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.
18 #include <linux/oom.h>
20 #include <linux/err.h>
21 #include <linux/sched.h>
22 #include <linux/swap.h>
23 #include <linux/timex.h>
24 #include <linux/jiffies.h>
25 #include <linux/cpuset.h>
26 #include <linux/module.h>
27 #include <linux/notifier.h>
28 #include <linux/memcontrol.h>
30 int sysctl_panic_on_oom
;
31 int sysctl_oom_kill_allocating_task
;
32 static DEFINE_SPINLOCK(zone_scan_mutex
);
36 * badness - calculate a numeric value for how bad this task has been
37 * @p: task struct of which task we should calculate
38 * @uptime: current uptime in seconds
40 * The formula used is relatively simple and documented inline in the
41 * function. The main rationale is that we want to select a good task
42 * to kill when we run out of memory.
44 * Good in this context means that:
45 * 1) we lose the minimum amount of work done
46 * 2) we recover a large amount of memory
47 * 3) we don't kill anything innocent of eating tons of memory
48 * 4) we want to kill the minimum amount of processes (one)
49 * 5) we try to kill the process the user expects us to kill, this
50 * algorithm has been meticulously tuned to meet the principle
51 * of least surprise ... (be careful when you change it)
54 unsigned long badness(struct task_struct
*p
, unsigned long uptime
,
55 struct mem_cgroup
*mem
)
57 unsigned long points
, cpu_time
, run_time
, s
;
59 struct task_struct
*child
;
69 * The memory size of the process is the basis for the badness.
71 points
= mm
->total_vm
;
74 * After this unlock we can no longer dereference local variable `mm'
79 * swapoff can easily use up all memory, so kill those first.
81 if (p
->flags
& PF_SWAPOFF
)
85 * Processes which fork a lot of child processes are likely
86 * a good choice. We add half the vmsize of the children if they
87 * have an own mm. This prevents forking servers to flood the
88 * machine with an endless amount of children. In case a single
89 * child is eating the vast majority of memory, adding only half
90 * to the parents will make the child our kill candidate of choice.
92 list_for_each_entry(child
, &p
->children
, sibling
) {
94 if (child
->mm
!= mm
&& child
->mm
)
95 points
+= child
->mm
->total_vm
/2 + 1;
100 * CPU time is in tens of seconds and run time is in thousands
101 * of seconds. There is no particular reason for this other than
102 * that it turned out to work very well in practice.
104 cpu_time
= (cputime_to_jiffies(p
->utime
) + cputime_to_jiffies(p
->stime
))
107 if (uptime
>= p
->start_time
.tv_sec
)
108 run_time
= (uptime
- p
->start_time
.tv_sec
) >> 10;
112 s
= int_sqrt(cpu_time
);
115 s
= int_sqrt(int_sqrt(run_time
));
120 * Niced processes are most likely less important, so double
121 * their badness points.
123 if (task_nice(p
) > 0)
127 * Superuser processes are usually more important, so we make it
128 * less likely that we kill those.
130 if (__capable(p
, CAP_SYS_ADMIN
) || __capable(p
, CAP_SYS_RESOURCE
))
134 * We don't want to kill a process with direct hardware access.
135 * Not only could that mess up the hardware, but usually users
136 * tend to only have this flag set on applications they think
139 if (__capable(p
, CAP_SYS_RAWIO
))
143 * If p's nodes don't overlap ours, it may still help to kill p
144 * because p may have allocated or otherwise mapped memory on
145 * this node before. However it will be less likely.
147 if (!cpuset_mems_allowed_intersects(current
, p
))
151 * Adjust the score by oomkilladj.
154 if (p
->oomkilladj
> 0) {
157 points
<<= p
->oomkilladj
;
159 points
>>= -(p
->oomkilladj
);
163 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %lu points\n",
164 p
->pid
, p
->comm
, points
);
170 * Determine the type of allocation constraint.
172 static inline enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
177 nodemask_t nodes
= node_states
[N_HIGH_MEMORY
];
179 for (z
= zonelist
->zones
; *z
; z
++)
180 if (cpuset_zone_allowed_softwall(*z
, gfp_mask
))
181 node_clear(zone_to_nid(*z
), nodes
);
183 return CONSTRAINT_CPUSET
;
185 if (!nodes_empty(nodes
))
186 return CONSTRAINT_MEMORY_POLICY
;
189 return CONSTRAINT_NONE
;
193 * Simple selection loop. We chose the process with the highest
194 * number of 'points'. We expect the caller will lock the tasklist.
196 * (not docbooked, we don't want this one cluttering up the manual)
198 static struct task_struct
*select_bad_process(unsigned long *ppoints
,
199 struct mem_cgroup
*mem
)
201 struct task_struct
*g
, *p
;
202 struct task_struct
*chosen
= NULL
;
203 struct timespec uptime
;
206 do_posix_clock_monotonic_gettime(&uptime
);
207 do_each_thread(g
, p
) {
208 unsigned long points
;
211 * skip kernel threads and tasks which have already released
216 /* skip the init task */
217 if (is_global_init(p
))
219 if (mem
&& !task_in_mem_cgroup(p
, mem
))
223 * This task already has access to memory reserves and is
224 * being killed. Don't allow any other task access to the
227 * Note: this may have a chance of deadlock if it gets
228 * blocked waiting for another task which itself is waiting
229 * for memory. Is there a better alternative?
231 if (test_tsk_thread_flag(p
, TIF_MEMDIE
))
232 return ERR_PTR(-1UL);
235 * This is in the process of releasing memory so wait for it
236 * to finish before killing some other task by mistake.
238 * However, if p is the current task, we allow the 'kill' to
239 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
240 * which will allow it to gain access to memory reserves in
241 * the process of exiting and releasing its resources.
242 * Otherwise we could get an easy OOM deadlock.
244 if (p
->flags
& PF_EXITING
) {
246 return ERR_PTR(-1UL);
249 *ppoints
= ULONG_MAX
;
252 if (p
->oomkilladj
== OOM_DISABLE
)
255 points
= badness(p
, uptime
.tv_sec
, mem
);
256 if (points
> *ppoints
|| !chosen
) {
260 } while_each_thread(g
, p
);
266 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
267 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
270 static void __oom_kill_task(struct task_struct
*p
, int verbose
)
272 if (is_global_init(p
)) {
274 printk(KERN_WARNING
"tried to kill init!\n");
280 printk(KERN_WARNING
"tried to kill an mm-less task!\n");
285 printk(KERN_ERR
"Killed process %d (%s)\n",
286 task_pid_nr(p
), p
->comm
);
289 * We give our sacrificial lamb high priority and access to
290 * all the memory it needs. That way it should be able to
291 * exit() and clear out its resources quickly...
293 p
->rt
.time_slice
= HZ
;
294 set_tsk_thread_flag(p
, TIF_MEMDIE
);
296 force_sig(SIGKILL
, p
);
299 static int oom_kill_task(struct task_struct
*p
)
301 struct mm_struct
*mm
;
302 struct task_struct
*g
, *q
;
306 /* WARNING: mm may not be dereferenced since we did not obtain its
307 * value from get_task_mm(p). This is OK since all we need to do is
308 * compare mm to q->mm below.
310 * Furthermore, even if mm contains a non-NULL value, p->mm may
311 * change to NULL at any time since we do not hold task_lock(p).
312 * However, this is of no concern to us.
319 * Don't kill the process if any threads are set to OOM_DISABLE
321 do_each_thread(g
, q
) {
322 if (q
->mm
== mm
&& q
->oomkilladj
== OOM_DISABLE
)
324 } while_each_thread(g
, q
);
326 __oom_kill_task(p
, 1);
329 * kill all processes that share the ->mm (i.e. all threads),
330 * but are in a different thread group. Don't let them have access
331 * to memory reserves though, otherwise we might deplete all memory.
333 do_each_thread(g
, q
) {
334 if (q
->mm
== mm
&& !same_thread_group(q
, p
))
335 force_sig(SIGKILL
, q
);
336 } while_each_thread(g
, q
);
341 static int oom_kill_process(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
342 unsigned long points
, const char *message
)
344 struct task_struct
*c
;
346 if (printk_ratelimit()) {
347 printk(KERN_WARNING
"%s invoked oom-killer: "
348 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
349 current
->comm
, gfp_mask
, order
, current
->oomkilladj
);
355 * If the task is already exiting, don't alarm the sysadmin or kill
356 * its children or threads, just set TIF_MEMDIE so it can die quickly
358 if (p
->flags
& PF_EXITING
) {
359 __oom_kill_task(p
, 0);
363 printk(KERN_ERR
"%s: kill process %d (%s) score %li or a child\n",
364 message
, task_pid_nr(p
), p
->comm
, points
);
366 /* Try to kill a child first */
367 list_for_each_entry(c
, &p
->children
, sibling
) {
370 if (!oom_kill_task(c
))
373 return oom_kill_task(p
);
376 #ifdef CONFIG_CGROUP_MEM_CONT
377 void mem_cgroup_out_of_memory(struct mem_cgroup
*mem
, gfp_t gfp_mask
)
379 unsigned long points
= 0;
380 struct task_struct
*p
;
385 p
= select_bad_process(&points
, mem
);
386 if (PTR_ERR(p
) == -1UL)
392 if (oom_kill_process(p
, gfp_mask
, 0, points
,
393 "Memory cgroup out of memory"))
401 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
403 int register_oom_notifier(struct notifier_block
*nb
)
405 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
407 EXPORT_SYMBOL_GPL(register_oom_notifier
);
409 int unregister_oom_notifier(struct notifier_block
*nb
)
411 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
413 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
416 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
417 * if a parallel OOM killing is already taking place that includes a zone in
418 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
420 int try_set_zone_oom(struct zonelist
*zonelist
)
427 spin_lock(&zone_scan_mutex
);
429 if (zone_is_oom_locked(*z
)) {
433 } while (*(++z
) != NULL
);
436 * Lock each zone in the zonelist under zone_scan_mutex so a parallel
437 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
441 zone_set_flag(*z
, ZONE_OOM_LOCKED
);
442 } while (*(++z
) != NULL
);
444 spin_unlock(&zone_scan_mutex
);
449 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
450 * allocation attempts with zonelists containing them may now recall the OOM
451 * killer, if necessary.
453 void clear_zonelist_oom(struct zonelist
*zonelist
)
459 spin_lock(&zone_scan_mutex
);
461 zone_clear_flag(*z
, ZONE_OOM_LOCKED
);
462 } while (*(++z
) != NULL
);
463 spin_unlock(&zone_scan_mutex
);
467 * out_of_memory - kill the "best" process when we run out of memory
469 * If we run out of memory, we have the choice between either
470 * killing a random task (bad), letting the system crash (worse)
471 * OR try to be smart about which process to kill. Note that we
472 * don't have to be perfect here, we just have to be good.
474 void out_of_memory(struct zonelist
*zonelist
, gfp_t gfp_mask
, int order
)
476 struct task_struct
*p
;
477 unsigned long points
= 0;
478 unsigned long freed
= 0;
479 enum oom_constraint constraint
;
481 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
483 /* Got some memory back in the last second. */
486 if (sysctl_panic_on_oom
== 2)
487 panic("out of memory. Compulsory panic_on_oom is selected.\n");
490 * Check if there were limitations on the allocation (only relevant for
491 * NUMA) that may require different handling.
493 constraint
= constrained_alloc(zonelist
, gfp_mask
);
494 read_lock(&tasklist_lock
);
496 switch (constraint
) {
497 case CONSTRAINT_MEMORY_POLICY
:
498 oom_kill_process(current
, gfp_mask
, order
, points
,
499 "No available memory (MPOL_BIND)");
502 case CONSTRAINT_NONE
:
503 if (sysctl_panic_on_oom
)
504 panic("out of memory. panic_on_oom is selected\n");
506 case CONSTRAINT_CPUSET
:
507 if (sysctl_oom_kill_allocating_task
) {
508 oom_kill_process(current
, gfp_mask
, order
, points
,
509 "Out of memory (oom_kill_allocating_task)");
514 * Rambo mode: Shoot down a process and hope it solves whatever
515 * issues we may have.
517 p
= select_bad_process(&points
, NULL
);
519 if (PTR_ERR(p
) == -1UL)
522 /* Found nothing?!?! Either we hang forever, or we panic. */
524 read_unlock(&tasklist_lock
);
525 panic("Out of memory and no killable processes...\n");
528 if (oom_kill_process(p
, gfp_mask
, order
, points
,
536 read_unlock(&tasklist_lock
);
539 * Give "p" a good chance of killing itself before we
540 * retry to allocate memory unless "p" is current
542 if (!test_thread_flag(TIF_MEMDIE
))
543 schedule_timeout_uninterruptible(1);