory is locked this can be very inaccurate
  18  * and processes may not get killed until the normal oom killer is triggered.
  19  *
  20  * Copyright (C) 2007-2008 Google, Inc.
  21  *
  22  * This software is licensed under the terms of the GNU General Public
  23  * License version 2, as published by the Free Software Foundation, and
  24  * may be copied, distributed, and modified under those terms.
  25  *
  26  * This program is distributed in the hope that it will be useful,
  27  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  28  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  29  * GNU General Public License for more details.
  30  *
  31  */
  32
  33 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  34
  35 #include <linux/init.h>
  36 #include <linux/moduleparam.h>
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/oom.h>
  40 #include <linux/sched.h>
  41 #include <linux/swap.h>
  42 #include <linux/rcupdate.h>
  43 #include <linux/profile.h>
  44 #include <linux/notifier.h>
  45
  46 #define CREATE_TRACE_POINTS
  47 #include "trace/lowmemorykiller.h"
  48
  49 static uint32_t lowmem_debug_level = 1;
  50 static short lowmem_adj[6] = {
  51         0,
  52         1,
  53         6,
  54         12,
  55 };
  56 static int lowmem_adj_size = 4;
  57 static int lowmem_minfree[6] = {
  58         3 * 512,        /* 6MB */
  59         2 * 1024,       /* 8MB */
  60         4 * 1024,       /* 16MB */
  61         16 * 1024,      /* 64MB */
  62 };
  63 static int lowmem_minfree_size = 4;
  64
  65 static unsigned long lowmem_deathpending_timeout;
  66
  67 #define lowmem_print(level, x...)                       \
  68         do {                                            \
  69                 if (lowmem_debug_level >= (level))      \
  70                         pr_info(x);                     \
  71         } while (0)
  72
  73 static unsigned long lowmem_count(struct shrinker *s,
  74                                   struct shrink_control *sc)
  75 {
  76         return global_page_state(NR_ACTIVE_ANON) +
  77                 global_page_state(NR_ACTIVE_FILE) +
  78                 global_page_state(NR_INACTIVE_ANON) +
  79                 global_page_state(NR_INACTIVE_FILE);
  80 }
  81
  82 static unsigned long lowmem_scan(struct shrinker *s, struct shrink_control *sc)
  83 {
  84         struct task_struct *tsk;
  85         struct task_struct *selected = NULL;
  86         unsigned long rem = 0;
  87         int tasksize;
  88         int i;
  89         short min_score_adj = OOM_SCORE_ADJ_MAX + 1;
  90         int minfree = 0;
  91         int selected_tasksize = 0;
  92         short selected_oom_score_adj;
  93         int array_size = ARRAY_SIZE(lowmem_adj);
  94         int other_free = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
  95         int other_file = global_page_state(NR_FILE_PAGES) -
  96                                                 global_page_state(NR_SHMEM) -
  97                                                 total_swapcache_pages();
  98
  99         if (lowmem_adj_size < array_size)
 100                 array_size = lowmem_adj_size;
 101         if (lowmem_minfree_size < array_size)
 102                 array_size = lowmem_minfree_size;
 103         for (i = 0; i < array_size; i++) {
 104                 minfree = lowmem_minfree[i];
 105                 if (other_free < minfree && other_file < minfree) {
 106                         min_score_adj = lowmem_adj[i];
 107                         break;
 108                 }
 109         }
 110
 111         lowmem_print(3, "lowmem_scan %lu, %x, ofree %d %d, ma %hd\n",
 112                         sc->nr_to_scan, sc->gfp_mask, other_free,
 113                         other_file, min_score_adj);
 114
 115         if (min_score_adj == OOM_SCORE_ADJ_MAX + 1) {
 116                 lowmem_print(5, "lowmem_scan %lu, %x, return 0\n",
 117                              sc->nr_to_scan, sc->gfp_mask);
 118                 return 0;
 119         }
 120
 121         selected_oom_score_adj = min_score_adj;
 122
 123         rcu_read_lock();
 124         for_each_process(tsk) {
 125                 struct task_struct *p;
 126                 short oom_score_adj;
 127
 128                 if (tsk->flags & PF_KTHREAD)
 129                         continue;
 130
 131                 p = find_lock_task_mm(tsk);
 132                 if (!p)
 133                         continue;
 134
 135                 if (test_tsk_thread_flag(p, TIF_MEMDIE) &&
 136                     time_before_eq(jiffies, lowmem_deathpending_timeout)) {
 137                         task_unlock(p);
 138                         rcu_read_unlock();
 139                         return 0;
 140                 }
 141                 oom_score_adj = p->signal->oom_score_adj;
 142                 if (oom_score_adj < min_score_adj) {
 143                         task_unlock(p);
 144                         continue;
 145                 }
 146                 tasksize = get_mm_rss(p->mm);
 147                 task_unlock(p);
 148                 if (tasksize <= 0)
 149                         continue;
 150                 if (selected) {
 151                         if (oom_score_adj < selected_oom_score_adj)
 152                                 continue;
 153                         if (oom_score_adj == selected_oom_score_adj &&
 154                             tasksize <= selected_tasksize)
 155                                 continue;
 156                 }
 157                 selected = p;
 158                 selected_tasksize = tasksize;
 159                 selected_oom_score_adj = oom_score_adj;
 160                 lowmem_print(2, "select '%s' (%d), adj %hd, size %d, to kill\n",
 161                              p->comm, p->pid, oom_score_adj, tasksize);
 162         }
 163         if (selected) {
 164                 task_lock(selected);
 165                 send_sig(SIGKILL, selected, 0);
 166                 /*
 167                  * FIXME: lowmemorykiller shouldn't abuse global OOM killer
 168                  * infrastructure. There is no real reason why the selected
 169                  * task should have access to the memory reserves.
 170                  */
 171                 if (selected->mm)
 172                         mark_oom_victim(selected);
 173                 task_unlock(selected);
 174                 long cache_size = other_file * (long)(PAGE_SIZE / 1024);
 175                 long cache_limit = minfree * (long)(PAGE_SIZE / 1024);
 176                 long free = other_free * (long)(PAGE_SIZE / 1024);
 177                 trace_lowmemory_kill(selected, cache_size, cache_limit, free);
 178                 lowmem_print(1, "Killing '%s' (%d), adj %hd,\n" \
 179                                 "   to free %ldkB on behalf of '%s' (%d) because\n" \
 180                                 "   cache %ldkB is below limit %ldkB for oom_score_adj %hd\n" \
 181                                 "   Free memory is %ldkB above reserved\n",
 182                              selected->comm, selected->pid,
 183                              selected_oom_score_adj,
 184                              selected_tasksize * (long)(PAGE_SIZE / 1024),
 185                              current->comm, current->pid,
 186                              cache_size, cache_limit,
 187                              min_score_adj,
 188                              free);
 189                 lowmem_deathpending_timeout = jiffies + HZ;
 190                 rem += selected_tasksize;
 191         }
 192
 193         lowmem_print(4, "lowmem_scan %lu, %x, return %lu\n",
 194                      sc->nr_to_scan, sc->gfp_mask, rem);
 195         rcu_read_unlock();
 196         return rem;
 197 }
 198
 199 static struct shrinker lowmem_shrinker = {
 200         .scan_objects = lowmem_scan,
 201         .count_objects = lowmem_count,
 202         .seeks = DEFAULT_SEEKS * 16
 203 };
 204
 205 static int __init lowmem_init(void)
 206 {
 207         register_shrinker(&lowmem_shrinker);
 208         return 0;
 209 }
 210 device_initcall(lowmem_init);
 211
 212 #ifdef CONFIG_ANDROID_LOW_MEMORY_KILLER_AUTODETECT_OOM_ADJ_VALUES
 213 static short lowmem_oom_adj_to_oom_score_adj(short oom_adj)
 214 {
 215         if (oom_adj == OOM_ADJUST_MAX)
 216                 return OOM_SCORE_ADJ_MAX;
 217         else
 218                 return (oom_adj * OOM_SCORE_ADJ_MAX) / -OOM_DISABLE;
 219 }
 220
 221 static void lowmem_autodetect_oom_adj_values(void)
 222 {
 223         int i;
 224         short oom_adj;
 225         short oom_score_adj;
 226         int array_size = ARRAY_SIZE(lowmem_adj);
 227
 228         if (lowmem_adj_size < array_size)
 229                 array_size = lowmem_adj_size;
 230
 231         if (array_size <= 0)
 232                 return;
 233
 234         oom_adj = lowmem_adj[array_size - 1];
 235         if (oom_adj > OOM_ADJUST_MAX)
 236                 return;
 237
 238         oom_score_adj = lowmem_oom_adj_to_oom_score_adj(oom_adj);
 239         if (oom_score_adj <= OOM_ADJUST_MAX)
 240                 return;
 241
 242         lowmem_print(1, "lowmem_shrink: convert oom_adj to oom_score_adj:\n");
 243         for (i = 0; i < array_size; i++) {
 244                 oom_adj = lowmem_adj[i];
 245                 oom_score_adj = lowmem_oom_adj_to_oom_score_adj(oom_adj);
 246                 lowmem_adj[i] = oom_score_adj;
 247                 lowmem_print(1, "oom_adj %d => oom_score_adj %d\n",
 248                              oom_adj, oom_score_adj);
 249         }
 250 }
 251
 252 static int lowmem_adj_array_set(const char *val, const struct kernel_param *kp)
 253 {
 254         int ret;
 255
 256         ret = param_array_ops.set(val, kp);
 257
 258         /* HACK: Autodetect oom_adj values in lowmem_adj array */
 259         lowmem_autodetect_oom_adj_values();
 260
 261         return ret;
 262 }
 263
 264 static int lowmem_adj_array_get(char *buffer, const struct kernel_param *kp)
 265 {
 266         return param_array_ops.get(buffer, kp);
 267 }
 268
 269 static void lowmem_adj_array_free(void *arg)
 270 {
 271         param_array_ops.free(arg);
 272 }
 273
 274 static struct kernel_param_ops lowmem_adj_array_ops = {
 275         .set = lowmem_adj_array_set,
 276         .get = lowmem_adj_array_get,
 277         .free = lowmem_adj_array_free,
 278 };
 279
 280 static const struct kparam_array __param_arr_adj = {
 281         .max = ARRAY_SIZE(lowmem_adj),
 282         .num = &lowmem_adj_size,
 283         .ops = &param_ops_short,
 284         .elemsize = sizeof(lowmem_adj[0]),
 285         .elem = lowmem_adj,
 286 };
 287 #endif
 288
 289 /*
 290  * not really modular, but the easiest way to keep compat with existing
 291  * bootargs behaviour is to continue using module_param here.
 292  */
 293 module_param_named(cost, lowmem_shrinker.seeks, int, S_IRUGO | S_IWUSR);
 294 #ifdef CONFIG_ANDROID_LOW_MEMORY_KILLER_AUTODETECT_OOM_ADJ_VALUES
 295 __module_param_call(MODULE_PARAM_PREFIX, adj,
 296                     &lowmem_adj_array_ops,
 297                     .arr = &__param_arr_adj,
 298                     S_IRUGO | S_IWUSR, -1);
 299 __MODULE_PARM_TYPE(adj, "array of short");
 300 #else
 301 module_param_array_named(adj, lowmem_adj, short, &lowmem_adj_size,
 302                          S_IRUGO | S_IWUSR);
 303 #endif
 304 module_param_array_named(minfree, lowmem_minfree, uint, &lowmem_minfree_size,
 305                          S_IRUGO | S_IWUSR);
 306 module_param_named(debug_level, lowmem_debug_level, uint, S_IRUGO | S_IWUSR);
 307