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/module.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>
36 int sysctl_panic_on_oom;
37 int sysctl_oom_kill_allocating_task;
38 int sysctl_oom_dump_tasks = 1;
39 static DEFINE_SPINLOCK(zone_scan_lock);
42 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
43 * @new_val: new oom_score_adj value
45 * Sets the oom_score_adj value for current to @new_val with proper
46 * synchronization and returns the old value. Usually used to temporarily
47 * set a value, save the old value in the caller, and then reinstate it later.
49 int test_set_oom_score_adj(int new_val)
51 struct sighand_struct *sighand = current->sighand;
54 spin_lock_irq(&sighand->siglock);
55 old_val = current->signal->oom_score_adj;
56 if (new_val != old_val) {
57 if (new_val == OOM_SCORE_ADJ_MIN)
58 atomic_inc(¤t->mm->oom_disable_count);
59 else if (old_val == OOM_SCORE_ADJ_MIN)
60 atomic_dec(¤t->mm->oom_disable_count);
61 current->signal->oom_score_adj = new_val;
63 spin_unlock_irq(&sighand->siglock);
70 * has_intersects_mems_allowed() - check task eligiblity for kill
71 * @tsk: task struct of which task to consider
72 * @mask: nodemask passed to page allocator for mempolicy ooms
74 * Task eligibility is determined by whether or not a candidate task, @tsk,
75 * shares the same mempolicy nodes as current if it is bound by such a policy
76 * and whether or not it has the same set of allowed cpuset nodes.
78 static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
81 struct task_struct *start = tsk;
86 * If this is a mempolicy constrained oom, tsk's
87 * cpuset is irrelevant. Only return true if its
88 * mempolicy intersects current, otherwise it may be
91 if (mempolicy_nodemask_intersects(tsk, mask))
95 * This is not a mempolicy constrained oom, so only
96 * check the mems of tsk's cpuset.
98 if (cpuset_mems_allowed_intersects(current, tsk))
101 } while_each_thread(start, tsk);
106 static bool has_intersects_mems_allowed(struct task_struct *tsk,
107 const nodemask_t *mask)
111 #endif /* CONFIG_NUMA */
114 * The process p may have detached its own ->mm while exiting or through
115 * use_mm(), but one or more of its subthreads may still have a valid
116 * pointer. Return p, or any of its subthreads with a valid ->mm, with
119 struct task_struct *find_lock_task_mm(struct task_struct *p)
121 struct task_struct *t = p;
128 } while_each_thread(p, t);
133 /* return true if the task is not adequate as candidate victim task. */
134 static bool oom_unkillable_task(struct task_struct *p,
135 const struct mem_cgroup *mem, const nodemask_t *nodemask)
137 if (is_global_init(p))
139 if (p->flags & PF_KTHREAD)
142 /* When mem_cgroup_out_of_memory() and p is not member of the group */
143 if (mem && !task_in_mem_cgroup(p, mem))
146 /* p may not have freeable memory in nodemask */
147 if (!has_intersects_mems_allowed(p, nodemask))
154 * oom_badness - heuristic function to determine which candidate task to kill
155 * @p: task struct of which task we should calculate
156 * @totalpages: total present RAM allowed for page allocation
158 * The heuristic for determining which task to kill is made to be as simple and
159 * predictable as possible. The goal is to return the highest value for the
160 * task consuming the most memory to avoid subsequent oom failures.
162 unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
163 const nodemask_t *nodemask, unsigned long totalpages)
167 if (oom_unkillable_task(p, mem, nodemask))
170 p = find_lock_task_mm(p);
175 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
176 * so the entire heuristic doesn't need to be executed for something
177 * that cannot be killed.
179 if (atomic_read(&p->mm->oom_disable_count)) {
185 * The memory controller may have a limit of 0 bytes, so avoid a divide
186 * by zero, if necessary.
192 * The baseline for the badness score is the proportion of RAM that each
193 * task's rss, pagetable and swap space use.
195 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
196 points += get_mm_counter(p->mm, MM_SWAPENTS);
199 points /= totalpages;
203 * Root processes get 3% bonus, just like the __vm_enough_memory()
204 * implementation used by LSMs.
206 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
210 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
211 * either completely disable oom killing or always prefer a certain
214 points += p->signal->oom_score_adj;
217 * Never return 0 for an eligible task that may be killed since it's
218 * possible that no single user task uses more than 0.1% of memory and
219 * no single admin tasks uses more than 3.0%.
223 return (points < 1000) ? points : 1000;
227 * Determine the type of allocation constraint.
230 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
231 gfp_t gfp_mask, nodemask_t *nodemask,
232 unsigned long *totalpages)
236 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
237 bool cpuset_limited = false;
240 /* Default to all available memory */
241 *totalpages = totalram_pages + total_swap_pages;
244 return CONSTRAINT_NONE;
246 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
247 * to kill current.We have to random task kill in this case.
248 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
250 if (gfp_mask & __GFP_THISNODE)
251 return CONSTRAINT_NONE;
254 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
255 * the page allocator means a mempolicy is in effect. Cpuset policy
256 * is enforced in get_page_from_freelist().
258 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
259 *totalpages = total_swap_pages;
260 for_each_node_mask(nid, *nodemask)
261 *totalpages += node_spanned_pages(nid);
262 return CONSTRAINT_MEMORY_POLICY;
265 /* Check this allocation failure is caused by cpuset's wall function */
266 for_each_zone_zonelist_nodemask(zone, z, zonelist,
267 high_zoneidx, nodemask)
268 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
269 cpuset_limited = true;
271 if (cpuset_limited) {
272 *totalpages = total_swap_pages;
273 for_each_node_mask(nid, cpuset_current_mems_allowed)
274 *totalpages += node_spanned_pages(nid);
275 return CONSTRAINT_CPUSET;
277 return CONSTRAINT_NONE;
280 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281 gfp_t gfp_mask, nodemask_t *nodemask,
282 unsigned long *totalpages)
284 *totalpages = totalram_pages + total_swap_pages;
285 return CONSTRAINT_NONE;
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. We expect the caller will lock the tasklist.
293 * (not docbooked, we don't want this one cluttering up the manual)
295 static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, struct mem_cgroup *mem,
297 const nodemask_t *nodemask)
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
303 do_each_thread(g, p) {
308 if (oom_unkillable_task(p, mem, nodemask))
312 * This task already has access to memory reserves and is
313 * being killed. Don't allow any other task access to the
316 * Note: this may have a chance of deadlock if it gets
317 * blocked waiting for another task which itself is waiting
318 * for memory. Is there a better alternative?
320 if (test_tsk_thread_flag(p, TIF_MEMDIE))
321 return ERR_PTR(-1UL);
323 if (p->flags & PF_EXITING) {
325 * If p is the current task and is in the process of
326 * releasing memory, we allow the "kill" to set
327 * TIF_MEMDIE, which will allow it to gain access to
328 * memory reserves. Otherwise, it may stall forever.
330 * The loop isn't broken here, however, in case other
331 * threads are found to have already been oom killed.
338 * If this task is not being ptraced on exit,
339 * then wait for it to finish before killing
340 * some other task unnecessarily.
342 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
343 return ERR_PTR(-1UL);
347 points = oom_badness(p, mem, nodemask, totalpages);
348 if (points > *ppoints) {
352 } while_each_thread(g, p);
358 * dump_tasks - dump current memory state of all system tasks
359 * @mem: current's memory controller, if constrained
360 * @nodemask: nodemask passed to page allocator for mempolicy ooms
362 * Dumps the current memory state of all eligible tasks. Tasks not in the same
363 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
365 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
366 * value, oom_score_adj value, and name.
368 * Call with tasklist_lock read-locked.
370 static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
372 struct task_struct *p;
373 struct task_struct *task;
375 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
376 for_each_process(p) {
377 if (oom_unkillable_task(p, mem, nodemask))
380 task = find_lock_task_mm(p);
383 * This is a kthread or all of p's threads have already
384 * detached their mm's. There's no need to report
385 * them; they can't be oom killed anyway.
390 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
391 task->pid, task_uid(task), task->tgid,
392 task->mm->total_vm, get_mm_rss(task->mm),
393 task_cpu(task), task->signal->oom_adj,
394 task->signal->oom_score_adj, task->comm);
399 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
400 struct mem_cgroup *mem, const nodemask_t *nodemask)
403 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
404 "oom_adj=%d, oom_score_adj=%d\n",
405 current->comm, gfp_mask, order, current->signal->oom_adj,
406 current->signal->oom_score_adj);
407 cpuset_print_task_mems_allowed(current);
408 task_unlock(current);
410 mem_cgroup_print_oom_info(mem, p);
411 show_mem(SHOW_MEM_FILTER_NODES);
412 if (sysctl_oom_dump_tasks)
413 dump_tasks(mem, nodemask);
416 #define K(x) ((x) << (PAGE_SHIFT-10))
417 static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
419 struct task_struct *q;
420 struct mm_struct *mm;
422 p = find_lock_task_mm(p);
426 /* mm cannot be safely dereferenced after task_unlock(p) */
429 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
430 task_pid_nr(p), p->comm, K(p->mm->total_vm),
431 K(get_mm_counter(p->mm, MM_ANONPAGES)),
432 K(get_mm_counter(p->mm, MM_FILEPAGES)));
436 * Kill all processes sharing p->mm in other thread groups, if any.
437 * They don't get access to memory reserves or a higher scheduler
438 * priority, though, to avoid depletion of all memory or task
439 * starvation. This prevents mm->mmap_sem livelock when an oom killed
440 * task cannot exit because it requires the semaphore and its contended
441 * by another thread trying to allocate memory itself. That thread will
442 * now get access to memory reserves since it has a pending fatal
446 if (q->mm == mm && !same_thread_group(q, p)) {
447 task_lock(q); /* Protect ->comm from prctl() */
448 pr_err("Kill process %d (%s) sharing same memory\n",
449 task_pid_nr(q), q->comm);
451 force_sig(SIGKILL, q);
454 set_tsk_thread_flag(p, TIF_MEMDIE);
455 force_sig(SIGKILL, p);
461 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
462 unsigned int points, unsigned long totalpages,
463 struct mem_cgroup *mem, nodemask_t *nodemask,
466 struct task_struct *victim = p;
467 struct task_struct *child;
468 struct task_struct *t = p;
469 unsigned int victim_points = 0;
471 if (printk_ratelimit())
472 dump_header(p, gfp_mask, order, mem, nodemask);
475 * If the task is already exiting, don't alarm the sysadmin or kill
476 * its children or threads, just set TIF_MEMDIE so it can die quickly
478 if (p->flags & PF_EXITING) {
479 set_tsk_thread_flag(p, TIF_MEMDIE);
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.
495 list_for_each_entry(child, &t->children, sibling) {
496 unsigned int child_points;
498 if (child->mm == p->mm)
501 * oom_badness() returns 0 if the thread is unkillable
503 child_points = oom_badness(child, mem, nodemask,
505 if (child_points > victim_points) {
507 victim_points = child_points;
510 } while_each_thread(p, t);
512 return oom_kill_task(victim, mem);
516 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
518 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
519 int order, const nodemask_t *nodemask)
521 if (likely(!sysctl_panic_on_oom))
523 if (sysctl_panic_on_oom != 2) {
525 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
526 * does not panic for cpuset, mempolicy, or memcg allocation
529 if (constraint != CONSTRAINT_NONE)
532 read_lock(&tasklist_lock);
533 dump_header(NULL, gfp_mask, order, NULL, nodemask);
534 read_unlock(&tasklist_lock);
535 panic("Out of memory: %s panic_on_oom is enabled\n",
536 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
539 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
540 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
543 unsigned int points = 0;
544 struct task_struct *p;
547 * If current has a pending SIGKILL, then automatically select it. The
548 * goal is to allow it to allocate so that it may quickly exit and free
551 if (fatal_signal_pending(current)) {
552 set_thread_flag(TIF_MEMDIE);
556 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
557 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
558 read_lock(&tasklist_lock);
560 p = select_bad_process(&points, limit, mem, NULL);
561 if (!p || PTR_ERR(p) == -1UL)
564 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
565 "Memory cgroup out of memory"))
568 read_unlock(&tasklist_lock);
572 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
574 int register_oom_notifier(struct notifier_block *nb)
576 return blocking_notifier_chain_register(&oom_notify_list, nb);
578 EXPORT_SYMBOL_GPL(register_oom_notifier);
580 int unregister_oom_notifier(struct notifier_block *nb)
582 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
584 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
587 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
588 * if a parallel OOM killing is already taking place that includes a zone in
589 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
591 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
597 spin_lock(&zone_scan_lock);
598 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
599 if (zone_is_oom_locked(zone)) {
605 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
607 * Lock each zone in the zonelist under zone_scan_lock so a
608 * parallel invocation of try_set_zonelist_oom() doesn't succeed
611 zone_set_flag(zone, ZONE_OOM_LOCKED);
615 spin_unlock(&zone_scan_lock);
620 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
621 * allocation attempts with zonelists containing them may now recall the OOM
622 * killer, if necessary.
624 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
629 spin_lock(&zone_scan_lock);
630 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
631 zone_clear_flag(zone, ZONE_OOM_LOCKED);
633 spin_unlock(&zone_scan_lock);
637 * Try to acquire the oom killer lock for all system zones. Returns zero if a
638 * parallel oom killing is taking place, otherwise locks all zones and returns
641 static int try_set_system_oom(void)
646 spin_lock(&zone_scan_lock);
647 for_each_populated_zone(zone)
648 if (zone_is_oom_locked(zone)) {
652 for_each_populated_zone(zone)
653 zone_set_flag(zone, ZONE_OOM_LOCKED);
655 spin_unlock(&zone_scan_lock);
660 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
661 * attempts or page faults may now recall the oom killer, if necessary.
663 static void clear_system_oom(void)
667 spin_lock(&zone_scan_lock);
668 for_each_populated_zone(zone)
669 zone_clear_flag(zone, ZONE_OOM_LOCKED);
670 spin_unlock(&zone_scan_lock);
674 * out_of_memory - kill the "best" process when we run out of memory
675 * @zonelist: zonelist pointer
676 * @gfp_mask: memory allocation flags
677 * @order: amount of memory being requested as a power of 2
678 * @nodemask: nodemask passed to page allocator
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)
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, then automatically select it. The
703 * goal is to allow it to allocate so that it may quickly exit and free
706 if (fatal_signal_pending(current)) {
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) &&
723 current->mm && !atomic_read(¤t->mm->oom_disable_count)) {
725 * oom_kill_process() needs tasklist_lock held. If it returns
726 * non-zero, current could not be killed so we must fallback to
729 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
731 "Out of memory (oom_kill_allocating_task)"))
736 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
737 if (PTR_ERR(p) == -1UL)
740 /* Found nothing?!?! Either we hang forever, or we panic. */
742 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
743 read_unlock(&tasklist_lock);
744 panic("Out of memory and no killable processes...\n");
747 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
748 nodemask, "Out of memory"))
752 read_unlock(&tasklist_lock);
755 * Give "p" a good chance of killing itself before we
756 * retry to allocate memory unless "p" is current
758 if (killed && !test_thread_flag(TIF_MEMDIE))
759 schedule_timeout_uninterruptible(1);
763 * The pagefault handler calls here because it is out of memory, so kill a
764 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
765 * oom killing is already in progress so do nothing. If a task is found with
766 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
768 void pagefault_out_of_memory(void)
770 if (try_set_system_oom()) {
771 out_of_memory(NULL, 0, 0, NULL);
774 if (!test_thread_flag(TIF_MEMDIE))
775 schedule_timeout_uninterruptible(1);