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>
35 #include <linux/freezer.h>
37 int sysctl_panic_on_oom;
38 int sysctl_oom_kill_allocating_task;
39 int sysctl_oom_dump_tasks = 1;
40 static DEFINE_SPINLOCK(zone_scan_lock);
43 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
44 * @new_val: new oom_score_adj value
46 * Sets the oom_score_adj value for current to @new_val with proper
47 * synchronization and returns the old value. Usually used to temporarily
48 * set a value, save the old value in the caller, and then reinstate it later.
50 int test_set_oom_score_adj(int new_val)
52 struct sighand_struct *sighand = current->sighand;
55 spin_lock_irq(&sighand->siglock);
56 old_val = current->signal->oom_score_adj;
57 current->signal->oom_score_adj = new_val;
58 spin_unlock_irq(&sighand->siglock);
65 * has_intersects_mems_allowed() - check task eligiblity for kill
66 * @tsk: task struct of which task to consider
67 * @mask: nodemask passed to page allocator for mempolicy ooms
69 * Task eligibility is determined by whether or not a candidate task, @tsk,
70 * shares the same mempolicy nodes as current if it is bound by such a policy
71 * and whether or not it has the same set of allowed cpuset nodes.
73 static bool has_intersects_mems_allowed(struct task_struct *tsk,
74 const nodemask_t *mask)
76 struct task_struct *start = tsk;
81 * If this is a mempolicy constrained oom, tsk's
82 * cpuset is irrelevant. Only return true if its
83 * mempolicy intersects current, otherwise it may be
86 if (mempolicy_nodemask_intersects(tsk, mask))
90 * This is not a mempolicy constrained oom, so only
91 * check the mems of tsk's cpuset.
93 if (cpuset_mems_allowed_intersects(current, tsk))
96 } while_each_thread(start, tsk);
101 static bool has_intersects_mems_allowed(struct task_struct *tsk,
102 const nodemask_t *mask)
106 #endif /* CONFIG_NUMA */
109 * The process p may have detached its own ->mm while exiting or through
110 * use_mm(), but one or more of its subthreads may still have a valid
111 * pointer. Return p, or any of its subthreads with a valid ->mm, with
114 struct task_struct *find_lock_task_mm(struct task_struct *p)
116 struct task_struct *t = p;
123 } while_each_thread(p, t);
128 /* return true if the task is not adequate as candidate victim task. */
129 static bool oom_unkillable_task(struct task_struct *p,
130 const struct mem_cgroup *mem, const nodemask_t *nodemask)
132 if (is_global_init(p))
134 if (p->flags & PF_KTHREAD)
137 /* When mem_cgroup_out_of_memory() and p is not member of the group */
138 if (mem && !task_in_mem_cgroup(p, mem))
141 /* p may not have freeable memory in nodemask */
142 if (!has_intersects_mems_allowed(p, nodemask))
149 * oom_badness - heuristic function to determine which candidate task to kill
150 * @p: task struct of which task we should calculate
151 * @totalpages: total present RAM allowed for page allocation
153 * The heuristic for determining which task to kill is made to be as simple and
154 * predictable as possible. The goal is to return the highest value for the
155 * task consuming the most memory to avoid subsequent oom failures.
157 unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
158 const nodemask_t *nodemask, unsigned long totalpages)
162 if (oom_unkillable_task(p, mem, nodemask))
165 p = find_lock_task_mm(p);
170 * The memory controller may have a limit of 0 bytes, so avoid a divide
171 * by zero, if necessary.
177 * The baseline for the badness score is the proportion of RAM that each
178 * task's rss, pagetable and swap space use.
180 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
181 points += get_mm_counter(p->mm, MM_SWAPENTS);
184 points /= totalpages;
188 * Root processes get 3% bonus, just like the __vm_enough_memory()
189 * implementation used by LSMs.
191 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
195 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
196 * either completely disable oom killing or always prefer a certain
199 points += p->signal->oom_score_adj;
202 * Never return 0 for an eligible task that may be killed since it's
203 * possible that no single user task uses more than 0.1% of memory and
204 * no single admin tasks uses more than 3.0%.
208 return (points < 1000) ? points : 1000;
212 * Determine the type of allocation constraint.
215 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
216 gfp_t gfp_mask, nodemask_t *nodemask,
217 unsigned long *totalpages)
221 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
222 bool cpuset_limited = false;
225 /* Default to all available memory */
226 *totalpages = totalram_pages + total_swap_pages;
229 return CONSTRAINT_NONE;
231 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
232 * to kill current.We have to random task kill in this case.
233 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
235 if (gfp_mask & __GFP_THISNODE)
236 return CONSTRAINT_NONE;
239 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
240 * the page allocator means a mempolicy is in effect. Cpuset policy
241 * is enforced in get_page_from_freelist().
243 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
244 *totalpages = total_swap_pages;
245 for_each_node_mask(nid, *nodemask)
246 *totalpages += node_spanned_pages(nid);
247 return CONSTRAINT_MEMORY_POLICY;
250 /* Check this allocation failure is caused by cpuset's wall function */
251 for_each_zone_zonelist_nodemask(zone, z, zonelist,
252 high_zoneidx, nodemask)
253 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
254 cpuset_limited = true;
256 if (cpuset_limited) {
257 *totalpages = total_swap_pages;
258 for_each_node_mask(nid, cpuset_current_mems_allowed)
259 *totalpages += node_spanned_pages(nid);
260 return CONSTRAINT_CPUSET;
262 return CONSTRAINT_NONE;
265 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
266 gfp_t gfp_mask, nodemask_t *nodemask,
267 unsigned long *totalpages)
269 *totalpages = totalram_pages + total_swap_pages;
270 return CONSTRAINT_NONE;
275 * Simple selection loop. We chose the process with the highest
276 * number of 'points'. We expect the caller will lock the tasklist.
278 * (not docbooked, we don't want this one cluttering up the manual)
280 static struct task_struct *select_bad_process(unsigned int *ppoints,
281 unsigned long totalpages, struct mem_cgroup *mem,
282 const nodemask_t *nodemask)
284 struct task_struct *g, *p;
285 struct task_struct *chosen = NULL;
288 do_each_thread(g, p) {
293 if (oom_unkillable_task(p, mem, nodemask))
297 * This task already has access to memory reserves and is
298 * being killed. Don't allow any other task access to the
301 * Note: this may have a chance of deadlock if it gets
302 * blocked waiting for another task which itself is waiting
303 * for memory. Is there a better alternative?
305 if (test_tsk_thread_flag(p, TIF_MEMDIE)) {
306 if (unlikely(frozen(p)))
308 return ERR_PTR(-1UL);
313 if (p->flags & PF_EXITING) {
315 * If p is the current task and is in the process of
316 * releasing memory, we allow the "kill" to set
317 * TIF_MEMDIE, which will allow it to gain access to
318 * memory reserves. Otherwise, it may stall forever.
320 * The loop isn't broken here, however, in case other
321 * threads are found to have already been oom killed.
328 * If this task is not being ptraced on exit,
329 * then wait for it to finish before killing
330 * some other task unnecessarily.
332 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
333 return ERR_PTR(-1UL);
337 points = oom_badness(p, mem, nodemask, totalpages);
338 if (points > *ppoints) {
342 } while_each_thread(g, p);
348 * dump_tasks - dump current memory state of all system tasks
349 * @mem: 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, cpu, oom_adj
356 * value, oom_score_adj value, and name.
358 * Call with tasklist_lock read-locked.
360 static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
362 struct task_struct *p;
363 struct task_struct *task;
365 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
366 for_each_process(p) {
367 if (oom_unkillable_task(p, mem, nodemask))
370 task = find_lock_task_mm(p);
373 * This is a kthread or all of p's threads have already
374 * detached their mm's. There's no need to report
375 * them; they can't be oom killed anyway.
380 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
381 task->pid, task_uid(task), task->tgid,
382 task->mm->total_vm, get_mm_rss(task->mm),
383 task_cpu(task), task->signal->oom_adj,
384 task->signal->oom_score_adj, task->comm);
389 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
390 struct mem_cgroup *mem, const nodemask_t *nodemask)
393 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
394 "oom_adj=%d, oom_score_adj=%d\n",
395 current->comm, gfp_mask, order, current->signal->oom_adj,
396 current->signal->oom_score_adj);
397 cpuset_print_task_mems_allowed(current);
398 task_unlock(current);
400 mem_cgroup_print_oom_info(mem, p);
401 show_mem(SHOW_MEM_FILTER_NODES);
402 if (sysctl_oom_dump_tasks)
403 dump_tasks(mem, nodemask);
406 #define K(x) ((x) << (PAGE_SHIFT-10))
407 static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
409 struct task_struct *q;
410 struct mm_struct *mm;
412 p = find_lock_task_mm(p);
416 /* mm cannot be safely dereferenced after task_unlock(p) */
419 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
420 task_pid_nr(p), p->comm, K(p->mm->total_vm),
421 K(get_mm_counter(p->mm, MM_ANONPAGES)),
422 K(get_mm_counter(p->mm, MM_FILEPAGES)));
426 * Kill all user processes sharing p->mm in other thread groups, if any.
427 * They don't get access to memory reserves or a higher scheduler
428 * priority, though, to avoid depletion of all memory or task
429 * starvation. This prevents mm->mmap_sem livelock when an oom killed
430 * task cannot exit because it requires the semaphore and its contended
431 * by another thread trying to allocate memory itself. That thread will
432 * now get access to memory reserves since it has a pending fatal
436 if (q->mm == mm && !same_thread_group(q, p) &&
437 !(q->flags & PF_KTHREAD)) {
438 if (q->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
441 task_lock(q); /* Protect ->comm from prctl() */
442 pr_err("Kill process %d (%s) sharing same memory\n",
443 task_pid_nr(q), q->comm);
445 force_sig(SIGKILL, q);
448 set_tsk_thread_flag(p, TIF_MEMDIE);
449 force_sig(SIGKILL, p);
455 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
456 unsigned int points, unsigned long totalpages,
457 struct mem_cgroup *mem, nodemask_t *nodemask,
460 struct task_struct *victim = p;
461 struct task_struct *child;
462 struct task_struct *t = p;
463 unsigned int victim_points = 0;
465 if (printk_ratelimit())
466 dump_header(p, gfp_mask, order, mem, nodemask);
469 * If the task is already exiting, don't alarm the sysadmin or kill
470 * its children or threads, just set TIF_MEMDIE so it can die quickly
472 if (p->flags & PF_EXITING) {
473 set_tsk_thread_flag(p, TIF_MEMDIE);
478 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
479 message, task_pid_nr(p), p->comm, points);
483 * If any of p's children has a different mm and is eligible for kill,
484 * the one with the highest oom_badness() score is sacrificed for its
485 * parent. This attempts to lose the minimal amount of work done while
486 * still freeing memory.
489 list_for_each_entry(child, &t->children, sibling) {
490 unsigned int child_points;
492 if (child->mm == p->mm)
495 * oom_badness() returns 0 if the thread is unkillable
497 child_points = oom_badness(child, mem, nodemask,
499 if (child_points > victim_points) {
501 victim_points = child_points;
504 } while_each_thread(p, t);
506 return oom_kill_task(victim, mem);
510 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
512 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
513 int order, const nodemask_t *nodemask)
515 if (likely(!sysctl_panic_on_oom))
517 if (sysctl_panic_on_oom != 2) {
519 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
520 * does not panic for cpuset, mempolicy, or memcg allocation
523 if (constraint != CONSTRAINT_NONE)
526 read_lock(&tasklist_lock);
527 dump_header(NULL, gfp_mask, order, NULL, nodemask);
528 read_unlock(&tasklist_lock);
529 panic("Out of memory: %s panic_on_oom is enabled\n",
530 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
533 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
534 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
537 unsigned int points = 0;
538 struct task_struct *p;
541 * If current has a pending SIGKILL, then automatically select it. The
542 * goal is to allow it to allocate so that it may quickly exit and free
545 if (fatal_signal_pending(current)) {
546 set_thread_flag(TIF_MEMDIE);
550 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
551 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
552 read_lock(&tasklist_lock);
554 p = select_bad_process(&points, limit, mem, NULL);
555 if (!p || PTR_ERR(p) == -1UL)
558 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
559 "Memory cgroup out of memory"))
562 read_unlock(&tasklist_lock);
566 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
568 int register_oom_notifier(struct notifier_block *nb)
570 return blocking_notifier_chain_register(&oom_notify_list, nb);
572 EXPORT_SYMBOL_GPL(register_oom_notifier);
574 int unregister_oom_notifier(struct notifier_block *nb)
576 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
578 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
581 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
582 * if a parallel OOM killing is already taking place that includes a zone in
583 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
585 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
591 spin_lock(&zone_scan_lock);
592 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
593 if (zone_is_oom_locked(zone)) {
599 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
601 * Lock each zone in the zonelist under zone_scan_lock so a
602 * parallel invocation of try_set_zonelist_oom() doesn't succeed
605 zone_set_flag(zone, ZONE_OOM_LOCKED);
609 spin_unlock(&zone_scan_lock);
614 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
615 * allocation attempts with zonelists containing them may now recall the OOM
616 * killer, if necessary.
618 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
623 spin_lock(&zone_scan_lock);
624 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
625 zone_clear_flag(zone, ZONE_OOM_LOCKED);
627 spin_unlock(&zone_scan_lock);
631 * Try to acquire the oom killer lock for all system zones. Returns zero if a
632 * parallel oom killing is taking place, otherwise locks all zones and returns
635 static int try_set_system_oom(void)
640 spin_lock(&zone_scan_lock);
641 for_each_populated_zone(zone)
642 if (zone_is_oom_locked(zone)) {
646 for_each_populated_zone(zone)
647 zone_set_flag(zone, ZONE_OOM_LOCKED);
649 spin_unlock(&zone_scan_lock);
654 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
655 * attempts or page faults may now recall the oom killer, if necessary.
657 static void clear_system_oom(void)
661 spin_lock(&zone_scan_lock);
662 for_each_populated_zone(zone)
663 zone_clear_flag(zone, ZONE_OOM_LOCKED);
664 spin_unlock(&zone_scan_lock);
668 * out_of_memory - kill the "best" process when we run out of memory
669 * @zonelist: zonelist pointer
670 * @gfp_mask: memory allocation flags
671 * @order: amount of memory being requested as a power of 2
672 * @nodemask: nodemask passed to page allocator
674 * If we run out of memory, we have the choice between either
675 * killing a random task (bad), letting the system crash (worse)
676 * OR try to be smart about which process to kill. Note that we
677 * don't have to be perfect here, we just have to be good.
679 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
680 int order, nodemask_t *nodemask)
682 const nodemask_t *mpol_mask;
683 struct task_struct *p;
684 unsigned long totalpages;
685 unsigned long freed = 0;
687 enum oom_constraint constraint = CONSTRAINT_NONE;
690 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
692 /* Got some memory back in the last second. */
696 * If current has a pending SIGKILL, then automatically select it. The
697 * goal is to allow it to allocate so that it may quickly exit and free
700 if (fatal_signal_pending(current)) {
701 set_thread_flag(TIF_MEMDIE);
706 * Check if there were limitations on the allocation (only relevant for
707 * NUMA) that may require different handling.
709 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
711 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
712 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
714 read_lock(&tasklist_lock);
715 if (sysctl_oom_kill_allocating_task &&
716 !oom_unkillable_task(current, NULL, nodemask) &&
719 * oom_kill_process() needs tasklist_lock held. If it returns
720 * non-zero, current could not be killed so we must fallback to
723 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
725 "Out of memory (oom_kill_allocating_task)"))
730 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
731 if (PTR_ERR(p) == -1UL)
734 /* Found nothing?!?! Either we hang forever, or we panic. */
736 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
737 read_unlock(&tasklist_lock);
738 panic("Out of memory and no killable processes...\n");
741 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
742 nodemask, "Out of memory"))
746 read_unlock(&tasklist_lock);
749 * Give "p" a good chance of killing itself before we
750 * retry to allocate memory unless "p" is current
752 if (killed && !test_thread_flag(TIF_MEMDIE))
753 schedule_timeout_uninterruptible(1);
757 * The pagefault handler calls here because it is out of memory, so kill a
758 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
759 * oom killing is already in progress so do nothing. If a task is found with
760 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
762 void pagefault_out_of_memory(void)
764 if (try_set_system_oom()) {
765 out_of_memory(NULL, 0, 0, NULL);
768 if (!test_thread_flag(TIF_MEMDIE))
769 schedule_timeout_uninterruptible(1);