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/swap.h>
26 #include <linux/timex.h>
27 #include <linux/jiffies.h>
28 #include <linux/cpuset.h>
29 #include <linux/export.h>
30 #include <linux/notifier.h>
31 #include <linux/memcontrol.h>
32 #include <linux/mempolicy.h>
33 #include <linux/security.h>
34 #include <linux/ptrace.h>
35 #include <linux/freezer.h>
36 #include <linux/ftrace.h>
37 #include <linux/ratelimit.h>
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/oom.h>
42 int sysctl_panic_on_oom;
43 int sysctl_oom_kill_allocating_task;
44 int sysctl_oom_dump_tasks = 1;
46 DEFINE_MUTEX(oom_lock);
50 * has_intersects_mems_allowed() - check task eligiblity for kill
51 * @start: task struct of which task to consider
52 * @mask: nodemask passed to page allocator for mempolicy ooms
54 * Task eligibility is determined by whether or not a candidate task, @tsk,
55 * shares the same mempolicy nodes as current if it is bound by such a policy
56 * and whether or not it has the same set of allowed cpuset nodes.
58 static bool has_intersects_mems_allowed(struct task_struct *start,
59 const nodemask_t *mask)
61 struct task_struct *tsk;
65 for_each_thread(start, tsk) {
68 * If this is a mempolicy constrained oom, tsk's
69 * cpuset is irrelevant. Only return true if its
70 * mempolicy intersects current, otherwise it may be
73 ret = mempolicy_nodemask_intersects(tsk, mask);
76 * This is not a mempolicy constrained oom, so only
77 * check the mems of tsk's cpuset.
79 ret = cpuset_mems_allowed_intersects(current, tsk);
89 static bool has_intersects_mems_allowed(struct task_struct *tsk,
90 const nodemask_t *mask)
94 #endif /* CONFIG_NUMA */
97 * The process p may have detached its own ->mm while exiting or through
98 * use_mm(), but one or more of its subthreads may still have a valid
99 * pointer. Return p, or any of its subthreads with a valid ->mm, with
102 struct task_struct *find_lock_task_mm(struct task_struct *p)
104 struct task_struct *t;
108 for_each_thread(p, t) {
122 * order == -1 means the oom kill is required by sysrq, otherwise only
123 * for display purposes.
125 static inline bool is_sysrq_oom(struct oom_control *oc)
127 return oc->order == -1;
130 /* return true if the task is not adequate as candidate victim task. */
131 static bool oom_unkillable_task(struct task_struct *p,
132 struct mem_cgroup *memcg, const nodemask_t *nodemask)
134 if (is_global_init(p))
136 if (p->flags & PF_KTHREAD)
139 /* When mem_cgroup_out_of_memory() and p is not member of the group */
140 if (memcg && !task_in_mem_cgroup(p, memcg))
143 /* p may not have freeable memory in nodemask */
144 if (!has_intersects_mems_allowed(p, nodemask))
151 * oom_badness - heuristic function to determine which candidate task to kill
152 * @p: task struct of which task we should calculate
153 * @totalpages: total present RAM allowed for page allocation
155 * The heuristic for determining which task to kill is made to be as simple and
156 * predictable as possible. The goal is to return the highest value for the
157 * task consuming the most memory to avoid subsequent oom failures.
159 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
160 const nodemask_t *nodemask, unsigned long totalpages)
165 if (oom_unkillable_task(p, memcg, nodemask))
168 p = find_lock_task_mm(p);
172 adj = (long)p->signal->oom_score_adj;
173 if (adj == OOM_SCORE_ADJ_MIN) {
179 * The baseline for the badness score is the proportion of RAM that each
180 * task's rss, pagetable and swap space use.
182 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
183 atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
187 * Root processes get 3% bonus, just like the __vm_enough_memory()
188 * implementation used by LSMs.
190 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
191 points -= (points * 3) / 100;
193 /* Normalize to oom_score_adj units */
194 adj *= totalpages / 1000;
198 * Never return 0 for an eligible task regardless of the root bonus and
199 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
201 return points > 0 ? points : 1;
205 * Determine the type of allocation constraint.
208 static enum oom_constraint constrained_alloc(struct oom_control *oc,
209 unsigned long *totalpages)
213 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
214 bool cpuset_limited = false;
217 /* Default to all available memory */
218 *totalpages = totalram_pages + total_swap_pages;
221 return CONSTRAINT_NONE;
223 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
224 * to kill current.We have to random task kill in this case.
225 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
227 if (oc->gfp_mask & __GFP_THISNODE)
228 return CONSTRAINT_NONE;
231 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
232 * the page allocator means a mempolicy is in effect. Cpuset policy
233 * is enforced in get_page_from_freelist().
236 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
237 *totalpages = total_swap_pages;
238 for_each_node_mask(nid, *oc->nodemask)
239 *totalpages += node_spanned_pages(nid);
240 return CONSTRAINT_MEMORY_POLICY;
243 /* Check this allocation failure is caused by cpuset's wall function */
244 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
245 high_zoneidx, oc->nodemask)
246 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
247 cpuset_limited = true;
249 if (cpuset_limited) {
250 *totalpages = total_swap_pages;
251 for_each_node_mask(nid, cpuset_current_mems_allowed)
252 *totalpages += node_spanned_pages(nid);
253 return CONSTRAINT_CPUSET;
255 return CONSTRAINT_NONE;
258 static enum oom_constraint constrained_alloc(struct oom_control *oc,
259 unsigned long *totalpages)
261 *totalpages = totalram_pages + total_swap_pages;
262 return CONSTRAINT_NONE;
266 enum oom_scan_t oom_scan_process_thread(struct oom_control *oc,
267 struct task_struct *task, unsigned long totalpages)
269 if (oom_unkillable_task(task, NULL, oc->nodemask))
270 return OOM_SCAN_CONTINUE;
273 * This task already has access to memory reserves and is being killed.
274 * Don't allow any other task to have access to the reserves.
276 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
277 if (!is_sysrq_oom(oc))
278 return OOM_SCAN_ABORT;
281 return OOM_SCAN_CONTINUE;
284 * If task is allocating a lot of memory and has been marked to be
285 * killed first if it triggers an oom, then select it.
287 if (oom_task_origin(task))
288 return OOM_SCAN_SELECT;
294 * Simple selection loop. We chose the process with the highest
295 * number of 'points'. Returns -1 on scan abort.
297 static struct task_struct *select_bad_process(struct oom_control *oc,
298 unsigned int *ppoints, unsigned long totalpages)
300 struct task_struct *g, *p;
301 struct task_struct *chosen = NULL;
302 unsigned long chosen_points = 0;
305 for_each_process_thread(g, p) {
308 switch (oom_scan_process_thread(oc, p, totalpages)) {
309 case OOM_SCAN_SELECT:
311 chosen_points = ULONG_MAX;
313 case OOM_SCAN_CONTINUE:
317 return (struct task_struct *)(-1UL);
321 points = oom_badness(p, NULL, oc->nodemask, totalpages);
322 if (!points || points < chosen_points)
324 /* Prefer thread group leaders for display purposes */
325 if (points == chosen_points && thread_group_leader(chosen))
329 chosen_points = points;
332 get_task_struct(chosen);
335 *ppoints = chosen_points * 1000 / totalpages;
340 * dump_tasks - dump current memory state of all system tasks
341 * @memcg: current's memory controller, if constrained
342 * @nodemask: nodemask passed to page allocator for mempolicy ooms
344 * Dumps the current memory state of all eligible tasks. Tasks not in the same
345 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
347 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
348 * swapents, oom_score_adj value, and name.
350 static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
352 struct task_struct *p;
353 struct task_struct *task;
355 pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n");
357 for_each_process(p) {
358 if (oom_unkillable_task(p, memcg, nodemask))
361 task = find_lock_task_mm(p);
364 * This is a kthread or all of p's threads have already
365 * detached their mm's. There's no need to report
366 * them; they can't be oom killed anyway.
371 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n",
372 task->pid, from_kuid(&init_user_ns, task_uid(task)),
373 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
374 atomic_long_read(&task->mm->nr_ptes),
375 mm_nr_pmds(task->mm),
376 get_mm_counter(task->mm, MM_SWAPENTS),
377 task->signal->oom_score_adj, task->comm);
383 static void dump_header(struct oom_control *oc, struct task_struct *p,
384 struct mem_cgroup *memcg)
386 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
387 current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
388 current->signal->oom_score_adj);
390 cpuset_print_current_mems_allowed();
393 mem_cgroup_print_oom_info(memcg, p);
395 show_mem(SHOW_MEM_FILTER_NODES);
396 if (sysctl_oom_dump_tasks)
397 dump_tasks(memcg, oc->nodemask);
401 * Number of OOM victims in flight
403 static atomic_t oom_victims = ATOMIC_INIT(0);
404 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
406 bool oom_killer_disabled __read_mostly;
409 * mark_oom_victim - mark the given task as OOM victim
412 * Has to be called with oom_lock held and never after
413 * oom has been disabled already.
415 void mark_oom_victim(struct task_struct *tsk)
417 WARN_ON(oom_killer_disabled);
418 /* OOM killer might race with memcg OOM */
419 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
422 * Make sure that the task is woken up from uninterruptible sleep
423 * if it is frozen because OOM killer wouldn't be able to free
424 * any memory and livelock. freezing_slow_path will tell the freezer
425 * that TIF_MEMDIE tasks should be ignored.
428 atomic_inc(&oom_victims);
432 * exit_oom_victim - note the exit of an OOM victim
434 void exit_oom_victim(void)
436 clear_thread_flag(TIF_MEMDIE);
438 if (!atomic_dec_return(&oom_victims))
439 wake_up_all(&oom_victims_wait);
443 * oom_killer_disable - disable OOM killer
445 * Forces all page allocations to fail rather than trigger OOM killer.
446 * Will block and wait until all OOM victims are killed.
448 * The function cannot be called when there are runnable user tasks because
449 * the userspace would see unexpected allocation failures as a result. Any
450 * new usage of this function should be consulted with MM people.
452 * Returns true if successful and false if the OOM killer cannot be
455 bool oom_killer_disable(void)
458 * Make sure to not race with an ongoing OOM killer. Check that the
459 * current is not killed (possibly due to sharing the victim's memory).
461 if (mutex_lock_killable(&oom_lock))
463 oom_killer_disabled = true;
464 mutex_unlock(&oom_lock);
466 wait_event(oom_victims_wait, !atomic_read(&oom_victims));
472 * oom_killer_enable - enable OOM killer
474 void oom_killer_enable(void)
476 oom_killer_disabled = false;
480 * task->mm can be NULL if the task is the exited group leader. So to
481 * determine whether the task is using a particular mm, we examine all the
482 * task's threads: if one of those is using this mm then this task was also
485 static bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
487 struct task_struct *t;
489 for_each_thread(p, t) {
490 struct mm_struct *t_mm = READ_ONCE(t->mm);
497 #define K(x) ((x) << (PAGE_SHIFT-10))
499 * Must be called while holding a reference to p, which will be released upon
502 void oom_kill_process(struct oom_control *oc, struct task_struct *p,
503 unsigned int points, unsigned long totalpages,
504 struct mem_cgroup *memcg, const char *message)
506 struct task_struct *victim = p;
507 struct task_struct *child;
508 struct task_struct *t;
509 struct mm_struct *mm;
510 unsigned int victim_points = 0;
511 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
512 DEFAULT_RATELIMIT_BURST);
515 * If the task is already exiting, don't alarm the sysadmin or kill
516 * its children or threads, just set TIF_MEMDIE so it can die quickly
519 if (p->mm && task_will_free_mem(p)) {
527 if (__ratelimit(&oom_rs))
528 dump_header(oc, p, memcg);
530 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
531 message, task_pid_nr(p), p->comm, points);
534 * If any of p's children has a different mm and is eligible for kill,
535 * the one with the highest oom_badness() score is sacrificed for its
536 * parent. This attempts to lose the minimal amount of work done while
537 * still freeing memory.
539 read_lock(&tasklist_lock);
540 for_each_thread(p, t) {
541 list_for_each_entry(child, &t->children, sibling) {
542 unsigned int child_points;
544 if (process_shares_mm(child, p->mm))
547 * oom_badness() returns 0 if the thread is unkillable
549 child_points = oom_badness(child, memcg, oc->nodemask,
551 if (child_points > victim_points) {
552 put_task_struct(victim);
554 victim_points = child_points;
555 get_task_struct(victim);
559 read_unlock(&tasklist_lock);
561 p = find_lock_task_mm(victim);
563 put_task_struct(victim);
565 } else if (victim != p) {
567 put_task_struct(victim);
571 /* Get a reference to safely compare mm after task_unlock(victim) */
573 atomic_inc(&mm->mm_count);
575 * We should send SIGKILL before setting TIF_MEMDIE in order to prevent
576 * the OOM victim from depleting the memory reserves from the user
577 * space under its control.
579 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
580 mark_oom_victim(victim);
581 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
582 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
583 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
584 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
585 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
589 * Kill all user processes sharing victim->mm in other thread groups, if
590 * any. They don't get access to memory reserves, though, to avoid
591 * depletion of all memory. This prevents mm->mmap_sem livelock when an
592 * oom killed thread cannot exit because it requires the semaphore and
593 * its contended by another thread trying to allocate memory itself.
594 * That thread will now get access to memory reserves since it has a
595 * pending fatal signal.
598 for_each_process(p) {
599 if (!process_shares_mm(p, mm))
601 if (same_thread_group(p, victim))
603 if (unlikely(p->flags & PF_KTHREAD))
605 if (is_global_init(p))
607 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
610 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
615 put_task_struct(victim);
620 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
622 void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint,
623 struct mem_cgroup *memcg)
625 if (likely(!sysctl_panic_on_oom))
627 if (sysctl_panic_on_oom != 2) {
629 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
630 * does not panic for cpuset, mempolicy, or memcg allocation
633 if (constraint != CONSTRAINT_NONE)
636 /* Do not panic for oom kills triggered by sysrq */
637 if (is_sysrq_oom(oc))
639 dump_header(oc, NULL, memcg);
640 panic("Out of memory: %s panic_on_oom is enabled\n",
641 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
644 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
646 int register_oom_notifier(struct notifier_block *nb)
648 return blocking_notifier_chain_register(&oom_notify_list, nb);
650 EXPORT_SYMBOL_GPL(register_oom_notifier);
652 int unregister_oom_notifier(struct notifier_block *nb)
654 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
656 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
659 * out_of_memory - kill the "best" process when we run out of memory
660 * @oc: pointer to struct oom_control
662 * If we run out of memory, we have the choice between either
663 * killing a random task (bad), letting the system crash (worse)
664 * OR try to be smart about which process to kill. Note that we
665 * don't have to be perfect here, we just have to be good.
667 bool out_of_memory(struct oom_control *oc)
669 struct task_struct *p;
670 unsigned long totalpages;
671 unsigned long freed = 0;
672 unsigned int uninitialized_var(points);
673 enum oom_constraint constraint = CONSTRAINT_NONE;
675 if (oom_killer_disabled)
678 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
680 /* Got some memory back in the last second. */
684 * If current has a pending SIGKILL or is exiting, then automatically
685 * select it. The goal is to allow it to allocate so that it may
686 * quickly exit and free its memory.
688 * But don't select if current has already released its mm and cleared
689 * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur.
692 (fatal_signal_pending(current) || task_will_free_mem(current))) {
693 mark_oom_victim(current);
698 * Check if there were limitations on the allocation (only relevant for
699 * NUMA) that may require different handling.
701 constraint = constrained_alloc(oc, &totalpages);
702 if (constraint != CONSTRAINT_MEMORY_POLICY)
704 check_panic_on_oom(oc, constraint, NULL);
706 if (sysctl_oom_kill_allocating_task && current->mm &&
707 !oom_unkillable_task(current, NULL, oc->nodemask) &&
708 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
709 get_task_struct(current);
710 oom_kill_process(oc, current, 0, totalpages, NULL,
711 "Out of memory (oom_kill_allocating_task)");
715 p = select_bad_process(oc, &points, totalpages);
716 /* Found nothing?!?! Either we hang forever, or we panic. */
717 if (!p && !is_sysrq_oom(oc)) {
718 dump_header(oc, NULL, NULL);
719 panic("Out of memory and no killable processes...\n");
721 if (p && p != (void *)-1UL) {
722 oom_kill_process(oc, p, points, totalpages, NULL,
725 * Give the killed process a good chance to exit before trying
726 * to allocate memory again.
728 schedule_timeout_killable(1);
734 * The pagefault handler calls here because it is out of memory, so kill a
735 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
736 * parallel oom killing is already in progress so do nothing.
738 void pagefault_out_of_memory(void)
740 struct oom_control oc = {
747 if (mem_cgroup_oom_synchronize(true))
750 if (!mutex_trylock(&oom_lock))
753 if (!out_of_memory(&oc)) {
755 * There shouldn't be any user tasks runnable while the
756 * OOM killer is disabled, so the current task has to
757 * be a racing OOM victim for which oom_killer_disable()
760 WARN_ON(test_thread_flag(TIF_MEMDIE));
763 mutex_unlock(&oom_lock);