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;
45 static DEFINE_SPINLOCK(zone_scan_lock);
49 * has_intersects_mems_allowed() - check task eligiblity for kill
50 * @start: task struct of which task to consider
51 * @mask: nodemask passed to page allocator for mempolicy ooms
53 * Task eligibility is determined by whether or not a candidate task, @tsk,
54 * shares the same mempolicy nodes as current if it is bound by such a policy
55 * and whether or not it has the same set of allowed cpuset nodes.
57 static bool has_intersects_mems_allowed(struct task_struct *start,
58 const nodemask_t *mask)
60 struct task_struct *tsk;
64 for_each_thread(start, tsk) {
67 * If this is a mempolicy constrained oom, tsk's
68 * cpuset is irrelevant. Only return true if its
69 * mempolicy intersects current, otherwise it may be
72 ret = mempolicy_nodemask_intersects(tsk, mask);
75 * This is not a mempolicy constrained oom, so only
76 * check the mems of tsk's cpuset.
78 ret = cpuset_mems_allowed_intersects(current, tsk);
88 static bool has_intersects_mems_allowed(struct task_struct *tsk,
89 const nodemask_t *mask)
93 #endif /* CONFIG_NUMA */
96 * The process p may have detached its own ->mm while exiting or through
97 * use_mm(), but one or more of its subthreads may still have a valid
98 * pointer. Return p, or any of its subthreads with a valid ->mm, with
101 struct task_struct *find_lock_task_mm(struct task_struct *p)
103 struct task_struct *t;
107 for_each_thread(p, t) {
120 /* return true if the task is not adequate as candidate victim task. */
121 static bool oom_unkillable_task(struct task_struct *p,
122 const struct mem_cgroup *memcg, const nodemask_t *nodemask)
124 if (is_global_init(p))
126 if (p->flags & PF_KTHREAD)
129 /* When mem_cgroup_out_of_memory() and p is not member of the group */
130 if (memcg && !task_in_mem_cgroup(p, memcg))
133 /* p may not have freeable memory in nodemask */
134 if (!has_intersects_mems_allowed(p, nodemask))
141 * oom_badness - heuristic function to determine which candidate task to kill
142 * @p: task struct of which task we should calculate
143 * @totalpages: total present RAM allowed for page allocation
145 * The heuristic for determining which task to kill is made to be as simple and
146 * predictable as possible. The goal is to return the highest value for the
147 * task consuming the most memory to avoid subsequent oom failures.
149 * ps. we don't want to kill the process still in the uninterruptible state
151 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
152 const nodemask_t *nodemask, unsigned long totalpages)
157 if (oom_unkillable_task(p, memcg, nodemask))
160 p = find_lock_task_mm(p);
164 adj = (long)p->signal->oom_score_adj;
165 if (adj == OOM_SCORE_ADJ_MIN) {
171 * The baseline for the badness score is the proportion of RAM that each
172 * task's rss, pagetable and swap space use.
174 points = get_mm_rss(p->mm) + p->mm->nr_ptes +
175 get_mm_counter(p->mm, MM_SWAPENTS);
179 * Root processes get 3% bonus, just like the __vm_enough_memory()
180 * implementation used by LSMs.
182 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
183 points -= (points * 3) / 100;
185 /* Normalize to oom_score_adj units */
186 adj *= totalpages / 1000;
190 * The unterruptible task cannot be killed,
191 * which couldn't receive any signal including SIGKILL.
193 if (p->state == TASK_UNINTERRUPTIBLE)
197 * Never return 0 for an eligible task regardless of the root bonus and
198 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
200 return points > 0 ? points : 1;
204 * Determine the type of allocation constraint.
207 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
208 gfp_t gfp_mask, nodemask_t *nodemask,
209 unsigned long *totalpages)
213 enum zone_type high_zoneidx = gfp_zone(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 (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().
235 if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
236 *totalpages = total_swap_pages;
237 for_each_node_mask(nid, *nodemask)
238 *totalpages += node_spanned_pages(nid);
239 return CONSTRAINT_MEMORY_POLICY;
242 /* Check this allocation failure is caused by cpuset's wall function */
243 for_each_zone_zonelist_nodemask(zone, z, zonelist,
244 high_zoneidx, nodemask)
245 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
246 cpuset_limited = true;
248 if (cpuset_limited) {
249 *totalpages = total_swap_pages;
250 for_each_node_mask(nid, cpuset_current_mems_allowed)
251 *totalpages += node_spanned_pages(nid);
252 return CONSTRAINT_CPUSET;
254 return CONSTRAINT_NONE;
257 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
258 gfp_t gfp_mask, nodemask_t *nodemask,
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 task_struct *task,
267 unsigned long totalpages, const nodemask_t *nodemask,
270 if (task->exit_state)
271 return OOM_SCAN_CONTINUE;
272 if (oom_unkillable_task(task, NULL, nodemask))
273 return OOM_SCAN_CONTINUE;
276 * This task already has access to memory reserves and is being killed.
277 * Don't allow any other task to have access to the reserves.
279 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
280 if (unlikely(frozen(task)))
283 return OOM_SCAN_ABORT;
286 return OOM_SCAN_CONTINUE;
289 * If task is allocating a lot of memory and has been marked to be
290 * killed first if it triggers an oom, then select it.
292 if (oom_task_origin(task))
293 return OOM_SCAN_SELECT;
295 if (task->flags & PF_EXITING && !force_kill) {
297 * If this task is not being ptraced on exit, then wait for it
298 * to finish before killing some other task unnecessarily.
300 if (!(task->group_leader->ptrace & PT_TRACE_EXIT))
301 return OOM_SCAN_ABORT;
307 * Simple selection loop. We chose the process with the highest
308 * number of 'points'.
310 * (not docbooked, we don't want this one cluttering up the manual)
312 static struct task_struct *select_bad_process(unsigned int *ppoints,
313 unsigned long totalpages, const nodemask_t *nodemask,
316 struct task_struct *g, *p;
317 struct task_struct *chosen = NULL;
318 unsigned long chosen_points = 0;
321 for_each_process_thread(g, p) {
324 switch (oom_scan_process_thread(p, totalpages, nodemask,
326 case OOM_SCAN_SELECT:
328 chosen_points = ULONG_MAX;
330 case OOM_SCAN_CONTINUE:
334 return ERR_PTR(-1UL);
338 points = oom_badness(p, NULL, nodemask, totalpages);
339 if (points > chosen_points) {
341 chosen_points = points;
345 get_task_struct(chosen);
348 *ppoints = chosen_points * 1000 / totalpages;
353 * dump_tasks - dump current memory state of all system tasks
354 * @memcg: current's memory controller, if constrained
355 * @nodemask: nodemask passed to page allocator for mempolicy ooms
357 * Dumps the current memory state of all eligible tasks. Tasks not in the same
358 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
360 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
361 * swapents, oom_score_adj value, and name.
363 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
365 struct task_struct *p;
366 struct task_struct *task;
368 pr_info("[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name\n");
370 for_each_process(p) {
371 if (oom_unkillable_task(p, memcg, nodemask))
374 task = find_lock_task_mm(p);
377 * This is a kthread or all of p's threads have already
378 * detached their mm's. There's no need to report
379 * them; they can't be oom killed anyway.
384 pr_info("[%5d] %5d %5d %8lu %8lu %7lu %8lu %5hd %s\n",
385 task->pid, from_kuid(&init_user_ns, task_uid(task)),
386 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
388 get_mm_counter(task->mm, MM_SWAPENTS),
389 task->signal->oom_score_adj, task->comm);
395 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
396 struct mem_cgroup *memcg, const nodemask_t *nodemask)
399 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
400 "oom_score_adj=%hd\n",
401 current->comm, gfp_mask, order,
402 current->signal->oom_score_adj);
403 cpuset_print_task_mems_allowed(current);
404 task_unlock(current);
407 mem_cgroup_print_oom_info(memcg, p);
409 show_mem(SHOW_MEM_FILTER_NODES);
410 if (sysctl_oom_dump_tasks)
411 dump_tasks(memcg, nodemask);
414 static void set_thread_group_flag(struct task_struct *p, int flag)
416 struct task_struct *t;
418 * In some situation, a thread group member holds the mm->mmap_sem
419 * as a writer and then blocks waiting for memory.
420 * After that, the oom-killer chooses this particular thread group
421 * to die to release some memory. since the mm->mmap_sem is held,
422 * the thread group cannot finish the exit routine,
423 * therefore no memory can be released, which leads to a deadlock
425 * Here grants TIF_MEMDIE to all thread group members to avoid the
429 for_each_thread(p, t) {
430 set_tsk_thread_flag(t, TIF_MEMDIE);
436 * Number of OOM killer invocations (including memcg OOM killer).
437 * Primarily used by PM freezer to check for potential races with
438 * OOM killed frozen task.
440 static atomic_t oom_kills = ATOMIC_INIT(0);
442 int oom_kills_count(void)
444 return atomic_read(&oom_kills);
447 void note_oom_kill(void)
449 atomic_inc(&oom_kills);
452 #define K(x) ((x) << (PAGE_SHIFT-10))
454 * Must be called while holding a reference to p, which will be released upon
457 void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
458 unsigned int points, unsigned long totalpages,
459 struct mem_cgroup *memcg, nodemask_t *nodemask,
462 struct task_struct *victim = p;
463 struct task_struct *child;
464 struct task_struct *t;
465 struct mm_struct *mm;
466 unsigned int victim_points = 0;
467 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
468 DEFAULT_RATELIMIT_BURST);
471 * If the task is already exiting, don't alarm the sysadmin or kill
472 * its children or threads, just set TIF_MEMDIE so it can die quickly
474 if (p->flags & PF_EXITING) {
475 set_thread_group_flag(p, TIF_MEMDIE);
480 if (__ratelimit(&oom_rs))
481 dump_header(p, gfp_mask, order, memcg, nodemask);
484 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
485 message, task_pid_nr(p), p->comm, points);
489 * If any of p's children has a different mm and is eligible for kill,
490 * the one with the highest oom_badness() score is sacrificed for its
491 * parent. This attempts to lose the minimal amount of work done while
492 * still freeing memory.
494 read_lock(&tasklist_lock);
495 for_each_thread(p, t) {
496 list_for_each_entry(child, &t->children, sibling) {
497 unsigned int child_points;
499 if (child->mm == p->mm)
502 * oom_badness() returns 0 if the thread is unkillable
504 child_points = oom_badness(child, memcg, nodemask,
506 if (child_points > victim_points) {
507 put_task_struct(victim);
509 victim_points = child_points;
510 get_task_struct(victim);
514 read_unlock(&tasklist_lock);
516 p = find_lock_task_mm(victim);
518 put_task_struct(victim);
520 } else if (victim != p) {
522 put_task_struct(victim);
526 /* mm cannot safely be dereferenced after task_unlock(victim) */
528 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
529 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
530 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
531 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
535 * Kill all user processes sharing victim->mm in other thread groups, if
536 * any. They don't get access to memory reserves, though, to avoid
537 * depletion of all memory. This prevents mm->mmap_sem livelock when an
538 * oom killed thread cannot exit because it requires the semaphore and
539 * its contended by another thread trying to allocate memory itself.
540 * That thread will now get access to memory reserves since it has a
541 * pending fatal signal.
545 if (p->mm == mm && !same_thread_group(p, victim) &&
546 !(p->flags & PF_KTHREAD)) {
547 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
550 task_lock(p); /* Protect ->comm from prctl() */
551 pr_err("Kill process %d (%s) sharing same memory\n",
552 task_pid_nr(p), p->comm);
554 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
558 set_thread_group_flag(victim, TIF_MEMDIE);
559 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
560 put_task_struct(victim);
565 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
567 void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
568 int order, const nodemask_t *nodemask)
570 if (likely(!sysctl_panic_on_oom))
572 if (sysctl_panic_on_oom != 2) {
574 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
575 * does not panic for cpuset, mempolicy, or memcg allocation
578 if (constraint != CONSTRAINT_NONE)
581 dump_header(NULL, gfp_mask, order, NULL, nodemask);
582 panic("Out of memory: %s panic_on_oom is enabled\n",
583 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
586 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
588 int register_oom_notifier(struct notifier_block *nb)
590 return blocking_notifier_chain_register(&oom_notify_list, nb);
592 EXPORT_SYMBOL_GPL(register_oom_notifier);
594 int unregister_oom_notifier(struct notifier_block *nb)
596 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
598 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
601 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
602 * if a parallel OOM killing is already taking place that includes a zone in
603 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
605 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
611 spin_lock(&zone_scan_lock);
612 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
613 if (zone_is_oom_locked(zone)) {
619 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
621 * Lock each zone in the zonelist under zone_scan_lock so a
622 * parallel invocation of try_set_zonelist_oom() doesn't succeed
625 zone_set_flag(zone, ZONE_OOM_LOCKED);
629 spin_unlock(&zone_scan_lock);
634 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
635 * allocation attempts with zonelists containing them may now recall the OOM
636 * killer, if necessary.
638 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
643 spin_lock(&zone_scan_lock);
644 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
645 zone_clear_flag(zone, ZONE_OOM_LOCKED);
647 spin_unlock(&zone_scan_lock);
651 * out_of_memory - kill the "best" process when we run out of memory
652 * @zonelist: zonelist pointer
653 * @gfp_mask: memory allocation flags
654 * @order: amount of memory being requested as a power of 2
655 * @nodemask: nodemask passed to page allocator
656 * @force_kill: true if a task must be killed, even if others are exiting
658 * If we run out of memory, we have the choice between either
659 * killing a random task (bad), letting the system crash (worse)
660 * OR try to be smart about which process to kill. Note that we
661 * don't have to be perfect here, we just have to be good.
663 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
664 int order, nodemask_t *nodemask, bool force_kill)
666 const nodemask_t *mpol_mask;
667 struct task_struct *p;
668 unsigned long totalpages;
669 unsigned long freed = 0;
670 unsigned int uninitialized_var(points);
671 enum oom_constraint constraint = CONSTRAINT_NONE;
674 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
676 /* Got some memory back in the last second. */
680 * If current has a pending SIGKILL or is exiting, then automatically
681 * select it. The goal is to allow it to allocate so that it may
682 * quickly exit and free its memory.
684 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
685 set_thread_group_flag(current, TIF_MEMDIE);
690 * Check if there were limitations on the allocation (only relevant for
691 * NUMA) that may require different handling.
693 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
695 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
696 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
698 if (sysctl_oom_kill_allocating_task && current->mm &&
699 !oom_unkillable_task(current, NULL, nodemask) &&
700 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
701 get_task_struct(current);
702 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
704 "Out of memory (oom_kill_allocating_task)");
708 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
709 /* Found nothing?!?! Either we hang forever, or we panic. */
711 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
712 panic("Out of memory and no killable processes...\n");
714 if (PTR_ERR(p) != -1UL) {
715 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
716 nodemask, "Out of memory");
721 * Give the killed threads a good chance of exiting before trying to
722 * allocate memory again.
725 schedule_timeout_killable(1);
729 * The pagefault handler calls here because it is out of memory, so kill a
730 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
731 * parallel oom killing is already in progress so do nothing.
733 void pagefault_out_of_memory(void)
735 struct zonelist *zonelist;
737 if (mem_cgroup_oom_synchronize(true))
740 zonelist = node_zonelist(first_online_node, GFP_KERNEL);
741 if (try_set_zonelist_oom(zonelist, GFP_KERNEL)) {
742 out_of_memory(NULL, 0, 0, NULL, false);
743 clear_zonelist_oom(zonelist, GFP_KERNEL);