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);
48 * compare_swap_oom_score_adj() - compare and swap current's oom_score_adj
49 * @old_val: old oom_score_adj for compare
50 * @new_val: new oom_score_adj for swap
52 * Sets the oom_score_adj value for current to @new_val iff its present value is
53 * @old_val. Usually used to reinstate a previous value to prevent racing with
54 * userspacing tuning the value in the interim.
56 void compare_swap_oom_score_adj(int old_val, int new_val)
58 struct sighand_struct *sighand = current->sighand;
60 spin_lock_irq(&sighand->siglock);
61 if (current->signal->oom_score_adj == old_val)
62 current->signal->oom_score_adj = new_val;
63 trace_oom_score_adj_update(current);
64 spin_unlock_irq(&sighand->siglock);
68 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
69 * @new_val: new oom_score_adj value
71 * Sets the oom_score_adj value for current to @new_val with proper
72 * synchronization and returns the old value. Usually used to temporarily
73 * set a value, save the old value in the caller, and then reinstate it later.
75 int test_set_oom_score_adj(int new_val)
77 struct sighand_struct *sighand = current->sighand;
80 spin_lock_irq(&sighand->siglock);
81 old_val = current->signal->oom_score_adj;
82 current->signal->oom_score_adj = new_val;
83 trace_oom_score_adj_update(current);
84 spin_unlock_irq(&sighand->siglock);
91 * has_intersects_mems_allowed() - check task eligiblity for kill
92 * @tsk: task struct of which task to consider
93 * @mask: nodemask passed to page allocator for mempolicy ooms
95 * Task eligibility is determined by whether or not a candidate task, @tsk,
96 * shares the same mempolicy nodes as current if it is bound by such a policy
97 * and whether or not it has the same set of allowed cpuset nodes.
99 static bool has_intersects_mems_allowed(struct task_struct *tsk,
100 const nodemask_t *mask)
102 struct task_struct *start = tsk;
107 * If this is a mempolicy constrained oom, tsk's
108 * cpuset is irrelevant. Only return true if its
109 * mempolicy intersects current, otherwise it may be
112 if (mempolicy_nodemask_intersects(tsk, mask))
116 * This is not a mempolicy constrained oom, so only
117 * check the mems of tsk's cpuset.
119 if (cpuset_mems_allowed_intersects(current, tsk))
122 } while_each_thread(start, tsk);
127 static bool has_intersects_mems_allowed(struct task_struct *tsk,
128 const nodemask_t *mask)
132 #endif /* CONFIG_NUMA */
135 * The process p may have detached its own ->mm while exiting or through
136 * use_mm(), but one or more of its subthreads may still have a valid
137 * pointer. Return p, or any of its subthreads with a valid ->mm, with
140 struct task_struct *find_lock_task_mm(struct task_struct *p)
142 struct task_struct *t = p;
149 } while_each_thread(p, t);
154 /* return true if the task is not adequate as candidate victim task. */
155 static bool oom_unkillable_task(struct task_struct *p,
156 const struct mem_cgroup *memcg, const nodemask_t *nodemask)
158 if (is_global_init(p))
160 if (p->flags & PF_KTHREAD)
163 /* When mem_cgroup_out_of_memory() and p is not member of the group */
164 if (memcg && !task_in_mem_cgroup(p, memcg))
167 /* p may not have freeable memory in nodemask */
168 if (!has_intersects_mems_allowed(p, nodemask))
175 * oom_badness - heuristic function to determine which candidate task to kill
176 * @p: task struct of which task we should calculate
177 * @totalpages: total present RAM allowed for page allocation
179 * The heuristic for determining which task to kill is made to be as simple and
180 * predictable as possible. The goal is to return the highest value for the
181 * task consuming the most memory to avoid subsequent oom failures.
183 unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
184 const nodemask_t *nodemask, unsigned long totalpages)
188 if (oom_unkillable_task(p, memcg, nodemask))
191 p = find_lock_task_mm(p);
195 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
201 * The memory controller may have a limit of 0 bytes, so avoid a divide
202 * by zero, if necessary.
208 * The baseline for the badness score is the proportion of RAM that each
209 * task's rss, pagetable and swap space use.
211 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
212 points += get_mm_counter(p->mm, MM_SWAPENTS);
215 points /= totalpages;
219 * Root processes get 3% bonus, just like the __vm_enough_memory()
220 * implementation used by LSMs.
222 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
226 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
227 * either completely disable oom killing or always prefer a certain
230 points += p->signal->oom_score_adj;
233 * Never return 0 for an eligible task that may be killed since it's
234 * possible that no single user task uses more than 0.1% of memory and
235 * no single admin tasks uses more than 3.0%.
239 return (points < 1000) ? points : 1000;
243 * Determine the type of allocation constraint.
246 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
247 gfp_t gfp_mask, nodemask_t *nodemask,
248 unsigned long *totalpages)
252 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
253 bool cpuset_limited = false;
256 /* Default to all available memory */
257 *totalpages = totalram_pages + total_swap_pages;
260 return CONSTRAINT_NONE;
262 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
263 * to kill current.We have to random task kill in this case.
264 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
266 if (gfp_mask & __GFP_THISNODE)
267 return CONSTRAINT_NONE;
270 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
271 * the page allocator means a mempolicy is in effect. Cpuset policy
272 * is enforced in get_page_from_freelist().
274 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
275 *totalpages = total_swap_pages;
276 for_each_node_mask(nid, *nodemask)
277 *totalpages += node_spanned_pages(nid);
278 return CONSTRAINT_MEMORY_POLICY;
281 /* Check this allocation failure is caused by cpuset's wall function */
282 for_each_zone_zonelist_nodemask(zone, z, zonelist,
283 high_zoneidx, nodemask)
284 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
285 cpuset_limited = true;
287 if (cpuset_limited) {
288 *totalpages = total_swap_pages;
289 for_each_node_mask(nid, cpuset_current_mems_allowed)
290 *totalpages += node_spanned_pages(nid);
291 return CONSTRAINT_CPUSET;
293 return CONSTRAINT_NONE;
296 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
297 gfp_t gfp_mask, nodemask_t *nodemask,
298 unsigned long *totalpages)
300 *totalpages = totalram_pages + total_swap_pages;
301 return CONSTRAINT_NONE;
306 * Simple selection loop. We chose the process with the highest
307 * number of 'points'. We expect the caller will lock the tasklist.
309 * (not docbooked, we don't want this one cluttering up the manual)
311 static struct task_struct *select_bad_process(unsigned int *ppoints,
312 unsigned long totalpages, struct mem_cgroup *memcg,
313 const nodemask_t *nodemask, bool force_kill)
315 struct task_struct *g, *p;
316 struct task_struct *chosen = NULL;
319 do_each_thread(g, p) {
324 if (oom_unkillable_task(p, memcg, nodemask))
328 * This task already has access to memory reserves and is
329 * being killed. Don't allow any other task access to the
332 * Note: this may have a chance of deadlock if it gets
333 * blocked waiting for another task which itself is waiting
334 * for memory. Is there a better alternative?
336 if (test_tsk_thread_flag(p, TIF_MEMDIE)) {
337 if (unlikely(frozen(p)))
340 return ERR_PTR(-1UL);
345 if (p->flags & PF_EXITING && !force_kill) {
347 * If this task is not being ptraced on exit,
348 * then wait for it to finish before killing
349 * some other task unnecessarily.
351 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
352 return ERR_PTR(-1UL);
355 points = oom_badness(p, memcg, nodemask, totalpages);
356 if (points > *ppoints) {
360 } while_each_thread(g, p);
366 * dump_tasks - dump current memory state of all system tasks
367 * @mem: current's memory controller, if constrained
368 * @nodemask: nodemask passed to page allocator for mempolicy ooms
370 * Dumps the current memory state of all eligible tasks. Tasks not in the same
371 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
373 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
374 * value, oom_score_adj value, and name.
376 * Call with tasklist_lock read-locked.
378 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
380 struct task_struct *p;
381 struct task_struct *task;
383 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
384 for_each_process(p) {
385 if (oom_unkillable_task(p, memcg, nodemask))
388 task = find_lock_task_mm(p);
391 * This is a kthread or all of p's threads have already
392 * detached their mm's. There's no need to report
393 * them; they can't be oom killed anyway.
398 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
399 task->pid, task_uid(task), task->tgid,
400 task->mm->total_vm, get_mm_rss(task->mm),
401 task_cpu(task), task->signal->oom_adj,
402 task->signal->oom_score_adj, task->comm);
407 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
408 struct mem_cgroup *memcg, const nodemask_t *nodemask)
411 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
412 "oom_adj=%d, oom_score_adj=%d\n",
413 current->comm, gfp_mask, order, current->signal->oom_adj,
414 current->signal->oom_score_adj);
415 cpuset_print_task_mems_allowed(current);
416 task_unlock(current);
418 mem_cgroup_print_oom_info(memcg, p);
419 show_mem(SHOW_MEM_FILTER_NODES);
420 if (sysctl_oom_dump_tasks)
421 dump_tasks(memcg, nodemask);
424 #define K(x) ((x) << (PAGE_SHIFT-10))
425 static void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
426 unsigned int points, unsigned long totalpages,
427 struct mem_cgroup *memcg, nodemask_t *nodemask,
430 struct task_struct *victim = p;
431 struct task_struct *child;
432 struct task_struct *t = p;
433 struct mm_struct *mm;
434 unsigned int victim_points = 0;
435 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
436 DEFAULT_RATELIMIT_BURST);
439 * If the task is already exiting, don't alarm the sysadmin or kill
440 * its children or threads, just set TIF_MEMDIE so it can die quickly
442 if (p->flags & PF_EXITING) {
443 set_tsk_thread_flag(p, TIF_MEMDIE);
447 if (__ratelimit(&oom_rs))
448 dump_header(p, gfp_mask, order, memcg, nodemask);
451 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
452 message, task_pid_nr(p), p->comm, points);
456 * If any of p's children has a different mm and is eligible for kill,
457 * the one with the highest oom_badness() score is sacrificed for its
458 * parent. This attempts to lose the minimal amount of work done while
459 * still freeing memory.
462 list_for_each_entry(child, &t->children, sibling) {
463 unsigned int child_points;
465 if (child->mm == p->mm)
468 * oom_badness() returns 0 if the thread is unkillable
470 child_points = oom_badness(child, memcg, nodemask,
472 if (child_points > victim_points) {
474 victim_points = child_points;
477 } while_each_thread(p, t);
479 victim = find_lock_task_mm(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.
501 if (p->mm == mm && !same_thread_group(p, victim) &&
502 !(p->flags & PF_KTHREAD)) {
503 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
506 task_lock(p); /* Protect ->comm from prctl() */
507 pr_err("Kill process %d (%s) sharing same memory\n",
508 task_pid_nr(p), p->comm);
510 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
513 set_tsk_thread_flag(victim, TIF_MEMDIE);
514 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
519 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
521 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
522 int order, const nodemask_t *nodemask)
524 if (likely(!sysctl_panic_on_oom))
526 if (sysctl_panic_on_oom != 2) {
528 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
529 * does not panic for cpuset, mempolicy, or memcg allocation
532 if (constraint != CONSTRAINT_NONE)
535 read_lock(&tasklist_lock);
536 dump_header(NULL, gfp_mask, order, NULL, nodemask);
537 read_unlock(&tasklist_lock);
538 panic("Out of memory: %s panic_on_oom is enabled\n",
539 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
542 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
543 void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
547 unsigned int points = 0;
548 struct task_struct *p;
551 * If current has a pending SIGKILL, then automatically select it. The
552 * goal is to allow it to allocate so that it may quickly exit and free
555 if (fatal_signal_pending(current)) {
556 set_thread_flag(TIF_MEMDIE);
560 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
561 limit = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT;
562 read_lock(&tasklist_lock);
563 p = select_bad_process(&points, limit, memcg, NULL, false);
564 if (p && PTR_ERR(p) != -1UL)
565 oom_kill_process(p, gfp_mask, order, points, limit, memcg, NULL,
566 "Memory cgroup out of memory");
567 read_unlock(&tasklist_lock);
571 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
573 int register_oom_notifier(struct notifier_block *nb)
575 return blocking_notifier_chain_register(&oom_notify_list, nb);
577 EXPORT_SYMBOL_GPL(register_oom_notifier);
579 int unregister_oom_notifier(struct notifier_block *nb)
581 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
583 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
586 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
587 * if a parallel OOM killing is already taking place that includes a zone in
588 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
590 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
596 spin_lock(&zone_scan_lock);
597 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
598 if (zone_is_oom_locked(zone)) {
604 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
606 * Lock each zone in the zonelist under zone_scan_lock so a
607 * parallel invocation of try_set_zonelist_oom() doesn't succeed
610 zone_set_flag(zone, ZONE_OOM_LOCKED);
614 spin_unlock(&zone_scan_lock);
619 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
620 * allocation attempts with zonelists containing them may now recall the OOM
621 * killer, if necessary.
623 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
628 spin_lock(&zone_scan_lock);
629 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
630 zone_clear_flag(zone, ZONE_OOM_LOCKED);
632 spin_unlock(&zone_scan_lock);
636 * Try to acquire the oom killer lock for all system zones. Returns zero if a
637 * parallel oom killing is taking place, otherwise locks all zones and returns
640 static int try_set_system_oom(void)
645 spin_lock(&zone_scan_lock);
646 for_each_populated_zone(zone)
647 if (zone_is_oom_locked(zone)) {
651 for_each_populated_zone(zone)
652 zone_set_flag(zone, ZONE_OOM_LOCKED);
654 spin_unlock(&zone_scan_lock);
659 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
660 * attempts or page faults may now recall the oom killer, if necessary.
662 static void clear_system_oom(void)
666 spin_lock(&zone_scan_lock);
667 for_each_populated_zone(zone)
668 zone_clear_flag(zone, ZONE_OOM_LOCKED);
669 spin_unlock(&zone_scan_lock);
673 * out_of_memory - kill the "best" process when we run out of memory
674 * @zonelist: zonelist pointer
675 * @gfp_mask: memory allocation flags
676 * @order: amount of memory being requested as a power of 2
677 * @nodemask: nodemask passed to page allocator
678 * @force_kill: true if a task must be killed, even if others are exiting
680 * If we run out of memory, we have the choice between either
681 * killing a random task (bad), letting the system crash (worse)
682 * OR try to be smart about which process to kill. Note that we
683 * don't have to be perfect here, we just have to be good.
685 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
686 int order, nodemask_t *nodemask, bool force_kill)
688 const nodemask_t *mpol_mask;
689 struct task_struct *p;
690 unsigned long totalpages;
691 unsigned long freed = 0;
693 enum oom_constraint constraint = CONSTRAINT_NONE;
696 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
698 /* Got some memory back in the last second. */
702 * If current has a pending SIGKILL or is exiting, then automatically
703 * select it. The goal is to allow it to allocate so that it may
704 * quickly exit and free its memory.
706 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
707 set_thread_flag(TIF_MEMDIE);
712 * Check if there were limitations on the allocation (only relevant for
713 * NUMA) that may require different handling.
715 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
717 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
718 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
720 read_lock(&tasklist_lock);
721 if (sysctl_oom_kill_allocating_task &&
722 !oom_unkillable_task(current, NULL, nodemask) &&
724 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
726 "Out of memory (oom_kill_allocating_task)");
730 p = select_bad_process(&points, totalpages, NULL, mpol_mask,
732 /* Found nothing?!?! Either we hang forever, or we panic. */
734 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
735 read_unlock(&tasklist_lock);
736 panic("Out of memory and no killable processes...\n");
738 if (PTR_ERR(p) != -1UL) {
739 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
740 nodemask, "Out of memory");
744 read_unlock(&tasklist_lock);
747 * Give "p" a good chance of killing itself before we
748 * retry to allocate memory unless "p" is current
750 if (killed && !test_thread_flag(TIF_MEMDIE))
751 schedule_timeout_uninterruptible(1);
755 * The pagefault handler calls here because it is out of memory, so kill a
756 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
757 * oom killing is already in progress so do nothing. If a task is found with
758 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
760 void pagefault_out_of_memory(void)
762 if (try_set_system_oom()) {
763 out_of_memory(NULL, 0, 0, NULL, false);
766 if (!test_thread_flag(TIF_MEMDIE))
767 schedule_timeout_uninterruptible(1);