blob: a5493a3b4851282b879590f653ed9aae1b36b08a [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
18#include <linux/mm.h>
19#include <linux/sched.h>
20#include <linux/swap.h>
21#include <linux/timex.h>
22#include <linux/jiffies.h>
Paul Jacksonef08e3b2005-09-06 15:18:13 -070023#include <linux/cpuset.h>
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -070024#include <linux/module.h>
25#include <linux/notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070027int sysctl_panic_on_oom;
Linus Torvalds1da177e2005-04-16 15:20:36 -070028/* #define DEBUG */
29
30/**
Dave Peterson6937a252006-06-23 02:03:13 -070031 * badness - calculate a numeric value for how bad this task has been
Linus Torvalds1da177e2005-04-16 15:20:36 -070032 * @p: task struct of which task we should calculate
Paul Jacksona49335c2005-09-06 15:18:09 -070033 * @uptime: current uptime in seconds
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 *
35 * The formula used is relatively simple and documented inline in the
36 * function. The main rationale is that we want to select a good task
37 * to kill when we run out of memory.
38 *
39 * Good in this context means that:
40 * 1) we lose the minimum amount of work done
41 * 2) we recover a large amount of memory
42 * 3) we don't kill anything innocent of eating tons of memory
43 * 4) we want to kill the minimum amount of processes (one)
44 * 5) we try to kill the process the user expects us to kill, this
45 * algorithm has been meticulously tuned to meet the principle
46 * of least surprise ... (be careful when you change it)
47 */
48
49unsigned long badness(struct task_struct *p, unsigned long uptime)
50{
51 unsigned long points, cpu_time, run_time, s;
Andrew Morton97c2c9b82006-04-18 22:20:38 -070052 struct mm_struct *mm;
53 struct task_struct *child;
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
Andrew Morton97c2c9b82006-04-18 22:20:38 -070055 task_lock(p);
56 mm = p->mm;
57 if (!mm) {
58 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 return 0;
Andrew Morton97c2c9b82006-04-18 22:20:38 -070060 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62 /*
Nick Pigginaf5b9122006-09-25 23:31:31 -070063 * swapoff can easily use up all memory, so kill those first.
64 */
65 if (p->flags & PF_SWAPOFF)
66 return ULONG_MAX;
67
68 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 * The memory size of the process is the basis for the badness.
70 */
Andrew Morton97c2c9b82006-04-18 22:20:38 -070071 points = mm->total_vm;
72
73 /*
74 * After this unlock we can no longer dereference local variable `mm'
75 */
76 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
78 /*
79 * Processes which fork a lot of child processes are likely
Kurt Garloff9827b782006-02-20 18:27:51 -080080 * a good choice. We add half the vmsize of the children if they
Linus Torvalds1da177e2005-04-16 15:20:36 -070081 * have an own mm. This prevents forking servers to flood the
Kurt Garloff9827b782006-02-20 18:27:51 -080082 * machine with an endless amount of children. In case a single
83 * child is eating the vast majority of memory, adding only half
84 * to the parents will make the child our kill candidate of choice.
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 */
Andrew Morton97c2c9b82006-04-18 22:20:38 -070086 list_for_each_entry(child, &p->children, sibling) {
87 task_lock(child);
88 if (child->mm != mm && child->mm)
89 points += child->mm->total_vm/2 + 1;
90 task_unlock(child);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 }
92
93 /*
94 * CPU time is in tens of seconds and run time is in thousands
95 * of seconds. There is no particular reason for this other than
96 * that it turned out to work very well in practice.
97 */
98 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
99 >> (SHIFT_HZ + 3);
100
101 if (uptime >= p->start_time.tv_sec)
102 run_time = (uptime - p->start_time.tv_sec) >> 10;
103 else
104 run_time = 0;
105
106 s = int_sqrt(cpu_time);
107 if (s)
108 points /= s;
109 s = int_sqrt(int_sqrt(run_time));
110 if (s)
111 points /= s;
112
113 /*
114 * Niced processes are most likely less important, so double
115 * their badness points.
116 */
117 if (task_nice(p) > 0)
118 points *= 2;
119
120 /*
121 * Superuser processes are usually more important, so we make it
122 * less likely that we kill those.
123 */
124 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
125 p->uid == 0 || p->euid == 0)
126 points /= 4;
127
128 /*
129 * We don't want to kill a process with direct hardware access.
130 * Not only could that mess up the hardware, but usually users
131 * tend to only have this flag set on applications they think
132 * of as important.
133 */
134 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
135 points /= 4;
136
137 /*
Nick Piggin7887a3d2006-09-25 23:31:29 -0700138 * If p's nodes don't overlap ours, it may still help to kill p
139 * because p may have allocated or otherwise mapped memory on
140 * this node before. However it will be less likely.
141 */
142 if (!cpuset_excl_nodes_overlap(p))
143 points /= 8;
144
145 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 * Adjust the score by oomkilladj.
147 */
148 if (p->oomkilladj) {
149 if (p->oomkilladj > 0)
150 points <<= p->oomkilladj;
151 else
152 points >>= -(p->oomkilladj);
153 }
154
155#ifdef DEBUG
156 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
157 p->pid, p->comm, points);
158#endif
159 return points;
160}
161
162/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800163 * Types of limitations to the nodes from which allocations may occur
164 */
165#define CONSTRAINT_NONE 1
166#define CONSTRAINT_MEMORY_POLICY 2
167#define CONSTRAINT_CPUSET 3
168
169/*
170 * Determine the type of allocation constraint.
171 */
172static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
173{
174#ifdef CONFIG_NUMA
175 struct zone **z;
176 nodemask_t nodes = node_online_map;
177
178 for (z = zonelist->zones; *z; z++)
179 if (cpuset_zone_allowed(*z, gfp_mask))
Christoph Lameter89fa3022006-09-25 23:31:55 -0700180 node_clear(zone_to_nid(*z), nodes);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800181 else
182 return CONSTRAINT_CPUSET;
183
184 if (!nodes_empty(nodes))
185 return CONSTRAINT_MEMORY_POLICY;
186#endif
187
188 return CONSTRAINT_NONE;
189}
190
191/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 * Simple selection loop. We chose the process with the highest
193 * number of 'points'. We expect the caller will lock the tasklist.
194 *
195 * (not docbooked, we don't want this one cluttering up the manual)
196 */
Kurt Garloff9827b782006-02-20 18:27:51 -0800197static struct task_struct *select_bad_process(unsigned long *ppoints)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 struct task_struct *g, *p;
200 struct task_struct *chosen = NULL;
201 struct timespec uptime;
Kurt Garloff9827b782006-02-20 18:27:51 -0800202 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
204 do_posix_clock_monotonic_gettime(&uptime);
Paul Jacksona49335c2005-09-06 15:18:09 -0700205 do_each_thread(g, p) {
206 unsigned long points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700207
Oleg Nesterov28324d12006-09-29 02:01:12 -0700208 /*
209 * skip kernel threads and tasks which have already released
210 * their mm.
211 */
Nick Piggin5081dde2006-09-25 23:31:32 -0700212 if (!p->mm)
213 continue;
Oleg Nesterov28324d12006-09-29 02:01:12 -0700214 /* skip the init task */
215 if (is_init(p))
Paul Jacksona49335c2005-09-06 15:18:09 -0700216 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700217
Paul Jacksona49335c2005-09-06 15:18:09 -0700218 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700219 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700220 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700221 *
222 * However, if p is the current task, we allow the 'kill' to
223 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
224 * which will allow it to gain access to memory reserves in
225 * the process of exiting and releasing its resources.
226 * Otherwise we could get an OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700227 */
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700228 if ((p->flags & PF_EXITING) && p == current) {
229 chosen = p;
230 *ppoints = ULONG_MAX;
231 break;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700232 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700233 if ((p->flags & PF_EXITING) ||
234 test_tsk_thread_flag(p, TIF_MEMDIE))
235 return ERR_PTR(-1UL);
236
Nick Piggin4a3ede12006-09-25 23:31:30 -0700237 if (p->oomkilladj == OOM_DISABLE)
238 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Paul Jacksona49335c2005-09-06 15:18:09 -0700240 points = badness(p, uptime.tv_sec);
Kurt Garloff9827b782006-02-20 18:27:51 -0800241 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700242 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800243 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 }
Paul Jacksona49335c2005-09-06 15:18:09 -0700245 } while_each_thread(g, p);
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700246
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 return chosen;
248}
249
250/**
Ram Gupta5a291b982006-09-25 23:31:54 -0700251 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
252 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
253 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700255static void __oom_kill_task(struct task_struct *p, const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256{
Sukadev Bhattiproluf400e192006-09-29 02:00:07 -0700257 if (is_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 WARN_ON(1);
259 printk(KERN_WARNING "tried to kill init!\n");
260 return;
261 }
262
Oleg Nesterov01017a22006-09-29 02:01:13 -0700263 if (!p->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 WARN_ON(1);
265 printk(KERN_WARNING "tried to kill an mm-less task!\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 return;
267 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700268
269 if (message) {
270 printk(KERN_ERR "%s: Killed process %d (%s).\n",
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800271 message, p->pid, p->comm);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700272 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274 /*
275 * We give our sacrificial lamb high priority and access to
276 * all the memory it needs. That way it should be able to
277 * exit() and clear out its resources quickly...
278 */
279 p->time_slice = HZ;
280 set_tsk_thread_flag(p, TIF_MEMDIE);
281
282 force_sig(SIGKILL, p);
283}
284
Ingo Molnar36c8b582006-07-03 00:25:41 -0700285static int oom_kill_task(struct task_struct *p, const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286{
Dave Peterson01315922006-04-18 22:20:44 -0700287 struct mm_struct *mm;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700288 struct task_struct *g, *q;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
Dave Peterson01315922006-04-18 22:20:44 -0700290 mm = p->mm;
291
292 /* WARNING: mm may not be dereferenced since we did not obtain its
293 * value from get_task_mm(p). This is OK since all we need to do is
294 * compare mm to q->mm below.
295 *
296 * Furthermore, even if mm contains a non-NULL value, p->mm may
297 * change to NULL at any time since we do not hold task_lock(p).
298 * However, this is of no concern to us.
299 */
300
Oleg Nesterov01017a22006-09-29 02:01:13 -0700301 if (mm == NULL)
Dave Peterson01315922006-04-18 22:20:44 -0700302 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800304 __oom_kill_task(p, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 /*
306 * kill all processes that share the ->mm (i.e. all threads),
307 * but are in a different thread group
308 */
309 do_each_thread(g, q)
310 if (q->mm == mm && q->tgid != p->tgid)
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800311 __oom_kill_task(q, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 while_each_thread(g, q);
313
Dave Peterson01315922006-04-18 22:20:44 -0700314 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315}
316
Dave Peterson01315922006-04-18 22:20:44 -0700317static int oom_kill_process(struct task_struct *p, unsigned long points,
318 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 struct task_struct *c;
321 struct list_head *tsk;
322
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700323 /*
324 * If the task is already exiting, don't alarm the sysadmin or kill
325 * its children or threads, just set TIF_MEMDIE so it can die quickly
326 */
327 if (p->flags & PF_EXITING) {
328 __oom_kill_task(p, NULL);
329 return 0;
330 }
331
332 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li"
333 " and children.\n", p->pid, p->comm, points);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 /* Try to kill a child first */
335 list_for_each(tsk, &p->children) {
336 c = list_entry(tsk, struct task_struct, sibling);
337 if (c->mm == p->mm)
338 continue;
Dave Peterson01315922006-04-18 22:20:44 -0700339 if (!oom_kill_task(c, message))
340 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 }
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800342 return oom_kill_task(p, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343}
344
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700345static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
346
347int register_oom_notifier(struct notifier_block *nb)
348{
349 return blocking_notifier_chain_register(&oom_notify_list, nb);
350}
351EXPORT_SYMBOL_GPL(register_oom_notifier);
352
353int unregister_oom_notifier(struct notifier_block *nb)
354{
355 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
356}
357EXPORT_SYMBOL_GPL(unregister_oom_notifier);
358
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359/**
Dave Peterson6937a252006-06-23 02:03:13 -0700360 * out_of_memory - kill the "best" process when we run out of memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 *
362 * If we run out of memory, we have the choice between either
363 * killing a random task (bad), letting the system crash (worse)
364 * OR try to be smart about which process to kill. Note that we
365 * don't have to be perfect here, we just have to be good.
366 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800367void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
Ingo Molnar36c8b582006-07-03 00:25:41 -0700369 struct task_struct *p;
Andrew Mortond6713e02006-02-28 16:59:19 -0800370 unsigned long points = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700371 unsigned long freed = 0;
372
373 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
374 if (freed > 0)
375 /* Got some memory back in the last second. */
376 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
Anton Blanchard42639262005-07-07 17:56:06 -0700378 if (printk_ratelimit()) {
Nick Pigginb72f1602006-09-25 23:31:32 -0700379 printk(KERN_WARNING "%s invoked oom-killer: "
380 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
381 current->comm, gfp_mask, order, current->oomkilladj);
Andrew Mortonb958f7d2006-02-01 03:05:51 -0800382 dump_stack();
Anton Blanchard42639262005-07-07 17:56:06 -0700383 show_mem();
384 }
Janet Morgan578c2fd2005-06-21 17:14:56 -0700385
Paul Jackson505970b2006-01-14 13:21:06 -0800386 cpuset_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 read_lock(&tasklist_lock);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800388
389 /*
390 * Check if there were limitations on the allocation (only relevant for
391 * NUMA) that may require different handling.
392 */
393 switch (constrained_alloc(zonelist, gfp_mask)) {
394 case CONSTRAINT_MEMORY_POLICY:
Dave Peterson01315922006-04-18 22:20:44 -0700395 oom_kill_process(current, points,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800396 "No available memory (MPOL_BIND)");
397 break;
398
399 case CONSTRAINT_CPUSET:
Dave Peterson01315922006-04-18 22:20:44 -0700400 oom_kill_process(current, points,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800401 "No available memory in cpuset");
402 break;
403
404 case CONSTRAINT_NONE:
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700405 if (sysctl_panic_on_oom)
406 panic("out of memory. panic_on_oom is selected\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407retry:
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800408 /*
409 * Rambo mode: Shoot down a process and hope it solves whatever
410 * issues we may have.
411 */
412 p = select_bad_process(&points);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800414 if (PTR_ERR(p) == -1UL)
415 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800417 /* Found nothing?!?! Either we hang forever, or we panic. */
418 if (!p) {
419 read_unlock(&tasklist_lock);
420 cpuset_unlock();
421 panic("Out of memory and no killable processes...\n");
422 }
423
Dave Peterson01315922006-04-18 22:20:44 -0700424 if (oom_kill_process(p, points, "Out of memory"))
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800425 goto retry;
426
427 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 }
429
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800430out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800431 read_unlock(&tasklist_lock);
Paul Jackson505970b2006-01-14 13:21:06 -0800432 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
434 /*
435 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800436 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800438 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800439 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440}