xref: /openbmc/linux/mm/oom_kill.c (revision 92a76f6d)
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 	adj = (long)p->signal->oom_score_adj;
178 	if (adj == OOM_SCORE_ADJ_MIN) {
179 		task_unlock(p);
180 		return 0;
181 	}
182 
183 	/*
184 	 * The baseline for the badness score is the proportion of RAM that each
185 	 * task's rss, pagetable and swap space use.
186 	 */
187 	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
188 		atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
189 	task_unlock(p);
190 
191 	/*
192 	 * Root processes get 3% bonus, just like the __vm_enough_memory()
193 	 * implementation used by LSMs.
194 	 */
195 	if (has_capability_noaudit(p, CAP_SYS_ADMIN))
196 		points -= (points * 3) / 100;
197 
198 	/* Normalize to oom_score_adj units */
199 	adj *= totalpages / 1000;
200 	points += adj;
201 
202 	/*
203 	 * Never return 0 for an eligible task regardless of the root bonus and
204 	 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
205 	 */
206 	return points > 0 ? points : 1;
207 }
208 
209 /*
210  * Determine the type of allocation constraint.
211  */
212 #ifdef CONFIG_NUMA
213 static enum oom_constraint constrained_alloc(struct oom_control *oc,
214 					     unsigned long *totalpages)
215 {
216 	struct zone *zone;
217 	struct zoneref *z;
218 	enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
219 	bool cpuset_limited = false;
220 	int nid;
221 
222 	/* Default to all available memory */
223 	*totalpages = totalram_pages + total_swap_pages;
224 
225 	if (!oc->zonelist)
226 		return CONSTRAINT_NONE;
227 	/*
228 	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
229 	 * to kill current.We have to random task kill in this case.
230 	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
231 	 */
232 	if (oc->gfp_mask & __GFP_THISNODE)
233 		return CONSTRAINT_NONE;
234 
235 	/*
236 	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
237 	 * the page allocator means a mempolicy is in effect.  Cpuset policy
238 	 * is enforced in get_page_from_freelist().
239 	 */
240 	if (oc->nodemask &&
241 	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
242 		*totalpages = total_swap_pages;
243 		for_each_node_mask(nid, *oc->nodemask)
244 			*totalpages += node_spanned_pages(nid);
245 		return CONSTRAINT_MEMORY_POLICY;
246 	}
247 
248 	/* Check this allocation failure is caused by cpuset's wall function */
249 	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
250 			high_zoneidx, oc->nodemask)
251 		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
252 			cpuset_limited = true;
253 
254 	if (cpuset_limited) {
255 		*totalpages = total_swap_pages;
256 		for_each_node_mask(nid, cpuset_current_mems_allowed)
257 			*totalpages += node_spanned_pages(nid);
258 		return CONSTRAINT_CPUSET;
259 	}
260 	return CONSTRAINT_NONE;
261 }
262 #else
263 static enum oom_constraint constrained_alloc(struct oom_control *oc,
264 					     unsigned long *totalpages)
265 {
266 	*totalpages = totalram_pages + total_swap_pages;
267 	return CONSTRAINT_NONE;
268 }
269 #endif
270 
271 enum oom_scan_t oom_scan_process_thread(struct oom_control *oc,
272 			struct task_struct *task, unsigned long totalpages)
273 {
274 	if (oom_unkillable_task(task, NULL, oc->nodemask))
275 		return OOM_SCAN_CONTINUE;
276 
277 	/*
278 	 * This task already has access to memory reserves and is being killed.
279 	 * Don't allow any other task to have access to the reserves.
280 	 */
281 	if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
282 		if (!is_sysrq_oom(oc))
283 			return OOM_SCAN_ABORT;
284 	}
285 	if (!task->mm)
286 		return OOM_SCAN_CONTINUE;
287 
288 	/*
289 	 * If task is allocating a lot of memory and has been marked to be
290 	 * killed first if it triggers an oom, then select it.
291 	 */
292 	if (oom_task_origin(task))
293 		return OOM_SCAN_SELECT;
294 
295 	return OOM_SCAN_OK;
296 }
297 
298 /*
299  * Simple selection loop. We chose the process with the highest
300  * number of 'points'.  Returns -1 on scan abort.
301  */
302 static struct task_struct *select_bad_process(struct oom_control *oc,
303 		unsigned int *ppoints, unsigned long totalpages)
304 {
305 	struct task_struct *g, *p;
306 	struct task_struct *chosen = NULL;
307 	unsigned long chosen_points = 0;
308 
309 	rcu_read_lock();
310 	for_each_process_thread(g, p) {
311 		unsigned int points;
312 
313 		switch (oom_scan_process_thread(oc, p, totalpages)) {
314 		case OOM_SCAN_SELECT:
315 			chosen = p;
316 			chosen_points = ULONG_MAX;
317 			/* fall through */
318 		case OOM_SCAN_CONTINUE:
319 			continue;
320 		case OOM_SCAN_ABORT:
321 			rcu_read_unlock();
322 			return (struct task_struct *)(-1UL);
323 		case OOM_SCAN_OK:
324 			break;
325 		};
326 		points = oom_badness(p, NULL, oc->nodemask, totalpages);
327 		if (!points || points < chosen_points)
328 			continue;
329 		/* Prefer thread group leaders for display purposes */
330 		if (points == chosen_points && thread_group_leader(chosen))
331 			continue;
332 
333 		chosen = p;
334 		chosen_points = points;
335 	}
336 	if (chosen)
337 		get_task_struct(chosen);
338 	rcu_read_unlock();
339 
340 	*ppoints = chosen_points * 1000 / totalpages;
341 	return chosen;
342 }
343 
344 /**
345  * dump_tasks - dump current memory state of all system tasks
346  * @memcg: current's memory controller, if constrained
347  * @nodemask: nodemask passed to page allocator for mempolicy ooms
348  *
349  * Dumps the current memory state of all eligible tasks.  Tasks not in the same
350  * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
351  * are not shown.
352  * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
353  * swapents, oom_score_adj value, and name.
354  */
355 static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
356 {
357 	struct task_struct *p;
358 	struct task_struct *task;
359 
360 	pr_info("[ pid ]   uid  tgid total_vm      rss nr_ptes nr_pmds swapents oom_score_adj name\n");
361 	rcu_read_lock();
362 	for_each_process(p) {
363 		if (oom_unkillable_task(p, memcg, nodemask))
364 			continue;
365 
366 		task = find_lock_task_mm(p);
367 		if (!task) {
368 			/*
369 			 * This is a kthread or all of p's threads have already
370 			 * detached their mm's.  There's no need to report
371 			 * them; they can't be oom killed anyway.
372 			 */
373 			continue;
374 		}
375 
376 		pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu         %5hd %s\n",
377 			task->pid, from_kuid(&init_user_ns, task_uid(task)),
378 			task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
379 			atomic_long_read(&task->mm->nr_ptes),
380 			mm_nr_pmds(task->mm),
381 			get_mm_counter(task->mm, MM_SWAPENTS),
382 			task->signal->oom_score_adj, task->comm);
383 		task_unlock(task);
384 	}
385 	rcu_read_unlock();
386 }
387 
388 static void dump_header(struct oom_control *oc, struct task_struct *p,
389 			struct mem_cgroup *memcg)
390 {
391 	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
392 		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
393 		current->signal->oom_score_adj);
394 
395 	cpuset_print_current_mems_allowed();
396 	dump_stack();
397 	if (memcg)
398 		mem_cgroup_print_oom_info(memcg, p);
399 	else
400 		show_mem(SHOW_MEM_FILTER_NODES);
401 	if (sysctl_oom_dump_tasks)
402 		dump_tasks(memcg, oc->nodemask);
403 }
404 
405 /*
406  * Number of OOM victims in flight
407  */
408 static atomic_t oom_victims = ATOMIC_INIT(0);
409 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
410 
411 bool oom_killer_disabled __read_mostly;
412 
413 #define K(x) ((x) << (PAGE_SHIFT-10))
414 
415 #ifdef CONFIG_MMU
416 /*
417  * OOM Reaper kernel thread which tries to reap the memory used by the OOM
418  * victim (if that is possible) to help the OOM killer to move on.
419  */
420 static struct task_struct *oom_reaper_th;
421 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
422 static struct task_struct *oom_reaper_list;
423 static DEFINE_SPINLOCK(oom_reaper_lock);
424 
425 
426 static bool __oom_reap_task(struct task_struct *tsk)
427 {
428 	struct mmu_gather tlb;
429 	struct vm_area_struct *vma;
430 	struct mm_struct *mm;
431 	struct task_struct *p;
432 	struct zap_details details = {.check_swap_entries = true,
433 				      .ignore_dirty = true};
434 	bool ret = true;
435 
436 	/*
437 	 * Make sure we find the associated mm_struct even when the particular
438 	 * thread has already terminated and cleared its mm.
439 	 * We might have race with exit path so consider our work done if there
440 	 * is no mm.
441 	 */
442 	p = find_lock_task_mm(tsk);
443 	if (!p)
444 		return true;
445 
446 	mm = p->mm;
447 	if (!atomic_inc_not_zero(&mm->mm_users)) {
448 		task_unlock(p);
449 		return true;
450 	}
451 
452 	task_unlock(p);
453 
454 	if (!down_read_trylock(&mm->mmap_sem)) {
455 		ret = false;
456 		goto out;
457 	}
458 
459 	tlb_gather_mmu(&tlb, mm, 0, -1);
460 	for (vma = mm->mmap ; vma; vma = vma->vm_next) {
461 		if (is_vm_hugetlb_page(vma))
462 			continue;
463 
464 		/*
465 		 * mlocked VMAs require explicit munlocking before unmap.
466 		 * Let's keep it simple here and skip such VMAs.
467 		 */
468 		if (vma->vm_flags & VM_LOCKED)
469 			continue;
470 
471 		/*
472 		 * Only anonymous pages have a good chance to be dropped
473 		 * without additional steps which we cannot afford as we
474 		 * are OOM already.
475 		 *
476 		 * We do not even care about fs backed pages because all
477 		 * which are reclaimable have already been reclaimed and
478 		 * we do not want to block exit_mmap by keeping mm ref
479 		 * count elevated without a good reason.
480 		 */
481 		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED))
482 			unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end,
483 					 &details);
484 	}
485 	tlb_finish_mmu(&tlb, 0, -1);
486 	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
487 			task_pid_nr(tsk), tsk->comm,
488 			K(get_mm_counter(mm, MM_ANONPAGES)),
489 			K(get_mm_counter(mm, MM_FILEPAGES)),
490 			K(get_mm_counter(mm, MM_SHMEMPAGES)));
491 	up_read(&mm->mmap_sem);
492 
493 	/*
494 	 * Clear TIF_MEMDIE because the task shouldn't be sitting on a
495 	 * reasonably reclaimable memory anymore. OOM killer can continue
496 	 * by selecting other victim if unmapping hasn't led to any
497 	 * improvements. This also means that selecting this task doesn't
498 	 * make any sense.
499 	 */
500 	tsk->signal->oom_score_adj = OOM_SCORE_ADJ_MIN;
501 	exit_oom_victim(tsk);
502 out:
503 	mmput(mm);
504 	return ret;
505 }
506 
507 #define MAX_OOM_REAP_RETRIES 10
508 static void oom_reap_task(struct task_struct *tsk)
509 {
510 	int attempts = 0;
511 
512 	/* Retry the down_read_trylock(mmap_sem) a few times */
513 	while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task(tsk))
514 		schedule_timeout_idle(HZ/10);
515 
516 	if (attempts > MAX_OOM_REAP_RETRIES) {
517 		pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
518 				task_pid_nr(tsk), tsk->comm);
519 		debug_show_all_locks();
520 	}
521 
522 	/* Drop a reference taken by wake_oom_reaper */
523 	put_task_struct(tsk);
524 }
525 
526 static int oom_reaper(void *unused)
527 {
528 	set_freezable();
529 
530 	while (true) {
531 		struct task_struct *tsk = NULL;
532 
533 		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
534 		spin_lock(&oom_reaper_lock);
535 		if (oom_reaper_list != NULL) {
536 			tsk = oom_reaper_list;
537 			oom_reaper_list = tsk->oom_reaper_list;
538 		}
539 		spin_unlock(&oom_reaper_lock);
540 
541 		if (tsk)
542 			oom_reap_task(tsk);
543 	}
544 
545 	return 0;
546 }
547 
548 static void wake_oom_reaper(struct task_struct *tsk)
549 {
550 	if (!oom_reaper_th || tsk->oom_reaper_list)
551 		return;
552 
553 	get_task_struct(tsk);
554 
555 	spin_lock(&oom_reaper_lock);
556 	tsk->oom_reaper_list = oom_reaper_list;
557 	oom_reaper_list = tsk;
558 	spin_unlock(&oom_reaper_lock);
559 	wake_up(&oom_reaper_wait);
560 }
561 
562 static int __init oom_init(void)
563 {
564 	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
565 	if (IS_ERR(oom_reaper_th)) {
566 		pr_err("Unable to start OOM reaper %ld. Continuing regardless\n",
567 				PTR_ERR(oom_reaper_th));
568 		oom_reaper_th = NULL;
569 	}
570 	return 0;
571 }
572 subsys_initcall(oom_init)
573 #else
574 static void wake_oom_reaper(struct task_struct *tsk)
575 {
576 }
577 #endif
578 
579 /**
580  * mark_oom_victim - mark the given task as OOM victim
581  * @tsk: task to mark
582  *
583  * Has to be called with oom_lock held and never after
584  * oom has been disabled already.
585  */
586 void mark_oom_victim(struct task_struct *tsk)
587 {
588 	WARN_ON(oom_killer_disabled);
589 	/* OOM killer might race with memcg OOM */
590 	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
591 		return;
592 	/*
593 	 * Make sure that the task is woken up from uninterruptible sleep
594 	 * if it is frozen because OOM killer wouldn't be able to free
595 	 * any memory and livelock. freezing_slow_path will tell the freezer
596 	 * that TIF_MEMDIE tasks should be ignored.
597 	 */
598 	__thaw_task(tsk);
599 	atomic_inc(&oom_victims);
600 }
601 
602 /**
603  * exit_oom_victim - note the exit of an OOM victim
604  */
605 void exit_oom_victim(struct task_struct *tsk)
606 {
607 	if (!test_and_clear_tsk_thread_flag(tsk, TIF_MEMDIE))
608 		return;
609 
610 	if (!atomic_dec_return(&oom_victims))
611 		wake_up_all(&oom_victims_wait);
612 }
613 
614 /**
615  * oom_killer_disable - disable OOM killer
616  *
617  * Forces all page allocations to fail rather than trigger OOM killer.
618  * Will block and wait until all OOM victims are killed.
619  *
620  * The function cannot be called when there are runnable user tasks because
621  * the userspace would see unexpected allocation failures as a result. Any
622  * new usage of this function should be consulted with MM people.
623  *
624  * Returns true if successful and false if the OOM killer cannot be
625  * disabled.
626  */
627 bool oom_killer_disable(void)
628 {
629 	/*
630 	 * Make sure to not race with an ongoing OOM killer. Check that the
631 	 * current is not killed (possibly due to sharing the victim's memory).
632 	 */
633 	if (mutex_lock_killable(&oom_lock))
634 		return false;
635 	oom_killer_disabled = true;
636 	mutex_unlock(&oom_lock);
637 
638 	wait_event(oom_victims_wait, !atomic_read(&oom_victims));
639 
640 	return true;
641 }
642 
643 /**
644  * oom_killer_enable - enable OOM killer
645  */
646 void oom_killer_enable(void)
647 {
648 	oom_killer_disabled = false;
649 }
650 
651 /*
652  * task->mm can be NULL if the task is the exited group leader.  So to
653  * determine whether the task is using a particular mm, we examine all the
654  * task's threads: if one of those is using this mm then this task was also
655  * using it.
656  */
657 static bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
658 {
659 	struct task_struct *t;
660 
661 	for_each_thread(p, t) {
662 		struct mm_struct *t_mm = READ_ONCE(t->mm);
663 		if (t_mm)
664 			return t_mm == mm;
665 	}
666 	return false;
667 }
668 
669 /*
670  * Must be called while holding a reference to p, which will be released upon
671  * returning.
672  */
673 void oom_kill_process(struct oom_control *oc, struct task_struct *p,
674 		      unsigned int points, unsigned long totalpages,
675 		      struct mem_cgroup *memcg, const char *message)
676 {
677 	struct task_struct *victim = p;
678 	struct task_struct *child;
679 	struct task_struct *t;
680 	struct mm_struct *mm;
681 	unsigned int victim_points = 0;
682 	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
683 					      DEFAULT_RATELIMIT_BURST);
684 	bool can_oom_reap = true;
685 
686 	/*
687 	 * If the task is already exiting, don't alarm the sysadmin or kill
688 	 * its children or threads, just set TIF_MEMDIE so it can die quickly
689 	 */
690 	task_lock(p);
691 	if (p->mm && task_will_free_mem(p)) {
692 		mark_oom_victim(p);
693 		task_unlock(p);
694 		put_task_struct(p);
695 		return;
696 	}
697 	task_unlock(p);
698 
699 	if (__ratelimit(&oom_rs))
700 		dump_header(oc, p, memcg);
701 
702 	pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
703 		message, task_pid_nr(p), p->comm, points);
704 
705 	/*
706 	 * If any of p's children has a different mm and is eligible for kill,
707 	 * the one with the highest oom_badness() score is sacrificed for its
708 	 * parent.  This attempts to lose the minimal amount of work done while
709 	 * still freeing memory.
710 	 */
711 	read_lock(&tasklist_lock);
712 	for_each_thread(p, t) {
713 		list_for_each_entry(child, &t->children, sibling) {
714 			unsigned int child_points;
715 
716 			if (process_shares_mm(child, p->mm))
717 				continue;
718 			/*
719 			 * oom_badness() returns 0 if the thread is unkillable
720 			 */
721 			child_points = oom_badness(child, memcg, oc->nodemask,
722 								totalpages);
723 			if (child_points > victim_points) {
724 				put_task_struct(victim);
725 				victim = child;
726 				victim_points = child_points;
727 				get_task_struct(victim);
728 			}
729 		}
730 	}
731 	read_unlock(&tasklist_lock);
732 
733 	p = find_lock_task_mm(victim);
734 	if (!p) {
735 		put_task_struct(victim);
736 		return;
737 	} else if (victim != p) {
738 		get_task_struct(p);
739 		put_task_struct(victim);
740 		victim = p;
741 	}
742 
743 	/* Get a reference to safely compare mm after task_unlock(victim) */
744 	mm = victim->mm;
745 	atomic_inc(&mm->mm_count);
746 	/*
747 	 * We should send SIGKILL before setting TIF_MEMDIE in order to prevent
748 	 * the OOM victim from depleting the memory reserves from the user
749 	 * space under its control.
750 	 */
751 	do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
752 	mark_oom_victim(victim);
753 	pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
754 		task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
755 		K(get_mm_counter(victim->mm, MM_ANONPAGES)),
756 		K(get_mm_counter(victim->mm, MM_FILEPAGES)),
757 		K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
758 	task_unlock(victim);
759 
760 	/*
761 	 * Kill all user processes sharing victim->mm in other thread groups, if
762 	 * any.  They don't get access to memory reserves, though, to avoid
763 	 * depletion of all memory.  This prevents mm->mmap_sem livelock when an
764 	 * oom killed thread cannot exit because it requires the semaphore and
765 	 * its contended by another thread trying to allocate memory itself.
766 	 * That thread will now get access to memory reserves since it has a
767 	 * pending fatal signal.
768 	 */
769 	rcu_read_lock();
770 	for_each_process(p) {
771 		if (!process_shares_mm(p, mm))
772 			continue;
773 		if (same_thread_group(p, victim))
774 			continue;
775 		if (unlikely(p->flags & PF_KTHREAD) || is_global_init(p) ||
776 		    p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
777 			/*
778 			 * We cannot use oom_reaper for the mm shared by this
779 			 * process because it wouldn't get killed and so the
780 			 * memory might be still used.
781 			 */
782 			can_oom_reap = false;
783 			continue;
784 		}
785 		do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
786 	}
787 	rcu_read_unlock();
788 
789 	if (can_oom_reap)
790 		wake_oom_reaper(victim);
791 
792 	mmdrop(mm);
793 	put_task_struct(victim);
794 }
795 #undef K
796 
797 /*
798  * Determines whether the kernel must panic because of the panic_on_oom sysctl.
799  */
800 void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint,
801 			struct mem_cgroup *memcg)
802 {
803 	if (likely(!sysctl_panic_on_oom))
804 		return;
805 	if (sysctl_panic_on_oom != 2) {
806 		/*
807 		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
808 		 * does not panic for cpuset, mempolicy, or memcg allocation
809 		 * failures.
810 		 */
811 		if (constraint != CONSTRAINT_NONE)
812 			return;
813 	}
814 	/* Do not panic for oom kills triggered by sysrq */
815 	if (is_sysrq_oom(oc))
816 		return;
817 	dump_header(oc, NULL, memcg);
818 	panic("Out of memory: %s panic_on_oom is enabled\n",
819 		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
820 }
821 
822 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
823 
824 int register_oom_notifier(struct notifier_block *nb)
825 {
826 	return blocking_notifier_chain_register(&oom_notify_list, nb);
827 }
828 EXPORT_SYMBOL_GPL(register_oom_notifier);
829 
830 int unregister_oom_notifier(struct notifier_block *nb)
831 {
832 	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
833 }
834 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
835 
836 /**
837  * out_of_memory - kill the "best" process when we run out of memory
838  * @oc: pointer to struct oom_control
839  *
840  * If we run out of memory, we have the choice between either
841  * killing a random task (bad), letting the system crash (worse)
842  * OR try to be smart about which process to kill. Note that we
843  * don't have to be perfect here, we just have to be good.
844  */
845 bool out_of_memory(struct oom_control *oc)
846 {
847 	struct task_struct *p;
848 	unsigned long totalpages;
849 	unsigned long freed = 0;
850 	unsigned int uninitialized_var(points);
851 	enum oom_constraint constraint = CONSTRAINT_NONE;
852 
853 	if (oom_killer_disabled)
854 		return false;
855 
856 	blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
857 	if (freed > 0)
858 		/* Got some memory back in the last second. */
859 		return true;
860 
861 	/*
862 	 * If current has a pending SIGKILL or is exiting, then automatically
863 	 * select it.  The goal is to allow it to allocate so that it may
864 	 * quickly exit and free its memory.
865 	 *
866 	 * But don't select if current has already released its mm and cleared
867 	 * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur.
868 	 */
869 	if (current->mm &&
870 	    (fatal_signal_pending(current) || task_will_free_mem(current))) {
871 		mark_oom_victim(current);
872 		return true;
873 	}
874 
875 	/*
876 	 * Check if there were limitations on the allocation (only relevant for
877 	 * NUMA) that may require different handling.
878 	 */
879 	constraint = constrained_alloc(oc, &totalpages);
880 	if (constraint != CONSTRAINT_MEMORY_POLICY)
881 		oc->nodemask = NULL;
882 	check_panic_on_oom(oc, constraint, NULL);
883 
884 	if (sysctl_oom_kill_allocating_task && current->mm &&
885 	    !oom_unkillable_task(current, NULL, oc->nodemask) &&
886 	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
887 		get_task_struct(current);
888 		oom_kill_process(oc, current, 0, totalpages, NULL,
889 				 "Out of memory (oom_kill_allocating_task)");
890 		return true;
891 	}
892 
893 	p = select_bad_process(oc, &points, totalpages);
894 	/* Found nothing?!?! Either we hang forever, or we panic. */
895 	if (!p && !is_sysrq_oom(oc)) {
896 		dump_header(oc, NULL, NULL);
897 		panic("Out of memory and no killable processes...\n");
898 	}
899 	if (p && p != (void *)-1UL) {
900 		oom_kill_process(oc, p, points, totalpages, NULL,
901 				 "Out of memory");
902 		/*
903 		 * Give the killed process a good chance to exit before trying
904 		 * to allocate memory again.
905 		 */
906 		schedule_timeout_killable(1);
907 	}
908 	return true;
909 }
910 
911 /*
912  * The pagefault handler calls here because it is out of memory, so kill a
913  * memory-hogging task.  If any populated zone has ZONE_OOM_LOCKED set, a
914  * parallel oom killing is already in progress so do nothing.
915  */
916 void pagefault_out_of_memory(void)
917 {
918 	struct oom_control oc = {
919 		.zonelist = NULL,
920 		.nodemask = NULL,
921 		.gfp_mask = 0,
922 		.order = 0,
923 	};
924 
925 	if (mem_cgroup_oom_synchronize(true))
926 		return;
927 
928 	if (!mutex_trylock(&oom_lock))
929 		return;
930 
931 	if (!out_of_memory(&oc)) {
932 		/*
933 		 * There shouldn't be any user tasks runnable while the
934 		 * OOM killer is disabled, so the current task has to
935 		 * be a racing OOM victim for which oom_killer_disable()
936 		 * is waiting for.
937 		 */
938 		WARN_ON(test_thread_flag(TIF_MEMDIE));
939 	}
940 
941 	mutex_unlock(&oom_lock);
942 }
943