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 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
150 const nodemask_t *nodemask, unsigned long totalpages)
155 if (oom_unkillable_task(p, memcg, nodemask))
158 p = find_lock_task_mm(p);
162 adj = (long)p->signal->oom_score_adj;
163 if (adj == OOM_SCORE_ADJ_MIN) {
169 * The baseline for the badness score is the proportion of RAM that each
170 * task's rss, pagetable and swap space use.
172 points = get_mm_rss(p->mm) + atomic_long_read(&p->mm->nr_ptes) +
173 get_mm_counter(p->mm, MM_SWAPENTS);
177 * Root processes get 3% bonus, just like the __vm_enough_memory()
178 * implementation used by LSMs.
180 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
183 /* Normalize to oom_score_adj units */
184 adj *= totalpages / 1000;
188 * Never return 0 for an eligible task regardless of the root bonus and
189 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
191 return points > 0 ? points : 1;
195 * Determine the type of allocation constraint.
198 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
199 gfp_t gfp_mask, nodemask_t *nodemask,
200 unsigned long *totalpages)
204 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
205 bool cpuset_limited = false;
208 /* Default to all available memory */
209 *totalpages = totalram_pages + total_swap_pages;
212 return CONSTRAINT_NONE;
214 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
215 * to kill current.We have to random task kill in this case.
216 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
218 if (gfp_mask & __GFP_THISNODE)
219 return CONSTRAINT_NONE;
222 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
223 * the page allocator means a mempolicy is in effect. Cpuset policy
224 * is enforced in get_page_from_freelist().
226 if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
227 *totalpages = total_swap_pages;
228 for_each_node_mask(nid, *nodemask)
229 *totalpages += node_spanned_pages(nid);
230 return CONSTRAINT_MEMORY_POLICY;
233 /* Check this allocation failure is caused by cpuset's wall function */
234 for_each_zone_zonelist_nodemask(zone, z, zonelist,
235 high_zoneidx, nodemask)
236 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
237 cpuset_limited = true;
239 if (cpuset_limited) {
240 *totalpages = total_swap_pages;
241 for_each_node_mask(nid, cpuset_current_mems_allowed)
242 *totalpages += node_spanned_pages(nid);
243 return CONSTRAINT_CPUSET;
245 return CONSTRAINT_NONE;
248 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
249 gfp_t gfp_mask, nodemask_t *nodemask,
250 unsigned long *totalpages)
252 *totalpages = totalram_pages + total_swap_pages;
253 return CONSTRAINT_NONE;
257 enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
258 unsigned long totalpages, const nodemask_t *nodemask,
261 if (task->exit_state)
262 return OOM_SCAN_CONTINUE;
263 if (oom_unkillable_task(task, NULL, nodemask))
264 return OOM_SCAN_CONTINUE;
267 * This task already has access to memory reserves and is being killed.
268 * Don't allow any other task to have access to the reserves.
270 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
271 if (unlikely(frozen(task)))
274 return OOM_SCAN_ABORT;
277 return OOM_SCAN_CONTINUE;
280 * If task is allocating a lot of memory and has been marked to be
281 * killed first if it triggers an oom, then select it.
283 if (oom_task_origin(task))
284 return OOM_SCAN_SELECT;
286 if (task->flags & PF_EXITING && !force_kill) {
288 * If this task is not being ptraced on exit, then wait for it
289 * to finish before killing some other task unnecessarily.
291 if (!(task->group_leader->ptrace & PT_TRACE_EXIT))
292 return OOM_SCAN_ABORT;
298 * Simple selection loop. We chose the process with the highest
299 * number of 'points'. Returns -1 on scan abort.
301 * (not docbooked, we don't want this one cluttering up the manual)
303 static struct task_struct *select_bad_process(unsigned int *ppoints,
304 unsigned long totalpages, const nodemask_t *nodemask,
307 struct task_struct *g, *p;
308 struct task_struct *chosen = NULL;
309 unsigned long chosen_points = 0;
312 for_each_process_thread(g, p) {
315 switch (oom_scan_process_thread(p, totalpages, nodemask,
317 case OOM_SCAN_SELECT:
319 chosen_points = ULONG_MAX;
321 case OOM_SCAN_CONTINUE:
325 return (struct task_struct *)(-1UL);
329 points = oom_badness(p, NULL, nodemask, totalpages);
330 if (!points || points < chosen_points)
332 /* Prefer thread group leaders for display purposes */
333 if (points == chosen_points && thread_group_leader(chosen))
337 chosen_points = points;
340 get_task_struct(chosen);
343 *ppoints = chosen_points * 1000 / totalpages;
348 * dump_tasks - dump current memory state of all system tasks
349 * @memcg: current's memory controller, if constrained
350 * @nodemask: nodemask passed to page allocator for mempolicy ooms
352 * Dumps the current memory state of all eligible tasks. Tasks not in the same
353 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
355 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
356 * swapents, oom_score_adj value, and name.
358 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
360 struct task_struct *p;
361 struct task_struct *task;
363 pr_info("[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name\n");
365 for_each_process(p) {
366 if (oom_unkillable_task(p, memcg, nodemask))
369 task = find_lock_task_mm(p);
372 * This is a kthread or all of p's threads have already
373 * detached their mm's. There's no need to report
374 * them; they can't be oom killed anyway.
379 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %8lu %5hd %s\n",
380 task->pid, from_kuid(&init_user_ns, task_uid(task)),
381 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
382 atomic_long_read(&task->mm->nr_ptes),
383 get_mm_counter(task->mm, MM_SWAPENTS),
384 task->signal->oom_score_adj, task->comm);
390 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
391 struct mem_cgroup *memcg, const nodemask_t *nodemask)
394 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
395 "oom_score_adj=%hd\n",
396 current->comm, gfp_mask, order,
397 current->signal->oom_score_adj);
398 cpuset_print_task_mems_allowed(current);
399 task_unlock(current);
402 mem_cgroup_print_oom_info(memcg, p);
404 show_mem(SHOW_MEM_FILTER_NODES);
405 if (sysctl_oom_dump_tasks)
406 dump_tasks(memcg, nodemask);
409 #define K(x) ((x) << (PAGE_SHIFT-10))
411 * Must be called while holding a reference to p, which will be released upon
414 void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
415 unsigned int points, unsigned long totalpages,
416 struct mem_cgroup *memcg, nodemask_t *nodemask,
419 struct task_struct *victim = p;
420 struct task_struct *child;
421 struct task_struct *t;
422 struct mm_struct *mm;
423 unsigned int victim_points = 0;
424 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
425 DEFAULT_RATELIMIT_BURST);
428 * If the task is already exiting, don't alarm the sysadmin or kill
429 * its children or threads, just set TIF_MEMDIE so it can die quickly
431 if (p->flags & PF_EXITING) {
432 set_tsk_thread_flag(p, TIF_MEMDIE);
437 if (__ratelimit(&oom_rs))
438 dump_header(p, gfp_mask, order, memcg, nodemask);
441 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
442 message, task_pid_nr(p), p->comm, points);
446 * If any of p's children has a different mm and is eligible for kill,
447 * the one with the highest oom_badness() score is sacrificed for its
448 * parent. This attempts to lose the minimal amount of work done while
449 * still freeing memory.
451 read_lock(&tasklist_lock);
452 for_each_thread(p, t) {
453 list_for_each_entry(child, &t->children, sibling) {
454 unsigned int child_points;
456 if (child->mm == p->mm)
459 * oom_badness() returns 0 if the thread is unkillable
461 child_points = oom_badness(child, memcg, nodemask,
463 if (child_points > victim_points) {
464 put_task_struct(victim);
466 victim_points = child_points;
467 get_task_struct(victim);
471 read_unlock(&tasklist_lock);
473 p = find_lock_task_mm(victim);
475 put_task_struct(victim);
477 } else if (victim != p) {
479 put_task_struct(victim);
483 /* mm cannot safely be dereferenced after task_unlock(victim) */
485 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
486 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
487 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
488 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
492 * Kill all user processes sharing victim->mm in other thread groups, if
493 * any. They don't get access to memory reserves, though, to avoid
494 * depletion of all memory. This prevents mm->mmap_sem livelock when an
495 * oom killed thread cannot exit because it requires the semaphore and
496 * its contended by another thread trying to allocate memory itself.
497 * That thread will now get access to memory reserves since it has a
498 * pending fatal signal.
502 if (p->mm == mm && !same_thread_group(p, victim) &&
503 !(p->flags & PF_KTHREAD)) {
504 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
507 task_lock(p); /* Protect ->comm from prctl() */
508 pr_err("Kill process %d (%s) sharing same memory\n",
509 task_pid_nr(p), p->comm);
511 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
515 set_tsk_thread_flag(victim, TIF_MEMDIE);
516 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
517 put_task_struct(victim);
522 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
524 void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
525 int order, const nodemask_t *nodemask)
527 if (likely(!sysctl_panic_on_oom))
529 if (sysctl_panic_on_oom != 2) {
531 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
532 * does not panic for cpuset, mempolicy, or memcg allocation
535 if (constraint != CONSTRAINT_NONE)
538 dump_header(NULL, gfp_mask, order, NULL, nodemask);
539 panic("Out of memory: %s panic_on_oom is enabled\n",
540 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
543 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
545 int register_oom_notifier(struct notifier_block *nb)
547 return blocking_notifier_chain_register(&oom_notify_list, nb);
549 EXPORT_SYMBOL_GPL(register_oom_notifier);
551 int unregister_oom_notifier(struct notifier_block *nb)
553 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
555 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
558 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
559 * if a parallel OOM killing is already taking place that includes a zone in
560 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
562 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
568 spin_lock(&zone_scan_lock);
569 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
570 if (zone_is_oom_locked(zone)) {
576 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
578 * Lock each zone in the zonelist under zone_scan_lock so a
579 * parallel invocation of try_set_zonelist_oom() doesn't succeed
582 zone_set_flag(zone, ZONE_OOM_LOCKED);
586 spin_unlock(&zone_scan_lock);
591 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
592 * allocation attempts with zonelists containing them may now recall the OOM
593 * killer, if necessary.
595 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
600 spin_lock(&zone_scan_lock);
601 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
602 zone_clear_flag(zone, ZONE_OOM_LOCKED);
604 spin_unlock(&zone_scan_lock);
608 * out_of_memory - kill the "best" process when we run out of memory
609 * @zonelist: zonelist pointer
610 * @gfp_mask: memory allocation flags
611 * @order: amount of memory being requested as a power of 2
612 * @nodemask: nodemask passed to page allocator
613 * @force_kill: true if a task must be killed, even if others are exiting
615 * If we run out of memory, we have the choice between either
616 * killing a random task (bad), letting the system crash (worse)
617 * OR try to be smart about which process to kill. Note that we
618 * don't have to be perfect here, we just have to be good.
620 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
621 int order, nodemask_t *nodemask, bool force_kill)
623 const nodemask_t *mpol_mask;
624 struct task_struct *p;
625 unsigned long totalpages;
626 unsigned long freed = 0;
627 unsigned int uninitialized_var(points);
628 enum oom_constraint constraint = CONSTRAINT_NONE;
631 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
633 /* Got some memory back in the last second. */
637 * If current has a pending SIGKILL or is exiting, then automatically
638 * select it. The goal is to allow it to allocate so that it may
639 * quickly exit and free its memory.
641 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
642 set_thread_flag(TIF_MEMDIE);
647 * Check if there were limitations on the allocation (only relevant for
648 * NUMA) that may require different handling.
650 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
652 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
653 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
655 if (sysctl_oom_kill_allocating_task && current->mm &&
656 !oom_unkillable_task(current, NULL, nodemask) &&
657 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
658 get_task_struct(current);
659 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
661 "Out of memory (oom_kill_allocating_task)");
665 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
666 /* Found nothing?!?! Either we hang forever, or we panic. */
668 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
669 panic("Out of memory and no killable processes...\n");
671 if (p != (void *)-1UL) {
672 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
673 nodemask, "Out of memory");
678 * Give the killed threads a good chance of exiting before trying to
679 * allocate memory again.
682 schedule_timeout_killable(1);
686 * The pagefault handler calls here because it is out of memory, so kill a
687 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
688 * parallel oom killing is already in progress so do nothing.
690 void pagefault_out_of_memory(void)
692 struct zonelist *zonelist;
694 if (mem_cgroup_oom_synchronize(true))
697 zonelist = node_zonelist(first_online_node, GFP_KERNEL);
698 if (try_set_zonelist_oom(zonelist, GFP_KERNEL)) {
699 out_of_memory(NULL, 0, 0, NULL, false);
700 clear_zonelist_oom(zonelist, GFP_KERNEL);