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>
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/oom.h>
52 int sysctl_panic_on_oom;
53 int sysctl_oom_kill_allocating_task;
54 int sysctl_oom_dump_tasks = 1;
57 * Serializes oom killer invocations (out_of_memory()) from all contexts to
58 * prevent from over eager oom killing (e.g. when the oom killer is invoked
59 * from different domains).
61 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
64 DEFINE_MUTEX(oom_lock);
65 /* Serializes oom_score_adj and oom_score_adj_min updates */
66 DEFINE_MUTEX(oom_adj_mutex);
70 * has_intersects_mems_allowed() - check task eligiblity for kill
71 * @start: task struct of which task to consider
72 * @mask: nodemask passed to page allocator for mempolicy ooms
74 * Task eligibility is determined by whether or not a candidate task, @tsk,
75 * shares the same mempolicy nodes as current if it is bound by such a policy
76 * and whether or not it has the same set of allowed cpuset nodes.
78 static bool has_intersects_mems_allowed(struct task_struct *start,
79 const nodemask_t *mask)
81 struct task_struct *tsk;
85 for_each_thread(start, tsk) {
88 * If this is a mempolicy constrained oom, tsk's
89 * cpuset is irrelevant. Only return true if its
90 * mempolicy intersects current, otherwise it may be
93 ret = mempolicy_nodemask_intersects(tsk, mask);
96 * This is not a mempolicy constrained oom, so only
97 * check the mems of tsk's cpuset.
99 ret = cpuset_mems_allowed_intersects(current, tsk);
109 static bool has_intersects_mems_allowed(struct task_struct *tsk,
110 const nodemask_t *mask)
114 #endif /* CONFIG_NUMA */
117 * The process p may have detached its own ->mm while exiting or through
118 * use_mm(), but one or more of its subthreads may still have a valid
119 * pointer. Return p, or any of its subthreads with a valid ->mm, with
122 struct task_struct *find_lock_task_mm(struct task_struct *p)
124 struct task_struct *t;
128 for_each_thread(p, t) {
142 * order == -1 means the oom kill is required by sysrq, otherwise only
143 * for display purposes.
145 static inline bool is_sysrq_oom(struct oom_control *oc)
147 return oc->order == -1;
150 static inline bool is_memcg_oom(struct oom_control *oc)
152 return oc->memcg != NULL;
155 /* return true if the task is not adequate as candidate victim task. */
156 static bool oom_unkillable_task(struct task_struct *p,
157 struct mem_cgroup *memcg, const nodemask_t *nodemask)
159 if (is_global_init(p))
161 if (p->flags & PF_KTHREAD)
164 /* When mem_cgroup_out_of_memory() and p is not member of the group */
165 if (memcg && !task_in_mem_cgroup(p, memcg))
168 /* p may not have freeable memory in nodemask */
169 if (!has_intersects_mems_allowed(p, nodemask))
176 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
177 * than all user memory (LRU pages)
179 static bool is_dump_unreclaim_slabs(void)
181 unsigned long nr_lru;
183 nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
184 global_node_page_state(NR_INACTIVE_ANON) +
185 global_node_page_state(NR_ACTIVE_FILE) +
186 global_node_page_state(NR_INACTIVE_FILE) +
187 global_node_page_state(NR_ISOLATED_ANON) +
188 global_node_page_state(NR_ISOLATED_FILE) +
189 global_node_page_state(NR_UNEVICTABLE);
191 return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
195 * oom_badness - heuristic function to determine which candidate task to kill
196 * @p: task struct of which task we should calculate
197 * @totalpages: total present RAM allowed for page allocation
198 * @memcg: task's memory controller, if constrained
199 * @nodemask: nodemask passed to page allocator for mempolicy ooms
201 * The heuristic for determining which task to kill is made to be as simple and
202 * predictable as possible. The goal is to return the highest value for the
203 * task consuming the most memory to avoid subsequent oom failures.
205 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
206 const nodemask_t *nodemask, unsigned long totalpages)
211 if (oom_unkillable_task(p, memcg, nodemask))
214 p = find_lock_task_mm(p);
219 * Do not even consider tasks which are explicitly marked oom
220 * unkillable or have been already oom reaped or the are in
221 * the middle of vfork
223 adj = (long)p->signal->oom_score_adj;
224 if (adj == OOM_SCORE_ADJ_MIN ||
225 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
232 * The baseline for the badness score is the proportion of RAM that each
233 * task's rss, pagetable and swap space use.
235 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
236 mm_pgtables_bytes(p->mm) / PAGE_SIZE;
239 /* Normalize to oom_score_adj units */
240 adj *= totalpages / 1000;
244 * Never return 0 for an eligible task regardless of the root bonus and
245 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
247 return points > 0 ? points : 1;
250 enum oom_constraint {
253 CONSTRAINT_MEMORY_POLICY,
258 * Determine the type of allocation constraint.
260 static enum oom_constraint constrained_alloc(struct oom_control *oc)
264 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
265 bool cpuset_limited = false;
268 if (is_memcg_oom(oc)) {
269 oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
270 return CONSTRAINT_MEMCG;
273 /* Default to all available memory */
274 oc->totalpages = totalram_pages + total_swap_pages;
276 if (!IS_ENABLED(CONFIG_NUMA))
277 return CONSTRAINT_NONE;
280 return CONSTRAINT_NONE;
282 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
283 * to kill current.We have to random task kill in this case.
284 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
286 if (oc->gfp_mask & __GFP_THISNODE)
287 return CONSTRAINT_NONE;
290 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
291 * the page allocator means a mempolicy is in effect. Cpuset policy
292 * is enforced in get_page_from_freelist().
295 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
296 oc->totalpages = total_swap_pages;
297 for_each_node_mask(nid, *oc->nodemask)
298 oc->totalpages += node_spanned_pages(nid);
299 return CONSTRAINT_MEMORY_POLICY;
302 /* Check this allocation failure is caused by cpuset's wall function */
303 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
304 high_zoneidx, oc->nodemask)
305 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
306 cpuset_limited = true;
308 if (cpuset_limited) {
309 oc->totalpages = total_swap_pages;
310 for_each_node_mask(nid, cpuset_current_mems_allowed)
311 oc->totalpages += node_spanned_pages(nid);
312 return CONSTRAINT_CPUSET;
314 return CONSTRAINT_NONE;
317 static int oom_evaluate_task(struct task_struct *task, void *arg)
319 struct oom_control *oc = arg;
320 unsigned long points;
322 if (oom_unkillable_task(task, NULL, oc->nodemask))
326 * This task already has access to memory reserves and is being killed.
327 * Don't allow any other task to have access to the reserves unless
328 * the task has MMF_OOM_SKIP because chances that it would release
329 * any memory is quite low.
331 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
332 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
338 * If task is allocating a lot of memory and has been marked to be
339 * killed first if it triggers an oom, then select it.
341 if (oom_task_origin(task)) {
346 points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
347 if (!points || points < oc->chosen_points)
350 /* Prefer thread group leaders for display purposes */
351 if (points == oc->chosen_points && thread_group_leader(oc->chosen))
355 put_task_struct(oc->chosen);
356 get_task_struct(task);
358 oc->chosen_points = points;
363 put_task_struct(oc->chosen);
364 oc->chosen = (void *)-1UL;
369 * Simple selection loop. We choose the process with the highest number of
370 * 'points'. In case scan was aborted, oc->chosen is set to -1.
372 static void select_bad_process(struct oom_control *oc)
374 if (is_memcg_oom(oc))
375 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
377 struct task_struct *p;
381 if (oom_evaluate_task(p, oc))
386 oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages;
390 * dump_tasks - dump current memory state of all system tasks
391 * @memcg: current's memory controller, if constrained
392 * @nodemask: nodemask passed to page allocator for mempolicy ooms
394 * Dumps the current memory state of all eligible tasks. Tasks not in the same
395 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
397 * State information includes task's pid, uid, tgid, vm size, rss,
398 * pgtables_bytes, swapents, oom_score_adj value, and name.
400 static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
402 struct task_struct *p;
403 struct task_struct *task;
405 pr_info("Tasks state (memory values in pages):\n");
406 pr_info("[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name\n");
408 for_each_process(p) {
409 if (oom_unkillable_task(p, memcg, nodemask))
412 task = find_lock_task_mm(p);
415 * This is a kthread or all of p's threads have already
416 * detached their mm's. There's no need to report
417 * them; they can't be oom killed anyway.
422 pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu %5hd %s\n",
423 task->pid, from_kuid(&init_user_ns, task_uid(task)),
424 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
425 mm_pgtables_bytes(task->mm),
426 get_mm_counter(task->mm, MM_SWAPENTS),
427 task->signal->oom_score_adj, task->comm);
433 static void dump_header(struct oom_control *oc, struct task_struct *p)
435 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n",
436 current->comm, oc->gfp_mask, &oc->gfp_mask,
437 nodemask_pr_args(oc->nodemask), oc->order,
438 current->signal->oom_score_adj);
439 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
440 pr_warn("COMPACTION is disabled!!!\n");
442 cpuset_print_current_mems_allowed();
444 if (is_memcg_oom(oc))
445 mem_cgroup_print_oom_info(oc->memcg, p);
447 show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
448 if (is_dump_unreclaim_slabs())
449 dump_unreclaimable_slab();
451 if (sysctl_oom_dump_tasks)
452 dump_tasks(oc->memcg, oc->nodemask);
456 * Number of OOM victims in flight
458 static atomic_t oom_victims = ATOMIC_INIT(0);
459 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
461 static bool oom_killer_disabled __read_mostly;
463 #define K(x) ((x) << (PAGE_SHIFT-10))
466 * task->mm can be NULL if the task is the exited group leader. So to
467 * determine whether the task is using a particular mm, we examine all the
468 * task's threads: if one of those is using this mm then this task was also
471 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
473 struct task_struct *t;
475 for_each_thread(p, t) {
476 struct mm_struct *t_mm = READ_ONCE(t->mm);
485 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
486 * victim (if that is possible) to help the OOM killer to move on.
488 static struct task_struct *oom_reaper_th;
489 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
490 static struct task_struct *oom_reaper_list;
491 static DEFINE_SPINLOCK(oom_reaper_lock);
493 bool __oom_reap_task_mm(struct mm_struct *mm)
495 struct vm_area_struct *vma;
499 * Tell all users of get_user/copy_from_user etc... that the content
500 * is no longer stable. No barriers really needed because unmapping
501 * should imply barriers already and the reader would hit a page fault
502 * if it stumbled over a reaped memory.
504 set_bit(MMF_UNSTABLE, &mm->flags);
506 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
507 if (!can_madv_dontneed_vma(vma))
511 * Only anonymous pages have a good chance to be dropped
512 * without additional steps which we cannot afford as we
515 * We do not even care about fs backed pages because all
516 * which are reclaimable have already been reclaimed and
517 * we do not want to block exit_mmap by keeping mm ref
518 * count elevated without a good reason.
520 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
521 const unsigned long start = vma->vm_start;
522 const unsigned long end = vma->vm_end;
523 struct mmu_gather tlb;
525 tlb_gather_mmu(&tlb, mm, start, end);
526 if (mmu_notifier_invalidate_range_start_nonblock(mm, start, end)) {
527 tlb_finish_mmu(&tlb, start, end);
531 unmap_page_range(&tlb, vma, start, end, NULL);
532 mmu_notifier_invalidate_range_end(mm, start, end);
533 tlb_finish_mmu(&tlb, start, end);
541 * Reaps the address space of the give task.
543 * Returns true on success and false if none or part of the address space
544 * has been reclaimed and the caller should retry later.
546 static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
550 if (!down_read_trylock(&mm->mmap_sem)) {
551 trace_skip_task_reaping(tsk->pid);
556 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
557 * work on the mm anymore. The check for MMF_OOM_SKIP must run
558 * under mmap_sem for reading because it serializes against the
559 * down_write();up_write() cycle in exit_mmap().
561 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
562 trace_skip_task_reaping(tsk->pid);
566 trace_start_task_reaping(tsk->pid);
568 /* failed to reap part of the address space. Try again later */
569 ret = __oom_reap_task_mm(mm);
573 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
574 task_pid_nr(tsk), tsk->comm,
575 K(get_mm_counter(mm, MM_ANONPAGES)),
576 K(get_mm_counter(mm, MM_FILEPAGES)),
577 K(get_mm_counter(mm, MM_SHMEMPAGES)));
579 trace_finish_task_reaping(tsk->pid);
581 up_read(&mm->mmap_sem);
586 #define MAX_OOM_REAP_RETRIES 10
587 static void oom_reap_task(struct task_struct *tsk)
590 struct mm_struct *mm = tsk->signal->oom_mm;
592 /* Retry the down_read_trylock(mmap_sem) a few times */
593 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
594 schedule_timeout_idle(HZ/10);
596 if (attempts <= MAX_OOM_REAP_RETRIES ||
597 test_bit(MMF_OOM_SKIP, &mm->flags))
600 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
601 task_pid_nr(tsk), tsk->comm);
602 debug_show_all_locks();
605 tsk->oom_reaper_list = NULL;
608 * Hide this mm from OOM killer because it has been either reaped or
609 * somebody can't call up_write(mmap_sem).
611 set_bit(MMF_OOM_SKIP, &mm->flags);
613 /* Drop a reference taken by wake_oom_reaper */
614 put_task_struct(tsk);
617 static int oom_reaper(void *unused)
620 struct task_struct *tsk = NULL;
622 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
623 spin_lock(&oom_reaper_lock);
624 if (oom_reaper_list != NULL) {
625 tsk = oom_reaper_list;
626 oom_reaper_list = tsk->oom_reaper_list;
628 spin_unlock(&oom_reaper_lock);
637 static void wake_oom_reaper(struct task_struct *tsk)
639 /* mm is already queued? */
640 if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
643 get_task_struct(tsk);
645 spin_lock(&oom_reaper_lock);
646 tsk->oom_reaper_list = oom_reaper_list;
647 oom_reaper_list = tsk;
648 spin_unlock(&oom_reaper_lock);
649 trace_wake_reaper(tsk->pid);
650 wake_up(&oom_reaper_wait);
653 static int __init oom_init(void)
655 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
658 subsys_initcall(oom_init)
660 static inline void wake_oom_reaper(struct task_struct *tsk)
663 #endif /* CONFIG_MMU */
666 * mark_oom_victim - mark the given task as OOM victim
669 * Has to be called with oom_lock held and never after
670 * oom has been disabled already.
672 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
673 * under task_lock or operate on the current).
675 static void mark_oom_victim(struct task_struct *tsk)
677 struct mm_struct *mm = tsk->mm;
679 WARN_ON(oom_killer_disabled);
680 /* OOM killer might race with memcg OOM */
681 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
684 /* oom_mm is bound to the signal struct life time. */
685 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
686 mmgrab(tsk->signal->oom_mm);
687 set_bit(MMF_OOM_VICTIM, &mm->flags);
691 * Make sure that the task is woken up from uninterruptible sleep
692 * if it is frozen because OOM killer wouldn't be able to free
693 * any memory and livelock. freezing_slow_path will tell the freezer
694 * that TIF_MEMDIE tasks should be ignored.
697 atomic_inc(&oom_victims);
698 trace_mark_victim(tsk->pid);
702 * exit_oom_victim - note the exit of an OOM victim
704 void exit_oom_victim(void)
706 clear_thread_flag(TIF_MEMDIE);
708 if (!atomic_dec_return(&oom_victims))
709 wake_up_all(&oom_victims_wait);
713 * oom_killer_enable - enable OOM killer
715 void oom_killer_enable(void)
717 oom_killer_disabled = false;
718 pr_info("OOM killer enabled.\n");
722 * oom_killer_disable - disable OOM killer
723 * @timeout: maximum timeout to wait for oom victims in jiffies
725 * Forces all page allocations to fail rather than trigger OOM killer.
726 * Will block and wait until all OOM victims are killed or the given
729 * The function cannot be called when there are runnable user tasks because
730 * the userspace would see unexpected allocation failures as a result. Any
731 * new usage of this function should be consulted with MM people.
733 * Returns true if successful and false if the OOM killer cannot be
736 bool oom_killer_disable(signed long timeout)
741 * Make sure to not race with an ongoing OOM killer. Check that the
742 * current is not killed (possibly due to sharing the victim's memory).
744 if (mutex_lock_killable(&oom_lock))
746 oom_killer_disabled = true;
747 mutex_unlock(&oom_lock);
749 ret = wait_event_interruptible_timeout(oom_victims_wait,
750 !atomic_read(&oom_victims), timeout);
755 pr_info("OOM killer disabled.\n");
760 static inline bool __task_will_free_mem(struct task_struct *task)
762 struct signal_struct *sig = task->signal;
765 * A coredumping process may sleep for an extended period in exit_mm(),
766 * so the oom killer cannot assume that the process will promptly exit
767 * and release memory.
769 if (sig->flags & SIGNAL_GROUP_COREDUMP)
772 if (sig->flags & SIGNAL_GROUP_EXIT)
775 if (thread_group_empty(task) && (task->flags & PF_EXITING))
782 * Checks whether the given task is dying or exiting and likely to
783 * release its address space. This means that all threads and processes
784 * sharing the same mm have to be killed or exiting.
785 * Caller has to make sure that task->mm is stable (hold task_lock or
786 * it operates on the current).
788 static bool task_will_free_mem(struct task_struct *task)
790 struct mm_struct *mm = task->mm;
791 struct task_struct *p;
795 * Skip tasks without mm because it might have passed its exit_mm and
796 * exit_oom_victim. oom_reaper could have rescued that but do not rely
797 * on that for now. We can consider find_lock_task_mm in future.
802 if (!__task_will_free_mem(task))
806 * This task has already been drained by the oom reaper so there are
807 * only small chances it will free some more
809 if (test_bit(MMF_OOM_SKIP, &mm->flags))
812 if (atomic_read(&mm->mm_users) <= 1)
816 * Make sure that all tasks which share the mm with the given tasks
817 * are dying as well to make sure that a) nobody pins its mm and
818 * b) the task is also reapable by the oom reaper.
821 for_each_process(p) {
822 if (!process_shares_mm(p, mm))
824 if (same_thread_group(task, p))
826 ret = __task_will_free_mem(p);
835 static void __oom_kill_process(struct task_struct *victim)
837 struct task_struct *p;
838 struct mm_struct *mm;
839 bool can_oom_reap = true;
841 p = find_lock_task_mm(victim);
843 put_task_struct(victim);
845 } else if (victim != p) {
847 put_task_struct(victim);
851 /* Get a reference to safely compare mm after task_unlock(victim) */
855 /* Raise event before sending signal: task reaper must see this */
856 count_vm_event(OOM_KILL);
857 memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
860 * We should send SIGKILL before granting access to memory reserves
861 * in order to prevent the OOM victim from depleting the memory
862 * reserves from the user space under its control.
864 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, PIDTYPE_TGID);
865 mark_oom_victim(victim);
866 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
867 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
868 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
869 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
870 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
874 * Kill all user processes sharing victim->mm in other thread groups, if
875 * any. They don't get access to memory reserves, though, to avoid
876 * depletion of all memory. This prevents mm->mmap_sem livelock when an
877 * oom killed thread cannot exit because it requires the semaphore and
878 * its contended by another thread trying to allocate memory itself.
879 * That thread will now get access to memory reserves since it has a
880 * pending fatal signal.
883 for_each_process(p) {
884 if (!process_shares_mm(p, mm))
886 if (same_thread_group(p, victim))
888 if (is_global_init(p)) {
889 can_oom_reap = false;
890 set_bit(MMF_OOM_SKIP, &mm->flags);
891 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
892 task_pid_nr(victim), victim->comm,
893 task_pid_nr(p), p->comm);
897 * No use_mm() user needs to read from the userspace so we are
900 if (unlikely(p->flags & PF_KTHREAD))
902 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, PIDTYPE_TGID);
907 wake_oom_reaper(victim);
910 put_task_struct(victim);
915 * Kill provided task unless it's secured by setting
916 * oom_score_adj to OOM_SCORE_ADJ_MIN.
918 static int oom_kill_memcg_member(struct task_struct *task, void *unused)
920 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
921 !is_global_init(task)) {
922 get_task_struct(task);
923 __oom_kill_process(task);
928 static void oom_kill_process(struct oom_control *oc, const char *message)
930 struct task_struct *p = oc->chosen;
931 unsigned int points = oc->chosen_points;
932 struct task_struct *victim = p;
933 struct task_struct *child;
934 struct task_struct *t;
935 struct mem_cgroup *oom_group;
936 unsigned int victim_points = 0;
937 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
938 DEFAULT_RATELIMIT_BURST);
941 * If the task is already exiting, don't alarm the sysadmin or kill
942 * its children or threads, just give it access to memory reserves
943 * so it can die quickly
946 if (task_will_free_mem(p)) {
955 if (__ratelimit(&oom_rs))
958 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
959 message, task_pid_nr(p), p->comm, points);
962 * If any of p's children has a different mm and is eligible for kill,
963 * the one with the highest oom_badness() score is sacrificed for its
964 * parent. This attempts to lose the minimal amount of work done while
965 * still freeing memory.
967 read_lock(&tasklist_lock);
970 * The task 'p' might have already exited before reaching here. The
971 * put_task_struct() will free task_struct 'p' while the loop still try
972 * to access the field of 'p', so, get an extra reference.
975 for_each_thread(p, t) {
976 list_for_each_entry(child, &t->children, sibling) {
977 unsigned int child_points;
979 if (process_shares_mm(child, p->mm))
982 * oom_badness() returns 0 if the thread is unkillable
984 child_points = oom_badness(child,
985 oc->memcg, oc->nodemask, oc->totalpages);
986 if (child_points > victim_points) {
987 put_task_struct(victim);
989 victim_points = child_points;
990 get_task_struct(victim);
995 read_unlock(&tasklist_lock);
998 * Do we need to kill the entire memory cgroup?
999 * Or even one of the ancestor memory cgroups?
1000 * Check this out before killing the victim task.
1002 oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1004 __oom_kill_process(victim);
1007 * If necessary, kill all tasks in the selected memory cgroup.
1010 mem_cgroup_print_oom_group(oom_group);
1011 mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member, NULL);
1012 mem_cgroup_put(oom_group);
1017 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1019 static void check_panic_on_oom(struct oom_control *oc,
1020 enum oom_constraint constraint)
1022 if (likely(!sysctl_panic_on_oom))
1024 if (sysctl_panic_on_oom != 2) {
1026 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1027 * does not panic for cpuset, mempolicy, or memcg allocation
1030 if (constraint != CONSTRAINT_NONE)
1033 /* Do not panic for oom kills triggered by sysrq */
1034 if (is_sysrq_oom(oc))
1036 dump_header(oc, NULL);
1037 panic("Out of memory: %s panic_on_oom is enabled\n",
1038 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1041 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1043 int register_oom_notifier(struct notifier_block *nb)
1045 return blocking_notifier_chain_register(&oom_notify_list, nb);
1047 EXPORT_SYMBOL_GPL(register_oom_notifier);
1049 int unregister_oom_notifier(struct notifier_block *nb)
1051 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1053 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1056 * out_of_memory - kill the "best" process when we run out of memory
1057 * @oc: pointer to struct oom_control
1059 * If we run out of memory, we have the choice between either
1060 * killing a random task (bad), letting the system crash (worse)
1061 * OR try to be smart about which process to kill. Note that we
1062 * don't have to be perfect here, we just have to be good.
1064 bool out_of_memory(struct oom_control *oc)
1066 unsigned long freed = 0;
1067 enum oom_constraint constraint = CONSTRAINT_NONE;
1069 if (oom_killer_disabled)
1072 if (!is_memcg_oom(oc)) {
1073 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1075 /* Got some memory back in the last second. */
1080 * If current has a pending SIGKILL or is exiting, then automatically
1081 * select it. The goal is to allow it to allocate so that it may
1082 * quickly exit and free its memory.
1084 if (task_will_free_mem(current)) {
1085 mark_oom_victim(current);
1086 wake_oom_reaper(current);
1091 * The OOM killer does not compensate for IO-less reclaim.
1092 * pagefault_out_of_memory lost its gfp context so we have to
1093 * make sure exclude 0 mask - all other users should have at least
1094 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1095 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1097 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1101 * Check if there were limitations on the allocation (only relevant for
1102 * NUMA and memcg) that may require different handling.
1104 constraint = constrained_alloc(oc);
1105 if (constraint != CONSTRAINT_MEMORY_POLICY)
1106 oc->nodemask = NULL;
1107 check_panic_on_oom(oc, constraint);
1109 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1110 current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) &&
1111 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1112 get_task_struct(current);
1113 oc->chosen = current;
1114 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1118 select_bad_process(oc);
1119 /* Found nothing?!?! */
1121 dump_header(oc, NULL);
1122 pr_warn("Out of memory and no killable processes...\n");
1124 * If we got here due to an actual allocation at the
1125 * system level, we cannot survive this and will enter
1126 * an endless loop in the allocator. Bail out now.
1128 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1129 panic("System is deadlocked on memory\n");
1131 if (oc->chosen && oc->chosen != (void *)-1UL)
1132 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1133 "Memory cgroup out of memory");
1134 return !!oc->chosen;
1138 * The pagefault handler calls here because some allocation has failed. We have
1139 * to take care of the memcg OOM here because this is the only safe context without
1140 * any locks held but let the oom killer triggered from the allocation context care
1141 * about the global OOM.
1143 void pagefault_out_of_memory(void)
1145 static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1146 DEFAULT_RATELIMIT_BURST);
1148 if (mem_cgroup_oom_synchronize(true))
1151 if (fatal_signal_pending(current))
1154 if (__ratelimit(&pfoom_rs))
1155 pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");