Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / mm / oom_kill.c
blobf2e7dfb81eeeeca2a6fe217bbe4d63457bdbc78b
1 /*
2 * linux/mm/oom_kill.c
3 *
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>
21 #include <linux/mm.h>
22 #include <linux/err.h>
23 #include <linux/gfp.h>
24 #include <linux/sched.h>
25 #include <linux/sched/mm.h>
26 #include <linux/sched/coredump.h>
27 #include <linux/sched/task.h>
28 #include <linux/swap.h>
29 #include <linux/timex.h>
30 #include <linux/jiffies.h>
31 #include <linux/cpuset.h>
32 #include <linux/export.h>
33 #include <linux/notifier.h>
34 #include <linux/memcontrol.h>
35 #include <linux/mempolicy.h>
36 #include <linux/security.h>
37 #include <linux/ptrace.h>
38 #include <linux/freezer.h>
39 #include <linux/ftrace.h>
40 #include <linux/ratelimit.h>
41 #include <linux/kthread.h>
42 #include <linux/init.h>
43 #include <linux/mmu_notifier.h>
45 #include <asm/tlb.h>
46 #include "internal.h"
47 #include "slab.h"
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/oom.h>
52 int sysctl_panic_on_oom;
53 int sysctl_oom_kill_allocating_task;
54 int sysctl_oom_dump_tasks = 1;
56 DEFINE_MUTEX(oom_lock);
58 #ifdef CONFIG_NUMA
59 /**
60 * has_intersects_mems_allowed() - check task eligiblity for kill
61 * @start: task struct of which task to consider
62 * @mask: nodemask passed to page allocator for mempolicy ooms
64 * Task eligibility is determined by whether or not a candidate task, @tsk,
65 * shares the same mempolicy nodes as current if it is bound by such a policy
66 * and whether or not it has the same set of allowed cpuset nodes.
68 static bool has_intersects_mems_allowed(struct task_struct *start,
69 const nodemask_t *mask)
71 struct task_struct *tsk;
72 bool ret = false;
74 rcu_read_lock();
75 for_each_thread(start, tsk) {
76 if (mask) {
78 * If this is a mempolicy constrained oom, tsk's
79 * cpuset is irrelevant. Only return true if its
80 * mempolicy intersects current, otherwise it may be
81 * needlessly killed.
83 ret = mempolicy_nodemask_intersects(tsk, mask);
84 } else {
86 * This is not a mempolicy constrained oom, so only
87 * check the mems of tsk's cpuset.
89 ret = cpuset_mems_allowed_intersects(current, tsk);
91 if (ret)
92 break;
94 rcu_read_unlock();
96 return ret;
98 #else
99 static bool has_intersects_mems_allowed(struct task_struct *tsk,
100 const nodemask_t *mask)
102 return true;
104 #endif /* CONFIG_NUMA */
107 * The process p may have detached its own ->mm while exiting or through
108 * use_mm(), but one or more of its subthreads may still have a valid
109 * pointer. Return p, or any of its subthreads with a valid ->mm, with
110 * task_lock() held.
112 struct task_struct *find_lock_task_mm(struct task_struct *p)
114 struct task_struct *t;
116 rcu_read_lock();
118 for_each_thread(p, t) {
119 task_lock(t);
120 if (likely(t->mm))
121 goto found;
122 task_unlock(t);
124 t = NULL;
125 found:
126 rcu_read_unlock();
128 return t;
132 * order == -1 means the oom kill is required by sysrq, otherwise only
133 * for display purposes.
135 static inline bool is_sysrq_oom(struct oom_control *oc)
137 return oc->order == -1;
140 static inline bool is_memcg_oom(struct oom_control *oc)
142 return oc->memcg != NULL;
145 /* return true if the task is not adequate as candidate victim task. */
146 static bool oom_unkillable_task(struct task_struct *p,
147 struct mem_cgroup *memcg, const nodemask_t *nodemask)
149 if (is_global_init(p))
150 return true;
151 if (p->flags & PF_KTHREAD)
152 return true;
154 /* When mem_cgroup_out_of_memory() and p is not member of the group */
155 if (memcg && !task_in_mem_cgroup(p, memcg))
156 return true;
158 /* p may not have freeable memory in nodemask */
159 if (!has_intersects_mems_allowed(p, nodemask))
160 return true;
162 return false;
166 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
167 * than all user memory (LRU pages)
169 static bool is_dump_unreclaim_slabs(void)
171 unsigned long nr_lru;
173 nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
174 global_node_page_state(NR_INACTIVE_ANON) +
175 global_node_page_state(NR_ACTIVE_FILE) +
176 global_node_page_state(NR_INACTIVE_FILE) +
177 global_node_page_state(NR_ISOLATED_ANON) +
178 global_node_page_state(NR_ISOLATED_FILE) +
179 global_node_page_state(NR_UNEVICTABLE);
181 return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
185 * oom_badness - heuristic function to determine which candidate task to kill
186 * @p: task struct of which task we should calculate
187 * @totalpages: total present RAM allowed for page allocation
189 * The heuristic for determining which task to kill is made to be as simple and
190 * predictable as possible. The goal is to return the highest value for the
191 * task consuming the most memory to avoid subsequent oom failures.
193 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
194 const nodemask_t *nodemask, unsigned long totalpages)
196 long points;
197 long adj;
199 if (oom_unkillable_task(p, memcg, nodemask))
200 return 0;
202 p = find_lock_task_mm(p);
203 if (!p)
204 return 0;
207 * Do not even consider tasks which are explicitly marked oom
208 * unkillable or have been already oom reaped or the are in
209 * the middle of vfork
211 adj = (long)p->signal->oom_score_adj;
212 if (adj == OOM_SCORE_ADJ_MIN ||
213 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
214 in_vfork(p)) {
215 task_unlock(p);
216 return 0;
220 * The baseline for the badness score is the proportion of RAM that each
221 * task's rss, pagetable and swap space use.
223 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
224 mm_pgtables_bytes(p->mm) / PAGE_SIZE;
225 task_unlock(p);
228 * Root processes get 3% bonus, just like the __vm_enough_memory()
229 * implementation used by LSMs.
231 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
232 points -= (points * 3) / 100;
234 /* Normalize to oom_score_adj units */
235 adj *= totalpages / 1000;
236 points += adj;
239 * Never return 0 for an eligible task regardless of the root bonus and
240 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
242 return points > 0 ? points : 1;
245 enum oom_constraint {
246 CONSTRAINT_NONE,
247 CONSTRAINT_CPUSET,
248 CONSTRAINT_MEMORY_POLICY,
249 CONSTRAINT_MEMCG,
253 * Determine the type of allocation constraint.
255 static enum oom_constraint constrained_alloc(struct oom_control *oc)
257 struct zone *zone;
258 struct zoneref *z;
259 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
260 bool cpuset_limited = false;
261 int nid;
263 if (is_memcg_oom(oc)) {
264 oc->totalpages = mem_cgroup_get_limit(oc->memcg) ?: 1;
265 return CONSTRAINT_MEMCG;
268 /* Default to all available memory */
269 oc->totalpages = totalram_pages + total_swap_pages;
271 if (!IS_ENABLED(CONFIG_NUMA))
272 return CONSTRAINT_NONE;
274 if (!oc->zonelist)
275 return CONSTRAINT_NONE;
277 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
278 * to kill current.We have to random task kill in this case.
279 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
281 if (oc->gfp_mask & __GFP_THISNODE)
282 return CONSTRAINT_NONE;
285 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
286 * the page allocator means a mempolicy is in effect. Cpuset policy
287 * is enforced in get_page_from_freelist().
289 if (oc->nodemask &&
290 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
291 oc->totalpages = total_swap_pages;
292 for_each_node_mask(nid, *oc->nodemask)
293 oc->totalpages += node_spanned_pages(nid);
294 return CONSTRAINT_MEMORY_POLICY;
297 /* Check this allocation failure is caused by cpuset's wall function */
298 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
299 high_zoneidx, oc->nodemask)
300 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
301 cpuset_limited = true;
303 if (cpuset_limited) {
304 oc->totalpages = total_swap_pages;
305 for_each_node_mask(nid, cpuset_current_mems_allowed)
306 oc->totalpages += node_spanned_pages(nid);
307 return CONSTRAINT_CPUSET;
309 return CONSTRAINT_NONE;
312 static int oom_evaluate_task(struct task_struct *task, void *arg)
314 struct oom_control *oc = arg;
315 unsigned long points;
317 if (oom_unkillable_task(task, NULL, oc->nodemask))
318 goto next;
321 * This task already has access to memory reserves and is being killed.
322 * Don't allow any other task to have access to the reserves unless
323 * the task has MMF_OOM_SKIP because chances that it would release
324 * any memory is quite low.
326 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
327 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
328 goto next;
329 goto abort;
333 * If task is allocating a lot of memory and has been marked to be
334 * killed first if it triggers an oom, then select it.
336 if (oom_task_origin(task)) {
337 points = ULONG_MAX;
338 goto select;
341 points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
342 if (!points || points < oc->chosen_points)
343 goto next;
345 /* Prefer thread group leaders for display purposes */
346 if (points == oc->chosen_points && thread_group_leader(oc->chosen))
347 goto next;
348 select:
349 if (oc->chosen)
350 put_task_struct(oc->chosen);
351 get_task_struct(task);
352 oc->chosen = task;
353 oc->chosen_points = points;
354 next:
355 return 0;
356 abort:
357 if (oc->chosen)
358 put_task_struct(oc->chosen);
359 oc->chosen = (void *)-1UL;
360 return 1;
364 * Simple selection loop. We choose the process with the highest number of
365 * 'points'. In case scan was aborted, oc->chosen is set to -1.
367 static void select_bad_process(struct oom_control *oc)
369 if (is_memcg_oom(oc))
370 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
371 else {
372 struct task_struct *p;
374 rcu_read_lock();
375 for_each_process(p)
376 if (oom_evaluate_task(p, oc))
377 break;
378 rcu_read_unlock();
381 oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages;
385 * dump_tasks - dump current memory state of all system tasks
386 * @memcg: current's memory controller, if constrained
387 * @nodemask: nodemask passed to page allocator for mempolicy ooms
389 * Dumps the current memory state of all eligible tasks. Tasks not in the same
390 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
391 * are not shown.
392 * State information includes task's pid, uid, tgid, vm size, rss,
393 * pgtables_bytes, swapents, oom_score_adj value, and name.
395 static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
397 struct task_struct *p;
398 struct task_struct *task;
400 pr_info("[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name\n");
401 rcu_read_lock();
402 for_each_process(p) {
403 if (oom_unkillable_task(p, memcg, nodemask))
404 continue;
406 task = find_lock_task_mm(p);
407 if (!task) {
409 * This is a kthread or all of p's threads have already
410 * detached their mm's. There's no need to report
411 * them; they can't be oom killed anyway.
413 continue;
416 pr_info("[%5d] %5d %5d %8lu %8lu %8ld %8lu %5hd %s\n",
417 task->pid, from_kuid(&init_user_ns, task_uid(task)),
418 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
419 mm_pgtables_bytes(task->mm),
420 get_mm_counter(task->mm, MM_SWAPENTS),
421 task->signal->oom_score_adj, task->comm);
422 task_unlock(task);
424 rcu_read_unlock();
427 static void dump_header(struct oom_control *oc, struct task_struct *p)
429 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n",
430 current->comm, oc->gfp_mask, &oc->gfp_mask,
431 nodemask_pr_args(oc->nodemask), oc->order,
432 current->signal->oom_score_adj);
433 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
434 pr_warn("COMPACTION is disabled!!!\n");
436 cpuset_print_current_mems_allowed();
437 dump_stack();
438 if (is_memcg_oom(oc))
439 mem_cgroup_print_oom_info(oc->memcg, p);
440 else {
441 show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
442 if (is_dump_unreclaim_slabs())
443 dump_unreclaimable_slab();
445 if (sysctl_oom_dump_tasks)
446 dump_tasks(oc->memcg, oc->nodemask);
450 * Number of OOM victims in flight
452 static atomic_t oom_victims = ATOMIC_INIT(0);
453 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
455 static bool oom_killer_disabled __read_mostly;
457 #define K(x) ((x) << (PAGE_SHIFT-10))
460 * task->mm can be NULL if the task is the exited group leader. So to
461 * determine whether the task is using a particular mm, we examine all the
462 * task's threads: if one of those is using this mm then this task was also
463 * using it.
465 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
467 struct task_struct *t;
469 for_each_thread(p, t) {
470 struct mm_struct *t_mm = READ_ONCE(t->mm);
471 if (t_mm)
472 return t_mm == mm;
474 return false;
478 #ifdef CONFIG_MMU
480 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
481 * victim (if that is possible) to help the OOM killer to move on.
483 static struct task_struct *oom_reaper_th;
484 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
485 static struct task_struct *oom_reaper_list;
486 static DEFINE_SPINLOCK(oom_reaper_lock);
488 static bool __oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
490 struct mmu_gather tlb;
491 struct vm_area_struct *vma;
492 bool ret = true;
495 * We have to make sure to not race with the victim exit path
496 * and cause premature new oom victim selection:
497 * __oom_reap_task_mm exit_mm
498 * mmget_not_zero
499 * mmput
500 * atomic_dec_and_test
501 * exit_oom_victim
502 * [...]
503 * out_of_memory
504 * select_bad_process
505 * # no TIF_MEMDIE task selects new victim
506 * unmap_page_range # frees some memory
508 mutex_lock(&oom_lock);
510 if (!down_read_trylock(&mm->mmap_sem)) {
511 ret = false;
512 trace_skip_task_reaping(tsk->pid);
513 goto unlock_oom;
517 * If the mm has invalidate_{start,end}() notifiers that could block,
518 * sleep to give the oom victim some more time.
519 * TODO: we really want to get rid of this ugly hack and make sure that
520 * notifiers cannot block for unbounded amount of time
522 if (mm_has_blockable_invalidate_notifiers(mm)) {
523 up_read(&mm->mmap_sem);
524 schedule_timeout_idle(HZ);
525 goto unlock_oom;
529 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
530 * work on the mm anymore. The check for MMF_OOM_SKIP must run
531 * under mmap_sem for reading because it serializes against the
532 * down_write();up_write() cycle in exit_mmap().
534 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
535 up_read(&mm->mmap_sem);
536 trace_skip_task_reaping(tsk->pid);
537 goto unlock_oom;
540 trace_start_task_reaping(tsk->pid);
543 * Tell all users of get_user/copy_from_user etc... that the content
544 * is no longer stable. No barriers really needed because unmapping
545 * should imply barriers already and the reader would hit a page fault
546 * if it stumbled over a reaped memory.
548 set_bit(MMF_UNSTABLE, &mm->flags);
550 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
551 if (!can_madv_dontneed_vma(vma))
552 continue;
555 * Only anonymous pages have a good chance to be dropped
556 * without additional steps which we cannot afford as we
557 * are OOM already.
559 * We do not even care about fs backed pages because all
560 * which are reclaimable have already been reclaimed and
561 * we do not want to block exit_mmap by keeping mm ref
562 * count elevated without a good reason.
564 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
565 const unsigned long start = vma->vm_start;
566 const unsigned long end = vma->vm_end;
568 tlb_gather_mmu(&tlb, mm, start, end);
569 mmu_notifier_invalidate_range_start(mm, start, end);
570 unmap_page_range(&tlb, vma, start, end, NULL);
571 mmu_notifier_invalidate_range_end(mm, start, end);
572 tlb_finish_mmu(&tlb, start, end);
575 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
576 task_pid_nr(tsk), tsk->comm,
577 K(get_mm_counter(mm, MM_ANONPAGES)),
578 K(get_mm_counter(mm, MM_FILEPAGES)),
579 K(get_mm_counter(mm, MM_SHMEMPAGES)));
580 up_read(&mm->mmap_sem);
582 trace_finish_task_reaping(tsk->pid);
583 unlock_oom:
584 mutex_unlock(&oom_lock);
585 return ret;
588 #define MAX_OOM_REAP_RETRIES 10
589 static void oom_reap_task(struct task_struct *tsk)
591 int attempts = 0;
592 struct mm_struct *mm = tsk->signal->oom_mm;
594 /* Retry the down_read_trylock(mmap_sem) a few times */
595 while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task_mm(tsk, mm))
596 schedule_timeout_idle(HZ/10);
598 if (attempts <= MAX_OOM_REAP_RETRIES)
599 goto done;
602 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
603 task_pid_nr(tsk), tsk->comm);
604 debug_show_all_locks();
606 done:
607 tsk->oom_reaper_list = NULL;
610 * Hide this mm from OOM killer because it has been either reaped or
611 * somebody can't call up_write(mmap_sem).
613 set_bit(MMF_OOM_SKIP, &mm->flags);
615 /* Drop a reference taken by wake_oom_reaper */
616 put_task_struct(tsk);
619 static int oom_reaper(void *unused)
621 while (true) {
622 struct task_struct *tsk = NULL;
624 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
625 spin_lock(&oom_reaper_lock);
626 if (oom_reaper_list != NULL) {
627 tsk = oom_reaper_list;
628 oom_reaper_list = tsk->oom_reaper_list;
630 spin_unlock(&oom_reaper_lock);
632 if (tsk)
633 oom_reap_task(tsk);
636 return 0;
639 static void wake_oom_reaper(struct task_struct *tsk)
641 /* tsk is already queued? */
642 if (tsk == oom_reaper_list || tsk->oom_reaper_list)
643 return;
645 get_task_struct(tsk);
647 spin_lock(&oom_reaper_lock);
648 tsk->oom_reaper_list = oom_reaper_list;
649 oom_reaper_list = tsk;
650 spin_unlock(&oom_reaper_lock);
651 trace_wake_reaper(tsk->pid);
652 wake_up(&oom_reaper_wait);
655 static int __init oom_init(void)
657 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
658 return 0;
660 subsys_initcall(oom_init)
661 #else
662 static inline void wake_oom_reaper(struct task_struct *tsk)
665 #endif /* CONFIG_MMU */
668 * mark_oom_victim - mark the given task as OOM victim
669 * @tsk: task to mark
671 * Has to be called with oom_lock held and never after
672 * oom has been disabled already.
674 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
675 * under task_lock or operate on the current).
677 static void mark_oom_victim(struct task_struct *tsk)
679 struct mm_struct *mm = tsk->mm;
681 WARN_ON(oom_killer_disabled);
682 /* OOM killer might race with memcg OOM */
683 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
684 return;
686 /* oom_mm is bound to the signal struct life time. */
687 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
688 mmgrab(tsk->signal->oom_mm);
689 set_bit(MMF_OOM_VICTIM, &mm->flags);
693 * Make sure that the task is woken up from uninterruptible sleep
694 * if it is frozen because OOM killer wouldn't be able to free
695 * any memory and livelock. freezing_slow_path will tell the freezer
696 * that TIF_MEMDIE tasks should be ignored.
698 __thaw_task(tsk);
699 atomic_inc(&oom_victims);
700 trace_mark_victim(tsk->pid);
704 * exit_oom_victim - note the exit of an OOM victim
706 void exit_oom_victim(void)
708 clear_thread_flag(TIF_MEMDIE);
710 if (!atomic_dec_return(&oom_victims))
711 wake_up_all(&oom_victims_wait);
715 * oom_killer_enable - enable OOM killer
717 void oom_killer_enable(void)
719 oom_killer_disabled = false;
720 pr_info("OOM killer enabled.\n");
724 * oom_killer_disable - disable OOM killer
725 * @timeout: maximum timeout to wait for oom victims in jiffies
727 * Forces all page allocations to fail rather than trigger OOM killer.
728 * Will block and wait until all OOM victims are killed or the given
729 * timeout expires.
731 * The function cannot be called when there are runnable user tasks because
732 * the userspace would see unexpected allocation failures as a result. Any
733 * new usage of this function should be consulted with MM people.
735 * Returns true if successful and false if the OOM killer cannot be
736 * disabled.
738 bool oom_killer_disable(signed long timeout)
740 signed long ret;
743 * Make sure to not race with an ongoing OOM killer. Check that the
744 * current is not killed (possibly due to sharing the victim's memory).
746 if (mutex_lock_killable(&oom_lock))
747 return false;
748 oom_killer_disabled = true;
749 mutex_unlock(&oom_lock);
751 ret = wait_event_interruptible_timeout(oom_victims_wait,
752 !atomic_read(&oom_victims), timeout);
753 if (ret <= 0) {
754 oom_killer_enable();
755 return false;
757 pr_info("OOM killer disabled.\n");
759 return true;
762 static inline bool __task_will_free_mem(struct task_struct *task)
764 struct signal_struct *sig = task->signal;
767 * A coredumping process may sleep for an extended period in exit_mm(),
768 * so the oom killer cannot assume that the process will promptly exit
769 * and release memory.
771 if (sig->flags & SIGNAL_GROUP_COREDUMP)
772 return false;
774 if (sig->flags & SIGNAL_GROUP_EXIT)
775 return true;
777 if (thread_group_empty(task) && (task->flags & PF_EXITING))
778 return true;
780 return false;
784 * Checks whether the given task is dying or exiting and likely to
785 * release its address space. This means that all threads and processes
786 * sharing the same mm have to be killed or exiting.
787 * Caller has to make sure that task->mm is stable (hold task_lock or
788 * it operates on the current).
790 static bool task_will_free_mem(struct task_struct *task)
792 struct mm_struct *mm = task->mm;
793 struct task_struct *p;
794 bool ret = true;
797 * Skip tasks without mm because it might have passed its exit_mm and
798 * exit_oom_victim. oom_reaper could have rescued that but do not rely
799 * on that for now. We can consider find_lock_task_mm in future.
801 if (!mm)
802 return false;
804 if (!__task_will_free_mem(task))
805 return false;
808 * This task has already been drained by the oom reaper so there are
809 * only small chances it will free some more
811 if (test_bit(MMF_OOM_SKIP, &mm->flags))
812 return false;
814 if (atomic_read(&mm->mm_users) <= 1)
815 return true;
818 * Make sure that all tasks which share the mm with the given tasks
819 * are dying as well to make sure that a) nobody pins its mm and
820 * b) the task is also reapable by the oom reaper.
822 rcu_read_lock();
823 for_each_process(p) {
824 if (!process_shares_mm(p, mm))
825 continue;
826 if (same_thread_group(task, p))
827 continue;
828 ret = __task_will_free_mem(p);
829 if (!ret)
830 break;
832 rcu_read_unlock();
834 return ret;
837 static void oom_kill_process(struct oom_control *oc, const char *message)
839 struct task_struct *p = oc->chosen;
840 unsigned int points = oc->chosen_points;
841 struct task_struct *victim = p;
842 struct task_struct *child;
843 struct task_struct *t;
844 struct mm_struct *mm;
845 unsigned int victim_points = 0;
846 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
847 DEFAULT_RATELIMIT_BURST);
848 bool can_oom_reap = true;
851 * If the task is already exiting, don't alarm the sysadmin or kill
852 * its children or threads, just give it access to memory reserves
853 * so it can die quickly
855 task_lock(p);
856 if (task_will_free_mem(p)) {
857 mark_oom_victim(p);
858 wake_oom_reaper(p);
859 task_unlock(p);
860 put_task_struct(p);
861 return;
863 task_unlock(p);
865 if (__ratelimit(&oom_rs))
866 dump_header(oc, p);
868 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
869 message, task_pid_nr(p), p->comm, points);
872 * If any of p's children has a different mm and is eligible for kill,
873 * the one with the highest oom_badness() score is sacrificed for its
874 * parent. This attempts to lose the minimal amount of work done while
875 * still freeing memory.
877 read_lock(&tasklist_lock);
878 for_each_thread(p, t) {
879 list_for_each_entry(child, &t->children, sibling) {
880 unsigned int child_points;
882 if (process_shares_mm(child, p->mm))
883 continue;
885 * oom_badness() returns 0 if the thread is unkillable
887 child_points = oom_badness(child,
888 oc->memcg, oc->nodemask, oc->totalpages);
889 if (child_points > victim_points) {
890 put_task_struct(victim);
891 victim = child;
892 victim_points = child_points;
893 get_task_struct(victim);
897 read_unlock(&tasklist_lock);
899 p = find_lock_task_mm(victim);
900 if (!p) {
901 put_task_struct(victim);
902 return;
903 } else if (victim != p) {
904 get_task_struct(p);
905 put_task_struct(victim);
906 victim = p;
909 /* Get a reference to safely compare mm after task_unlock(victim) */
910 mm = victim->mm;
911 mmgrab(mm);
913 /* Raise event before sending signal: task reaper must see this */
914 count_vm_event(OOM_KILL);
915 count_memcg_event_mm(mm, OOM_KILL);
918 * We should send SIGKILL before granting access to memory reserves
919 * in order to prevent the OOM victim from depleting the memory
920 * reserves from the user space under its control.
922 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
923 mark_oom_victim(victim);
924 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
925 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
926 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
927 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
928 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
929 task_unlock(victim);
932 * Kill all user processes sharing victim->mm in other thread groups, if
933 * any. They don't get access to memory reserves, though, to avoid
934 * depletion of all memory. This prevents mm->mmap_sem livelock when an
935 * oom killed thread cannot exit because it requires the semaphore and
936 * its contended by another thread trying to allocate memory itself.
937 * That thread will now get access to memory reserves since it has a
938 * pending fatal signal.
940 rcu_read_lock();
941 for_each_process(p) {
942 if (!process_shares_mm(p, mm))
943 continue;
944 if (same_thread_group(p, victim))
945 continue;
946 if (is_global_init(p)) {
947 can_oom_reap = false;
948 set_bit(MMF_OOM_SKIP, &mm->flags);
949 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
950 task_pid_nr(victim), victim->comm,
951 task_pid_nr(p), p->comm);
952 continue;
955 * No use_mm() user needs to read from the userspace so we are
956 * ok to reap it.
958 if (unlikely(p->flags & PF_KTHREAD))
959 continue;
960 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
962 rcu_read_unlock();
964 if (can_oom_reap)
965 wake_oom_reaper(victim);
967 mmdrop(mm);
968 put_task_struct(victim);
970 #undef K
973 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
975 static void check_panic_on_oom(struct oom_control *oc,
976 enum oom_constraint constraint)
978 if (likely(!sysctl_panic_on_oom))
979 return;
980 if (sysctl_panic_on_oom != 2) {
982 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
983 * does not panic for cpuset, mempolicy, or memcg allocation
984 * failures.
986 if (constraint != CONSTRAINT_NONE)
987 return;
989 /* Do not panic for oom kills triggered by sysrq */
990 if (is_sysrq_oom(oc))
991 return;
992 dump_header(oc, NULL);
993 panic("Out of memory: %s panic_on_oom is enabled\n",
994 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
997 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
999 int register_oom_notifier(struct notifier_block *nb)
1001 return blocking_notifier_chain_register(&oom_notify_list, nb);
1003 EXPORT_SYMBOL_GPL(register_oom_notifier);
1005 int unregister_oom_notifier(struct notifier_block *nb)
1007 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1009 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1012 * out_of_memory - kill the "best" process when we run out of memory
1013 * @oc: pointer to struct oom_control
1015 * If we run out of memory, we have the choice between either
1016 * killing a random task (bad), letting the system crash (worse)
1017 * OR try to be smart about which process to kill. Note that we
1018 * don't have to be perfect here, we just have to be good.
1020 bool out_of_memory(struct oom_control *oc)
1022 unsigned long freed = 0;
1023 enum oom_constraint constraint = CONSTRAINT_NONE;
1025 if (oom_killer_disabled)
1026 return false;
1028 if (!is_memcg_oom(oc)) {
1029 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1030 if (freed > 0)
1031 /* Got some memory back in the last second. */
1032 return true;
1036 * If current has a pending SIGKILL or is exiting, then automatically
1037 * select it. The goal is to allow it to allocate so that it may
1038 * quickly exit and free its memory.
1040 if (task_will_free_mem(current)) {
1041 mark_oom_victim(current);
1042 wake_oom_reaper(current);
1043 return true;
1047 * The OOM killer does not compensate for IO-less reclaim.
1048 * pagefault_out_of_memory lost its gfp context so we have to
1049 * make sure exclude 0 mask - all other users should have at least
1050 * ___GFP_DIRECT_RECLAIM to get here.
1052 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS))
1053 return true;
1056 * Check if there were limitations on the allocation (only relevant for
1057 * NUMA and memcg) that may require different handling.
1059 constraint = constrained_alloc(oc);
1060 if (constraint != CONSTRAINT_MEMORY_POLICY)
1061 oc->nodemask = NULL;
1062 check_panic_on_oom(oc, constraint);
1064 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1065 current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) &&
1066 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1067 get_task_struct(current);
1068 oc->chosen = current;
1069 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1070 return true;
1073 select_bad_process(oc);
1074 /* Found nothing?!?! Either we hang forever, or we panic. */
1075 if (!oc->chosen && !is_sysrq_oom(oc) && !is_memcg_oom(oc)) {
1076 dump_header(oc, NULL);
1077 panic("Out of memory and no killable processes...\n");
1079 if (oc->chosen && oc->chosen != (void *)-1UL) {
1080 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1081 "Memory cgroup out of memory");
1083 * Give the killed process a good chance to exit before trying
1084 * to allocate memory again.
1086 schedule_timeout_killable(1);
1088 return !!oc->chosen;
1092 * The pagefault handler calls here because it is out of memory, so kill a
1093 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1094 * killing is already in progress so do nothing.
1096 void pagefault_out_of_memory(void)
1098 struct oom_control oc = {
1099 .zonelist = NULL,
1100 .nodemask = NULL,
1101 .memcg = NULL,
1102 .gfp_mask = 0,
1103 .order = 0,
1106 if (mem_cgroup_oom_synchronize(true))
1107 return;
1109 if (!mutex_trylock(&oom_lock))
1110 return;
1111 out_of_memory(&oc);
1112 mutex_unlock(&oom_lock);