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>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/oom.h>
41 int sysctl_panic_on_oom;
42 int sysctl_oom_kill_allocating_task;
43 int sysctl_oom_dump_tasks = 1;
44 static DEFINE_SPINLOCK(zone_scan_lock);
47 * compare_swap_oom_score_adj() - compare and swap current's oom_score_adj
48 * @old_val: old oom_score_adj for compare
49 * @new_val: new oom_score_adj for swap
51 * Sets the oom_score_adj value for current to @new_val iff its present value is
52 * @old_val. Usually used to reinstate a previous value to prevent racing with
53 * userspacing tuning the value in the interim.
55 void compare_swap_oom_score_adj(int old_val, int new_val)
57 struct sighand_struct *sighand = current->sighand;
59 spin_lock_irq(&sighand->siglock);
60 if (current->signal->oom_score_adj == old_val)
61 current->signal->oom_score_adj = new_val;
62 trace_oom_score_adj_update(current);
63 spin_unlock_irq(&sighand->siglock);
67 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
68 * @new_val: new oom_score_adj value
70 * Sets the oom_score_adj value for current to @new_val with proper
71 * synchronization and returns the old value. Usually used to temporarily
72 * set a value, save the old value in the caller, and then reinstate it later.
74 int test_set_oom_score_adj(int new_val)
76 struct sighand_struct *sighand = current->sighand;
79 spin_lock_irq(&sighand->siglock);
80 old_val = current->signal->oom_score_adj;
81 current->signal->oom_score_adj = new_val;
82 trace_oom_score_adj_update(current);
83 spin_unlock_irq(&sighand->siglock);
90 * has_intersects_mems_allowed() - check task eligiblity for kill
91 * @tsk: task struct of which task to consider
92 * @mask: nodemask passed to page allocator for mempolicy ooms
94 * Task eligibility is determined by whether or not a candidate task, @tsk,
95 * shares the same mempolicy nodes as current if it is bound by such a policy
96 * and whether or not it has the same set of allowed cpuset nodes.
98 static bool has_intersects_mems_allowed(struct task_struct *tsk,
99 const nodemask_t *mask)
101 struct task_struct *start = tsk;
106 * If this is a mempolicy constrained oom, tsk's
107 * cpuset is irrelevant. Only return true if its
108 * mempolicy intersects current, otherwise it may be
111 if (mempolicy_nodemask_intersects(tsk, mask))
115 * This is not a mempolicy constrained oom, so only
116 * check the mems of tsk's cpuset.
118 if (cpuset_mems_allowed_intersects(current, tsk))
121 } while_each_thread(start, tsk);
126 static bool has_intersects_mems_allowed(struct task_struct *tsk,
127 const nodemask_t *mask)
131 #endif /* CONFIG_NUMA */
134 * The process p may have detached its own ->mm while exiting or through
135 * use_mm(), but one or more of its subthreads may still have a valid
136 * pointer. Return p, or any of its subthreads with a valid ->mm, with
139 struct task_struct *find_lock_task_mm(struct task_struct *p)
141 struct task_struct *t = p;
148 } while_each_thread(p, t);
153 /* return true if the task is not adequate as candidate victim task. */
154 static bool oom_unkillable_task(struct task_struct *p,
155 const struct mem_cgroup *mem, const nodemask_t *nodemask)
157 if (is_global_init(p))
159 if (p->flags & PF_KTHREAD)
162 /* When mem_cgroup_out_of_memory() and p is not member of the group */
163 if (mem && !task_in_mem_cgroup(p, mem))
166 /* p may not have freeable memory in nodemask */
167 if (!has_intersects_mems_allowed(p, nodemask))
174 * oom_badness - heuristic function to determine which candidate task to kill
175 * @p: task struct of which task we should calculate
176 * @totalpages: total present RAM allowed for page allocation
178 * The heuristic for determining which task to kill is made to be as simple and
179 * predictable as possible. The goal is to return the highest value for the
180 * task consuming the most memory to avoid subsequent oom failures.
182 unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
183 const nodemask_t *nodemask, unsigned long totalpages)
187 if (oom_unkillable_task(p, mem, nodemask))
190 p = find_lock_task_mm(p);
194 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
200 * The memory controller may have a limit of 0 bytes, so avoid a divide
201 * by zero, if necessary.
207 * The baseline for the badness score is the proportion of RAM that each
208 * task's rss, pagetable and swap space use.
210 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
211 points += get_mm_counter(p->mm, MM_SWAPENTS);
214 points /= totalpages;
218 * Root processes get 3% bonus, just like the __vm_enough_memory()
219 * implementation used by LSMs.
221 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
225 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
226 * either completely disable oom killing or always prefer a certain
229 points += p->signal->oom_score_adj;
232 * Never return 0 for an eligible task that may be killed since it's
233 * possible that no single user task uses more than 0.1% of memory and
234 * no single admin tasks uses more than 3.0%.
238 return (points < 1000) ? points : 1000;
242 * Determine the type of allocation constraint.
245 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
246 gfp_t gfp_mask, nodemask_t *nodemask,
247 unsigned long *totalpages)
251 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
252 bool cpuset_limited = false;
255 /* Default to all available memory */
256 *totalpages = totalram_pages + total_swap_pages;
259 return CONSTRAINT_NONE;
261 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
262 * to kill current.We have to random task kill in this case.
263 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
265 if (gfp_mask & __GFP_THISNODE)
266 return CONSTRAINT_NONE;
269 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
270 * the page allocator means a mempolicy is in effect. Cpuset policy
271 * is enforced in get_page_from_freelist().
273 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
274 *totalpages = total_swap_pages;
275 for_each_node_mask(nid, *nodemask)
276 *totalpages += node_spanned_pages(nid);
277 return CONSTRAINT_MEMORY_POLICY;
280 /* Check this allocation failure is caused by cpuset's wall function */
281 for_each_zone_zonelist_nodemask(zone, z, zonelist,
282 high_zoneidx, nodemask)
283 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
284 cpuset_limited = true;
286 if (cpuset_limited) {
287 *totalpages = total_swap_pages;
288 for_each_node_mask(nid, cpuset_current_mems_allowed)
289 *totalpages += node_spanned_pages(nid);
290 return CONSTRAINT_CPUSET;
292 return CONSTRAINT_NONE;
295 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
296 gfp_t gfp_mask, nodemask_t *nodemask,
297 unsigned long *totalpages)
299 *totalpages = totalram_pages + total_swap_pages;
300 return CONSTRAINT_NONE;
305 * Simple selection loop. We chose the process with the highest
306 * number of 'points'. We expect the caller will lock the tasklist.
308 * (not docbooked, we don't want this one cluttering up the manual)
310 static struct task_struct *select_bad_process(unsigned int *ppoints,
311 unsigned long totalpages, struct mem_cgroup *mem,
312 const nodemask_t *nodemask)
314 struct task_struct *g, *p;
315 struct task_struct *chosen = NULL;
318 do_each_thread(g, p) {
323 if (oom_unkillable_task(p, mem, nodemask))
327 * This task already has access to memory reserves and is
328 * being killed. Don't allow any other task access to the
331 * Note: this may have a chance of deadlock if it gets
332 * blocked waiting for another task which itself is waiting
333 * for memory. Is there a better alternative?
335 if (test_tsk_thread_flag(p, TIF_MEMDIE)) {
336 if (unlikely(frozen(p)))
338 return ERR_PTR(-1UL);
343 if (p->flags & PF_EXITING) {
345 * If p is the current task and is in the process of
346 * releasing memory, we allow the "kill" to set
347 * TIF_MEMDIE, which will allow it to gain access to
348 * memory reserves. Otherwise, it may stall forever.
350 * The loop isn't broken here, however, in case other
351 * threads are found to have already been oom killed.
358 * If this task is not being ptraced on exit,
359 * then wait for it to finish before killing
360 * some other task unnecessarily.
362 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
363 return ERR_PTR(-1UL);
367 points = oom_badness(p, mem, nodemask, totalpages);
368 if (points > *ppoints) {
372 } while_each_thread(g, p);
378 * dump_tasks - dump current memory state of all system tasks
379 * @mem: current's memory controller, if constrained
380 * @nodemask: nodemask passed to page allocator for mempolicy ooms
382 * Dumps the current memory state of all eligible tasks. Tasks not in the same
383 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
385 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
386 * value, oom_score_adj value, and name.
388 * Call with tasklist_lock read-locked.
390 static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
392 struct task_struct *p;
393 struct task_struct *task;
395 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
396 for_each_process(p) {
397 if (oom_unkillable_task(p, mem, nodemask))
400 task = find_lock_task_mm(p);
403 * This is a kthread or all of p's threads have already
404 * detached their mm's. There's no need to report
405 * them; they can't be oom killed anyway.
410 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
411 task->pid, task_uid(task), task->tgid,
412 task->mm->total_vm, get_mm_rss(task->mm),
413 task_cpu(task), task->signal->oom_adj,
414 task->signal->oom_score_adj, task->comm);
419 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
420 struct mem_cgroup *mem, const nodemask_t *nodemask)
423 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
424 "oom_adj=%d, oom_score_adj=%d\n",
425 current->comm, gfp_mask, order, current->signal->oom_adj,
426 current->signal->oom_score_adj);
427 cpuset_print_task_mems_allowed(current);
428 task_unlock(current);
430 mem_cgroup_print_oom_info(mem, p);
431 show_mem(SHOW_MEM_FILTER_NODES);
432 if (sysctl_oom_dump_tasks)
433 dump_tasks(mem, nodemask);
436 #define K(x) ((x) << (PAGE_SHIFT-10))
437 static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
439 struct task_struct *q;
440 struct mm_struct *mm;
442 p = find_lock_task_mm(p);
446 /* mm cannot be safely dereferenced after task_unlock(p) */
449 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
450 task_pid_nr(p), p->comm, K(p->mm->total_vm),
451 K(get_mm_counter(p->mm, MM_ANONPAGES)),
452 K(get_mm_counter(p->mm, MM_FILEPAGES)));
456 * Kill all user processes sharing p->mm in other thread groups, if any.
457 * They don't get access to memory reserves or a higher scheduler
458 * priority, though, to avoid depletion of all memory or task
459 * starvation. This prevents mm->mmap_sem livelock when an oom killed
460 * task cannot exit because it requires the semaphore and its contended
461 * by another thread trying to allocate memory itself. That thread will
462 * now get access to memory reserves since it has a pending fatal
466 if (q->mm == mm && !same_thread_group(q, p) &&
467 !(q->flags & PF_KTHREAD)) {
468 if (q->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
471 task_lock(q); /* Protect ->comm from prctl() */
472 pr_err("Kill process %d (%s) sharing same memory\n",
473 task_pid_nr(q), q->comm);
475 force_sig(SIGKILL, q);
478 set_tsk_thread_flag(p, TIF_MEMDIE);
479 force_sig(SIGKILL, p);
485 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
486 unsigned int points, unsigned long totalpages,
487 struct mem_cgroup *mem, nodemask_t *nodemask,
490 struct task_struct *victim = p;
491 struct task_struct *child;
492 struct task_struct *t = p;
493 unsigned int victim_points = 0;
495 if (printk_ratelimit())
496 dump_header(p, gfp_mask, order, mem, nodemask);
499 * If the task is already exiting, don't alarm the sysadmin or kill
500 * its children or threads, just set TIF_MEMDIE so it can die quickly
502 if (p->flags & PF_EXITING) {
503 set_tsk_thread_flag(p, TIF_MEMDIE);
508 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
509 message, task_pid_nr(p), p->comm, points);
513 * If any of p's children has a different mm and is eligible for kill,
514 * the one with the highest oom_badness() score is sacrificed for its
515 * parent. This attempts to lose the minimal amount of work done while
516 * still freeing memory.
519 list_for_each_entry(child, &t->children, sibling) {
520 unsigned int child_points;
522 if (child->mm == p->mm)
525 * oom_badness() returns 0 if the thread is unkillable
527 child_points = oom_badness(child, mem, nodemask,
529 if (child_points > victim_points) {
531 victim_points = child_points;
534 } while_each_thread(p, t);
536 return oom_kill_task(victim, mem);
540 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
542 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
543 int order, const nodemask_t *nodemask)
545 if (likely(!sysctl_panic_on_oom))
547 if (sysctl_panic_on_oom != 2) {
549 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
550 * does not panic for cpuset, mempolicy, or memcg allocation
553 if (constraint != CONSTRAINT_NONE)
556 read_lock(&tasklist_lock);
557 dump_header(NULL, gfp_mask, order, NULL, nodemask);
558 read_unlock(&tasklist_lock);
559 panic("Out of memory: %s panic_on_oom is enabled\n",
560 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
563 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
564 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
567 unsigned int points = 0;
568 struct task_struct *p;
571 * If current has a pending SIGKILL, then automatically select it. The
572 * goal is to allow it to allocate so that it may quickly exit and free
575 if (fatal_signal_pending(current)) {
576 set_thread_flag(TIF_MEMDIE);
580 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
581 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
582 read_lock(&tasklist_lock);
584 p = select_bad_process(&points, limit, mem, NULL);
585 if (!p || PTR_ERR(p) == -1UL)
588 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
589 "Memory cgroup out of memory"))
592 read_unlock(&tasklist_lock);
596 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
598 int register_oom_notifier(struct notifier_block *nb)
600 return blocking_notifier_chain_register(&oom_notify_list, nb);
602 EXPORT_SYMBOL_GPL(register_oom_notifier);
604 int unregister_oom_notifier(struct notifier_block *nb)
606 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
608 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
611 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
612 * if a parallel OOM killing is already taking place that includes a zone in
613 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
615 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
621 spin_lock(&zone_scan_lock);
622 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
623 if (zone_is_oom_locked(zone)) {
629 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
631 * Lock each zone in the zonelist under zone_scan_lock so a
632 * parallel invocation of try_set_zonelist_oom() doesn't succeed
635 zone_set_flag(zone, ZONE_OOM_LOCKED);
639 spin_unlock(&zone_scan_lock);
644 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
645 * allocation attempts with zonelists containing them may now recall the OOM
646 * killer, if necessary.
648 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
653 spin_lock(&zone_scan_lock);
654 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
655 zone_clear_flag(zone, ZONE_OOM_LOCKED);
657 spin_unlock(&zone_scan_lock);
661 * Try to acquire the oom killer lock for all system zones. Returns zero if a
662 * parallel oom killing is taking place, otherwise locks all zones and returns
665 static int try_set_system_oom(void)
670 spin_lock(&zone_scan_lock);
671 for_each_populated_zone(zone)
672 if (zone_is_oom_locked(zone)) {
676 for_each_populated_zone(zone)
677 zone_set_flag(zone, ZONE_OOM_LOCKED);
679 spin_unlock(&zone_scan_lock);
684 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
685 * attempts or page faults may now recall the oom killer, if necessary.
687 static void clear_system_oom(void)
691 spin_lock(&zone_scan_lock);
692 for_each_populated_zone(zone)
693 zone_clear_flag(zone, ZONE_OOM_LOCKED);
694 spin_unlock(&zone_scan_lock);
698 * out_of_memory - kill the "best" process when we run out of memory
699 * @zonelist: zonelist pointer
700 * @gfp_mask: memory allocation flags
701 * @order: amount of memory being requested as a power of 2
702 * @nodemask: nodemask passed to page allocator
704 * If we run out of memory, we have the choice between either
705 * killing a random task (bad), letting the system crash (worse)
706 * OR try to be smart about which process to kill. Note that we
707 * don't have to be perfect here, we just have to be good.
709 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
710 int order, nodemask_t *nodemask)
712 const nodemask_t *mpol_mask;
713 struct task_struct *p;
714 unsigned long totalpages;
715 unsigned long freed = 0;
717 enum oom_constraint constraint = CONSTRAINT_NONE;
720 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
722 /* Got some memory back in the last second. */
726 * If current has a pending SIGKILL, then automatically select it. The
727 * goal is to allow it to allocate so that it may quickly exit and free
730 if (fatal_signal_pending(current)) {
731 set_thread_flag(TIF_MEMDIE);
736 * Check if there were limitations on the allocation (only relevant for
737 * NUMA) that may require different handling.
739 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
741 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
742 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
744 read_lock(&tasklist_lock);
745 if (sysctl_oom_kill_allocating_task &&
746 !oom_unkillable_task(current, NULL, nodemask) &&
749 * oom_kill_process() needs tasklist_lock held. If it returns
750 * non-zero, current could not be killed so we must fallback to
753 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
755 "Out of memory (oom_kill_allocating_task)"))
760 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
761 if (PTR_ERR(p) == -1UL)
764 /* Found nothing?!?! Either we hang forever, or we panic. */
766 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
767 read_unlock(&tasklist_lock);
768 panic("Out of memory and no killable processes...\n");
771 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
772 nodemask, "Out of memory"))
776 read_unlock(&tasklist_lock);
779 * Give "p" a good chance of killing itself before we
780 * retry to allocate memory unless "p" is current
782 if (killed && !test_thread_flag(TIF_MEMDIE))
783 schedule_timeout_uninterruptible(1);
787 * The pagefault handler calls here because it is out of memory, so kill a
788 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
789 * oom killing is already in progress so do nothing. If a task is found with
790 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
792 void pagefault_out_of_memory(void)
794 if (try_set_system_oom()) {
795 out_of_memory(NULL, 0, 0, NULL);
798 if (!test_thread_flag(TIF_MEMDIE))
799 schedule_timeout_uninterruptible(1);