]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/oom_kill.c
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[mirror_ubuntu-artful-kernel.git] / mm / oom_kill.c
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
9 *
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.
13 *
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.
18 */
19
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/swap.h>
26 #include <linux/timex.h>
27 #include <linux/jiffies.h>
28 #include <linux/cpuset.h>
29 #include <linux/export.h>
30 #include <linux/notifier.h>
31 #include <linux/memcontrol.h>
32 #include <linux/mempolicy.h>
33 #include <linux/security.h>
34 #include <linux/ptrace.h>
35 #include <linux/freezer.h>
36 #include <linux/ftrace.h>
37 #include <linux/ratelimit.h>
38 #include <linux/kthread.h>
39 #include <linux/init.h>
40
41 #include <asm/tlb.h>
42 #include "internal.h"
43
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/oom.h>
46
47 int sysctl_panic_on_oom;
48 int sysctl_oom_kill_allocating_task;
49 int sysctl_oom_dump_tasks = 1;
50
51 DEFINE_MUTEX(oom_lock);
52
53 #ifdef CONFIG_NUMA
54 /**
55 * has_intersects_mems_allowed() - check task eligiblity for kill
56 * @start: task struct of which task to consider
57 * @mask: nodemask passed to page allocator for mempolicy ooms
58 *
59 * Task eligibility is determined by whether or not a candidate task, @tsk,
60 * shares the same mempolicy nodes as current if it is bound by such a policy
61 * and whether or not it has the same set of allowed cpuset nodes.
62 */
63 static bool has_intersects_mems_allowed(struct task_struct *start,
64 const nodemask_t *mask)
65 {
66 struct task_struct *tsk;
67 bool ret = false;
68
69 rcu_read_lock();
70 for_each_thread(start, tsk) {
71 if (mask) {
72 /*
73 * If this is a mempolicy constrained oom, tsk's
74 * cpuset is irrelevant. Only return true if its
75 * mempolicy intersects current, otherwise it may be
76 * needlessly killed.
77 */
78 ret = mempolicy_nodemask_intersects(tsk, mask);
79 } else {
80 /*
81 * This is not a mempolicy constrained oom, so only
82 * check the mems of tsk's cpuset.
83 */
84 ret = cpuset_mems_allowed_intersects(current, tsk);
85 }
86 if (ret)
87 break;
88 }
89 rcu_read_unlock();
90
91 return ret;
92 }
93 #else
94 static bool has_intersects_mems_allowed(struct task_struct *tsk,
95 const nodemask_t *mask)
96 {
97 return true;
98 }
99 #endif /* CONFIG_NUMA */
100
101 /*
102 * The process p may have detached its own ->mm while exiting or through
103 * use_mm(), but one or more of its subthreads may still have a valid
104 * pointer. Return p, or any of its subthreads with a valid ->mm, with
105 * task_lock() held.
106 */
107 struct task_struct *find_lock_task_mm(struct task_struct *p)
108 {
109 struct task_struct *t;
110
111 rcu_read_lock();
112
113 for_each_thread(p, t) {
114 task_lock(t);
115 if (likely(t->mm))
116 goto found;
117 task_unlock(t);
118 }
119 t = NULL;
120 found:
121 rcu_read_unlock();
122
123 return t;
124 }
125
126 /*
127 * order == -1 means the oom kill is required by sysrq, otherwise only
128 * for display purposes.
129 */
130 static inline bool is_sysrq_oom(struct oom_control *oc)
131 {
132 return oc->order == -1;
133 }
134
135 /* return true if the task is not adequate as candidate victim task. */
136 static bool oom_unkillable_task(struct task_struct *p,
137 struct mem_cgroup *memcg, const nodemask_t *nodemask)
138 {
139 if (is_global_init(p))
140 return true;
141 if (p->flags & PF_KTHREAD)
142 return true;
143
144 /* When mem_cgroup_out_of_memory() and p is not member of the group */
145 if (memcg && !task_in_mem_cgroup(p, memcg))
146 return true;
147
148 /* p may not have freeable memory in nodemask */
149 if (!has_intersects_mems_allowed(p, nodemask))
150 return true;
151
152 return false;
153 }
154
155 /**
156 * oom_badness - heuristic function to determine which candidate task to kill
157 * @p: task struct of which task we should calculate
158 * @totalpages: total present RAM allowed for page allocation
159 *
160 * The heuristic for determining which task to kill is made to be as simple and
161 * predictable as possible. The goal is to return the highest value for the
162 * task consuming the most memory to avoid subsequent oom failures.
163 */
164 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
165 const nodemask_t *nodemask, unsigned long totalpages)
166 {
167 long points;
168 long adj;
169
170 if (oom_unkillable_task(p, memcg, nodemask))
171 return 0;
172
173 p = find_lock_task_mm(p);
174 if (!p)
175 return 0;
176
177 /*
178 * Do not even consider tasks which are explicitly marked oom
179 * unkillable or have been already oom reaped.
180 */
181 adj = (long)p->signal->oom_score_adj;
182 if (adj == OOM_SCORE_ADJ_MIN ||
183 test_bit(MMF_OOM_REAPED, &p->mm->flags)) {
184 task_unlock(p);
185 return 0;
186 }
187
188 /*
189 * The baseline for the badness score is the proportion of RAM that each
190 * task's rss, pagetable and swap space use.
191 */
192 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
193 atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
194 task_unlock(p);
195
196 /*
197 * Root processes get 3% bonus, just like the __vm_enough_memory()
198 * implementation used by LSMs.
199 */
200 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
201 points -= (points * 3) / 100;
202
203 /* Normalize to oom_score_adj units */
204 adj *= totalpages / 1000;
205 points += adj;
206
207 /*
208 * Never return 0 for an eligible task regardless of the root bonus and
209 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
210 */
211 return points > 0 ? points : 1;
212 }
213
214 /*
215 * Determine the type of allocation constraint.
216 */
217 #ifdef CONFIG_NUMA
218 static enum oom_constraint constrained_alloc(struct oom_control *oc,
219 unsigned long *totalpages)
220 {
221 struct zone *zone;
222 struct zoneref *z;
223 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
224 bool cpuset_limited = false;
225 int nid;
226
227 /* Default to all available memory */
228 *totalpages = totalram_pages + total_swap_pages;
229
230 if (!oc->zonelist)
231 return CONSTRAINT_NONE;
232 /*
233 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
234 * to kill current.We have to random task kill in this case.
235 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
236 */
237 if (oc->gfp_mask & __GFP_THISNODE)
238 return CONSTRAINT_NONE;
239
240 /*
241 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
242 * the page allocator means a mempolicy is in effect. Cpuset policy
243 * is enforced in get_page_from_freelist().
244 */
245 if (oc->nodemask &&
246 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
247 *totalpages = total_swap_pages;
248 for_each_node_mask(nid, *oc->nodemask)
249 *totalpages += node_spanned_pages(nid);
250 return CONSTRAINT_MEMORY_POLICY;
251 }
252
253 /* Check this allocation failure is caused by cpuset's wall function */
254 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
255 high_zoneidx, oc->nodemask)
256 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
257 cpuset_limited = true;
258
259 if (cpuset_limited) {
260 *totalpages = total_swap_pages;
261 for_each_node_mask(nid, cpuset_current_mems_allowed)
262 *totalpages += node_spanned_pages(nid);
263 return CONSTRAINT_CPUSET;
264 }
265 return CONSTRAINT_NONE;
266 }
267 #else
268 static enum oom_constraint constrained_alloc(struct oom_control *oc,
269 unsigned long *totalpages)
270 {
271 *totalpages = totalram_pages + total_swap_pages;
272 return CONSTRAINT_NONE;
273 }
274 #endif
275
276 enum oom_scan_t oom_scan_process_thread(struct oom_control *oc,
277 struct task_struct *task, unsigned long totalpages)
278 {
279 if (oom_unkillable_task(task, NULL, oc->nodemask))
280 return OOM_SCAN_CONTINUE;
281
282 /*
283 * This task already has access to memory reserves and is being killed.
284 * Don't allow any other task to have access to the reserves.
285 */
286 if (!is_sysrq_oom(oc) && atomic_read(&task->signal->oom_victims))
287 return OOM_SCAN_ABORT;
288
289 /*
290 * If task is allocating a lot of memory and has been marked to be
291 * killed first if it triggers an oom, then select it.
292 */
293 if (oom_task_origin(task))
294 return OOM_SCAN_SELECT;
295
296 return OOM_SCAN_OK;
297 }
298
299 /*
300 * Simple selection loop. We chose the process with the highest
301 * number of 'points'. Returns -1 on scan abort.
302 */
303 static struct task_struct *select_bad_process(struct oom_control *oc,
304 unsigned int *ppoints, unsigned long totalpages)
305 {
306 struct task_struct *p;
307 struct task_struct *chosen = NULL;
308 unsigned long chosen_points = 0;
309
310 rcu_read_lock();
311 for_each_process(p) {
312 unsigned int points;
313
314 switch (oom_scan_process_thread(oc, p, totalpages)) {
315 case OOM_SCAN_SELECT:
316 chosen = p;
317 chosen_points = ULONG_MAX;
318 /* fall through */
319 case OOM_SCAN_CONTINUE:
320 continue;
321 case OOM_SCAN_ABORT:
322 rcu_read_unlock();
323 return (struct task_struct *)(-1UL);
324 case OOM_SCAN_OK:
325 break;
326 };
327 points = oom_badness(p, NULL, oc->nodemask, totalpages);
328 if (!points || points < chosen_points)
329 continue;
330
331 chosen = p;
332 chosen_points = points;
333 }
334 if (chosen)
335 get_task_struct(chosen);
336 rcu_read_unlock();
337
338 *ppoints = chosen_points * 1000 / totalpages;
339 return chosen;
340 }
341
342 /**
343 * dump_tasks - dump current memory state of all system tasks
344 * @memcg: current's memory controller, if constrained
345 * @nodemask: nodemask passed to page allocator for mempolicy ooms
346 *
347 * Dumps the current memory state of all eligible tasks. Tasks not in the same
348 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
349 * are not shown.
350 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
351 * swapents, oom_score_adj value, and name.
352 */
353 static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
354 {
355 struct task_struct *p;
356 struct task_struct *task;
357
358 pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n");
359 rcu_read_lock();
360 for_each_process(p) {
361 if (oom_unkillable_task(p, memcg, nodemask))
362 continue;
363
364 task = find_lock_task_mm(p);
365 if (!task) {
366 /*
367 * This is a kthread or all of p's threads have already
368 * detached their mm's. There's no need to report
369 * them; they can't be oom killed anyway.
370 */
371 continue;
372 }
373
374 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n",
375 task->pid, from_kuid(&init_user_ns, task_uid(task)),
376 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
377 atomic_long_read(&task->mm->nr_ptes),
378 mm_nr_pmds(task->mm),
379 get_mm_counter(task->mm, MM_SWAPENTS),
380 task->signal->oom_score_adj, task->comm);
381 task_unlock(task);
382 }
383 rcu_read_unlock();
384 }
385
386 static void dump_header(struct oom_control *oc, struct task_struct *p,
387 struct mem_cgroup *memcg)
388 {
389 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
390 current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
391 current->signal->oom_score_adj);
392
393 cpuset_print_current_mems_allowed();
394 dump_stack();
395 if (memcg)
396 mem_cgroup_print_oom_info(memcg, p);
397 else
398 show_mem(SHOW_MEM_FILTER_NODES);
399 if (sysctl_oom_dump_tasks)
400 dump_tasks(memcg, oc->nodemask);
401 }
402
403 /*
404 * Number of OOM victims in flight
405 */
406 static atomic_t oom_victims = ATOMIC_INIT(0);
407 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
408
409 bool oom_killer_disabled __read_mostly;
410
411 #define K(x) ((x) << (PAGE_SHIFT-10))
412
413 /*
414 * task->mm can be NULL if the task is the exited group leader. So to
415 * determine whether the task is using a particular mm, we examine all the
416 * task's threads: if one of those is using this mm then this task was also
417 * using it.
418 */
419 static bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
420 {
421 struct task_struct *t;
422
423 for_each_thread(p, t) {
424 struct mm_struct *t_mm = READ_ONCE(t->mm);
425 if (t_mm)
426 return t_mm == mm;
427 }
428 return false;
429 }
430
431
432 #ifdef CONFIG_MMU
433 /*
434 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
435 * victim (if that is possible) to help the OOM killer to move on.
436 */
437 static struct task_struct *oom_reaper_th;
438 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
439 static struct task_struct *oom_reaper_list;
440 static DEFINE_SPINLOCK(oom_reaper_lock);
441
442 static bool __oom_reap_task(struct task_struct *tsk)
443 {
444 struct mmu_gather tlb;
445 struct vm_area_struct *vma;
446 struct mm_struct *mm;
447 struct task_struct *p;
448 struct zap_details details = {.check_swap_entries = true,
449 .ignore_dirty = true};
450 bool ret = true;
451
452 /*
453 * Make sure we find the associated mm_struct even when the particular
454 * thread has already terminated and cleared its mm.
455 * We might have race with exit path so consider our work done if there
456 * is no mm.
457 */
458 p = find_lock_task_mm(tsk);
459 if (!p)
460 return true;
461
462 mm = p->mm;
463 if (!atomic_inc_not_zero(&mm->mm_users)) {
464 task_unlock(p);
465 return true;
466 }
467
468 task_unlock(p);
469
470 if (!down_read_trylock(&mm->mmap_sem)) {
471 ret = false;
472 goto out;
473 }
474
475 tlb_gather_mmu(&tlb, mm, 0, -1);
476 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
477 if (is_vm_hugetlb_page(vma))
478 continue;
479
480 /*
481 * mlocked VMAs require explicit munlocking before unmap.
482 * Let's keep it simple here and skip such VMAs.
483 */
484 if (vma->vm_flags & VM_LOCKED)
485 continue;
486
487 /*
488 * Only anonymous pages have a good chance to be dropped
489 * without additional steps which we cannot afford as we
490 * are OOM already.
491 *
492 * We do not even care about fs backed pages because all
493 * which are reclaimable have already been reclaimed and
494 * we do not want to block exit_mmap by keeping mm ref
495 * count elevated without a good reason.
496 */
497 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED))
498 unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end,
499 &details);
500 }
501 tlb_finish_mmu(&tlb, 0, -1);
502 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
503 task_pid_nr(tsk), tsk->comm,
504 K(get_mm_counter(mm, MM_ANONPAGES)),
505 K(get_mm_counter(mm, MM_FILEPAGES)),
506 K(get_mm_counter(mm, MM_SHMEMPAGES)));
507 up_read(&mm->mmap_sem);
508
509 /*
510 * This task can be safely ignored because we cannot do much more
511 * to release its memory.
512 */
513 set_bit(MMF_OOM_REAPED, &mm->flags);
514 out:
515 /*
516 * Drop our reference but make sure the mmput slow path is called from a
517 * different context because we shouldn't risk we get stuck there and
518 * put the oom_reaper out of the way.
519 */
520 mmput_async(mm);
521 return ret;
522 }
523
524 #define MAX_OOM_REAP_RETRIES 10
525 static void oom_reap_task(struct task_struct *tsk)
526 {
527 int attempts = 0;
528
529 /* Retry the down_read_trylock(mmap_sem) a few times */
530 while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task(tsk))
531 schedule_timeout_idle(HZ/10);
532
533 if (attempts > MAX_OOM_REAP_RETRIES) {
534 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
535 task_pid_nr(tsk), tsk->comm);
536 debug_show_all_locks();
537 }
538
539 /*
540 * Clear TIF_MEMDIE because the task shouldn't be sitting on a
541 * reasonably reclaimable memory anymore or it is not a good candidate
542 * for the oom victim right now because it cannot release its memory
543 * itself nor by the oom reaper.
544 */
545 tsk->oom_reaper_list = NULL;
546 exit_oom_victim(tsk);
547
548 /* Drop a reference taken by wake_oom_reaper */
549 put_task_struct(tsk);
550 }
551
552 static int oom_reaper(void *unused)
553 {
554 set_freezable();
555
556 while (true) {
557 struct task_struct *tsk = NULL;
558
559 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
560 spin_lock(&oom_reaper_lock);
561 if (oom_reaper_list != NULL) {
562 tsk = oom_reaper_list;
563 oom_reaper_list = tsk->oom_reaper_list;
564 }
565 spin_unlock(&oom_reaper_lock);
566
567 if (tsk)
568 oom_reap_task(tsk);
569 }
570
571 return 0;
572 }
573
574 static void wake_oom_reaper(struct task_struct *tsk)
575 {
576 if (!oom_reaper_th)
577 return;
578
579 /* tsk is already queued? */
580 if (tsk == oom_reaper_list || tsk->oom_reaper_list)
581 return;
582
583 get_task_struct(tsk);
584
585 spin_lock(&oom_reaper_lock);
586 tsk->oom_reaper_list = oom_reaper_list;
587 oom_reaper_list = tsk;
588 spin_unlock(&oom_reaper_lock);
589 wake_up(&oom_reaper_wait);
590 }
591
592 /* Check if we can reap the given task. This has to be called with stable
593 * tsk->mm
594 */
595 void try_oom_reaper(struct task_struct *tsk)
596 {
597 struct mm_struct *mm = tsk->mm;
598 struct task_struct *p;
599
600 if (!mm)
601 return;
602
603 /*
604 * There might be other threads/processes which are either not
605 * dying or even not killable.
606 */
607 if (atomic_read(&mm->mm_users) > 1) {
608 rcu_read_lock();
609 for_each_process(p) {
610 bool exiting;
611
612 if (!process_shares_mm(p, mm))
613 continue;
614 if (same_thread_group(p, tsk))
615 continue;
616 if (fatal_signal_pending(p))
617 continue;
618
619 /*
620 * If the task is exiting make sure the whole thread group
621 * is exiting and cannot acces mm anymore.
622 */
623 spin_lock_irq(&p->sighand->siglock);
624 exiting = signal_group_exit(p->signal);
625 spin_unlock_irq(&p->sighand->siglock);
626 if (exiting)
627 continue;
628
629 /* Give up */
630 rcu_read_unlock();
631 return;
632 }
633 rcu_read_unlock();
634 }
635
636 wake_oom_reaper(tsk);
637 }
638
639 static int __init oom_init(void)
640 {
641 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
642 if (IS_ERR(oom_reaper_th)) {
643 pr_err("Unable to start OOM reaper %ld. Continuing regardless\n",
644 PTR_ERR(oom_reaper_th));
645 oom_reaper_th = NULL;
646 }
647 return 0;
648 }
649 subsys_initcall(oom_init)
650 #else
651 static void wake_oom_reaper(struct task_struct *tsk)
652 {
653 }
654 #endif
655
656 /**
657 * mark_oom_victim - mark the given task as OOM victim
658 * @tsk: task to mark
659 *
660 * Has to be called with oom_lock held and never after
661 * oom has been disabled already.
662 */
663 void mark_oom_victim(struct task_struct *tsk)
664 {
665 WARN_ON(oom_killer_disabled);
666 /* OOM killer might race with memcg OOM */
667 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
668 return;
669 atomic_inc(&tsk->signal->oom_victims);
670 /*
671 * Make sure that the task is woken up from uninterruptible sleep
672 * if it is frozen because OOM killer wouldn't be able to free
673 * any memory and livelock. freezing_slow_path will tell the freezer
674 * that TIF_MEMDIE tasks should be ignored.
675 */
676 __thaw_task(tsk);
677 atomic_inc(&oom_victims);
678 }
679
680 /**
681 * exit_oom_victim - note the exit of an OOM victim
682 */
683 void exit_oom_victim(struct task_struct *tsk)
684 {
685 if (!test_and_clear_tsk_thread_flag(tsk, TIF_MEMDIE))
686 return;
687 atomic_dec(&tsk->signal->oom_victims);
688
689 if (!atomic_dec_return(&oom_victims))
690 wake_up_all(&oom_victims_wait);
691 }
692
693 /**
694 * oom_killer_disable - disable OOM killer
695 *
696 * Forces all page allocations to fail rather than trigger OOM killer.
697 * Will block and wait until all OOM victims are killed.
698 *
699 * The function cannot be called when there are runnable user tasks because
700 * the userspace would see unexpected allocation failures as a result. Any
701 * new usage of this function should be consulted with MM people.
702 *
703 * Returns true if successful and false if the OOM killer cannot be
704 * disabled.
705 */
706 bool oom_killer_disable(void)
707 {
708 /*
709 * Make sure to not race with an ongoing OOM killer. Check that the
710 * current is not killed (possibly due to sharing the victim's memory).
711 */
712 if (mutex_lock_killable(&oom_lock))
713 return false;
714 oom_killer_disabled = true;
715 mutex_unlock(&oom_lock);
716
717 wait_event(oom_victims_wait, !atomic_read(&oom_victims));
718
719 return true;
720 }
721
722 /**
723 * oom_killer_enable - enable OOM killer
724 */
725 void oom_killer_enable(void)
726 {
727 oom_killer_disabled = false;
728 }
729
730 /*
731 * Must be called while holding a reference to p, which will be released upon
732 * returning.
733 */
734 void oom_kill_process(struct oom_control *oc, struct task_struct *p,
735 unsigned int points, unsigned long totalpages,
736 struct mem_cgroup *memcg, const char *message)
737 {
738 struct task_struct *victim = p;
739 struct task_struct *child;
740 struct task_struct *t;
741 struct mm_struct *mm;
742 unsigned int victim_points = 0;
743 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
744 DEFAULT_RATELIMIT_BURST);
745 bool can_oom_reap = true;
746
747 /*
748 * If the task is already exiting, don't alarm the sysadmin or kill
749 * its children or threads, just set TIF_MEMDIE so it can die quickly
750 */
751 task_lock(p);
752 if (p->mm && task_will_free_mem(p)) {
753 mark_oom_victim(p);
754 try_oom_reaper(p);
755 task_unlock(p);
756 put_task_struct(p);
757 return;
758 }
759 task_unlock(p);
760
761 if (__ratelimit(&oom_rs))
762 dump_header(oc, p, memcg);
763
764 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
765 message, task_pid_nr(p), p->comm, points);
766
767 /*
768 * If any of p's children has a different mm and is eligible for kill,
769 * the one with the highest oom_badness() score is sacrificed for its
770 * parent. This attempts to lose the minimal amount of work done while
771 * still freeing memory.
772 */
773 read_lock(&tasklist_lock);
774 for_each_thread(p, t) {
775 list_for_each_entry(child, &t->children, sibling) {
776 unsigned int child_points;
777
778 if (process_shares_mm(child, p->mm))
779 continue;
780 /*
781 * oom_badness() returns 0 if the thread is unkillable
782 */
783 child_points = oom_badness(child, memcg, oc->nodemask,
784 totalpages);
785 if (child_points > victim_points) {
786 put_task_struct(victim);
787 victim = child;
788 victim_points = child_points;
789 get_task_struct(victim);
790 }
791 }
792 }
793 read_unlock(&tasklist_lock);
794
795 p = find_lock_task_mm(victim);
796 if (!p) {
797 put_task_struct(victim);
798 return;
799 } else if (victim != p) {
800 get_task_struct(p);
801 put_task_struct(victim);
802 victim = p;
803 }
804
805 /* Get a reference to safely compare mm after task_unlock(victim) */
806 mm = victim->mm;
807 atomic_inc(&mm->mm_count);
808 /*
809 * We should send SIGKILL before setting TIF_MEMDIE in order to prevent
810 * the OOM victim from depleting the memory reserves from the user
811 * space under its control.
812 */
813 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
814 mark_oom_victim(victim);
815 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
816 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
817 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
818 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
819 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
820 task_unlock(victim);
821
822 /*
823 * Kill all user processes sharing victim->mm in other thread groups, if
824 * any. They don't get access to memory reserves, though, to avoid
825 * depletion of all memory. This prevents mm->mmap_sem livelock when an
826 * oom killed thread cannot exit because it requires the semaphore and
827 * its contended by another thread trying to allocate memory itself.
828 * That thread will now get access to memory reserves since it has a
829 * pending fatal signal.
830 */
831 rcu_read_lock();
832 for_each_process(p) {
833 if (!process_shares_mm(p, mm))
834 continue;
835 if (same_thread_group(p, victim))
836 continue;
837 if (unlikely(p->flags & PF_KTHREAD) || is_global_init(p) ||
838 p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
839 /*
840 * We cannot use oom_reaper for the mm shared by this
841 * process because it wouldn't get killed and so the
842 * memory might be still used.
843 */
844 can_oom_reap = false;
845 continue;
846 }
847 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
848 }
849 rcu_read_unlock();
850
851 if (can_oom_reap)
852 wake_oom_reaper(victim);
853
854 mmdrop(mm);
855 put_task_struct(victim);
856 }
857 #undef K
858
859 /*
860 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
861 */
862 void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint,
863 struct mem_cgroup *memcg)
864 {
865 if (likely(!sysctl_panic_on_oom))
866 return;
867 if (sysctl_panic_on_oom != 2) {
868 /*
869 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
870 * does not panic for cpuset, mempolicy, or memcg allocation
871 * failures.
872 */
873 if (constraint != CONSTRAINT_NONE)
874 return;
875 }
876 /* Do not panic for oom kills triggered by sysrq */
877 if (is_sysrq_oom(oc))
878 return;
879 dump_header(oc, NULL, memcg);
880 panic("Out of memory: %s panic_on_oom is enabled\n",
881 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
882 }
883
884 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
885
886 int register_oom_notifier(struct notifier_block *nb)
887 {
888 return blocking_notifier_chain_register(&oom_notify_list, nb);
889 }
890 EXPORT_SYMBOL_GPL(register_oom_notifier);
891
892 int unregister_oom_notifier(struct notifier_block *nb)
893 {
894 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
895 }
896 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
897
898 /**
899 * out_of_memory - kill the "best" process when we run out of memory
900 * @oc: pointer to struct oom_control
901 *
902 * If we run out of memory, we have the choice between either
903 * killing a random task (bad), letting the system crash (worse)
904 * OR try to be smart about which process to kill. Note that we
905 * don't have to be perfect here, we just have to be good.
906 */
907 bool out_of_memory(struct oom_control *oc)
908 {
909 struct task_struct *p;
910 unsigned long totalpages;
911 unsigned long freed = 0;
912 unsigned int uninitialized_var(points);
913 enum oom_constraint constraint = CONSTRAINT_NONE;
914
915 if (oom_killer_disabled)
916 return false;
917
918 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
919 if (freed > 0)
920 /* Got some memory back in the last second. */
921 return true;
922
923 /*
924 * If current has a pending SIGKILL or is exiting, then automatically
925 * select it. The goal is to allow it to allocate so that it may
926 * quickly exit and free its memory.
927 *
928 * But don't select if current has already released its mm and cleared
929 * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur.
930 */
931 if (current->mm &&
932 (fatal_signal_pending(current) || task_will_free_mem(current))) {
933 mark_oom_victim(current);
934 try_oom_reaper(current);
935 return true;
936 }
937
938 /*
939 * The OOM killer does not compensate for IO-less reclaim.
940 * pagefault_out_of_memory lost its gfp context so we have to
941 * make sure exclude 0 mask - all other users should have at least
942 * ___GFP_DIRECT_RECLAIM to get here.
943 */
944 if (oc->gfp_mask && !(oc->gfp_mask & (__GFP_FS|__GFP_NOFAIL)))
945 return true;
946
947 /*
948 * Check if there were limitations on the allocation (only relevant for
949 * NUMA) that may require different handling.
950 */
951 constraint = constrained_alloc(oc, &totalpages);
952 if (constraint != CONSTRAINT_MEMORY_POLICY)
953 oc->nodemask = NULL;
954 check_panic_on_oom(oc, constraint, NULL);
955
956 if (sysctl_oom_kill_allocating_task && current->mm &&
957 !oom_unkillable_task(current, NULL, oc->nodemask) &&
958 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
959 get_task_struct(current);
960 oom_kill_process(oc, current, 0, totalpages, NULL,
961 "Out of memory (oom_kill_allocating_task)");
962 return true;
963 }
964
965 p = select_bad_process(oc, &points, totalpages);
966 /* Found nothing?!?! Either we hang forever, or we panic. */
967 if (!p && !is_sysrq_oom(oc)) {
968 dump_header(oc, NULL, NULL);
969 panic("Out of memory and no killable processes...\n");
970 }
971 if (p && p != (void *)-1UL) {
972 oom_kill_process(oc, p, points, totalpages, NULL,
973 "Out of memory");
974 /*
975 * Give the killed process a good chance to exit before trying
976 * to allocate memory again.
977 */
978 schedule_timeout_killable(1);
979 }
980 return true;
981 }
982
983 /*
984 * The pagefault handler calls here because it is out of memory, so kill a
985 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
986 * parallel oom killing is already in progress so do nothing.
987 */
988 void pagefault_out_of_memory(void)
989 {
990 struct oom_control oc = {
991 .zonelist = NULL,
992 .nodemask = NULL,
993 .gfp_mask = 0,
994 .order = 0,
995 };
996
997 if (mem_cgroup_oom_synchronize(true))
998 return;
999
1000 if (!mutex_trylock(&oom_lock))
1001 return;
1002
1003 if (!out_of_memory(&oc)) {
1004 /*
1005 * There shouldn't be any user tasks runnable while the
1006 * OOM killer is disabled, so the current task has to
1007 * be a racing OOM victim for which oom_killer_disable()
1008 * is waiting for.
1009 */
1010 WARN_ON(test_thread_flag(TIF_MEMDIE));
1011 }
1012
1013 mutex_unlock(&oom_lock);
1014 }