xref: /linux/mm/oom_kill.c (revision 37744feebc086908fd89760650f458ab19071750)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/oom_kill.c
4  *
5  *  Copyright (C)  1998,2000  Rik van Riel
6  *	Thanks go out to Claus Fischer for some serious inspiration and
7  *	for goading me into coding this file...
8  *  Copyright (C)  2010  Google, Inc.
9  *	Rewritten by David Rientjes
10  *
11  *  The routines in this file are used to kill a process when
12  *  we're seriously out of memory. This gets called from __alloc_pages()
13  *  in mm/page_alloc.c when we really run out of memory.
14  *
15  *  Since we won't call these routines often (on a well-configured
16  *  machine) this file will double as a 'coding guide' and a signpost
17  *  for newbie kernel hackers. It features several pointers to major
18  *  kernel subsystems and hints as to where to find out what things do.
19  */
20 
21 #include <linux/oom.h>
22 #include <linux/mm.h>
23 #include <linux/err.h>
24 #include <linux/gfp.h>
25 #include <linux/sched.h>
26 #include <linux/sched/mm.h>
27 #include <linux/sched/coredump.h>
28 #include <linux/sched/task.h>
29 #include <linux/sched/debug.h>
30 #include <linux/swap.h>
31 #include <linux/timex.h>
32 #include <linux/jiffies.h>
33 #include <linux/cpuset.h>
34 #include <linux/export.h>
35 #include <linux/notifier.h>
36 #include <linux/memcontrol.h>
37 #include <linux/mempolicy.h>
38 #include <linux/security.h>
39 #include <linux/ptrace.h>
40 #include <linux/freezer.h>
41 #include <linux/ftrace.h>
42 #include <linux/ratelimit.h>
43 #include <linux/kthread.h>
44 #include <linux/init.h>
45 #include <linux/mmu_notifier.h>
46 
47 #include <asm/tlb.h>
48 #include "internal.h"
49 #include "slab.h"
50 
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/oom.h>
53 
54 int sysctl_panic_on_oom;
55 int sysctl_oom_kill_allocating_task;
56 int sysctl_oom_dump_tasks = 1;
57 
58 /*
59  * Serializes oom killer invocations (out_of_memory()) from all contexts to
60  * prevent from over eager oom killing (e.g. when the oom killer is invoked
61  * from different domains).
62  *
63  * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
64  * and mark_oom_victim
65  */
66 DEFINE_MUTEX(oom_lock);
67 
68 static inline bool is_memcg_oom(struct oom_control *oc)
69 {
70 	return oc->memcg != NULL;
71 }
72 
73 #ifdef CONFIG_NUMA
74 /**
75  * oom_cpuset_eligible() - check task eligiblity for kill
76  * @start: task struct of which task to consider
77  * @oc: pointer to struct oom_control
78  *
79  * Task eligibility is determined by whether or not a candidate task, @tsk,
80  * shares the same mempolicy nodes as current if it is bound by such a policy
81  * and whether or not it has the same set of allowed cpuset nodes.
82  *
83  * This function is assuming oom-killer context and 'current' has triggered
84  * the oom-killer.
85  */
86 static bool oom_cpuset_eligible(struct task_struct *start,
87 				struct oom_control *oc)
88 {
89 	struct task_struct *tsk;
90 	bool ret = false;
91 	const nodemask_t *mask = oc->nodemask;
92 
93 	if (is_memcg_oom(oc))
94 		return true;
95 
96 	rcu_read_lock();
97 	for_each_thread(start, tsk) {
98 		if (mask) {
99 			/*
100 			 * If this is a mempolicy constrained oom, tsk's
101 			 * cpuset is irrelevant.  Only return true if its
102 			 * mempolicy intersects current, otherwise it may be
103 			 * needlessly killed.
104 			 */
105 			ret = mempolicy_nodemask_intersects(tsk, mask);
106 		} else {
107 			/*
108 			 * This is not a mempolicy constrained oom, so only
109 			 * check the mems of tsk's cpuset.
110 			 */
111 			ret = cpuset_mems_allowed_intersects(current, tsk);
112 		}
113 		if (ret)
114 			break;
115 	}
116 	rcu_read_unlock();
117 
118 	return ret;
119 }
120 #else
121 static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
122 {
123 	return true;
124 }
125 #endif /* CONFIG_NUMA */
126 
127 /*
128  * The process p may have detached its own ->mm while exiting or through
129  * use_mm(), but one or more of its subthreads may still have a valid
130  * pointer.  Return p, or any of its subthreads with a valid ->mm, with
131  * task_lock() held.
132  */
133 struct task_struct *find_lock_task_mm(struct task_struct *p)
134 {
135 	struct task_struct *t;
136 
137 	rcu_read_lock();
138 
139 	for_each_thread(p, t) {
140 		task_lock(t);
141 		if (likely(t->mm))
142 			goto found;
143 		task_unlock(t);
144 	}
145 	t = NULL;
146 found:
147 	rcu_read_unlock();
148 
149 	return t;
150 }
151 
152 /*
153  * order == -1 means the oom kill is required by sysrq, otherwise only
154  * for display purposes.
155  */
156 static inline bool is_sysrq_oom(struct oom_control *oc)
157 {
158 	return oc->order == -1;
159 }
160 
161 /* return true if the task is not adequate as candidate victim task. */
162 static bool oom_unkillable_task(struct task_struct *p)
163 {
164 	if (is_global_init(p))
165 		return true;
166 	if (p->flags & PF_KTHREAD)
167 		return true;
168 	return false;
169 }
170 
171 /*
172  * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
173  * than all user memory (LRU pages)
174  */
175 static bool is_dump_unreclaim_slabs(void)
176 {
177 	unsigned long nr_lru;
178 
179 	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
180 		 global_node_page_state(NR_INACTIVE_ANON) +
181 		 global_node_page_state(NR_ACTIVE_FILE) +
182 		 global_node_page_state(NR_INACTIVE_FILE) +
183 		 global_node_page_state(NR_ISOLATED_ANON) +
184 		 global_node_page_state(NR_ISOLATED_FILE) +
185 		 global_node_page_state(NR_UNEVICTABLE);
186 
187 	return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
188 }
189 
190 /**
191  * oom_badness - heuristic function to determine which candidate task to kill
192  * @p: task struct of which task we should calculate
193  * @totalpages: total present RAM allowed for page allocation
194  *
195  * The heuristic for determining which task to kill is made to be as simple and
196  * predictable as possible.  The goal is to return the highest value for the
197  * task consuming the most memory to avoid subsequent oom failures.
198  */
199 unsigned long oom_badness(struct task_struct *p, unsigned long totalpages)
200 {
201 	long points;
202 	long adj;
203 
204 	if (oom_unkillable_task(p))
205 		return 0;
206 
207 	p = find_lock_task_mm(p);
208 	if (!p)
209 		return 0;
210 
211 	/*
212 	 * Do not even consider tasks which are explicitly marked oom
213 	 * unkillable or have been already oom reaped or the are in
214 	 * the middle of vfork
215 	 */
216 	adj = (long)p->signal->oom_score_adj;
217 	if (adj == OOM_SCORE_ADJ_MIN ||
218 			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
219 			in_vfork(p)) {
220 		task_unlock(p);
221 		return 0;
222 	}
223 
224 	/*
225 	 * The baseline for the badness score is the proportion of RAM that each
226 	 * task's rss, pagetable and swap space use.
227 	 */
228 	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
229 		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
230 	task_unlock(p);
231 
232 	/* Normalize to oom_score_adj units */
233 	adj *= totalpages / 1000;
234 	points += adj;
235 
236 	/*
237 	 * Never return 0 for an eligible task regardless of the root bonus and
238 	 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
239 	 */
240 	return points > 0 ? points : 1;
241 }
242 
243 static const char * const oom_constraint_text[] = {
244 	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
245 	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
246 	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
247 	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
248 };
249 
250 /*
251  * Determine the type of allocation constraint.
252  */
253 static enum oom_constraint constrained_alloc(struct oom_control *oc)
254 {
255 	struct zone *zone;
256 	struct zoneref *z;
257 	enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
258 	bool cpuset_limited = false;
259 	int nid;
260 
261 	if (is_memcg_oom(oc)) {
262 		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
263 		return CONSTRAINT_MEMCG;
264 	}
265 
266 	/* Default to all available memory */
267 	oc->totalpages = totalram_pages() + total_swap_pages;
268 
269 	if (!IS_ENABLED(CONFIG_NUMA))
270 		return CONSTRAINT_NONE;
271 
272 	if (!oc->zonelist)
273 		return CONSTRAINT_NONE;
274 	/*
275 	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
276 	 * to kill current.We have to random task kill in this case.
277 	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
278 	 */
279 	if (oc->gfp_mask & __GFP_THISNODE)
280 		return CONSTRAINT_NONE;
281 
282 	/*
283 	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
284 	 * the page allocator means a mempolicy is in effect.  Cpuset policy
285 	 * is enforced in get_page_from_freelist().
286 	 */
287 	if (oc->nodemask &&
288 	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
289 		oc->totalpages = total_swap_pages;
290 		for_each_node_mask(nid, *oc->nodemask)
291 			oc->totalpages += node_present_pages(nid);
292 		return CONSTRAINT_MEMORY_POLICY;
293 	}
294 
295 	/* Check this allocation failure is caused by cpuset's wall function */
296 	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
297 			high_zoneidx, oc->nodemask)
298 		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
299 			cpuset_limited = true;
300 
301 	if (cpuset_limited) {
302 		oc->totalpages = total_swap_pages;
303 		for_each_node_mask(nid, cpuset_current_mems_allowed)
304 			oc->totalpages += node_present_pages(nid);
305 		return CONSTRAINT_CPUSET;
306 	}
307 	return CONSTRAINT_NONE;
308 }
309 
310 static int oom_evaluate_task(struct task_struct *task, void *arg)
311 {
312 	struct oom_control *oc = arg;
313 	unsigned long points;
314 
315 	if (oom_unkillable_task(task))
316 		goto next;
317 
318 	/* p may not have freeable memory in nodemask */
319 	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
320 		goto next;
321 
322 	/*
323 	 * This task already has access to memory reserves and is being killed.
324 	 * Don't allow any other task to have access to the reserves unless
325 	 * the task has MMF_OOM_SKIP because chances that it would release
326 	 * any memory is quite low.
327 	 */
328 	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
329 		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
330 			goto next;
331 		goto abort;
332 	}
333 
334 	/*
335 	 * If task is allocating a lot of memory and has been marked to be
336 	 * killed first if it triggers an oom, then select it.
337 	 */
338 	if (oom_task_origin(task)) {
339 		points = ULONG_MAX;
340 		goto select;
341 	}
342 
343 	points = oom_badness(task, oc->totalpages);
344 	if (!points || points < oc->chosen_points)
345 		goto next;
346 
347 select:
348 	if (oc->chosen)
349 		put_task_struct(oc->chosen);
350 	get_task_struct(task);
351 	oc->chosen = task;
352 	oc->chosen_points = points;
353 next:
354 	return 0;
355 abort:
356 	if (oc->chosen)
357 		put_task_struct(oc->chosen);
358 	oc->chosen = (void *)-1UL;
359 	return 1;
360 }
361 
362 /*
363  * Simple selection loop. We choose the process with the highest number of
364  * 'points'. In case scan was aborted, oc->chosen is set to -1.
365  */
366 static void select_bad_process(struct oom_control *oc)
367 {
368 	if (is_memcg_oom(oc))
369 		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
370 	else {
371 		struct task_struct *p;
372 
373 		rcu_read_lock();
374 		for_each_process(p)
375 			if (oom_evaluate_task(p, oc))
376 				break;
377 		rcu_read_unlock();
378 	}
379 }
380 
381 static int dump_task(struct task_struct *p, void *arg)
382 {
383 	struct oom_control *oc = arg;
384 	struct task_struct *task;
385 
386 	if (oom_unkillable_task(p))
387 		return 0;
388 
389 	/* p may not have freeable memory in nodemask */
390 	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
391 		return 0;
392 
393 	task = find_lock_task_mm(p);
394 	if (!task) {
395 		/*
396 		 * This is a kthread or all of p's threads have already
397 		 * detached their mm's.  There's no need to report
398 		 * them; they can't be oom killed anyway.
399 		 */
400 		return 0;
401 	}
402 
403 	pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu         %5hd %s\n",
404 		task->pid, from_kuid(&init_user_ns, task_uid(task)),
405 		task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
406 		mm_pgtables_bytes(task->mm),
407 		get_mm_counter(task->mm, MM_SWAPENTS),
408 		task->signal->oom_score_adj, task->comm);
409 	task_unlock(task);
410 
411 	return 0;
412 }
413 
414 /**
415  * dump_tasks - dump current memory state of all system tasks
416  * @oc: pointer to struct oom_control
417  *
418  * Dumps the current memory state of all eligible tasks.  Tasks not in the same
419  * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
420  * are not shown.
421  * State information includes task's pid, uid, tgid, vm size, rss,
422  * pgtables_bytes, swapents, oom_score_adj value, and name.
423  */
424 static void dump_tasks(struct oom_control *oc)
425 {
426 	pr_info("Tasks state (memory values in pages):\n");
427 	pr_info("[  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name\n");
428 
429 	if (is_memcg_oom(oc))
430 		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
431 	else {
432 		struct task_struct *p;
433 
434 		rcu_read_lock();
435 		for_each_process(p)
436 			dump_task(p, oc);
437 		rcu_read_unlock();
438 	}
439 }
440 
441 static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
442 {
443 	/* one line summary of the oom killer context. */
444 	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
445 			oom_constraint_text[oc->constraint],
446 			nodemask_pr_args(oc->nodemask));
447 	cpuset_print_current_mems_allowed();
448 	mem_cgroup_print_oom_context(oc->memcg, victim);
449 	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
450 		from_kuid(&init_user_ns, task_uid(victim)));
451 }
452 
453 static void dump_header(struct oom_control *oc, struct task_struct *p)
454 {
455 	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
456 		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
457 			current->signal->oom_score_adj);
458 	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
459 		pr_warn("COMPACTION is disabled!!!\n");
460 
461 	dump_stack();
462 	if (is_memcg_oom(oc))
463 		mem_cgroup_print_oom_meminfo(oc->memcg);
464 	else {
465 		show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
466 		if (is_dump_unreclaim_slabs())
467 			dump_unreclaimable_slab();
468 	}
469 	if (sysctl_oom_dump_tasks)
470 		dump_tasks(oc);
471 	if (p)
472 		dump_oom_summary(oc, p);
473 }
474 
475 /*
476  * Number of OOM victims in flight
477  */
478 static atomic_t oom_victims = ATOMIC_INIT(0);
479 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
480 
481 static bool oom_killer_disabled __read_mostly;
482 
483 #define K(x) ((x) << (PAGE_SHIFT-10))
484 
485 /*
486  * task->mm can be NULL if the task is the exited group leader.  So to
487  * determine whether the task is using a particular mm, we examine all the
488  * task's threads: if one of those is using this mm then this task was also
489  * using it.
490  */
491 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
492 {
493 	struct task_struct *t;
494 
495 	for_each_thread(p, t) {
496 		struct mm_struct *t_mm = READ_ONCE(t->mm);
497 		if (t_mm)
498 			return t_mm == mm;
499 	}
500 	return false;
501 }
502 
503 #ifdef CONFIG_MMU
504 /*
505  * OOM Reaper kernel thread which tries to reap the memory used by the OOM
506  * victim (if that is possible) to help the OOM killer to move on.
507  */
508 static struct task_struct *oom_reaper_th;
509 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
510 static struct task_struct *oom_reaper_list;
511 static DEFINE_SPINLOCK(oom_reaper_lock);
512 
513 bool __oom_reap_task_mm(struct mm_struct *mm)
514 {
515 	struct vm_area_struct *vma;
516 	bool ret = true;
517 
518 	/*
519 	 * Tell all users of get_user/copy_from_user etc... that the content
520 	 * is no longer stable. No barriers really needed because unmapping
521 	 * should imply barriers already and the reader would hit a page fault
522 	 * if it stumbled over a reaped memory.
523 	 */
524 	set_bit(MMF_UNSTABLE, &mm->flags);
525 
526 	for (vma = mm->mmap ; vma; vma = vma->vm_next) {
527 		if (!can_madv_lru_vma(vma))
528 			continue;
529 
530 		/*
531 		 * Only anonymous pages have a good chance to be dropped
532 		 * without additional steps which we cannot afford as we
533 		 * are OOM already.
534 		 *
535 		 * We do not even care about fs backed pages because all
536 		 * which are reclaimable have already been reclaimed and
537 		 * we do not want to block exit_mmap by keeping mm ref
538 		 * count elevated without a good reason.
539 		 */
540 		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
541 			struct mmu_notifier_range range;
542 			struct mmu_gather tlb;
543 
544 			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
545 						vma, mm, vma->vm_start,
546 						vma->vm_end);
547 			tlb_gather_mmu(&tlb, mm, range.start, range.end);
548 			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
549 				tlb_finish_mmu(&tlb, range.start, range.end);
550 				ret = false;
551 				continue;
552 			}
553 			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
554 			mmu_notifier_invalidate_range_end(&range);
555 			tlb_finish_mmu(&tlb, range.start, range.end);
556 		}
557 	}
558 
559 	return ret;
560 }
561 
562 /*
563  * Reaps the address space of the give task.
564  *
565  * Returns true on success and false if none or part of the address space
566  * has been reclaimed and the caller should retry later.
567  */
568 static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
569 {
570 	bool ret = true;
571 
572 	if (!down_read_trylock(&mm->mmap_sem)) {
573 		trace_skip_task_reaping(tsk->pid);
574 		return false;
575 	}
576 
577 	/*
578 	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
579 	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
580 	 * under mmap_sem for reading because it serializes against the
581 	 * down_write();up_write() cycle in exit_mmap().
582 	 */
583 	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
584 		trace_skip_task_reaping(tsk->pid);
585 		goto out_unlock;
586 	}
587 
588 	trace_start_task_reaping(tsk->pid);
589 
590 	/* failed to reap part of the address space. Try again later */
591 	ret = __oom_reap_task_mm(mm);
592 	if (!ret)
593 		goto out_finish;
594 
595 	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
596 			task_pid_nr(tsk), tsk->comm,
597 			K(get_mm_counter(mm, MM_ANONPAGES)),
598 			K(get_mm_counter(mm, MM_FILEPAGES)),
599 			K(get_mm_counter(mm, MM_SHMEMPAGES)));
600 out_finish:
601 	trace_finish_task_reaping(tsk->pid);
602 out_unlock:
603 	up_read(&mm->mmap_sem);
604 
605 	return ret;
606 }
607 
608 #define MAX_OOM_REAP_RETRIES 10
609 static void oom_reap_task(struct task_struct *tsk)
610 {
611 	int attempts = 0;
612 	struct mm_struct *mm = tsk->signal->oom_mm;
613 
614 	/* Retry the down_read_trylock(mmap_sem) a few times */
615 	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
616 		schedule_timeout_idle(HZ/10);
617 
618 	if (attempts <= MAX_OOM_REAP_RETRIES ||
619 	    test_bit(MMF_OOM_SKIP, &mm->flags))
620 		goto done;
621 
622 	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
623 		task_pid_nr(tsk), tsk->comm);
624 	sched_show_task(tsk);
625 	debug_show_all_locks();
626 
627 done:
628 	tsk->oom_reaper_list = NULL;
629 
630 	/*
631 	 * Hide this mm from OOM killer because it has been either reaped or
632 	 * somebody can't call up_write(mmap_sem).
633 	 */
634 	set_bit(MMF_OOM_SKIP, &mm->flags);
635 
636 	/* Drop a reference taken by wake_oom_reaper */
637 	put_task_struct(tsk);
638 }
639 
640 static int oom_reaper(void *unused)
641 {
642 	while (true) {
643 		struct task_struct *tsk = NULL;
644 
645 		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
646 		spin_lock(&oom_reaper_lock);
647 		if (oom_reaper_list != NULL) {
648 			tsk = oom_reaper_list;
649 			oom_reaper_list = tsk->oom_reaper_list;
650 		}
651 		spin_unlock(&oom_reaper_lock);
652 
653 		if (tsk)
654 			oom_reap_task(tsk);
655 	}
656 
657 	return 0;
658 }
659 
660 static void wake_oom_reaper(struct task_struct *tsk)
661 {
662 	/* mm is already queued? */
663 	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
664 		return;
665 
666 	get_task_struct(tsk);
667 
668 	spin_lock(&oom_reaper_lock);
669 	tsk->oom_reaper_list = oom_reaper_list;
670 	oom_reaper_list = tsk;
671 	spin_unlock(&oom_reaper_lock);
672 	trace_wake_reaper(tsk->pid);
673 	wake_up(&oom_reaper_wait);
674 }
675 
676 static int __init oom_init(void)
677 {
678 	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
679 	return 0;
680 }
681 subsys_initcall(oom_init)
682 #else
683 static inline void wake_oom_reaper(struct task_struct *tsk)
684 {
685 }
686 #endif /* CONFIG_MMU */
687 
688 /**
689  * mark_oom_victim - mark the given task as OOM victim
690  * @tsk: task to mark
691  *
692  * Has to be called with oom_lock held and never after
693  * oom has been disabled already.
694  *
695  * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
696  * under task_lock or operate on the current).
697  */
698 static void mark_oom_victim(struct task_struct *tsk)
699 {
700 	struct mm_struct *mm = tsk->mm;
701 
702 	WARN_ON(oom_killer_disabled);
703 	/* OOM killer might race with memcg OOM */
704 	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
705 		return;
706 
707 	/* oom_mm is bound to the signal struct life time. */
708 	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
709 		mmgrab(tsk->signal->oom_mm);
710 		set_bit(MMF_OOM_VICTIM, &mm->flags);
711 	}
712 
713 	/*
714 	 * Make sure that the task is woken up from uninterruptible sleep
715 	 * if it is frozen because OOM killer wouldn't be able to free
716 	 * any memory and livelock. freezing_slow_path will tell the freezer
717 	 * that TIF_MEMDIE tasks should be ignored.
718 	 */
719 	__thaw_task(tsk);
720 	atomic_inc(&oom_victims);
721 	trace_mark_victim(tsk->pid);
722 }
723 
724 /**
725  * exit_oom_victim - note the exit of an OOM victim
726  */
727 void exit_oom_victim(void)
728 {
729 	clear_thread_flag(TIF_MEMDIE);
730 
731 	if (!atomic_dec_return(&oom_victims))
732 		wake_up_all(&oom_victims_wait);
733 }
734 
735 /**
736  * oom_killer_enable - enable OOM killer
737  */
738 void oom_killer_enable(void)
739 {
740 	oom_killer_disabled = false;
741 	pr_info("OOM killer enabled.\n");
742 }
743 
744 /**
745  * oom_killer_disable - disable OOM killer
746  * @timeout: maximum timeout to wait for oom victims in jiffies
747  *
748  * Forces all page allocations to fail rather than trigger OOM killer.
749  * Will block and wait until all OOM victims are killed or the given
750  * timeout expires.
751  *
752  * The function cannot be called when there are runnable user tasks because
753  * the userspace would see unexpected allocation failures as a result. Any
754  * new usage of this function should be consulted with MM people.
755  *
756  * Returns true if successful and false if the OOM killer cannot be
757  * disabled.
758  */
759 bool oom_killer_disable(signed long timeout)
760 {
761 	signed long ret;
762 
763 	/*
764 	 * Make sure to not race with an ongoing OOM killer. Check that the
765 	 * current is not killed (possibly due to sharing the victim's memory).
766 	 */
767 	if (mutex_lock_killable(&oom_lock))
768 		return false;
769 	oom_killer_disabled = true;
770 	mutex_unlock(&oom_lock);
771 
772 	ret = wait_event_interruptible_timeout(oom_victims_wait,
773 			!atomic_read(&oom_victims), timeout);
774 	if (ret <= 0) {
775 		oom_killer_enable();
776 		return false;
777 	}
778 	pr_info("OOM killer disabled.\n");
779 
780 	return true;
781 }
782 
783 static inline bool __task_will_free_mem(struct task_struct *task)
784 {
785 	struct signal_struct *sig = task->signal;
786 
787 	/*
788 	 * A coredumping process may sleep for an extended period in exit_mm(),
789 	 * so the oom killer cannot assume that the process will promptly exit
790 	 * and release memory.
791 	 */
792 	if (sig->flags & SIGNAL_GROUP_COREDUMP)
793 		return false;
794 
795 	if (sig->flags & SIGNAL_GROUP_EXIT)
796 		return true;
797 
798 	if (thread_group_empty(task) && (task->flags & PF_EXITING))
799 		return true;
800 
801 	return false;
802 }
803 
804 /*
805  * Checks whether the given task is dying or exiting and likely to
806  * release its address space. This means that all threads and processes
807  * sharing the same mm have to be killed or exiting.
808  * Caller has to make sure that task->mm is stable (hold task_lock or
809  * it operates on the current).
810  */
811 static bool task_will_free_mem(struct task_struct *task)
812 {
813 	struct mm_struct *mm = task->mm;
814 	struct task_struct *p;
815 	bool ret = true;
816 
817 	/*
818 	 * Skip tasks without mm because it might have passed its exit_mm and
819 	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
820 	 * on that for now. We can consider find_lock_task_mm in future.
821 	 */
822 	if (!mm)
823 		return false;
824 
825 	if (!__task_will_free_mem(task))
826 		return false;
827 
828 	/*
829 	 * This task has already been drained by the oom reaper so there are
830 	 * only small chances it will free some more
831 	 */
832 	if (test_bit(MMF_OOM_SKIP, &mm->flags))
833 		return false;
834 
835 	if (atomic_read(&mm->mm_users) <= 1)
836 		return true;
837 
838 	/*
839 	 * Make sure that all tasks which share the mm with the given tasks
840 	 * are dying as well to make sure that a) nobody pins its mm and
841 	 * b) the task is also reapable by the oom reaper.
842 	 */
843 	rcu_read_lock();
844 	for_each_process(p) {
845 		if (!process_shares_mm(p, mm))
846 			continue;
847 		if (same_thread_group(task, p))
848 			continue;
849 		ret = __task_will_free_mem(p);
850 		if (!ret)
851 			break;
852 	}
853 	rcu_read_unlock();
854 
855 	return ret;
856 }
857 
858 static void __oom_kill_process(struct task_struct *victim, const char *message)
859 {
860 	struct task_struct *p;
861 	struct mm_struct *mm;
862 	bool can_oom_reap = true;
863 
864 	p = find_lock_task_mm(victim);
865 	if (!p) {
866 		put_task_struct(victim);
867 		return;
868 	} else if (victim != p) {
869 		get_task_struct(p);
870 		put_task_struct(victim);
871 		victim = p;
872 	}
873 
874 	/* Get a reference to safely compare mm after task_unlock(victim) */
875 	mm = victim->mm;
876 	mmgrab(mm);
877 
878 	/* Raise event before sending signal: task reaper must see this */
879 	count_vm_event(OOM_KILL);
880 	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
881 
882 	/*
883 	 * We should send SIGKILL before granting access to memory reserves
884 	 * in order to prevent the OOM victim from depleting the memory
885 	 * reserves from the user space under its control.
886 	 */
887 	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
888 	mark_oom_victim(victim);
889 	pr_err("%s: Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB, UID:%u pgtables:%lukB oom_score_adj:%hd\n",
890 		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
891 		K(get_mm_counter(mm, MM_ANONPAGES)),
892 		K(get_mm_counter(mm, MM_FILEPAGES)),
893 		K(get_mm_counter(mm, MM_SHMEMPAGES)),
894 		from_kuid(&init_user_ns, task_uid(victim)),
895 		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
896 	task_unlock(victim);
897 
898 	/*
899 	 * Kill all user processes sharing victim->mm in other thread groups, if
900 	 * any.  They don't get access to memory reserves, though, to avoid
901 	 * depletion of all memory.  This prevents mm->mmap_sem livelock when an
902 	 * oom killed thread cannot exit because it requires the semaphore and
903 	 * its contended by another thread trying to allocate memory itself.
904 	 * That thread will now get access to memory reserves since it has a
905 	 * pending fatal signal.
906 	 */
907 	rcu_read_lock();
908 	for_each_process(p) {
909 		if (!process_shares_mm(p, mm))
910 			continue;
911 		if (same_thread_group(p, victim))
912 			continue;
913 		if (is_global_init(p)) {
914 			can_oom_reap = false;
915 			set_bit(MMF_OOM_SKIP, &mm->flags);
916 			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
917 					task_pid_nr(victim), victim->comm,
918 					task_pid_nr(p), p->comm);
919 			continue;
920 		}
921 		/*
922 		 * No use_mm() user needs to read from the userspace so we are
923 		 * ok to reap it.
924 		 */
925 		if (unlikely(p->flags & PF_KTHREAD))
926 			continue;
927 		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
928 	}
929 	rcu_read_unlock();
930 
931 	if (can_oom_reap)
932 		wake_oom_reaper(victim);
933 
934 	mmdrop(mm);
935 	put_task_struct(victim);
936 }
937 #undef K
938 
939 /*
940  * Kill provided task unless it's secured by setting
941  * oom_score_adj to OOM_SCORE_ADJ_MIN.
942  */
943 static int oom_kill_memcg_member(struct task_struct *task, void *message)
944 {
945 	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
946 	    !is_global_init(task)) {
947 		get_task_struct(task);
948 		__oom_kill_process(task, message);
949 	}
950 	return 0;
951 }
952 
953 static void oom_kill_process(struct oom_control *oc, const char *message)
954 {
955 	struct task_struct *victim = oc->chosen;
956 	struct mem_cgroup *oom_group;
957 	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
958 					      DEFAULT_RATELIMIT_BURST);
959 
960 	/*
961 	 * If the task is already exiting, don't alarm the sysadmin or kill
962 	 * its children or threads, just give it access to memory reserves
963 	 * so it can die quickly
964 	 */
965 	task_lock(victim);
966 	if (task_will_free_mem(victim)) {
967 		mark_oom_victim(victim);
968 		wake_oom_reaper(victim);
969 		task_unlock(victim);
970 		put_task_struct(victim);
971 		return;
972 	}
973 	task_unlock(victim);
974 
975 	if (__ratelimit(&oom_rs))
976 		dump_header(oc, victim);
977 
978 	/*
979 	 * Do we need to kill the entire memory cgroup?
980 	 * Or even one of the ancestor memory cgroups?
981 	 * Check this out before killing the victim task.
982 	 */
983 	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
984 
985 	__oom_kill_process(victim, message);
986 
987 	/*
988 	 * If necessary, kill all tasks in the selected memory cgroup.
989 	 */
990 	if (oom_group) {
991 		mem_cgroup_print_oom_group(oom_group);
992 		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
993 				      (void*)message);
994 		mem_cgroup_put(oom_group);
995 	}
996 }
997 
998 /*
999  * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1000  */
1001 static void check_panic_on_oom(struct oom_control *oc)
1002 {
1003 	if (likely(!sysctl_panic_on_oom))
1004 		return;
1005 	if (sysctl_panic_on_oom != 2) {
1006 		/*
1007 		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1008 		 * does not panic for cpuset, mempolicy, or memcg allocation
1009 		 * failures.
1010 		 */
1011 		if (oc->constraint != CONSTRAINT_NONE)
1012 			return;
1013 	}
1014 	/* Do not panic for oom kills triggered by sysrq */
1015 	if (is_sysrq_oom(oc))
1016 		return;
1017 	dump_header(oc, NULL);
1018 	panic("Out of memory: %s panic_on_oom is enabled\n",
1019 		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1020 }
1021 
1022 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1023 
1024 int register_oom_notifier(struct notifier_block *nb)
1025 {
1026 	return blocking_notifier_chain_register(&oom_notify_list, nb);
1027 }
1028 EXPORT_SYMBOL_GPL(register_oom_notifier);
1029 
1030 int unregister_oom_notifier(struct notifier_block *nb)
1031 {
1032 	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1033 }
1034 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1035 
1036 /**
1037  * out_of_memory - kill the "best" process when we run out of memory
1038  * @oc: pointer to struct oom_control
1039  *
1040  * If we run out of memory, we have the choice between either
1041  * killing a random task (bad), letting the system crash (worse)
1042  * OR try to be smart about which process to kill. Note that we
1043  * don't have to be perfect here, we just have to be good.
1044  */
1045 bool out_of_memory(struct oom_control *oc)
1046 {
1047 	unsigned long freed = 0;
1048 
1049 	if (oom_killer_disabled)
1050 		return false;
1051 
1052 	if (!is_memcg_oom(oc)) {
1053 		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1054 		if (freed > 0)
1055 			/* Got some memory back in the last second. */
1056 			return true;
1057 	}
1058 
1059 	/*
1060 	 * If current has a pending SIGKILL or is exiting, then automatically
1061 	 * select it.  The goal is to allow it to allocate so that it may
1062 	 * quickly exit and free its memory.
1063 	 */
1064 	if (task_will_free_mem(current)) {
1065 		mark_oom_victim(current);
1066 		wake_oom_reaper(current);
1067 		return true;
1068 	}
1069 
1070 	/*
1071 	 * The OOM killer does not compensate for IO-less reclaim.
1072 	 * pagefault_out_of_memory lost its gfp context so we have to
1073 	 * make sure exclude 0 mask - all other users should have at least
1074 	 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1075 	 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1076 	 */
1077 	if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1078 		return true;
1079 
1080 	/*
1081 	 * Check if there were limitations on the allocation (only relevant for
1082 	 * NUMA and memcg) that may require different handling.
1083 	 */
1084 	oc->constraint = constrained_alloc(oc);
1085 	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1086 		oc->nodemask = NULL;
1087 	check_panic_on_oom(oc);
1088 
1089 	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1090 	    current->mm && !oom_unkillable_task(current) &&
1091 	    oom_cpuset_eligible(current, oc) &&
1092 	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1093 		get_task_struct(current);
1094 		oc->chosen = current;
1095 		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1096 		return true;
1097 	}
1098 
1099 	select_bad_process(oc);
1100 	/* Found nothing?!?! */
1101 	if (!oc->chosen) {
1102 		dump_header(oc, NULL);
1103 		pr_warn("Out of memory and no killable processes...\n");
1104 		/*
1105 		 * If we got here due to an actual allocation at the
1106 		 * system level, we cannot survive this and will enter
1107 		 * an endless loop in the allocator. Bail out now.
1108 		 */
1109 		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1110 			panic("System is deadlocked on memory\n");
1111 	}
1112 	if (oc->chosen && oc->chosen != (void *)-1UL)
1113 		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1114 				 "Memory cgroup out of memory");
1115 	return !!oc->chosen;
1116 }
1117 
1118 /*
1119  * The pagefault handler calls here because it is out of memory, so kill a
1120  * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1121  * killing is already in progress so do nothing.
1122  */
1123 void pagefault_out_of_memory(void)
1124 {
1125 	struct oom_control oc = {
1126 		.zonelist = NULL,
1127 		.nodemask = NULL,
1128 		.memcg = NULL,
1129 		.gfp_mask = 0,
1130 		.order = 0,
1131 	};
1132 
1133 	if (mem_cgroup_oom_synchronize(true))
1134 		return;
1135 
1136 	if (!mutex_trylock(&oom_lock))
1137 		return;
1138 	out_of_memory(&oc);
1139 	mutex_unlock(&oom_lock);
1140 }
1141