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 * Adjust the score by oom_adj.
195 points
>>= -(oom_adj
);
199 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %lu points\n",
200 p
->pid
, p
->comm
, points
);
206 * Determine the type of allocation constraint.
209 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
210 gfp_t gfp_mask
, nodemask_t
*nodemask
)
214 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
217 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
218 * to kill current.We have to random task kill in this case.
219 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
221 if (gfp_mask
& __GFP_THISNODE
)
222 return CONSTRAINT_NONE
;
225 * The nodemask here is a nodemask passed to alloc_pages(). Now,
226 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
227 * feature. mempolicy is an only user of nodemask here.
228 * check mempolicy's nodemask contains all N_HIGH_MEMORY
230 if (nodemask
&& !nodes_subset(node_states
[N_HIGH_MEMORY
], *nodemask
))
231 return CONSTRAINT_MEMORY_POLICY
;
233 /* Check this allocation failure is caused by cpuset's wall function */
234 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
235 high_zoneidx
, nodemask
)
236 if (!cpuset_zone_allowed_softwall(zone
, gfp_mask
))
237 return CONSTRAINT_CPUSET
;
239 return CONSTRAINT_NONE
;
242 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
243 gfp_t gfp_mask
, nodemask_t
*nodemask
)
245 return CONSTRAINT_NONE
;
250 * Simple selection loop. We chose the process with the highest
251 * number of 'points'. We expect the caller will lock the tasklist.
253 * (not docbooked, we don't want this one cluttering up the manual)
255 static struct task_struct
*select_bad_process(unsigned long *ppoints
,
256 struct mem_cgroup
*mem
)
258 struct task_struct
*p
;
259 struct task_struct
*chosen
= NULL
;
260 struct timespec uptime
;
263 do_posix_clock_monotonic_gettime(&uptime
);
264 for_each_process(p
) {
265 unsigned long points
;
267 /* skip the init task and kthreads */
268 if (is_global_init(p
) || (p
->flags
& PF_KTHREAD
))
270 if (mem
&& !task_in_mem_cgroup(p
, mem
))
272 if (!has_intersects_mems_allowed(p
))
276 * This task already has access to memory reserves and is
277 * being killed. Don't allow any other task access to the
280 * Note: this may have a chance of deadlock if it gets
281 * blocked waiting for another task which itself is waiting
282 * for memory. Is there a better alternative?
284 if (test_tsk_thread_flag(p
, TIF_MEMDIE
))
285 return ERR_PTR(-1UL);
288 * This is in the process of releasing memory so wait for it
289 * to finish before killing some other task by mistake.
291 * However, if p is the current task, we allow the 'kill' to
292 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
293 * which will allow it to gain access to memory reserves in
294 * the process of exiting and releasing its resources.
295 * Otherwise we could get an easy OOM deadlock.
297 if ((p
->flags
& PF_EXITING
) && p
->mm
) {
299 return ERR_PTR(-1UL);
302 *ppoints
= ULONG_MAX
;
305 if (p
->signal
->oom_adj
== OOM_DISABLE
)
308 points
= badness(p
, uptime
.tv_sec
);
309 if (points
> *ppoints
|| !chosen
) {
319 * dump_tasks - dump current memory state of all system tasks
320 * @mem: current's memory controller, if constrained
322 * Dumps the current memory state of all system tasks, excluding kernel threads.
323 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
326 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
329 * Call with tasklist_lock read-locked.
331 static void dump_tasks(const struct mem_cgroup
*mem
)
333 struct task_struct
*p
;
334 struct task_struct
*task
;
336 printk(KERN_INFO
"[ pid ] uid tgid total_vm rss cpu oom_adj "
338 for_each_process(p
) {
339 if (p
->flags
& PF_KTHREAD
)
341 if (mem
&& !task_in_mem_cgroup(p
, mem
))
344 task
= find_lock_task_mm(p
);
347 * This is a kthread or all of p's threads have already
348 * detached their mm's. There's no need to report
349 * them; they can't be oom killed anyway.
354 printk(KERN_INFO
"[%5d] %5d %5d %8lu %8lu %3u %3d %s\n",
355 task
->pid
, __task_cred(task
)->uid
, task
->tgid
,
356 task
->mm
->total_vm
, get_mm_rss(task
->mm
),
357 task_cpu(task
), task
->signal
->oom_adj
, task
->comm
);
362 static void dump_header(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
363 struct mem_cgroup
*mem
)
366 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
368 current
->comm
, gfp_mask
, order
, current
->signal
->oom_adj
);
369 cpuset_print_task_mems_allowed(current
);
370 task_unlock(current
);
372 mem_cgroup_print_oom_info(mem
, p
);
374 if (sysctl_oom_dump_tasks
)
378 #define K(x) ((x) << (PAGE_SHIFT-10))
381 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
382 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
385 static void __oom_kill_task(struct task_struct
*p
, int verbose
)
387 if (is_global_init(p
)) {
389 printk(KERN_WARNING
"tried to kill init!\n");
393 p
= find_lock_task_mm(p
);
398 printk(KERN_ERR
"Killed process %d (%s) "
399 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
400 task_pid_nr(p
), p
->comm
,
402 K(get_mm_counter(p
->mm
, MM_ANONPAGES
)),
403 K(get_mm_counter(p
->mm
, MM_FILEPAGES
)));
407 * We give our sacrificial lamb high priority and access to
408 * all the memory it needs. That way it should be able to
409 * exit() and clear out its resources quickly...
411 p
->rt
.time_slice
= HZ
;
412 set_tsk_thread_flag(p
, TIF_MEMDIE
);
414 force_sig(SIGKILL
, p
);
417 static int oom_kill_task(struct task_struct
*p
)
419 /* WARNING: mm may not be dereferenced since we did not obtain its
420 * value from get_task_mm(p). This is OK since all we need to do is
421 * compare mm to q->mm below.
423 * Furthermore, even if mm contains a non-NULL value, p->mm may
424 * change to NULL at any time since we do not hold task_lock(p).
425 * However, this is of no concern to us.
427 if (!p
->mm
|| p
->signal
->oom_adj
== OOM_DISABLE
)
430 __oom_kill_task(p
, 1);
435 static int oom_kill_process(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
436 unsigned long points
, struct mem_cgroup
*mem
,
439 struct task_struct
*victim
= p
;
440 struct task_struct
*child
;
441 struct task_struct
*t
= p
;
442 unsigned long victim_points
= 0;
443 struct timespec uptime
;
445 if (printk_ratelimit())
446 dump_header(p
, gfp_mask
, order
, mem
);
449 * If the task is already exiting, don't alarm the sysadmin or kill
450 * its children or threads, just set TIF_MEMDIE so it can die quickly
452 if (p
->flags
& PF_EXITING
) {
453 set_tsk_thread_flag(p
, TIF_MEMDIE
);
458 pr_err("%s: Kill process %d (%s) score %lu or sacrifice child\n",
459 message
, task_pid_nr(p
), p
->comm
, points
);
463 * If any of p's children has a different mm and is eligible for kill,
464 * the one with the highest badness() score is sacrificed for its
465 * parent. This attempts to lose the minimal amount of work done while
466 * still freeing memory.
468 do_posix_clock_monotonic_gettime(&uptime
);
470 list_for_each_entry(child
, &t
->children
, sibling
) {
471 unsigned long child_points
;
473 if (child
->mm
== p
->mm
)
475 if (mem
&& !task_in_mem_cgroup(child
, mem
))
478 /* badness() returns 0 if the thread is unkillable */
479 child_points
= badness(child
, uptime
.tv_sec
);
480 if (child_points
> victim_points
) {
482 victim_points
= child_points
;
485 } while_each_thread(p
, t
);
487 return oom_kill_task(victim
);
490 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
491 void mem_cgroup_out_of_memory(struct mem_cgroup
*mem
, gfp_t gfp_mask
)
493 unsigned long points
= 0;
494 struct task_struct
*p
;
496 if (sysctl_panic_on_oom
== 2)
497 panic("out of memory(memcg). panic_on_oom is selected.\n");
498 read_lock(&tasklist_lock
);
500 p
= select_bad_process(&points
, mem
);
501 if (!p
|| PTR_ERR(p
) == -1UL)
504 if (oom_kill_process(p
, gfp_mask
, 0, points
, mem
,
505 "Memory cgroup out of memory"))
508 read_unlock(&tasklist_lock
);
512 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
514 int register_oom_notifier(struct notifier_block
*nb
)
516 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
518 EXPORT_SYMBOL_GPL(register_oom_notifier
);
520 int unregister_oom_notifier(struct notifier_block
*nb
)
522 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
524 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
527 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
528 * if a parallel OOM killing is already taking place that includes a zone in
529 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
531 int try_set_zone_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
537 spin_lock(&zone_scan_lock
);
538 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
539 if (zone_is_oom_locked(zone
)) {
545 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
547 * Lock each zone in the zonelist under zone_scan_lock so a
548 * parallel invocation of try_set_zone_oom() doesn't succeed
551 zone_set_flag(zone
, ZONE_OOM_LOCKED
);
555 spin_unlock(&zone_scan_lock
);
560 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
561 * allocation attempts with zonelists containing them may now recall the OOM
562 * killer, if necessary.
564 void clear_zonelist_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
569 spin_lock(&zone_scan_lock
);
570 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
571 zone_clear_flag(zone
, ZONE_OOM_LOCKED
);
573 spin_unlock(&zone_scan_lock
);
577 * Must be called with tasklist_lock held for read.
579 static void __out_of_memory(gfp_t gfp_mask
, int order
)
581 struct task_struct
*p
;
582 unsigned long points
;
584 if (sysctl_oom_kill_allocating_task
)
585 if (!oom_kill_process(current
, gfp_mask
, order
, 0, NULL
,
586 "Out of memory (oom_kill_allocating_task)"))
590 * Rambo mode: Shoot down a process and hope it solves whatever
591 * issues we may have.
593 p
= select_bad_process(&points
, NULL
);
595 if (PTR_ERR(p
) == -1UL)
598 /* Found nothing?!?! Either we hang forever, or we panic. */
600 read_unlock(&tasklist_lock
);
601 dump_header(NULL
, gfp_mask
, order
, NULL
);
602 panic("Out of memory and no killable processes...\n");
605 if (oom_kill_process(p
, gfp_mask
, order
, points
, NULL
,
611 * pagefault handler calls into here because it is out of memory but
612 * doesn't know exactly how or why.
614 void pagefault_out_of_memory(void)
616 unsigned long freed
= 0;
618 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
620 /* Got some memory back in the last second. */
623 if (sysctl_panic_on_oom
)
624 panic("out of memory from page fault. panic_on_oom is selected.\n");
626 read_lock(&tasklist_lock
);
627 __out_of_memory(0, 0); /* unknown gfp_mask and order */
628 read_unlock(&tasklist_lock
);
631 * Give "p" a good chance of killing itself before we
632 * retry to allocate memory.
634 if (!test_thread_flag(TIF_MEMDIE
))
635 schedule_timeout_uninterruptible(1);
639 * out_of_memory - kill the "best" process when we run out of memory
640 * @zonelist: zonelist pointer
641 * @gfp_mask: memory allocation flags
642 * @order: amount of memory being requested as a power of 2
644 * If we run out of memory, we have the choice between either
645 * killing a random task (bad), letting the system crash (worse)
646 * OR try to be smart about which process to kill. Note that we
647 * don't have to be perfect here, we just have to be good.
649 void out_of_memory(struct zonelist
*zonelist
, gfp_t gfp_mask
,
650 int order
, nodemask_t
*nodemask
)
652 unsigned long freed
= 0;
653 enum oom_constraint constraint
;
655 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
657 /* Got some memory back in the last second. */
661 * If current has a pending SIGKILL, then automatically select it. The
662 * goal is to allow it to allocate so that it may quickly exit and free
665 if (fatal_signal_pending(current
)) {
666 set_thread_flag(TIF_MEMDIE
);
670 if (sysctl_panic_on_oom
== 2) {
671 dump_header(NULL
, gfp_mask
, order
, NULL
);
672 panic("out of memory. Compulsory panic_on_oom is selected.\n");
676 * Check if there were limitations on the allocation (only relevant for
677 * NUMA) that may require different handling.
679 constraint
= constrained_alloc(zonelist
, gfp_mask
, nodemask
);
680 read_lock(&tasklist_lock
);
682 switch (constraint
) {
683 case CONSTRAINT_MEMORY_POLICY
:
684 oom_kill_process(current
, gfp_mask
, order
, 0, NULL
,
685 "No available memory (MPOL_BIND)");
688 case CONSTRAINT_NONE
:
689 if (sysctl_panic_on_oom
) {
690 dump_header(NULL
, gfp_mask
, order
, NULL
);
691 panic("out of memory. panic_on_oom is selected\n");
694 case CONSTRAINT_CPUSET
:
695 __out_of_memory(gfp_mask
, order
);
699 read_unlock(&tasklist_lock
);
702 * Give "p" a good chance of killing itself before we
703 * retry to allocate memory unless "p" is current
705 if (!test_thread_flag(TIF_MEMDIE
))
706 schedule_timeout_uninterruptible(1);