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/mempolicy.h>
31 #include <linux/security.h>
33 int sysctl_panic_on_oom
;
34 int sysctl_oom_kill_allocating_task
;
35 int sysctl_oom_dump_tasks
= 1;
36 static DEFINE_SPINLOCK(zone_scan_lock
);
41 * has_intersects_mems_allowed() - check task eligiblity for kill
42 * @tsk: task struct of which task to consider
43 * @mask: nodemask passed to page allocator for mempolicy ooms
45 * Task eligibility is determined by whether or not a candidate task, @tsk,
46 * shares the same mempolicy nodes as current if it is bound by such a policy
47 * and whether or not it has the same set of allowed cpuset nodes.
49 static bool has_intersects_mems_allowed(struct task_struct
*tsk
,
50 const nodemask_t
*mask
)
52 struct task_struct
*start
= tsk
;
57 * If this is a mempolicy constrained oom, tsk's
58 * cpuset is irrelevant. Only return true if its
59 * mempolicy intersects current, otherwise it may be
62 if (mempolicy_nodemask_intersects(tsk
, mask
))
66 * This is not a mempolicy constrained oom, so only
67 * check the mems of tsk's cpuset.
69 if (cpuset_mems_allowed_intersects(current
, tsk
))
72 tsk
= next_thread(tsk
);
73 } while (tsk
!= start
);
77 static bool has_intersects_mems_allowed(struct task_struct
*tsk
,
78 const nodemask_t
*mask
)
82 #endif /* CONFIG_NUMA */
85 * The process p may have detached its own ->mm while exiting or through
86 * use_mm(), but one or more of its subthreads may still have a valid
87 * pointer. Return p, or any of its subthreads with a valid ->mm, with
90 static struct task_struct
*find_lock_task_mm(struct task_struct
*p
)
92 struct task_struct
*t
= p
;
99 } while_each_thread(p
, t
);
104 /* return true if the task is not adequate as candidate victim task. */
105 static bool oom_unkillable_task(struct task_struct
*p
, struct mem_cgroup
*mem
,
106 const nodemask_t
*nodemask
)
108 if (is_global_init(p
))
110 if (p
->flags
& PF_KTHREAD
)
113 /* When mem_cgroup_out_of_memory() and p is not member of the group */
114 if (mem
&& !task_in_mem_cgroup(p
, mem
))
117 /* p may not have freeable memory in nodemask */
118 if (!has_intersects_mems_allowed(p
, nodemask
))
125 * badness - calculate a numeric value for how bad this task has been
126 * @p: task struct of which task we should calculate
127 * @uptime: current uptime in seconds
129 * The formula used is relatively simple and documented inline in the
130 * function. The main rationale is that we want to select a good task
131 * to kill when we run out of memory.
133 * Good in this context means that:
134 * 1) we lose the minimum amount of work done
135 * 2) we recover a large amount of memory
136 * 3) we don't kill anything innocent of eating tons of memory
137 * 4) we want to kill the minimum amount of processes (one)
138 * 5) we try to kill the process the user expects us to kill, this
139 * algorithm has been meticulously tuned to meet the principle
140 * of least surprise ... (be careful when you change it)
143 unsigned long badness(struct task_struct
*p
, unsigned long uptime
)
145 unsigned long points
, cpu_time
, run_time
;
146 struct task_struct
*child
;
147 struct task_struct
*c
, *t
;
148 int oom_adj
= p
->signal
->oom_adj
;
149 struct task_cputime task_time
;
153 if (oom_adj
== OOM_DISABLE
)
156 p
= find_lock_task_mm(p
);
161 * The memory size of the process is the basis for the badness.
163 points
= p
->mm
->total_vm
;
167 * swapoff can easily use up all memory, so kill those first.
169 if (p
->flags
& PF_OOM_ORIGIN
)
173 * Processes which fork a lot of child processes are likely
174 * a good choice. We add half the vmsize of the children if they
175 * have an own mm. This prevents forking servers to flood the
176 * machine with an endless amount of children. In case a single
177 * child is eating the vast majority of memory, adding only half
178 * to the parents will make the child our kill candidate of choice.
182 list_for_each_entry(c
, &t
->children
, sibling
) {
183 child
= find_lock_task_mm(c
);
185 if (child
->mm
!= p
->mm
)
186 points
+= child
->mm
->total_vm
/2 + 1;
190 } while_each_thread(p
, t
);
193 * CPU time is in tens of seconds and run time is in thousands
194 * of seconds. There is no particular reason for this other than
195 * that it turned out to work very well in practice.
197 thread_group_cputime(p
, &task_time
);
198 utime
= cputime_to_jiffies(task_time
.utime
);
199 stime
= cputime_to_jiffies(task_time
.stime
);
200 cpu_time
= (utime
+ stime
) >> (SHIFT_HZ
+ 3);
203 if (uptime
>= p
->start_time
.tv_sec
)
204 run_time
= (uptime
- p
->start_time
.tv_sec
) >> 10;
209 points
/= int_sqrt(cpu_time
);
211 points
/= int_sqrt(int_sqrt(run_time
));
214 * Niced processes are most likely less important, so double
215 * their badness points.
217 if (task_nice(p
) > 0)
221 * Superuser processes are usually more important, so we make it
222 * less likely that we kill those.
224 if (has_capability_noaudit(p
, CAP_SYS_ADMIN
) ||
225 has_capability_noaudit(p
, CAP_SYS_RESOURCE
))
229 * We don't want to kill a process with direct hardware access.
230 * Not only could that mess up the hardware, but usually users
231 * tend to only have this flag set on applications they think
234 if (has_capability_noaudit(p
, CAP_SYS_RAWIO
))
238 * Adjust the score by oom_adj.
246 points
>>= -(oom_adj
);
250 printk(KERN_DEBUG
"OOMkill: task %d (%s) got %lu points\n",
251 p
->pid
, p
->comm
, points
);
257 * Determine the type of allocation constraint.
260 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
261 gfp_t gfp_mask
, nodemask_t
*nodemask
)
265 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
268 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
269 * to kill current.We have to random task kill in this case.
270 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
272 if (gfp_mask
& __GFP_THISNODE
)
273 return CONSTRAINT_NONE
;
276 * The nodemask here is a nodemask passed to alloc_pages(). Now,
277 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
278 * feature. mempolicy is an only user of nodemask here.
279 * check mempolicy's nodemask contains all N_HIGH_MEMORY
281 if (nodemask
&& !nodes_subset(node_states
[N_HIGH_MEMORY
], *nodemask
))
282 return CONSTRAINT_MEMORY_POLICY
;
284 /* Check this allocation failure is caused by cpuset's wall function */
285 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
286 high_zoneidx
, nodemask
)
287 if (!cpuset_zone_allowed_softwall(zone
, gfp_mask
))
288 return CONSTRAINT_CPUSET
;
290 return CONSTRAINT_NONE
;
293 static enum oom_constraint
constrained_alloc(struct zonelist
*zonelist
,
294 gfp_t gfp_mask
, nodemask_t
*nodemask
)
296 return CONSTRAINT_NONE
;
301 * Simple selection loop. We chose the process with the highest
302 * number of 'points'. We expect the caller will lock the tasklist.
304 * (not docbooked, we don't want this one cluttering up the manual)
306 static struct task_struct
*select_bad_process(unsigned long *ppoints
,
307 struct mem_cgroup
*mem
, const nodemask_t
*nodemask
)
309 struct task_struct
*p
;
310 struct task_struct
*chosen
= NULL
;
311 struct timespec uptime
;
314 do_posix_clock_monotonic_gettime(&uptime
);
315 for_each_process(p
) {
316 unsigned long points
;
318 if (oom_unkillable_task(p
, mem
, nodemask
))
322 * This task already has access to memory reserves and is
323 * being killed. Don't allow any other task access to the
326 * Note: this may have a chance of deadlock if it gets
327 * blocked waiting for another task which itself is waiting
328 * for memory. Is there a better alternative?
330 if (test_tsk_thread_flag(p
, TIF_MEMDIE
))
331 return ERR_PTR(-1UL);
334 * This is in the process of releasing memory so wait for it
335 * to finish before killing some other task by mistake.
337 * However, if p is the current task, we allow the 'kill' to
338 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
339 * which will allow it to gain access to memory reserves in
340 * the process of exiting and releasing its resources.
341 * Otherwise we could get an easy OOM deadlock.
343 if ((p
->flags
& PF_EXITING
) && p
->mm
) {
345 return ERR_PTR(-1UL);
348 *ppoints
= ULONG_MAX
;
351 if (p
->signal
->oom_adj
== OOM_DISABLE
)
354 points
= badness(p
, uptime
.tv_sec
);
355 if (points
> *ppoints
|| !chosen
) {
365 * dump_tasks - dump current memory state of all system tasks
366 * @mem: current's memory controller, if constrained
368 * Dumps the current memory state of all system tasks, excluding kernel threads.
369 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
372 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
375 * Call with tasklist_lock read-locked.
377 static void dump_tasks(const struct mem_cgroup
*mem
)
379 struct task_struct
*p
;
380 struct task_struct
*task
;
382 printk(KERN_INFO
"[ pid ] uid tgid total_vm rss cpu oom_adj "
384 for_each_process(p
) {
385 if (p
->flags
& PF_KTHREAD
)
387 if (mem
&& !task_in_mem_cgroup(p
, mem
))
390 task
= find_lock_task_mm(p
);
393 * This is a kthread or all of p's threads have already
394 * detached their mm's. There's no need to report
395 * them; they can't be oom killed anyway.
400 printk(KERN_INFO
"[%5d] %5d %5d %8lu %8lu %3u %3d %s\n",
401 task
->pid
, __task_cred(task
)->uid
, task
->tgid
,
402 task
->mm
->total_vm
, get_mm_rss(task
->mm
),
403 task_cpu(task
), task
->signal
->oom_adj
, task
->comm
);
408 static void dump_header(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
409 struct mem_cgroup
*mem
)
412 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
414 current
->comm
, gfp_mask
, order
, current
->signal
->oom_adj
);
415 cpuset_print_task_mems_allowed(current
);
416 task_unlock(current
);
418 mem_cgroup_print_oom_info(mem
, p
);
420 if (sysctl_oom_dump_tasks
)
424 #define K(x) ((x) << (PAGE_SHIFT-10))
425 static int oom_kill_task(struct task_struct
*p
)
427 p
= find_lock_task_mm(p
);
428 if (!p
|| p
->signal
->oom_adj
== OOM_DISABLE
) {
432 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
433 task_pid_nr(p
), p
->comm
, K(p
->mm
->total_vm
),
434 K(get_mm_counter(p
->mm
, MM_ANONPAGES
)),
435 K(get_mm_counter(p
->mm
, MM_FILEPAGES
)));
438 p
->rt
.time_slice
= HZ
;
439 set_tsk_thread_flag(p
, TIF_MEMDIE
);
440 force_sig(SIGKILL
, p
);
445 static int oom_kill_process(struct task_struct
*p
, gfp_t gfp_mask
, int order
,
446 unsigned long points
, struct mem_cgroup
*mem
,
447 nodemask_t
*nodemask
, const char *message
)
449 struct task_struct
*victim
= p
;
450 struct task_struct
*child
;
451 struct task_struct
*t
= p
;
452 unsigned long victim_points
= 0;
453 struct timespec uptime
;
455 if (printk_ratelimit())
456 dump_header(p
, gfp_mask
, order
, mem
);
459 * If the task is already exiting, don't alarm the sysadmin or kill
460 * its children or threads, just set TIF_MEMDIE so it can die quickly
462 if (p
->flags
& PF_EXITING
) {
463 set_tsk_thread_flag(p
, TIF_MEMDIE
);
468 pr_err("%s: Kill process %d (%s) score %lu or sacrifice child\n",
469 message
, task_pid_nr(p
), p
->comm
, points
);
473 * If any of p's children has a different mm and is eligible for kill,
474 * the one with the highest badness() score is sacrificed for its
475 * parent. This attempts to lose the minimal amount of work done while
476 * still freeing memory.
478 do_posix_clock_monotonic_gettime(&uptime
);
480 list_for_each_entry(child
, &t
->children
, sibling
) {
481 unsigned long child_points
;
483 if (child
->mm
== p
->mm
)
485 if (oom_unkillable_task(p
, mem
, nodemask
))
488 /* badness() returns 0 if the thread is unkillable */
489 child_points
= badness(child
, uptime
.tv_sec
);
490 if (child_points
> victim_points
) {
492 victim_points
= child_points
;
495 } while_each_thread(p
, t
);
497 return oom_kill_task(victim
);
501 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
503 static void check_panic_on_oom(enum oom_constraint constraint
, gfp_t gfp_mask
,
506 if (likely(!sysctl_panic_on_oom
))
508 if (sysctl_panic_on_oom
!= 2) {
510 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
511 * does not panic for cpuset, mempolicy, or memcg allocation
514 if (constraint
!= CONSTRAINT_NONE
)
517 read_lock(&tasklist_lock
);
518 dump_header(NULL
, gfp_mask
, order
, NULL
);
519 read_unlock(&tasklist_lock
);
520 panic("Out of memory: %s panic_on_oom is enabled\n",
521 sysctl_panic_on_oom
== 2 ? "compulsory" : "system-wide");
524 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
525 void mem_cgroup_out_of_memory(struct mem_cgroup
*mem
, gfp_t gfp_mask
)
527 unsigned long points
= 0;
528 struct task_struct
*p
;
530 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, 0);
531 read_lock(&tasklist_lock
);
533 p
= select_bad_process(&points
, mem
, NULL
);
534 if (!p
|| PTR_ERR(p
) == -1UL)
537 if (oom_kill_process(p
, gfp_mask
, 0, points
, mem
, NULL
,
538 "Memory cgroup out of memory"))
541 read_unlock(&tasklist_lock
);
545 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
547 int register_oom_notifier(struct notifier_block
*nb
)
549 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
551 EXPORT_SYMBOL_GPL(register_oom_notifier
);
553 int unregister_oom_notifier(struct notifier_block
*nb
)
555 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
557 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
560 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
561 * if a parallel OOM killing is already taking place that includes a zone in
562 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
564 int try_set_zonelist_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
570 spin_lock(&zone_scan_lock
);
571 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
572 if (zone_is_oom_locked(zone
)) {
578 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
580 * Lock each zone in the zonelist under zone_scan_lock so a
581 * parallel invocation of try_set_zonelist_oom() doesn't succeed
584 zone_set_flag(zone
, ZONE_OOM_LOCKED
);
588 spin_unlock(&zone_scan_lock
);
593 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
594 * allocation attempts with zonelists containing them may now recall the OOM
595 * killer, if necessary.
597 void clear_zonelist_oom(struct zonelist
*zonelist
, gfp_t gfp_mask
)
602 spin_lock(&zone_scan_lock
);
603 for_each_zone_zonelist(zone
, z
, zonelist
, gfp_zone(gfp_mask
)) {
604 zone_clear_flag(zone
, ZONE_OOM_LOCKED
);
606 spin_unlock(&zone_scan_lock
);
610 * Try to acquire the oom killer lock for all system zones. Returns zero if a
611 * parallel oom killing is taking place, otherwise locks all zones and returns
614 static int try_set_system_oom(void)
619 spin_lock(&zone_scan_lock
);
620 for_each_populated_zone(zone
)
621 if (zone_is_oom_locked(zone
)) {
625 for_each_populated_zone(zone
)
626 zone_set_flag(zone
, ZONE_OOM_LOCKED
);
628 spin_unlock(&zone_scan_lock
);
633 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
634 * attempts or page faults may now recall the oom killer, if necessary.
636 static void clear_system_oom(void)
640 spin_lock(&zone_scan_lock
);
641 for_each_populated_zone(zone
)
642 zone_clear_flag(zone
, ZONE_OOM_LOCKED
);
643 spin_unlock(&zone_scan_lock
);
647 * out_of_memory - kill the "best" process when we run out of memory
648 * @zonelist: zonelist pointer
649 * @gfp_mask: memory allocation flags
650 * @order: amount of memory being requested as a power of 2
651 * @nodemask: nodemask passed to page allocator
653 * If we run out of memory, we have the choice between either
654 * killing a random task (bad), letting the system crash (worse)
655 * OR try to be smart about which process to kill. Note that we
656 * don't have to be perfect here, we just have to be good.
658 void out_of_memory(struct zonelist
*zonelist
, gfp_t gfp_mask
,
659 int order
, nodemask_t
*nodemask
)
661 struct task_struct
*p
;
662 unsigned long freed
= 0;
663 unsigned long points
;
664 enum oom_constraint constraint
= CONSTRAINT_NONE
;
666 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
668 /* Got some memory back in the last second. */
672 * If current has a pending SIGKILL, then automatically select it. The
673 * goal is to allow it to allocate so that it may quickly exit and free
676 if (fatal_signal_pending(current
)) {
677 set_thread_flag(TIF_MEMDIE
);
682 * Check if there were limitations on the allocation (only relevant for
683 * NUMA) that may require different handling.
686 constraint
= constrained_alloc(zonelist
, gfp_mask
, nodemask
);
687 check_panic_on_oom(constraint
, gfp_mask
, order
);
689 read_lock(&tasklist_lock
);
690 if (sysctl_oom_kill_allocating_task
) {
692 * oom_kill_process() needs tasklist_lock held. If it returns
693 * non-zero, current could not be killed so we must fallback to
696 if (!oom_kill_process(current
, gfp_mask
, order
, 0, NULL
,
698 "Out of memory (oom_kill_allocating_task)"))
703 p
= select_bad_process(&points
, NULL
,
704 constraint
== CONSTRAINT_MEMORY_POLICY
? nodemask
:
706 if (PTR_ERR(p
) == -1UL)
709 /* Found nothing?!?! Either we hang forever, or we panic. */
711 dump_header(NULL
, gfp_mask
, order
, NULL
);
712 read_unlock(&tasklist_lock
);
713 panic("Out of memory and no killable processes...\n");
716 if (oom_kill_process(p
, gfp_mask
, order
, points
, NULL
, nodemask
,
719 read_unlock(&tasklist_lock
);
722 * Give "p" a good chance of killing itself before we
723 * retry to allocate memory unless "p" is current
725 if (!test_thread_flag(TIF_MEMDIE
))
726 schedule_timeout_uninterruptible(1);
730 * The pagefault handler calls here because it is out of memory, so kill a
731 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
732 * oom killing is already in progress so do nothing. If a task is found with
733 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
735 void pagefault_out_of_memory(void)
737 if (try_set_system_oom()) {
738 out_of_memory(NULL
, 0, 0, NULL
);
741 if (!test_thread_flag(TIF_MEMDIE
))
742 schedule_timeout_uninterruptible(1);