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) {
121 /* return true if the task is not adequate as candidate victim task. */
122 static bool oom_unkillable_task(struct task_struct *p,
123 struct mem_cgroup *memcg, const nodemask_t *nodemask)
125 if (is_global_init(p))
127 if (p->flags & PF_KTHREAD)
130 /* When mem_cgroup_out_of_memory() and p is not member of the group */
131 if (memcg && !task_in_mem_cgroup(p, memcg))
134 /* p may not have freeable memory in nodemask */
135 if (!has_intersects_mems_allowed(p, nodemask))
142 * oom_badness - heuristic function to determine which candidate task to kill
143 * @p: task struct of which task we should calculate
144 * @totalpages: total present RAM allowed for page allocation
146 * The heuristic for determining which task to kill is made to be as simple and
147 * predictable as possible. The goal is to return the highest value for the
148 * task consuming the most memory to avoid subsequent oom failures.
150 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
151 const nodemask_t *nodemask, unsigned long totalpages)
156 if (oom_unkillable_task(p, memcg, nodemask))
159 p = find_lock_task_mm(p);
163 adj = (long)p->signal->oom_score_adj;
164 if (adj == OOM_SCORE_ADJ_MIN) {
170 * The baseline for the badness score is the proportion of RAM that each
171 * task's rss, pagetable and swap space use.
173 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
174 atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
178 * Root processes get 3% bonus, just like the __vm_enough_memory()
179 * implementation used by LSMs.
181 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
182 points -= (points * 3) / 100;
184 /* Normalize to oom_score_adj units */
185 adj *= totalpages / 1000;
189 * Never return 0 for an eligible task regardless of the root bonus and
190 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
192 return points > 0 ? points : 1;
196 * Determine the type of allocation constraint.
199 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
200 gfp_t gfp_mask, nodemask_t *nodemask,
201 unsigned long *totalpages)
205 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
206 bool cpuset_limited = false;
209 /* Default to all available memory */
210 *totalpages = totalram_pages + total_swap_pages;
213 return CONSTRAINT_NONE;
215 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
216 * to kill current.We have to random task kill in this case.
217 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
219 if (gfp_mask & __GFP_THISNODE)
220 return CONSTRAINT_NONE;
223 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
224 * the page allocator means a mempolicy is in effect. Cpuset policy
225 * is enforced in get_page_from_freelist().
227 if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
228 *totalpages = total_swap_pages;
229 for_each_node_mask(nid, *nodemask)
230 *totalpages += node_spanned_pages(nid);
231 return CONSTRAINT_MEMORY_POLICY;
234 /* Check this allocation failure is caused by cpuset's wall function */
235 for_each_zone_zonelist_nodemask(zone, z, zonelist,
236 high_zoneidx, nodemask)
237 if (!cpuset_zone_allowed(zone, gfp_mask))
238 cpuset_limited = true;
240 if (cpuset_limited) {
241 *totalpages = total_swap_pages;
242 for_each_node_mask(nid, cpuset_current_mems_allowed)
243 *totalpages += node_spanned_pages(nid);
244 return CONSTRAINT_CPUSET;
246 return CONSTRAINT_NONE;
249 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
250 gfp_t gfp_mask, nodemask_t *nodemask,
251 unsigned long *totalpages)
253 *totalpages = totalram_pages + total_swap_pages;
254 return CONSTRAINT_NONE;
258 enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
259 unsigned long totalpages, const nodemask_t *nodemask,
262 if (oom_unkillable_task(task, NULL, nodemask))
263 return OOM_SCAN_CONTINUE;
266 * This task already has access to memory reserves and is being killed.
267 * Don't allow any other task to have access to the reserves.
269 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
271 return OOM_SCAN_ABORT;
274 return OOM_SCAN_CONTINUE;
277 * If task is allocating a lot of memory and has been marked to be
278 * killed first if it triggers an oom, then select it.
280 if (oom_task_origin(task))
281 return OOM_SCAN_SELECT;
283 if (task_will_free_mem(task) && !force_kill)
284 return OOM_SCAN_ABORT;
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. Returns -1 on scan abort.
293 * (not docbooked, we don't want this one cluttering up the manual)
295 static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, const nodemask_t *nodemask,
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
301 unsigned long chosen_points = 0;
304 for_each_process_thread(g, p) {
307 switch (oom_scan_process_thread(p, totalpages, nodemask,
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, 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 task_struct *p, gfp_t gfp_mask, int order,
384 struct mem_cgroup *memcg, const nodemask_t *nodemask)
387 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
388 "oom_score_adj=%hd\n",
389 current->comm, gfp_mask, order,
390 current->signal->oom_score_adj);
391 cpuset_print_task_mems_allowed(current);
392 task_unlock(current);
395 mem_cgroup_print_oom_info(memcg, p);
397 show_mem(SHOW_MEM_FILTER_NODES);
398 if (sysctl_oom_dump_tasks)
399 dump_tasks(memcg, nodemask);
403 * Number of OOM victims in flight
405 static atomic_t oom_victims = ATOMIC_INIT(0);
406 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
408 bool oom_killer_disabled __read_mostly;
411 * mark_oom_victim - mark the given task as OOM victim
414 * Has to be called with oom_lock held and never after
415 * oom has been disabled already.
417 void mark_oom_victim(struct task_struct *tsk)
419 WARN_ON(oom_killer_disabled);
420 /* OOM killer might race with memcg OOM */
421 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
424 * Make sure that the task is woken up from uninterruptible sleep
425 * if it is frozen because OOM killer wouldn't be able to free
426 * any memory and livelock. freezing_slow_path will tell the freezer
427 * that TIF_MEMDIE tasks should be ignored.
430 atomic_inc(&oom_victims);
434 * exit_oom_victim - note the exit of an OOM victim
436 void exit_oom_victim(void)
438 clear_thread_flag(TIF_MEMDIE);
440 if (!atomic_dec_return(&oom_victims))
441 wake_up_all(&oom_victims_wait);
445 * oom_killer_disable - disable OOM killer
447 * Forces all page allocations to fail rather than trigger OOM killer.
448 * Will block and wait until all OOM victims are killed.
450 * The function cannot be called when there are runnable user tasks because
451 * the userspace would see unexpected allocation failures as a result. Any
452 * new usage of this function should be consulted with MM people.
454 * Returns true if successful and false if the OOM killer cannot be
457 bool oom_killer_disable(void)
460 * Make sure to not race with an ongoing OOM killer
461 * and that the current is not the victim.
463 mutex_lock(&oom_lock);
464 if (test_thread_flag(TIF_MEMDIE)) {
465 mutex_unlock(&oom_lock);
469 oom_killer_disabled = true;
470 mutex_unlock(&oom_lock);
472 wait_event(oom_victims_wait, !atomic_read(&oom_victims));
478 * oom_killer_enable - enable OOM killer
480 void oom_killer_enable(void)
482 oom_killer_disabled = false;
485 #define K(x) ((x) << (PAGE_SHIFT-10))
487 * Must be called while holding a reference to p, which will be released upon
490 void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
491 unsigned int points, unsigned long totalpages,
492 struct mem_cgroup *memcg, nodemask_t *nodemask,
495 struct task_struct *victim = p;
496 struct task_struct *child;
497 struct task_struct *t;
498 struct mm_struct *mm;
499 unsigned int victim_points = 0;
500 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
501 DEFAULT_RATELIMIT_BURST);
504 * If the task is already exiting, don't alarm the sysadmin or kill
505 * its children or threads, just set TIF_MEMDIE so it can die quickly
508 if (p->mm && task_will_free_mem(p)) {
516 if (__ratelimit(&oom_rs))
517 dump_header(p, gfp_mask, order, memcg, nodemask);
520 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
521 message, task_pid_nr(p), p->comm, points);
525 * If any of p's children has a different mm and is eligible for kill,
526 * the one with the highest oom_badness() score is sacrificed for its
527 * parent. This attempts to lose the minimal amount of work done while
528 * still freeing memory.
530 read_lock(&tasklist_lock);
531 for_each_thread(p, t) {
532 list_for_each_entry(child, &t->children, sibling) {
533 unsigned int child_points;
535 if (child->mm == p->mm)
538 * oom_badness() returns 0 if the thread is unkillable
540 child_points = oom_badness(child, memcg, nodemask,
542 if (child_points > victim_points) {
543 put_task_struct(victim);
545 victim_points = child_points;
546 get_task_struct(victim);
550 read_unlock(&tasklist_lock);
552 p = find_lock_task_mm(victim);
554 put_task_struct(victim);
556 } else if (victim != p) {
558 put_task_struct(victim);
562 /* mm cannot safely be dereferenced after task_unlock(victim) */
564 mark_oom_victim(victim);
565 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
566 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
567 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
568 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
572 * Kill all user processes sharing victim->mm in other thread groups, if
573 * any. They don't get access to memory reserves, though, to avoid
574 * depletion of all memory. This prevents mm->mmap_sem livelock when an
575 * oom killed thread cannot exit because it requires the semaphore and
576 * its contended by another thread trying to allocate memory itself.
577 * That thread will now get access to memory reserves since it has a
578 * pending fatal signal.
582 if (p->mm == mm && !same_thread_group(p, victim) &&
583 !(p->flags & PF_KTHREAD)) {
584 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
587 task_lock(p); /* Protect ->comm from prctl() */
588 pr_err("Kill process %d (%s) sharing same memory\n",
589 task_pid_nr(p), p->comm);
591 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
595 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
596 put_task_struct(victim);
601 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
603 void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
604 int order, const nodemask_t *nodemask,
605 struct mem_cgroup *memcg)
607 if (likely(!sysctl_panic_on_oom))
609 if (sysctl_panic_on_oom != 2) {
611 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
612 * does not panic for cpuset, mempolicy, or memcg allocation
615 if (constraint != CONSTRAINT_NONE)
618 dump_header(NULL, gfp_mask, order, memcg, nodemask);
619 panic("Out of memory: %s panic_on_oom is enabled\n",
620 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
623 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
625 int register_oom_notifier(struct notifier_block *nb)
627 return blocking_notifier_chain_register(&oom_notify_list, nb);
629 EXPORT_SYMBOL_GPL(register_oom_notifier);
631 int unregister_oom_notifier(struct notifier_block *nb)
633 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
635 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
638 * __out_of_memory - kill the "best" process when we run out of memory
639 * @zonelist: zonelist pointer
640 * @gfp_mask: memory allocation flags
641 * @order: amount of memory being requested as a power of 2
642 * @nodemask: nodemask passed to page allocator
643 * @force_kill: true if a task must be killed, even if others are exiting
645 * If we run out of memory, we have the choice between either
646 * killing a random task (bad), letting the system crash (worse)
647 * OR try to be smart about which process to kill. Note that we
648 * don't have to be perfect here, we just have to be good.
650 bool out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
651 int order, nodemask_t *nodemask, bool force_kill)
653 const nodemask_t *mpol_mask;
654 struct task_struct *p;
655 unsigned long totalpages;
656 unsigned long freed = 0;
657 unsigned int uninitialized_var(points);
658 enum oom_constraint constraint = CONSTRAINT_NONE;
661 if (oom_killer_disabled)
664 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
666 /* Got some memory back in the last second. */
670 * If current has a pending SIGKILL or is exiting, then automatically
671 * select it. The goal is to allow it to allocate so that it may
672 * quickly exit and free its memory.
674 * But don't select if current has already released its mm and cleared
675 * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur.
678 (fatal_signal_pending(current) || task_will_free_mem(current))) {
679 mark_oom_victim(current);
684 * Check if there were limitations on the allocation (only relevant for
685 * NUMA) that may require different handling.
687 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
689 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
690 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask, NULL);
692 if (sysctl_oom_kill_allocating_task && current->mm &&
693 !oom_unkillable_task(current, NULL, nodemask) &&
694 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
695 get_task_struct(current);
696 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
698 "Out of memory (oom_kill_allocating_task)");
702 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
703 /* Found nothing?!?! Either we hang forever, or we panic. */
705 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
706 panic("Out of memory and no killable processes...\n");
708 if (p != (void *)-1UL) {
709 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
710 nodemask, "Out of memory");
715 * Give the killed threads a good chance of exiting before trying to
716 * allocate memory again.
719 schedule_timeout_killable(1);
725 * The pagefault handler calls here because it is out of memory, so kill a
726 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
727 * parallel oom killing is already in progress so do nothing.
729 void pagefault_out_of_memory(void)
731 if (mem_cgroup_oom_synchronize(true))
734 if (!mutex_trylock(&oom_lock))
737 if (!out_of_memory(NULL, 0, 0, NULL, false)) {
739 * There shouldn't be any user tasks runnable while the
740 * OOM killer is disabled, so the current task has to
741 * be a racing OOM victim for which oom_killer_disable()
744 WARN_ON(test_thread_flag(TIF_MEMDIE));
747 mutex_unlock(&oom_lock);