blob: 8a5467ee6265deb11a48fe9861348dd182dc9f7f [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
Randy Dunlap1b578df2008-03-19 17:00:42 -070040 * @mem: target memory controller
Linus Torvalds1da177e2005-04-16 15:20:36 -070041 *
42 * The formula used is relatively simple and documented inline in the
43 * function. The main rationale is that we want to select a good task
44 * to kill when we run out of memory.
45 *
46 * Good in this context means that:
47 * 1) we lose the minimum amount of work done
48 * 2) we recover a large amount of memory
49 * 3) we don't kill anything innocent of eating tons of memory
50 * 4) we want to kill the minimum amount of processes (one)
51 * 5) we try to kill the process the user expects us to kill, this
52 * algorithm has been meticulously tuned to meet the principle
53 * of least surprise ... (be careful when you change it)
54 */
55
Li Zefan97d87c92008-04-28 02:13:35 -070056unsigned long badness(struct task_struct *p, unsigned long uptime)
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
Mel Gorman54a6eb52008-04-28 02:12:16 -0700177 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -0700178 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -0700179 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
Christoph Lameteree31af52007-10-16 01:25:30 -0700180 nodemask_t nodes = node_states[N_HIGH_MEMORY];
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800181
Mel Gorman54a6eb52008-04-28 02:12:16 -0700182 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
183 if (cpuset_zone_allowed_softwall(zone, gfp_mask))
184 node_clear(zone_to_nid(zone), nodes);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800185 else
186 return CONSTRAINT_CPUSET;
187
188 if (!nodes_empty(nodes))
189 return CONSTRAINT_MEMORY_POLICY;
190#endif
191
192 return CONSTRAINT_NONE;
193}
194
195/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196 * Simple selection loop. We chose the process with the highest
197 * number of 'points'. We expect the caller will lock the tasklist.
198 *
199 * (not docbooked, we don't want this one cluttering up the manual)
200 */
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800201static struct task_struct *select_bad_process(unsigned long *ppoints,
202 struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 struct task_struct *g, *p;
205 struct task_struct *chosen = NULL;
206 struct timespec uptime;
Kurt Garloff9827b782006-02-20 18:27:51 -0800207 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
209 do_posix_clock_monotonic_gettime(&uptime);
Paul Jacksona49335c2005-09-06 15:18:09 -0700210 do_each_thread(g, p) {
211 unsigned long points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700212
Oleg Nesterov28324d12006-09-29 02:01:12 -0700213 /*
214 * skip kernel threads and tasks which have already released
215 * their mm.
216 */
Nick Piggin5081dde2006-09-25 23:31:32 -0700217 if (!p->mm)
218 continue;
Oleg Nesterov28324d12006-09-29 02:01:12 -0700219 /* skip the init task */
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700220 if (is_global_init(p))
Paul Jacksona49335c2005-09-06 15:18:09 -0700221 continue;
David Rientjes4c4a2212008-02-07 00:14:06 -0800222 if (mem && !task_in_mem_cgroup(p, mem))
223 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700224
Paul Jacksona49335c2005-09-06 15:18:09 -0700225 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700226 * This task already has access to memory reserves and is
227 * being killed. Don't allow any other task access to the
228 * memory reserve.
229 *
230 * Note: this may have a chance of deadlock if it gets
231 * blocked waiting for another task which itself is waiting
232 * for memory. Is there a better alternative?
233 */
234 if (test_tsk_thread_flag(p, TIF_MEMDIE))
235 return ERR_PTR(-1UL);
236
237 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700238 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700239 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700240 *
241 * However, if p is the current task, we allow the 'kill' to
242 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
243 * which will allow it to gain access to memory reserves in
244 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700245 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700246 */
Nick Pigginb78483a2006-09-29 02:01:14 -0700247 if (p->flags & PF_EXITING) {
248 if (p != current)
249 return ERR_PTR(-1UL);
250
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700251 chosen = p;
252 *ppoints = ULONG_MAX;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700253 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700254
Nick Piggin4a3ede12006-09-25 23:31:30 -0700255 if (p->oomkilladj == OOM_DISABLE)
256 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
Li Zefan97d87c92008-04-28 02:13:35 -0700258 points = badness(p, uptime.tv_sec);
Kurt Garloff9827b782006-02-20 18:27:51 -0800259 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700260 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800261 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 }
Paul Jacksona49335c2005-09-06 15:18:09 -0700263 } while_each_thread(g, p);
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700264
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 return chosen;
266}
267
268/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700269 * dump_tasks - dump current memory state of all system tasks
270 * @mem: target memory controller
271 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800272 * Dumps the current memory state of all system tasks, excluding kernel threads.
273 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
274 * score, and name.
275 *
276 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
277 * shown.
278 *
279 * Call with tasklist_lock read-locked.
280 */
281static void dump_tasks(const struct mem_cgroup *mem)
282{
283 struct task_struct *g, *p;
284
285 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
286 "name\n");
287 do_each_thread(g, p) {
288 /*
289 * total_vm and rss sizes do not exist for tasks with a
290 * detached mm so there's no need to report them.
291 */
292 if (!p->mm)
293 continue;
294 if (mem && !task_in_mem_cgroup(p, mem))
295 continue;
296
297 task_lock(p);
298 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
299 p->pid, p->uid, p->tgid, p->mm->total_vm,
300 get_mm_rss(p->mm), (int)task_cpu(p), p->oomkilladj,
301 p->comm);
302 task_unlock(p);
303 } while_each_thread(g, p);
304}
305
Randy Dunlap1b578df2008-03-19 17:00:42 -0700306/*
Ram Gupta5a291b982006-09-25 23:31:54 -0700307 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
308 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
309 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310 */
Nick Pigginf3af38d2006-12-06 20:31:51 -0800311static void __oom_kill_task(struct task_struct *p, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312{
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700313 if (is_global_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 WARN_ON(1);
315 printk(KERN_WARNING "tried to kill init!\n");
316 return;
317 }
318
Oleg Nesterov01017a22006-09-29 02:01:13 -0700319 if (!p->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 WARN_ON(1);
321 printk(KERN_WARNING "tried to kill an mm-less task!\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 return;
323 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700324
Nick Pigginf3af38d2006-12-06 20:31:51 -0800325 if (verbose)
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700326 printk(KERN_ERR "Killed process %d (%s)\n",
327 task_pid_nr(p), p->comm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329 /*
330 * We give our sacrificial lamb high priority and access to
331 * all the memory it needs. That way it should be able to
332 * exit() and clear out its resources quickly...
333 */
Peter Zijlstrafa717062008-01-25 21:08:27 +0100334 p->rt.time_slice = HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 set_tsk_thread_flag(p, TIF_MEMDIE);
336
337 force_sig(SIGKILL, p);
338}
339
Nick Pigginf3af38d2006-12-06 20:31:51 -0800340static int oom_kill_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
Dave Peterson01315922006-04-18 22:20:44 -0700342 struct mm_struct *mm;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700343 struct task_struct *g, *q;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
Dave Peterson01315922006-04-18 22:20:44 -0700345 mm = p->mm;
346
347 /* WARNING: mm may not be dereferenced since we did not obtain its
348 * value from get_task_mm(p). This is OK since all we need to do is
349 * compare mm to q->mm below.
350 *
351 * Furthermore, even if mm contains a non-NULL value, p->mm may
352 * change to NULL at any time since we do not hold task_lock(p).
353 * However, this is of no concern to us.
354 */
355
Oleg Nesterov01017a22006-09-29 02:01:13 -0700356 if (mm == NULL)
Dave Peterson01315922006-04-18 22:20:44 -0700357 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800359 /*
360 * Don't kill the process if any threads are set to OOM_DISABLE
361 */
362 do_each_thread(g, q) {
Ankita Garg35ae8342007-03-16 13:38:33 -0800363 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800364 return 1;
365 } while_each_thread(g, q);
366
Nick Pigginf3af38d2006-12-06 20:31:51 -0800367 __oom_kill_task(p, 1);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800368
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 /*
370 * kill all processes that share the ->mm (i.e. all threads),
Nick Pigginf2a2a712006-12-06 20:31:52 -0800371 * but are in a different thread group. Don't let them have access
372 * to memory reserves though, otherwise we might deplete all memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 */
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800374 do_each_thread(g, q) {
Pavel Emelyanovbac0abd2007-10-18 23:40:18 -0700375 if (q->mm == mm && !same_thread_group(q, p))
David Rientjes650a7c92007-04-23 21:36:13 -0700376 force_sig(SIGKILL, q);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800377 } while_each_thread(g, q);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378
Dave Peterson01315922006-04-18 22:20:44 -0700379 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380}
381
David Rientjes7213f502007-10-16 23:25:57 -0700382static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesfef1bdd2008-02-07 00:14:07 -0800383 unsigned long points, struct mem_cgroup *mem,
384 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 struct task_struct *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
David Rientjes7213f502007-10-16 23:25:57 -0700388 if (printk_ratelimit()) {
389 printk(KERN_WARNING "%s invoked oom-killer: "
390 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
391 current->comm, gfp_mask, order, current->oomkilladj);
392 dump_stack();
393 show_mem();
David Rientjesfef1bdd2008-02-07 00:14:07 -0800394 if (sysctl_oom_dump_tasks)
395 dump_tasks(mem);
David Rientjes7213f502007-10-16 23:25:57 -0700396 }
397
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700398 /*
399 * If the task is already exiting, don't alarm the sysadmin or kill
400 * its children or threads, just set TIF_MEMDIE so it can die quickly
401 */
402 if (p->flags & PF_EXITING) {
Nick Pigginf3af38d2006-12-06 20:31:51 -0800403 __oom_kill_task(p, 0);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700404 return 0;
405 }
406
Nick Pigginf3af38d2006-12-06 20:31:51 -0800407 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700408 message, task_pid_nr(p), p->comm, points);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800409
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 /* Try to kill a child first */
Matthias Kaehlcke7b1915a2007-10-18 23:39:58 -0700411 list_for_each_entry(c, &p->children, sibling) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 if (c->mm == p->mm)
413 continue;
Nick Pigginf3af38d2006-12-06 20:31:51 -0800414 if (!oom_kill_task(c))
Dave Peterson01315922006-04-18 22:20:44 -0700415 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 }
Nick Pigginf3af38d2006-12-06 20:31:51 -0800417 return oom_kill_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418}
419
Balbir Singh00f0b822008-03-04 14:28:39 -0800420#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800421void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
422{
423 unsigned long points = 0;
424 struct task_struct *p;
425
426 cgroup_lock();
Li Zefane115f2d2008-04-15 14:34:37 -0700427 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800428retry:
429 p = select_bad_process(&points, mem);
430 if (PTR_ERR(p) == -1UL)
431 goto out;
432
433 if (!p)
434 p = current;
435
David Rientjesfef1bdd2008-02-07 00:14:07 -0800436 if (oom_kill_process(p, gfp_mask, 0, points, mem,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800437 "Memory cgroup out of memory"))
438 goto retry;
439out:
Li Zefane115f2d2008-04-15 14:34:37 -0700440 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800441 cgroup_unlock();
442}
443#endif
444
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700445static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
446
447int register_oom_notifier(struct notifier_block *nb)
448{
449 return blocking_notifier_chain_register(&oom_notify_list, nb);
450}
451EXPORT_SYMBOL_GPL(register_oom_notifier);
452
453int unregister_oom_notifier(struct notifier_block *nb)
454{
455 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
456}
457EXPORT_SYMBOL_GPL(unregister_oom_notifier);
458
David Rientjes098d7f12007-10-16 23:25:55 -0700459/*
460 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
461 * if a parallel OOM killing is already taking place that includes a zone in
462 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
463 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700464int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700465{
Mel Gormandd1a2392008-04-28 02:12:17 -0700466 struct zoneref *z;
467 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700468 int ret = 1;
469
David Rientjesae741382007-10-16 23:26:00 -0700470 spin_lock(&zone_scan_mutex);
Mel Gormandd1a2392008-04-28 02:12:17 -0700471 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
472 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700473 ret = 0;
474 goto out;
475 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700476 }
David Rientjes098d7f12007-10-16 23:25:55 -0700477
Mel Gormandd1a2392008-04-28 02:12:17 -0700478 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
479 /*
480 * Lock each zone in the zonelist under zone_scan_mutex so a
481 * parallel invocation of try_set_zone_oom() doesn't succeed
482 * when it shouldn't.
483 */
484 zone_set_flag(zone, ZONE_OOM_LOCKED);
485 }
486
David Rientjes098d7f12007-10-16 23:25:55 -0700487out:
David Rientjesae741382007-10-16 23:26:00 -0700488 spin_unlock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700489 return ret;
490}
491
492/*
493 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
494 * allocation attempts with zonelists containing them may now recall the OOM
495 * killer, if necessary.
496 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700497void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700498{
Mel Gormandd1a2392008-04-28 02:12:17 -0700499 struct zoneref *z;
500 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700501
David Rientjesae741382007-10-16 23:26:00 -0700502 spin_lock(&zone_scan_mutex);
Mel Gormandd1a2392008-04-28 02:12:17 -0700503 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
504 zone_clear_flag(zone, ZONE_OOM_LOCKED);
505 }
David Rientjesae741382007-10-16 23:26:00 -0700506 spin_unlock(&zone_scan_mutex);
David Rientjes098d7f12007-10-16 23:25:55 -0700507}
508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509/**
Dave Peterson6937a252006-06-23 02:03:13 -0700510 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700511 * @zonelist: zonelist pointer
512 * @gfp_mask: memory allocation flags
513 * @order: amount of memory being requested as a power of 2
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 *
515 * If we run out of memory, we have the choice between either
516 * killing a random task (bad), letting the system crash (worse)
517 * OR try to be smart about which process to kill. Note that we
518 * don't have to be perfect here, we just have to be good.
519 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800520void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
Ingo Molnar36c8b582006-07-03 00:25:41 -0700522 struct task_struct *p;
Andrew Mortond6713e02006-02-28 16:59:19 -0800523 unsigned long points = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700524 unsigned long freed = 0;
David Rientjes70e24bd2007-10-16 23:25:53 -0700525 enum oom_constraint constraint;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700526
527 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
528 if (freed > 0)
529 /* Got some memory back in the last second. */
530 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531
Yasunori Goto2b744c02007-05-06 14:49:59 -0700532 if (sysctl_panic_on_oom == 2)
533 panic("out of memory. Compulsory panic_on_oom is selected.\n");
534
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800535 /*
536 * Check if there were limitations on the allocation (only relevant for
537 * NUMA) that may require different handling.
538 */
David Rientjes2b45ab32007-05-06 14:50:00 -0700539 constraint = constrained_alloc(zonelist, gfp_mask);
David Rientjes2b45ab32007-05-06 14:50:00 -0700540 read_lock(&tasklist_lock);
541
542 switch (constraint) {
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800543 case CONSTRAINT_MEMORY_POLICY:
David Rientjesfef1bdd2008-02-07 00:14:07 -0800544 oom_kill_process(current, gfp_mask, order, points, NULL,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800545 "No available memory (MPOL_BIND)");
546 break;
547
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800548 case CONSTRAINT_NONE:
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700549 if (sysctl_panic_on_oom)
550 panic("out of memory. panic_on_oom is selected\n");
David Rientjesfe071d72007-10-16 23:25:56 -0700551 /* Fall-through */
552 case CONSTRAINT_CPUSET:
553 if (sysctl_oom_kill_allocating_task) {
David Rientjesfef1bdd2008-02-07 00:14:07 -0800554 oom_kill_process(current, gfp_mask, order, points, NULL,
David Rientjesfe071d72007-10-16 23:25:56 -0700555 "Out of memory (oom_kill_allocating_task)");
556 break;
557 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558retry:
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800559 /*
560 * Rambo mode: Shoot down a process and hope it solves whatever
561 * issues we may have.
562 */
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800563 p = select_bad_process(&points, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800565 if (PTR_ERR(p) == -1UL)
566 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800568 /* Found nothing?!?! Either we hang forever, or we panic. */
569 if (!p) {
570 read_unlock(&tasklist_lock);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800571 panic("Out of memory and no killable processes...\n");
572 }
573
David Rientjesfef1bdd2008-02-07 00:14:07 -0800574 if (oom_kill_process(p, gfp_mask, order, points, NULL,
David Rientjes7213f502007-10-16 23:25:57 -0700575 "Out of memory"))
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800576 goto retry;
577
578 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 }
580
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800581out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800582 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
584 /*
585 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800586 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800588 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800589 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590}