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/gfp.h>
22 #include <linux/sched.h>
23 #include <linux/swap.h>
24 #include <linux/timex.h>
25 #include <linux/jiffies.h>
26 #include <linux/cpuset.h>
27 #include <linux/module.h>
28 #include <linux/notifier.h>
29 #include <linux/memcontrol.h>
30 #include <linux/security.h>
32 int sysctl_panic_on_oom
;
33 int sysctl_oom_kill_allocating_task
;
34 int sysctl_oom_dump_tasks
;
35 static DEFINE_SPINLOCK(zone_scan_lock
);
39 * Is all threads of the target process nodes overlap ours?
41 static int has_intersects_mems_allowed(struct task_struct
*tsk
)
43 struct task_struct
*t
;
47 if (cpuset_mems_allowed_intersects(current
, t
))
55 static struct task_struct
*find_lock_task_mm(struct task_struct
*p
)
57 struct task_struct
*t
= p
;
64 } while_each_thread(p
, t
);
70 * badness - calculate a numeric value for how bad this task has been
71 * @p: task struct of which task we should calculate
72 * @uptime: current uptime in seconds
74 * The formula used is relatively simple and documented inline in the
75 * function. The main rationale is that we want to select a good task
76 * to kill when we run out of memory.
78 * Good in this context means that:
79 * 1) we lose the minimum amount of work done
80 * 2) we recover a large amount of memory
81 * 3) we don't kill anything innocent of eating tons of memory
82 * 4) we want to kill the minimum amount of processes (one)
83 * 5) we try to kill the process the user expects us to kill, this
84 * algorithm has been meticulously tuned to meet the principle
85 * of least surprise ... (be careful when you change it)
88 unsigned long badness(struct task_struct
*p
, unsigned long uptime
)
90 unsigned long points
, cpu_time
, run_time
;
91 struct task_struct
*child
;
92 struct task_struct
*c
, *t
;
93 int oom_adj
= p
->signal
->oom_adj
;
94 struct task_cputime task_time
;
98 if (oom_adj
== OOM_DISABLE
)
101 p
= find_lock_task_mm(p
);
106 * The memory size of the process is the basis for the badness.
108 points
= p
->mm
->total_vm
;
111 * After this unlock we can no longer dereference local variable `mm'
116 * swapoff can easily use up all memory, so kill those first.
118 if (p
->flags
& PF_OOM_ORIGIN
)
122 * Processes which fork a lot of child processes are likely
123 * a good choice. We add half the vmsize of the children if they
124 * have an own mm. This prevents forking servers to flood the
125 * machine with an endless amount of children. In case a single
126 * child is eating the vast majority of memory, adding only half
127 * to the parents will make the child our kill candidate of choice.
131 list_for_each_entry(c
, &t
->children
, sibling
) {
132 child
= find_lock_task_mm(c
);
134 if (child
->mm
!= p
->mm
)
135 points
+= child
->mm
->total_vm
/2 + 1;
139 } while_each_thread(p
, t
);
142 * CPU time is in tens of seconds and run time is in thousands
143 * of seconds. There is no particular reason for this other than
144 * that it turned out to work very well in practice.
146 thread_group_cputime(p
, &task_time
);
147 utime
= cputime_to_jiffies(task_time
.utime
);
148 stime
= cputime_to_jiffies(task_time
.stime
);
149 cpu_time
= (utime
+ stime
) >> (SHIFT_HZ
+ 3);
152 if (uptime
>= p
->start_time
.tv_sec
)
153 run_time
= (uptime
- p
->start_time
.tv_sec
) >> 10;
158 points
/= int_sqrt(cpu_time
);
160 points
/= int_sqrt(int_sqrt(run_time
));
163 * Niced processes are most likely less important, so double
164 * their badness points.
166 if (task_nice(p
) > 0)
170 * Superuser processes are usually more important, so we make it
171 * less likely that we kill those.
173 if (has_capability_noaudit(p
, CAP_SYS_ADMIN
) ||
174 has_capability_noaudit(p
, CAP_SYS_RESOURCE
))
178 * We don't want to kill a process with direct hardware access.
179 * Not only could that mess up the hardware, but usually users
180 * tend to only have this flag set on applications they think
183 if (has_capability_noaudit(p
, CAP_SYS_RAWIO
))
187 * If p's nodes don't overlap ours, it may still help to kill p
188 * because p may have allocated or otherwise mapped memory on
189 * this node before. However it will be less likely.
191 if (!has_intersects_mems_allowed(p
))
195 * Adjust the score by oom_adj.
203 points
>>= -(oom_adj
);
207 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %lu points\n",
208 p
->pid
, p
->comm
, points
);
214 * Determine the type of allocation constraint.
217 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
218 gfp_t gfp_mask
, nodemask_t
*nodemask
)
222 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
225 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
226 * to kill current.We have to random task kill in this case.
227 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
229 if (gfp_mask
& __GFP_THISNODE
)
230 return CONSTRAINT_NONE
;
233 * The nodemask here is a nodemask passed to alloc_pages(). Now,
234 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
235 * feature. mempolicy is an only user of nodemask here.
236 * check mempolicy's nodemask contains all N_HIGH_MEMORY
238 if (nodemask
&& !nodes_subset(node_states
[N_HIGH_MEMORY
], *nodemask
))
239 return CONSTRAINT_MEMORY_POLICY
;
241 /* Check this allocation failure is caused by cpuset's wall function */
242 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
243 high_zoneidx
, nodemask
)
244 if (!cpuset_zone_allowed_softwall(zone
, gfp_mask
))
245 return CONSTRAINT_CPUSET
;
247 return CONSTRAINT_NONE
;
250 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
251 gfp_t gfp_mask
, nodemask_t
*nodemask
)
253 return CONSTRAINT_NONE
;
258 * Simple selection loop. We chose the process with the highest
259 * number of 'points'. We expect the caller will lock the tasklist.
261 * (not docbooked, we don't want this one cluttering up the manual)
263 static struct task_struct
*select_bad_process(unsigned long *ppoints
,
264 struct mem_cgroup
*mem
)
266 struct task_struct
*p
;
267 struct task_struct
*chosen
= NULL
;
268 struct timespec uptime
;
271 do_posix_clock_monotonic_gettime(&uptime
);
272 for_each_process(p
) {
273 unsigned long points
;
275 /* skip the init task and kthreads */
276 if (is_global_init(p
) || (p
->flags
& PF_KTHREAD
))
278 if (mem
&& !task_in_mem_cgroup(p
, mem
))
282 * This task already has access to memory reserves and is
283 * being killed. Don't allow any other task access to the
286 * Note: this may have a chance of deadlock if it gets
287 * blocked waiting for another task which itself is waiting
288 * for memory. Is there a better alternative?
290 if (test_tsk_thread_flag(p
, TIF_MEMDIE
))
291 return ERR_PTR(-1UL);
294 * This is in the process of releasing memory so wait for it
295 * to finish before killing some other task by mistake.
297 * However, if p is the current task, we allow the 'kill' to
298 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
299 * which will allow it to gain access to memory reserves in
300 * the process of exiting and releasing its resources.
301 * Otherwise we could get an easy OOM deadlock.
303 if ((p
->flags
& PF_EXITING
) && p
->mm
) {
305 return ERR_PTR(-1UL);
308 *ppoints
= ULONG_MAX
;
311 if (p
->signal
->oom_adj
== OOM_DISABLE
)
314 points
= badness(p
, uptime
.tv_sec
);
315 if (points
> *ppoints
|| !chosen
) {
325 * dump_tasks - dump current memory state of all system tasks
326 * @mem: target memory controller
328 * Dumps the current memory state of all system tasks, excluding kernel threads.
329 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
332 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
335 * Call with tasklist_lock read-locked.
337 static void dump_tasks(const struct mem_cgroup
*mem
)
339 struct task_struct
*g
, *p
;
341 printk(KERN_INFO
"[ pid ] uid tgid total_vm rss cpu oom_adj "
343 do_each_thread(g
, p
) {
344 struct mm_struct
*mm
;
346 if (mem
&& !task_in_mem_cgroup(p
, mem
))
348 if (!thread_group_leader(p
))
355 * total_vm and rss sizes do not exist for tasks with no
356 * mm so there's no need to report them; they can't be
362 printk(KERN_INFO
"[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
363 p
->pid
, __task_cred(p
)->uid
, p
->tgid
, mm
->total_vm
,
364 get_mm_rss(mm
), (int)task_cpu(p
), p
->signal
->oom_adj
,
367 } while_each_thread(g
, p
);
370 static void dump_header(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
371 struct mem_cgroup
*mem
)
373 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
375 current
->comm
, gfp_mask
, order
, current
->signal
->oom_adj
);
377 cpuset_print_task_mems_allowed(current
);
378 task_unlock(current
);
380 mem_cgroup_print_oom_info(mem
, p
);
382 if (sysctl_oom_dump_tasks
)
386 #define K(x) ((x) << (PAGE_SHIFT-10))
389 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
390 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
393 static void __oom_kill_task(struct task_struct
*p
, int verbose
)
395 if (is_global_init(p
)) {
397 printk(KERN_WARNING
"tried to kill init!\n");
401 p
= find_lock_task_mm(p
);
406 printk(KERN_ERR
"Killed process %d (%s) "
407 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
408 task_pid_nr(p
), p
->comm
,
410 K(get_mm_counter(p
->mm
, MM_ANONPAGES
)),
411 K(get_mm_counter(p
->mm
, MM_FILEPAGES
)));
415 * We give our sacrificial lamb high priority and access to
416 * all the memory it needs. That way it should be able to
417 * exit() and clear out its resources quickly...
419 p
->rt
.time_slice
= HZ
;
420 set_tsk_thread_flag(p
, TIF_MEMDIE
);
422 force_sig(SIGKILL
, p
);
425 static int oom_kill_task(struct task_struct
*p
)
427 /* WARNING: mm may not be dereferenced since we did not obtain its
428 * value from get_task_mm(p). This is OK since all we need to do is
429 * compare mm to q->mm below.
431 * Furthermore, even if mm contains a non-NULL value, p->mm may
432 * change to NULL at any time since we do not hold task_lock(p).
433 * However, this is of no concern to us.
435 if (!p
->mm
|| p
->signal
->oom_adj
== OOM_DISABLE
)
438 __oom_kill_task(p
, 1);
443 static int oom_kill_process(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
444 unsigned long points
, struct mem_cgroup
*mem
,
447 struct task_struct
*c
;
448 struct task_struct
*t
= p
;
450 if (printk_ratelimit())
451 dump_header(p
, gfp_mask
, order
, mem
);
454 * If the task is already exiting, don't alarm the sysadmin or kill
455 * its children or threads, just set TIF_MEMDIE so it can die quickly
457 if (p
->flags
& PF_EXITING
) {
458 __oom_kill_task(p
, 0);
462 printk(KERN_ERR
"%s: kill process %d (%s) score %li or a child\n",
463 message
, task_pid_nr(p
), p
->comm
, points
);
465 /* Try to kill a child first */
467 list_for_each_entry(c
, &t
->children
, sibling
) {
470 if (mem
&& !task_in_mem_cgroup(c
, mem
))
472 if (!oom_kill_task(c
))
475 } while_each_thread(p
, t
);
477 return oom_kill_task(p
);
480 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
481 void mem_cgroup_out_of_memory(struct mem_cgroup
*mem
, gfp_t gfp_mask
)
483 unsigned long points
= 0;
484 struct task_struct
*p
;
486 if (sysctl_panic_on_oom
== 2)
487 panic("out of memory(memcg). panic_on_oom is selected.\n");
488 read_lock(&tasklist_lock
);
490 p
= select_bad_process(&points
, mem
);
491 if (!p
|| PTR_ERR(p
) == -1UL)
494 if (oom_kill_process(p
, gfp_mask
, 0, points
, mem
,
495 "Memory cgroup out of memory"))
498 read_unlock(&tasklist_lock
);
502 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
504 int register_oom_notifier(struct notifier_block
*nb
)
506 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
508 EXPORT_SYMBOL_GPL(register_oom_notifier
);
510 int unregister_oom_notifier(struct notifier_block
*nb
)
512 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
514 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
517 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
518 * if a parallel OOM killing is already taking place that includes a zone in
519 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
521 int try_set_zone_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
527 spin_lock(&zone_scan_lock
);
528 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
529 if (zone_is_oom_locked(zone
)) {
535 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
537 * Lock each zone in the zonelist under zone_scan_lock so a
538 * parallel invocation of try_set_zone_oom() doesn't succeed
541 zone_set_flag(zone
, ZONE_OOM_LOCKED
);
545 spin_unlock(&zone_scan_lock
);
550 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
551 * allocation attempts with zonelists containing them may now recall the OOM
552 * killer, if necessary.
554 void clear_zonelist_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
559 spin_lock(&zone_scan_lock
);
560 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
561 zone_clear_flag(zone
, ZONE_OOM_LOCKED
);
563 spin_unlock(&zone_scan_lock
);
567 * Must be called with tasklist_lock held for read.
569 static void __out_of_memory(gfp_t gfp_mask
, int order
)
571 struct task_struct
*p
;
572 unsigned long points
;
574 if (sysctl_oom_kill_allocating_task
)
575 if (!oom_kill_process(current
, gfp_mask
, order
, 0, NULL
,
576 "Out of memory (oom_kill_allocating_task)"))
580 * Rambo mode: Shoot down a process and hope it solves whatever
581 * issues we may have.
583 p
= select_bad_process(&points
, NULL
);
585 if (PTR_ERR(p
) == -1UL)
588 /* Found nothing?!?! Either we hang forever, or we panic. */
590 read_unlock(&tasklist_lock
);
591 dump_header(NULL
, gfp_mask
, order
, NULL
);
592 panic("Out of memory and no killable processes...\n");
595 if (oom_kill_process(p
, gfp_mask
, order
, points
, NULL
,
601 * pagefault handler calls into here because it is out of memory but
602 * doesn't know exactly how or why.
604 void pagefault_out_of_memory(void)
606 unsigned long freed
= 0;
608 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
610 /* Got some memory back in the last second. */
613 if (sysctl_panic_on_oom
)
614 panic("out of memory from page fault. panic_on_oom is selected.\n");
616 read_lock(&tasklist_lock
);
617 __out_of_memory(0, 0); /* unknown gfp_mask and order */
618 read_unlock(&tasklist_lock
);
621 * Give "p" a good chance of killing itself before we
622 * retry to allocate memory.
624 if (!test_thread_flag(TIF_MEMDIE
))
625 schedule_timeout_uninterruptible(1);
629 * out_of_memory - kill the "best" process when we run out of memory
630 * @zonelist: zonelist pointer
631 * @gfp_mask: memory allocation flags
632 * @order: amount of memory being requested as a power of 2
634 * If we run out of memory, we have the choice between either
635 * killing a random task (bad), letting the system crash (worse)
636 * OR try to be smart about which process to kill. Note that we
637 * don't have to be perfect here, we just have to be good.
639 void out_of_memory(struct zonelist
*zonelist
, gfp_t gfp_mask
,
640 int order
, nodemask_t
*nodemask
)
642 unsigned long freed
= 0;
643 enum oom_constraint constraint
;
645 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
647 /* Got some memory back in the last second. */
650 if (sysctl_panic_on_oom
== 2) {
651 dump_header(NULL
, gfp_mask
, order
, NULL
);
652 panic("out of memory. Compulsory panic_on_oom is selected.\n");
656 * Check if there were limitations on the allocation (only relevant for
657 * NUMA) that may require different handling.
659 constraint
= constrained_alloc(zonelist
, gfp_mask
, nodemask
);
660 read_lock(&tasklist_lock
);
662 switch (constraint
) {
663 case CONSTRAINT_MEMORY_POLICY
:
664 oom_kill_process(current
, gfp_mask
, order
, 0, NULL
,
665 "No available memory (MPOL_BIND)");
668 case CONSTRAINT_NONE
:
669 if (sysctl_panic_on_oom
) {
670 dump_header(NULL
, gfp_mask
, order
, NULL
);
671 panic("out of memory. panic_on_oom is selected\n");
674 case CONSTRAINT_CPUSET
:
675 __out_of_memory(gfp_mask
, order
);
679 read_unlock(&tasklist_lock
);
682 * Give "p" a good chance of killing itself before we
683 * retry to allocate memory unless "p" is current
685 if (!test_thread_flag(TIF_MEMDIE
))
686 schedule_timeout_uninterruptible(1);