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 int sysctl_oom_dump_tasks
;
33 static DEFINE_SPINLOCK(zone_scan_mutex
);
37 * badness - calculate a numeric value for how bad this task has been
38 * @p: task struct of which task we should calculate
39 * @uptime: current uptime in seconds
40 * @mem: target memory controller
42 * The formula used is relatively simple and documented inline in the
43 * function. The main rationale is that we want to select a good task
44 * to kill when we run out of memory.
46 * Good in this context means that:
47 * 1) we lose the minimum amount of work done
48 * 2) we recover a large amount of memory
49 * 3) we don't kill anything innocent of eating tons of memory
50 * 4) we want to kill the minimum amount of processes (one)
51 * 5) we try to kill the process the user expects us to kill, this
52 * algorithm has been meticulously tuned to meet the principle
53 * of least surprise ... (be careful when you change it)
56 unsigned long badness(struct task_struct
*p
, unsigned long uptime
,
57 struct mem_cgroup
*mem
)
59 unsigned long points
, cpu_time
, run_time
, s
;
61 struct task_struct
*child
;
71 * The memory size of the process is the basis for the badness.
73 points
= mm
->total_vm
;
76 * After this unlock we can no longer dereference local variable `mm'
81 * swapoff can easily use up all memory, so kill those first.
83 if (p
->flags
& PF_SWAPOFF
)
87 * Processes which fork a lot of child processes are likely
88 * a good choice. We add half the vmsize of the children if they
89 * have an own mm. This prevents forking servers to flood the
90 * machine with an endless amount of children. In case a single
91 * child is eating the vast majority of memory, adding only half
92 * to the parents will make the child our kill candidate of choice.
94 list_for_each_entry(child
, &p
->children
, sibling
) {
96 if (child
->mm
!= mm
&& child
->mm
)
97 points
+= child
->mm
->total_vm
/2 + 1;
102 * CPU time is in tens of seconds and run time is in thousands
103 * of seconds. There is no particular reason for this other than
104 * that it turned out to work very well in practice.
106 cpu_time
= (cputime_to_jiffies(p
->utime
) + cputime_to_jiffies(p
->stime
))
109 if (uptime
>= p
->start_time
.tv_sec
)
110 run_time
= (uptime
- p
->start_time
.tv_sec
) >> 10;
114 s
= int_sqrt(cpu_time
);
117 s
= int_sqrt(int_sqrt(run_time
));
122 * Niced processes are most likely less important, so double
123 * their badness points.
125 if (task_nice(p
) > 0)
129 * Superuser processes are usually more important, so we make it
130 * less likely that we kill those.
132 if (__capable(p
, CAP_SYS_ADMIN
) || __capable(p
, CAP_SYS_RESOURCE
))
136 * We don't want to kill a process with direct hardware access.
137 * Not only could that mess up the hardware, but usually users
138 * tend to only have this flag set on applications they think
141 if (__capable(p
, CAP_SYS_RAWIO
))
145 * If p's nodes don't overlap ours, it may still help to kill p
146 * because p may have allocated or otherwise mapped memory on
147 * this node before. However it will be less likely.
149 if (!cpuset_mems_allowed_intersects(current
, p
))
153 * Adjust the score by oomkilladj.
156 if (p
->oomkilladj
> 0) {
159 points
<<= p
->oomkilladj
;
161 points
>>= -(p
->oomkilladj
);
165 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %lu points\n",
166 p
->pid
, p
->comm
, points
);
172 * Determine the type of allocation constraint.
174 static inline enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
179 nodemask_t nodes
= node_states
[N_HIGH_MEMORY
];
181 for (z
= zonelist
->zones
; *z
; z
++)
182 if (cpuset_zone_allowed_softwall(*z
, gfp_mask
))
183 node_clear(zone_to_nid(*z
), nodes
);
185 return CONSTRAINT_CPUSET
;
187 if (!nodes_empty(nodes
))
188 return CONSTRAINT_MEMORY_POLICY
;
191 return CONSTRAINT_NONE
;
195 * Simple selection loop. We chose the process with the highest
196 * number of 'points'. We expect the caller will lock the tasklist.
198 * (not docbooked, we don't want this one cluttering up the manual)
200 static struct task_struct
*select_bad_process(unsigned long *ppoints
,
201 struct mem_cgroup
*mem
)
203 struct task_struct
*g
, *p
;
204 struct task_struct
*chosen
= NULL
;
205 struct timespec uptime
;
208 do_posix_clock_monotonic_gettime(&uptime
);
209 do_each_thread(g
, p
) {
210 unsigned long points
;
213 * skip kernel threads and tasks which have already released
218 /* skip the init task */
219 if (is_global_init(p
))
221 if (mem
&& !task_in_mem_cgroup(p
, mem
))
225 * This task already has access to memory reserves and is
226 * being killed. Don't allow any other task access to the
229 * Note: this may have a chance of deadlock if it gets
230 * blocked waiting for another task which itself is waiting
231 * for memory. Is there a better alternative?
233 if (test_tsk_thread_flag(p
, TIF_MEMDIE
))
234 return ERR_PTR(-1UL);
237 * This is in the process of releasing memory so wait for it
238 * to finish before killing some other task by mistake.
240 * However, if p is the current task, we allow the 'kill' to
241 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
242 * which will allow it to gain access to memory reserves in
243 * the process of exiting and releasing its resources.
244 * Otherwise we could get an easy OOM deadlock.
246 if (p
->flags
& PF_EXITING
) {
248 return ERR_PTR(-1UL);
251 *ppoints
= ULONG_MAX
;
254 if (p
->oomkilladj
== OOM_DISABLE
)
257 points
= badness(p
, uptime
.tv_sec
, mem
);
258 if (points
> *ppoints
|| !chosen
) {
262 } while_each_thread(g
, p
);
268 * dump_tasks - dump current memory state of all system tasks
269 * @mem: target memory controller
271 * Dumps the current memory state of all system tasks, excluding kernel threads.
272 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
275 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
278 * Call with tasklist_lock read-locked.
280 static void dump_tasks(const struct mem_cgroup
*mem
)
282 struct task_struct
*g
, *p
;
284 printk(KERN_INFO
"[ pid ] uid tgid total_vm rss cpu oom_adj "
286 do_each_thread(g
, p
) {
288 * total_vm and rss sizes do not exist for tasks with a
289 * detached mm so there's no need to report them.
293 if (mem
&& !task_in_mem_cgroup(p
, mem
))
297 printk(KERN_INFO
"[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
298 p
->pid
, p
->uid
, p
->tgid
, p
->mm
->total_vm
,
299 get_mm_rss(p
->mm
), (int)task_cpu(p
), p
->oomkilladj
,
302 } while_each_thread(g
, p
);
306 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
307 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
310 static void __oom_kill_task(struct task_struct
*p
, int verbose
)
312 if (is_global_init(p
)) {
314 printk(KERN_WARNING
"tried to kill init!\n");
320 printk(KERN_WARNING
"tried to kill an mm-less task!\n");
325 printk(KERN_ERR
"Killed process %d (%s)\n",
326 task_pid_nr(p
), p
->comm
);
329 * We give our sacrificial lamb high priority and access to
330 * all the memory it needs. That way it should be able to
331 * exit() and clear out its resources quickly...
333 p
->rt
.time_slice
= HZ
;
334 set_tsk_thread_flag(p
, TIF_MEMDIE
);
336 force_sig(SIGKILL
, p
);
339 static int oom_kill_task(struct task_struct
*p
)
341 struct mm_struct
*mm
;
342 struct task_struct
*g
, *q
;
346 /* WARNING: mm may not be dereferenced since we did not obtain its
347 * value from get_task_mm(p). This is OK since all we need to do is
348 * compare mm to q->mm below.
350 * Furthermore, even if mm contains a non-NULL value, p->mm may
351 * change to NULL at any time since we do not hold task_lock(p).
352 * However, this is of no concern to us.
359 * Don't kill the process if any threads are set to OOM_DISABLE
361 do_each_thread(g
, q
) {
362 if (q
->mm
== mm
&& q
->oomkilladj
== OOM_DISABLE
)
364 } while_each_thread(g
, q
);
366 __oom_kill_task(p
, 1);
369 * kill all processes that share the ->mm (i.e. all threads),
370 * but are in a different thread group. Don't let them have access
371 * to memory reserves though, otherwise we might deplete all memory.
373 do_each_thread(g
, q
) {
374 if (q
->mm
== mm
&& !same_thread_group(q
, p
))
375 force_sig(SIGKILL
, q
);
376 } while_each_thread(g
, q
);
381 static int oom_kill_process(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
382 unsigned long points
, struct mem_cgroup
*mem
,
385 struct task_struct
*c
;
387 if (printk_ratelimit()) {
388 printk(KERN_WARNING
"%s invoked oom-killer: "
389 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
390 current
->comm
, gfp_mask
, order
, current
->oomkilladj
);
393 if (sysctl_oom_dump_tasks
)
398 * If the task is already exiting, don't alarm the sysadmin or kill
399 * its children or threads, just set TIF_MEMDIE so it can die quickly
401 if (p
->flags
& PF_EXITING
) {
402 __oom_kill_task(p
, 0);
406 printk(KERN_ERR
"%s: kill process %d (%s) score %li or a child\n",
407 message
, task_pid_nr(p
), p
->comm
, points
);
409 /* Try to kill a child first */
410 list_for_each_entry(c
, &p
->children
, sibling
) {
413 if (!oom_kill_task(c
))
416 return oom_kill_task(p
);
419 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
420 void mem_cgroup_out_of_memory(struct mem_cgroup
*mem
, gfp_t gfp_mask
)
422 unsigned long points
= 0;
423 struct task_struct
*p
;
426 read_lock(&tasklist_lock
);
428 p
= select_bad_process(&points
, mem
);
429 if (PTR_ERR(p
) == -1UL)
435 if (oom_kill_process(p
, gfp_mask
, 0, points
, mem
,
436 "Memory cgroup out of memory"))
439 read_unlock(&tasklist_lock
);
444 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
446 int register_oom_notifier(struct notifier_block
*nb
)
448 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
450 EXPORT_SYMBOL_GPL(register_oom_notifier
);
452 int unregister_oom_notifier(struct notifier_block
*nb
)
454 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
456 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
459 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
460 * if a parallel OOM killing is already taking place that includes a zone in
461 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
463 int try_set_zone_oom(struct zonelist
*zonelist
)
470 spin_lock(&zone_scan_mutex
);
472 if (zone_is_oom_locked(*z
)) {
476 } while (*(++z
) != NULL
);
479 * Lock each zone in the zonelist under zone_scan_mutex so a parallel
480 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
484 zone_set_flag(*z
, ZONE_OOM_LOCKED
);
485 } while (*(++z
) != NULL
);
487 spin_unlock(&zone_scan_mutex
);
492 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
493 * allocation attempts with zonelists containing them may now recall the OOM
494 * killer, if necessary.
496 void clear_zonelist_oom(struct zonelist
*zonelist
)
502 spin_lock(&zone_scan_mutex
);
504 zone_clear_flag(*z
, ZONE_OOM_LOCKED
);
505 } while (*(++z
) != NULL
);
506 spin_unlock(&zone_scan_mutex
);
510 * out_of_memory - kill the "best" process when we run out of memory
511 * @zonelist: zonelist pointer
512 * @gfp_mask: memory allocation flags
513 * @order: amount of memory being requested as a power of 2
515 * If we run out of memory, we have the choice between either
516 * killing a random task (bad), letting the system crash (worse)
517 * OR try to be smart about which process to kill. Note that we
518 * don't have to be perfect here, we just have to be good.
520 void out_of_memory(struct zonelist
*zonelist
, gfp_t gfp_mask
, int order
)
522 struct task_struct
*p
;
523 unsigned long points
= 0;
524 unsigned long freed
= 0;
525 enum oom_constraint constraint
;
527 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
529 /* Got some memory back in the last second. */
532 if (sysctl_panic_on_oom
== 2)
533 panic("out of memory. Compulsory panic_on_oom is selected.\n");
536 * Check if there were limitations on the allocation (only relevant for
537 * NUMA) that may require different handling.
539 constraint
= constrained_alloc(zonelist
, gfp_mask
);
540 read_lock(&tasklist_lock
);
542 switch (constraint
) {
543 case CONSTRAINT_MEMORY_POLICY
:
544 oom_kill_process(current
, gfp_mask
, order
, points
, NULL
,
545 "No available memory (MPOL_BIND)");
548 case CONSTRAINT_NONE
:
549 if (sysctl_panic_on_oom
)
550 panic("out of memory. panic_on_oom is selected\n");
552 case CONSTRAINT_CPUSET
:
553 if (sysctl_oom_kill_allocating_task
) {
554 oom_kill_process(current
, gfp_mask
, order
, points
, NULL
,
555 "Out of memory (oom_kill_allocating_task)");
560 * Rambo mode: Shoot down a process and hope it solves whatever
561 * issues we may have.
563 p
= select_bad_process(&points
, NULL
);
565 if (PTR_ERR(p
) == -1UL)
568 /* Found nothing?!?! Either we hang forever, or we panic. */
570 read_unlock(&tasklist_lock
);
571 panic("Out of memory and no killable processes...\n");
574 if (oom_kill_process(p
, gfp_mask
, order
, points
, NULL
,
582 read_unlock(&tasklist_lock
);
585 * Give "p" a good chance of killing itself before we
586 * retry to allocate memory unless "p" is current
588 if (!test_thread_flag(TIF_MEMDIE
))
589 schedule_timeout_uninterruptible(1);