blob: 7c0dc414dba896486bf786a2342ade250d14ea23 [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/sched.h>
23#include <linux/swap.h>
24#include <linux/timex.h>
25#include <linux/jiffies.h>
Paul Jacksonef08e3b2005-09-06 15:18:13 -070026#include <linux/cpuset.h>
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -070027#include <linux/module.h>
28#include <linux/notifier.h>
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -080029#include <linux/memcontrol.h>
David Howells5cd9c582008-08-14 11:37:28 +010030#include <linux/security.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070032int sysctl_panic_on_oom;
David Rientjesfe071d72007-10-16 23:25:56 -070033int sysctl_oom_kill_allocating_task;
David Rientjesfef1bdd2008-02-07 00:14:07 -080034int sysctl_oom_dump_tasks;
David Rientjesc7d4cae2009-01-06 14:39:00 -080035static DEFINE_SPINLOCK(zone_scan_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070036/* #define DEBUG */
37
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070038/*
39 * Is all threads of the target process nodes overlap ours?
40 */
41static int has_intersects_mems_allowed(struct task_struct *tsk)
42{
43 struct task_struct *t;
44
45 t = tsk;
46 do {
47 if (cpuset_mems_allowed_intersects(current, t))
48 return 1;
49 t = next_thread(t);
50 } while (t != tsk);
51
52 return 0;
53}
54
Linus Torvalds1da177e2005-04-16 15:20:36 -070055/**
Dave Peterson6937a252006-06-23 02:03:13 -070056 * badness - calculate a numeric value for how bad this task has been
Linus Torvalds1da177e2005-04-16 15:20:36 -070057 * @p: task struct of which task we should calculate
Paul Jacksona49335c2005-09-06 15:18:09 -070058 * @uptime: current uptime in seconds
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 *
60 * The formula used is relatively simple and documented inline in the
61 * function. The main rationale is that we want to select a good task
62 * to kill when we run out of memory.
63 *
64 * Good in this context means that:
65 * 1) we lose the minimum amount of work done
66 * 2) we recover a large amount of memory
67 * 3) we don't kill anything innocent of eating tons of memory
68 * 4) we want to kill the minimum amount of processes (one)
69 * 5) we try to kill the process the user expects us to kill, this
70 * algorithm has been meticulously tuned to meet the principle
71 * of least surprise ... (be careful when you change it)
72 */
73
Li Zefan97d87c92008-04-28 02:13:35 -070074unsigned long badness(struct task_struct *p, unsigned long uptime)
Linus Torvalds1da177e2005-04-16 15:20:36 -070075{
Cyrill Gorcunova12888f2009-03-31 15:19:27 -070076 unsigned long points, cpu_time, run_time;
Andrew Morton97c2c9b82006-04-18 22:20:38 -070077 struct mm_struct *mm;
78 struct task_struct *child;
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -070079 int oom_adj = p->signal->oom_adj;
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070080 struct task_cputime task_time;
81 unsigned long utime;
82 unsigned long stime;
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -070083
84 if (oom_adj == OOM_DISABLE)
85 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
Andrew Morton97c2c9b82006-04-18 22:20:38 -070087 task_lock(p);
88 mm = p->mm;
89 if (!mm) {
90 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 return 0;
Andrew Morton97c2c9b82006-04-18 22:20:38 -070092 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
94 /*
95 * The memory size of the process is the basis for the badness.
96 */
Andrew Morton97c2c9b82006-04-18 22:20:38 -070097 points = mm->total_vm;
98
99 /*
100 * After this unlock we can no longer dereference local variable `mm'
101 */
102 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
104 /*
Hugh Dickins7ba34852007-01-05 16:37:03 -0800105 * swapoff can easily use up all memory, so kill those first.
106 */
Hugh Dickins35451be2009-09-21 17:02:27 -0700107 if (p->flags & PF_OOM_ORIGIN)
Hugh Dickins7ba34852007-01-05 16:37:03 -0800108 return ULONG_MAX;
109
110 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 * Processes which fork a lot of child processes are likely
Kurt Garloff9827b782006-02-20 18:27:51 -0800112 * a good choice. We add half the vmsize of the children if they
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 * have an own mm. This prevents forking servers to flood the
Kurt Garloff9827b782006-02-20 18:27:51 -0800114 * machine with an endless amount of children. In case a single
115 * child is eating the vast majority of memory, adding only half
116 * to the parents will make the child our kill candidate of choice.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 */
Andrew Morton97c2c9b82006-04-18 22:20:38 -0700118 list_for_each_entry(child, &p->children, sibling) {
119 task_lock(child);
120 if (child->mm != mm && child->mm)
121 points += child->mm->total_vm/2 + 1;
122 task_unlock(child);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 }
124
125 /*
126 * CPU time is in tens of seconds and run time is in thousands
127 * of seconds. There is no particular reason for this other than
128 * that it turned out to work very well in practice.
129 */
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700130 thread_group_cputime(p, &task_time);
131 utime = cputime_to_jiffies(task_time.utime);
132 stime = cputime_to_jiffies(task_time.stime);
133 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
134
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135
136 if (uptime >= p->start_time.tv_sec)
137 run_time = (uptime - p->start_time.tv_sec) >> 10;
138 else
139 run_time = 0;
140
Cyrill Gorcunova12888f2009-03-31 15:19:27 -0700141 if (cpu_time)
142 points /= int_sqrt(cpu_time);
143 if (run_time)
144 points /= int_sqrt(int_sqrt(run_time));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146 /*
147 * Niced processes are most likely less important, so double
148 * their badness points.
149 */
150 if (task_nice(p) > 0)
151 points *= 2;
152
153 /*
154 * Superuser processes are usually more important, so we make it
155 * less likely that we kill those.
156 */
Eric Parisa2f29452008-11-11 22:02:54 +1100157 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
158 has_capability_noaudit(p, CAP_SYS_RESOURCE))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 points /= 4;
160
161 /*
162 * We don't want to kill a process with direct hardware access.
163 * Not only could that mess up the hardware, but usually users
164 * tend to only have this flag set on applications they think
165 * of as important.
166 */
Eric Parisa2f29452008-11-11 22:02:54 +1100167 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 points /= 4;
169
170 /*
Nick Piggin7887a3d2006-09-25 23:31:29 -0700171 * If p's nodes don't overlap ours, it may still help to kill p
172 * because p may have allocated or otherwise mapped memory on
173 * this node before. However it will be less likely.
174 */
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700175 if (!has_intersects_mems_allowed(p))
Nick Piggin7887a3d2006-09-25 23:31:29 -0700176 points /= 8;
177
178 /*
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700179 * Adjust the score by oom_adj.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 */
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700181 if (oom_adj) {
182 if (oom_adj > 0) {
Joshua N Pritikin9a827822007-05-06 14:49:07 -0700183 if (!points)
184 points = 1;
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700185 points <<= oom_adj;
Joshua N Pritikin9a827822007-05-06 14:49:07 -0700186 } else
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700187 points >>= -(oom_adj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 }
189
190#ifdef DEBUG
David Rientjesa5e58a62007-07-31 00:37:32 -0700191 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 p->pid, p->comm, points);
193#endif
194 return points;
195}
196
197/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800198 * Determine the type of allocation constraint.
199 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800200#ifdef CONFIG_NUMA
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800201static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
202 gfp_t gfp_mask, nodemask_t *nodemask)
203{
Mel Gorman54a6eb52008-04-28 02:12:16 -0700204 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -0700205 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -0700206 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800207
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800208 /*
209 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
210 * to kill current.We have to random task kill in this case.
211 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
212 */
213 if (gfp_mask & __GFP_THISNODE)
214 return CONSTRAINT_NONE;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800215
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800216 /*
217 * The nodemask here is a nodemask passed to alloc_pages(). Now,
218 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
219 * feature. mempolicy is an only user of nodemask here.
220 * check mempolicy's nodemask contains all N_HIGH_MEMORY
221 */
222 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask))
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800223 return CONSTRAINT_MEMORY_POLICY;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800224
225 /* Check this allocation failure is caused by cpuset's wall function */
226 for_each_zone_zonelist_nodemask(zone, z, zonelist,
227 high_zoneidx, nodemask)
228 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
229 return CONSTRAINT_CPUSET;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800230
231 return CONSTRAINT_NONE;
232}
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800233#else
234static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
235 gfp_t gfp_mask, nodemask_t *nodemask)
236{
237 return CONSTRAINT_NONE;
238}
239#endif
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800240
241/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 * Simple selection loop. We chose the process with the highest
243 * number of 'points'. We expect the caller will lock the tasklist.
244 *
245 * (not docbooked, we don't want this one cluttering up the manual)
246 */
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800247static struct task_struct *select_bad_process(unsigned long *ppoints,
248 struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249{
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700250 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 struct task_struct *chosen = NULL;
252 struct timespec uptime;
Kurt Garloff9827b782006-02-20 18:27:51 -0800253 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254
255 do_posix_clock_monotonic_gettime(&uptime);
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700256 for_each_process(p) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700257 unsigned long points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700258
Oleg Nesterov455c0e52010-08-09 17:18:43 -0700259 /* skip tasks that have already released their mm */
Nick Piggin5081dde2006-09-25 23:31:32 -0700260 if (!p->mm)
261 continue;
Oleg Nesterov455c0e52010-08-09 17:18:43 -0700262 /* skip the init task and kthreads */
263 if (is_global_init(p) || (p->flags & PF_KTHREAD))
Paul Jacksona49335c2005-09-06 15:18:09 -0700264 continue;
David Rientjes4c4a2212008-02-07 00:14:06 -0800265 if (mem && !task_in_mem_cgroup(p, mem))
266 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700267
Paul Jacksona49335c2005-09-06 15:18:09 -0700268 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700269 * This task already has access to memory reserves and is
270 * being killed. Don't allow any other task access to the
271 * memory reserve.
272 *
273 * Note: this may have a chance of deadlock if it gets
274 * blocked waiting for another task which itself is waiting
275 * for memory. Is there a better alternative?
276 */
277 if (test_tsk_thread_flag(p, TIF_MEMDIE))
278 return ERR_PTR(-1UL);
279
280 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700281 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700282 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700283 *
284 * However, if p is the current task, we allow the 'kill' to
285 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
286 * which will allow it to gain access to memory reserves in
287 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700288 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700289 */
Nick Pigginb78483a2006-09-29 02:01:14 -0700290 if (p->flags & PF_EXITING) {
291 if (p != current)
292 return ERR_PTR(-1UL);
293
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700294 chosen = p;
295 *ppoints = ULONG_MAX;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700296 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700297
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700298 if (p->signal->oom_adj == OOM_DISABLE)
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700299 continue;
300
Li Zefan97d87c92008-04-28 02:13:35 -0700301 points = badness(p, uptime.tv_sec);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700302 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700303 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800304 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 }
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700306 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700307
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308 return chosen;
309}
310
311/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700312 * dump_tasks - dump current memory state of all system tasks
313 * @mem: target memory controller
314 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800315 * Dumps the current memory state of all system tasks, excluding kernel threads.
316 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
317 * score, and name.
318 *
319 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
320 * shown.
321 *
322 * Call with tasklist_lock read-locked.
323 */
324static void dump_tasks(const struct mem_cgroup *mem)
325{
326 struct task_struct *g, *p;
327
328 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
329 "name\n");
330 do_each_thread(g, p) {
David Rientjes6d2661e2009-05-28 14:34:19 -0700331 struct mm_struct *mm;
332
David Rientjesfef1bdd2008-02-07 00:14:07 -0800333 if (mem && !task_in_mem_cgroup(p, mem))
334 continue;
David Rientjesb4416d22008-11-06 12:53:29 -0800335 if (!thread_group_leader(p))
336 continue;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800337
338 task_lock(p);
David Rientjes6d2661e2009-05-28 14:34:19 -0700339 mm = p->mm;
340 if (!mm) {
341 /*
342 * total_vm and rss sizes do not exist for tasks with no
343 * mm so there's no need to report them; they can't be
344 * oom killed anyway.
345 */
346 task_unlock(p);
347 continue;
348 }
David Rientjesfef1bdd2008-02-07 00:14:07 -0800349 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
David Rientjes6d2661e2009-05-28 14:34:19 -0700350 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700351 get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj,
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700352 p->comm);
David Rientjesfef1bdd2008-02-07 00:14:07 -0800353 task_unlock(p);
354 } while_each_thread(g, p);
355}
356
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800357static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
358 struct mem_cgroup *mem)
David Rientjes1b604d72009-12-14 17:57:47 -0800359{
360 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
361 "oom_adj=%d\n",
362 current->comm, gfp_mask, order, current->signal->oom_adj);
363 task_lock(current);
364 cpuset_print_task_mems_allowed(current);
365 task_unlock(current);
366 dump_stack();
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800367 mem_cgroup_print_oom_info(mem, p);
David Rientjes1b604d72009-12-14 17:57:47 -0800368 show_mem();
369 if (sysctl_oom_dump_tasks)
370 dump_tasks(mem);
371}
372
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800373#define K(x) ((x) << (PAGE_SHIFT-10))
374
Randy Dunlap1b578df2008-03-19 17:00:42 -0700375/*
Ram Gupta5a291b982006-09-25 23:31:54 -0700376 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
377 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
378 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 */
Nick Pigginf3af38d2006-12-06 20:31:51 -0800380static void __oom_kill_task(struct task_struct *p, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381{
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700382 if (is_global_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 WARN_ON(1);
384 printk(KERN_WARNING "tried to kill init!\n");
385 return;
386 }
387
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800388 task_lock(p);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700389 if (!p->mm) {
390 WARN_ON(1);
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800391 printk(KERN_WARNING "tried to kill an mm-less task %d (%s)!\n",
392 task_pid_nr(p), p->comm);
393 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 return;
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700395 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700396
Nick Pigginf3af38d2006-12-06 20:31:51 -0800397 if (verbose)
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800398 printk(KERN_ERR "Killed process %d (%s) "
399 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
400 task_pid_nr(p), p->comm,
401 K(p->mm->total_vm),
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800402 K(get_mm_counter(p->mm, MM_ANONPAGES)),
403 K(get_mm_counter(p->mm, MM_FILEPAGES)));
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800404 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405
406 /*
407 * We give our sacrificial lamb high priority and access to
408 * all the memory it needs. That way it should be able to
409 * exit() and clear out its resources quickly...
410 */
Peter Zijlstrafa717062008-01-25 21:08:27 +0100411 p->rt.time_slice = HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 set_tsk_thread_flag(p, TIF_MEMDIE);
413
414 force_sig(SIGKILL, p);
415}
416
Nick Pigginf3af38d2006-12-06 20:31:51 -0800417static int oom_kill_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418{
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700419 /* WARNING: mm may not be dereferenced since we did not obtain its
420 * value from get_task_mm(p). This is OK since all we need to do is
421 * compare mm to q->mm below.
422 *
423 * Furthermore, even if mm contains a non-NULL value, p->mm may
424 * change to NULL at any time since we do not hold task_lock(p).
425 * However, this is of no concern to us.
426 */
KOSAKI Motohiro8c5cd6f2009-09-21 17:03:15 -0700427 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
Dave Peterson01315922006-04-18 22:20:44 -0700428 return 1;
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700429
Nick Pigginf3af38d2006-12-06 20:31:51 -0800430 __oom_kill_task(p, 1);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800431
Dave Peterson01315922006-04-18 22:20:44 -0700432 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433}
434
David Rientjes7213f502007-10-16 23:25:57 -0700435static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesfef1bdd2008-02-07 00:14:07 -0800436 unsigned long points, struct mem_cgroup *mem,
437 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 struct task_struct *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
David Rientjes1b604d72009-12-14 17:57:47 -0800441 if (printk_ratelimit())
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800442 dump_header(p, gfp_mask, order, mem);
David Rientjes7213f502007-10-16 23:25:57 -0700443
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700444 /*
445 * If the task is already exiting, don't alarm the sysadmin or kill
446 * its children or threads, just set TIF_MEMDIE so it can die quickly
447 */
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700448 if (p->flags & PF_EXITING) {
Nick Pigginf3af38d2006-12-06 20:31:51 -0800449 __oom_kill_task(p, 0);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700450 return 0;
451 }
452
Nick Pigginf3af38d2006-12-06 20:31:51 -0800453 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700454 message, task_pid_nr(p), p->comm, points);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800455
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 /* Try to kill a child first */
Matthias Kaehlcke7b1915a2007-10-18 23:39:58 -0700457 list_for_each_entry(c, &p->children, sibling) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 if (c->mm == p->mm)
459 continue;
KAMEZAWA Hiroyuki5a2d4192010-02-22 12:44:14 -0800460 if (mem && !task_in_mem_cgroup(c, mem))
461 continue;
Nick Pigginf3af38d2006-12-06 20:31:51 -0800462 if (!oom_kill_task(c))
Dave Peterson01315922006-04-18 22:20:44 -0700463 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 }
Nick Pigginf3af38d2006-12-06 20:31:51 -0800465 return oom_kill_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466}
467
Balbir Singh00f0b822008-03-04 14:28:39 -0800468#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800469void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
470{
471 unsigned long points = 0;
472 struct task_struct *p;
473
KAMEZAWA Hiroyukidaaf1e62010-03-10 15:22:32 -0800474 if (sysctl_panic_on_oom == 2)
475 panic("out of memory(memcg). panic_on_oom is selected.\n");
Li Zefane115f2d2008-04-15 14:34:37 -0700476 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800477retry:
478 p = select_bad_process(&points, mem);
David Rientjesdf64f812010-05-26 14:42:41 -0700479 if (!p || PTR_ERR(p) == -1UL)
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800480 goto out;
481
David Rientjesfef1bdd2008-02-07 00:14:07 -0800482 if (oom_kill_process(p, gfp_mask, 0, points, mem,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800483 "Memory cgroup out of memory"))
484 goto retry;
485out:
Li Zefane115f2d2008-04-15 14:34:37 -0700486 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800487}
488#endif
489
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700490static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
491
492int register_oom_notifier(struct notifier_block *nb)
493{
494 return blocking_notifier_chain_register(&oom_notify_list, nb);
495}
496EXPORT_SYMBOL_GPL(register_oom_notifier);
497
498int unregister_oom_notifier(struct notifier_block *nb)
499{
500 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
501}
502EXPORT_SYMBOL_GPL(unregister_oom_notifier);
503
David Rientjes098d7f12007-10-16 23:25:55 -0700504/*
505 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
506 * if a parallel OOM killing is already taking place that includes a zone in
507 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
508 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700509int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700510{
Mel Gormandd1a2392008-04-28 02:12:17 -0700511 struct zoneref *z;
512 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700513 int ret = 1;
514
David Rientjesc7d4cae2009-01-06 14:39:00 -0800515 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700516 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
517 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700518 ret = 0;
519 goto out;
520 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700521 }
David Rientjes098d7f12007-10-16 23:25:55 -0700522
Mel Gormandd1a2392008-04-28 02:12:17 -0700523 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
524 /*
David Rientjesc7d4cae2009-01-06 14:39:00 -0800525 * Lock each zone in the zonelist under zone_scan_lock so a
Mel Gormandd1a2392008-04-28 02:12:17 -0700526 * parallel invocation of try_set_zone_oom() doesn't succeed
527 * when it shouldn't.
528 */
529 zone_set_flag(zone, ZONE_OOM_LOCKED);
530 }
531
David Rientjes098d7f12007-10-16 23:25:55 -0700532out:
David Rientjesc7d4cae2009-01-06 14:39:00 -0800533 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700534 return ret;
535}
536
537/*
538 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
539 * allocation attempts with zonelists containing them may now recall the OOM
540 * killer, if necessary.
541 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700542void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700543{
Mel Gormandd1a2392008-04-28 02:12:17 -0700544 struct zoneref *z;
545 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700546
David Rientjesc7d4cae2009-01-06 14:39:00 -0800547 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700548 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
549 zone_clear_flag(zone, ZONE_OOM_LOCKED);
550 }
David Rientjesc7d4cae2009-01-06 14:39:00 -0800551 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700552}
553
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800554/*
555 * Must be called with tasklist_lock held for read.
556 */
557static void __out_of_memory(gfp_t gfp_mask, int order)
558{
David Rientjes184101b2009-05-06 16:02:55 -0700559 struct task_struct *p;
560 unsigned long points;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800561
David Rientjes184101b2009-05-06 16:02:55 -0700562 if (sysctl_oom_kill_allocating_task)
563 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
564 "Out of memory (oom_kill_allocating_task)"))
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800565 return;
David Rientjes184101b2009-05-06 16:02:55 -0700566retry:
567 /*
568 * Rambo mode: Shoot down a process and hope it solves whatever
569 * issues we may have.
570 */
571 p = select_bad_process(&points, NULL);
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800572
David Rientjes184101b2009-05-06 16:02:55 -0700573 if (PTR_ERR(p) == -1UL)
574 return;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800575
David Rientjes184101b2009-05-06 16:02:55 -0700576 /* Found nothing?!?! Either we hang forever, or we panic. */
577 if (!p) {
578 read_unlock(&tasklist_lock);
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800579 dump_header(NULL, gfp_mask, order, NULL);
David Rientjes184101b2009-05-06 16:02:55 -0700580 panic("Out of memory and no killable processes...\n");
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800581 }
David Rientjes184101b2009-05-06 16:02:55 -0700582
583 if (oom_kill_process(p, gfp_mask, order, points, NULL,
584 "Out of memory"))
585 goto retry;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800586}
587
588/*
589 * pagefault handler calls into here because it is out of memory but
590 * doesn't know exactly how or why.
591 */
592void pagefault_out_of_memory(void)
593{
594 unsigned long freed = 0;
595
596 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
597 if (freed > 0)
598 /* Got some memory back in the last second. */
599 return;
600
601 if (sysctl_panic_on_oom)
602 panic("out of memory from page fault. panic_on_oom is selected.\n");
603
604 read_lock(&tasklist_lock);
605 __out_of_memory(0, 0); /* unknown gfp_mask and order */
606 read_unlock(&tasklist_lock);
607
608 /*
609 * Give "p" a good chance of killing itself before we
610 * retry to allocate memory.
611 */
612 if (!test_thread_flag(TIF_MEMDIE))
613 schedule_timeout_uninterruptible(1);
614}
615
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616/**
Dave Peterson6937a252006-06-23 02:03:13 -0700617 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700618 * @zonelist: zonelist pointer
619 * @gfp_mask: memory allocation flags
620 * @order: amount of memory being requested as a power of 2
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 *
622 * If we run out of memory, we have the choice between either
623 * killing a random task (bad), letting the system crash (worse)
624 * OR try to be smart about which process to kill. Note that we
625 * don't have to be perfect here, we just have to be good.
626 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800627void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
628 int order, nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629{
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700630 unsigned long freed = 0;
David Rientjes70e24bd2007-10-16 23:25:53 -0700631 enum oom_constraint constraint;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700632
633 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
634 if (freed > 0)
635 /* Got some memory back in the last second. */
636 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
David Rientjes1b604d72009-12-14 17:57:47 -0800638 if (sysctl_panic_on_oom == 2) {
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800639 dump_header(NULL, gfp_mask, order, NULL);
Yasunori Goto2b744c02007-05-06 14:49:59 -0700640 panic("out of memory. Compulsory panic_on_oom is selected.\n");
David Rientjes1b604d72009-12-14 17:57:47 -0800641 }
Yasunori Goto2b744c02007-05-06 14:49:59 -0700642
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800643 /*
644 * Check if there were limitations on the allocation (only relevant for
645 * NUMA) that may require different handling.
646 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800647 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
David Rientjes2b45ab32007-05-06 14:50:00 -0700648 read_lock(&tasklist_lock);
649
650 switch (constraint) {
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800651 case CONSTRAINT_MEMORY_POLICY:
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800652 oom_kill_process(current, gfp_mask, order, 0, NULL,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800653 "No available memory (MPOL_BIND)");
654 break;
655
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800656 case CONSTRAINT_NONE:
David Rientjes1b604d72009-12-14 17:57:47 -0800657 if (sysctl_panic_on_oom) {
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800658 dump_header(NULL, gfp_mask, order, NULL);
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700659 panic("out of memory. panic_on_oom is selected\n");
David Rientjes1b604d72009-12-14 17:57:47 -0800660 }
David Rientjesfe071d72007-10-16 23:25:56 -0700661 /* Fall-through */
662 case CONSTRAINT_CPUSET:
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800663 __out_of_memory(gfp_mask, order);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800664 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 }
666
Andrew Morton140ffce2006-03-02 02:54:28 -0800667 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
669 /*
670 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800671 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800673 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800674 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675}