blob: b68e802a7a7d6961a0c1f7709a9cc606d022371c [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 Morton97c2c9b2006-04-18 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 Morton97c2c9b2006-04-18 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 Morton97c2c9b2006-04-18 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 Morton97c2c9b2006-04-18 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 Morton97c2c9b2006-04-18 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 Nesterov28324d12006-09-29 02:01:12 -0700259 /*
260 * skip kernel threads and tasks which have already released
261 * their mm.
262 */
Nick Piggin5081dde2006-09-25 23:31:32 -0700263 if (!p->mm)
264 continue;
Oleg Nesterov28324d12006-09-29 02:01:12 -0700265 /* skip the init task */
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700266 if (is_global_init(p))
Paul Jacksona49335c2005-09-06 15:18:09 -0700267 continue;
David Rientjes4c4a2212008-02-07 00:14:06 -0800268 if (mem && !task_in_mem_cgroup(p, mem))
269 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700270
Paul Jacksona49335c2005-09-06 15:18:09 -0700271 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700272 * This task already has access to memory reserves and is
273 * being killed. Don't allow any other task access to the
274 * memory reserve.
275 *
276 * Note: this may have a chance of deadlock if it gets
277 * blocked waiting for another task which itself is waiting
278 * for memory. Is there a better alternative?
279 */
280 if (test_tsk_thread_flag(p, TIF_MEMDIE))
281 return ERR_PTR(-1UL);
282
283 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700284 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700285 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700286 *
287 * However, if p is the current task, we allow the 'kill' to
288 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
289 * which will allow it to gain access to memory reserves in
290 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700291 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700292 */
Nick Pigginb78483a2006-09-29 02:01:14 -0700293 if (p->flags & PF_EXITING) {
294 if (p != current)
295 return ERR_PTR(-1UL);
296
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700297 chosen = p;
298 *ppoints = ULONG_MAX;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700299 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700300
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700301 if (p->signal->oom_adj == OOM_DISABLE)
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700302 continue;
303
Li Zefan97d87c92008-04-28 02:13:35 -0700304 points = badness(p, uptime.tv_sec);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700305 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700306 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800307 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308 }
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700309 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700310
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 return chosen;
312}
313
314/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700315 * dump_tasks - dump current memory state of all system tasks
316 * @mem: target memory controller
317 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800318 * Dumps the current memory state of all system tasks, excluding kernel threads.
319 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
320 * score, and name.
321 *
322 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
323 * shown.
324 *
325 * Call with tasklist_lock read-locked.
326 */
327static void dump_tasks(const struct mem_cgroup *mem)
328{
329 struct task_struct *g, *p;
330
331 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
332 "name\n");
333 do_each_thread(g, p) {
David Rientjes6d2661e2009-05-28 14:34:19 -0700334 struct mm_struct *mm;
335
David Rientjesfef1bdd2008-02-07 00:14:07 -0800336 if (mem && !task_in_mem_cgroup(p, mem))
337 continue;
David Rientjesb4416d22008-11-06 12:53:29 -0800338 if (!thread_group_leader(p))
339 continue;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800340
341 task_lock(p);
David Rientjes6d2661e2009-05-28 14:34:19 -0700342 mm = p->mm;
343 if (!mm) {
344 /*
345 * total_vm and rss sizes do not exist for tasks with no
346 * mm so there's no need to report them; they can't be
347 * oom killed anyway.
348 */
349 task_unlock(p);
350 continue;
351 }
David Rientjesfef1bdd2008-02-07 00:14:07 -0800352 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
David Rientjes6d2661e2009-05-28 14:34:19 -0700353 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700354 get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj,
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700355 p->comm);
David Rientjesfef1bdd2008-02-07 00:14:07 -0800356 task_unlock(p);
357 } while_each_thread(g, p);
358}
359
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800360static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
361 struct mem_cgroup *mem)
David Rientjes1b604d72009-12-14 17:57:47 -0800362{
363 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
364 "oom_adj=%d\n",
365 current->comm, gfp_mask, order, current->signal->oom_adj);
366 task_lock(current);
367 cpuset_print_task_mems_allowed(current);
368 task_unlock(current);
369 dump_stack();
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800370 mem_cgroup_print_oom_info(mem, p);
David Rientjes1b604d72009-12-14 17:57:47 -0800371 show_mem();
372 if (sysctl_oom_dump_tasks)
373 dump_tasks(mem);
374}
375
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800376#define K(x) ((x) << (PAGE_SHIFT-10))
377
Randy Dunlap1b578df2008-03-19 17:00:42 -0700378/*
Ram Gupta5a291b982006-09-25 23:31:54 -0700379 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
380 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
381 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 */
Nick Pigginf3af38d2006-12-06 20:31:51 -0800383static void __oom_kill_task(struct task_struct *p, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384{
Serge E. Hallynb460cbc2007-10-18 23:39:52 -0700385 if (is_global_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 WARN_ON(1);
387 printk(KERN_WARNING "tried to kill init!\n");
388 return;
389 }
390
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800391 task_lock(p);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700392 if (!p->mm) {
393 WARN_ON(1);
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800394 printk(KERN_WARNING "tried to kill an mm-less task %d (%s)!\n",
395 task_pid_nr(p), p->comm);
396 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 return;
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700398 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700399
Nick Pigginf3af38d2006-12-06 20:31:51 -0800400 if (verbose)
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800401 printk(KERN_ERR "Killed process %d (%s) "
402 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
403 task_pid_nr(p), p->comm,
404 K(p->mm->total_vm),
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800405 K(get_mm_counter(p->mm, MM_ANONPAGES)),
406 K(get_mm_counter(p->mm, MM_FILEPAGES)));
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800407 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409 /*
410 * We give our sacrificial lamb high priority and access to
411 * all the memory it needs. That way it should be able to
412 * exit() and clear out its resources quickly...
413 */
Peter Zijlstrafa717062008-01-25 21:08:27 +0100414 p->rt.time_slice = HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 set_tsk_thread_flag(p, TIF_MEMDIE);
416
417 force_sig(SIGKILL, p);
418}
419
Nick Pigginf3af38d2006-12-06 20:31:51 -0800420static int oom_kill_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421{
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700422 /* WARNING: mm may not be dereferenced since we did not obtain its
423 * value from get_task_mm(p). This is OK since all we need to do is
424 * compare mm to q->mm below.
425 *
426 * Furthermore, even if mm contains a non-NULL value, p->mm may
427 * change to NULL at any time since we do not hold task_lock(p).
428 * However, this is of no concern to us.
429 */
KOSAKI Motohiro8c5cd6f2009-09-21 17:03:15 -0700430 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
Dave Peterson01315922006-04-18 22:20:44 -0700431 return 1;
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700432
Nick Pigginf3af38d2006-12-06 20:31:51 -0800433 __oom_kill_task(p, 1);
Nick Pigginc33e0fc2006-12-06 20:31:50 -0800434
Dave Peterson01315922006-04-18 22:20:44 -0700435 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436}
437
David Rientjes7213f502007-10-16 23:25:57 -0700438static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesfef1bdd2008-02-07 00:14:07 -0800439 unsigned long points, struct mem_cgroup *mem,
440 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 struct task_struct *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
David Rientjes1b604d72009-12-14 17:57:47 -0800444 if (printk_ratelimit())
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800445 dump_header(p, gfp_mask, order, mem);
David Rientjes7213f502007-10-16 23:25:57 -0700446
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700447 /*
448 * If the task is already exiting, don't alarm the sysadmin or kill
449 * its children or threads, just set TIF_MEMDIE so it can die quickly
450 */
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700451 if (p->flags & PF_EXITING) {
Nick Pigginf3af38d2006-12-06 20:31:51 -0800452 __oom_kill_task(p, 0);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700453 return 0;
454 }
455
Nick Pigginf3af38d2006-12-06 20:31:51 -0800456 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700457 message, task_pid_nr(p), p->comm, points);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 /* Try to kill a child first */
Matthias Kaehlcke7b1915a2007-10-18 23:39:58 -0700460 list_for_each_entry(c, &p->children, sibling) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 if (c->mm == p->mm)
462 continue;
KAMEZAWA Hiroyuki5a2d4192010-02-22 12:44:14 -0800463 if (mem && !task_in_mem_cgroup(c, mem))
464 continue;
Nick Pigginf3af38d2006-12-06 20:31:51 -0800465 if (!oom_kill_task(c))
Dave Peterson01315922006-04-18 22:20:44 -0700466 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 }
Nick Pigginf3af38d2006-12-06 20:31:51 -0800468 return oom_kill_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469}
470
Balbir Singh00f0b822008-03-04 14:28:39 -0800471#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800472void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
473{
474 unsigned long points = 0;
475 struct task_struct *p;
476
KAMEZAWA Hiroyukidaaf1e62010-03-10 15:22:32 -0800477 if (sysctl_panic_on_oom == 2)
478 panic("out of memory(memcg). panic_on_oom is selected.\n");
Li Zefane115f2d2008-04-15 14:34:37 -0700479 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800480retry:
481 p = select_bad_process(&points, mem);
482 if (PTR_ERR(p) == -1UL)
483 goto out;
484
485 if (!p)
486 p = current;
487
David Rientjesfef1bdd2008-02-07 00:14:07 -0800488 if (oom_kill_process(p, gfp_mask, 0, points, mem,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800489 "Memory cgroup out of memory"))
490 goto retry;
491out:
Li Zefane115f2d2008-04-15 14:34:37 -0700492 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800493}
494#endif
495
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700496static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
497
498int register_oom_notifier(struct notifier_block *nb)
499{
500 return blocking_notifier_chain_register(&oom_notify_list, nb);
501}
502EXPORT_SYMBOL_GPL(register_oom_notifier);
503
504int unregister_oom_notifier(struct notifier_block *nb)
505{
506 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
507}
508EXPORT_SYMBOL_GPL(unregister_oom_notifier);
509
David Rientjes098d7f12007-10-16 23:25:55 -0700510/*
511 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
512 * if a parallel OOM killing is already taking place that includes a zone in
513 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
514 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700515int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700516{
Mel Gormandd1a2392008-04-28 02:12:17 -0700517 struct zoneref *z;
518 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700519 int ret = 1;
520
David Rientjesc7d4cae2009-01-06 14:39:00 -0800521 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700522 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
523 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700524 ret = 0;
525 goto out;
526 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700527 }
David Rientjes098d7f12007-10-16 23:25:55 -0700528
Mel Gormandd1a2392008-04-28 02:12:17 -0700529 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
530 /*
David Rientjesc7d4cae2009-01-06 14:39:00 -0800531 * Lock each zone in the zonelist under zone_scan_lock so a
Mel Gormandd1a2392008-04-28 02:12:17 -0700532 * parallel invocation of try_set_zone_oom() doesn't succeed
533 * when it shouldn't.
534 */
535 zone_set_flag(zone, ZONE_OOM_LOCKED);
536 }
537
David Rientjes098d7f12007-10-16 23:25:55 -0700538out:
David Rientjesc7d4cae2009-01-06 14:39:00 -0800539 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700540 return ret;
541}
542
543/*
544 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
545 * allocation attempts with zonelists containing them may now recall the OOM
546 * killer, if necessary.
547 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700548void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700549{
Mel Gormandd1a2392008-04-28 02:12:17 -0700550 struct zoneref *z;
551 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700552
David Rientjesc7d4cae2009-01-06 14:39:00 -0800553 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700554 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
555 zone_clear_flag(zone, ZONE_OOM_LOCKED);
556 }
David Rientjesc7d4cae2009-01-06 14:39:00 -0800557 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700558}
559
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800560/*
561 * Must be called with tasklist_lock held for read.
562 */
563static void __out_of_memory(gfp_t gfp_mask, int order)
564{
David Rientjes184101b2009-05-06 16:02:55 -0700565 struct task_struct *p;
566 unsigned long points;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800567
David Rientjes184101b2009-05-06 16:02:55 -0700568 if (sysctl_oom_kill_allocating_task)
569 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
570 "Out of memory (oom_kill_allocating_task)"))
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800571 return;
David Rientjes184101b2009-05-06 16:02:55 -0700572retry:
573 /*
574 * Rambo mode: Shoot down a process and hope it solves whatever
575 * issues we may have.
576 */
577 p = select_bad_process(&points, NULL);
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800578
David Rientjes184101b2009-05-06 16:02:55 -0700579 if (PTR_ERR(p) == -1UL)
580 return;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800581
David Rientjes184101b2009-05-06 16:02:55 -0700582 /* Found nothing?!?! Either we hang forever, or we panic. */
583 if (!p) {
584 read_unlock(&tasklist_lock);
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800585 dump_header(NULL, gfp_mask, order, NULL);
David Rientjes184101b2009-05-06 16:02:55 -0700586 panic("Out of memory and no killable processes...\n");
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800587 }
David Rientjes184101b2009-05-06 16:02:55 -0700588
589 if (oom_kill_process(p, gfp_mask, order, points, NULL,
590 "Out of memory"))
591 goto retry;
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800592}
593
594/*
595 * pagefault handler calls into here because it is out of memory but
596 * doesn't know exactly how or why.
597 */
598void pagefault_out_of_memory(void)
599{
600 unsigned long freed = 0;
601
602 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
603 if (freed > 0)
604 /* Got some memory back in the last second. */
605 return;
606
607 if (sysctl_panic_on_oom)
608 panic("out of memory from page fault. panic_on_oom is selected.\n");
609
610 read_lock(&tasklist_lock);
611 __out_of_memory(0, 0); /* unknown gfp_mask and order */
612 read_unlock(&tasklist_lock);
613
614 /*
615 * Give "p" a good chance of killing itself before we
616 * retry to allocate memory.
617 */
618 if (!test_thread_flag(TIF_MEMDIE))
619 schedule_timeout_uninterruptible(1);
620}
621
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622/**
Dave Peterson6937a252006-06-23 02:03:13 -0700623 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700624 * @zonelist: zonelist pointer
625 * @gfp_mask: memory allocation flags
626 * @order: amount of memory being requested as a power of 2
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 *
628 * If we run out of memory, we have the choice between either
629 * killing a random task (bad), letting the system crash (worse)
630 * OR try to be smart about which process to kill. Note that we
631 * don't have to be perfect here, we just have to be good.
632 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800633void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
634 int order, nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635{
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700636 unsigned long freed = 0;
David Rientjes70e24bd2007-10-16 23:25:53 -0700637 enum oom_constraint constraint;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700638
639 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
640 if (freed > 0)
641 /* Got some memory back in the last second. */
642 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
David Rientjes1b604d72009-12-14 17:57:47 -0800644 if (sysctl_panic_on_oom == 2) {
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800645 dump_header(NULL, gfp_mask, order, NULL);
Yasunori Goto2b744c02007-05-06 14:49:59 -0700646 panic("out of memory. Compulsory panic_on_oom is selected.\n");
David Rientjes1b604d72009-12-14 17:57:47 -0800647 }
Yasunori Goto2b744c02007-05-06 14:49:59 -0700648
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800649 /*
650 * Check if there were limitations on the allocation (only relevant for
651 * NUMA) that may require different handling.
652 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800653 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
David Rientjes2b45ab32007-05-06 14:50:00 -0700654 read_lock(&tasklist_lock);
655
656 switch (constraint) {
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800657 case CONSTRAINT_MEMORY_POLICY:
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800658 oom_kill_process(current, gfp_mask, order, 0, NULL,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800659 "No available memory (MPOL_BIND)");
660 break;
661
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800662 case CONSTRAINT_NONE:
David Rientjes1b604d72009-12-14 17:57:47 -0800663 if (sysctl_panic_on_oom) {
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800664 dump_header(NULL, gfp_mask, order, NULL);
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700665 panic("out of memory. panic_on_oom is selected\n");
David Rientjes1b604d72009-12-14 17:57:47 -0800666 }
David Rientjesfe071d72007-10-16 23:25:56 -0700667 /* Fall-through */
668 case CONSTRAINT_CPUSET:
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800669 __out_of_memory(gfp_mask, order);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800670 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 }
672
Andrew Morton140ffce2006-03-02 02:54:28 -0800673 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
675 /*
676 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800677 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800679 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800680 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}