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 long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
184 const nodemask_t *nodemask, unsigned long totalpages)
189 if (oom_unkillable_task(p, memcg, nodemask))
192 p = find_lock_task_mm(p);
196 adj = p->signal->oom_score_adj;
197 if (adj == OOM_SCORE_ADJ_MIN) {
203 * The baseline for the badness score is the proportion of RAM that each
204 * task's rss, pagetable and swap space use.
206 points = get_mm_rss(p->mm) + p->mm->nr_ptes +
207 get_mm_counter(p->mm, MM_SWAPENTS);
211 * Root processes get 3% bonus, just like the __vm_enough_memory()
212 * implementation used by LSMs.
214 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
217 /* Normalize to oom_score_adj units */
218 adj *= totalpages / 1000;
222 * Never return 0 for an eligible task regardless of the root bonus and
223 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
225 return points > 0 ? points : 1;
229 * Determine the type of allocation constraint.
232 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
233 gfp_t gfp_mask, nodemask_t *nodemask,
234 unsigned long *totalpages)
238 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
239 bool cpuset_limited = false;
242 /* Default to all available memory */
243 *totalpages = totalram_pages + total_swap_pages;
246 return CONSTRAINT_NONE;
248 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
249 * to kill current.We have to random task kill in this case.
250 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
252 if (gfp_mask & __GFP_THISNODE)
253 return CONSTRAINT_NONE;
256 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
257 * the page allocator means a mempolicy is in effect. Cpuset policy
258 * is enforced in get_page_from_freelist().
260 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
261 *totalpages = total_swap_pages;
262 for_each_node_mask(nid, *nodemask)
263 *totalpages += node_spanned_pages(nid);
264 return CONSTRAINT_MEMORY_POLICY;
267 /* Check this allocation failure is caused by cpuset's wall function */
268 for_each_zone_zonelist_nodemask(zone, z, zonelist,
269 high_zoneidx, nodemask)
270 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
271 cpuset_limited = true;
273 if (cpuset_limited) {
274 *totalpages = total_swap_pages;
275 for_each_node_mask(nid, cpuset_current_mems_allowed)
276 *totalpages += node_spanned_pages(nid);
277 return CONSTRAINT_CPUSET;
279 return CONSTRAINT_NONE;
282 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
283 gfp_t gfp_mask, nodemask_t *nodemask,
284 unsigned long *totalpages)
286 *totalpages = totalram_pages + total_swap_pages;
287 return CONSTRAINT_NONE;
291 enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
292 unsigned long totalpages, const nodemask_t *nodemask,
295 if (task->exit_state)
296 return OOM_SCAN_CONTINUE;
297 if (oom_unkillable_task(task, NULL, nodemask))
298 return OOM_SCAN_CONTINUE;
301 * This task already has access to memory reserves and is being killed.
302 * Don't allow any other task to have access to the reserves.
304 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
305 if (unlikely(frozen(task)))
308 return OOM_SCAN_ABORT;
311 return OOM_SCAN_CONTINUE;
313 if (task->flags & PF_EXITING) {
315 * If task is current and is in the process of releasing memory,
316 * allow the "kill" to set TIF_MEMDIE, which will allow it to
317 * access memory reserves. Otherwise, it may stall forever.
319 * The iteration isn't broken here, however, in case other
320 * threads are found to have already been oom killed.
323 return OOM_SCAN_SELECT;
324 else if (!force_kill) {
326 * If this task is not being ptraced on exit, then wait
327 * for it to finish before killing some other task
330 if (!(task->group_leader->ptrace & PT_TRACE_EXIT))
331 return OOM_SCAN_ABORT;
338 * Simple selection loop. We chose the process with the highest
339 * number of 'points'.
341 * (not docbooked, we don't want this one cluttering up the manual)
343 static struct task_struct *select_bad_process(unsigned int *ppoints,
344 unsigned long totalpages, const nodemask_t *nodemask,
347 struct task_struct *g, *p;
348 struct task_struct *chosen = NULL;
349 unsigned long chosen_points = 0;
352 do_each_thread(g, p) {
355 switch (oom_scan_process_thread(p, totalpages, nodemask,
357 case OOM_SCAN_SELECT:
359 chosen_points = ULONG_MAX;
361 case OOM_SCAN_CONTINUE:
365 return ERR_PTR(-1UL);
369 points = oom_badness(p, NULL, nodemask, totalpages);
370 if (points > chosen_points) {
372 chosen_points = points;
374 } while_each_thread(g, p);
376 get_task_struct(chosen);
379 *ppoints = chosen_points * 1000 / totalpages;
384 * dump_tasks - dump current memory state of all system tasks
385 * @memcg: current's memory controller, if constrained
386 * @nodemask: nodemask passed to page allocator for mempolicy ooms
388 * Dumps the current memory state of all eligible tasks. Tasks not in the same
389 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
391 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
392 * swapents, oom_score_adj value, and name.
394 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
396 struct task_struct *p;
397 struct task_struct *task;
399 pr_info("[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name\n");
401 for_each_process(p) {
402 if (oom_unkillable_task(p, memcg, nodemask))
405 task = find_lock_task_mm(p);
408 * This is a kthread or all of p's threads have already
409 * detached their mm's. There's no need to report
410 * them; they can't be oom killed anyway.
415 pr_info("[%5d] %5d %5d %8lu %8lu %7lu %8lu %5d %s\n",
416 task->pid, from_kuid(&init_user_ns, task_uid(task)),
417 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
419 get_mm_counter(task->mm, MM_SWAPENTS),
420 task->signal->oom_score_adj, task->comm);
426 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
427 struct mem_cgroup *memcg, const nodemask_t *nodemask)
430 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
431 "oom_adj=%d, oom_score_adj=%d\n",
432 current->comm, gfp_mask, order, current->signal->oom_adj,
433 current->signal->oom_score_adj);
434 cpuset_print_task_mems_allowed(current);
435 task_unlock(current);
437 mem_cgroup_print_oom_info(memcg, p);
438 show_mem(SHOW_MEM_FILTER_NODES);
439 if (sysctl_oom_dump_tasks)
440 dump_tasks(memcg, nodemask);
443 #define K(x) ((x) << (PAGE_SHIFT-10))
445 * Must be called while holding a reference to p, which will be released upon
448 void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
449 unsigned int points, unsigned long totalpages,
450 struct mem_cgroup *memcg, nodemask_t *nodemask,
453 struct task_struct *victim = p;
454 struct task_struct *child;
455 struct task_struct *t = p;
456 struct mm_struct *mm;
457 unsigned int victim_points = 0;
458 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
459 DEFAULT_RATELIMIT_BURST);
462 * If the task is already exiting, don't alarm the sysadmin or kill
463 * its children or threads, just set TIF_MEMDIE so it can die quickly
465 if (p->flags & PF_EXITING) {
466 set_tsk_thread_flag(p, TIF_MEMDIE);
471 if (__ratelimit(&oom_rs))
472 dump_header(p, gfp_mask, order, memcg, nodemask);
475 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
476 message, task_pid_nr(p), p->comm, points);
480 * If any of p's children has a different mm and is eligible for kill,
481 * the one with the highest oom_badness() score is sacrificed for its
482 * parent. This attempts to lose the minimal amount of work done while
483 * still freeing memory.
485 read_lock(&tasklist_lock);
487 list_for_each_entry(child, &t->children, sibling) {
488 unsigned int child_points;
490 if (child->mm == p->mm)
493 * oom_badness() returns 0 if the thread is unkillable
495 child_points = oom_badness(child, memcg, nodemask,
497 if (child_points > victim_points) {
498 put_task_struct(victim);
500 victim_points = child_points;
501 get_task_struct(victim);
504 } while_each_thread(p, t);
505 read_unlock(&tasklist_lock);
508 p = find_lock_task_mm(victim);
511 put_task_struct(victim);
513 } else if (victim != p) {
515 put_task_struct(victim);
519 /* mm cannot safely be dereferenced after task_unlock(victim) */
521 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
522 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
523 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
524 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
528 * Kill all user processes sharing victim->mm in other thread groups, if
529 * any. They don't get access to memory reserves, though, to avoid
530 * depletion of all memory. This prevents mm->mmap_sem livelock when an
531 * oom killed thread cannot exit because it requires the semaphore and
532 * its contended by another thread trying to allocate memory itself.
533 * That thread will now get access to memory reserves since it has a
534 * pending fatal signal.
537 if (p->mm == mm && !same_thread_group(p, victim) &&
538 !(p->flags & PF_KTHREAD)) {
539 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
542 task_lock(p); /* Protect ->comm from prctl() */
543 pr_err("Kill process %d (%s) sharing same memory\n",
544 task_pid_nr(p), p->comm);
546 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
550 set_tsk_thread_flag(victim, TIF_MEMDIE);
551 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
552 put_task_struct(victim);
557 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
559 void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
560 int order, const nodemask_t *nodemask)
562 if (likely(!sysctl_panic_on_oom))
564 if (sysctl_panic_on_oom != 2) {
566 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
567 * does not panic for cpuset, mempolicy, or memcg allocation
570 if (constraint != CONSTRAINT_NONE)
573 dump_header(NULL, gfp_mask, order, NULL, nodemask);
574 panic("Out of memory: %s panic_on_oom is enabled\n",
575 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
578 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
580 int register_oom_notifier(struct notifier_block *nb)
582 return blocking_notifier_chain_register(&oom_notify_list, nb);
584 EXPORT_SYMBOL_GPL(register_oom_notifier);
586 int unregister_oom_notifier(struct notifier_block *nb)
588 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
590 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
593 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
594 * if a parallel OOM killing is already taking place that includes a zone in
595 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
597 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
603 spin_lock(&zone_scan_lock);
604 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
605 if (zone_is_oom_locked(zone)) {
611 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
613 * Lock each zone in the zonelist under zone_scan_lock so a
614 * parallel invocation of try_set_zonelist_oom() doesn't succeed
617 zone_set_flag(zone, ZONE_OOM_LOCKED);
621 spin_unlock(&zone_scan_lock);
626 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
627 * allocation attempts with zonelists containing them may now recall the OOM
628 * killer, if necessary.
630 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
635 spin_lock(&zone_scan_lock);
636 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
637 zone_clear_flag(zone, ZONE_OOM_LOCKED);
639 spin_unlock(&zone_scan_lock);
643 * Try to acquire the oom killer lock for all system zones. Returns zero if a
644 * parallel oom killing is taking place, otherwise locks all zones and returns
647 static int try_set_system_oom(void)
652 spin_lock(&zone_scan_lock);
653 for_each_populated_zone(zone)
654 if (zone_is_oom_locked(zone)) {
658 for_each_populated_zone(zone)
659 zone_set_flag(zone, ZONE_OOM_LOCKED);
661 spin_unlock(&zone_scan_lock);
666 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
667 * attempts or page faults may now recall the oom killer, if necessary.
669 static void clear_system_oom(void)
673 spin_lock(&zone_scan_lock);
674 for_each_populated_zone(zone)
675 zone_clear_flag(zone, ZONE_OOM_LOCKED);
676 spin_unlock(&zone_scan_lock);
680 * out_of_memory - kill the "best" process when we run out of memory
681 * @zonelist: zonelist pointer
682 * @gfp_mask: memory allocation flags
683 * @order: amount of memory being requested as a power of 2
684 * @nodemask: nodemask passed to page allocator
685 * @force_kill: true if a task must be killed, even if others are exiting
687 * If we run out of memory, we have the choice between either
688 * killing a random task (bad), letting the system crash (worse)
689 * OR try to be smart about which process to kill. Note that we
690 * don't have to be perfect here, we just have to be good.
692 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
693 int order, nodemask_t *nodemask, bool force_kill)
695 const nodemask_t *mpol_mask;
696 struct task_struct *p;
697 unsigned long totalpages;
698 unsigned long freed = 0;
699 unsigned int uninitialized_var(points);
700 enum oom_constraint constraint = CONSTRAINT_NONE;
703 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
705 /* Got some memory back in the last second. */
709 * If current has a pending SIGKILL, then automatically select it. The
710 * goal is to allow it to allocate so that it may quickly exit and free
713 if (fatal_signal_pending(current)) {
714 set_thread_flag(TIF_MEMDIE);
719 * Check if there were limitations on the allocation (only relevant for
720 * NUMA) that may require different handling.
722 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
724 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
725 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
727 if (sysctl_oom_kill_allocating_task && current->mm &&
728 !oom_unkillable_task(current, NULL, nodemask) &&
729 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
730 get_task_struct(current);
731 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
733 "Out of memory (oom_kill_allocating_task)");
737 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
738 /* Found nothing?!?! Either we hang forever, or we panic. */
740 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
741 panic("Out of memory and no killable processes...\n");
743 if (PTR_ERR(p) != -1UL) {
744 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
745 nodemask, "Out of memory");
750 * Give the killed threads a good chance of exiting before trying to
751 * allocate memory again.
754 schedule_timeout_killable(1);
758 * The pagefault handler calls here because it is out of memory, so kill a
759 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
760 * oom killing is already in progress so do nothing. If a task is found with
761 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
763 void pagefault_out_of_memory(void)
765 if (try_set_system_oom()) {
766 out_of_memory(NULL, 0, 0, NULL, false);
769 schedule_timeout_killable(1);