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...
8 * The routines in this file are used to kill a process when
9 * 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.
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.
19 #include <linux/sched.h>
20 #include <linux/swap.h>
21 #include <linux/timex.h>
22 #include <linux/jiffies.h>
23 #include <linux/cpuset.h>
28 * oom_badness - calculate a numeric value for how bad this task has been
29 * @p: task struct of which task we should calculate
30 * @uptime: current uptime in seconds
32 * The formula used is relatively simple and documented inline in the
33 * function. The main rationale is that we want to select a good task
34 * to kill when we run out of memory.
36 * Good in this context means that:
37 * 1) we lose the minimum amount of work done
38 * 2) we recover a large amount of memory
39 * 3) we don't kill anything innocent of eating tons of memory
40 * 4) we want to kill the minimum amount of processes (one)
41 * 5) we try to kill the process the user expects us to kill, this
42 * algorithm has been meticulously tuned to meet the principle
43 * of least surprise ... (be careful when you change it)
46 unsigned long badness(struct task_struct
*p
, unsigned long uptime
)
48 unsigned long points
, cpu_time
, run_time
, s
;
49 struct list_head
*tsk
;
55 * The memory size of the process is the basis for the badness.
57 points
= p
->mm
->total_vm
;
60 * Processes which fork a lot of child processes are likely
61 * a good choice. We add half the vmsize of the children if they
62 * have an own mm. This prevents forking servers to flood the
63 * machine with an endless amount of children. In case a single
64 * child is eating the vast majority of memory, adding only half
65 * to the parents will make the child our kill candidate of choice.
67 list_for_each(tsk
, &p
->children
) {
68 struct task_struct
*chld
;
69 chld
= list_entry(tsk
, struct task_struct
, sibling
);
70 if (chld
->mm
!= p
->mm
&& chld
->mm
)
71 points
+= chld
->mm
->total_vm
/2 + 1;
75 * CPU time is in tens of seconds and run time is in thousands
76 * of seconds. There is no particular reason for this other than
77 * that it turned out to work very well in practice.
79 cpu_time
= (cputime_to_jiffies(p
->utime
) + cputime_to_jiffies(p
->stime
))
82 if (uptime
>= p
->start_time
.tv_sec
)
83 run_time
= (uptime
- p
->start_time
.tv_sec
) >> 10;
87 s
= int_sqrt(cpu_time
);
90 s
= int_sqrt(int_sqrt(run_time
));
95 * Niced processes are most likely less important, so double
96 * their badness points.
102 * Superuser processes are usually more important, so we make it
103 * less likely that we kill those.
105 if (cap_t(p
->cap_effective
) & CAP_TO_MASK(CAP_SYS_ADMIN
) ||
106 p
->uid
== 0 || p
->euid
== 0)
110 * We don't want to kill a process with direct hardware access.
111 * Not only could that mess up the hardware, but usually users
112 * tend to only have this flag set on applications they think
115 if (cap_t(p
->cap_effective
) & CAP_TO_MASK(CAP_SYS_RAWIO
))
119 * Adjust the score by oomkilladj.
122 if (p
->oomkilladj
> 0)
123 points
<<= p
->oomkilladj
;
125 points
>>= -(p
->oomkilladj
);
129 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %d points\n",
130 p
->pid
, p
->comm
, points
);
136 * Simple selection loop. We chose the process with the highest
137 * number of 'points'. We expect the caller will lock the tasklist.
139 * (not docbooked, we don't want this one cluttering up the manual)
141 static struct task_struct
*select_bad_process(unsigned long *ppoints
)
143 struct task_struct
*g
, *p
;
144 struct task_struct
*chosen
= NULL
;
145 struct timespec uptime
;
148 do_posix_clock_monotonic_gettime(&uptime
);
149 do_each_thread(g
, p
) {
150 unsigned long points
;
153 /* skip the init task with pid == 1 */
156 if (p
->oomkilladj
== OOM_DISABLE
)
158 /* If p's nodes don't overlap ours, it won't help to kill p. */
159 if (!cpuset_excl_nodes_overlap(p
))
163 * This is in the process of releasing memory so for wait it
164 * to finish before killing some other task by mistake.
166 releasing
= test_tsk_thread_flag(p
, TIF_MEMDIE
) ||
167 p
->flags
& PF_EXITING
;
168 if (releasing
&& !(p
->flags
& PF_DEAD
))
169 return ERR_PTR(-1UL);
170 if (p
->flags
& PF_SWAPOFF
)
173 points
= badness(p
, uptime
.tv_sec
);
174 if (points
> *ppoints
|| !chosen
) {
178 } while_each_thread(g
, p
);
183 * We must be careful though to never send SIGKILL a process with
184 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
185 * we select a process with CAP_SYS_RAW_IO set).
187 static void __oom_kill_task(task_t
*p
)
191 printk(KERN_WARNING
"tried to kill init!\n");
196 if (!p
->mm
|| p
->mm
== &init_mm
) {
198 printk(KERN_WARNING
"tried to kill an mm-less task!\n");
203 printk(KERN_ERR
"Out of Memory: Killed process %d (%s).\n",
207 * We give our sacrificial lamb high priority and access to
208 * all the memory it needs. That way it should be able to
209 * exit() and clear out its resources quickly...
212 set_tsk_thread_flag(p
, TIF_MEMDIE
);
214 force_sig(SIGKILL
, p
);
217 static struct mm_struct
*oom_kill_task(task_t
*p
)
219 struct mm_struct
*mm
= get_task_mm(p
);
224 if (mm
== &init_mm
) {
231 * kill all processes that share the ->mm (i.e. all threads),
232 * but are in a different thread group
235 if (q
->mm
== mm
&& q
->tgid
!= p
->tgid
)
237 while_each_thread(g
, q
);
242 static struct mm_struct
*oom_kill_process(struct task_struct
*p
,
243 unsigned long points
)
245 struct mm_struct
*mm
;
246 struct task_struct
*c
;
247 struct list_head
*tsk
;
249 printk(KERN_ERR
"Out of Memory: Kill process %d (%s) score %li and "
250 "children.\n", p
->pid
, p
->comm
, points
);
251 /* Try to kill a child first */
252 list_for_each(tsk
, &p
->children
) {
253 c
= list_entry(tsk
, struct task_struct
, sibling
);
256 mm
= oom_kill_task(c
);
260 return oom_kill_task(p
);
264 * oom_kill - kill the "best" process when we run out of memory
266 * If we run out of memory, we have the choice between either
267 * killing a random task (bad), letting the system crash (worse)
268 * OR try to be smart about which process to kill. Note that we
269 * don't have to be perfect here, we just have to be good.
271 void out_of_memory(gfp_t gfp_mask
, int order
)
273 struct mm_struct
*mm
= NULL
;
275 unsigned long points
;
277 if (printk_ratelimit()) {
278 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
285 read_lock(&tasklist_lock
);
287 p
= select_bad_process(&points
);
289 if (PTR_ERR(p
) == -1UL)
292 /* Found nothing?!?! Either we hang forever, or we panic. */
294 read_unlock(&tasklist_lock
);
296 panic("Out of memory and no killable processes...\n");
299 mm
= oom_kill_process(p
, points
);
304 read_unlock(&tasklist_lock
);
310 * Give "p" a good chance of killing itself before we
311 * retry to allocate memory unless "p" is current
313 if (!test_thread_flag(TIF_MEMDIE
))
314 schedule_timeout_interruptible(1);