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...
7 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
20 #include <linux/oom.h>
22 #include <linux/err.h>
23 #include <linux/gfp.h>
24 #include <linux/sched.h>
25 #include <linux/sched/mm.h>
26 #include <linux/sched/coredump.h>
27 #include <linux/sched/task.h>
28 #include <linux/swap.h>
29 #include <linux/timex.h>
30 #include <linux/jiffies.h>
31 #include <linux/cpuset.h>
32 #include <linux/export.h>
33 #include <linux/notifier.h>
34 #include <linux/memcontrol.h>
35 #include <linux/mempolicy.h>
36 #include <linux/security.h>
37 #include <linux/ptrace.h>
38 #include <linux/freezer.h>
39 #include <linux/ftrace.h>
40 #include <linux/ratelimit.h>
41 #include <linux/kthread.h>
42 #include <linux/init.h>
43 #include <linux/mmu_notifier.h>
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/oom.h>
51 int sysctl_panic_on_oom
;
52 int sysctl_oom_kill_allocating_task
;
53 int sysctl_oom_dump_tasks
= 1;
55 DEFINE_MUTEX(oom_lock
);
59 * has_intersects_mems_allowed() - check task eligiblity for kill
60 * @start: task struct of which task to consider
61 * @mask: nodemask passed to page allocator for mempolicy ooms
63 * Task eligibility is determined by whether or not a candidate task, @tsk,
64 * shares the same mempolicy nodes as current if it is bound by such a policy
65 * and whether or not it has the same set of allowed cpuset nodes.
67 static bool has_intersects_mems_allowed(struct task_struct
*start
,
68 const nodemask_t
*mask
)
70 struct task_struct
*tsk
;
74 for_each_thread(start
, tsk
) {
77 * If this is a mempolicy constrained oom, tsk's
78 * cpuset is irrelevant. Only return true if its
79 * mempolicy intersects current, otherwise it may be
82 ret
= mempolicy_nodemask_intersects(tsk
, mask
);
85 * This is not a mempolicy constrained oom, so only
86 * check the mems of tsk's cpuset.
88 ret
= cpuset_mems_allowed_intersects(current
, tsk
);
98 static bool has_intersects_mems_allowed(struct task_struct
*tsk
,
99 const nodemask_t
*mask
)
103 #endif /* CONFIG_NUMA */
106 * The process p may have detached its own ->mm while exiting or through
107 * use_mm(), but one or more of its subthreads may still have a valid
108 * pointer. Return p, or any of its subthreads with a valid ->mm, with
111 struct task_struct
*find_lock_task_mm(struct task_struct
*p
)
113 struct task_struct
*t
;
117 for_each_thread(p
, t
) {
131 * order == -1 means the oom kill is required by sysrq, otherwise only
132 * for display purposes.
134 static inline bool is_sysrq_oom(struct oom_control
*oc
)
136 return oc
->order
== -1;
139 static inline bool is_memcg_oom(struct oom_control
*oc
)
141 return oc
->memcg
!= NULL
;
144 /* return true if the task is not adequate as candidate victim task. */
145 static bool oom_unkillable_task(struct task_struct
*p
,
146 struct mem_cgroup
*memcg
, const nodemask_t
*nodemask
)
148 if (is_global_init(p
))
150 if (p
->flags
& PF_KTHREAD
)
153 /* When mem_cgroup_out_of_memory() and p is not member of the group */
154 if (memcg
&& !task_in_mem_cgroup(p
, memcg
))
157 /* p may not have freeable memory in nodemask */
158 if (!has_intersects_mems_allowed(p
, nodemask
))
165 * oom_badness - heuristic function to determine which candidate task to kill
166 * @p: task struct of which task we should calculate
167 * @totalpages: total present RAM allowed for page allocation
169 * The heuristic for determining which task to kill is made to be as simple and
170 * predictable as possible. The goal is to return the highest value for the
171 * task consuming the most memory to avoid subsequent oom failures.
173 unsigned long oom_badness(struct task_struct
*p
, struct mem_cgroup
*memcg
,
174 const nodemask_t
*nodemask
, unsigned long totalpages
)
179 if (oom_unkillable_task(p
, memcg
, nodemask
))
182 p
= find_lock_task_mm(p
);
187 * Do not even consider tasks which are explicitly marked oom
188 * unkillable or have been already oom reaped or the are in
189 * the middle of vfork
191 adj
= (long)p
->signal
->oom_score_adj
;
192 if (adj
== OOM_SCORE_ADJ_MIN
||
193 test_bit(MMF_OOM_SKIP
, &p
->mm
->flags
) ||
200 * The baseline for the badness score is the proportion of RAM that each
201 * task's rss, pagetable and swap space use.
203 points
= get_mm_rss(p
->mm
) + get_mm_counter(p
->mm
, MM_SWAPENTS
) +
204 atomic_long_read(&p
->mm
->nr_ptes
) + mm_nr_pmds(p
->mm
);
208 * Root processes get 3% bonus, just like the __vm_enough_memory()
209 * implementation used by LSMs.
211 if (has_capability_noaudit(p
, CAP_SYS_ADMIN
))
212 points
-= (points
* 3) / 100;
214 /* Normalize to oom_score_adj units */
215 adj
*= totalpages
/ 1000;
219 * Never return 0 for an eligible task regardless of the root bonus and
220 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
222 return points
> 0 ? points
: 1;
225 enum oom_constraint
{
228 CONSTRAINT_MEMORY_POLICY
,
233 * Determine the type of allocation constraint.
235 static enum oom_constraint
constrained_alloc(struct oom_control
*oc
)
239 enum zone_type high_zoneidx
= gfp_zone(oc
->gfp_mask
);
240 bool cpuset_limited
= false;
243 if (is_memcg_oom(oc
)) {
244 oc
->totalpages
= mem_cgroup_get_limit(oc
->memcg
) ?: 1;
245 return CONSTRAINT_MEMCG
;
248 /* Default to all available memory */
249 oc
->totalpages
= totalram_pages
+ total_swap_pages
;
251 if (!IS_ENABLED(CONFIG_NUMA
))
252 return CONSTRAINT_NONE
;
255 return CONSTRAINT_NONE
;
257 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
258 * to kill current.We have to random task kill in this case.
259 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
261 if (oc
->gfp_mask
& __GFP_THISNODE
)
262 return CONSTRAINT_NONE
;
265 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
266 * the page allocator means a mempolicy is in effect. Cpuset policy
267 * is enforced in get_page_from_freelist().
270 !nodes_subset(node_states
[N_MEMORY
], *oc
->nodemask
)) {
271 oc
->totalpages
= total_swap_pages
;
272 for_each_node_mask(nid
, *oc
->nodemask
)
273 oc
->totalpages
+= node_spanned_pages(nid
);
274 return CONSTRAINT_MEMORY_POLICY
;
277 /* Check this allocation failure is caused by cpuset's wall function */
278 for_each_zone_zonelist_nodemask(zone
, z
, oc
->zonelist
,
279 high_zoneidx
, oc
->nodemask
)
280 if (!cpuset_zone_allowed(zone
, oc
->gfp_mask
))
281 cpuset_limited
= true;
283 if (cpuset_limited
) {
284 oc
->totalpages
= total_swap_pages
;
285 for_each_node_mask(nid
, cpuset_current_mems_allowed
)
286 oc
->totalpages
+= node_spanned_pages(nid
);
287 return CONSTRAINT_CPUSET
;
289 return CONSTRAINT_NONE
;
292 static int oom_evaluate_task(struct task_struct
*task
, void *arg
)
294 struct oom_control
*oc
= arg
;
295 unsigned long points
;
297 if (oom_unkillable_task(task
, NULL
, oc
->nodemask
))
301 * This task already has access to memory reserves and is being killed.
302 * Don't allow any other task to have access to the reserves unless
303 * the task has MMF_OOM_SKIP because chances that it would release
304 * any memory is quite low.
306 if (!is_sysrq_oom(oc
) && tsk_is_oom_victim(task
)) {
307 if (test_bit(MMF_OOM_SKIP
, &task
->signal
->oom_mm
->flags
))
313 * If task is allocating a lot of memory and has been marked to be
314 * killed first if it triggers an oom, then select it.
316 if (oom_task_origin(task
)) {
321 points
= oom_badness(task
, NULL
, oc
->nodemask
, oc
->totalpages
);
322 if (!points
|| points
< oc
->chosen_points
)
325 /* Prefer thread group leaders for display purposes */
326 if (points
== oc
->chosen_points
&& thread_group_leader(oc
->chosen
))
330 put_task_struct(oc
->chosen
);
331 get_task_struct(task
);
333 oc
->chosen_points
= points
;
338 put_task_struct(oc
->chosen
);
339 oc
->chosen
= (void *)-1UL;
344 * Simple selection loop. We choose the process with the highest number of
345 * 'points'. In case scan was aborted, oc->chosen is set to -1.
347 static void select_bad_process(struct oom_control
*oc
)
349 if (is_memcg_oom(oc
))
350 mem_cgroup_scan_tasks(oc
->memcg
, oom_evaluate_task
, oc
);
352 struct task_struct
*p
;
356 if (oom_evaluate_task(p
, oc
))
361 oc
->chosen_points
= oc
->chosen_points
* 1000 / oc
->totalpages
;
365 * dump_tasks - dump current memory state of all system tasks
366 * @memcg: current's memory controller, if constrained
367 * @nodemask: nodemask passed to page allocator for mempolicy ooms
369 * Dumps the current memory state of all eligible tasks. Tasks not in the same
370 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
372 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
373 * swapents, oom_score_adj value, and name.
375 static void dump_tasks(struct mem_cgroup
*memcg
, const nodemask_t
*nodemask
)
377 struct task_struct
*p
;
378 struct task_struct
*task
;
380 pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n");
382 for_each_process(p
) {
383 if (oom_unkillable_task(p
, memcg
, nodemask
))
386 task
= find_lock_task_mm(p
);
389 * This is a kthread or all of p's threads have already
390 * detached their mm's. There's no need to report
391 * them; they can't be oom killed anyway.
396 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n",
397 task
->pid
, from_kuid(&init_user_ns
, task_uid(task
)),
398 task
->tgid
, task
->mm
->total_vm
, get_mm_rss(task
->mm
),
399 atomic_long_read(&task
->mm
->nr_ptes
),
400 mm_nr_pmds(task
->mm
),
401 get_mm_counter(task
->mm
, MM_SWAPENTS
),
402 task
->signal
->oom_score_adj
, task
->comm
);
408 static void dump_header(struct oom_control
*oc
, struct task_struct
*p
)
410 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=",
411 current
->comm
, oc
->gfp_mask
, &oc
->gfp_mask
);
413 pr_cont("%*pbl", nodemask_pr_args(oc
->nodemask
));
416 pr_cont(", order=%d, oom_score_adj=%hd\n",
417 oc
->order
, current
->signal
->oom_score_adj
);
418 if (!IS_ENABLED(CONFIG_COMPACTION
) && oc
->order
)
419 pr_warn("COMPACTION is disabled!!!\n");
421 cpuset_print_current_mems_allowed();
424 mem_cgroup_print_oom_info(oc
->memcg
, p
);
426 show_mem(SHOW_MEM_FILTER_NODES
, oc
->nodemask
);
427 if (sysctl_oom_dump_tasks
)
428 dump_tasks(oc
->memcg
, oc
->nodemask
);
432 * Number of OOM victims in flight
434 static atomic_t oom_victims
= ATOMIC_INIT(0);
435 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait
);
437 static bool oom_killer_disabled __read_mostly
;
439 #define K(x) ((x) << (PAGE_SHIFT-10))
442 * task->mm can be NULL if the task is the exited group leader. So to
443 * determine whether the task is using a particular mm, we examine all the
444 * task's threads: if one of those is using this mm then this task was also
447 bool process_shares_mm(struct task_struct
*p
, struct mm_struct
*mm
)
449 struct task_struct
*t
;
451 for_each_thread(p
, t
) {
452 struct mm_struct
*t_mm
= READ_ONCE(t
->mm
);
461 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
462 * victim (if that is possible) to help the OOM killer to move on.
464 static struct task_struct
*oom_reaper_th
;
465 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait
);
466 static struct task_struct
*oom_reaper_list
;
467 static DEFINE_SPINLOCK(oom_reaper_lock
);
469 void __oom_reap_task_mm(struct mm_struct
*mm
)
471 struct vm_area_struct
*vma
;
474 * Tell all users of get_user/copy_from_user etc... that the content
475 * is no longer stable. No barriers really needed because unmapping
476 * should imply barriers already and the reader would hit a page fault
477 * if it stumbled over a reaped memory.
479 set_bit(MMF_UNSTABLE
, &mm
->flags
);
481 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
482 if (!can_madv_dontneed_vma(vma
))
486 * Only anonymous pages have a good chance to be dropped
487 * without additional steps which we cannot afford as we
490 * We do not even care about fs backed pages because all
491 * which are reclaimable have already been reclaimed and
492 * we do not want to block exit_mmap by keeping mm ref
493 * count elevated without a good reason.
495 if (vma_is_anonymous(vma
) || !(vma
->vm_flags
& VM_SHARED
)) {
496 struct mmu_gather tlb
;
498 tlb_gather_mmu(&tlb
, mm
, vma
->vm_start
, vma
->vm_end
);
499 unmap_page_range(&tlb
, vma
, vma
->vm_start
, vma
->vm_end
,
501 tlb_finish_mmu(&tlb
, vma
->vm_start
, vma
->vm_end
);
506 static bool oom_reap_task_mm(struct task_struct
*tsk
, struct mm_struct
*mm
)
511 * We have to make sure to not race with the victim exit path
512 * and cause premature new oom victim selection:
513 * oom_reap_task_mm exit_mm
516 * atomic_dec_and_test
521 * # no TIF_MEMDIE task selects new victim
522 * unmap_page_range # frees some memory
524 mutex_lock(&oom_lock
);
526 if (!down_read_trylock(&mm
->mmap_sem
)) {
528 trace_skip_task_reaping(tsk
->pid
);
533 * If the mm has notifiers then we would need to invalidate them around
534 * unmap_page_range and that is risky because notifiers can sleep and
535 * what they do is basically undeterministic. So let's have a short
536 * sleep to give the oom victim some more time.
537 * TODO: we really want to get rid of this ugly hack and make sure that
538 * notifiers cannot block for unbounded amount of time and add
539 * mmu_notifier_invalidate_range_{start,end} around unmap_page_range
541 if (mm_has_notifiers(mm
)) {
542 up_read(&mm
->mmap_sem
);
543 schedule_timeout_idle(HZ
);
548 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
549 * work on the mm anymore. The check for MMF_OOM_SKIP must run
550 * under mmap_sem for reading because it serializes against the
551 * down_write();up_write() cycle in exit_mmap().
553 if (test_bit(MMF_OOM_SKIP
, &mm
->flags
)) {
554 up_read(&mm
->mmap_sem
);
555 trace_skip_task_reaping(tsk
->pid
);
559 trace_start_task_reaping(tsk
->pid
);
561 __oom_reap_task_mm(mm
);
563 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
564 task_pid_nr(tsk
), tsk
->comm
,
565 K(get_mm_counter(mm
, MM_ANONPAGES
)),
566 K(get_mm_counter(mm
, MM_FILEPAGES
)),
567 K(get_mm_counter(mm
, MM_SHMEMPAGES
)));
568 up_read(&mm
->mmap_sem
);
570 trace_finish_task_reaping(tsk
->pid
);
572 mutex_unlock(&oom_lock
);
576 #define MAX_OOM_REAP_RETRIES 10
577 static void oom_reap_task(struct task_struct
*tsk
)
580 struct mm_struct
*mm
= tsk
->signal
->oom_mm
;
582 /* Retry the down_read_trylock(mmap_sem) a few times */
583 while (attempts
++ < MAX_OOM_REAP_RETRIES
&& !oom_reap_task_mm(tsk
, mm
))
584 schedule_timeout_idle(HZ
/10);
586 if (attempts
<= MAX_OOM_REAP_RETRIES
)
589 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
590 task_pid_nr(tsk
), tsk
->comm
);
591 debug_show_all_locks();
594 tsk
->oom_reaper_list
= NULL
;
597 * Hide this mm from OOM killer because it has been either reaped or
598 * somebody can't call up_write(mmap_sem).
600 set_bit(MMF_OOM_SKIP
, &mm
->flags
);
602 /* Drop a reference taken by wake_oom_reaper */
603 put_task_struct(tsk
);
606 static int oom_reaper(void *unused
)
609 struct task_struct
*tsk
= NULL
;
611 wait_event_freezable(oom_reaper_wait
, oom_reaper_list
!= NULL
);
612 spin_lock(&oom_reaper_lock
);
613 if (oom_reaper_list
!= NULL
) {
614 tsk
= oom_reaper_list
;
615 oom_reaper_list
= tsk
->oom_reaper_list
;
617 spin_unlock(&oom_reaper_lock
);
626 static void wake_oom_reaper(struct task_struct
*tsk
)
631 /* mm is already queued? */
632 if (test_and_set_bit(MMF_OOM_REAP_QUEUED
, &tsk
->signal
->oom_mm
->flags
))
635 get_task_struct(tsk
);
637 spin_lock(&oom_reaper_lock
);
638 tsk
->oom_reaper_list
= oom_reaper_list
;
639 oom_reaper_list
= tsk
;
640 spin_unlock(&oom_reaper_lock
);
641 trace_wake_reaper(tsk
->pid
);
642 wake_up(&oom_reaper_wait
);
645 static int __init
oom_init(void)
647 oom_reaper_th
= kthread_run(oom_reaper
, NULL
, "oom_reaper");
648 if (IS_ERR(oom_reaper_th
)) {
649 pr_err("Unable to start OOM reaper %ld. Continuing regardless\n",
650 PTR_ERR(oom_reaper_th
));
651 oom_reaper_th
= NULL
;
655 subsys_initcall(oom_init
)
657 static inline void wake_oom_reaper(struct task_struct
*tsk
)
660 #endif /* CONFIG_MMU */
663 * mark_oom_victim - mark the given task as OOM victim
666 * Has to be called with oom_lock held and never after
667 * oom has been disabled already.
669 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
670 * under task_lock or operate on the current).
672 static void mark_oom_victim(struct task_struct
*tsk
)
674 struct mm_struct
*mm
= tsk
->mm
;
676 WARN_ON(oom_killer_disabled
);
677 /* OOM killer might race with memcg OOM */
678 if (test_and_set_tsk_thread_flag(tsk
, TIF_MEMDIE
))
681 /* oom_mm is bound to the signal struct life time. */
682 if (!cmpxchg(&tsk
->signal
->oom_mm
, NULL
, mm
)) {
683 mmgrab(tsk
->signal
->oom_mm
);
684 set_bit(MMF_OOM_VICTIM
, &mm
->flags
);
688 * Make sure that the task is woken up from uninterruptible sleep
689 * if it is frozen because OOM killer wouldn't be able to free
690 * any memory and livelock. freezing_slow_path will tell the freezer
691 * that TIF_MEMDIE tasks should be ignored.
694 atomic_inc(&oom_victims
);
695 trace_mark_victim(tsk
->pid
);
699 * exit_oom_victim - note the exit of an OOM victim
701 void exit_oom_victim(void)
703 clear_thread_flag(TIF_MEMDIE
);
705 if (!atomic_dec_return(&oom_victims
))
706 wake_up_all(&oom_victims_wait
);
710 * oom_killer_enable - enable OOM killer
712 void oom_killer_enable(void)
714 oom_killer_disabled
= false;
715 pr_info("OOM killer enabled.\n");
719 * oom_killer_disable - disable OOM killer
720 * @timeout: maximum timeout to wait for oom victims in jiffies
722 * Forces all page allocations to fail rather than trigger OOM killer.
723 * Will block and wait until all OOM victims are killed or the given
726 * The function cannot be called when there are runnable user tasks because
727 * the userspace would see unexpected allocation failures as a result. Any
728 * new usage of this function should be consulted with MM people.
730 * Returns true if successful and false if the OOM killer cannot be
733 bool oom_killer_disable(signed long timeout
)
738 * Make sure to not race with an ongoing OOM killer. Check that the
739 * current is not killed (possibly due to sharing the victim's memory).
741 if (mutex_lock_killable(&oom_lock
))
743 oom_killer_disabled
= true;
744 mutex_unlock(&oom_lock
);
746 ret
= wait_event_interruptible_timeout(oom_victims_wait
,
747 !atomic_read(&oom_victims
), timeout
);
752 pr_info("OOM killer disabled.\n");
757 static inline bool __task_will_free_mem(struct task_struct
*task
)
759 struct signal_struct
*sig
= task
->signal
;
762 * A coredumping process may sleep for an extended period in exit_mm(),
763 * so the oom killer cannot assume that the process will promptly exit
764 * and release memory.
766 if (sig
->flags
& SIGNAL_GROUP_COREDUMP
)
769 if (sig
->flags
& SIGNAL_GROUP_EXIT
)
772 if (thread_group_empty(task
) && (task
->flags
& PF_EXITING
))
779 * Checks whether the given task is dying or exiting and likely to
780 * release its address space. This means that all threads and processes
781 * sharing the same mm have to be killed or exiting.
782 * Caller has to make sure that task->mm is stable (hold task_lock or
783 * it operates on the current).
785 static bool task_will_free_mem(struct task_struct
*task
)
787 struct mm_struct
*mm
= task
->mm
;
788 struct task_struct
*p
;
792 * Skip tasks without mm because it might have passed its exit_mm and
793 * exit_oom_victim. oom_reaper could have rescued that but do not rely
794 * on that for now. We can consider find_lock_task_mm in future.
799 if (!__task_will_free_mem(task
))
803 * This task has already been drained by the oom reaper so there are
804 * only small chances it will free some more
806 if (test_bit(MMF_OOM_SKIP
, &mm
->flags
))
809 if (atomic_read(&mm
->mm_users
) <= 1)
813 * Make sure that all tasks which share the mm with the given tasks
814 * are dying as well to make sure that a) nobody pins its mm and
815 * b) the task is also reapable by the oom reaper.
818 for_each_process(p
) {
819 if (!process_shares_mm(p
, mm
))
821 if (same_thread_group(task
, p
))
823 ret
= __task_will_free_mem(p
);
832 static void oom_kill_process(struct oom_control
*oc
, const char *message
)
834 struct task_struct
*p
= oc
->chosen
;
835 unsigned int points
= oc
->chosen_points
;
836 struct task_struct
*victim
= p
;
837 struct task_struct
*child
;
838 struct task_struct
*t
;
839 struct mm_struct
*mm
;
840 unsigned int victim_points
= 0;
841 static DEFINE_RATELIMIT_STATE(oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
842 DEFAULT_RATELIMIT_BURST
);
843 bool can_oom_reap
= true;
846 * If the task is already exiting, don't alarm the sysadmin or kill
847 * its children or threads, just give it access to memory reserves
848 * so it can die quickly
851 if (task_will_free_mem(p
)) {
860 if (__ratelimit(&oom_rs
))
863 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
864 message
, task_pid_nr(p
), p
->comm
, points
);
867 * If any of p's children has a different mm and is eligible for kill,
868 * the one with the highest oom_badness() score is sacrificed for its
869 * parent. This attempts to lose the minimal amount of work done while
870 * still freeing memory.
872 read_lock(&tasklist_lock
);
875 * The task 'p' might have already exited before reaching here. The
876 * put_task_struct() will free task_struct 'p' while the loop still try
877 * to access the field of 'p', so, get an extra reference.
880 for_each_thread(p
, t
) {
881 list_for_each_entry(child
, &t
->children
, sibling
) {
882 unsigned int child_points
;
884 if (process_shares_mm(child
, p
->mm
))
887 * oom_badness() returns 0 if the thread is unkillable
889 child_points
= oom_badness(child
,
890 oc
->memcg
, oc
->nodemask
, oc
->totalpages
);
891 if (child_points
> victim_points
) {
892 put_task_struct(victim
);
894 victim_points
= child_points
;
895 get_task_struct(victim
);
900 read_unlock(&tasklist_lock
);
902 p
= find_lock_task_mm(victim
);
904 put_task_struct(victim
);
906 } else if (victim
!= p
) {
908 put_task_struct(victim
);
912 /* Get a reference to safely compare mm after task_unlock(victim) */
916 /* Raise event before sending signal: task reaper must see this */
917 count_vm_event(OOM_KILL
);
918 count_memcg_event_mm(mm
, OOM_KILL
);
921 * We should send SIGKILL before granting access to memory reserves
922 * in order to prevent the OOM victim from depleting the memory
923 * reserves from the user space under its control.
925 do_send_sig_info(SIGKILL
, SEND_SIG_FORCED
, victim
, true);
926 mark_oom_victim(victim
);
927 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
928 task_pid_nr(victim
), victim
->comm
, K(victim
->mm
->total_vm
),
929 K(get_mm_counter(victim
->mm
, MM_ANONPAGES
)),
930 K(get_mm_counter(victim
->mm
, MM_FILEPAGES
)),
931 K(get_mm_counter(victim
->mm
, MM_SHMEMPAGES
)));
935 * Kill all user processes sharing victim->mm in other thread groups, if
936 * any. They don't get access to memory reserves, though, to avoid
937 * depletion of all memory. This prevents mm->mmap_sem livelock when an
938 * oom killed thread cannot exit because it requires the semaphore and
939 * its contended by another thread trying to allocate memory itself.
940 * That thread will now get access to memory reserves since it has a
941 * pending fatal signal.
944 for_each_process(p
) {
945 if (!process_shares_mm(p
, mm
))
947 if (same_thread_group(p
, victim
))
949 if (is_global_init(p
)) {
950 can_oom_reap
= false;
951 set_bit(MMF_OOM_SKIP
, &mm
->flags
);
952 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
953 task_pid_nr(victim
), victim
->comm
,
954 task_pid_nr(p
), p
->comm
);
958 * No use_mm() user needs to read from the userspace so we are
961 if (unlikely(p
->flags
& PF_KTHREAD
))
963 do_send_sig_info(SIGKILL
, SEND_SIG_FORCED
, p
, true);
968 wake_oom_reaper(victim
);
971 put_task_struct(victim
);
976 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
978 static void check_panic_on_oom(struct oom_control
*oc
,
979 enum oom_constraint constraint
)
981 if (likely(!sysctl_panic_on_oom
))
983 if (sysctl_panic_on_oom
!= 2) {
985 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
986 * does not panic for cpuset, mempolicy, or memcg allocation
989 if (constraint
!= CONSTRAINT_NONE
)
992 /* Do not panic for oom kills triggered by sysrq */
993 if (is_sysrq_oom(oc
))
995 dump_header(oc
, NULL
);
996 panic("Out of memory: %s panic_on_oom is enabled\n",
997 sysctl_panic_on_oom
== 2 ? "compulsory" : "system-wide");
1000 static BLOCKING_NOTIFIER_HEAD(oom_notify_list
);
1002 int register_oom_notifier(struct notifier_block
*nb
)
1004 return blocking_notifier_chain_register(&oom_notify_list
, nb
);
1006 EXPORT_SYMBOL_GPL(register_oom_notifier
);
1008 int unregister_oom_notifier(struct notifier_block
*nb
)
1010 return blocking_notifier_chain_unregister(&oom_notify_list
, nb
);
1012 EXPORT_SYMBOL_GPL(unregister_oom_notifier
);
1015 * out_of_memory - kill the "best" process when we run out of memory
1016 * @oc: pointer to struct oom_control
1018 * If we run out of memory, we have the choice between either
1019 * killing a random task (bad), letting the system crash (worse)
1020 * OR try to be smart about which process to kill. Note that we
1021 * don't have to be perfect here, we just have to be good.
1023 bool out_of_memory(struct oom_control
*oc
)
1025 unsigned long freed
= 0;
1026 enum oom_constraint constraint
= CONSTRAINT_NONE
;
1028 if (oom_killer_disabled
)
1031 if (!is_memcg_oom(oc
)) {
1032 blocking_notifier_call_chain(&oom_notify_list
, 0, &freed
);
1034 /* Got some memory back in the last second. */
1039 * If current has a pending SIGKILL or is exiting, then automatically
1040 * select it. The goal is to allow it to allocate so that it may
1041 * quickly exit and free its memory.
1043 if (task_will_free_mem(current
)) {
1044 mark_oom_victim(current
);
1045 wake_oom_reaper(current
);
1050 * The OOM killer does not compensate for IO-less reclaim.
1051 * pagefault_out_of_memory lost its gfp context so we have to
1052 * make sure exclude 0 mask - all other users should have at least
1053 * ___GFP_DIRECT_RECLAIM to get here.
1055 if (oc
->gfp_mask
&& !(oc
->gfp_mask
& __GFP_FS
))
1059 * Check if there were limitations on the allocation (only relevant for
1060 * NUMA and memcg) that may require different handling.
1062 constraint
= constrained_alloc(oc
);
1063 if (constraint
!= CONSTRAINT_MEMORY_POLICY
)
1064 oc
->nodemask
= NULL
;
1065 check_panic_on_oom(oc
, constraint
);
1067 if (!is_memcg_oom(oc
) && sysctl_oom_kill_allocating_task
&&
1068 current
->mm
&& !oom_unkillable_task(current
, NULL
, oc
->nodemask
) &&
1069 current
->signal
->oom_score_adj
!= OOM_SCORE_ADJ_MIN
) {
1070 get_task_struct(current
);
1071 oc
->chosen
= current
;
1072 oom_kill_process(oc
, "Out of memory (oom_kill_allocating_task)");
1076 select_bad_process(oc
);
1077 /* Found nothing?!?! Either we hang forever, or we panic. */
1078 if (!oc
->chosen
&& !is_sysrq_oom(oc
) && !is_memcg_oom(oc
)) {
1079 dump_header(oc
, NULL
);
1080 panic("Out of memory and no killable processes...\n");
1082 if (oc
->chosen
&& oc
->chosen
!= (void *)-1UL) {
1083 oom_kill_process(oc
, !is_memcg_oom(oc
) ? "Out of memory" :
1084 "Memory cgroup out of memory");
1086 * Give the killed process a good chance to exit before trying
1087 * to allocate memory again.
1089 schedule_timeout_killable(1);
1091 return !!oc
->chosen
;
1095 * The pagefault handler calls here because it is out of memory, so kill a
1096 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1097 * killing is already in progress so do nothing.
1099 void pagefault_out_of_memory(void)
1101 struct oom_control oc
= {
1109 if (mem_cgroup_oom_synchronize(true))
1112 if (!mutex_trylock(&oom_lock
))
1115 mutex_unlock(&oom_lock
);