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) {
347 * If p is the current task and is in the process of
348 * releasing memory, we allow the "kill" to set
349 * TIF_MEMDIE, which will allow it to gain access to
350 * memory reserves. Otherwise, it may stall forever.
352 * The loop isn't broken here, however, in case other
353 * threads are found to have already been oom killed.
358 } else if (!force_kill) {
360 * If this task is not being ptraced on exit,
361 * then wait for it to finish before killing
362 * some other task unnecessarily.
364 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
365 return ERR_PTR(-1UL);
369 points = oom_badness(p, memcg, nodemask, totalpages);
370 if (points > *ppoints) {
374 } while_each_thread(g, p);
380 * dump_tasks - dump current memory state of all system tasks
381 * @mem: current's memory controller, if constrained
382 * @nodemask: nodemask passed to page allocator for mempolicy ooms
384 * Dumps the current memory state of all eligible tasks. Tasks not in the same
385 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
387 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
388 * value, oom_score_adj value, and name.
390 * Call with tasklist_lock read-locked.
392 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
394 struct task_struct *p;
395 struct task_struct *task;
397 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
398 for_each_process(p) {
399 if (oom_unkillable_task(p, memcg, nodemask))
402 task = find_lock_task_mm(p);
405 * This is a kthread or all of p's threads have already
406 * detached their mm's. There's no need to report
407 * them; they can't be oom killed anyway.
412 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
413 task->pid, task_uid(task), task->tgid,
414 task->mm->total_vm, get_mm_rss(task->mm),
415 task_cpu(task), task->signal->oom_adj,
416 task->signal->oom_score_adj, task->comm);
421 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
422 struct mem_cgroup *memcg, const nodemask_t *nodemask)
425 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
426 "oom_adj=%d, oom_score_adj=%d\n",
427 current->comm, gfp_mask, order, current->signal->oom_adj,
428 current->signal->oom_score_adj);
429 cpuset_print_task_mems_allowed(current);
430 task_unlock(current);
432 mem_cgroup_print_oom_info(memcg, p);
433 show_mem(SHOW_MEM_FILTER_NODES);
434 if (sysctl_oom_dump_tasks)
435 dump_tasks(memcg, nodemask);
438 #define K(x) ((x) << (PAGE_SHIFT-10))
439 static void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
440 unsigned int points, unsigned long totalpages,
441 struct mem_cgroup *memcg, nodemask_t *nodemask,
444 struct task_struct *victim = p;
445 struct task_struct *child;
446 struct task_struct *t = p;
447 struct mm_struct *mm;
448 unsigned int victim_points = 0;
449 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
450 DEFAULT_RATELIMIT_BURST);
453 * If the task is already exiting, don't alarm the sysadmin or kill
454 * its children or threads, just set TIF_MEMDIE so it can die quickly
456 if (p->flags & PF_EXITING) {
457 set_tsk_thread_flag(p, TIF_MEMDIE);
461 if (__ratelimit(&oom_rs))
462 dump_header(p, gfp_mask, order, memcg, nodemask);
465 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
466 message, task_pid_nr(p), p->comm, points);
470 * If any of p's children has a different mm and is eligible for kill,
471 * the one with the highest oom_badness() score is sacrificed for its
472 * parent. This attempts to lose the minimal amount of work done while
473 * still freeing memory.
476 list_for_each_entry(child, &t->children, sibling) {
477 unsigned int child_points;
479 if (child->mm == p->mm)
482 * oom_badness() returns 0 if the thread is unkillable
484 child_points = oom_badness(child, memcg, nodemask,
486 if (child_points > victim_points) {
488 victim_points = child_points;
491 } while_each_thread(p, t);
493 victim = find_lock_task_mm(victim);
497 /* mm cannot safely be dereferenced after task_unlock(victim) */
499 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
500 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
501 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
502 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
506 * Kill all user processes sharing victim->mm in other thread groups, if
507 * any. They don't get access to memory reserves, though, to avoid
508 * depletion of all memory. This prevents mm->mmap_sem livelock when an
509 * oom killed thread cannot exit because it requires the semaphore and
510 * its contended by another thread trying to allocate memory itself.
511 * That thread will now get access to memory reserves since it has a
512 * pending fatal signal.
515 if (p->mm == mm && !same_thread_group(p, victim) &&
516 !(p->flags & PF_KTHREAD)) {
517 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
520 task_lock(p); /* Protect ->comm from prctl() */
521 pr_err("Kill process %d (%s) sharing same memory\n",
522 task_pid_nr(p), p->comm);
524 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
527 set_tsk_thread_flag(victim, TIF_MEMDIE);
528 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
533 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
535 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
536 int order, const nodemask_t *nodemask)
538 if (likely(!sysctl_panic_on_oom))
540 if (sysctl_panic_on_oom != 2) {
542 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
543 * does not panic for cpuset, mempolicy, or memcg allocation
546 if (constraint != CONSTRAINT_NONE)
549 read_lock(&tasklist_lock);
550 dump_header(NULL, gfp_mask, order, NULL, nodemask);
551 read_unlock(&tasklist_lock);
552 panic("Out of memory: %s panic_on_oom is enabled\n",
553 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
556 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
557 void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
561 unsigned int points = 0;
562 struct task_struct *p;
565 * If current has a pending SIGKILL, then automatically select it. The
566 * goal is to allow it to allocate so that it may quickly exit and free
569 if (fatal_signal_pending(current)) {
570 set_thread_flag(TIF_MEMDIE);
574 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
575 limit = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT;
576 read_lock(&tasklist_lock);
577 p = select_bad_process(&points, limit, memcg, NULL, false);
578 if (p && PTR_ERR(p) != -1UL)
579 oom_kill_process(p, gfp_mask, order, points, limit, memcg, NULL,
580 "Memory cgroup out of memory");
581 read_unlock(&tasklist_lock);
585 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
587 int register_oom_notifier(struct notifier_block *nb)
589 return blocking_notifier_chain_register(&oom_notify_list, nb);
591 EXPORT_SYMBOL_GPL(register_oom_notifier);
593 int unregister_oom_notifier(struct notifier_block *nb)
595 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
597 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
600 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
601 * if a parallel OOM killing is already taking place that includes a zone in
602 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
604 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
610 spin_lock(&zone_scan_lock);
611 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
612 if (zone_is_oom_locked(zone)) {
618 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
620 * Lock each zone in the zonelist under zone_scan_lock so a
621 * parallel invocation of try_set_zonelist_oom() doesn't succeed
624 zone_set_flag(zone, ZONE_OOM_LOCKED);
628 spin_unlock(&zone_scan_lock);
633 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
634 * allocation attempts with zonelists containing them may now recall the OOM
635 * killer, if necessary.
637 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
642 spin_lock(&zone_scan_lock);
643 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
644 zone_clear_flag(zone, ZONE_OOM_LOCKED);
646 spin_unlock(&zone_scan_lock);
650 * Try to acquire the oom killer lock for all system zones. Returns zero if a
651 * parallel oom killing is taking place, otherwise locks all zones and returns
654 static int try_set_system_oom(void)
659 spin_lock(&zone_scan_lock);
660 for_each_populated_zone(zone)
661 if (zone_is_oom_locked(zone)) {
665 for_each_populated_zone(zone)
666 zone_set_flag(zone, ZONE_OOM_LOCKED);
668 spin_unlock(&zone_scan_lock);
673 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
674 * attempts or page faults may now recall the oom killer, if necessary.
676 static void clear_system_oom(void)
680 spin_lock(&zone_scan_lock);
681 for_each_populated_zone(zone)
682 zone_clear_flag(zone, ZONE_OOM_LOCKED);
683 spin_unlock(&zone_scan_lock);
687 * out_of_memory - kill the "best" process when we run out of memory
688 * @zonelist: zonelist pointer
689 * @gfp_mask: memory allocation flags
690 * @order: amount of memory being requested as a power of 2
691 * @nodemask: nodemask passed to page allocator
692 * @force_kill: true if a task must be killed, even if others are exiting
694 * If we run out of memory, we have the choice between either
695 * killing a random task (bad), letting the system crash (worse)
696 * OR try to be smart about which process to kill. Note that we
697 * don't have to be perfect here, we just have to be good.
699 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
700 int order, nodemask_t *nodemask, bool force_kill)
702 const nodemask_t *mpol_mask;
703 struct task_struct *p;
704 unsigned long totalpages;
705 unsigned long freed = 0;
707 enum oom_constraint constraint = CONSTRAINT_NONE;
710 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
712 /* Got some memory back in the last second. */
716 * If current has a pending SIGKILL, then automatically select it. The
717 * goal is to allow it to allocate so that it may quickly exit and free
720 if (fatal_signal_pending(current)) {
721 set_thread_flag(TIF_MEMDIE);
726 * Check if there were limitations on the allocation (only relevant for
727 * NUMA) that may require different handling.
729 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
731 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
732 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
734 read_lock(&tasklist_lock);
735 if (sysctl_oom_kill_allocating_task &&
736 !oom_unkillable_task(current, NULL, nodemask) &&
738 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
740 "Out of memory (oom_kill_allocating_task)");
744 p = select_bad_process(&points, totalpages, NULL, mpol_mask,
746 /* Found nothing?!?! Either we hang forever, or we panic. */
748 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
749 read_unlock(&tasklist_lock);
750 panic("Out of memory and no killable processes...\n");
752 if (PTR_ERR(p) != -1UL) {
753 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
754 nodemask, "Out of memory");
758 read_unlock(&tasklist_lock);
761 * Give "p" a good chance of killing itself before we
762 * retry to allocate memory unless "p" is current
764 if (killed && !test_thread_flag(TIF_MEMDIE))
765 schedule_timeout_uninterruptible(1);
769 * The pagefault handler calls here because it is out of memory, so kill a
770 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
771 * oom killing is already in progress so do nothing. If a task is found with
772 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
774 void pagefault_out_of_memory(void)
776 if (try_set_system_oom()) {
777 out_of_memory(NULL, 0, 0, NULL, false);
780 if (!test_thread_flag(TIF_MEMDIE))
781 schedule_timeout_uninterruptible(1);