blob: 44b2da11bf43f75a3714761ab09217339579977d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/oom_kill.c
3 *
4 * Copyright (C) 1998,2000 Rik van Riel
5 * Thanks go out to Claus Fischer for some serious inspiration and
6 * for goading me into coding this file...
7 *
8 * The routines in this file are used to kill a process when
Paul Jacksona49335c2005-09-06 15:18:09 -07009 * we're seriously out of memory. This gets called from __alloc_pages()
10 * in mm/page_alloc.c when we really run out of memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 *
12 * Since we won't call these routines often (on a well-configured
13 * machine) this file will double as a 'coding guide' and a signpost
14 * for newbie kernel hackers. It features several pointers to major
15 * kernel subsystems and hints as to where to find out what things do.
16 */
17
Alexey Dobriyan8ac773b2006-10-19 23:28:32 -070018#include <linux/oom.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/mm.h>
Alexey Dobriyan4e950f62007-07-30 02:36:13 +040020#include <linux/err.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/sched.h>
22#include <linux/swap.h>
23#include <linux/timex.h>
24#include <linux/jiffies.h>
Paul Jacksonef08e3b2005-09-06 15:18:13 -070025#include <linux/cpuset.h>
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -070026#include <linux/module.h>
27#include <linux/notifier.h>
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -080028#include <linux/memcontrol.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070030int sysctl_panic_on_oom;
David Rientjesfe071d72007-10-16 23:25:56 -070031int sysctl_oom_kill_allocating_task;
David Rientjesfef1bdd2008-02-07 00:14:07 -080032int sysctl_oom_dump_tasks;
David Rientjesae741382007-10-16 23:26:00 -070033static DEFINE_SPINLOCK(zone_scan_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -070034/* #define DEBUG */
35
36/**
Dave Peterson6937a252006-06-23 02:03:13 -070037 * badness - calculate a numeric value for how bad this task has been
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 * @p: task struct of which task we should calculate
Paul Jacksona49335c2005-09-06 15:18:09 -070039 * @uptime: current uptime in seconds
Linus Torvalds1da177e2005-04-16 15:20:36 -070040 *
41 * The formula used is relatively simple and documented inline in the
42 * function. The main rationale is that we want to select a good task
43 * to kill when we run out of memory.
44 *
45 * Good in this context means that:
46 * 1) we lose the minimum amount of work done
47 * 2) we recover a large amount of memory
48 * 3) we don't kill anything innocent of eating tons of memory
49 * 4) we want to kill the minimum amount of processes (one)
50 * 5) we try to kill the process the user expects us to kill, this
51 * algorithm has been meticulously tuned to meet the principle
52 * of least surprise ... (be careful when you change it)
53 */
54
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -080055unsigned long badness(struct task_struct *p, unsigned long uptime,
56 struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -070057{
58 unsigned long points, cpu_time, run_time, s;
Andrew Morton97c2c9b2006-04-18 22:20:38 -070059 struct mm_struct *mm;
60 struct task_struct *child;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
Andrew Morton97c2c9b2006-04-18 22:20:38 -070062 task_lock(p);
63 mm = p->mm;
64 if (!mm) {
65 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 return 0;
Andrew Morton97c2c9b2006-04-18 22:20:38 -070067 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
69 /*
70 * The memory size of the process is the basis for the badness.
71 */
Andrew Morton97c2c9b2006-04-18 22:20:38 -070072 points = mm->total_vm;
73
74 /*
75 * After this unlock we can no longer dereference local variable `mm'
76 */
77 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
79 /*
Hugh Dickins7ba34852007-01-05 16:37:03 -080080 * swapoff can easily use up all memory, so kill those first.
81 */
82 if (p->flags & PF_SWAPOFF)
83 return ULONG_MAX;
84
85 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 * Processes which fork a lot of child processes are likely
Kurt Garloff9827b782006-02-20 18:27:51 -080087 * a good choice. We add half the vmsize of the children if they
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 * have an own mm. This prevents forking servers to flood the
Kurt Garloff9827b782006-02-20 18:27:51 -080089 * machine with an endless amount of children. In case a single
90 * child is eating the vast majority of memory, adding only half
91 * to the parents will make the child our kill candidate of choice.
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 */
Andrew Morton97c2c9b2006-04-18 22:20:38 -070093 list_for_each_entry(child, &p->children, sibling) {
94 task_lock(child);
95 if (child->mm != mm && child->mm)
96 points += child->mm->total_vm/2 + 1;
97 task_unlock(child);
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 }
99
100 /*
101 * CPU time is in tens of seconds and run time is in thousands
102 * of seconds. There is no particular reason for this other than
103 * that it turned out to work very well in practice.
104 */
105 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
106 >> (SHIFT_HZ + 3);
107
108 if (uptime >= p->start_time.tv_sec)
109 run_time = (uptime - p->start_time.tv_sec) >> 10;
110 else
111 run_time = 0;
112
113 s = int_sqrt(cpu_time);
114 if (s)
115 points /= s;
116 s = int_sqrt(int_sqrt(run_time));
117 if (s)
118 points /= s;
119
120 /*
121 * Niced processes are most likely less important, so double
122 * their badness points.
123 */
124 if (task_nice(p) > 0)
125 points *= 2;
126
127 /*
128 * Superuser processes are usually more important, so we make it
129 * less likely that we kill those.
130 */
Serge E. Hallyn97829952008-02-04 22:29:47 -0800131 if (__capable(p, CAP_SYS_ADMIN) || __capable(p, CAP_SYS_RESOURCE))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 points /= 4;
133
134 /*
135 * We don't want to kill a process with direct hardware access.
136 * Not only could that mess up the hardware, but usually users
137 * tend to only have this flag set on applications they think
138 * of as important.
139 */
Andrew Morgane338d262008-02-04 22:29:42 -0800140 if (__capable(p, CAP_SYS_RAWIO))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 points /= 4;
142
143 /*
Nick Piggin7887a3d2006-09-25 23:31:29 -0700144 * If p's nodes don't overlap ours, it may still help to kill p
145 * because p may have allocated or otherwise mapped memory on
146 * this node before. However it will be less likely.
147 */
David Rientjesbbe373f2007-10-16 23:25:58 -0700148 if (!cpuset_mems_allowed_intersects(current, p))
Nick Piggin7887a3d2006-09-25 23:31:29 -0700149 points /= 8;
150
151 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 * Adjust the score by oomkilladj.
153 */
154 if (p->oomkilladj) {
Joshua N Pritikin9a827822007-05-06 14:49:07 -0700155 if (p->oomkilladj > 0) {
156 if (!points)
157 points = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 points <<= p->oomkilladj;
Joshua N Pritikin9a827822007-05-06 14:49:07 -0700159 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 points >>= -(p->oomkilladj);
161 }
162
163#ifdef DEBUG
David Rientjesa5e58a62007-07-31 00:37:32 -0700164 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 p->pid, p->comm, points);
166#endif
167 return points;
168}
169
170/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800171 * Determine the type of allocation constraint.
172 */
David Rientjes70e24bd2007-10-16 23:25:53 -0700173static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
174 gfp_t gfp_mask)
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800175{
176#ifdef CONFIG_NUMA
177 struct zone **z;
Christoph Lameteree31af52007-10-16 01:25:30 -0700178 nodemask_t nodes = node_states[N_HIGH_MEMORY];
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800179
180 for (z = zonelist->zones; *z; z++)
Paul Jackson02a0e532006-12-13 00:34:25 -0800181 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
Christoph Lameter89fa3022006-09-25 23:31:55 -0700182 node_clear(zone_to_nid(*z), nodes);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800183 else
184 return CONSTRAINT_CPUSET;
185
186 if (!nodes_empty(nodes))
187 return CONSTRAINT_MEMORY_POLICY;
188#endif
189
190 return CONSTRAINT_NONE;
191}
192
193/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 * Simple selection loop. We chose the process with the highest
195 * number of 'points'. We expect the caller will lock the tasklist.
196 *
197 * (not docbooked, we don't want this one cluttering up the manual)
198 */
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800199static struct task_struct *select_bad_process(unsigned long *ppoints,
200 struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 struct task_struct *g, *p;
203 struct task_struct *chosen = NULL;
204 struct timespec uptime;
Kurt Garloff9827b782006-02-20 18:27:51 -0800205 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 do_posix_clock_monotonic_gettime(&uptime);
Paul Jacksona49335c2005-09-06 15:18:09 -0700208 do_each_thread(g, p) {
209 unsigned long points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700210
Oleg Nesterov28324d12006-09-29 02:01:12 -0700211 /*
212 * skip kernel threads and tasks which have already released
213 * their mm.
214 */
Nick Piggin5081dde2006-09-25 23:31:32 -0700215 if (!p->mm)
216 continue;
Oleg Nesterov28324d12006-09-29 02:01:12 -0700217 /* skip the init task */
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700218 if (is_global_init(p))
Paul Jacksona49335c2005-09-06 15:18:09 -0700219 continue;
David Rientjes4c4a2212008-02-07 00:14:06 -0800220 if (mem && !task_in_mem_cgroup(p, mem))
221 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700222
Paul Jacksona49335c2005-09-06 15:18:09 -0700223 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700224 * This task already has access to memory reserves and is
225 * being killed. Don't allow any other task access to the
226 * memory reserve.
227 *
228 * Note: this may have a chance of deadlock if it gets
229 * blocked waiting for another task which itself is waiting
230 * for memory. Is there a better alternative?
231 */
232 if (test_tsk_thread_flag(p, TIF_MEMDIE))
233 return ERR_PTR(-1UL);
234
235 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700236 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700237 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700238 *
239 * However, if p is the current task, we allow the 'kill' to
240 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
241 * which will allow it to gain access to memory reserves in
242 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700243 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700244 */
Nick Pigginb78483a2006-09-29 02:01:14 -0700245 if (p->flags & PF_EXITING) {
246 if (p != current)
247 return ERR_PTR(-1UL);
248
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700249 chosen = p;
250 *ppoints = ULONG_MAX;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700251 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700252
Nick Piggin4a3ede12006-09-25 23:31:30 -0700253 if (p->oomkilladj == OOM_DISABLE)
254 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800256 points = badness(p, uptime.tv_sec, mem);
Kurt Garloff9827b782006-02-20 18:27:51 -0800257 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700258 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800259 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 }
Paul Jacksona49335c2005-09-06 15:18:09 -0700261 } while_each_thread(g, p);
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 return chosen;
264}
265
266/**
David Rientjesfef1bdd2008-02-07 00:14:07 -0800267 * Dumps the current memory state of all system tasks, excluding kernel threads.
268 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
269 * score, and name.
270 *
271 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
272 * shown.
273 *
274 * Call with tasklist_lock read-locked.
275 */
276static void dump_tasks(const struct mem_cgroup *mem)
277{
278 struct task_struct *g, *p;
279
280 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
281 "name\n");
282 do_each_thread(g, p) {
283 /*
284 * total_vm and rss sizes do not exist for tasks with a
285 * detached mm so there's no need to report them.
286 */
287 if (!p->mm)
288 continue;
289 if (mem && !task_in_mem_cgroup(p, mem))
290 continue;
291
292 task_lock(p);
293 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
294 p->pid, p->uid, p->tgid, p->mm->total_vm,
295 get_mm_rss(p->mm), (int)task_cpu(p), p->oomkilladj,
296 p->comm);
297 task_unlock(p);
298 } while_each_thread(g, p);
299}
300
301/**
Ram Gupta5a291b982006-09-25 23:31:54 -0700302 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
303 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
304 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 */
Nick Pigginf3af38d2006-12-06 20:31:51 -0800306static void __oom_kill_task(struct task_struct *p, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307{
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700308 if (is_global_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 WARN_ON(1);
310 printk(KERN_WARNING "tried to kill init!\n");
311 return;
312 }
313
Oleg Nesterov01017a22006-09-29 02:01:13 -0700314 if (!p->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 WARN_ON(1);
316 printk(KERN_WARNING "tried to kill an mm-less task!\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 return;
318 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700319
Nick Pigginf3af38d2006-12-06 20:31:51 -0800320 if (verbose)
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700321 printk(KERN_ERR "Killed process %d (%s)\n",
322 task_pid_nr(p), p->comm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323
324 /*
325 * We give our sacrificial lamb high priority and access to
326 * all the memory it needs. That way it should be able to
327 * exit() and clear out its resources quickly...
328 */
Peter Zijlstrafa717062008-01-25 21:08:27 +0100329 p->rt.time_slice = HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 set_tsk_thread_flag(p, TIF_MEMDIE);
331
332 force_sig(SIGKILL, p);
333}
334
Nick Pigginf3af38d2006-12-06 20:31:51 -0800335static int oom_kill_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
Dave Peterson01315922006-04-18 22:20:44 -0700337 struct mm_struct *mm;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700338 struct task_struct *g, *q;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
Dave Peterson01315922006-04-18 22:20:44 -0700340 mm = p->mm;
341
342 /* WARNING: mm may not be dereferenced since we did not obtain its
343 * value from get_task_mm(p). This is OK since all we need to do is
344 * compare mm to q->mm below.
345 *
346 * Furthermore, even if mm contains a non-NULL value, p->mm may
347 * change to NULL at any time since we do not hold task_lock(p).
348 * However, this is of no concern to us.
349 */
350
Oleg Nesterov01017a22006-09-29 02:01:13 -0700351 if (mm == NULL)
Dave Peterson01315922006-04-18 22:20:44 -0700352 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800354 /*
355 * Don't kill the process if any threads are set to OOM_DISABLE
356 */
357 do_each_thread(g, q) {
Ankita Garg35ae8342007-03-16 13:38:33 -0800358 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800359 return 1;
360 } while_each_thread(g, q);
361
Nick Pigginf3af38d2006-12-06 20:31:51 -0800362 __oom_kill_task(p, 1);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800363
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 /*
365 * kill all processes that share the ->mm (i.e. all threads),
Nick Pigginf2a2a712006-12-06 20:31:52 -0800366 * but are in a different thread group. Don't let them have access
367 * to memory reserves though, otherwise we might deplete all memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 */
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800369 do_each_thread(g, q) {
Pavel Emelyanovbac0abd2007-10-18 23:40:18 -0700370 if (q->mm == mm && !same_thread_group(q, p))
David Rientjes650a7c92007-04-23 21:36:13 -0700371 force_sig(SIGKILL, q);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800372 } while_each_thread(g, q);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
Dave Peterson01315922006-04-18 22:20:44 -0700374 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
David Rientjes7213f502007-10-16 23:25:57 -0700377static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesfef1bdd2008-02-07 00:14:07 -0800378 unsigned long points, struct mem_cgroup *mem,
379 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 struct task_struct *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
David Rientjes7213f502007-10-16 23:25:57 -0700383 if (printk_ratelimit()) {
384 printk(KERN_WARNING "%s invoked oom-killer: "
385 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
386 current->comm, gfp_mask, order, current->oomkilladj);
387 dump_stack();
388 show_mem();
David Rientjesfef1bdd2008-02-07 00:14:07 -0800389 if (sysctl_oom_dump_tasks)
390 dump_tasks(mem);
David Rientjes7213f502007-10-16 23:25:57 -0700391 }
392
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700393 /*
394 * If the task is already exiting, don't alarm the sysadmin or kill
395 * its children or threads, just set TIF_MEMDIE so it can die quickly
396 */
397 if (p->flags & PF_EXITING) {
Nick Pigginf3af38d2006-12-06 20:31:51 -0800398 __oom_kill_task(p, 0);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700399 return 0;
400 }
401
Nick Pigginf3af38d2006-12-06 20:31:51 -0800402 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700403 message, task_pid_nr(p), p->comm, points);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800404
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 /* Try to kill a child first */
Matthias Kaehlcke7b1915a2007-10-18 23:39:58 -0700406 list_for_each_entry(c, &p->children, sibling) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 if (c->mm == p->mm)
408 continue;
Nick Pigginf3af38d2006-12-06 20:31:51 -0800409 if (!oom_kill_task(c))
Dave Peterson01315922006-04-18 22:20:44 -0700410 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 }
Nick Pigginf3af38d2006-12-06 20:31:51 -0800412 return oom_kill_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413}
414
Balbir Singh00f0b822008-03-04 14:28:39 -0800415#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800416void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
417{
418 unsigned long points = 0;
419 struct task_struct *p;
420
421 cgroup_lock();
422 rcu_read_lock();
423retry:
424 p = select_bad_process(&points, mem);
425 if (PTR_ERR(p) == -1UL)
426 goto out;
427
428 if (!p)
429 p = current;
430
David Rientjesfef1bdd2008-02-07 00:14:07 -0800431 if (oom_kill_process(p, gfp_mask, 0, points, mem,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800432 "Memory cgroup out of memory"))
433 goto retry;
434out:
435 rcu_read_unlock();
436 cgroup_unlock();
437}
438#endif
439
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700440static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
441
442int register_oom_notifier(struct notifier_block *nb)
443{
444 return blocking_notifier_chain_register(&oom_notify_list, nb);
445}
446EXPORT_SYMBOL_GPL(register_oom_notifier);
447
448int unregister_oom_notifier(struct notifier_block *nb)
449{
450 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
451}
452EXPORT_SYMBOL_GPL(unregister_oom_notifier);
453
David Rientjes098d7f12007-10-16 23:25:55 -0700454/*
455 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
456 * if a parallel OOM killing is already taking place that includes a zone in
457 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
458 */
459int try_set_zone_oom(struct zonelist *zonelist)
460{
461 struct zone **z;
462 int ret = 1;
463
464 z = zonelist->zones;
465
David Rientjesae741382007-10-16 23:26:00 -0700466 spin_lock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700467 do {
468 if (zone_is_oom_locked(*z)) {
469 ret = 0;
470 goto out;
471 }
472 } while (*(++z) != NULL);
473
474 /*
475 * Lock each zone in the zonelist under zone_scan_mutex so a parallel
476 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
477 */
478 z = zonelist->zones;
479 do {
480 zone_set_flag(*z, ZONE_OOM_LOCKED);
481 } while (*(++z) != NULL);
482out:
David Rientjesae741382007-10-16 23:26:00 -0700483 spin_unlock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700484 return ret;
485}
486
487/*
488 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
489 * allocation attempts with zonelists containing them may now recall the OOM
490 * killer, if necessary.
491 */
492void clear_zonelist_oom(struct zonelist *zonelist)
493{
494 struct zone **z;
495
496 z = zonelist->zones;
497
David Rientjesae741382007-10-16 23:26:00 -0700498 spin_lock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700499 do {
500 zone_clear_flag(*z, ZONE_OOM_LOCKED);
501 } while (*(++z) != NULL);
David Rientjesae741382007-10-16 23:26:00 -0700502 spin_unlock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700503}
504
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505/**
Dave Peterson6937a252006-06-23 02:03:13 -0700506 * out_of_memory - kill the "best" process when we run out of memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 *
508 * If we run out of memory, we have the choice between either
509 * killing a random task (bad), letting the system crash (worse)
510 * OR try to be smart about which process to kill. Note that we
511 * don't have to be perfect here, we just have to be good.
512 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800513void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514{
Ingo Molnar36c8b582006-07-03 00:25:41 -0700515 struct task_struct *p;
Andrew Mortond6713e02006-02-28 16:59:19 -0800516 unsigned long points = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700517 unsigned long freed = 0;
David Rientjes70e24bd2007-10-16 23:25:53 -0700518 enum oom_constraint constraint;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700519
520 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
521 if (freed > 0)
522 /* Got some memory back in the last second. */
523 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524
Yasunori Goto2b744c02007-05-06 14:49:59 -0700525 if (sysctl_panic_on_oom == 2)
526 panic("out of memory. Compulsory panic_on_oom is selected.\n");
527
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800528 /*
529 * Check if there were limitations on the allocation (only relevant for
530 * NUMA) that may require different handling.
531 */
David Rientjes2b45ab32007-05-06 14:50:00 -0700532 constraint = constrained_alloc(zonelist, gfp_mask);
David Rientjes2b45ab32007-05-06 14:50:00 -0700533 read_lock(&tasklist_lock);
534
535 switch (constraint) {
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800536 case CONSTRAINT_MEMORY_POLICY:
David Rientjesfef1bdd2008-02-07 00:14:07 -0800537 oom_kill_process(current, gfp_mask, order, points, NULL,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800538 "No available memory (MPOL_BIND)");
539 break;
540
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800541 case CONSTRAINT_NONE:
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700542 if (sysctl_panic_on_oom)
543 panic("out of memory. panic_on_oom is selected\n");
David Rientjesfe071d72007-10-16 23:25:56 -0700544 /* Fall-through */
545 case CONSTRAINT_CPUSET:
546 if (sysctl_oom_kill_allocating_task) {
David Rientjesfef1bdd2008-02-07 00:14:07 -0800547 oom_kill_process(current, gfp_mask, order, points, NULL,
David Rientjesfe071d72007-10-16 23:25:56 -0700548 "Out of memory (oom_kill_allocating_task)");
549 break;
550 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551retry:
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800552 /*
553 * Rambo mode: Shoot down a process and hope it solves whatever
554 * issues we may have.
555 */
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800556 p = select_bad_process(&points, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800558 if (PTR_ERR(p) == -1UL)
559 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800561 /* Found nothing?!?! Either we hang forever, or we panic. */
562 if (!p) {
563 read_unlock(&tasklist_lock);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800564 panic("Out of memory and no killable processes...\n");
565 }
566
David Rientjesfef1bdd2008-02-07 00:14:07 -0800567 if (oom_kill_process(p, gfp_mask, order, points, NULL,
David Rientjes7213f502007-10-16 23:25:57 -0700568 "Out of memory"))
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800569 goto retry;
570
571 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 }
573
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800574out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800575 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
577 /*
578 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800579 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800581 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800582 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583}